From f09b9b0ab706cfe25994a7ed35066e61e5af5d8f Mon Sep 17 00:00:00 2001 From: nikochiko Date: Mon, 18 Nov 2019 23:36:54 +0530 Subject: [PATCH 001/867] Revert "Update save.py" This reverts commit fdadd0e5e524df6488cd763c4ab7595d469ed1ef. --- tensorflow/python/keras/saving/save.py | 44 +++++--------------------- 1 file changed, 8 insertions(+), 36 deletions(-) diff --git a/tensorflow/python/keras/saving/save.py b/tensorflow/python/keras/saving/save.py index 9f7f5778afe..4be3aa0bbda 100644 --- a/tensorflow/python/keras/saving/save.py +++ b/tensorflow/python/keras/saving/save.py @@ -23,7 +23,6 @@ import os import six from tensorflow.python import tf2 -from tensorflow.python.keras.engine.network import _is_hdf5_filepath from tensorflow.python.keras.saving import hdf5_format from tensorflow.python.keras.saving.saved_model import load as saved_model_load from tensorflow.python.keras.saving.saved_model import save as saved_model_save @@ -37,6 +36,9 @@ except ImportError: h5py = None # pylint: enable=g-import-not-at-top +_HDF5_EXTENSIONS = ['.h5', '.hdf5', '.keras'] + + # TODO(kathywu): Remove this when Keras SavedModel is not experimental. _KERAS_SAVED_MODEL_STILL_EXPERIMENTAL = True @@ -90,42 +92,12 @@ def save_model(model, """ from tensorflow.python.keras.engine import sequential # pylint: disable=g-import-not-at-top - if type(filepath) != str and not isinstance(filepath, h5py.File): - raise ValueError( - 'Expected `filepath` to be a String or `h5py.File` object. Got' - 'unsupported value %s of type %s' - % (filepath, type(filepath))) + default_format = 'tf' if tf2.enabled() else 'h5' + save_format = save_format or default_format - filepath_is_h5py_file = h5py is not None and isinstance(filepath, h5py.File) - filepath_is_h5 = type(filepath) == str and _is_hdf5_filepath(filepath) - if save_format is None: - if (filepath_is_h5 or - (filepath_is_h5py_file)): - save_format = 'h5' - else: - save_format = 'tf' if tf2.enabled() else 'h5' - else: - user_format = save_format.lower().strip() - if user_format in ('tensorflow', 'tf'): - save_format = 'tf' - elif user_format in ('hdf5', 'h5', 'keras'): - save_format = 'h5' - else: - raise ValueError( - 'Unknown format "%s". Was expecting one of {"tf", "h5"}.' % ( - save_format,)) - if save_format == 'tf' and filepath_is_h5: - raise ValueError( - ('`save` got save_format="tf"/"tensorflow", but the ' - 'filepath ("%s") looks like an HDF5 file. Omit the ".h5"/".keras" ' - 'when saving in TensorFlow format.') - % filepath) - if save_format == 'tf' and filepath_is_h5py_file: - raise ValueError( - '`save` got save_format="tf"/"tensorflow", but the given `filepath`' - 'is an `h5py.File` object.') - - if save_format == 'h5': + if (save_format == 'h5' or + (h5py is not None and isinstance(filepath, h5py.File)) or + os.path.splitext(filepath)[1] in _HDF5_EXTENSIONS): # TODO(b/130258301): add utility method for detecting model type. if (not model._is_graph_network and # pylint:disable=protected-access not isinstance(model, sequential.Sequential)): From ae86fd4b96ee934a8af239b5d01381bba8a470f1 Mon Sep 17 00:00:00 2001 From: nikochiko Date: Wed, 20 Nov 2019 15:48:54 +0530 Subject: [PATCH 002/867] Update docstring for tf.ensure_shape Update docstring Enhance example. --- tensorflow/python/ops/check_ops.py | 42 ++++++++++++++++++------------ 1 file changed, 25 insertions(+), 17 deletions(-) diff --git a/tensorflow/python/ops/check_ops.py b/tensorflow/python/ops/check_ops.py index 34106f61fd8..68546d792a8 100644 --- a/tensorflow/python/ops/check_ops.py +++ b/tensorflow/python/ops/check_ops.py @@ -144,7 +144,7 @@ def _unary_assert_doc(sym, sym_name): Raises: InvalidArgumentError: if the check can be performed immediately and - `x {sym}` is False. The check can be performed immediately during + `x {sym}` is False. The check can be performed immediately during eager execution or if `x` is statically known. """.format( sym=sym, sym_name=cap_sym_name, opname=opname) @@ -207,7 +207,7 @@ def _binary_assert_doc(sym): Raises: InvalidArgumentError: if the check can be performed immediately and - `x {sym} y` is False. The check can be performed immediately during + `x {sym} y` is False. The check can be performed immediately during eager execution or if `x` and `y` are statically known. """.format( sym=sym, opname=opname) @@ -2107,24 +2107,30 @@ def ensure_shape(x, shape, name=None): """Updates the shape of a tensor and checks at runtime that the shape holds. For example: - ```python - x = tf.compat.v1.placeholder(tf.int32) - print(x.shape) - ==> TensorShape(None) - y = x * 2 - print(y.shape) - ==> TensorShape(None) + + >>> # tf.placeholder() is not compatible with eager execution + ... + >>> tf.compat.v1.disable_eager_execution() + >>> x = tf.compat.v1.placeholder(tf.int32) + >>> print(x.shape) + TensorShape(None) + >>> y = x * 2 + >>> print(y.shape) + TensorShape(None) y = tf.ensure_shape(y, (None, 3, 3)) print(y.shape) - ==> TensorShape([Dimension(None), Dimension(3), Dimension(3)]) + TensorShape([None, 3, 3]) - with tf.compat.v1.Session() as sess: - # Raises tf.errors.InvalidArgumentError, because the shape (3,) is not - # compatible with the shape (None, 3, 3) - sess.run(y, feed_dict={x: [1, 2, 3]}) + >>> with tf.compat.v1.Session() as sess: + >>> sess.run(y, feed_dict={x: [1, 2, 3]}) + Traceback (most recent call last): + ... + InvalidArgumentError: Shape of tensor mul [3] is not compatible with + expected shape [?,3,3]. - ``` + The above example raises `tf.errors.InvalidArgumentError`, + because the shape (3,) is not compatible with the shape (None, 3, 3) NOTE: This differs from `Tensor.set_shape` in that it sets the static shape of the resulting tensor and enforces it at runtime, raising an error if the @@ -2140,8 +2146,10 @@ def ensure_shape(x, shape, name=None): name: A name for this operation (optional). Defaults to "EnsureShape". Returns: - A `Tensor`. Has the same type and contents as `x`. At runtime, raises a - `tf.errors.InvalidArgumentError` if `shape` is incompatible with the shape + A `Tensor`. Has the same type and contents as `x`. + + Raises: + tf.errors.InvalidArgumentError: If `shape` is incompatible with the shape of `x`. """ if not isinstance(shape, tensor_shape.TensorShape): From c6281f660b3f593054737de0bc9388113181aa4d Mon Sep 17 00:00:00 2001 From: nikochiko Date: Thu, 21 Nov 2019 00:06:38 +0530 Subject: [PATCH 003/867] Make requested changes --- tensorflow/python/ops/array_ops.py | 14 +++++++------- 1 file changed, 7 insertions(+), 7 deletions(-) diff --git a/tensorflow/python/ops/array_ops.py b/tensorflow/python/ops/array_ops.py index 6a18d08f22f..b31c28deb28 100644 --- a/tensorflow/python/ops/array_ops.py +++ b/tensorflow/python/ops/array_ops.py @@ -444,7 +444,7 @@ def broadcast_dynamic_shape(shape_x, shape_y): >>> shape_x = [1, 2, 3] >>> shape_y = [5, 1, 3] - >>> broadcast_dynamic_shape(shape_x, shape_y) + >>> tf.broadcast_dynamic_shape(shape_x, shape_y) Args: @@ -479,7 +479,7 @@ def broadcast_static_shape(shape_x, shape_y): >>> shape_x = tf.TensorShape([1, 2, 3]) >>> shape_y = tf.TensorShape([5, 1 ,3]) - >>> broadcast_static_shape(shape_x, shape_y) + >>> tf.broadcast_static_shape(shape_x, shape_y) TensorShape([Dimension(5), Dimension(2), Dimension(3)]) Args: @@ -1556,12 +1556,12 @@ def boolean_mask(tensor, mask, name="boolean_mask", axis=None): # 1-D example tensor = [0, 1, 2, 3] mask = np.array([True, False, True, False]) - boolean_mask(tensor, mask) # [0, 2] + tf.boolean_mask(tensor, mask) # [0, 2] # 2-D example tensor = [[1, 2], [3, 4], [5, 6]] mask = np.array([True, False, True]) - boolean_mask(tensor, mask) # [[1, 2], [5, 6]] + tf.boolean_mask(tensor, mask) # [[1, 2], [5, 6]] ``` Args: @@ -1640,12 +1640,12 @@ def boolean_mask_v2(tensor, mask, axis=None, name="boolean_mask"): >>> tensor = [0, 1, 2, 3] # 1-D example >>> mask = np.array([True, False, True, False]) - >>> boolean_mask(tensor, mask) + >>> tf.boolean_mask(tensor, mask) >>> tensor = [[1, 2], [3, 4], [5, 6]] # 2-D example >>> mask = np.array([True, False, True]) - >>> boolean_mask(tensor, mask) + >>> tf.boolean_mask(tensor, mask) Args: @@ -3175,7 +3175,7 @@ def edit_distance(hypothesis, truth, normalize=True, name="edit_distance"): ... [1, 1, 0]], ... ["a", "b", "c", "a"], ... (2, 2, 2)) - >>> edit_distance(hypothesis, truth, normalize=True) + >>> tf.edit_distance(hypothesis, truth, normalize=True) From d6cff22e271692d770ebc94a8aec6b83e2d533b9 Mon Sep 17 00:00:00 2001 From: nikochiko Date: Thu, 21 Nov 2019 14:47:02 +0530 Subject: [PATCH 004/867] Make requested changes - Shapes from lists to tuples - TensorShapes to `TensorShape([...])` --- tensorflow/python/ops/array_ops.py | 21 +++++++++++---------- tensorflow/python/ops/check_ops.py | 10 ++++------ 2 files changed, 15 insertions(+), 16 deletions(-) diff --git a/tensorflow/python/ops/array_ops.py b/tensorflow/python/ops/array_ops.py index b31c28deb28..a345d290f69 100644 --- a/tensorflow/python/ops/array_ops.py +++ b/tensorflow/python/ops/array_ops.py @@ -442,8 +442,8 @@ def broadcast_dynamic_shape(shape_x, shape_y): Example: - >>> shape_x = [1, 2, 3] - >>> shape_y = [5, 1, 3] + >>> shape_x = (1, 2, 3) + >>> shape_y = (5, 1, 3) >>> tf.broadcast_dynamic_shape(shape_x, shape_y) @@ -469,8 +469,9 @@ def broadcast_static_shape(shape_x, shape_y): `TensorShape` which is the shape of the result of a broadcasting op applied in tensors of shapes `shape_x` and `shape_y`. - For example, if shape_x is [1, 2, 3] and shape_y is [5, 1, 3], the result is a - TensorShape whose value is [5, 2, 3]. + For example, if shape_x is `TensorShape([1, 2, 3])` and shape_y is + `TensorShape([5, 1, 3])`, the result is a TensorShape whose value is + `TensorShape([5, 2, 3])`. This is useful when validating the result of a broadcasting operation when the tensors have statically known shapes. @@ -3160,8 +3161,8 @@ def edit_distance(hypothesis, truth, normalize=True, name="edit_distance"): For example: Given the following input, - * `hypothesis` is a `tf.SparseTensor` of shape `[2, 1, 1]` - * `truth` is a `tf.SparseTensor` of shape `[2, 2, 2]` + * `hypothesis` is a `tf.SparseTensor` of shape `(2, 1, 1)` + * `truth` is a `tf.SparseTensor` of shape `(2, 2, 2)` >>> hypothesis = tf.SparseTensor( ... [[0, 0, 0], @@ -3180,7 +3181,7 @@ def edit_distance(hypothesis, truth, normalize=True, name="edit_distance"): array([[inf, 1. ], [0.5, 1. ]], dtype=float32)> - The operaton returns a dense Tensor of shape `[2, 2]` with + The operaton returns a dense Tensor of shape `(2, 2)` with edit distances normalized by `truth` lengths. **Note**: It is possible to calculate edit distance between two @@ -3190,14 +3191,14 @@ def edit_distance(hypothesis, truth, normalize=True, name="edit_distance"): For the following inputs, ```python - # 'hypothesis' is a tensor of shape `[2, 1]` with variable-length values: + # 'hypothesis' is a tensor of shape `(2, 1)` with variable-length values: hypothesis = tf.SparseTensor( [[0, 0], [1,0]], ["a", "b"], (2, 1)) - # 'truth' is a tensor of shape `[2, 2]` with variable-length values: + # 'truth' is a tensor of shape `(2, 2)` with variable-length values: truth = tf.SparseTensor( [[0, 1], [1, 0], @@ -3207,7 +3208,7 @@ def edit_distance(hypothesis, truth, normalize=True, name="edit_distance"): normalize = True - # The output would be a dense Tensor of shape `[2,]`, with edit distances + # The output would be a dense Tensor of shape `(2,)`, with edit distances noramlized by 'truth' lengths. # output => array([0., 0.5], dtype=float32) ``` diff --git a/tensorflow/python/ops/check_ops.py b/tensorflow/python/ops/check_ops.py index 68546d792a8..c76f3939f1e 100644 --- a/tensorflow/python/ops/check_ops.py +++ b/tensorflow/python/ops/check_ops.py @@ -2117,13 +2117,11 @@ def ensure_shape(x, shape, name=None): >>> y = x * 2 >>> print(y.shape) TensorShape(None) - - y = tf.ensure_shape(y, (None, 3, 3)) - print(y.shape) - TensorShape([None, 3, 3]) - + >>> y = tf.ensure_shape(y, (None, 3, 3)) + >>> print(y.shape) + TensorShape([Dimension(None), Dimension(3), Dimension(3)]) >>> with tf.compat.v1.Session() as sess: - >>> sess.run(y, feed_dict={x: [1, 2, 3]}) + >>> sess.run(y, feed_dict={x: [1, 2, 3]}) Traceback (most recent call last): ... InvalidArgumentError: Shape of tensor mul [3] is not compatible with From 9ccf3973b0af7232e83d18a8ea6c6fc27632f0fc Mon Sep 17 00:00:00 2001 From: nikochiko Date: Tue, 26 Nov 2019 21:47:35 +0530 Subject: [PATCH 005/867] Minor changes Made few more minor changes to documentation of `tf.extract_volume_patches` and `tf.fill` --- .../api_def_ExtractVolumePatches.pbtxt | 8 +++---- tensorflow/python/ops/array_ops.py | 21 ++++++++++--------- 2 files changed, 15 insertions(+), 14 deletions(-) diff --git a/tensorflow/core/api_def/base_api/api_def_ExtractVolumePatches.pbtxt b/tensorflow/core/api_def/base_api/api_def_ExtractVolumePatches.pbtxt index 9c4015eaa4c..32a3c33dc35 100644 --- a/tensorflow/core/api_def/base_api/api_def_ExtractVolumePatches.pbtxt +++ b/tensorflow/core/api_def/base_api/api_def_ExtractVolumePatches.pbtxt @@ -34,13 +34,13 @@ END description: < Date: Sat, 14 Dec 2019 22:23:21 +0530 Subject: [PATCH 006/867] Update example to use tf.function --- tensorflow/python/ops/check_ops.py | 27 +++++++++------------------ 1 file changed, 9 insertions(+), 18 deletions(-) diff --git a/tensorflow/python/ops/check_ops.py b/tensorflow/python/ops/check_ops.py index e53a00bc232..ed28b1b093a 100644 --- a/tensorflow/python/ops/check_ops.py +++ b/tensorflow/python/ops/check_ops.py @@ -2121,25 +2121,16 @@ def ensure_shape(x, shape, name=None): For example: - >>> # tf.placeholder() is not compatible with eager execution - ... - >>> tf.compat.v1.disable_eager_execution() - >>> x = tf.compat.v1.placeholder(tf.int32) - >>> print(x.shape) - TensorShape(None) - >>> y = x * 2 - >>> print(y.shape) - TensorShape(None) - >>> y = tf.ensure_shape(y, (None, 3, 3)) - >>> print(y.shape) - TensorShape([Dimension(None), Dimension(3), Dimension(3)]) - >>> with tf.compat.v1.Session() as sess: - >>> sess.run(y, feed_dict={x: [1, 2, 3]}) + >>> @tf.function(input_signature=[tf.TensorSpec(dtype=tf.float32, shape=None)]) + >>> def f(tensor): + >>> return tf.ensure_shape(x, [3, 3]) + >>> + >>> f(tf.zeros([3, 3])) # Passes + >>> f([1, 2, 3]) # fails Traceback (most recent call last): - ... - InvalidArgumentError: Shape of tensor mul [3] is not compatible with - expected shape [?,3,3]. - + ... + InvalidArgumentError: Shape of tensor x [3] is not compatible with expected shape [3,3]. + The above example raises `tf.errors.InvalidArgumentError`, because the shape (3,) is not compatible with the shape (None, 3, 3) From e96d18ac0fa564b666db80bba1e044fb88d7251a Mon Sep 17 00:00:00 2001 From: Stephan Uphoff Date: Sun, 3 Nov 2019 17:11:03 -0700 Subject: [PATCH 007/867] Set g_pdm_dma_error_reporter pointer to the error_reporter passed to InitAudioRecording() --- .../micro/examples/micro_speech/apollo3evb/audio_provider.cc | 4 ++-- 1 file changed, 2 insertions(+), 2 deletions(-) diff --git a/tensorflow/lite/micro/examples/micro_speech/apollo3evb/audio_provider.cc b/tensorflow/lite/micro/examples/micro_speech/apollo3evb/audio_provider.cc index 0f9a91a9dba..cf10785071b 100644 --- a/tensorflow/lite/micro/examples/micro_speech/apollo3evb/audio_provider.cc +++ b/tensorflow/lite/micro/examples/micro_speech/apollo3evb/audio_provider.cc @@ -247,7 +247,6 @@ void pdm_start_dma(tflite::ErrorReporter* error_reporter) { // Reset the PDM DMA flags. g_pdm_dma_error = false; - g_pdm_dma_error_reporter = error_reporter; } #if USE_MAYA @@ -460,11 +459,12 @@ TfLiteStatus InitAudioRecording(tflite::ErrorReporter* error_reporter) { #endif // USE_TIME_STAMP // Configure, turn on PDM + g_pdm_dma_error_reporter = error_reporter; pdm_init(); am_hal_interrupt_master_enable(); am_hal_pdm_fifo_flush(g_pdm_handle); // Trigger the PDM DMA for the first time manually. - pdm_start_dma(g_pdm_dma_error_reporter); + pdm_start_dma(error_reporter); error_reporter->Report("\nPDM DMA Threshold = %d", PDMn(0)->FIFOTHR); From 6d6411b339ef2b14f38c57e38c05b2f24122364a Mon Sep 17 00:00:00 2001 From: nikochiko Date: Tue, 17 Dec 2019 16:50:45 +0530 Subject: [PATCH 008/867] Fix sanity build errors --- tensorflow/python/ops/array_ops.py | 14 ++++++++------ 1 file changed, 8 insertions(+), 6 deletions(-) diff --git a/tensorflow/python/ops/array_ops.py b/tensorflow/python/ops/array_ops.py index 52c265608fe..65a55b5b2bd 100644 --- a/tensorflow/python/ops/array_ops.py +++ b/tensorflow/python/ops/array_ops.py @@ -490,10 +490,10 @@ setdiff1d.__doc__ = gen_array_ops.list_diff.__doc__ def broadcast_dynamic_shape(shape_x, shape_y): """Computes the shape of a broadcast given symbolic shapes. - When `shape_x` and `shape_y` are Tensors representing shapes (i.e. the result of - calling tf.shape on another Tensor) this computes a Tensor which is the shape - of the result of a broadcasting op applied in tensors of shapes `shape_x` and - `shape_y`. + When `shape_x` and `shape_y` are Tensors representing shapes (i.e. the result + of calling tf.shape on another Tensor) this computes a Tensor which is the + shape of the result of a broadcasting op applied in tensors of shapes + `shape_x` and `shape_y`. This is useful when validating the result of a broadcasting operation when the tensors do not have statically known shapes. @@ -503,7 +503,8 @@ def broadcast_dynamic_shape(shape_x, shape_y): >>> shape_x = (1, 2, 3) >>> shape_y = (5, 1, 3) >>> tf.broadcast_dynamic_shape(shape_x, shape_y) - + Args: shape_x: A rank 1 integer `Tensor`, representing the shape of x. @@ -1732,7 +1733,8 @@ def boolean_mask_v2(tensor, mask, axis=None, name="boolean_mask"): >>> tensor = [[1, 2], [3, 4], [5, 6]] # 2-D example >>> mask = np.array([True, False, True]) >>> tf.boolean_mask(tensor, mask) - + Args: tensor: N-D Tensor. From 2f3f3ee15bbc837a22b33c6177e064b5b92d5fec Mon Sep 17 00:00:00 2001 From: nikochiko Date: Tue, 17 Dec 2019 16:58:07 +0530 Subject: [PATCH 009/867] Fix pylint, wrap at 80 chars --- tensorflow/python/ops/array_ops.py | 8 ++++---- 1 file changed, 4 insertions(+), 4 deletions(-) diff --git a/tensorflow/python/ops/array_ops.py b/tensorflow/python/ops/array_ops.py index 65a55b5b2bd..16dda049420 100644 --- a/tensorflow/python/ops/array_ops.py +++ b/tensorflow/python/ops/array_ops.py @@ -503,8 +503,8 @@ def broadcast_dynamic_shape(shape_x, shape_y): >>> shape_x = (1, 2, 3) >>> shape_y = (5, 1, 3) >>> tf.broadcast_dynamic_shape(shape_x, shape_y) - + Args: shape_x: A rank 1 integer `Tensor`, representing the shape of x. @@ -1733,8 +1733,8 @@ def boolean_mask_v2(tensor, mask, axis=None, name="boolean_mask"): >>> tensor = [[1, 2], [3, 4], [5, 6]] # 2-D example >>> mask = np.array([True, False, True]) >>> tf.boolean_mask(tensor, mask) - + Args: tensor: N-D Tensor. From d25cff86db264b870439dbf2b6f37ae30096af0b Mon Sep 17 00:00:00 2001 From: Kaustubh Maske Patil <37668193+nikochiko@users.noreply.github.com> Date: Sat, 21 Dec 2019 20:25:12 +0530 Subject: [PATCH 010/867] Update array_ops.py --- tensorflow/python/ops/array_ops.py | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/tensorflow/python/ops/array_ops.py b/tensorflow/python/ops/array_ops.py index 16dda049420..d980babf99d 100644 --- a/tensorflow/python/ops/array_ops.py +++ b/tensorflow/python/ops/array_ops.py @@ -503,7 +503,7 @@ def broadcast_dynamic_shape(shape_x, shape_y): >>> shape_x = (1, 2, 3) >>> shape_y = (5, 1, 3) >>> tf.broadcast_dynamic_shape(shape_x, shape_y) - Args: From 6746a6b8317313d952c8c8083be2476c75d7cc23 Mon Sep 17 00:00:00 2001 From: Kaustubh Maske Patil <37668193+nikochiko@users.noreply.github.com> Date: Sat, 21 Dec 2019 20:26:41 +0530 Subject: [PATCH 011/867] Update array_ops.py --- tensorflow/python/ops/array_ops.py | 6 +++--- 1 file changed, 3 insertions(+), 3 deletions(-) diff --git a/tensorflow/python/ops/array_ops.py b/tensorflow/python/ops/array_ops.py index d980babf99d..57d953481a9 100644 --- a/tensorflow/python/ops/array_ops.py +++ b/tensorflow/python/ops/array_ops.py @@ -1728,12 +1728,12 @@ def boolean_mask_v2(tensor, mask, axis=None, name="boolean_mask"): >>> tensor = [0, 1, 2, 3] # 1-D example >>> mask = np.array([True, False, True, False]) >>> tf.boolean_mask(tensor, mask) - + >>> tensor = [[1, 2], [3, 4], [5, 6]] # 2-D example >>> mask = np.array([True, False, True]) >>> tf.boolean_mask(tensor, mask) - Args: @@ -3459,7 +3459,7 @@ def edit_distance(hypothesis, truth, normalize=True, name="edit_distance"): ... ["a", "b", "c", "a"], ... (2, 2, 2)) >>> tf.edit_distance(hypothesis, truth, normalize=True) - From 6a89b01c85893ecba281447e4d54c77889cd4d83 Mon Sep 17 00:00:00 2001 From: nikochiko Date: Sat, 4 Jan 2020 13:01:11 +0530 Subject: [PATCH 012/867] Fix docstrings --- tensorflow/python/ops/array_ops.py | 7 ++++--- tensorflow/python/ops/check_ops.py | 14 +++++++------- 2 files changed, 11 insertions(+), 10 deletions(-) diff --git a/tensorflow/python/ops/array_ops.py b/tensorflow/python/ops/array_ops.py index 57d953481a9..a208272a7e2 100644 --- a/tensorflow/python/ops/array_ops.py +++ b/tensorflow/python/ops/array_ops.py @@ -540,7 +540,7 @@ def broadcast_static_shape(shape_x, shape_y): >>> shape_x = tf.TensorShape([1, 2, 3]) >>> shape_y = tf.TensorShape([5, 1 ,3]) >>> tf.broadcast_static_shape(shape_x, shape_y) - TensorShape([Dimension(5), Dimension(2), Dimension(3)]) + TensorShape([5, 2, 3]) Args: shape_x: A `TensorShape` @@ -1733,8 +1733,9 @@ def boolean_mask_v2(tensor, mask, axis=None, name="boolean_mask"): >>> tensor = [[1, 2], [3, 4], [5, 6]] # 2-D example >>> mask = np.array([True, False, True]) >>> tf.boolean_mask(tensor, mask) - + Args: tensor: N-D Tensor. diff --git a/tensorflow/python/ops/check_ops.py b/tensorflow/python/ops/check_ops.py index ed28b1b093a..2088dea7c12 100644 --- a/tensorflow/python/ops/check_ops.py +++ b/tensorflow/python/ops/check_ops.py @@ -1595,7 +1595,7 @@ def assert_shapes_v2(shapes, data=None, summarize=None, message=None, >>> n = 10 >>> q = 3 >>> d = 7 - >>> x = tf.zeros([n,q]) + >>> x = tf.zeros([n,q]) >>> y = tf.ones([n,d]) >>> param = tf.Variable([1.0, 2.0, 3.0]) >>> scalar = 1.0 @@ -1605,9 +1605,9 @@ def assert_shapes_v2(shapes, data=None, summarize=None, message=None, ... (param, ('Q',)), ... (scalar, ()), ... ]) - + >>> tf.debugging.assert_shapes([ - ... (x, ('N', 'D')), + ... (x, ('N', 'D')), ... (y, ('N', 'D')) ... ]) Traceback (most recent call last): @@ -2121,16 +2121,16 @@ def ensure_shape(x, shape, name=None): For example: - >>> @tf.function(input_signature=[tf.TensorSpec(dtype=tf.float32, shape=None)]) - >>> def f(tensor): - >>> return tf.ensure_shape(x, [3, 3]) + >>> @tf.function(input_signature=[tf.TensorSpec(shape=None, dtype=tf.float32)]) + ... def f(tensor): + ... return tf.ensure_shape(tensor, [3, 3]) >>> >>> f(tf.zeros([3, 3])) # Passes >>> f([1, 2, 3]) # fails Traceback (most recent call last): ... InvalidArgumentError: Shape of tensor x [3] is not compatible with expected shape [3,3]. - + The above example raises `tf.errors.InvalidArgumentError`, because the shape (3,) is not compatible with the shape (None, 3, 3) From aee9e604872ecb6bdd15cd6a1864b2d1108b671b Mon Sep 17 00:00:00 2001 From: Kaustubh Maske Patil <37668193+nikochiko@users.noreply.github.com> Date: Sun, 12 Jan 2020 10:41:37 +0530 Subject: [PATCH 013/867] Fix doctest for ensure_shape --- tensorflow/python/ops/check_ops.py | 4 ++++ 1 file changed, 4 insertions(+) diff --git a/tensorflow/python/ops/check_ops.py b/tensorflow/python/ops/check_ops.py index 2088dea7c12..d38b3958205 100644 --- a/tensorflow/python/ops/check_ops.py +++ b/tensorflow/python/ops/check_ops.py @@ -2126,6 +2126,10 @@ def ensure_shape(x, shape, name=None): ... return tf.ensure_shape(tensor, [3, 3]) >>> >>> f(tf.zeros([3, 3])) # Passes + >>> f([1, 2, 3]) # fails Traceback (most recent call last): ... From a7899d7544230fce8dae4895733d82623af2b934 Mon Sep 17 00:00:00 2001 From: Elena Zhelezina Date: Tue, 21 Jan 2020 13:18:55 +0000 Subject: [PATCH 014/867] Added an option TFLITE_BUILTINS_ACTIVATIONS_INT16_WEIGHTS_INT8 to enable sym quantization with activations in 16-bit and weigths in 8-bit. --- tensorflow/lite/python/convert.py | 6 + tensorflow/lite/python/lite.py | 13 +- tensorflow/lite/python/lite_constants.py | 3 + tensorflow/lite/python/lite_test.py | 14 +- .../python/optimize/calibration_wrapper.cc | 8 +- .../python/optimize/calibration_wrapper.h | 3 +- tensorflow/lite/python/optimize/calibrator.py | 6 +- .../lite/python/optimize/calibrator_test.py | 39 ++- .../lite/tools/optimize/operator_property.cc | 17 +- .../lite/tools/optimize/operator_property.h | 10 +- .../lite/tools/optimize/quantization_utils.cc | 102 +++++-- .../lite/tools/optimize/quantization_utils.h | 10 +- .../tools/optimize/quantization_utils_test.cc | 4 +- .../tools/optimize/quantization_wrapper.cc | 4 +- .../lite/tools/optimize/quantize_model.cc | 175 +++++++----- .../lite/tools/optimize/quantize_model.h | 7 +- .../tools/optimize/quantize_model_test.cc | 258 ++++++++++++------ 17 files changed, 477 insertions(+), 202 deletions(-) diff --git a/tensorflow/lite/python/convert.py b/tensorflow/lite/python/convert.py index 2fe4d172487..494f32a515c 100644 --- a/tensorflow/lite/python/convert.py +++ b/tensorflow/lite/python/convert.py @@ -93,6 +93,12 @@ class OpsSet(enum.Enum): # quantized implementations. TFLITE_BUILTINS_INT8 = "TFLITE_BUILTINS_INT8" + # Convert model using only TensorFlow Lite operations with quantized int8 weights + # and int16 activations. + # Specifying this will throw an error for operations that do not yet have + # quantized implementations. + TFLITE_BUILTINS_ACTIVATIONS_INT16_WEIGHTS_INT8 = "TFLITE_BUILTINS_ACTIVATIONS_INT16_WEIGHTS_INT8" + def __str__(self): return self.value diff --git a/tensorflow/lite/python/lite.py b/tensorflow/lite/python/lite.py index 657cfea1bb8..fc9c064faf0 100644 --- a/tensorflow/lite/python/lite.py +++ b/tensorflow/lite/python/lite.py @@ -224,6 +224,10 @@ class TFLiteConverterBase(object): self.target_spec.supported_ops) or self._smallest_supported_type() == constants.INT8) + def _is_int16x8_target_required(self): + return (set([OpsSet.TFLITE_BUILTINS_ACTIVATIONS_INT16_WEIGHTS_INT8]) == + set(self.target_spec.supported_ops)) + def _smallest_supported_type(self): if self.target_spec.supported_types: return min(self.target_spec.supported_types, key=lambda x: x.size) @@ -238,7 +242,9 @@ class TFLiteConverterBase(object): ])) def _is_post_training_optimize(self): - return self._is_int8_target_required() or self._any_optimization_enabled() + return self._is_int8_target_required() or \ + self._is_int16x8_target_required() or \ + self._any_optimization_enabled() def _is_int8_weight_only_quantize(self): return (self._is_post_training_optimize() and @@ -255,11 +261,12 @@ class TFLiteConverterBase(object): def _calibrate_quantize_model(self, result, inference_input_type, inference_output_type, enable_mlir_quantizer): - allow_float = not self._is_int8_target_required() + allow_float = not self._is_int8_target_required() and not self._is_int16x8_target_required() calibrate_quantize = _calibrator.Calibrator(result) + activations_type = constants.INT16 if self._is_int16x8_target_required() else constants.INT8 return calibrate_quantize.calibrate_and_quantize( self.representative_dataset.input_gen, inference_input_type, - inference_output_type, allow_float, enable_mlir_quantizer) + inference_output_type, allow_float, activations_type, enable_mlir_quantizer) def _get_base_converter_args(self): """Returns the base converter args. diff --git a/tensorflow/lite/python/lite_constants.py b/tensorflow/lite/python/lite_constants.py index d43452c775b..4902f23795e 100644 --- a/tensorflow/lite/python/lite_constants.py +++ b/tensorflow/lite/python/lite_constants.py @@ -30,6 +30,7 @@ INT64 = dtypes.int64 STRING = dtypes.string QUANTIZED_UINT8 = dtypes.uint8 INT8 = dtypes.int8 +INT16 = dtypes.int16 COMPLEX64 = dtypes.complex64 TENSORFLOW_GRAPHDEF = _toco_flags_pb2.TENSORFLOW_GRAPHDEF TFLITE = _toco_flags_pb2.TFLITE @@ -43,6 +44,7 @@ _tf_export(v1=["lite.constants.STRING"]).export_constant(__name__, "STRING") _tf_export(v1=["lite.constants.QUANTIZED_UINT8"]).export_constant( __name__, "QUANTIZED_UINT8") _tf_export(v1=["lite.constants.INT8"]).export_constant(__name__, "INT8") +_tf_export(v1=["lite.constants.INT16"]).export_constant(__name__, "INT16") _tf_export(v1=["lite.constants.TFLITE"]).export_constant(__name__, "TFLITE") _tf_export(v1=["lite.constants.GRAPHVIZ_DOT"]).export_constant( __name__, "GRAPHVIZ_DOT") @@ -62,6 +64,7 @@ _allowed_symbols = [ "STRING", "QUANTIZED_UINT8", "INT8", + "INT16", "COMPLEX64", "TENSORFLOW_GRAPHDEF", "TFLITE", diff --git a/tensorflow/lite/python/lite_test.py b/tensorflow/lite/python/lite_test.py index 16959c84146..ef5e5d1cdf4 100644 --- a/tensorflow/lite/python/lite_test.py +++ b/tensorflow/lite/python/lite_test.py @@ -769,9 +769,13 @@ class FromSessionTest(TestModels, parameterized.TestCase): self.assertLess(len(quantized_tflite), len(float_tflite)) @parameterized.named_parameters( - ('EnableMlirConverter', True), # enable mlir - ('DisableMlirConverter', False)) # disable mlir - def testCalibrateAndQuantizeBuiltinInt8(self, enable_mlir): + # Quantize model to Int8: with enable mlir + ('UseTfliteBuiltinsIntEnableMLIR', [lite.OpsSet.TFLITE_BUILTINS_INT8], True), + # Quantize model to Int8: with disable mlir + ('UseTfliteBuiltinsIntDisableMLIR', [lite.OpsSet.TFLITE_BUILTINS_INT8], False), + # Quantize model to Int16: with disable mlir + ('UseTfliteBuiltinsInt16DisableMLIR', [lite.OpsSet.TFLITE_BUILTINS_ACTIVATIONS_INT16_WEIGHTS_INT8], False)) + def testCalibrateAndQuantizeBuiltinInt(self, supported_ops, enable_mlir): with ops.Graph().as_default(): inp, output, calibration_gen = self._getCalibrationQuantizeModel() sess = session.Session() @@ -787,9 +791,7 @@ class FromSessionTest(TestModels, parameterized.TestCase): quantized_converter = lite.TFLiteConverter.from_session( sess, [inp], [output]) quantized_converter.experimental_new_converter = enable_mlir - quantized_converter.target_spec.supported_ops = [ - lite.OpsSet.TFLITE_BUILTINS_INT8 - ] + quantized_converter.target_spec.supported_ops = supported_ops quantized_converter.representative_dataset = calibration_gen quantized_tflite = quantized_converter.convert() self.assertTrue(quantized_tflite) diff --git a/tensorflow/lite/python/optimize/calibration_wrapper.cc b/tensorflow/lite/python/optimize/calibration_wrapper.cc index 89ffb3430ea..88995136726 100644 --- a/tensorflow/lite/python/optimize/calibration_wrapper.cc +++ b/tensorflow/lite/python/optimize/calibration_wrapper.cc @@ -204,6 +204,7 @@ PyObject* CalibrationWrapper::SetTensor(int index, PyObject* value) { PyObject* CalibrationWrapper::QuantizeModel(int input_py_type, int output_py_type, bool allow_float, + int activations_py_type, bool enable_mlir_quantizer) { if (NoOpModel(*model_)) { return python_utils::ConvertToPyString(model_str_->data(), @@ -212,6 +213,9 @@ PyObject* CalibrationWrapper::QuantizeModel(int input_py_type, TfLiteType input_type = python_utils::TfLiteTypeFromPyType(input_py_type); TfLiteType output_type = python_utils::TfLiteTypeFromPyType(output_py_type); + TfLiteType activations_type = + python_utils::TfLiteTypeFromPyType(activations_py_type); + if (input_type == kTfLiteNoType || output_type == kTfLiteNoType) { PyErr_SetString(PyExc_ValueError, "Input/output type cannot be kTfLiteNoType"); @@ -230,7 +234,7 @@ PyObject* CalibrationWrapper::QuantizeModel(int input_py_type, status = tflite::optimize::QuantizeModel( &builder, tflite_model.get(), TfLiteTypeToSchemaType(input_type), TfLiteTypeToSchemaType(output_type), allow_float, - error_reporter_.get()); + TfLiteTypeToSchemaType(activations_type), error_reporter_.get()); } if (status != kTfLiteOk) { @@ -262,7 +266,7 @@ PyObject* CalibrationWrapper::QuantizeModel(int input_py_type, auto status = tflite::optimize::QuantizeModel( &builder, tflite_model.get(), TfLiteTypeToSchemaType(input_type), TfLiteTypeToSchemaType(output_type), allow_float, {op_name}, - error_reporter_.get()); + TensorType_INT8, error_reporter_.get()); if (status != kTfLiteOk) { error_reporter_->exception(); return nullptr; diff --git a/tensorflow/lite/python/optimize/calibration_wrapper.h b/tensorflow/lite/python/optimize/calibration_wrapper.h index 0fefc29dd81..e72fe15e958 100644 --- a/tensorflow/lite/python/optimize/calibration_wrapper.h +++ b/tensorflow/lite/python/optimize/calibration_wrapper.h @@ -60,7 +60,8 @@ class CalibrationWrapper { PyObject* FeedTensor(PyObject* input_value); PyObject* QuantizeModel(int input_py_type, int output_py_type, - bool allow_float, bool enable_mlir_quantizer = false); + bool allow_float, int activations_py_type, + bool enable_mlir_quantizer = false); // Allows quantizing only the operator that produces the tensor with name // operator_output_name. (This can be used to help debug.). diff --git a/tensorflow/lite/python/optimize/calibrator.py b/tensorflow/lite/python/optimize/calibrator.py index 6d9a29236f0..1f962917551 100644 --- a/tensorflow/lite/python/optimize/calibrator.py +++ b/tensorflow/lite/python/optimize/calibrator.py @@ -19,6 +19,7 @@ from __future__ import print_function import numpy as np from tensorflow.python.util.lazy_loader import LazyLoader +from tensorflow.lite.python import lite_constants # Lazy load since some of the performance benchmark skylark rules # break dependencies. Must use double quotes to match code internal rewrite @@ -55,7 +56,8 @@ class Calibrator(object): raise ValueError("Failed to parse the model.") def calibrate_and_quantize(self, dataset_gen, input_type, output_type, - allow_float, enable_mlir_quantizer=False): + allow_float, activations_type = lite_constants.INT8, + enable_mlir_quantizer=False): """Calibrates the model with specified generator and then quantizes it. Returns: @@ -69,6 +71,7 @@ class Calibrator(object): computation, useful when targeting an integer-only backend. If False, an error will be thrown if an operation cannot be quantized, otherwise the model will fallback to float ops. + activations_type: A tf.dtype representing the desired type for activations enable_mlir_quantizer: A boolean. True if wants to use mlir quantizer to quantize the calibrated model. """ @@ -78,6 +81,7 @@ class Calibrator(object): return self._calibrator.QuantizeModel( np.dtype(input_type.as_numpy_dtype()).num, np.dtype(output_type.as_numpy_dtype()).num, allow_float, + np.dtype(activations_type.as_numpy_dtype()).num, enable_mlir_quantizer) def calibrate_and_quantize_single(self, dataset_gen, input_type, output_type, diff --git a/tensorflow/lite/python/optimize/calibrator_test.py b/tensorflow/lite/python/optimize/calibrator_test.py index 28e8723f23d..7ec5f8f526c 100644 --- a/tensorflow/lite/python/optimize/calibrator_test.py +++ b/tensorflow/lite/python/optimize/calibrator_test.py @@ -33,9 +33,13 @@ from tensorflow.python.platform import test class CalibratorTest(test_util.TensorFlowTestCase, parameterized.TestCase): @parameterized.named_parameters( - ('EnableMlirQuantizer', True), # enable mlir quantizer - ('DisableMlirQuantizer', False)) # disable mlir quantizer - def test_calibration_with_quantization(self, enable_mlir): + # Activation type Int8 - enable mlir quantizer + ('UseActivationTypeInt8EnabledMlir', constants.INT8, True), + # Activation type Int8 - disable mlir quantizer + ('UseActivationTypeInt8DisabledMlir', constants.INT8, False), + # Activation type Int16 + ('UseActivationTypeInt16', constants.INT16, False)) + def test_calibration_with_quantization(self, activations_type, enable_mlir): model_path = resource_loader.get_path_to_datafile( 'test_data/mobilenet_like_model.bin') float_model = open(model_path, 'rb').read() @@ -49,13 +53,18 @@ class CalibratorTest(test_util.TensorFlowTestCase, parameterized.TestCase): quantized_model = quantizer.calibrate_and_quantize(input_gen, constants.FLOAT, constants.FLOAT, False, + activations_type, enable_mlir) self.assertIsNotNone(quantized_model) @parameterized.named_parameters( - ('EnableMlirQuantizer', True), # enable mlir quantizer - ('DisableMlirQuantizer', False)) # disable mlir quantizer - def test_calibration_with_quantization_allow_float(self, enable_mlir): + # Activation type Int8 - enable mlir quantizer + ('UseActivationTypeInt8EnabledMlir', constants.INT8, True), + # Activation type Int8 - disable mlir quantizer + ('UseActivationTypeInt8DisableMlir', constants.INT8, False), + # Activation type Int16 - disable mlir quantizer + ('UseActivationTypeInt16', constants.INT16, False)) + def test_calibration_with_quantization_allow_float(self, activations_type, enable_mlir): model_path = resource_loader.get_path_to_datafile( 'test_data/mobilenet_like_model.bin') float_model = open(model_path, 'rb').read() @@ -69,6 +78,7 @@ class CalibratorTest(test_util.TensorFlowTestCase, parameterized.TestCase): quantized_model = quantizer.calibrate_and_quantize(input_gen, constants.FLOAT, constants.FLOAT, True, + activations_type, enable_mlir) self.assertIsNotNone(quantized_model) @@ -88,9 +98,13 @@ class CalibratorTest(test_util.TensorFlowTestCase, parameterized.TestCase): self.assertIsNotNone(quantized_model) @parameterized.named_parameters( - ('EnableMlirQuantizer', True), # enable mlir quantizer - ('DisableMlirQuantizer', False)) # disable mlir quantizer - def test_calibration_with_quantization_multiple_inputs(self, enable_mlir): + # Activation type Int8 - enable mlir quantizer + ('UseActivationTypeInt8 - EnableMlirQuantizer', constants.INT8, True), + # Activation type Int8 - disable mlir quantizer + ('UseActivationTypeInt8 - DisableMlirQuantizer', constants.INT8, False), + # Activation type Int16 - disable mlir quantizer + ('UseActivationTypeInt16 - DisableEnableMlirQuantizer', constants.INT16, False)) + def test_calibration_with_quantization_multiple_inputs(self, activations_type, enable_mlir): # Load multi add model from test data. # This model has 4 inputs of size (1, 8, 8, 3). model_path = resource_loader.get_path_to_datafile( @@ -106,6 +120,7 @@ class CalibratorTest(test_util.TensorFlowTestCase, parameterized.TestCase): quantized_model = quantizer.calibrate_and_quantize(input_gen, constants.FLOAT, constants.FLOAT, False, + activations_type, enable_mlir) self.assertIsNotNone(quantized_model) @@ -148,7 +163,8 @@ class CalibratorTest(test_util.TensorFlowTestCase, parameterized.TestCase): with self.assertRaisesRegex(ValueError, 'Size mismatch'): quantizer.calibrate_and_quantize(input_gen, constants.FLOAT, - constants.FLOAT, False, enable_mlir) + constants.FLOAT, False, + enable_mlir) @parameterized.named_parameters( ('EnableMlirQuantizer', True), # enable mlir quantizer @@ -166,7 +182,8 @@ class CalibratorTest(test_util.TensorFlowTestCase, parameterized.TestCase): with self.assertRaises(ValueError): quantizer.calibrate_and_quantize(input_gen, constants.FLOAT, - constants.FLOAT, False, enable_mlir) + constants.FLOAT, False, + constants.INT8, enable_mlir) if __name__ == '__main__': diff --git a/tensorflow/lite/tools/optimize/operator_property.cc b/tensorflow/lite/tools/optimize/operator_property.cc index 13f63092761..1f2d8bb4a4d 100644 --- a/tensorflow/lite/tools/optimize/operator_property.cc +++ b/tensorflow/lite/tools/optimize/operator_property.cc @@ -64,6 +64,7 @@ OperatorProperty GetOperatorProperty(const ModelT* model, int subgraph_index, property.inputs = {{0, {}}, {1, {}}}; property.outputs = {{0, {}}}; property.version = 2; + property.quantize_input_as_activations = true; break; case BuiltinOperator_ARG_MAX: property.inputs = {{0, {}}}; @@ -176,7 +177,7 @@ OperatorProperty GetOperatorProperty(const ModelT* model, int subgraph_index, // LogSoftmax requires output with 16/256 as scale and 127 as zero point. TensorProperty tensor_property; tensor_property.restriction = true; - tensor_property.restricted_value = {16.0 / 256.0, 127}; + tensor_property.restricted_value_int8 = {16.0 / 256.0, 127}; property.outputs = {{0, tensor_property}}; property.version = 2; break; @@ -186,7 +187,8 @@ OperatorProperty GetOperatorProperty(const ModelT* model, int subgraph_index, // Logistic requires output with 1/256 as scale and -128 as zero point. TensorProperty tensor_property; tensor_property.restriction = true; - tensor_property.restricted_value = {1 / 256.0, -128}; + tensor_property.restricted_value_int8 = {1 / 256.0, -128}; + tensor_property.restricted_value_int16 = {1 / 32768.0, 0}; property.outputs = {{0, tensor_property}}; property.version = 2; break; @@ -741,7 +743,7 @@ OperatorProperty GetOperatorProperty(const ModelT* model, int subgraph_index, // L2 Norm requires output with 1/128 as scale and 0 as zero point. TensorProperty tensor_property; tensor_property.restriction = true; - tensor_property.restricted_value = {1 / 128.0, 0}; + tensor_property.restricted_value_int8 = {1 / 128.0, 0}; property.outputs = {{0, tensor_property}}; property.version = 2; break; @@ -756,6 +758,7 @@ OperatorProperty GetOperatorProperty(const ModelT* model, int subgraph_index, property.arbitrary_inputs = true; property.outputs = {{0, {}}}; property.restrict_same_input_output_scale = true; + property.quantize_input_as_activations = true; property.version = 2; break; case BuiltinOperator_MEAN: @@ -767,6 +770,7 @@ OperatorProperty GetOperatorProperty(const ModelT* model, int subgraph_index, property.arbitrary_inputs = true; property.outputs = {{0, {}}}; property.restrict_same_input_output_scale = true; + property.quantize_input_as_activations = true; property.version = 2; break; case BuiltinOperator_MUL: @@ -778,6 +782,7 @@ OperatorProperty GetOperatorProperty(const ModelT* model, int subgraph_index, property.arbitrary_inputs = true; property.outputs = {{0, {}}}; property.restrict_same_input_output_scale = true; + property.restrict_same_input_output_scale = true; property.version = 2; break; case BuiltinOperator_PAD: @@ -840,7 +845,8 @@ OperatorProperty GetOperatorProperty(const ModelT* model, int subgraph_index, // Softmax requires output with 1/256 as scale and -128 as zero point. TensorProperty tensor_property; tensor_property.restriction = true; - tensor_property.restricted_value = {1 / 256.0, -128}; + tensor_property.restricted_value_int8 = {1 / 256.0, -128}; + tensor_property.restricted_value_int16 = {1 / 32768.0, 0}; property.outputs = {{0, tensor_property}}; property.version = 2; break; @@ -866,7 +872,8 @@ OperatorProperty GetOperatorProperty(const ModelT* model, int subgraph_index, // Tanh requires output with 1/128 as scale and 0 as zero point. TensorProperty tensor_property; tensor_property.restriction = true; - tensor_property.restricted_value = {1 / 128.0, 0}; + tensor_property.restricted_value_int8 = {1 / 128.0, 0}; + tensor_property.restricted_value_int16 = {1 / 32768.0, 0}; property.outputs = {{0, tensor_property}}; property.version = 2; break; diff --git a/tensorflow/lite/tools/optimize/operator_property.h b/tensorflow/lite/tools/optimize/operator_property.h index 5d37aa304e5..23052308568 100644 --- a/tensorflow/lite/tools/optimize/operator_property.h +++ b/tensorflow/lite/tools/optimize/operator_property.h @@ -43,7 +43,8 @@ struct TensorProperty { // Constraints. bool restriction = false; // scale/zero_point hardcoded. - std::pair restricted_value = {0.0, 0}; + std::pair restricted_value_int8 = {0.0, 0}; + std::pair restricted_value_int16 = {0.0, 0}; // Use derived scale. bool use_derived_scale = false; @@ -93,6 +94,13 @@ struct OperatorProperty { // Op version. int version = 1; + + // When we quantize activations into 16 bit and weights into 8 bit, + // we want to quantize all inputs, including constant tensors, + // for the operators like Add, Mul into 16-bit as well. The constant + // inputs are quantized as weights and this variable indicates + // that we want to do quantizations of these tensors as activations. + bool quantize_input_as_activations = false; }; OperatorProperty GetOperatorProperty(const ModelT* model, int subgraph_index, diff --git a/tensorflow/lite/tools/optimize/quantization_utils.cc b/tensorflow/lite/tools/optimize/quantization_utils.cc index 10680758d72..4bc9686ec2c 100644 --- a/tensorflow/lite/tools/optimize/quantization_utils.cc +++ b/tensorflow/lite/tools/optimize/quantization_utils.cc @@ -20,7 +20,6 @@ limitations under the License. #include #include "absl/memory/memory.h" -#include "third_party/eigen3/Eigen/Core" #include "tensorflow/lite/c/common.h" #include "tensorflow/lite/core/api/error_reporter.h" #include "tensorflow/lite/kernels/internal/quantization_util.h" @@ -30,6 +29,7 @@ limitations under the License. #include "tensorflow/lite/minimal_logging.h" #include "tensorflow/lite/schema/schema_generated.h" #include "tensorflow/lite/tools/optimize/model_utils.h" +#include "third_party/eigen3/Eigen/Core" namespace tflite { namespace optimize { @@ -85,6 +85,46 @@ void GetAsymmetricQuantizationParams( quantization_params->zero_point = std::vector(1, zero_point); } +void GetSymmetricQuantizationParams( + float min, float max, const int half_quant_range, + QuantizationParametersT* quantization_params) { + // Adjust the boundaries to guarantee 0 is included. + min = std::min(min, 0.0f); + max = std::max(max, 0.0f); + const float scale = std::max(std::abs(max), std::abs(min)) / half_quant_range; + int64_t zero_point = 0; + quantization_params->min = std::vector(1, min); + quantization_params->max = std::vector(1, max); + quantization_params->scale = std::vector(1, scale); + quantization_params->zero_point = std::vector(1, 0); +} + +TfLiteStatus GetQuantizationParams(TensorT* tensor, TensorType activations_type, + QuantizationParametersT* quantization_params, + ErrorReporter* error_reporter) { + if (activations_type == TensorType_INT8) { + GetAsymmetricQuantizationParams( + tensor->quantization->min[0], tensor->quantization->max[0], + std::numeric_limits::min(), std::numeric_limits::max(), + quantization_params); + } else if (activations_type == TensorType_INT16) { + float range = std::max(std::abs(tensor->quantization->min[0]), + std::abs(tensor->quantization->max[0])); + const float quantized_range = 32767.0; + const float scale = range / quantized_range; + quantization_params->min = std::vector(1, -range); + quantization_params->max = std::vector(1, range); + quantization_params->scale = std::vector(1, scale); + quantization_params->zero_point = std::vector(1, 0); + } else { + error_reporter->Report( + "Unsupported activation type for quantize-activation: %s", + activations_type); + return kTfLiteError; + } + return kTfLiteOk; +} + // Set the max and min quantization parameter for a single tensor given its // values. void FillSingleMinMax(const float* const input, const uint64_t input_size, @@ -536,6 +576,7 @@ TfLiteStatus SymmetricQuantizeTensorPerChannel(ModelT* model, TensorT* tensor, model, tensor, error_reporter); } +template TfLiteStatus SymmetricPerLayerBiasQuantize(ModelT* model, TensorT* tensor, float scaling_factor, ErrorReporter* error_reporter) { @@ -548,25 +589,38 @@ TfLiteStatus SymmetricPerLayerBiasQuantize(ModelT* model, TensorT* tensor, uint64_t num_elements; TF_LITE_ENSURE_STATUS(NumElements(*tensor, &num_elements)); - std::vector final_buffer(num_elements); - const int32_t kScale = std::numeric_limits::max(); + std::vector final_buffer(num_elements); + const BiasType kScale = std::numeric_limits::max(); for (size_t i = 0; i < num_elements; i++) { - const int32_t quantized_value = tflite::SafeCast( + const BiasType quantized_value = tflite::SafeCast( TfLiteRound(float_data[i] * scaling_factor_inv)); final_buffer[i] = std::min(kScale, std::max(-kScale, quantized_value)); } // Set the buffers and output type. uint8_t* uint8_buffer = reinterpret_cast(final_buffer.data()); - size_t buffer_size = num_elements * sizeof(int32_t); + size_t buffer_size = num_elements * sizeof(BiasType); std::vector scales(1, scaling_factor); std::vector zero_points(1, 0); + + auto output_type = std::is_same::value + ? TensorType_INT32 + : TensorType_INT64; return AddQuantizationParams(scales, zero_points, 0, uint8_buffer, - buffer_size, TensorType_INT32, model, tensor, + buffer_size, output_type, model, tensor, error_reporter); } +template TfLiteStatus SymmetricPerLayerBiasQuantize( + ModelT* model, TensorT* tensor, float scaling_factor, + ErrorReporter* error_reporter); + +template TfLiteStatus SymmetricPerLayerBiasQuantize( + ModelT* model, TensorT* tensor, float scaling_factor, + ErrorReporter* error_reporter); + +template TfLiteStatus SymmetricPerChannelBiasQuantize(ModelT* model, TensorT* tensor, float input_scale, const float* weight_scales, @@ -583,14 +637,14 @@ TfLiteStatus SymmetricPerChannelBiasQuantize(ModelT* model, TensorT* tensor, uint64_t num_elements; TF_LITE_ENSURE_STATUS(NumElements(*tensor, &num_elements)); - std::vector final_buffer(num_elements); - const int32_t kScale = std::numeric_limits::max(); + std::vector final_buffer(num_elements); + const BiasType kScale = std::numeric_limits::max(); for (int32_t channel_idx = 0; channel_idx < number_of_dimension; channel_idx++) { float scaling_factor = scales[channel_idx]; float scaling_factor_inv = (scaling_factor == 0) ? 0 : 1.0 / scaling_factor; - const int32_t quantized_value = tflite::SafeCast( + const BiasType quantized_value = tflite::SafeCast( TfLiteRound(float_data[channel_idx] * scaling_factor_inv)); final_buffer[channel_idx] = std::min(kScale, std::max(-kScale, quantized_value)); @@ -598,12 +652,26 @@ TfLiteStatus SymmetricPerChannelBiasQuantize(ModelT* model, TensorT* tensor, // Set the buffers and output type. uint8_t* uint8_buffer = reinterpret_cast(final_buffer.data()); - size_t buffer_size = num_elements * sizeof(int32_t); + size_t buffer_size = num_elements * sizeof(BiasType); std::vector zero_point(scales.size(), 0); + + auto output_type = std::is_same::value + ? TensorType_INT32 + : TensorType_INT64; return AddQuantizationParams(scales, zero_point, 0, uint8_buffer, buffer_size, - TensorType_INT32, model, tensor, error_reporter); + output_type, model, tensor, error_reporter); } +template TfLiteStatus SymmetricPerChannelBiasQuantize( + ModelT* model, TensorT* tensor, float input_scale, + const float* weight_scales, int number_of_dimension, + ErrorReporter* error_reporter); + +template TfLiteStatus SymmetricPerChannelBiasQuantize( + ModelT* model, TensorT* tensor, float input_scale, + const float* weight_scales, int number_of_dimension, + ErrorReporter* error_reporter); + TfLiteStatus QuantizeWeight(ModelT* model, TensorT* tensor, bool per_channel, int per_axis_index, ErrorReporter* error_reporter) { // TODO(suharshs): Currently we conflate quantizing weights and constants. Its @@ -645,12 +713,12 @@ float GetEffectiveScale(ModelT* model, SubGraphT* subgraph, int op_idx, return scale; } -void QuantizeActivation(TensorT* tensor) { - GetAsymmetricQuantizationParams( - tensor->quantization->min[0], tensor->quantization->max[0], - std::numeric_limits::min(), std::numeric_limits::max(), - tensor->quantization.get()); - tensor->type = TensorType_INT8; +TfLiteStatus QuantizeActivation(TensorT* tensor, TensorType activations_type, + ErrorReporter* error_reporter) { + TF_LITE_ENSURE_STATUS(GetQuantizationParams( + tensor, activations_type, tensor->quantization.get(), error_reporter)); + tensor->type = activations_type; + return kTfLiteOk; } TfLiteStatus QuantizeActivationToInt16(TensorT* tensor, float scale) { diff --git a/tensorflow/lite/tools/optimize/quantization_utils.h b/tensorflow/lite/tools/optimize/quantization_utils.h index 18ed707e175..752b4253250 100644 --- a/tensorflow/lite/tools/optimize/quantization_utils.h +++ b/tensorflow/lite/tools/optimize/quantization_utils.h @@ -113,12 +113,14 @@ TfLiteStatus SymmetricQuantizeFloatsToInt16(ModelT* model, TensorT* tensor, ErrorReporter* error_reporter); // Symmetrically quantized the bias for per-layer ops (i.e. FullyConnected). +template TfLiteStatus SymmetricPerLayerBiasQuantize(ModelT* model, TensorT* tensor, float scaling_factor, ErrorReporter* error_reporter); // Symmetrically quantizes the bias for ops like Conv and DepthwiseConv. // The scale of bias if weight_per_channel_scale[channel] * input_scale. +template TfLiteStatus SymmetricPerChannelBiasQuantize(ModelT* model, TensorT* tensor, float input_scale, const float* weight_scales, @@ -135,8 +137,14 @@ float GetEffectiveScale(ModelT* model, SubGraphT* subgraph, int op_idx, std::vector intermediate_index, std::vector factors); +// Return quantization parameters depending on activations type. +TfLiteStatus GetQuantizationParams(TensorT* tensor, TensorType activations_type, + QuantizationParametersT* quantization_params, + ErrorReporter* error_reporter); + // Quantize activation. -void QuantizeActivation(TensorT* tensor); +TfLiteStatus QuantizeActivation(TensorT* tensor, TensorType activations_type, + ErrorReporter* error_reporter); // Quantize activation to 16bit. TfLiteStatus QuantizeActivationToInt16(TensorT* tensor, float scale); diff --git a/tensorflow/lite/tools/optimize/quantization_utils_test.cc b/tensorflow/lite/tools/optimize/quantization_utils_test.cc index ece0123d166..49009e49600 100644 --- a/tensorflow/lite/tools/optimize/quantization_utils_test.cc +++ b/tensorflow/lite/tools/optimize/quantization_utils_test.cc @@ -701,7 +701,7 @@ TEST_F(QuantizationUtilsTest, SymmetricPerLayerBiasQuantize) { model->buffers.push_back(std::move(buffer)); // Call and verify. - EXPECT_EQ(SymmetricPerLayerBiasQuantize( + EXPECT_EQ(SymmetricPerLayerBiasQuantize( model.get(), model->subgraphs[0]->tensors[0].get(), input_scale * weight_scale, &error_reporter_), kTfLiteOk); @@ -759,7 +759,7 @@ TEST_F(QuantizationUtilsTest, SymmetricPerChannelBiasQuantize) { model->buffers.push_back(std::move(buffer)); // Call and verify. - EXPECT_EQ(SymmetricPerChannelBiasQuantize( + EXPECT_EQ(SymmetricPerChannelBiasQuantize( model.get(), model->subgraphs[0]->tensors[0].get(), input_scale, weight_scales.data(), 2, &error_reporter_), kTfLiteOk); diff --git a/tensorflow/lite/tools/optimize/quantization_wrapper.cc b/tensorflow/lite/tools/optimize/quantization_wrapper.cc index bd3331da6bf..5002c382bc7 100644 --- a/tensorflow/lite/tools/optimize/quantization_wrapper.cc +++ b/tensorflow/lite/tools/optimize/quantization_wrapper.cc @@ -42,7 +42,9 @@ bool CreateQuantizedModel(const std::string& path) { tflite::StderrReporter error_reporter; if (tflite::optimize::QuantizeModel( &builder, &model, tflite::TensorType_FLOAT32, - tflite::TensorType_FLOAT32, &error_reporter) != kTfLiteOk) { + tflite::TensorType_FLOAT32, + // TODO: Pass required activation type if needed + tflite::TensorType_INT8, &error_reporter) != kTfLiteOk) { return false; } return WriteFile(path, builder.GetBufferPointer(), builder.GetSize()); diff --git a/tensorflow/lite/tools/optimize/quantize_model.cc b/tensorflow/lite/tools/optimize/quantize_model.cc index 6fc19ff2a56..ee562fe9c4c 100644 --- a/tensorflow/lite/tools/optimize/quantize_model.cc +++ b/tensorflow/lite/tools/optimize/quantize_model.cc @@ -64,6 +64,7 @@ operator_property::OperatorProperty GetOperatorProperty( TfLiteStatus QuantizeBias(ModelT* model, const TensorT* input_tensor, const TensorT* weight_tensor, TensorT* bias_tensor, bool is_per_channel, int channel_dim_index, + const TensorType& activations_type, ErrorReporter* error_reporter) { if (bias_tensor->shape.size() != 1) { error_reporter->Report("Expected bias tensor shape to be 1."); @@ -92,9 +93,15 @@ TfLiteStatus QuantizeBias(ModelT* model, const TensorT* input_tensor, weight_scales.size()); return kTfLiteError; } - return utils::SymmetricPerChannelBiasQuantize( - model, bias_tensor, input_tensor->quantization->scale[0], - weight_scales.data(), channel_dim_size, error_reporter); + if (activations_type == tflite::TensorType_INT16) { + return utils::SymmetricPerChannelBiasQuantize( + model, bias_tensor, input_tensor->quantization->scale[0], + weight_scales.data(), channel_dim_size, error_reporter); + } else { + return utils::SymmetricPerChannelBiasQuantize( + model, bias_tensor, input_tensor->quantization->scale[0], + weight_scales.data(), channel_dim_size, error_reporter); + } } else { if (weight_scales.size() != 1) { error_reporter->Report( @@ -102,40 +109,54 @@ TfLiteStatus QuantizeBias(ModelT* model, const TensorT* input_tensor, weight_scales.size()); return kTfLiteError; } - return utils::SymmetricPerLayerBiasQuantize( - model, bias_tensor, - input_tensor->quantization->scale[0] * weight_scales[0], - error_reporter); + if (activations_type == tflite::TensorType_INT16) { + return utils::SymmetricPerLayerBiasQuantize( + model, bias_tensor, + input_tensor->quantization->scale[0] * weight_scales[0], + error_reporter); + } else { + return utils::SymmetricPerLayerBiasQuantize( + model, bias_tensor, + input_tensor->quantization->scale[0] * weight_scales[0], + error_reporter); + } } return kTfLiteError; } // True if the tensor type has to be modified. bool TensorTypeChangeRequired(const TensorT* tensor, const TensorType& type) { - // The quantized model is type INT8, so if the user provided type is INT8, we - // do not have to do any custom logic. Additionally, if the current tensor - // isn't INT8 quantized, the custom type doesn't apply. - return (type != TensorType_INT8 && tensor->type == TensorType_INT8 && - !tensor->quantization->scale.empty()); + // The quantized model is type INT8/INT16, so if the user provided type is + // INT8/INT16, we do not have to do any custom logic. Additionally, if the + // current tensor isn't INT8/INT16 quantized, the custom type doesn't apply. + bool int8check = type != TensorType_INT8 && tensor->type == TensorType_INT8 && + !tensor->quantization->scale.empty(); + bool int16check = type != TensorType_INT16 && + tensor->type == TensorType_INT16 && + !tensor->quantization->scale.empty(); + return (int8check || int16check); } // Sets the input type, adding a Leading Op node at the start of the model if // necessary. // Returns the new input tensor index. int32_t SetInputType(ModelT* model, SubGraphT* subgraph, - const int32_t tensor_idx, const TensorType& input_type) { + const int32_t tensor_idx, const TensorType& input_type, + const TensorType& activations_type) { TensorT* tensor = subgraph->tensors[tensor_idx].get(); if (!TensorTypeChangeRequired(tensor, input_type)) { return -1; } if (input_type == TensorType_FLOAT32 || input_type == TensorType_UINT8) { + std::string type_string = + activations_type == TensorType_INT16 ? "int16" : "int8"; // Create a new tensor to be the input of the leading Op. std::unique_ptr leading_op_input; if (input_type == TensorType_FLOAT32) { // Add tensor for quantize operator. Scales and zero points are not // needed. const string leading_op_name = tensor->name; - const string new_name_original_input = tensor->name + "_int8"; + const string new_name_original_input = tensor->name + "_" + type_string; tensor->name = new_name_original_input; utils::MakeTensor(leading_op_name, tensor->shape, input_type, &leading_op_input); @@ -150,7 +171,7 @@ int32_t SetInputType(ModelT* model, SubGraphT* subgraph, TFLITE_DCHECK_GE(zero_point, -128); TFLITE_DCHECK_LE(zero_point, 127); const string leading_op_name = tensor->name; - const string new_name_original_input = tensor->name + "_int8"; + const string new_name_original_input = tensor->name + "_" + type_string; tensor->name = new_name_original_input; utils::MakeTensorWithQuantParam(leading_op_name, tensor->shape, input_type, scale, zero_point + 128, @@ -177,17 +198,20 @@ int32_t SetInputType(ModelT* model, SubGraphT* subgraph, // necessary. // Returns the new output tensor index. int32_t SetOutputType(ModelT* model, SubGraphT* subgraph, - const int32_t tensor_idx, const TensorType& output_type) { + const int32_t tensor_idx, const TensorType& output_type, + const TensorType& activations_type) { TensorT* tensor = subgraph->tensors[tensor_idx].get(); if (!TensorTypeChangeRequired(tensor, output_type)) { return -1; } if (output_type == TensorType_FLOAT32 || output_type == TensorType_UINT8) { + std::string type_string = + activations_type == TensorType_INT16 ? "int16" : "int8"; // Create a new tensor to be the output of the tailing op. std::unique_ptr tailing_op_output; if (output_type == TensorType_FLOAT32) { const string tailing_op_name = tensor->name; - const string new_name_original_output = tensor->name + "_int8"; + const string new_name_original_output = tensor->name + "_" + type_string; tensor->name = new_name_original_output; utils::MakeTensor(tailing_op_name, tensor->shape, output_type, &tailing_op_output); @@ -202,7 +226,7 @@ int32_t SetOutputType(ModelT* model, SubGraphT* subgraph, TFLITE_DCHECK_GE(zero_point, -128); TFLITE_DCHECK_LE(zero_point, 127); const string tailing_op_name = tensor->name; - const string new_name_original_output = tensor->name + "_int8"; + const string new_name_original_output = tensor->name + "_" + type_string; tensor->name = new_name_original_output; utils::MakeTensorWithQuantParam(tailing_op_name, tensor->shape, output_type, scale, zero_point + 128, @@ -238,6 +262,7 @@ int32_t SetOutputType(ModelT* model, SubGraphT* subgraph, // uint8, can be thought as "requant"). TfLiteStatus SetInputAndOutputTypes(ModelT* model, const TensorType& input_type, const TensorType& output_type, + const TensorType& activations_type, ErrorReporter* error_reporter) { for (int subgraph_idx = 0; subgraph_idx < model->subgraphs.size(); subgraph_idx++) { @@ -253,8 +278,8 @@ TfLiteStatus SetInputAndOutputTypes(ModelT* model, const TensorType& input_type, EnumNameTensorType(tensor->type)); return kTfLiteError; } - const int32_t input_idx = - SetInputType(model, subgraph, subgraph->inputs[i], input_type); + const int32_t input_idx = SetInputType( + model, subgraph, subgraph->inputs[i], input_type, activations_type); if (input_idx < 0) { continue; } @@ -270,8 +295,8 @@ TfLiteStatus SetInputAndOutputTypes(ModelT* model, const TensorType& input_type, EnumNameTensorType(tensor->type)); return kTfLiteError; } - const int32_t output_idx = - SetOutputType(model, subgraph, subgraph->outputs[i], output_type); + const int32_t output_idx = SetOutputType( + model, subgraph, subgraph->outputs[i], output_type, activations_type); if (output_idx < 0) { continue; } @@ -287,6 +312,7 @@ TfLiteStatus SetInputAndOutputTypes(ModelT* model, const TensorType& input_type, // The other ones with constraints are handled in QuantizeWeightsAndInput. TfLiteStatus ApplyConstraints(ModelT* model, const std::unordered_set& operator_names, + TensorType activations_type, ErrorReporter* error_reporter) { for (int subgraph_idx = 0; subgraph_idx < model->subgraphs.size(); subgraph_idx++) { @@ -332,7 +358,7 @@ TfLiteStatus ApplyConstraints(ModelT* model, std::unique_ptr additional_tensor; const string requant_tensor_name = input_tensor->name + "_requantized"; utils::MakeTensorWithQuantParam( - requant_tensor_name, input_tensor->shape, TensorType_INT8, + requant_tensor_name, input_tensor->shape, activations_type, output_scale, output_zp, &additional_tensor); const int32_t additional_tensor_idx = subgraph->tensors.size(); subgraph->tensors.push_back(std::move(additional_tensor)); @@ -382,7 +408,8 @@ std::vector> GetOutputs( bool ShouldRestrictSameInputOutputScale( operator_property::OperatorProperty property) { - // Ops with multiple inputs (i.e. concat) gets restricted in ApplyConstraints. + // Ops with multiple inputs (i.e. concat, max and min) gets restricted in + // ApplyConstraints. return (!property.arbitrary_inputs && property.restrict_same_input_output_scale); } @@ -401,7 +428,7 @@ TfLiteStatus QuantizeOpInput( ModelT* model, int32_t subgraph_idx, size_t* op_idx, operator_property::OperatorProperty property, const std::pair& input, - ErrorReporter* error_reporter) { + const TensorType& activations_type, ErrorReporter* error_reporter) { int32_t input_idx = input.first; operator_property::TensorProperty tensor_property = input.second; SubGraphT* subgraph = model->subgraphs.at(subgraph_idx).get(); @@ -429,7 +456,9 @@ TfLiteStatus QuantizeOpInput( if (utils::HasBuffer(model, subgraph, tensor_idx)) { // TODO(suharshs): Look at consumers, throw error if one consumer is // per-channel and one per-layer. - if (tensor_property.number_of_bits == 8) { + bool quantize_const_input = property.quantize_input_as_activations && + activations_type == TensorType_INT16; + if (tensor_property.number_of_bits == 8 && !quantize_const_input) { if (tensor_property.use_derived_scale) { // Currently 8bit tensors in input do not accept derived scale. return kTfLiteError; @@ -444,7 +473,7 @@ TfLiteStatus QuantizeOpInput( *op_idx); return kTfLiteError; } - } else if (tensor_property.number_of_bits == 16) { + } else if (tensor_property.number_of_bits == 16 || quantize_const_input) { if (tensor_property.use_derived_scale) { // Currently 16bit tensors in input do not accept derived scale. return kTfLiteError; @@ -476,8 +505,8 @@ TfLiteStatus QuantizeOpInput( tensor_property.derived_scale.input_tensors, tensor_property.derived_scale.intermediate_tensors, tensor_property.derived_scale.factors); - return utils::SymmetricPerLayerBiasQuantize(model, tensor, scale, - error_reporter); + return utils::SymmetricPerLayerBiasQuantize( + model, tensor, scale, error_reporter); } else if (tensor_property.number_of_bits == 10) { // When the number of bits is 10 (instead of 16), quantize the tensor to @@ -514,7 +543,8 @@ TfLiteStatus QuantizeOpInput( // Currently 8bit tensors in input do not accept derived scale. return kTfLiteError; } - utils::QuantizeActivation(tensor); + TF_LITE_ENSURE_STATUS(utils::QuantizeActivation( + tensor, activations_type, error_reporter)); } else if (tensor_property.number_of_bits == 16) { TensorT* tensor = subgraph->tensors[tensor_idx].get(); float quantized_range = 32767.0; @@ -532,13 +562,16 @@ TfLiteStatus QuantizeOpInput( } else { // If the tensor is not a model input, we need to add a Quantize // operation since the preceding op may require a float output. + std::string type_string = + activations_type == TensorType_INT16 ? "int16" : "int8"; std::unique_ptr op_output; - utils::MakeTensor(tensor->name + "_int8", tensor->shape, - TensorType_INT8, &op_output); + utils::MakeTensor(tensor->name + "_" + type_string, tensor->shape, + activations_type, &op_output); op_output->quantization = absl::make_unique(); op_output->quantization->min.push_back(tensor->quantization->min[0]); op_output->quantization->max.push_back(tensor->quantization->max[0]); - utils::QuantizeActivation(op_output.get()); + TF_LITE_ENSURE_STATUS(utils::QuantizeActivation( + op_output.get(), activations_type, error_reporter)); const int32_t quant_op_output_idx = subgraph->tensors.size(); subgraph->tensors.push_back(std::move(op_output)); std::unique_ptr quant_op; @@ -580,7 +613,7 @@ TfLiteStatus QuantizeOpOutput( ModelT* model, int32_t subgraph_idx, int32_t op_idx, operator_property::OperatorProperty property, const std::pair& output, - ErrorReporter* error_reporter) { + TensorType activations_type, ErrorReporter* error_reporter) { int32_t output_idx = output.first; operator_property::TensorProperty tensor_property = output.second; // If the operator is not quantizable, we don't need to do anything for the @@ -644,18 +677,22 @@ TfLiteStatus QuantizeOpOutput( const float max = input_tensor->quantization->max[0]; output_tensor->quantization->max = {max}; } - output_tensor->type = TensorType_INT8; + output_tensor->type = activations_type; } else if (tensor_property.restriction) { - const auto scale_and_zp = tensor_property.restricted_value; + const auto scale_and_zp = activations_type == TensorType_INT16 + ? tensor_property.restricted_value_int16 + : tensor_property.restricted_value_int8; + // Apply to output. output_tensor->quantization = absl::make_unique(); output_tensor->quantization->scale.push_back(scale_and_zp.first); output_tensor->quantization->zero_point.push_back(scale_and_zp.second); - output_tensor->type = TensorType_INT8; + output_tensor->type = activations_type; } else { // Process regular output that doesn't have any restrictions. if (utils::HasMinMax(output_tensor)) { - utils::QuantizeActivation(output_tensor); + utils::QuantizeActivation(output_tensor, activations_type, + error_reporter); } else { error_reporter->Report( "Unable to find min/max value for output %d in %s in " @@ -668,6 +705,7 @@ TfLiteStatus QuantizeOpOutput( } TfLiteStatus QuantizeIntemediateTensors(ModelT* model, + TensorType activations_type, ErrorReporter* error_reporter) { for (size_t subgraph_idx = 0; subgraph_idx < model->subgraphs.size(); subgraph_idx++) { @@ -691,7 +729,8 @@ TfLiteStatus QuantizeIntemediateTensors(ModelT* model, input.second.symmetric == false) { TensorT* tensor = subgraph->tensors[index_global].get(); if (utils::HasMinMax(tensor)) { - utils::QuantizeActivation(tensor); + utils::QuantizeActivation(tensor, activations_type, + error_reporter); } else { error_reporter->Report( "Unable to find min/max value for output %d in %s in " @@ -793,7 +832,7 @@ TfLiteStatus QuantizeSharedRange(ModelT* model, ErrorReporter* error_reporter) { TfLiteStatus QuantizeWeightsInputOutput( ModelT* model, bool allow_float, const std::unordered_set& operator_names, - ErrorReporter* error_reporter) { + const TensorType& activations_type, ErrorReporter* error_reporter) { for (size_t subgraph_idx = 0; subgraph_idx < model->subgraphs.size(); subgraph_idx++) { SubGraphT* subgraph = model->subgraphs.at(subgraph_idx).get(); @@ -815,14 +854,16 @@ TfLiteStatus QuantizeWeightsInputOutput( for (const std::pair& input : GetInputs(op, property)) { TF_LITE_ENSURE_STATUS(QuantizeOpInput(model, subgraph_idx, &op_idx, - property, input, error_reporter)); + property, input, activations_type, + error_reporter)); } // Quantize operator outputs. for (const std::pair& output : GetOutputs(op, property)) { - TF_LITE_ENSURE_STATUS(QuantizeOpOutput( - model, subgraph_idx, op_idx, property, output, error_reporter)); + TF_LITE_ENSURE_STATUS( + QuantizeOpOutput(model, subgraph_idx, op_idx, property, output, + activations_type, error_reporter)); } } } @@ -832,6 +873,7 @@ TfLiteStatus QuantizeWeightsInputOutput( // Quantize bias. TfLiteStatus QuantizeBiases(ModelT* model, const std::unordered_set& operator_names, + const TensorType& activations_type, ErrorReporter* error_reporter) { for (size_t subgraph_idx = 0; subgraph_idx < model->subgraphs.size(); subgraph_idx++) { @@ -877,10 +919,10 @@ TfLiteStatus QuantizeBiases(ModelT* model, subgraph->tensors[op->inputs[property.inputs[1].first]].get(); operator_property::TensorProperty weight_property = property.inputs[1].second; - TF_LITE_ENSURE_STATUS( - QuantizeBias(model, input_tensor, weight_tensor, bias_tensor, - weight_property.per_axis, - weight_property.per_axis_index, error_reporter)); + TF_LITE_ENSURE_STATUS(QuantizeBias( + model, input_tensor, weight_tensor, bias_tensor, + weight_property.per_axis, weight_property.per_axis_index, + activations_type, error_reporter)); } } } @@ -1000,7 +1042,7 @@ TfLiteStatus FillQuantizationParams( // Check compatibility of activation, weight and bias scales. Adjust if needed. TfLiteStatus EnsureBiasScaleCompatibility( ModelT* model, const std::unordered_set& operator_names, - ErrorReporter* error_reporter) { + TensorType activations_type, ErrorReporter* error_reporter) { for (size_t subgraph_idx = 0; subgraph_idx < model->subgraphs.size(); subgraph_idx++) { SubGraphT* subgraph = model->subgraphs.at(subgraph_idx).get(); @@ -1049,11 +1091,9 @@ TfLiteStatus EnsureBiasScaleCompatibility( // Get input scale for assymmetric quantization. QuantizationParametersT temp_quant_params = QuantizationParametersT(); - utils::GetAsymmetricQuantizationParams( - input_tensor->quantization->min[0], - input_tensor->quantization->max[0], - std::numeric_limits::min(), - std::numeric_limits::max(), &temp_quant_params); + TF_LITE_ENSURE_STATUS( + utils::GetQuantizationParams(input_tensor, activations_type, + &temp_quant_params, error_reporter)); if (temp_quant_params.scale.size() != 1) { error_reporter->Report("Unexpected input quantization scale size."); return kTfLiteError; @@ -1132,21 +1172,24 @@ TfLiteStatus QuantizeModel(flatbuffers::FlatBufferBuilder* builder, ModelT* model, const TensorType& input_type, const TensorType& output_type, bool allow_float, const std::unordered_set& operator_names, + const TensorType& activations_type, ErrorReporter* error_reporter) { TF_LITE_ENSURE_STATUS( FillQuantizationParams(model, operator_names, error_reporter)); + TF_LITE_ENSURE_STATUS(EnsureBiasScaleCompatibility( + model, operator_names, activations_type, error_reporter)); TF_LITE_ENSURE_STATUS( - EnsureBiasScaleCompatibility(model, operator_names, error_reporter)); - TF_LITE_ENSURE_STATUS(QuantizeIntemediateTensors(model, error_reporter)); + QuantizeIntemediateTensors(model, activations_type, error_reporter)); TF_LITE_ENSURE_STATUS(QuantizeSharedRange(model, error_reporter)); TF_LITE_ENSURE_STATUS(QuantizeWeightsInputOutput( - model, allow_float, operator_names, error_reporter)); + model, allow_float, operator_names, activations_type, error_reporter)); + TF_LITE_ENSURE_STATUS(ApplyConstraints(model, operator_names, + activations_type, error_reporter)); TF_LITE_ENSURE_STATUS( - ApplyConstraints(model, operator_names, error_reporter)); - TF_LITE_ENSURE_STATUS(QuantizeBiases(model, operator_names, error_reporter)); + QuantizeBiases(model, operator_names, activations_type, error_reporter)); utils::SetOperatorCodeVersion(model); - TF_LITE_ENSURE_STATUS( - SetInputAndOutputTypes(model, input_type, output_type, error_reporter)); + TF_LITE_ENSURE_STATUS(SetInputAndOutputTypes( + model, input_type, output_type, activations_type, error_reporter)); flatbuffers::Offset output_model_location = Model::Pack(*builder, model); @@ -1158,23 +1201,27 @@ TfLiteStatus QuantizeModel(flatbuffers::FlatBufferBuilder* builder, TfLiteStatus QuantizeModel(flatbuffers::FlatBufferBuilder* builder, ModelT* model, const TensorType& input_type, const TensorType& output_type, bool allow_float, + const TensorType& activations_type, ErrorReporter* error_reporter) { return QuantizeModel(builder, model, input_type, output_type, allow_float, - GetAllOperatorOutputs(model), error_reporter); + GetAllOperatorOutputs(model), activations_type, + error_reporter); } TfLiteStatus QuantizeModel(flatbuffers::FlatBufferBuilder* builder, ModelT* model, const TensorType& input_type, const TensorType& output_type, + const TensorType& activations_type, ErrorReporter* error_reporter) { return QuantizeModel(builder, model, input_type, output_type, - /*allow_float=*/false, error_reporter); + /*allow_float=*/false, activations_type, error_reporter); } TfLiteStatus QuantizeModel(flatbuffers::FlatBufferBuilder* builder, - ModelT* model, ErrorReporter* error_reporter) { + ModelT* model, const TensorType& activations_type, + ErrorReporter* error_reporter) { return QuantizeModel(builder, model, TensorType_FLOAT32, TensorType_FLOAT32, - /*allow_float=*/false, error_reporter); + /*allow_float=*/false, activations_type, error_reporter); } } // namespace optimize diff --git a/tensorflow/lite/tools/optimize/quantize_model.h b/tensorflow/lite/tools/optimize/quantize_model.h index 9b0353f6b6b..cc801ec9870 100644 --- a/tensorflow/lite/tools/optimize/quantize_model.h +++ b/tensorflow/lite/tools/optimize/quantize_model.h @@ -35,7 +35,9 @@ namespace optimize { // // Note: This is a private API, subject to change. TfLiteStatus QuantizeModel(flatbuffers::FlatBufferBuilder* builder, - ModelT* input_model, ErrorReporter* error_reporter); + ModelT* input_model, + const TensorType& activations_type, + ErrorReporter* error_reporter); // Same as above, but the types of quantized inputs and outputs are // configurable. @@ -44,6 +46,7 @@ TfLiteStatus QuantizeModel(flatbuffers::FlatBufferBuilder* builder, TfLiteStatus QuantizeModel(flatbuffers::FlatBufferBuilder* builder, ModelT* input_model, const TensorType& input_type, const TensorType& output_type, + const TensorType& activations_type, ErrorReporter* error_reporter); // Same as above, but can enable allowing float intermediate operations for ops @@ -53,6 +56,7 @@ TfLiteStatus QuantizeModel(flatbuffers::FlatBufferBuilder* builder, TfLiteStatus QuantizeModel(flatbuffers::FlatBufferBuilder* builder, ModelT* input_model, const TensorType& input_type, const TensorType& output_type, bool allow_float, + const TensorType& activations_type, ErrorReporter* error_reporter); // Same as above, but enables only quantizing a whitelist of operations, @@ -63,6 +67,7 @@ TfLiteStatus QuantizeModel(flatbuffers::FlatBufferBuilder* builder, ModelT* input_model, const TensorType& input_type, const TensorType& output_type, bool allow_float, const std::unordered_set& operator_names, + const TensorType& activations_type, ErrorReporter* error_reporter); } // namespace optimize diff --git a/tensorflow/lite/tools/optimize/quantize_model_test.cc b/tensorflow/lite/tools/optimize/quantize_model_test.cc index da1b293c84b..166d60ecc66 100644 --- a/tensorflow/lite/tools/optimize/quantize_model_test.cc +++ b/tensorflow/lite/tools/optimize/quantize_model_test.cc @@ -80,28 +80,35 @@ class QuantizeModelTest : public testing::Test { internal::FailOnErrorReporter error_reporter_; }; -class QuantizeConvModelTest : public QuantizeModelTest { +class QuantizeConvModelTest : public QuantizeModelTest, + public testing::WithParamInterface { protected: QuantizeConvModelTest() { + tensor_type_ = GetParam(); input_model_ = ReadModel(internal::kConvModelWith0Plus10Weights); readonly_model_ = input_model_->GetModel(); readonly_model_->UnPackTo(&model_); } + TensorType tensor_type_; }; -TEST_F(QuantizeConvModelTest, QuantizationSucceeds) { - auto status = QuantizeModel(&builder_, &model_, TensorType_INT8, - TensorType_INT8, &error_reporter_); +INSTANTIATE_TEST_SUITE_P(QuantizeConvModelTestInst, QuantizeConvModelTest, + testing::ValuesIn({TensorType_INT8, + TensorType_INT16})); + +TEST_P(QuantizeConvModelTest, QuantizationSucceeds) { + auto status = QuantizeModel(&builder_, &model_, tensor_type_, tensor_type_, + tensor_type_, &error_reporter_); EXPECT_EQ(status, kTfLiteOk); const uint8_t* buffer = builder_.GetBufferPointer(); const Model* output_model = GetModel(buffer); ASSERT_TRUE(output_model); } -TEST_F(QuantizeConvModelTest, SkipUnspecifiedLayer) { +TEST_P(QuantizeConvModelTest, SkipUnspecifiedLayer) { auto status = QuantizeModel(&builder_, &model_, TensorType_FLOAT32, TensorType_FLOAT32, - /*allow_float=*/true, {}, &error_reporter_); + /*allow_float=*/true, {}, tensor_type_, &error_reporter_); EXPECT_EQ(status, kTfLiteOk); ASSERT_EQ(model_.subgraphs.size(), readonly_model_->subgraphs()->size()); // The resulting model should be the same. @@ -123,9 +130,9 @@ TEST_F(QuantizeConvModelTest, SkipUnspecifiedLayer) { } } -TEST_F(QuantizeConvModelTest, TensorShapesAndStructureIsUnchanged) { - auto status = QuantizeModel(&builder_, &model_, TensorType_INT8, - TensorType_INT8, &error_reporter_); +TEST_P(QuantizeConvModelTest, TensorShapesAndStructureIsUnchanged) { + auto status = QuantizeModel(&builder_, &model_, tensor_type_, tensor_type_, + tensor_type_, &error_reporter_); EXPECT_EQ(status, kTfLiteOk); ASSERT_EQ(model_.subgraphs.size(), readonly_model_->subgraphs()->size()); for (size_t subgraph_idx = 0; subgraph_idx < model_.subgraphs.size(); @@ -148,9 +155,9 @@ TEST_F(QuantizeConvModelTest, TensorShapesAndStructureIsUnchanged) { EXPECT_EQ(model_.operator_codes[0]->version, 3); } -TEST_F(QuantizeConvModelTest, OperatorsAreUnchanged) { - auto status = QuantizeModel(&builder_, &model_, TensorType_INT8, - TensorType_INT8, &error_reporter_); +TEST_P(QuantizeConvModelTest, OperatorsAreUnchanged) { + auto status = QuantizeModel(&builder_, &model_, tensor_type_, tensor_type_, + tensor_type_, &error_reporter_); EXPECT_EQ(status, kTfLiteOk); ASSERT_EQ(model_.operator_codes.size(), readonly_model_->operator_codes()->size()); @@ -182,20 +189,28 @@ TEST_F(QuantizeConvModelTest, OperatorsAreUnchanged) { } } -TEST_F(QuantizeConvModelTest, GraphIsFullyQuantized) { - auto status = QuantizeModel(&builder_, &model_, TensorType_INT8, - TensorType_INT8, &error_reporter_); +TEST_P(QuantizeConvModelTest, GraphIsFullyQuantized) { + auto status = QuantizeModel(&builder_, &model_, tensor_type_, tensor_type_, + tensor_type_, &error_reporter_); EXPECT_EQ(status, kTfLiteOk); for (const auto& subgraph : model_.subgraphs) { for (const auto& tensor : subgraph->tensors) { - EXPECT_TRUE(tensor->type == TensorType_INT32 || - tensor->type == TensorType_INT8); + if (tensor_type_ == TensorType_INT8) { + EXPECT_TRUE(tensor->type == TensorType_INT32 || + tensor->type == TensorType_INT8); + } else if (tensor_type_ == TensorType_INT16) { + EXPECT_TRUE(tensor->type == TensorType_INT64 || // bias + tensor->type == TensorType_INT8 || // weights + tensor->type == TensorType_INT16); // activations + } } } } -TEST_F(QuantizeConvModelTest, FloatInputAndOutput) { - auto status = QuantizeModel(&builder_, &model_, &error_reporter_); +TEST_P(QuantizeConvModelTest, FloatInputAndOutput) { + auto status = + QuantizeModel(&builder_, &model_, TensorType_FLOAT32, TensorType_FLOAT32, + tensor_type_, &error_reporter_); EXPECT_EQ(status, kTfLiteOk); for (int32_t subgraph_idx = 0; subgraph_idx < model_.subgraphs.size(); @@ -234,22 +249,33 @@ TEST_F(QuantizeConvModelTest, FloatInputAndOutput) { EXPECT_EQ(subgraph->tensors[output_idx]->type, TensorType_FLOAT32); EXPECT_EQ(subgraph->tensors[output_idx]->name, "output"); // The original input and output has been renamed. - EXPECT_EQ(subgraph->tensors[quant_op->outputs[0]]->name, "input_int8"); - EXPECT_EQ(subgraph->tensors[dequant_op->inputs[0]]->name, "output_int8"); + std::string control_suffix = + (tensor_type_ == TensorType_INT16) ? "int16" : "int8"; + EXPECT_EQ(subgraph->tensors[quant_op->outputs[0]]->name, + "input_" + control_suffix); + EXPECT_EQ(subgraph->tensors[dequant_op->inputs[0]]->name, + "output_" + control_suffix); for (int tensor_idx = 0; tensor_idx < subgraph->tensors.size(); ++tensor_idx) { const auto& tensor = subgraph->tensors[tensor_idx]; if (input_idx != tensor_idx && output_idx != tensor_idx) { - EXPECT_TRUE(tensor->type == TensorType_INT32 || - tensor->type == TensorType_INT8); + if (tensor_type_ == TensorType_INT8) { + EXPECT_TRUE(tensor->type == TensorType_INT32 || + tensor->type == TensorType_INT8); + } else if (tensor_type_ == TensorType_INT16) { + EXPECT_TRUE(tensor->type == TensorType_INT64 || // bias + tensor->type == TensorType_INT8 || // weights + tensor->type == TensorType_INT16); // activations + } } } } } -TEST_F(QuantizeConvModelTest, Uint8InputAndOutput) { - auto status = QuantizeModel(&builder_, &model_, TensorType_UINT8, - TensorType_UINT8, &error_reporter_); +TEST_P(QuantizeConvModelTest, Uint8InputAndOutput) { + auto status = + QuantizeModel(&builder_, &model_, TensorType_UINT8, TensorType_UINT8, + TensorType_INT8, &error_reporter_); EXPECT_EQ(status, kTfLiteOk); for (int32_t subgraph_idx = 0; subgraph_idx < model_.subgraphs.size(); @@ -326,21 +352,25 @@ class QuantizeConvNoBiasModelTest : public QuantizeModelTest { }; TEST_F(QuantizeConvNoBiasModelTest, QuantizationSucceeds) { - auto status = QuantizeModel(&builder_, &model_, TensorType_INT8, - TensorType_INT8, &error_reporter_); + auto status = + QuantizeModel(&builder_, &model_, TensorType_INT8, TensorType_INT8, + TensorType_INT8, &error_reporter_); EXPECT_EQ(status, kTfLiteOk); const uint8_t* buffer = builder_.GetBufferPointer(); const Model* output_model = GetModel(buffer); ASSERT_TRUE(output_model); } -class QuantizeConcatModelTest : public QuantizeModelTest { +class QuantizeConcatModelTest : public QuantizeModelTest, + public testing::WithParamInterface { protected: QuantizeConcatModelTest() { input_model_ = ReadModel(internal::kFloatConcatMax5Max10Max10); readonly_model_ = input_model_->GetModel(); readonly_model_->UnPackTo(&model_); } + + TensorType tensor_type_; }; // There are two inputs for concat, "input0" and "input1". "input0" has [0, 5] @@ -352,9 +382,9 @@ class QuantizeConcatModelTest : public QuantizeModelTest { // input0 -> requant -> input0_requant \ // concat - output // input1 / -TEST_F(QuantizeConcatModelTest, AddRequantBeforeConcat) { - auto status = QuantizeModel(&builder_, &model_, TensorType_INT8, - TensorType_INT8, &error_reporter_); +TEST_P(QuantizeConcatModelTest, AddRequantBeforeConcat) { + auto status = QuantizeModel(&builder_, &model_, tensor_type_, tensor_type_, + tensor_type_, &error_reporter_); EXPECT_EQ(status, kTfLiteOk); // There is only one subgraph. @@ -373,32 +403,51 @@ TEST_F(QuantizeConcatModelTest, AddRequantBeforeConcat) { EXPECT_EQ(model_.operator_codes[concat->opcode_index]->builtin_code, BuiltinOperator_CONCATENATION); + auto zero_point_control = tensor_type_ == TensorType_INT8 ? -128 : 0; + /* + input0_scale_control + INT8: (5-0) / (2^8 - 1) + INT16: (5-0) / (2^16 / 2 - 1) + input1_scale + INT8: (10-0) / (2^8 - 1) + INT16: (10-0) / (2^16 / 2 - 1) + */ + auto input0_scale_control = + tensor_type_ == TensorType_INT8 ? 0.019607844 : 0.00015259254; + auto input1_scale = + tensor_type_ == TensorType_INT8 ? 0.039215688 : 0.00030518509; + // There should be 4 tensors: input0, input1, input0_requantized, output. EXPECT_EQ(subgraph->tensors.size(), 4); - EXPECT_EQ(subgraph->tensors[0]->type, TensorType_INT8); + EXPECT_EQ(subgraph->tensors[0]->type, tensor_type_); EXPECT_EQ(subgraph->tensors[0]->name, "input0"); EXPECT_EQ(subgraph->tensors[0]->quantization->scale.size(), 1); EXPECT_EQ(subgraph->tensors[0]->quantization->zero_point.size(), 1); - EXPECT_FLOAT_EQ(subgraph->tensors[0]->quantization->scale[0], 0.019607844); - EXPECT_FLOAT_EQ(subgraph->tensors[0]->quantization->zero_point[0], -128); - EXPECT_EQ(subgraph->tensors[1]->type, TensorType_INT8); + EXPECT_FLOAT_EQ(subgraph->tensors[0]->quantization->scale[0], + input0_scale_control); + EXPECT_FLOAT_EQ(subgraph->tensors[0]->quantization->zero_point[0], + zero_point_control); + EXPECT_EQ(subgraph->tensors[1]->type, tensor_type_); EXPECT_EQ(subgraph->tensors[1]->name, "input1"); EXPECT_EQ(subgraph->tensors[1]->quantization->scale.size(), 1); EXPECT_EQ(subgraph->tensors[1]->quantization->zero_point.size(), 1); - EXPECT_FLOAT_EQ(subgraph->tensors[1]->quantization->scale[0], 0.039215688); - EXPECT_FLOAT_EQ(subgraph->tensors[1]->quantization->zero_point[0], -128); - EXPECT_EQ(subgraph->tensors[2]->type, TensorType_INT8); + EXPECT_FLOAT_EQ(subgraph->tensors[1]->quantization->scale[0], input1_scale); + EXPECT_FLOAT_EQ(subgraph->tensors[1]->quantization->zero_point[0], + zero_point_control); + EXPECT_EQ(subgraph->tensors[2]->type, tensor_type_); EXPECT_EQ(subgraph->tensors[2]->name, "output"); EXPECT_EQ(subgraph->tensors[2]->quantization->scale.size(), 1); EXPECT_EQ(subgraph->tensors[2]->quantization->zero_point.size(), 1); - EXPECT_FLOAT_EQ(subgraph->tensors[2]->quantization->scale[0], 0.039215688); - EXPECT_FLOAT_EQ(subgraph->tensors[2]->quantization->zero_point[0], -128); - EXPECT_EQ(subgraph->tensors[3]->type, TensorType_INT8); + EXPECT_FLOAT_EQ(subgraph->tensors[2]->quantization->scale[0], input1_scale); + EXPECT_FLOAT_EQ(subgraph->tensors[2]->quantization->zero_point[0], + zero_point_control); + EXPECT_EQ(subgraph->tensors[3]->type, tensor_type_); EXPECT_EQ(subgraph->tensors[3]->name, "input0_requantized"); EXPECT_EQ(subgraph->tensors[3]->quantization->scale.size(), 1); EXPECT_EQ(subgraph->tensors[3]->quantization->zero_point.size(), 1); - EXPECT_FLOAT_EQ(subgraph->tensors[3]->quantization->scale[0], 0.039215688); - EXPECT_FLOAT_EQ(subgraph->tensors[3]->quantization->zero_point[0], -128); + EXPECT_FLOAT_EQ(subgraph->tensors[3]->quantization->scale[0], input1_scale); + EXPECT_FLOAT_EQ(subgraph->tensors[3]->quantization->zero_point[0], + zero_point_control); // The connection should be what is described in the comment. EXPECT_EQ(requant->inputs.size(), 1); @@ -419,7 +468,9 @@ TEST_F(QuantizeConcatModelTest, AddRequantBeforeConcat) { EXPECT_EQ(model_.operator_codes[1]->builtin_code, BuiltinOperator_QUANTIZE); EXPECT_EQ(model_.operator_codes[1]->version, 2); } - +INSTANTIATE_TEST_SUITE_P(QuantizeConcatModelInst, QuantizeConcatModelTest, + testing::ValuesIn({TensorType_INT8, + TensorType_INT16})); class QuantizeSplitModelTest : public QuantizeModelTest { protected: QuantizeSplitModelTest() { @@ -432,8 +483,9 @@ class QuantizeSplitModelTest : public QuantizeModelTest { // There are two outputs for split with different scales, the resulting model // should have the scales be hardcodes to the input scale value. TEST_F(QuantizeSplitModelTest, QuantizeSplit) { - auto status = QuantizeModel(&builder_, &model_, TensorType_INT8, - TensorType_INT8, &error_reporter_); + auto status = + QuantizeModel(&builder_, &model_, TensorType_INT8, TensorType_INT8, + TensorType_INT8, &error_reporter_); EXPECT_EQ(status, kTfLiteOk); // There is only one subgraph. @@ -496,8 +548,9 @@ class QuantizeConvModel1Test : public QuantizeModelTest { }; TEST_F(QuantizeConvModel1Test, VerifyConvQuantizationWithUnitScale) { - auto status = QuantizeModel(&builder_, &model_, TensorType_INT8, - TensorType_INT8, &error_reporter_); + auto status = + QuantizeModel(&builder_, &model_, TensorType_INT8, TensorType_INT8, + TensorType_INT8, &error_reporter_); EXPECT_EQ(status, kTfLiteOk); const auto& subgraph = model_.subgraphs[0]; @@ -587,18 +640,25 @@ TEST_F(QuantizeConvModel1Test, VerifyConvQuantizationWithUnitScale) { EXPECT_EQ(model_.operator_codes[0]->version, 3); } -class QuantizeConvModel2Test : public QuantizeModelTest { +class QuantizeConvModel2Test : public QuantizeModelTest, + public testing::WithParamInterface { protected: QuantizeConvModel2Test() { + tensor_type_ = GetParam(); input_model_ = ReadModel(internal::kConvModelWith0Plus10Weights); readonly_model_ = input_model_->GetModel(); readonly_model_->UnPackTo(&model_); } -}; -TEST_F(QuantizeConvModel2Test, VerifyConvQuantization) { - auto status = QuantizeModel(&builder_, &model_, TensorType_INT8, - TensorType_INT8, &error_reporter_); + TensorType tensor_type_; +}; +INSTANTIATE_TEST_SUITE_P(QuantizeConvModel2TestInst, QuantizeConvModel2Test, + testing::ValuesIn({TensorType_INT8, + TensorType_INT16})); + +TEST_P(QuantizeConvModel2Test, VerifyConvQuantization) { + auto status = QuantizeModel(&builder_, &model_, tensor_type_, tensor_type_, + tensor_type_, &error_reporter_); ASSERT_EQ(kTfLiteOk, status); const auto& subgraph = model_.subgraphs[0]; auto conv_op = subgraph->operators[0].get(); @@ -615,8 +675,10 @@ TEST_F(QuantizeConvModel2Test, VerifyConvQuantization) { const auto output_tensor = subgraph->tensors[conv_op->outputs[output_tensor_idx]].get(); - EXPECT_EQ(bias_tensor->type, TensorType_INT32); - EXPECT_EQ(input_tensor->type, TensorType_INT8); + EXPECT_EQ(bias_tensor->type, tensor_type_ == TensorType_INT8 + ? TensorType_INT32 + : TensorType_INT64); + EXPECT_EQ(input_tensor->type, tensor_type_); EXPECT_EQ(weights_tensor->type, TensorType_INT8); ASSERT_TRUE(weights_tensor->quantization); @@ -644,17 +706,28 @@ TEST_F(QuantizeConvModel2Test, VerifyConvQuantization) { } const auto bias_buffer = model_.buffers[bias_tensor->buffer].get(); - ASSERT_EQ(bias_buffer->data.size(), sizeof(int32_t) * bias_tensor->shape[0]); - const int32_t* bias_values = - reinterpret_cast(bias_buffer->data.data()); + auto control_size = tensor_type_ == TensorType_INT8 + ? sizeof(int32_t) * bias_tensor->shape[0] + : sizeof(int64_t) * bias_tensor->shape[0]; + + ASSERT_EQ(bias_buffer->data.size(), control_size); const auto original_bias_buffer = readonly_model_->buffers()->Get(bias_tensor->buffer); const float* bias_float_buffer = reinterpret_cast(original_bias_buffer->data()->data()); - for (size_t i = 0; i < out_channel_size; i++) { - auto dequantized_value = bias_values[i] * bias_scales[i]; - EXPECT_NEAR(dequantized_value, bias_float_buffer[i], bias_scales[i] / 2); + if (tensor_type_ == TensorType_INT8) { + int32_t* bias_values = reinterpret_cast(bias_buffer->data.data()); + for (size_t i = 0; i < out_channel_size; i++) { + auto dequantized_value = bias_values[i] * bias_scales[i]; + EXPECT_NEAR(dequantized_value, bias_float_buffer[i], bias_scales[i] / 2); + } + } else if (tensor_type_ == TensorType_INT16) { + int64_t* bias_values = reinterpret_cast(bias_buffer->data.data()); + for (size_t i = 0; i < out_channel_size; i++) { + auto dequantized_value = bias_values[i] * bias_scales[i]; + EXPECT_NEAR(dequantized_value, bias_float_buffer[i], bias_scales[i] / 2); + } } const auto weights_buffer = model_.buffers[weights_tensor->buffer].get(); @@ -695,8 +768,9 @@ class QuantizeSoftmaxTest : public QuantizeModelTest { }; TEST_F(QuantizeSoftmaxTest, VerifySoftmaxQuantization) { - auto status = QuantizeModel(&builder_, &model_, TensorType_INT8, - TensorType_INT8, &error_reporter_); + auto status = + QuantizeModel(&builder_, &model_, TensorType_INT8, TensorType_INT8, + TensorType_INT8, &error_reporter_); ASSERT_EQ(kTfLiteOk, status); const auto& subgraph = model_.subgraphs[0]; @@ -755,8 +829,9 @@ class QuantizeAvgPoolTest : public QuantizeModelTest { }; TEST_F(QuantizeAvgPoolTest, VerifyAvgPoolQuantization) { - auto status = QuantizeModel(&builder_, &model_, TensorType_INT8, - TensorType_INT8, &error_reporter_); + auto status = + QuantizeModel(&builder_, &model_, TensorType_INT8, TensorType_INT8, + TensorType_INT8, &error_reporter_); ASSERT_EQ(kTfLiteOk, status); const auto& subgraph = model_.subgraphs[0]; @@ -816,8 +891,9 @@ class QuantizeMultiInputAddWithReshapeTest : public QuantizeModelTest { }; TEST_F(QuantizeMultiInputAddWithReshapeTest, VerifyReshapeQuantization) { - auto status = QuantizeModel(&builder_, &model_, TensorType_INT8, - TensorType_INT8, &error_reporter_); + auto status = + QuantizeModel(&builder_, &model_, TensorType_INT8, TensorType_INT8, + TensorType_INT8, &error_reporter_); ASSERT_EQ(kTfLiteOk, status); // Verify Reshape is quantized. @@ -863,8 +939,9 @@ TEST_F(QuantizeMultiInputAddWithReshapeTest, VerifyReshapeQuantization) { } TEST_F(QuantizeMultiInputAddWithReshapeTest, VerifyAddQuantization) { - auto status = QuantizeModel(&builder_, &model_, TensorType_INT8, - TensorType_INT8, &error_reporter_); + auto status = + QuantizeModel(&builder_, &model_, TensorType_INT8, TensorType_INT8, + TensorType_INT8, &error_reporter_); ASSERT_EQ(kTfLiteOk, status); // Verify ADD is quantized. @@ -923,8 +1000,9 @@ class QuantizeConstInputTest : public QuantizeModelTest { }; TEST_F(QuantizeConstInputTest, VerifyConstOpInput) { - auto status = QuantizeModel(&builder_, &model_, TensorType_INT8, - TensorType_INT8, &error_reporter_); + auto status = + QuantizeModel(&builder_, &model_, TensorType_INT8, TensorType_INT8, + TensorType_INT8, &error_reporter_); ASSERT_EQ(kTfLiteOk, status); // Verify ConstOp is quantized. @@ -965,8 +1043,9 @@ class QuantizeArgMaxTest : public QuantizeModelTest { }; TEST_F(QuantizeArgMaxTest, VerifyArgMax) { - auto status = QuantizeModel(&builder_, &model_, TensorType_INT8, - TensorType_INT8, &error_reporter_); + auto status = + QuantizeModel(&builder_, &model_, TensorType_INT8, TensorType_INT8, + TensorType_INT8, &error_reporter_); ASSERT_EQ(kTfLiteOk, status); const auto& subgraph = model_.subgraphs[0]; @@ -1008,8 +1087,9 @@ class QuantizeLSTMTest : public QuantizeModelTest { TEST_F(QuantizeLSTMTest, VerifyLSTM) { // Quantize model. - auto status = QuantizeModel(&builder_, &model_, TensorType_FLOAT32, - TensorType_FLOAT32, &error_reporter_); + auto status = + QuantizeModel(&builder_, &model_, TensorType_FLOAT32, TensorType_FLOAT32, + TensorType_INT8, &error_reporter_); ASSERT_EQ(kTfLiteOk, status); // Read expected model. @@ -1067,8 +1147,9 @@ class QuantizeLSTM2Test : public QuantizeModelTest { TEST_F(QuantizeLSTM2Test, VerifyLSTM) { // Quantize model. - auto status = QuantizeModel(&builder_, &model_, TensorType_FLOAT32, - TensorType_FLOAT32, &error_reporter_); + auto status = + QuantizeModel(&builder_, &model_, TensorType_FLOAT32, TensorType_FLOAT32, + TensorType_INT8, &error_reporter_); ASSERT_EQ(kTfLiteOk, status); // Read expected model. @@ -1126,8 +1207,9 @@ class QuantizeSVDFTest : public QuantizeModelTest { TEST_F(QuantizeSVDFTest, VerifySVDF) { // Quantize model. - auto status = QuantizeModel(&builder_, &model_, TensorType_INT8, - TensorType_INT8, &error_reporter_); + auto status = + QuantizeModel(&builder_, &model_, TensorType_INT8, TensorType_INT8, + TensorType_INT8, &error_reporter_); ASSERT_EQ(kTfLiteOk, status); // Read expected model. @@ -1184,8 +1266,9 @@ class QuantizeFCTest : public QuantizeModelTest { }; TEST_F(QuantizeFCTest, VerifyFC) { - auto status = QuantizeModel(&builder_, &model_, TensorType_INT8, - TensorType_INT8, &error_reporter_); + auto status = + QuantizeModel(&builder_, &model_, TensorType_INT8, TensorType_INT8, + TensorType_INT8, &error_reporter_); ASSERT_EQ(kTfLiteOk, status); const auto& subgraph = model_.subgraphs[0]; @@ -1236,7 +1319,7 @@ class QuantizeCustomOpTest : public QuantizeModelTest { TEST_F(QuantizeCustomOpTest, VerifyMixedQuantization) { auto status = QuantizeModel(&builder_, &model_, TensorType_INT8, TensorType_INT8, - /*allow_float=*/true, &error_reporter_); + /*allow_float=*/true, TensorType_INT8, &error_reporter_); ASSERT_EQ(kTfLiteOk, status); const auto& subgraph = model_.subgraphs[0]; auto float_graph = readonly_model_->subgraphs()->Get(0); @@ -1270,7 +1353,8 @@ class QuantizePackTest : public QuantizeModelTest { }; TEST_F(QuantizePackTest, VerifyPack) { - auto status = QuantizeModel(&builder_, &model_, &error_reporter_); + auto status = + QuantizeModel(&builder_, &model_, TensorType_INT8, &error_reporter_); ASSERT_EQ(kTfLiteOk, status); @@ -1334,7 +1418,8 @@ class QuantizeMinimumMaximumTest }; TEST_P(QuantizeMinimumMaximumTest, VerifyMinimumMaximum) { - auto status = QuantizeModel(&builder_, &model_, &error_reporter_); + auto status = + QuantizeModel(&builder_, &model_, TensorType_INT8, &error_reporter_); ASSERT_EQ(kTfLiteOk, status); const auto& subgraph = model_.subgraphs[0]; @@ -1415,7 +1500,8 @@ class QuantizeUnpackTest : public QuantizeModelTest { } }; TEST_F(QuantizeUnpackTest, VerifyUnpack) { - auto status = QuantizeModel(&builder_, &model_, &error_reporter_); + auto status = + QuantizeModel(&builder_, &model_, TensorType_INT8, &error_reporter_); ASSERT_EQ(kTfLiteOk, status); From de6afc5d6b509c8f3d709bf1e275373864ec0936 Mon Sep 17 00:00:00 2001 From: Elena Zhelezina Date: Wed, 12 Feb 2020 13:51:24 +0000 Subject: [PATCH 015/867] Changed per reviewer comments. --- .../python/optimize/calibration_wrapper.cc | 3 +- .../lite/tools/optimize/quantization_utils.cc | 10 ++----- .../lite/tools/optimize/quantize_model.cc | 19 +++++++++---- .../lite/tools/optimize/quantize_model.h | 25 +++++++++++++---- .../tools/optimize/quantize_model_test.cc | 28 +++++++++---------- 5 files changed, 51 insertions(+), 34 deletions(-) diff --git a/tensorflow/lite/python/optimize/calibration_wrapper.cc b/tensorflow/lite/python/optimize/calibration_wrapper.cc index 88995136726..cdc8adaaf2b 100644 --- a/tensorflow/lite/python/optimize/calibration_wrapper.cc +++ b/tensorflow/lite/python/optimize/calibration_wrapper.cc @@ -266,7 +266,8 @@ PyObject* CalibrationWrapper::QuantizeModel(int input_py_type, auto status = tflite::optimize::QuantizeModel( &builder, tflite_model.get(), TfLiteTypeToSchemaType(input_type), TfLiteTypeToSchemaType(output_type), allow_float, {op_name}, - TensorType_INT8, error_reporter_.get()); + TensorType_INT8, + error_reporter_.get()); if (status != kTfLiteOk) { error_reporter_->exception(); return nullptr; diff --git a/tensorflow/lite/tools/optimize/quantization_utils.cc b/tensorflow/lite/tools/optimize/quantization_utils.cc index 4bc9686ec2c..ba43416cf04 100644 --- a/tensorflow/lite/tools/optimize/quantization_utils.cc +++ b/tensorflow/lite/tools/optimize/quantization_utils.cc @@ -108,14 +108,10 @@ TfLiteStatus GetQuantizationParams(TensorT* tensor, TensorType activations_type, std::numeric_limits::min(), std::numeric_limits::max(), quantization_params); } else if (activations_type == TensorType_INT16) { - float range = std::max(std::abs(tensor->quantization->min[0]), - std::abs(tensor->quantization->max[0])); const float quantized_range = 32767.0; - const float scale = range / quantized_range; - quantization_params->min = std::vector(1, -range); - quantization_params->max = std::vector(1, range); - quantization_params->scale = std::vector(1, scale); - quantization_params->zero_point = std::vector(1, 0); + GetSymmetricQuantizationParams(tensor->quantization->min[0], + tensor->quantization->max[0], + quantized_range, quantization_params); } else { error_reporter->Report( "Unsupported activation type for quantize-activation: %s", diff --git a/tensorflow/lite/tools/optimize/quantize_model.cc b/tensorflow/lite/tools/optimize/quantize_model.cc index ee562fe9c4c..bbb40080fbc 100644 --- a/tensorflow/lite/tools/optimize/quantize_model.cc +++ b/tensorflow/lite/tools/optimize/quantize_model.cc @@ -1210,18 +1210,25 @@ TfLiteStatus QuantizeModel(flatbuffers::FlatBufferBuilder* builder, TfLiteStatus QuantizeModel(flatbuffers::FlatBufferBuilder* builder, ModelT* model, const TensorType& input_type, - const TensorType& output_type, - const TensorType& activations_type, + const TensorType& output_type, bool allow_float, ErrorReporter* error_reporter) { - return QuantizeModel(builder, model, input_type, output_type, - /*allow_float=*/false, activations_type, error_reporter); + return QuantizeModel(builder, model, input_type, output_type, allow_float, + GetAllOperatorOutputs(model), TensorType_INT8, + error_reporter); } TfLiteStatus QuantizeModel(flatbuffers::FlatBufferBuilder* builder, - ModelT* model, const TensorType& activations_type, + ModelT* model, const TensorType& input_type, + const TensorType& output_type, ErrorReporter* error_reporter) { + return QuantizeModel(builder, model, input_type, output_type, + /*allow_float=*/false, error_reporter); +} + +TfLiteStatus QuantizeModel(flatbuffers::FlatBufferBuilder* builder, + ModelT* model, ErrorReporter* error_reporter) { return QuantizeModel(builder, model, TensorType_FLOAT32, TensorType_FLOAT32, - /*allow_float=*/false, activations_type, error_reporter); + /*allow_float=*/false, error_reporter); } } // namespace optimize diff --git a/tensorflow/lite/tools/optimize/quantize_model.h b/tensorflow/lite/tools/optimize/quantize_model.h index cc801ec9870..06c30b88fd0 100644 --- a/tensorflow/lite/tools/optimize/quantize_model.h +++ b/tensorflow/lite/tools/optimize/quantize_model.h @@ -35,9 +35,7 @@ namespace optimize { // // Note: This is a private API, subject to change. TfLiteStatus QuantizeModel(flatbuffers::FlatBufferBuilder* builder, - ModelT* input_model, - const TensorType& activations_type, - ErrorReporter* error_reporter); + ModelT* input_model, ErrorReporter* error_reporter); // Same as above, but the types of quantized inputs and outputs are // configurable. @@ -46,7 +44,6 @@ TfLiteStatus QuantizeModel(flatbuffers::FlatBufferBuilder* builder, TfLiteStatus QuantizeModel(flatbuffers::FlatBufferBuilder* builder, ModelT* input_model, const TensorType& input_type, const TensorType& output_type, - const TensorType& activations_type, ErrorReporter* error_reporter); // Same as above, but can enable allowing float intermediate operations for ops @@ -56,7 +53,6 @@ TfLiteStatus QuantizeModel(flatbuffers::FlatBufferBuilder* builder, TfLiteStatus QuantizeModel(flatbuffers::FlatBufferBuilder* builder, ModelT* input_model, const TensorType& input_type, const TensorType& output_type, bool allow_float, - const TensorType& activations_type, ErrorReporter* error_reporter); // Same as above, but enables only quantizing a whitelist of operations, @@ -67,6 +63,25 @@ TfLiteStatus QuantizeModel(flatbuffers::FlatBufferBuilder* builder, ModelT* input_model, const TensorType& input_type, const TensorType& output_type, bool allow_float, const std::unordered_set& operator_names, + ErrorReporter* error_reporter); + +// Same as above, but enables to provide activation type, which +// could be TensorType_INT16 or TensorType_INT8. +// +// Note: This is a private API, subject to change. +TfLiteStatus QuantizeModel(flatbuffers::FlatBufferBuilder* builder, + ModelT* model, const TensorType& input_type, + const TensorType& output_type, bool allow_float, + const TensorType& activations_type, + ErrorReporter* error_reporter); + +// Quantizes input_model and populates the provided builder with the new model +// with all possible input parameters. +// All functions above call this function underneath. +TfLiteStatus QuantizeModel(flatbuffers::FlatBufferBuilder* builder, + ModelT* model, const TensorType& input_type, + const TensorType& output_type, bool allow_float, + const std::unordered_set& operator_names, const TensorType& activations_type, ErrorReporter* error_reporter); diff --git a/tensorflow/lite/tools/optimize/quantize_model_test.cc b/tensorflow/lite/tools/optimize/quantize_model_test.cc index 166d60ecc66..ef46b3fbd5d 100644 --- a/tensorflow/lite/tools/optimize/quantize_model_test.cc +++ b/tensorflow/lite/tools/optimize/quantize_model_test.cc @@ -106,9 +106,9 @@ TEST_P(QuantizeConvModelTest, QuantizationSucceeds) { } TEST_P(QuantizeConvModelTest, SkipUnspecifiedLayer) { - auto status = - QuantizeModel(&builder_, &model_, TensorType_FLOAT32, TensorType_FLOAT32, - /*allow_float=*/true, {}, tensor_type_, &error_reporter_); + auto status = QuantizeModel( + &builder_, &model_, TensorType_FLOAT32, TensorType_FLOAT32, + /*allow_float=*/true, {}, TensorType_FLOAT32, &error_reporter_); EXPECT_EQ(status, kTfLiteOk); ASSERT_EQ(model_.subgraphs.size(), readonly_model_->subgraphs()->size()); // The resulting model should be the same. @@ -190,8 +190,9 @@ TEST_P(QuantizeConvModelTest, OperatorsAreUnchanged) { } TEST_P(QuantizeConvModelTest, GraphIsFullyQuantized) { - auto status = QuantizeModel(&builder_, &model_, tensor_type_, tensor_type_, - tensor_type_, &error_reporter_); + auto status = + QuantizeModel(&builder_, &model_, tensor_type_, tensor_type_, + /*allow_float*/ false, tensor_type_, &error_reporter_); EXPECT_EQ(status, kTfLiteOk); for (const auto& subgraph : model_.subgraphs) { for (const auto& tensor : subgraph->tensors) { @@ -210,7 +211,7 @@ TEST_P(QuantizeConvModelTest, GraphIsFullyQuantized) { TEST_P(QuantizeConvModelTest, FloatInputAndOutput) { auto status = QuantizeModel(&builder_, &model_, TensorType_FLOAT32, TensorType_FLOAT32, - tensor_type_, &error_reporter_); + /*allow_float*/ false, tensor_type_, &error_reporter_); EXPECT_EQ(status, kTfLiteOk); for (int32_t subgraph_idx = 0; subgraph_idx < model_.subgraphs.size(); @@ -384,7 +385,7 @@ class QuantizeConcatModelTest : public QuantizeModelTest, // input1 / TEST_P(QuantizeConcatModelTest, AddRequantBeforeConcat) { auto status = QuantizeModel(&builder_, &model_, tensor_type_, tensor_type_, - tensor_type_, &error_reporter_); + false, tensor_type_, &error_reporter_); EXPECT_EQ(status, kTfLiteOk); // There is only one subgraph. @@ -549,7 +550,7 @@ class QuantizeConvModel1Test : public QuantizeModelTest { TEST_F(QuantizeConvModel1Test, VerifyConvQuantizationWithUnitScale) { auto status = - QuantizeModel(&builder_, &model_, TensorType_INT8, TensorType_INT8, + QuantizeModel(&builder_, &model_, TensorType_INT8, TensorType_INT8, false, TensorType_INT8, &error_reporter_); EXPECT_EQ(status, kTfLiteOk); const auto& subgraph = model_.subgraphs[0]; @@ -658,7 +659,7 @@ INSTANTIATE_TEST_SUITE_P(QuantizeConvModel2TestInst, QuantizeConvModel2Test, TEST_P(QuantizeConvModel2Test, VerifyConvQuantization) { auto status = QuantizeModel(&builder_, &model_, tensor_type_, tensor_type_, - tensor_type_, &error_reporter_); + false, tensor_type_, &error_reporter_); ASSERT_EQ(kTfLiteOk, status); const auto& subgraph = model_.subgraphs[0]; auto conv_op = subgraph->operators[0].get(); @@ -1353,8 +1354,7 @@ class QuantizePackTest : public QuantizeModelTest { }; TEST_F(QuantizePackTest, VerifyPack) { - auto status = - QuantizeModel(&builder_, &model_, TensorType_INT8, &error_reporter_); + auto status = QuantizeModel(&builder_, &model_, &error_reporter_); ASSERT_EQ(kTfLiteOk, status); @@ -1418,8 +1418,7 @@ class QuantizeMinimumMaximumTest }; TEST_P(QuantizeMinimumMaximumTest, VerifyMinimumMaximum) { - auto status = - QuantizeModel(&builder_, &model_, TensorType_INT8, &error_reporter_); + auto status = QuantizeModel(&builder_, &model_, &error_reporter_); ASSERT_EQ(kTfLiteOk, status); const auto& subgraph = model_.subgraphs[0]; @@ -1500,8 +1499,7 @@ class QuantizeUnpackTest : public QuantizeModelTest { } }; TEST_F(QuantizeUnpackTest, VerifyUnpack) { - auto status = - QuantizeModel(&builder_, &model_, TensorType_INT8, &error_reporter_); + auto status = QuantizeModel(&builder_, &model_, &error_reporter_); ASSERT_EQ(kTfLiteOk, status); From 792f553fd078a425d66c81567ca8f3588d44fcdc Mon Sep 17 00:00:00 2001 From: Elena Zhelezina Date: Wed, 5 Feb 2020 11:55:27 +0000 Subject: [PATCH 016/867] Added non-strict mode for 16x8 quantization --- tensorflow/lite/python/lite.py | 16 ++++-- tensorflow/lite/tools/optimize/BUILD | 1 + .../lite/tools/optimize/operator_property.cc | 16 ++++++ .../lite/tools/optimize/operator_property.h | 3 +- .../lite/tools/optimize/quantize_model.cc | 50 +++++++++++------- .../tools/optimize/quantize_model_test.cc | 50 +++++++++++++++--- tensorflow/lite/tools/optimize/test_util.cc | 1 + tensorflow/lite/tools/optimize/test_util.h | 5 ++ .../tools/optimize/testdata/mixed16x8.bin | Bin 0 -> 1184 bytes 9 files changed, 111 insertions(+), 31 deletions(-) create mode 100644 tensorflow/lite/tools/optimize/testdata/mixed16x8.bin diff --git a/tensorflow/lite/python/lite.py b/tensorflow/lite/python/lite.py index fc9c064faf0..1e0c89d3aa5 100644 --- a/tensorflow/lite/python/lite.py +++ b/tensorflow/lite/python/lite.py @@ -220,13 +220,16 @@ class TFLiteConverterBase(object): "type to be INT8.") def _is_int8_target_required(self): - return (set([OpsSet.TFLITE_BUILTINS_INT8]) == set( + return ((set([OpsSet.TFLITE_BUILTINS_INT8]) == set( self.target_spec.supported_ops) or - self._smallest_supported_type() == constants.INT8) + self._smallest_supported_type() == constants.INT8) and + not self._is_int16x8_target_required()) def _is_int16x8_target_required(self): - return (set([OpsSet.TFLITE_BUILTINS_ACTIVATIONS_INT16_WEIGHTS_INT8]) == - set(self.target_spec.supported_ops)) + return bool( + set(self.target_spec.supported_ops).intersection([ + OpsSet.TFLITE_BUILTINS_ACTIVATIONS_INT16_WEIGHTS_INT8 + ])) def _smallest_supported_type(self): if self.target_spec.supported_types: @@ -262,6 +265,11 @@ class TFLiteConverterBase(object): def _calibrate_quantize_model(self, result, inference_input_type, inference_output_type, enable_mlir_quantizer): allow_float = not self._is_int8_target_required() and not self._is_int16x8_target_required() + if (self._is_int16x8_target_required()): + allow_float = bool( + set(self.target_spec.supported_ops).intersection([ + OpsSet.TFLITE_BUILTINS + ])) calibrate_quantize = _calibrator.Calibrator(result) activations_type = constants.INT16 if self._is_int16x8_target_required() else constants.INT8 return calibrate_quantize.calibrate_and_quantize( diff --git a/tensorflow/lite/tools/optimize/BUILD b/tensorflow/lite/tools/optimize/BUILD index 27be0f829ba..ee5e845b96b 100644 --- a/tensorflow/lite/tools/optimize/BUILD +++ b/tensorflow/lite/tools/optimize/BUILD @@ -245,6 +245,7 @@ tf_cc_test( "//tensorflow/lite/tools/optimize:testdata/maximum.bin", "//tensorflow/lite/tools/optimize:testdata/minimum.bin", "//tensorflow/lite/tools/optimize:testdata/mixed.bin", + "//tensorflow/lite/tools/optimize:testdata/mixed16x8.bin", "//tensorflow/lite/tools/optimize:testdata/multi_input_add_reshape.bin", "//tensorflow/lite/tools/optimize:testdata/pack.bin", "//tensorflow/lite/tools/optimize:testdata/single_avg_pool_min_minus_5_max_plus_5.bin", diff --git a/tensorflow/lite/tools/optimize/operator_property.cc b/tensorflow/lite/tools/optimize/operator_property.cc index 1f2d8bb4a4d..c31ad9dbb1e 100644 --- a/tensorflow/lite/tools/optimize/operator_property.cc +++ b/tensorflow/lite/tools/optimize/operator_property.cc @@ -70,6 +70,7 @@ OperatorProperty GetOperatorProperty(const ModelT* model, int subgraph_index, property.inputs = {{0, {}}}; // ArgMax has no quantizable output. property.version = 2; + property.quantizable_int16 = false; break; case BuiltinOperator_AVERAGE_POOL_2D: property.inputs = {{0, {}}}; @@ -85,6 +86,7 @@ OperatorProperty GetOperatorProperty(const ModelT* model, int subgraph_index, property.outputs = {{0, {}}}; property.restrict_same_input_output_scale = true; property.version = 2; + property.quantizable_int16 = false; break; case BuiltinOperator_SPLIT: // We skip input 0 since it is the split dim which is not real valued. @@ -143,6 +145,7 @@ OperatorProperty GetOperatorProperty(const ModelT* model, int subgraph_index, property.inputs = {{0, {}}, {1, {}}}; // Comparisons have no quantizable outputs. property.version = 2; + property.quantizable_int16 = false; break; case BuiltinOperator_EXPAND_DIMS: // We skip input 1 as it is not real valued (it's the index of axis) and @@ -165,11 +168,13 @@ OperatorProperty GetOperatorProperty(const ModelT* model, int subgraph_index, property.outputs = {{0, {}}}; property.restrict_same_input_output_scale = true; property.version = 2; + property.quantizable_int16 = false; break; case BuiltinOperator_HARD_SWISH: { property.inputs = {{0, {}}}; property.outputs = {{0, {}}}; property.version = 1; + property.quantizable_int16 = false; break; } case BuiltinOperator_LOG_SOFTMAX: { @@ -180,6 +185,7 @@ OperatorProperty GetOperatorProperty(const ModelT* model, int subgraph_index, tensor_property.restricted_value_int8 = {16.0 / 256.0, 127}; property.outputs = {{0, tensor_property}}; property.version = 2; + property.quantizable_int16 = false; break; } case BuiltinOperator_LOGISTIC: { @@ -736,6 +742,7 @@ OperatorProperty GetOperatorProperty(const ModelT* model, int subgraph_index, property.restrict_scale = {{18, 0}}; property.version = 2; } + property.quantizable_int16 = false; break; } case BuiltinOperator_L2_NORMALIZATION: { @@ -746,6 +753,7 @@ OperatorProperty GetOperatorProperty(const ModelT* model, int subgraph_index, tensor_property.restricted_value_int8 = {1 / 128.0, 0}; property.outputs = {{0, tensor_property}}; property.version = 2; + property.quantizable_int16 = false; break; } case BuiltinOperator_MAX_POOL_2D: @@ -765,6 +773,7 @@ OperatorProperty GetOperatorProperty(const ModelT* model, int subgraph_index, property.inputs = {{0, {}}}; property.outputs = {{0, {}}}; property.version = 2; + property.quantizable_int16 = false; break; case BuiltinOperator_MINIMUM: property.arbitrary_inputs = true; @@ -791,6 +800,7 @@ OperatorProperty GetOperatorProperty(const ModelT* model, int subgraph_index, property.outputs = {{0, {}}}; property.restrict_same_input_output_scale = true; property.version = 2; + property.quantizable_int16 = false; break; case BuiltinOperator_QUANTIZE: property.inputs = {{0, {}}}; @@ -802,11 +812,13 @@ OperatorProperty GetOperatorProperty(const ModelT* model, int subgraph_index, property.inputs = {{0, {}}}; property.outputs = {{0, {}}}; property.version = 2; + property.quantizable_int16 = false; break; case BuiltinOperator_RELU_N1_TO_1: property.inputs = {{0, {}}}; property.outputs = {{0, {}}}; property.version = 1; + property.quantizable_int16 = false; break; case BuiltinOperator_RESHAPE: property.inputs = {{0, {}}}; @@ -820,6 +832,7 @@ OperatorProperty GetOperatorProperty(const ModelT* model, int subgraph_index, property.outputs = {{0, {}}}; property.restrict_same_input_output_scale = true; property.version = 2; + property.quantizable_int16 = false; break; case BuiltinOperator_SHAPE: property.inputs = {{0, {}}}; @@ -866,6 +879,7 @@ OperatorProperty GetOperatorProperty(const ModelT* model, int subgraph_index, property.inputs = {{0, {}}}; property.outputs = {{0, {}}}; property.version = 2; + property.quantizable_int16 = false; break; case BuiltinOperator_TANH: { property.inputs = {{0, {}}}; @@ -899,6 +913,7 @@ OperatorProperty GetOperatorProperty(const ModelT* model, int subgraph_index, {3, tensor_property_bias}}; property.outputs = {{0, {}}}; property.version = 3; + property.quantizable_int16 = false; break; } case BuiltinOperator_TRANSPOSE: @@ -916,6 +931,7 @@ OperatorProperty GetOperatorProperty(const ModelT* model, int subgraph_index, default: // No quantized implementation exists for this operation. property.quantizable = false; + property.quantizable_int16 = false; } return property; } diff --git a/tensorflow/lite/tools/optimize/operator_property.h b/tensorflow/lite/tools/optimize/operator_property.h index 23052308568..151e314f335 100644 --- a/tensorflow/lite/tools/optimize/operator_property.h +++ b/tensorflow/lite/tools/optimize/operator_property.h @@ -65,7 +65,8 @@ struct TensorProperty { struct OperatorProperty { // Is a quantized operations currently supported. bool quantizable = true; - + // Is a quantized operations currently supported for 16x8 + bool quantizable_int16 = true; // Op has arbitrary number of inputs, such as concat. bool arbitrary_inputs = false; // Op has arbitrary number of outputs, such as slice. diff --git a/tensorflow/lite/tools/optimize/quantize_model.cc b/tensorflow/lite/tools/optimize/quantize_model.cc index bbb40080fbc..ceae3c29d9e 100644 --- a/tensorflow/lite/tools/optimize/quantize_model.cc +++ b/tensorflow/lite/tools/optimize/quantize_model.cc @@ -43,13 +43,17 @@ namespace { // operator_names. operator_property::OperatorProperty GetOperatorProperty( const std::unordered_set& operator_names, const ModelT* model, - int subgraph_index, int op_idx, const string& operator_name) { + int subgraph_index, int op_idx, const string& operator_name, + const TensorType& activations_type) { operator_property::OperatorProperty property = operator_property::GetOperatorProperty(model, subgraph_index, op_idx); const OperatorT* op = model->subgraphs[subgraph_index]->operators[op_idx].get(); const BuiltinOperator op_code = model->operator_codes[op->opcode_index]->builtin_code; + if (activations_type == TensorType_INT16 && !property.quantizable_int16) { + property.quantizable = false; + } // The algorithm adds Dequantize and Quantize, so we don't require them to be // in the operator_names. if (op_code != BuiltinOperator_DEQUANTIZE && @@ -320,9 +324,9 @@ TfLiteStatus ApplyConstraints(ModelT* model, // Iterate backward to avoid messing with index. for (int op_idx = subgraph->operators.size() - 1; op_idx >= 0; op_idx--) { OperatorT* op = subgraph->operators[op_idx].get(); - operator_property::OperatorProperty property = - GetOperatorProperty(operator_names, model, subgraph_idx, op_idx, - subgraph->tensors[op->outputs[0]]->name); + operator_property::OperatorProperty property = GetOperatorProperty( + operator_names, model, subgraph_idx, op_idx, + subgraph->tensors[op->outputs[0]]->name, activations_type); if (!property.quantizable) { continue; } @@ -840,11 +844,17 @@ TfLiteStatus QuantizeWeightsInputOutput( OperatorT* op = subgraph->operators[op_idx].get(); const BuiltinOperator op_code = model->operator_codes[op->opcode_index]->builtin_code; - operator_property::OperatorProperty property = - GetOperatorProperty(operator_names, model, subgraph_idx, op_idx, - subgraph->tensors[op->outputs[0]]->name); + operator_property::OperatorProperty property = GetOperatorProperty( + operator_names, model, subgraph_idx, op_idx, + subgraph->tensors[op->outputs[0]]->name, activations_type); - if (!property.quantizable && !allow_float) { + if (activations_type == TensorType_INT16 && !property.quantizable && + !allow_float) { + error_reporter->Report( + "Quantization to 16x8-bit not yet supported for op: %s", + EnumNameBuiltinOperator(op_code)); + return kTfLiteError; + } else if (!property.quantizable && !allow_float) { error_reporter->Report("Quantization not yet supported for op: %s", EnumNameBuiltinOperator(op_code)); return kTfLiteError; @@ -882,9 +892,9 @@ TfLiteStatus QuantizeBiases(ModelT* model, OperatorT* op = subgraph->operators[op_idx].get(); const BuiltinOperator op_code = model->operator_codes[op->opcode_index]->builtin_code; - operator_property::OperatorProperty property = - GetOperatorProperty(operator_names, model, subgraph_idx, op_idx, - subgraph->tensors[op->outputs[0]]->name); + operator_property::OperatorProperty property = GetOperatorProperty( + operator_names, model, subgraph_idx, op_idx, + subgraph->tensors[op->outputs[0]]->name, activations_type); if (!property.quantizable) { continue; } @@ -951,15 +961,15 @@ std::unordered_set GetAllOperatorOutputs(ModelT* model) { // will not be filled by this function. TfLiteStatus FillQuantizationParams( ModelT* model, const std::unordered_set& operator_names, - ErrorReporter* error_reporter) { + const TensorType& activations_type, ErrorReporter* error_reporter) { for (size_t subgraph_idx = 0; subgraph_idx < model->subgraphs.size(); subgraph_idx++) { SubGraphT* subgraph = model->subgraphs.at(subgraph_idx).get(); for (size_t op_idx = 0; op_idx < subgraph->operators.size(); op_idx++) { OperatorT* op = subgraph->operators[op_idx].get(); - operator_property::OperatorProperty property = - GetOperatorProperty(operator_names, model, subgraph_idx, op_idx, - subgraph->tensors[op->outputs[0]]->name); + operator_property::OperatorProperty property = GetOperatorProperty( + operator_names, model, subgraph_idx, op_idx, + subgraph->tensors[op->outputs[0]]->name, activations_type); // Populate max, min for each input tensor. for (const std::pair& input : @@ -1048,9 +1058,9 @@ TfLiteStatus EnsureBiasScaleCompatibility( SubGraphT* subgraph = model->subgraphs.at(subgraph_idx).get(); for (size_t op_idx = 0; op_idx < subgraph->operators.size(); op_idx++) { OperatorT* op = subgraph->operators[op_idx].get(); - operator_property::OperatorProperty property = - GetOperatorProperty(operator_names, model, subgraph_idx, op_idx, - subgraph->tensors[op->outputs[0]]->name); + operator_property::OperatorProperty property = GetOperatorProperty( + operator_names, model, subgraph_idx, op_idx, + subgraph->tensors[op->outputs[0]]->name, activations_type); // Loop over all bias tensors. for (const int bias_idx : property.biases) { @@ -1174,8 +1184,8 @@ TfLiteStatus QuantizeModel(flatbuffers::FlatBufferBuilder* builder, const std::unordered_set& operator_names, const TensorType& activations_type, ErrorReporter* error_reporter) { - TF_LITE_ENSURE_STATUS( - FillQuantizationParams(model, operator_names, error_reporter)); + TF_LITE_ENSURE_STATUS(FillQuantizationParams( + model, operator_names, activations_type, error_reporter)); TF_LITE_ENSURE_STATUS(EnsureBiasScaleCompatibility( model, operator_names, activations_type, error_reporter)); TF_LITE_ENSURE_STATUS( diff --git a/tensorflow/lite/tools/optimize/quantize_model_test.cc b/tensorflow/lite/tools/optimize/quantize_model_test.cc index ef46b3fbd5d..b73cb9a79ca 100644 --- a/tensorflow/lite/tools/optimize/quantize_model_test.cc +++ b/tensorflow/lite/tools/optimize/quantize_model_test.cc @@ -1308,7 +1308,8 @@ TEST_F(QuantizeFCTest, VerifyFC) { EXPECT_EQ(model_.operator_codes[1]->version, 1); } -class QuantizeCustomOpTest : public QuantizeModelTest { +class QuantizeCustomOpTest : public QuantizeModelTest, + public ::testing::WithParamInterface { protected: QuantizeCustomOpTest() { input_model_ = ReadModel(internal::kModelMixed); @@ -1317,10 +1318,10 @@ class QuantizeCustomOpTest : public QuantizeModelTest { } }; -TEST_F(QuantizeCustomOpTest, VerifyMixedQuantization) { +TEST_P(QuantizeCustomOpTest, VerifyMixedQuantization) { auto status = - QuantizeModel(&builder_, &model_, TensorType_INT8, TensorType_INT8, - /*allow_float=*/true, TensorType_INT8, &error_reporter_); + QuantizeModel(&builder_, &model_, GetParam(), GetParam(), + /*allow_float=*/true, GetParam(), &error_reporter_); ASSERT_EQ(kTfLiteOk, status); const auto& subgraph = model_.subgraphs[0]; auto float_graph = readonly_model_->subgraphs()->Get(0); @@ -1334,8 +1335,45 @@ TEST_F(QuantizeCustomOpTest, VerifyMixedQuantization) { BuiltinOperator_CUSTOM, BuiltinOperator_CUSTOM, BuiltinOperator_QUANTIZE, BuiltinOperator_SQUEEZE}; const std::vector op_input_types = { - TensorType_INT8, TensorType_INT8, TensorType_FLOAT32, - TensorType_FLOAT32, TensorType_FLOAT32, TensorType_INT8}; + GetParam(), GetParam(), TensorType_FLOAT32, + TensorType_FLOAT32, TensorType_FLOAT32, GetParam()}; + for (int i = 0; i < subgraph->operators.size(); ++i) { + OperatorT* op = subgraph->operators[i].get(); + ASSERT_EQ(model_.operator_codes[op->opcode_index]->builtin_code, + op_codes[i]); + ASSERT_EQ(subgraph->tensors[op->inputs[0]]->type, op_input_types[i]); + } +} + +INSTANTIATE_TEST_SUITE_P(QuantizeCustomOpTest, QuantizeCustomOpTest, + ::testing::Values(TensorType_INT8, TensorType_INT16)); + +class QuantizeOp16x8Test : public QuantizeModelTest { + protected: + QuantizeOp16x8Test() { + input_model_ = ReadModel(internal::kModelMixed16x8); + readonly_model_ = input_model_->GetModel(); + readonly_model_->UnPackTo(&model_); + } +}; + +TEST_F(QuantizeOp16x8Test, VerifyMixedQuantization16x8) { + auto status = + QuantizeModel(&builder_, &model_, TensorType_INT16, TensorType_FLOAT32, + /*allow_float=*/true, TensorType_INT16, &error_reporter_); + ASSERT_EQ(kTfLiteOk, status); + const auto& subgraph = model_.subgraphs[0]; + auto float_graph = readonly_model_->subgraphs()->Get(0); + // The original model conv_2d->log_softmax + ASSERT_EQ(float_graph->operators()->size(), 2); + // The resulting model should be: + // conv_2d->dequantize->log_softmax + ASSERT_EQ(subgraph->operators.size(), 3); + const std::vector op_codes = { + BuiltinOperator_CONV_2D, BuiltinOperator_DEQUANTIZE, + BuiltinOperator_LOG_SOFTMAX}; + const std::vector op_input_types = { + TensorType_INT16, TensorType_INT16, TensorType_FLOAT32}; for (int i = 0; i < subgraph->operators.size(); ++i) { OperatorT* op = subgraph->operators[i].get(); ASSERT_EQ(model_.operator_codes[op->opcode_index]->builtin_code, diff --git a/tensorflow/lite/tools/optimize/test_util.cc b/tensorflow/lite/tools/optimize/test_util.cc index 7d5e9d65f06..379be64059f 100644 --- a/tensorflow/lite/tools/optimize/test_util.cc +++ b/tensorflow/lite/tools/optimize/test_util.cc @@ -48,6 +48,7 @@ const char* kModelWithArgMaxOp = "argmax.bin"; const char* kModelWithFCOp = "fc.bin"; const char* kModelMixed = "mixed.bin"; +const char* kModelMixed16x8 = "mixed16x8.bin"; const char* kModelSplit = "split.bin"; diff --git a/tensorflow/lite/tools/optimize/test_util.h b/tensorflow/lite/tools/optimize/test_util.h index abcdbc21d36..a49f3500288 100644 --- a/tensorflow/lite/tools/optimize/test_util.h +++ b/tensorflow/lite/tools/optimize/test_util.h @@ -76,6 +76,11 @@ extern const char* kModelWithFCOp; // reshape->custom->custom->squeeze. extern const char* kModelMixed; +// Test model with mixed quantizable and +// and un-quantizable ops for +// activations in 16-bit. +extern const char* kModelMixed16x8; + // Test model with split op. extern const char* kModelSplit; diff --git a/tensorflow/lite/tools/optimize/testdata/mixed16x8.bin b/tensorflow/lite/tools/optimize/testdata/mixed16x8.bin new file mode 100644 index 0000000000000000000000000000000000000000..c1f615e966eb164341d8ba4d56e1e359ef516388 GIT binary patch literal 1184 zcmb7E&ubG=5T2wl2_@9f2FxK8g;sg<_q<}Jbu!5~elfPXzj(rrE0@_xm+3R;4t8F&7%-*Ke z?MMvr&%hhtIgkZrfDAxiYp&~((nm2@EawU$tB@JMEnjCt_u<}P&Y0ZMYI|w5Y_-(Y z?}Kw!60=(O(}H?o7SzJdGcC3BTKlp4bnfJZ+^4yov`huYh4Sv}6>)b{(EDuQg76SInA$K1#dWu*K>jslH( zwSK(X?clzSY2>`xBq2|rjzL8e;CF#{zyRRcFt7rofz-$rjJJkeV4bPw`5-oBT%<(G zvtZ|9=p7fmc`I}VBn!|-a9(5_b}_Hl0$%}T--8^vr<{wF@*}`kFb#nBTJh!c ze_WR_X*1)~GGxq)vZ0^b)gcvcHmuG@-)WXcdio|9_W{lt{<*;4odf(8VSCxp9kbQ9 zFX=tUaxj6WQPpib-|iZw-7MzU(W*k84J*}}YnHz0`)5vJ%$p6$*g?FX2gIwu?lAH4 zOQvot>pe?fv(5U(=+2B6;4X4l1 Date: Tue, 17 Mar 2020 16:30:32 +0000 Subject: [PATCH 017/867] Corrected after merge with master. Tested: strict mode and non-strict mode. Change-Id: I7e03d08133f39cc65a18875e65ce5cdddaf2d6a4 --- tensorflow/lite/python/lite.py | 62 ++++++++++++++++--- .../optimize/calibration_wrapper_pybind11.cc | 8 +-- 2 files changed, 59 insertions(+), 11 deletions(-) diff --git a/tensorflow/lite/python/lite.py b/tensorflow/lite/python/lite.py index 7e5f8ce704f..900398d7a6f 100644 --- a/tensorflow/lite/python/lite.py +++ b/tensorflow/lite/python/lite.py @@ -180,11 +180,13 @@ class QuantizationMode(object): def post_training_int8_no_float(self): """Post training int8 quantize, disallow float fallback.""" return (self._is_int8_target_required() and + not self._is_int16x8_target_required() and self._representative_dataset is not None) def post_training_int8_allow_float(self): """Post training int8 quantize, allow float fallback.""" return (self._any_optimization_enabled() and + not self._is_int16x8_target_required() and self._representative_dataset is not None and self._smallest_supported_type() == constants.INT8) @@ -193,6 +195,18 @@ class QuantizationMode(object): return (self._any_optimization_enabled() and self._contains_training_quant_op()) + def post_training_int16x8_no_float(self): + """Post training int16x8 quantize, disallow float fallback.""" + return (not self._is_int8_target_required() and + self._is_int16x8_target_required() and + not self._is_allow_float() and + self._representative_dataset is not None) + + def post_training_int16x8_allow_float(self): + """Post training int16x8 quantize, allow float fallback.""" + return (self._is_int16x8_target_required() and + self._is_allow_float()) + def post_training_dynamic_range_int8(self): """Post training int8 const, on-the-fly int8 quantize of dynamic tensors.""" # Post-training dynamic range quantization is only enabled if post-training @@ -212,9 +226,14 @@ class QuantizationMode(object): return not (self.post_training_int8_no_float() or self.post_training_int8_allow_float() or self.training_time_int8_allow_float() or + self.post_training_int16x8_no_float() or + self.post_training_int16x8_allow_float() or self.post_training_dynamic_range_int8() or self.post_training_fp16()) + def activations_type(self): + return constants.INT16 if self._is_int16x8_target_required() else constants.INT8 + # Below are helpers for the above functions. def _validate_int8_required(self): @@ -244,6 +263,18 @@ class QuantizationMode(object): self._target_spec.supported_ops) or set(self._target_spec.supported_types) == set([constants.INT8])) + def _is_int16x8_target_required(self): + return bool( + set(self._target_spec.supported_ops).intersection([ + OpsSet.TFLITE_BUILTINS_ACTIVATIONS_INT16_WEIGHTS_INT8 + ])) + + def _is_allow_float(self): + return bool( + set(self._target_spec.supported_ops).intersection([ + OpsSet.TFLITE_BUILTINS + ])) + def _any_optimization_enabled(self): return bool( set(self._optimizations).intersection([ @@ -309,13 +340,13 @@ class TFLiteConverterBase(object): return _get_grappler_config(optimizers) def _calibrate_quantize_model(self, result, inference_input_type, - inference_output_type, allow_float): + inference_output_type, activations_type, allow_float): if not isinstance(self.representative_dataset, RepresentativeDataset): self.representative_dataset = RepresentativeDataset( self.representative_dataset) calibrate_quantize = _calibrator.Calibrator(result) - activations_type = constants.INT16 if self._is_int16x8_target_required() else constants.INT8 + if (self.experimental_calibrate_only: return calibrate_quantize.calibrate(self.representative_dataset.input_gen) else: @@ -608,12 +639,20 @@ class TFLiteConverterV2(TFLiteConverterBase): output_tensors=output_tensors, **converter_kwargs) + activations_type = quant_mode.activations_type() + if quant_mode.post_training_int8_no_float(): result = self._calibrate_quantize_model(result, constants.FLOAT, - constants.FLOAT, False) + constants.FLOAT, activations_type, False) elif quant_mode.post_training_int8_allow_float(): result = self._calibrate_quantize_model(result, constants.FLOAT, - constants.FLOAT, True) + constants.FLOAT, activations_type, True) + elif quant_mode.post_training_int16x8_no_float(): + result = self._calibrate_quantize_model(result, constants.FLOAT, + constants.FLOAT, activations_type, False) + elif quant_mode.post_training_int16x8_allow_float(): + result = self._calibrate_quantize_model(result, constants.FLOAT, + constants.FLOAT, activations_type, True) return result @@ -1114,6 +1153,8 @@ class TFLiteConverter(TFLiteConverterBase): quant_mode.post_training_int8_no_float() or quant_mode.post_training_int8_allow_float() or quant_mode.post_training_dynamic_range_int8() or + quant_mode.post_training_int16x8_no_float() or + quant_mode.post_training_int16x8_allow_float() or quant_mode.post_training_fp16()) if post_training_optimize: # Post training optimizations require that TOCO outputs a float model. @@ -1223,12 +1264,20 @@ class TFLiteConverter(TFLiteConverterBase): output_arrays=self._output_arrays, **converter_kwargs) + activations_type = quant_mode.activations_type() + if quant_mode.post_training_int8_no_float(): result = self._calibrate_quantize_model(result, inference_input_type, - inference_output_type, False) + inference_output_type, activations_type, False) elif quant_mode.post_training_int8_allow_float(): result = self._calibrate_quantize_model(result, inference_input_type, - inference_output_type, True) + inference_output_type, activations_type, True) + elif quant_mode.post_training_int16x8_no_float(): + result = self._calibrate_quantize_model(result, inference_input_type, + inference_output_type, activations_type, False) + elif quant_mode.post_training_int16x8_allow_float(): + result = self._calibrate_quantize_model(result, inference_input_type, + inference_output_type, activations_type, True) return result @@ -1334,7 +1383,6 @@ class TocoConverter(object): @classmethod @_deprecation.deprecated( - None, "Use `lite.TFLiteConverter.from_keras_model_file` instead.") def from_keras_model_file(cls, model_file, input_arrays=None, diff --git a/tensorflow/lite/python/optimize/calibration_wrapper_pybind11.cc b/tensorflow/lite/python/optimize/calibration_wrapper_pybind11.cc index f56b23090b9..9a8fea5d1f6 100644 --- a/tensorflow/lite/python/optimize/calibration_wrapper_pybind11.cc +++ b/tensorflow/lite/python/optimize/calibration_wrapper_pybind11.cc @@ -40,17 +40,17 @@ PYBIND11_MODULE(_pywrap_tensorflow_lite_calibration_wrapper, m) { }) .def("QuantizeModel", [](CalibrationWrapper& self, int input_py_type, int output_py_type, - bool allow_float, bool enable_mlir_quantizer) { + bool allow_float, int activations_py_type, bool enable_mlir_quantizer) { return tensorflow::pyo_or_throw( self.QuantizeModel(input_py_type, output_py_type, allow_float, - enable_mlir_quantizer)); + activations_py_type, enable_mlir_quantizer)); }) .def("QuantizeModel", [](CalibrationWrapper& self, int input_py_type, int output_py_type, - bool allow_float) { + bool allow_float, int activations_py_type) { return tensorflow::pyo_or_throw( self.QuantizeModel(input_py_type, output_py_type, allow_float, - /*enable_mlir_quantizer=*/false)); + activations_py_type, /*enable_mlir_quantizer=*/false)); }) .def("QuantizeModel", [](CalibrationWrapper& self, int input_py_type, int output_py_type, bool allow_float, From 6da9fa6111ab1be251b1138a0a591bc022b58cdf Mon Sep 17 00:00:00 2001 From: Peng Meng Date: Thu, 26 Mar 2020 00:05:02 +0800 Subject: [PATCH 018/867] fixOpFuse --- .../core/grappler/optimizers/remapper.cc | 31 ++++++++++++------- 1 file changed, 20 insertions(+), 11 deletions(-) diff --git a/tensorflow/core/grappler/optimizers/remapper.cc b/tensorflow/core/grappler/optimizers/remapper.cc index 5b41ad38089..b47c71fd5c8 100644 --- a/tensorflow/core/grappler/optimizers/remapper.cc +++ b/tensorflow/core/grappler/optimizers/remapper.cc @@ -1476,6 +1476,13 @@ bool RequiresInferredShapes(const RemapperContext& ctx, int node_index) { return true; }; + const auto is_conv2d_candidate = [&]() -> bool { + if (!IsConv2D(*node_def)) return false; + if (GetDataTypeFromAttr(*node_def, "T") != DT_FLOAT) return false; + + return true; + }; + // Candidate for a FusedBatchNorm fusion. const auto is_batch_norm_fusion_candidate = [&]() -> bool { if (!IsRelu(*node_def)) return false; @@ -1506,7 +1513,7 @@ bool RequiresInferredShapes(const RemapperContext& ctx, int node_index) { return false; }; - return is_batch_norm_candidate() || is_batch_norm_fusion_candidate(); + return is_conv2d_candidate() || is_batch_norm_candidate() || is_batch_norm_fusion_candidate(); } } // namespace @@ -1564,6 +1571,17 @@ Status Remapper::Optimize(Cluster* cluster, const GrapplerItem& item, } #endif //! INTEL_MKL + // Infer properties lazily in case they are not needed. + if (!ctx.inferred_graph_properties && RequiresInferredShapes(ctx, i)) { + const bool assume_valid_feeds = opt_level_ == RewriterConfig::AGGRESSIVE; + TF_RETURN_IF_ERROR(ctx.graph_properties.InferStatically( + assume_valid_feeds, + /*aggressive_shape_inference=*/false, + /*include_input_tensor_values=*/true, + /*include_output_tensor_values=*/false)); + ctx.inferred_graph_properties = true; + } + // Remap {Conv2D,MatMul}+BiasAdd into the _Fused{Conv2D,MatMul} ContractionWithBiasAdd contract_with_bias; if (allow_non_differentiable_rewrites && @@ -1592,6 +1610,7 @@ Status Remapper::Optimize(Cluster* cluster, const GrapplerItem& item, // Remove this once TF-MKL supports _FusedConv2D with these operations. #ifndef INTEL_MKL // Remap Conv2D+Squeeze+BiasAdd into the _FusedConv2D+Squeeze. + ContractionWithSqueezeAndBiasAdd contract_with_squeeze_and_bias; if (allow_non_differentiable_rewrites && FindConv2DWithSqueezeAndBias(ctx, i, &contract_with_squeeze_and_bias)) { @@ -1624,16 +1643,6 @@ Status Remapper::Optimize(Cluster* cluster, const GrapplerItem& item, } #endif // !INTEL_MKL - // Infer properties lazily in case they are not needed. - if (!ctx.inferred_graph_properties && RequiresInferredShapes(ctx, i)) { - const bool assume_valid_feeds = opt_level_ == RewriterConfig::AGGRESSIVE; - TF_RETURN_IF_ERROR(ctx.graph_properties.InferStatically( - assume_valid_feeds, - /*aggressive_shape_inference=*/false, - /*include_input_tensor_values=*/true, - /*include_output_tensor_values=*/false)); - ctx.inferred_graph_properties = true; - } // Remap FusedBatchNorm++ into the _FusedBatchNormEx. FusedBatchNormEx fused_batch_norm_ex; From 5673a09ff777ee08660ce9f71d90958e18bd995c Mon Sep 17 00:00:00 2001 From: Peng Meng Date: Thu, 26 Mar 2020 00:30:54 +0800 Subject: [PATCH 019/867] fix code format --- tensorflow/core/grappler/optimizers/remapper.cc | 3 ++- 1 file changed, 2 insertions(+), 1 deletion(-) diff --git a/tensorflow/core/grappler/optimizers/remapper.cc b/tensorflow/core/grappler/optimizers/remapper.cc index b47c71fd5c8..fdfb8e379b2 100644 --- a/tensorflow/core/grappler/optimizers/remapper.cc +++ b/tensorflow/core/grappler/optimizers/remapper.cc @@ -1513,7 +1513,8 @@ bool RequiresInferredShapes(const RemapperContext& ctx, int node_index) { return false; }; - return is_conv2d_candidate() || is_batch_norm_candidate() || is_batch_norm_fusion_candidate(); + return is_conv2d_candidate() || is_batch_norm_candidate() || + is_batch_norm_fusion_candidate(); } } // namespace From 69ee4de053a14bdf883a0e6726bb2b374b71c973 Mon Sep 17 00:00:00 2001 From: Elena Zhelezina Date: Fri, 27 Mar 2020 15:59:37 +0000 Subject: [PATCH 020/867] Fix for the broken 16-bit interface after latest changes to master. --- .../python/optimize/calibration_wrapper.cc | 14 ++--- .../lite/tools/optimize/quantize_model.cc | 12 ++-- .../lite/tools/optimize/quantize_model.h | 12 ++-- .../tools/optimize/quantize_model_test.cc | 55 ++++++++++--------- 4 files changed, 49 insertions(+), 44 deletions(-) diff --git a/tensorflow/lite/python/optimize/calibration_wrapper.cc b/tensorflow/lite/python/optimize/calibration_wrapper.cc index 660ea8d2d1b..ad82581bcba 100644 --- a/tensorflow/lite/python/optimize/calibration_wrapper.cc +++ b/tensorflow/lite/python/optimize/calibration_wrapper.cc @@ -233,12 +233,11 @@ PyObject* CalibrationWrapper::QuantizeModel(int input_py_type, reader_->AddCalibrationToModel(tflite_model.get(), /*update=*/false); flatbuffers::FlatBufferBuilder builder; auto status = kTfLiteOk; - - status = tflite::optimize::QuantizeModel( - &builder, tflite_model.get(), TfLiteTypeToSchemaType(input_type), - TfLiteTypeToSchemaType(output_type), allow_float, - TfLiteTypeToSchemaType(activations_type), error_reporter_.get()); - } + + status = tflite::optimize::QuantizeModelAllOperators( + &builder, tflite_model.get(), TfLiteTypeToSchemaType(input_type), + TfLiteTypeToSchemaType(output_type), allow_float, + TfLiteTypeToSchemaType(activations_type), error_reporter_.get()); if (status != kTfLiteOk) { error_reporter_->exception(); @@ -269,8 +268,7 @@ PyObject* CalibrationWrapper::QuantizeModel(int input_py_type, auto status = tflite::optimize::QuantizeModel( &builder, tflite_model.get(), TfLiteTypeToSchemaType(input_type), TfLiteTypeToSchemaType(output_type), allow_float, {op_name}, - TensorType_INT8, - error_reporter_.get()); + TensorType_INT8, error_reporter_.get()); if (status != kTfLiteOk) { error_reporter_->exception(); return nullptr; diff --git a/tensorflow/lite/tools/optimize/quantize_model.cc b/tensorflow/lite/tools/optimize/quantize_model.cc index 32cd2b8c25a..0892e7ae52a 100644 --- a/tensorflow/lite/tools/optimize/quantize_model.cc +++ b/tensorflow/lite/tools/optimize/quantize_model.cc @@ -1240,11 +1240,13 @@ TfLiteStatus QuantizeModel(flatbuffers::FlatBufferBuilder* builder, return kTfLiteOk; } -TfLiteStatus QuantizeModel(flatbuffers::FlatBufferBuilder* builder, - ModelT* model, const TensorType& input_type, - const TensorType& output_type, bool allow_float, - const TensorType& activations_type, - ErrorReporter* error_reporter) { +TfLiteStatus QuantizeModelAllOperators(flatbuffers::FlatBufferBuilder* builder, + ModelT* model, + const TensorType& input_type, + const TensorType& output_type, + bool allow_float, + const TensorType& activations_type, + ErrorReporter* error_reporter) { return QuantizeModel(builder, model, input_type, output_type, allow_float, GetAllOperatorOutputs(model), activations_type, error_reporter); diff --git a/tensorflow/lite/tools/optimize/quantize_model.h b/tensorflow/lite/tools/optimize/quantize_model.h index 06c30b88fd0..29f581d2b35 100644 --- a/tensorflow/lite/tools/optimize/quantize_model.h +++ b/tensorflow/lite/tools/optimize/quantize_model.h @@ -69,11 +69,13 @@ TfLiteStatus QuantizeModel(flatbuffers::FlatBufferBuilder* builder, // could be TensorType_INT16 or TensorType_INT8. // // Note: This is a private API, subject to change. -TfLiteStatus QuantizeModel(flatbuffers::FlatBufferBuilder* builder, - ModelT* model, const TensorType& input_type, - const TensorType& output_type, bool allow_float, - const TensorType& activations_type, - ErrorReporter* error_reporter); +TfLiteStatus QuantizeModelAllOperators(flatbuffers::FlatBufferBuilder* builder, + ModelT* model, + const TensorType& input_type, + const TensorType& output_type, + bool allow_float, + const TensorType& activations_type, + ErrorReporter* error_reporter); // Quantizes input_model and populates the provided builder with the new model // with all possible input parameters. diff --git a/tensorflow/lite/tools/optimize/quantize_model_test.cc b/tensorflow/lite/tools/optimize/quantize_model_test.cc index 0f780e4d3da..885fa98992c 100644 --- a/tensorflow/lite/tools/optimize/quantize_model_test.cc +++ b/tensorflow/lite/tools/optimize/quantize_model_test.cc @@ -190,9 +190,9 @@ TEST_P(QuantizeConvModelTest, OperatorsAreUnchanged) { } TEST_P(QuantizeConvModelTest, GraphIsFullyQuantized) { - auto status = - QuantizeModel(&builder_, &model_, tensor_type_, tensor_type_, - /*allow_float*/ false, tensor_type_, &error_reporter_); + auto status = QuantizeModelAllOperators( + &builder_, &model_, tensor_type_, tensor_type_, + /*allow_float*/ false, tensor_type_, &error_reporter_); EXPECT_EQ(status, kTfLiteOk); for (const auto& subgraph : model_.subgraphs) { for (const auto& tensor : subgraph->tensors) { @@ -209,9 +209,9 @@ TEST_P(QuantizeConvModelTest, GraphIsFullyQuantized) { } TEST_P(QuantizeConvModelTest, FloatInputAndOutput) { - auto status = - QuantizeModel(&builder_, &model_, TensorType_FLOAT32, TensorType_FLOAT32, - /*allow_float*/ false, tensor_type_, &error_reporter_); + auto status = QuantizeModelAllOperators( + &builder_, &model_, TensorType_FLOAT32, TensorType_FLOAT32, + /*allow_float*/ false, tensor_type_, &error_reporter_); EXPECT_EQ(status, kTfLiteOk); for (int32_t subgraph_idx = 0; subgraph_idx < model_.subgraphs.size(); @@ -384,8 +384,9 @@ class QuantizeConcatModelTest : public QuantizeModelTest, // concat - output // input1 / TEST_P(QuantizeConcatModelTest, AddRequantBeforeConcat) { - auto status = QuantizeModel(&builder_, &model_, tensor_type_, tensor_type_, - false, tensor_type_, &error_reporter_); + auto status = + QuantizeModelAllOperators(&builder_, &model_, tensor_type_, tensor_type_, + false, tensor_type_, &error_reporter_); EXPECT_EQ(status, kTfLiteOk); // There is only one subgraph. @@ -549,9 +550,9 @@ class QuantizeConvModel1Test : public QuantizeModelTest { }; TEST_F(QuantizeConvModel1Test, VerifyConvQuantizationWithUnitScale) { - auto status = - QuantizeModel(&builder_, &model_, TensorType_INT8, TensorType_INT8, false, - TensorType_INT8, &error_reporter_); + auto status = QuantizeModelAllOperators(&builder_, &model_, TensorType_INT8, + TensorType_INT8, false, + TensorType_INT8, &error_reporter_); EXPECT_EQ(status, kTfLiteOk); const auto& subgraph = model_.subgraphs[0]; @@ -658,8 +659,9 @@ INSTANTIATE_TEST_SUITE_P(QuantizeConvModel2TestInst, QuantizeConvModel2Test, TensorType_INT16})); TEST_P(QuantizeConvModel2Test, VerifyConvQuantization) { - auto status = QuantizeModel(&builder_, &model_, tensor_type_, tensor_type_, - false, tensor_type_, &error_reporter_); + auto status = + QuantizeModelAllOperators(&builder_, &model_, tensor_type_, tensor_type_, + false, tensor_type_, &error_reporter_); ASSERT_EQ(kTfLiteOk, status); const auto& subgraph = model_.subgraphs[0]; auto conv_op = subgraph->operators[0].get(); @@ -1308,8 +1310,9 @@ TEST_F(QuantizeFCTest, VerifyFC) { EXPECT_EQ(model_.operator_codes[1]->version, 1); } -class QuantizeCustomOpTest : public QuantizeModelTest, - public ::testing::WithParamInterface { +class QuantizeCustomOpTest + : public QuantizeModelTest, + public ::testing::WithParamInterface { protected: QuantizeCustomOpTest() { input_model_ = ReadModel(internal::kModelMixed); @@ -1319,9 +1322,9 @@ class QuantizeCustomOpTest : public QuantizeModelTest, }; TEST_P(QuantizeCustomOpTest, VerifyMixedQuantization) { - auto status = - QuantizeModel(&builder_, &model_, GetParam(), GetParam(), - /*allow_float=*/true, GetParam(), &error_reporter_); + auto status = QuantizeModelAllOperators( + &builder_, &model_, GetParam(), GetParam(), + /*allow_float=*/true, GetParam(), &error_reporter_); ASSERT_EQ(kTfLiteOk, status); const auto& subgraph = model_.subgraphs[0]; auto float_graph = readonly_model_->subgraphs()->Get(0); @@ -1335,7 +1338,7 @@ TEST_P(QuantizeCustomOpTest, VerifyMixedQuantization) { BuiltinOperator_CUSTOM, BuiltinOperator_CUSTOM, BuiltinOperator_QUANTIZE, BuiltinOperator_SQUEEZE}; const std::vector op_input_types = { - GetParam(), GetParam(), TensorType_FLOAT32, + GetParam(), GetParam(), TensorType_FLOAT32, TensorType_FLOAT32, TensorType_FLOAT32, GetParam()}; for (int i = 0; i < subgraph->operators.size(); ++i) { OperatorT* op = subgraph->operators[i].get(); @@ -1358,9 +1361,9 @@ class QuantizeOp16x8Test : public QuantizeModelTest { }; TEST_F(QuantizeOp16x8Test, VerifyMixedQuantization16x8) { - auto status = - QuantizeModel(&builder_, &model_, TensorType_INT16, TensorType_FLOAT32, - /*allow_float=*/true, TensorType_INT16, &error_reporter_); + auto status = QuantizeModelAllOperators( + &builder_, &model_, TensorType_INT16, TensorType_FLOAT32, + /*allow_float=*/true, TensorType_INT16, &error_reporter_); ASSERT_EQ(kTfLiteOk, status); const auto& subgraph = model_.subgraphs[0]; auto float_graph = readonly_model_->subgraphs()->Get(0); @@ -1369,11 +1372,11 @@ TEST_F(QuantizeOp16x8Test, VerifyMixedQuantization16x8) { // The resulting model should be: // conv_2d->dequantize->log_softmax ASSERT_EQ(subgraph->operators.size(), 3); - const std::vector op_codes = { - BuiltinOperator_CONV_2D, BuiltinOperator_DEQUANTIZE, - BuiltinOperator_LOG_SOFTMAX}; + const std::vector op_codes = {BuiltinOperator_CONV_2D, + BuiltinOperator_DEQUANTIZE, + BuiltinOperator_LOG_SOFTMAX}; const std::vector op_input_types = { - TensorType_INT16, TensorType_INT16, TensorType_FLOAT32}; + TensorType_INT16, TensorType_INT16, TensorType_FLOAT32}; for (int i = 0; i < subgraph->operators.size(); ++i) { OperatorT* op = subgraph->operators[i].get(); ASSERT_EQ(model_.operator_codes[op->opcode_index]->builtin_code, From 26e4ac1d76b43d9cf5288c914985e42fd0bfdbbf Mon Sep 17 00:00:00 2001 From: Mihai Maruseac Date: Mon, 6 Apr 2020 16:50:17 +0000 Subject: [PATCH 021/867] Update tensorflow/python/ops/array_ops.py --- tensorflow/python/ops/array_ops.py | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/tensorflow/python/ops/array_ops.py b/tensorflow/python/ops/array_ops.py index 9367374717e..fc29aafb7de 100644 --- a/tensorflow/python/ops/array_ops.py +++ b/tensorflow/python/ops/array_ops.py @@ -3507,7 +3507,7 @@ def edit_distance(hypothesis, truth, normalize=True, name="edit_distance"): For example: Given the following input, - * `hypothesis` is a `tf.SparseTensor` of shape `(2, 1, 1)` + * `hypothesis` is a `tf.SparseTensor` of shape `[2, 1, 1]` * `truth` is a `tf.SparseTensor` of shape `(2, 2, 2)` >>> hypothesis = tf.SparseTensor( From 05edf7e11bb3d91f0e2b75e67914c2379cf0cd89 Mon Sep 17 00:00:00 2001 From: Mihai Maruseac Date: Mon, 6 Apr 2020 16:50:26 +0000 Subject: [PATCH 022/867] Update tensorflow/python/ops/array_ops.py --- tensorflow/python/ops/array_ops.py | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/tensorflow/python/ops/array_ops.py b/tensorflow/python/ops/array_ops.py index fc29aafb7de..004bbeac7f6 100644 --- a/tensorflow/python/ops/array_ops.py +++ b/tensorflow/python/ops/array_ops.py @@ -3527,7 +3527,7 @@ def edit_distance(hypothesis, truth, normalize=True, name="edit_distance"): array([[inf, 1. ], [0.5, 1. ]], dtype=float32)> - The operaton returns a dense Tensor of shape `(2, 2)` with + The operaton returns a dense Tensor of shape `[2, 2]` with edit distances normalized by `truth` lengths. **Note**: It is possible to calculate edit distance between two From 28769b2b746570883ae8a57e042f50c4defba2b5 Mon Sep 17 00:00:00 2001 From: Mihai Maruseac Date: Mon, 6 Apr 2020 16:50:35 +0000 Subject: [PATCH 023/867] Update tensorflow/python/ops/array_ops.py --- tensorflow/python/ops/array_ops.py | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/tensorflow/python/ops/array_ops.py b/tensorflow/python/ops/array_ops.py index 004bbeac7f6..e508165a820 100644 --- a/tensorflow/python/ops/array_ops.py +++ b/tensorflow/python/ops/array_ops.py @@ -3508,7 +3508,7 @@ def edit_distance(hypothesis, truth, normalize=True, name="edit_distance"): Given the following input, * `hypothesis` is a `tf.SparseTensor` of shape `[2, 1, 1]` - * `truth` is a `tf.SparseTensor` of shape `(2, 2, 2)` + * `truth` is a `tf.SparseTensor` of shape `[2, 2, 2]` >>> hypothesis = tf.SparseTensor( ... [[0, 0, 0], From 2759cdca671dee4b7a2035710cf08725a97ce73c Mon Sep 17 00:00:00 2001 From: Elena Zhelezina Date: Tue, 7 Apr 2020 16:25:36 +0100 Subject: [PATCH 024/867] Fix after merging with master. --- tensorflow/lite/python/optimize/calibrator.py | 4 ++-- tensorflow/lite/python/optimize/calibrator_test.py | 4 +++- 2 files changed, 5 insertions(+), 3 deletions(-) diff --git a/tensorflow/lite/python/optimize/calibrator.py b/tensorflow/lite/python/optimize/calibrator.py index e31983b834e..90c43fcddfa 100644 --- a/tensorflow/lite/python/optimize/calibrator.py +++ b/tensorflow/lite/python/optimize/calibrator.py @@ -60,8 +60,8 @@ class Calibrator(object): input_type, output_type, allow_float, - resize_input=True, - activations_type = lite_constants.INT8): + activations_type=lite_constants.INT8, + resize_input=True): """Calibrates the model with specified generator and then quantizes it. The input shapes of the calibrator are resized with the calibration data if diff --git a/tensorflow/lite/python/optimize/calibrator_test.py b/tensorflow/lite/python/optimize/calibrator_test.py index f8a1171a629..f778c8a555d 100644 --- a/tensorflow/lite/python/optimize/calibrator_test.py +++ b/tensorflow/lite/python/optimize/calibrator_test.py @@ -148,7 +148,9 @@ class CalibratorTest(test_util.TensorFlowTestCase, parameterized.TestCase): with self.assertRaisesRegex(ValueError, 'Size mismatch'): quantizer.calibrate_and_quantize(input_gen, constants.FLOAT, - constants.FLOAT, False, False) + constants.FLOAT, False, + constants.INT8, + False) def test_invalid_type_calibrator_gen(self): model_path = resource_loader.get_path_to_datafile( From cc08b5dff7346b2e2cd9fb80409c8f09bfebc089 Mon Sep 17 00:00:00 2001 From: Jens Elofsson Date: Tue, 7 Apr 2020 11:20:02 +0200 Subject: [PATCH 025/867] Update network_tester. New features: Multiple inputs Multiple outputs Output in json format Can call invoke() more than once Updated README --- .../examples/network_tester/Makefile.inc | 4 + .../micro/examples/network_tester/README.md | 34 ++++++- .../network_tester/expected_output_data.h | 2 +- .../examples/network_tester/input_data.h | 4 +- .../examples/network_tester/network_model.h | 5 +- .../network_tester/network_tester_test.cc | 88 +++++++++++++------ 6 files changed, 104 insertions(+), 33 deletions(-) diff --git a/tensorflow/lite/micro/examples/network_tester/Makefile.inc b/tensorflow/lite/micro/examples/network_tester/Makefile.inc index 27f54a66763..a5c911238c8 100644 --- a/tensorflow/lite/micro/examples/network_tester/Makefile.inc +++ b/tensorflow/lite/micro/examples/network_tester/Makefile.inc @@ -33,6 +33,10 @@ ifeq ($(COMPARE_OUTPUT_DATA),no) CXXFLAGS += -DNO_COMPARE_OUTPUT_DATA endif +ifdef NUM_INFERENCES + CXXFLAGS += -DNUM_INFERENCES=$(NUM_INFERENCES) +endif + # Builds a standalone object recognition binary. $(eval $(call microlite_test,network_tester_test,\ $(NETWORK_TESTER_TEST_SRCS),$(NETWORK_TESTER_TEST_HDRS))) diff --git a/tensorflow/lite/micro/examples/network_tester/README.md b/tensorflow/lite/micro/examples/network_tester/README.md index 7c4c48e4eb1..0cb709dce0a 100644 --- a/tensorflow/lite/micro/examples/network_tester/README.md +++ b/tensorflow/lite/micro/examples/network_tester/README.md @@ -34,8 +34,40 @@ make -f tensorflow/lite/micro/tools/make/Makefile network_tester_test \ `ARENA_SIZE`: The size of the memory to be allocated (in bytes) by the interpreter. \ `NUM_BYTES_TO_PRINT`: The number of bytes of the output data to print. \ -Defaults to 0 if not specified. \ +If set to 0, all bytes of the output are printed. \ `COMPARE_OUTPUT_DATA`: If set to "no" the output data is not compared to the expected output data. This could be useful e.g. if the execution time needs to be minimized, or there is no expected output data. If omitted, the output data is compared to the expected output. +`NUM_INFERENCES`: Define how many inferences that are made. Defaults to 1. \ + +The output is printed in JSON format using printf: +``` +num_of_outputs: 1 +output_begin +[ +{ +"dims": [4,1,2,2,1], +"data_address": "0x000000", +"data":"0x06,0x08,0x0e,0x10" +}] +output_end +``` + +If there are multiple output tensors, the output will look like this: +``` +num_of_outputs: 2 +output_begin +[ +{ +"dims": [4,1,2,2,1], +"data_address": "0x000000", +"data":"0x06,0x08,0x0e,0x10" +}, +{ +"dims": [4,1,2,2,1], +"data_address": "0x111111", +"data":"0x06,0x08,0x0e,0x10" +}] +output_end +``` diff --git a/tensorflow/lite/micro/examples/network_tester/expected_output_data.h b/tensorflow/lite/micro/examples/network_tester/expected_output_data.h index 03e21954b7f..934722bad94 100644 --- a/tensorflow/lite/micro/examples/network_tester/expected_output_data.h +++ b/tensorflow/lite/micro/examples/network_tester/expected_output_data.h @@ -17,6 +17,6 @@ limitations under the License. #define TENSORFLOW_LITE_MICRO_EXAMPLES_NETWORK_TESTER_EXPECTED_OUTPUT_DATA_H_ static unsigned int expected_output_data_len = 4; -static unsigned char expected_output_data[] = {6, 8, 14, 16}; +static unsigned char expected_output_data[1][4] = {6, 8, 14, 16}; #endif // TENSORFLOW_LITE_MICRO_EXAMPLES_NETWORK_TESTER_EXPECTED_OUTPUT_DATA_H_ diff --git a/tensorflow/lite/micro/examples/network_tester/input_data.h b/tensorflow/lite/micro/examples/network_tester/input_data.h index b47277cca93..b3710313dd2 100644 --- a/tensorflow/lite/micro/examples/network_tester/input_data.h +++ b/tensorflow/lite/micro/examples/network_tester/input_data.h @@ -17,7 +17,7 @@ limitations under the License. #define TENSORFLOW_LITE_MICRO_EXAMPLES_NETWORK_TESTER_INPUT_DATA_H_ static const int input_data_len = 16; -static const unsigned char input_data[] = {1, 2, 3, 4, 5, 6, 7, 8, - 9, 10, 11, 12, 13, 14, 15, 16}; +static const unsigned char input_data[1][16] = {{1, 2, 3, 4, 5, 6, 7, 8, + 9, 10, 11, 12, 13, 14, 15, 16}}; #endif // TENSORFLOW_LITE_MICRO_EXAMPLES_NETWORK_TESTER_INPUT_DATA_H_ diff --git a/tensorflow/lite/micro/examples/network_tester/network_model.h b/tensorflow/lite/micro/examples/network_tester/network_model.h index 4c275dbfbba..0431d7deee7 100644 --- a/tensorflow/lite/micro/examples/network_tester/network_model.h +++ b/tensorflow/lite/micro/examples/network_tester/network_model.h @@ -1,8 +1,11 @@ /* Copyright 2019 The TensorFlow Authors. All Rights Reserved. + Licensed under the Apache License, Version 2.0 (the "License"); you may not use this file except in compliance with the License. You may obtain a copy of the License at + http://www.apache.org/licenses/LICENSE-2.0 + Unless required by applicable law or agreed to in writing, software distributed under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. @@ -64,4 +67,4 @@ const unsigned char network_model[] = { 0x08, 0x00, 0x07, 0x00, 0x06, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x11}; const unsigned int network_model_len = 576; -#endif // TENSORFLOW_LITE_MICRO_EXAMPLES_NETWORK_TESTER_NETWORK_MODEL_H_ +#endif diff --git a/tensorflow/lite/micro/examples/network_tester/network_tester_test.cc b/tensorflow/lite/micro/examples/network_tester/network_tester_test.cc index 0650222b970..5a307fb5c2a 100644 --- a/tensorflow/lite/micro/examples/network_tester/network_tester_test.cc +++ b/tensorflow/lite/micro/examples/network_tester/network_tester_test.cc @@ -1,8 +1,11 @@ /* Copyright 2019 The TensorFlow Authors. All Rights Reserved. + Licensed under the Apache License, Version 2.0 (the "License"); you may not use this file except in compliance with the License. You may obtain a copy of the License at + http://www.apache.org/licenses/LICENSE-2.0 + Unless required by applicable law or agreed to in writing, software distributed under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. @@ -10,44 +13,54 @@ See the License for the specific language governing permissions and limitations under the License. ==============================================================================*/ -#include "tensorflow/lite/micro/examples/network_tester/expected_output_data.h" -#include "tensorflow/lite/micro/examples/network_tester/input_data.h" -#include "tensorflow/lite/micro/examples/network_tester/network_model.h" #include "tensorflow/lite/micro/kernels/all_ops_resolver.h" #include "tensorflow/lite/micro/micro_error_reporter.h" #include "tensorflow/lite/micro/micro_interpreter.h" -#include "tensorflow/lite/micro/testing/micro_test.h" -#include "tensorflow/lite/micro/testing/test_utils.h" #include "tensorflow/lite/schema/schema_generated.h" #include "tensorflow/lite/version.h" +#include "tensorflow/lite/micro/examples/network_tester/expected_output_data.h" +#include "tensorflow/lite/micro/examples/network_tester/input_data.h" +#include "tensorflow/lite/micro/examples/network_tester/network_model.h" + +#include "tensorflow/lite/micro/testing/micro_test.h" +#include "tensorflow/lite/micro/testing/test_utils.h" + #ifndef TENSOR_ARENA_SIZE #define TENSOR_ARENA_SIZE (1024) #endif +#ifndef NUM_INFERENCES +#define NUM_INFERENCES 1 +#endif + uint8_t tensor_arena[TENSOR_ARENA_SIZE]; #ifdef NUM_BYTES_TO_PRINT inline void print_output_data(TfLiteTensor* output) { int num_bytes_to_print = - (output->bytes < NUM_BYTES_TO_PRINT) ? output->bytes : NUM_BYTES_TO_PRINT; + ((output->bytes < NUM_BYTES_TO_PRINT) || NUM_BYTES_TO_PRINT == 0) + ? output->bytes + : NUM_BYTES_TO_PRINT; int dims_size = output->dims->size; - printf("dims: {%d,", dims_size); + printf("{\n"); + printf("\"dims\": [%d,", dims_size); for (int i = 0; i < output->dims->size - 1; ++i) { printf("%d,", output->dims->data[i]); } - printf("%d}\n", output->dims->data[dims_size - 1]); + printf("%d],\n", output->dims->data[dims_size - 1]); - printf("data_address: %p\n", output->data.raw); - printf("data:\n{"); + printf("\"data_address\": \"%p\",\n", output->data.raw); + printf("\"data\":\""); for (int i = 0; i < num_bytes_to_print - 1; ++i) { - if (i % 16 == 0) { + if (i % 16 == 0 && i != 0) { printf("\n"); } printf("0x%02x,", output->data.uint8[i]); } - printf("0x%02x\n}\n", output->data.uint8[num_bytes_to_print - 1]); + printf("0x%02x\"\n", output->data.uint8[num_bytes_to_print - 1]); + printf("}"); } #endif @@ -63,7 +76,7 @@ TF_LITE_MICRO_TEST(TestInvoke) { "Model provided is schema version %d not equal " "to supported version %d.\n", model->version(), TFLITE_SCHEMA_VERSION); - return 1; + return kTfLiteError; } tflite::ops::micro::AllOpsResolver resolver; @@ -74,29 +87,48 @@ TF_LITE_MICRO_TEST(TestInvoke) { TfLiteStatus allocate_status = interpreter.AllocateTensors(); if (allocate_status != kTfLiteOk) { TF_LITE_REPORT_ERROR(error_reporter, "Tensor allocation failed\n"); + return kTfLiteError; } - TF_LITE_MICRO_EXPECT_EQ(kTfLiteOk, allocate_status); - TfLiteTensor* input = interpreter.input(0); - memcpy(input->data.uint8, input_data, input->bytes); - - TfLiteStatus invoke_status = interpreter.Invoke(); - if (invoke_status != kTfLiteOk) { - TF_LITE_REPORT_ERROR(error_reporter, "Invoke failed\n"); - } - TF_LITE_MICRO_EXPECT_EQ(kTfLiteOk, invoke_status); - - TfLiteTensor* output = interpreter.output(0); + for (int n = 0; n < NUM_INFERENCES; n++) { + for (int i = 0; i < interpreter.inputs_size(); ++i) { + TfLiteTensor* input = interpreter.input(i); + memcpy(input->data.uint8, input_data[i], input->bytes); + } + TfLiteStatus invoke_status = interpreter.Invoke(); + if (invoke_status != kTfLiteOk) { + TF_LITE_REPORT_ERROR(error_reporter, "Invoke failed\n"); + return kTfLiteError; + } + TF_LITE_MICRO_EXPECT_EQ(kTfLiteOk, invoke_status); #ifdef NUM_BYTES_TO_PRINT - print_output_data(output); + // Print all of the output data, or the first NUM_BYTES_TO_PRINT bytes, + // whichever comes first as well as the output shape. + printf("num_of_outputs: %d\n", interpreter.outputs_size()); + printf("output_begin\n"); + printf("[\n"); + for (int i = 0; i < interpreter.outputs_size(); i++) { + TfLiteTensor* output = interpreter.output(i); + print_output_data(output); + if (i != interpreter.outputs_size() - 1) { + printf(",\n"); + } + } + printf("]\n"); + printf("output_end\n"); #endif #ifndef NO_COMPARE_OUTPUT_DATA - for (int i = 0; i < output->bytes; ++i) { - TF_LITE_MICRO_EXPECT_EQ(output->data.uint8[i], expected_output_data[i]); - } + for (int i = 0; i < interpreter.outputs_size(); i++) { + TfLiteTensor* output = interpreter.output(i); + for (int j = 0; j < output->bytes; ++j) { + TF_LITE_MICRO_EXPECT_EQ(output->data.uint8[j], + expected_output_data[i][j]); + } + } #endif + } TF_LITE_REPORT_ERROR(error_reporter, "Ran successfully\n"); } From 43036088fc3cfda3bd8d9c3a5df114d7d393618a Mon Sep 17 00:00:00 2001 From: Peng Meng Date: Mon, 13 Apr 2020 16:36:59 +0800 Subject: [PATCH 026/867] add unit test --- .../core/grappler/optimizers/remapper.cc | 6 +- .../core/grappler/optimizers/remapper_test.cc | 70 +++++++++++++++++++ 2 files changed, 73 insertions(+), 3 deletions(-) diff --git a/tensorflow/core/grappler/optimizers/remapper.cc b/tensorflow/core/grappler/optimizers/remapper.cc index fdfb8e379b2..239002c9da3 100644 --- a/tensorflow/core/grappler/optimizers/remapper.cc +++ b/tensorflow/core/grappler/optimizers/remapper.cc @@ -1476,8 +1476,8 @@ bool RequiresInferredShapes(const RemapperContext& ctx, int node_index) { return true; }; - const auto is_conv2d_candidate = [&]() -> bool { - if (!IsConv2D(*node_def)) return false; + const auto is_relu_candidate = [&]() -> bool { + if (!IsRelu(*node_def)) return false; if (GetDataTypeFromAttr(*node_def, "T") != DT_FLOAT) return false; return true; @@ -1513,7 +1513,7 @@ bool RequiresInferredShapes(const RemapperContext& ctx, int node_index) { return false; }; - return is_conv2d_candidate() || is_batch_norm_candidate() || + return is_relu_candidate() || is_batch_norm_candidate() || is_batch_norm_fusion_candidate(); } diff --git a/tensorflow/core/grappler/optimizers/remapper_test.cc b/tensorflow/core/grappler/optimizers/remapper_test.cc index 35e09b28205..831aa51ee72 100644 --- a/tensorflow/core/grappler/optimizers/remapper_test.cc +++ b/tensorflow/core/grappler/optimizers/remapper_test.cc @@ -449,6 +449,76 @@ TEST_F(RemapperTest, FuseMatMulWithBias) { test::ExpectTensorNear(tensors[0], tensors_expected[0], 1e-6); } +TEST_F(RemapperTest, FuseConv2DWithBiasAndActivationOnGPU) { + using ::tensorflow::ops::Placeholder; + + tensorflow::Scope s = tensorflow::Scope::NewRootScope(); + + auto input_shape = Placeholder::Shape({8, 32, 32, 3}); + auto filter_shape = Placeholder::Shape({3, 3, 3, 128}); + auto bias_shape = Placeholder::Shape({128}); + + auto input = Placeholder(s.WithOpName("input"), DT_FLOAT, input_shape); + auto filter = Placeholder(s.WithOpName("filter"), DT_FLOAT, filter_shape); + auto bias = Placeholder(s.WithOpName("bias"), DT_FLOAT, bias_shape); + + std::vector strides = {1, 1, 1, 1}; + auto conv = + ops::Conv2D(s.WithOpName("conv"), input, filter, strides, "SAME"); + auto bias_add = ops::BiasAdd(s.WithOpName("bias_add"), conv, bias); + + ops::Identity fetch = [&]() -> ops::Identity { + auto activate = s.WithOpName("activation"); + auto fetch = s.WithOpName("fetch"); + return ops::Identity(fetch, ops::Relu(activate, bias_add)); + }(); + + auto input_t = GenerateRandomTensor({8, 32, 32, 3}); + auto filter_t = GenerateRandomTensor({3, 3, 3, 128}); + auto bias_t = GenerateRandomTensor({128}); + + GrapplerItem item; + item.fetch = {"fetch"}; + item.feed = {{"input", input_t}, {"filter", filter_t}, {"bias", bias_t}}; + TF_ASSERT_OK(s.ToGraphDef(&item.graph)); + + // Place all nodes on GPU. + for (int i = 0; i < item.graph.node_size(); ++i) { + item.graph.mutable_node(i)->set_device("/device:GPU:0"); + } + + Remapper optimizer(RewriterConfig::AGGRESSIVE); // trust placeholders shape + //Remapper optimizer(RewriterConfig::ON); + GraphDef output; + TF_ASSERT_OK(optimizer.Optimize(nullptr, item, &output)); + + int found = 0; + for (const NodeDef& node : output.node()) { + if (node.name() == "activation") { + EXPECT_EQ(node.op(), "_FusedConv2D"); + ASSERT_GE(node.input_size(), 3); + EXPECT_EQ(node.input(0), "input"); + EXPECT_EQ(node.input(1), "filter"); + + EXPECT_EQ(node.attr().at("num_args").i(), 1); + EXPECT_EQ(node.input(2), "bias"); + + const auto fused_ops = node.attr().at("fused_ops").list().s(); + ASSERT_EQ(fused_ops.size(), 2); + EXPECT_EQ(fused_ops[0], "BiasAdd"); + EXPECT_EQ(fused_ops[1], "Relu"); + found++; + } + } + EXPECT_EQ(found, 1); + + auto tensors_expected = EvaluateNodes(item.graph, item.fetch, item.feed); + ASSERT_EQ(tensors_expected.size(), 1); + auto tensors = EvaluateNodes(output, item.fetch, item.feed); + ASSERT_EQ(tensors.size(), 1); + test::ExpectTensorNear(tensors[0], tensors_expected[0], 1e-6); +} + TEST_F(RemapperTest, FuseConv2DWithBiasAndActivation) { using ::tensorflow::ops::Placeholder; From 8b791cbf71aa68a91c69080b4313f115cb60fbf5 Mon Sep 17 00:00:00 2001 From: Peng Meng Date: Mon, 13 Apr 2020 17:36:16 +0800 Subject: [PATCH 027/867] optimize shape infer condition --- .../core/grappler/optimizers/remapper.cc | 25 +++++++++++++++++-- 1 file changed, 23 insertions(+), 2 deletions(-) diff --git a/tensorflow/core/grappler/optimizers/remapper.cc b/tensorflow/core/grappler/optimizers/remapper.cc index 239002c9da3..427a3f14aca 100644 --- a/tensorflow/core/grappler/optimizers/remapper.cc +++ b/tensorflow/core/grappler/optimizers/remapper.cc @@ -1461,6 +1461,7 @@ Status AddBatchNormNodes(RemapperContext* ctx, const FusedBatchNorm& matched) { // shapes: // (1) Splitting FusedBatchNorm into primitives. // (2) Fusing side input and/or activation into FusedBatchNorm. +// (3) Fusing Conv2D biasadd and relu on GPU bool RequiresInferredShapes(const RemapperContext& ctx, int node_index) { // Candidate for a FusedBatchNorm splitting. const auto* node_view = ctx.graph_view.GetNode(node_index); @@ -1476,10 +1477,30 @@ bool RequiresInferredShapes(const RemapperContext& ctx, int node_index) { return true; }; - const auto is_relu_candidate = [&]() -> bool { + const auto is_relu_biasadd_conv2d_candidate = [&]() -> bool { if (!IsRelu(*node_def)) return false; if (GetDataTypeFromAttr(*node_def, "T") != DT_FLOAT) return false; + if (node_view->NumRegularFanins() < 1) return false; + const auto& relu_fanin_0 = node_view->GetRegularFanin(0); + const auto* relu_fanin_0_node_view = relu_fanin_0.node_view(); + const auto* relu_fanin_0_node_def = relu_fanin_0_node_view->node(); + + if (!IsBiasAdd(*relu_fanin_0_node_def)) return false; + if (GetDataTypeFromAttr(*relu_fanin_0_node_def, "T") != DT_FLOAT) + return false; + + if (relu_fanin_0_node_view->NumRegularFanins() < 1) return false; + + const auto& biasadd_fanin_0 = + relu_fanin_0_node_view->GetRegularFanin(0); + const auto* biasadd_fanin_0_node_def = + biasadd_fanin_0.node_view()->node(); + + if (!IsConv2D(*biasadd_fanin_0_node_def)) return false; + if (GetDataTypeFromAttr(*biasadd_fanin_0_node_def, "T") != DT_FLOAT) + return false; + return true; }; @@ -1513,7 +1534,7 @@ bool RequiresInferredShapes(const RemapperContext& ctx, int node_index) { return false; }; - return is_relu_candidate() || is_batch_norm_candidate() || + return is_relu_biasadd_conv2d_candidate() || is_batch_norm_candidate() || is_batch_norm_fusion_candidate(); } From d46acc6016666abfc27d28e35936ed70979c822e Mon Sep 17 00:00:00 2001 From: "902449@58880@bigcat_chen" Date: Wed, 29 Apr 2020 15:05:06 +0800 Subject: [PATCH 028/867] debug_message --- tensorflow/lite/micro/we_i/debug_log.cc | 33 +++++++++++++++++++++++++ 1 file changed, 33 insertions(+) create mode 100644 tensorflow/lite/micro/we_i/debug_log.cc diff --git a/tensorflow/lite/micro/we_i/debug_log.cc b/tensorflow/lite/micro/we_i/debug_log.cc new file mode 100644 index 00000000000..a115d476aff --- /dev/null +++ b/tensorflow/lite/micro/we_i/debug_log.cc @@ -0,0 +1,33 @@ +/* Copyright 2018 The TensorFlow Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. +==============================================================================*/ + +// Implementation for the DebugLog() function that prints to the UART on the +// SparkFun Edge microcontroller. The same should work for other targets using +// the Ambiq Apollo 3. + +#include "tensorflow/lite/micro/debug_log.h" +#include "xprintf.h" +#include "console_io.h" +#include + +extern "C" void DebugLog(const char* s) { + static bool is_initialized = false; + if (!is_initialized) { + xprintf_setup(); + is_initialized = true; + } + + xprintf("%s", s); +} From c8a2f59e2e2775808b82c877433ead9545fd7a84 Mon Sep 17 00:00:00 2001 From: Eugene Kuznetsov Date: Wed, 22 Jan 2020 12:53:00 -0800 Subject: [PATCH 029/867] Fixing and enabling //tensorflow/core/util:gpu_kernel_helper_test_gpu --- tensorflow/core/util/gpu_device_functions.h | 4 + .../core/util/gpu_kernel_helper_test.cu.cc | 127 +++++++++++------- 2 files changed, 84 insertions(+), 47 deletions(-) diff --git a/tensorflow/core/util/gpu_device_functions.h b/tensorflow/core/util/gpu_device_functions.h index 61d1e3c9453..083d42b7de4 100644 --- a/tensorflow/core/util/gpu_device_functions.h +++ b/tensorflow/core/util/gpu_device_functions.h @@ -53,6 +53,8 @@ using gpuEvent_t = cudaEvent_t; #define gpuEventCreate cudaEventCreate #define gpuEventCreateWithFlags cudaEventCreateWithFlags #define gpuEventDisableTiming cudaEventDisableTiming +#define gpuDeviceSynchronize cudaDeviceSynchronize +#define gpuFree cudaFree #elif TENSORFLOW_USE_ROCM using gpuFloatComplex = hipFloatComplex; using gpuDoubleComplex = hipDoubleComplex; @@ -68,6 +70,8 @@ using cudaError_t = int; #define gpuEventCreate hipEventCreate #define gpuEventCreateWithFlags hipEventCreateWithFlags #define gpuEventDisableTiming hipEventDisableTiming +#define gpuDeviceSynchronize hipDeviceSynchronize +#define gpuFree hipFree static std::string cudaGetErrorString(int err) { return std::to_string(err); } #endif diff --git a/tensorflow/core/util/gpu_kernel_helper_test.cu.cc b/tensorflow/core/util/gpu_kernel_helper_test.cu.cc index c089511e964..3135d25d1b8 100644 --- a/tensorflow/core/util/gpu_kernel_helper_test.cu.cc +++ b/tensorflow/core/util/gpu_kernel_helper_test.cu.cc @@ -13,9 +13,10 @@ See the License for the specific language governing permissions and limitations under the License. ==============================================================================*/ -#if GOOGLE_CUDA +#if GOOGLE_CUDA || TENSORFLOW_USE_ROCM #define EIGEN_USE_GPU +#include #include #include "tensorflow/core/lib/core/status_test_util.h" @@ -25,14 +26,14 @@ limitations under the License. #define CUDA_EXPECT_SUCCESS \ { \ - cudaDeviceSynchronize(); \ + gpuDeviceSynchronize(); \ cudaError_t err = cudaGetLastError(); \ EXPECT_EQ(cudaSuccess, err) << cudaGetErrorString(err); \ } #define CUDA_ASSERT_SUCCESS \ { \ - cudaDeviceSynchronize(); \ + gpuDeviceSynchronize(); \ cudaError_t err = cudaGetLastError(); \ ASSERT_EQ(cudaSuccess, err) << cudaGetErrorString(err); \ } @@ -94,8 +95,7 @@ __global__ void Count3D(Gpu3DLaunchConfig config, int bufsize, } } -__global__ void CudaShuffleGetSrcLaneTest( - unsigned* __restrict__ failure_count) { +__global__ void GpuShuffleGetSrcLaneTest(unsigned* __restrict__ failure_count) { unsigned lane_id = GpuLaneId(); for (int width = warpSize; width > 1; width /= 2) { auto check_result = [&](const char* op_name, int param, unsigned actual, @@ -103,31 +103,38 @@ __global__ void CudaShuffleGetSrcLaneTest( if (actual != expected) { printf("Cuda%sGetSrcLane(%d, %d) for lane %d returned %d, not %d\n", op_name, param, width, lane_id, actual, expected); - CudaAtomicAdd(failure_count, 1); + GpuAtomicAdd(failure_count, 1); } }; + for (int src_lane = -warpSize; src_lane <= warpSize; ++src_lane) { - unsigned actual_lane = detail::CudaShuffleGetSrcLane(src_lane, width); +#if TENSORFLOW_USE_ROCM + if (src_lane < 0 || src_lane >= width) continue; +#endif + unsigned actual_lane = detail::GpuShuffleGetSrcLane(src_lane, width); unsigned expect_lane = - CudaShuffleSync(kCudaWarpAll, lane_id, src_lane, width); + GpuShuffleSync(kCudaWarpAll, lane_id, src_lane, width); check_result("Shuffle", src_lane, actual_lane, expect_lane); } + for (unsigned delta = 0; delta <= warpSize; ++delta) { - unsigned actual_lane = detail::CudaShuffleUpGetSrcLane(delta, width); + unsigned actual_lane = detail::GpuShuffleUpGetSrcLane(delta, width); unsigned expect_lane = - CudaShuffleUpSync(kCudaWarpAll, lane_id, delta, width); + GpuShuffleUpSync(kCudaWarpAll, lane_id, delta, width); check_result("ShuffleUp", delta, actual_lane, expect_lane); } + for (unsigned delta = 0; delta <= warpSize; ++delta) { - unsigned actual_lane = detail::CudaShuffleDownGetSrcLane(delta, width); + unsigned actual_lane = detail::GpuShuffleDownGetSrcLane(delta, width); unsigned expect_lane = - CudaShuffleDownSync(kCudaWarpAll, lane_id, delta, width); + GpuShuffleDownSync(kCudaWarpAll, lane_id, delta, width); check_result("ShuffleDown", delta, actual_lane, expect_lane); } + for (int lane_lane = warpSize; lane_lane > 0; lane_lane /= 2) { - unsigned actual_lane = detail::CudaShuffleXorGetSrcLane(lane_lane, width); + unsigned actual_lane = detail::GpuShuffleXorGetSrcLane(lane_lane, width); unsigned expect_lane = - CudaShuffleXorSync(kCudaWarpAll, lane_id, lane_lane, width); + GpuShuffleXorSync(kCudaWarpAll, lane_id, lane_lane, width); check_result("ShuffleXor", lane_lane, actual_lane, expect_lane); } } @@ -137,19 +144,32 @@ __global__ void CudaShuffleGetSrcLaneTest( class GpuLaunchConfigTest : public ::testing::Test { protected: - const int bufsize = 1024; + static const int bufsize = 1024; int* outbuf = nullptr; + int* outbuf_host = nullptr; + int hostbuf[bufsize]; Eigen::GpuStreamDevice stream; Eigen::GpuDevice d = Eigen::GpuDevice(&stream); + void copyToHost() { +#if TENSORFLOW_USE_ROCM + hipMemcpy(hostbuf, outbuf, sizeof(int) * bufsize, hipMemcpyDeviceToHost); +#endif + } virtual void SetUp() { +#if GOOGLE_CUDA cudaError_t err = cudaMallocManaged(&outbuf, sizeof(int) * bufsize); + outbuf_host = outbuf; +#else + cudaError_t err = hipMalloc(&outbuf, sizeof(int) * bufsize); + outbuf_host = hostbuf; +#endif ASSERT_EQ(cudaSuccess, err) << cudaGetErrorString(err); } virtual void TearDown() { - cudaDeviceSynchronize(); - cudaFree(outbuf); + gpuDeviceSynchronize(); + gpuFree(outbuf); outbuf = nullptr; } }; @@ -158,28 +178,32 @@ TEST_F(GpuLaunchConfigTest, GetGpuLaunchConfig) { GpuLaunchConfig cfg; // test valid inputs -#define TEST_LAUNCH_PARAMETER(work_element_count) \ - cfg = GetGpuLaunchConfig(bufsize, d); \ - TF_CHECK_OK(GpuLaunchKernel(SetOutbufZero, cfg.block_count, \ - cfg.thread_per_block, 0, d.stream(), cfg, \ - outbuf)); \ - CUDA_ASSERT_SUCCESS \ - cfg = GetGpuLaunchConfig(work_element_count, d); \ - TF_CHECK_OK(GpuLaunchKernel(Count1D, cfg.block_count, cfg.thread_per_block, \ - 0, d.stream(), cfg, bufsize, outbuf)); \ - CUDA_EXPECT_SUCCESS \ - EXPECT_EQ(work_element_count, std::accumulate(outbuf, outbuf + bufsize, 0)); \ - \ - cfg = GetGpuLaunchConfig(bufsize, d, SetOutbufZero, 0, 0); \ - TF_CHECK_OK(GpuLaunchKernel(SetOutbufZero, cfg.block_count, \ - cfg.thread_per_block, 0, d.stream(), cfg, \ - outbuf)); \ - CUDA_ASSERT_SUCCESS \ - cfg = GetGpuLaunchConfig(work_element_count, d, Count1D, 0, 0); \ - TF_CHECK_OK(GpuLaunchKernel(Count1D, cfg.block_count, cfg.thread_per_block, \ - 0, d.stream(), cfg, bufsize, outbuf)); \ - CUDA_EXPECT_SUCCESS \ - EXPECT_EQ(work_element_count, std::accumulate(outbuf, outbuf + bufsize, 0)) +#define TEST_LAUNCH_PARAMETER(work_element_count) \ + cfg = GetGpuLaunchConfig(bufsize, d); \ + TF_CHECK_OK(GpuLaunchKernel(SetOutbufZero, cfg.block_count, \ + cfg.thread_per_block, 0, d.stream(), cfg, \ + outbuf)); \ + CUDA_ASSERT_SUCCESS \ + cfg = GetGpuLaunchConfig(work_element_count, d); \ + TF_CHECK_OK(GpuLaunchKernel(Count1D, cfg.block_count, cfg.thread_per_block, \ + 0, d.stream(), cfg, bufsize, outbuf)); \ + CUDA_EXPECT_SUCCESS \ + copyToHost(); \ + EXPECT_EQ(work_element_count, \ + std::accumulate(outbuf_host, outbuf_host + bufsize, 0)); \ + \ + cfg = GetGpuLaunchConfig(bufsize, d, SetOutbufZero, 0, 0); \ + TF_CHECK_OK(GpuLaunchKernel(SetOutbufZero, cfg.block_count, \ + cfg.thread_per_block, 0, d.stream(), cfg, \ + outbuf)); \ + CUDA_ASSERT_SUCCESS \ + cfg = GetGpuLaunchConfig(work_element_count, d, Count1D, 0, 0); \ + TF_CHECK_OK(GpuLaunchKernel(Count1D, cfg.block_count, cfg.thread_per_block, \ + 0, d.stream(), cfg, bufsize, outbuf)); \ + CUDA_EXPECT_SUCCESS \ + copyToHost(); \ + EXPECT_EQ(work_element_count, \ + std::accumulate(outbuf_host, outbuf_host + bufsize, 0)); TEST_LAUNCH_PARAMETER(128); TEST_LAUNCH_PARAMETER(129); @@ -221,7 +245,9 @@ TEST_F(GpuLaunchConfigTest, GetGpu2DLaunchConfig) { TF_EXPECT_OK(GpuLaunchKernel(Count2D, cfg.block_count, cfg.thread_per_block, \ 0, d.stream(), cfg, bufsize, outbuf)); \ CUDA_EXPECT_SUCCESS \ - EXPECT_EQ(dimx* dimy, std::accumulate(outbuf, outbuf + bufsize, 0)); \ + copyToHost(); \ + EXPECT_EQ(dimx* dimy, \ + std::accumulate(outbuf_host, outbuf_host + bufsize, 0)); \ \ cfg1d = GetGpuLaunchConfig(bufsize, d, SetOutbufZero, 0, 0); \ TF_EXPECT_OK(GpuLaunchKernel(SetOutbufZero, cfg1d.block_count, \ @@ -232,7 +258,8 @@ TEST_F(GpuLaunchConfigTest, GetGpu2DLaunchConfig) { TF_EXPECT_OK(GpuLaunchKernel(Count2D, cfg.block_count, cfg.thread_per_block, \ 0, d.stream(), cfg, bufsize, outbuf)); \ CUDA_EXPECT_SUCCESS \ - EXPECT_EQ(dimx* dimy, std::accumulate(outbuf, outbuf + bufsize, 0)) + copyToHost(); \ + EXPECT_EQ(dimx* dimy, std::accumulate(outbuf_host, outbuf_host + bufsize, 0)) TEST_LAUNCH_PARAMETER(128, 128); TEST_LAUNCH_PARAMETER(129, 64); @@ -263,7 +290,9 @@ TEST_F(GpuLaunchConfigTest, GetGpu3DLaunchConfig) { TF_EXPECT_OK(GpuLaunchKernel(Count3D, cfg.block_count, cfg.thread_per_block, \ 0, d.stream(), cfg, bufsize, outbuf)); \ CUDA_EXPECT_SUCCESS \ - EXPECT_EQ(dimx* dimy* dimz, std::accumulate(outbuf, outbuf + bufsize, 0)) + copyToHost(); \ + EXPECT_EQ(dimx* dimy* dimz, \ + std::accumulate(outbuf_host, outbuf_host + bufsize, 0)) TEST_LAUNCH_PARAMETER(128, 128, 128); TEST_LAUNCH_PARAMETER(129, 64, 1024); @@ -282,15 +311,19 @@ TEST_F(GpuLaunchConfigTest, GetGpu3DLaunchConfig) { TEST(CudaDeviceFunctionsTest, ShuffleGetSrcLane) { unsigned* failure_count; +#if GOOGLE_CUDA ASSERT_EQ(cudaMallocManaged(&failure_count, sizeof(unsigned)), cudaSuccess); +#else + ASSERT_EQ(hipHostMalloc(&failure_count, sizeof(unsigned), 0), cudaSuccess); +#endif *failure_count = 0; - TF_EXPECT_OK(GpuLaunchKernel(CudaShuffleGetSrcLaneTest, 1, 32, 0, nullptr, - failure_count)); - ASSERT_EQ(cudaDeviceSynchronize(), cudaSuccess); + TF_EXPECT_OK(GpuLaunchKernel(GpuShuffleGetSrcLaneTest, 1, TF_RED_WARPSIZE, 0, + nullptr, failure_count)); + ASSERT_EQ(gpuDeviceSynchronize(), cudaSuccess); ASSERT_EQ(*failure_count, 0); - cudaFree(failure_count); + gpuFree(failure_count); } } // namespace tensorflow -#endif // GOOGLE_CUDA +#endif // GOOGLE_CUDA || TENSORFLOW_USE_ROCM From e5d5522d827e1c60b3ac830000b4489206480f95 Mon Sep 17 00:00:00 2001 From: Elena Zhelezina Date: Mon, 4 May 2020 14:12:51 +0100 Subject: [PATCH 030/867] Both inputs should be in int16 for MUL operator. Some networks have one of inputs as a constant. --- tensorflow/lite/tools/optimize/operator_property.cc | 1 + 1 file changed, 1 insertion(+) diff --git a/tensorflow/lite/tools/optimize/operator_property.cc b/tensorflow/lite/tools/optimize/operator_property.cc index 38c34706fbe..2ffe9fa3671 100644 --- a/tensorflow/lite/tools/optimize/operator_property.cc +++ b/tensorflow/lite/tools/optimize/operator_property.cc @@ -807,6 +807,7 @@ OperatorProperty GetOperatorProperty(const ModelT* model, int subgraph_index, case BuiltinOperator_MUL: property.inputs = {{0, {}}, {1, {}}}; property.outputs = {{0, {}}}; + property.quantize_input_as_activations = true; property.version = 2; break; case BuiltinOperator_PACK: From b4062c71c659bde06e5e15c53424a6853aa1ee3c Mon Sep 17 00:00:00 2001 From: rangjiaheng Date: Wed, 27 May 2020 13:27:18 +0800 Subject: [PATCH 031/867] add TF_LOCKS_EXCLUDED to MutableDenseHashTable::MemoryUsed --- tensorflow/core/kernels/lookup_table_op.cc | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/tensorflow/core/kernels/lookup_table_op.cc b/tensorflow/core/kernels/lookup_table_op.cc index 9807247ed4f..f269aa65b4e 100644 --- a/tensorflow/core/kernels/lookup_table_op.cc +++ b/tensorflow/core/kernels/lookup_table_op.cc @@ -557,7 +557,7 @@ class MutableDenseHashTable final : public LookupInterface { TensorShape value_shape() const override { return value_shape_; } - int64 MemoryUsed() const override { + int64 MemoryUsed() const override TF_LOCKS_EXCLUDED(mu_) { tf_shared_lock l(mu_); return sizeof(MutableDenseHashTable) + key_buckets_.AllocatedBytes() + value_buckets_.AllocatedBytes() + empty_key_.AllocatedBytes(); From 0c7160babc2213269798ea64a292801123edd0d4 Mon Sep 17 00:00:00 2001 From: =?UTF-8?q?M=C3=A5ns=20Nilsson?= Date: Tue, 21 Apr 2020 15:25:11 +0200 Subject: [PATCH 032/867] TFLu: Add MVE flag to cmsis-nn glue for clarity __ARM_FEATURE_MVE is now autodetected so ARM_MATH_MVEI is no longer needed. --- tensorflow/lite/micro/kernels/cmsis-nn/conv.cc | 4 ++-- tensorflow/lite/micro/kernels/cmsis-nn/depthwise_conv.cc | 6 +++--- tensorflow/lite/micro/kernels/cmsis-nn/fully_connected.cc | 7 ++++--- tensorflow/lite/micro/kernels/cmsis-nn/pooling.cc | 4 ++-- tensorflow/lite/micro/tools/make/ext_libs/cmsis.inc | 5 ----- 5 files changed, 11 insertions(+), 15 deletions(-) diff --git a/tensorflow/lite/micro/kernels/cmsis-nn/conv.cc b/tensorflow/lite/micro/kernels/cmsis-nn/conv.cc index 6e8272b221a..c505670d747 100644 --- a/tensorflow/lite/micro/kernels/cmsis-nn/conv.cc +++ b/tensorflow/lite/micro/kernels/cmsis-nn/conv.cc @@ -115,7 +115,7 @@ void* Init(TfLiteContext* context, const char* buffer, size_t length) { } TfLiteStatus Prepare(TfLiteContext* context, TfLiteNode* node) { -#if defined(__ARM_FEATURE_DSP) +#if defined(__ARM_FEATURE_DSP) || defined(__ARM_FEATURE_MVE) OpData data; int32_t buf_size = 0; @@ -240,7 +240,7 @@ TfLiteStatus EvalQuantizedPerChannel( quant_params.multiplier = data->per_channel_output_multiplier; quant_params.shift = data->per_channel_output_shift; -#if defined(__ARM_FEATURE_DSP) +#if defined(__ARM_FEATURE_DSP) || defined(__ARM_FEATURE_MVE) RuntimeShape filter_shape = GetTensorShape(filter); RuntimeShape input_shape = GetTensorShape(input); RuntimeShape output_shape = GetTensorShape(output); diff --git a/tensorflow/lite/micro/kernels/cmsis-nn/depthwise_conv.cc b/tensorflow/lite/micro/kernels/cmsis-nn/depthwise_conv.cc index 7ba03d3890d..f18c3170174 100644 --- a/tensorflow/lite/micro/kernels/cmsis-nn/depthwise_conv.cc +++ b/tensorflow/lite/micro/kernels/cmsis-nn/depthwise_conv.cc @@ -104,7 +104,7 @@ void* Init(TfLiteContext* context, const char* buffer, size_t length) { } TfLiteStatus Prepare(TfLiteContext* context, TfLiteNode* node) { -#if defined(__ARM_FEATURE_DSP) +#if defined(__ARM_FEATURE_DSP) || defined(__ARM_FEATURE_MVE) auto* params = reinterpret_cast(node->builtin_data); @@ -186,7 +186,7 @@ TfLiteStatus EvalQuantizedPerChannel(TfLiteContext* context, TfLiteNode* node, op_params.quantized_activation_min = std::numeric_limits::min(); op_params.quantized_activation_max = std::numeric_limits::max(); -#if defined(__ARM_FEATURE_DSP) +#if defined(__ARM_FEATURE_DSP) || defined(__ARM_FEATURE_MVE) RuntimeShape filter_shape = GetTensorShape(filter); const int filter_height = filter_shape.Dims(1); const int filter_width = filter_shape.Dims(2); @@ -284,7 +284,7 @@ TfLiteStatus EvalQuantized(TfLiteContext* context, TfLiteNode* node, // Legacy ops used mixed left and right shifts. Now all are +ve-means-left. op_params.output_shift = -data->output_shift; -#if defined(__ARM_FEATURE_DSP) +#if defined(__ARM_FEATURE_DSP) || defined(__ARM_FEATURE_MVE) // optimizations utilize loop unrolling which requires the following power // of two kernel dimensions RuntimeShape filter_shape = GetTensorShape(filter); diff --git a/tensorflow/lite/micro/kernels/cmsis-nn/fully_connected.cc b/tensorflow/lite/micro/kernels/cmsis-nn/fully_connected.cc index 78787ea2547..0ab32ecc3c3 100644 --- a/tensorflow/lite/micro/kernels/cmsis-nn/fully_connected.cc +++ b/tensorflow/lite/micro/kernels/cmsis-nn/fully_connected.cc @@ -84,11 +84,11 @@ TfLiteStatus Prepare(TfLiteContext* context, TfLiteNode* node) { TF_LITE_ENSURE_EQ(context, input->type, output->type); TF_LITE_ENSURE_MSG(context, input->type == filter->type, "Hybrid models are not supported on TFLite Micro."); -#if defined(__ARM_FEATURE_DSP) + +#if defined(__ARM_FEATURE_DSP) || defined(__ARM_FEATURE_MVE) RuntimeShape filter_shape = GetTensorShape(filter); const int filter_dim_count = filter_shape.DimensionsCount(); const int accum_depth = filter_shape.Dims(filter_dim_count - 1); - const int32_t buf_size = arm_fully_connected_s8_get_buffer_size(accum_depth); int* buffer_idx = reinterpret_cast(node->user_data); @@ -101,6 +101,7 @@ TfLiteStatus Prepare(TfLiteContext* context, TfLiteNode* node) { *buffer_idx = -1; } #endif + return kTfLiteOk; } @@ -116,7 +117,7 @@ TfLiteStatus EvalQuantizedInt8(TfLiteContext* context, TfLiteNode* node, const int filter_dim_count = filter_shape.DimensionsCount(); const int accum_depth = filter_shape.Dims(filter_dim_count - 1); -#if defined(__ARM_FEATURE_DSP) +#if defined(__ARM_FEATURE_DSP) || defined(__ARM_FEATURE_MVE) int16_t* buf = nullptr; auto* buffer_idx = reinterpret_cast(node->user_data); diff --git a/tensorflow/lite/micro/kernels/cmsis-nn/pooling.cc b/tensorflow/lite/micro/kernels/cmsis-nn/pooling.cc index a12f628e721..8447ad041cc 100644 --- a/tensorflow/lite/micro/kernels/cmsis-nn/pooling.cc +++ b/tensorflow/lite/micro/kernels/cmsis-nn/pooling.cc @@ -106,7 +106,7 @@ TfLiteStatus AverageEvalInt8(TfLiteContext* context, const TfLiteNode* node, TFLITE_DCHECK_LE(activation_min, activation_max); -#if defined(__ARM_FEATURE_DSP) +#if defined(__ARM_FEATURE_DSP) || defined(__ARM_FEATURE_MVE) RuntimeShape input_shape = GetTensorShape(input); TFLITE_DCHECK_EQ(input_shape.DimensionsCount(), 4); @@ -283,7 +283,7 @@ void* Init(TfLiteContext* context, const char* buffer, size_t length) { } TfLiteStatus Prepare(TfLiteContext* context, TfLiteNode* node) { -#if defined(__ARM_FEATURE_DSP) +#if defined(__ARM_FEATURE_DSP) || defined(__ARM_FEATURE_MVE) const TfLiteTensor* input = GetInput(context, node, kInputTensor); const TfLiteTensor* output = GetOutput(context, node, kOutputTensor); diff --git a/tensorflow/lite/micro/tools/make/ext_libs/cmsis.inc b/tensorflow/lite/micro/tools/make/ext_libs/cmsis.inc index 657d4fa87cf..cfd87089a84 100644 --- a/tensorflow/lite/micro/tools/make/ext_libs/cmsis.inc +++ b/tensorflow/lite/micro/tools/make/ext_libs/cmsis.inc @@ -8,11 +8,6 @@ ifneq ($(filter cmsis-nn,$(ALL_TAGS)),) THIRD_PARTY_DOWNLOADS += \ $(eval $(call add_third_party_download,$(CMSIS_URL),$(CMSIS_MD5),cmsis,)) - ifneq (,$(filter $(TARGET_ARCH), cortex-m55)) - CCFLAGS += -DARM_MATH_MVEI - CXXFLAGS += -DARM_MATH_MVEI - endif - CMSIS_PATH = $(MAKEFILE_DIR)/downloads/cmsis/ # Include CMSIS-NN files From 1faa442c02c6bb479b958a46eb2f858c1985966a Mon Sep 17 00:00:00 2001 From: storypku Date: Fri, 29 May 2020 23:00:06 +0800 Subject: [PATCH 033/867] Make third_party/gpus/... build pass. fix issue #39759 --- third_party/gpus/cuda/BUILD | 10 +--------- third_party/gpus/rocm/BUILD | 7 +------ 2 files changed, 2 insertions(+), 15 deletions(-) diff --git a/third_party/gpus/cuda/BUILD b/third_party/gpus/cuda/BUILD index 413d28a2723..9c988668ebc 100644 --- a/third_party/gpus/cuda/BUILD +++ b/third_party/gpus/cuda/BUILD @@ -1,9 +1 @@ -load("@bazel_skylib//:bzl_library.bzl", "bzl_library") - -bzl_library( - name = "build_defs_bzl", - srcs = ["build_defs.bzl"], - deps = [ - "@bazel_skylib//lib:selects", - ], -) +# Dummy BUILD for Bazel diff --git a/third_party/gpus/rocm/BUILD b/third_party/gpus/rocm/BUILD index bc2dd419259..dd64bfb0f36 100644 --- a/third_party/gpus/rocm/BUILD +++ b/third_party/gpus/rocm/BUILD @@ -1,6 +1 @@ -load("@bazel_skylib//:bzl_library.bzl", "bzl_library") - -bzl_library( - name = "build_defs_bzl", - srcs = ["build_defs.bzl"], -) +# placeholder for bazel From cbf60b5223997eceb6c4221ef4868fd6c792622c Mon Sep 17 00:00:00 2001 From: Elena Zhelezina Date: Wed, 3 Jun 2020 12:38:44 +0100 Subject: [PATCH 034/867] Addressed reviewer's comments. Change-Id: I3b7842c42b8c905ed44e0cd556134210cb45479c --- tensorflow/lite/python/lite.py | 37 ++++++++++++++++------------------ 1 file changed, 17 insertions(+), 20 deletions(-) diff --git a/tensorflow/lite/python/lite.py b/tensorflow/lite/python/lite.py index 92c0d5a95d9..010952820b9 100644 --- a/tensorflow/lite/python/lite.py +++ b/tensorflow/lite/python/lite.py @@ -258,7 +258,7 @@ class QuantizationMode(object): if self.training_time_int8_allow_float(): return { - "inference_type": inference_ty if inference_ty else constants.INT8, + "inference_type": inference_ty if inference_ty else self.activations_type(), "inference_input_type": inference_input_ty if inference_input_ty else constants.FLOAT, "post_training_quantize": False, # disable dynamic range quantization @@ -297,12 +297,28 @@ class QuantizationMode(object): return True, { "inference_input_type": inference_input_type, "inference_output_type": inference_output_type, + "activations_type": constants.INT8, "allow_float": False } elif self.post_training_int8_allow_float(): return True, { "inference_input_type": inference_input_type, "inference_output_type": inference_output_type, + "activations_type": constants.INT8, + "allow_float": True + } + elif self.post_training_int16x8_no_float(): + return True, { + "inference_input_type": inference_input_type, + "inference_output_type": inference_output_type, + "activations_type": constants.INT16, + "allow_float": False + } + elif self.post_training_int16x8_allow_float(): + return True, { + "inference_input_type": inference_input_type, + "inference_output_type": inference_output_type, + "activations_type": constants.INT16, "allow_float": True } else: @@ -573,25 +589,6 @@ class TFLiteConverterBaseV2(TFLiteConverterBase): output_tensors=output_tensors, **converter_kwargs) - activations_type = quant_mode.activations_type() - - if quant_mode.post_training_int8_no_float(): - result = self._calibrate_quantize_model(result, constants.FLOAT, - constants.FLOAT, activations_type, - False) - elif quant_mode.post_training_int8_allow_float(): - result = self._calibrate_quantize_model(result, constants.FLOAT, - constants.FLOAT, activations_type, - True) - elif quant_mode.post_training_int16x8_no_float(): - result = self._calibrate_quantize_model(result, constants.FLOAT, - constants.FLOAT, activations_type, - False) - elif quant_mode.post_training_int16x8_allow_float(): - result = self._calibrate_quantize_model(result, constants.FLOAT, - constants.FLOAT, activations_type, - True) - calibrate_and_quantize, flags = quant_mode.quantizer_flags() if calibrate_and_quantize: result = self._calibrate_quantize_model(result, **flags) From 67ea57b15bb223e72a60265c24082ae5a31d0f0e Mon Sep 17 00:00:00 2001 From: Elena Zhelezina Date: Wed, 3 Jun 2020 15:57:18 +0100 Subject: [PATCH 035/867] Small fix for inference. Change-Id: Ifd8670ccb9604ecced3d013f529ddbe16fcd75cf --- tensorflow/lite/python/util.py | 1 + 1 file changed, 1 insertion(+) diff --git a/tensorflow/lite/python/util.py b/tensorflow/lite/python/util.py index 32a2d596629..b56b9f49b7a 100644 --- a/tensorflow/lite/python/util.py +++ b/tensorflow/lite/python/util.py @@ -49,6 +49,7 @@ _MAP_TF_TO_TFLITE_TYPES = { dtypes.string: _types_pb2.STRING, dtypes.uint8: _types_pb2.QUANTIZED_UINT8, dtypes.int8: _types_pb2.INT8, + dtypes.int16: _types_pb2.QUANTIZED_INT16, dtypes.complex64: _types_pb2.COMPLEX64, dtypes.bool: _types_pb2.BOOL, } From 29fdee8e85e750d04f6e9d378e85443ba5c7a239 Mon Sep 17 00:00:00 2001 From: Elena Zhelezina Date: Wed, 3 Jun 2020 17:50:28 +0100 Subject: [PATCH 036/867] Fix for error_reporter. Change-Id: I58745cc97872af74b1ad5b0af3ad778b39f01555 --- .../lite/tools/optimize/quantization_utils.cc | 3 ++- tensorflow/lite/tools/optimize/quantize_model.cc | 16 +++++++++------- 2 files changed, 11 insertions(+), 8 deletions(-) diff --git a/tensorflow/lite/tools/optimize/quantization_utils.cc b/tensorflow/lite/tools/optimize/quantization_utils.cc index abbcb642287..cdf2743585e 100644 --- a/tensorflow/lite/tools/optimize/quantization_utils.cc +++ b/tensorflow/lite/tools/optimize/quantization_utils.cc @@ -113,7 +113,8 @@ TfLiteStatus GetQuantizationParams(TensorT* tensor, TensorType activations_type, tensor->quantization->max[0], quantized_range, quantization_params); } else { - error_reporter->Report( + TF_LITE_REPORT_ERROR( + error_reporter, "Unsupported activation type for quantize-activation: %s", activations_type); return kTfLiteError; diff --git a/tensorflow/lite/tools/optimize/quantize_model.cc b/tensorflow/lite/tools/optimize/quantize_model.cc index 6dd8ddd2d8c..0cf69eee3b4 100644 --- a/tensorflow/lite/tools/optimize/quantize_model.cc +++ b/tensorflow/lite/tools/optimize/quantize_model.cc @@ -370,9 +370,9 @@ TfLiteStatus ApplyConstraints(ModelT* model, std::unique_ptr additional_tensor; const string requant_tensor_name = input_tensor->name + "_requantized"; utils::MakeTensorWithQuantParam( - requant_tensor_name, input_tensor->shape, - input_tensor->shape_signature, activations_type, - output_scale, output_zp, &additional_tensor); + requant_tensor_name, input_tensor->shape, + input_tensor->shape_signature, activations_type, output_scale, + output_zp, &additional_tensor); const int32_t additional_tensor_idx = subgraph->tensors.size(); subgraph->tensors.push_back(std::move(additional_tensor)); @@ -869,13 +869,15 @@ TfLiteStatus QuantizeWeightsInputOutput( if (activations_type == TensorType_INT16 && !property.quantizable && !allow_float) { - error_reporter->Report( - "Quantization to 16x8-bit not yet supported for op: %s", + TF_LITE_REPORT_ERROR( + error_reporter, + "Quantization to 16x8-bit not yet supported for op: %", EnumNameBuiltinOperator(op_code)); return kTfLiteError; } else if (!property.quantizable && !allow_float) { - error_reporter->Report("Quantization not yet supported for op: %s", - EnumNameBuiltinOperator(op_code)); + TF_LITE_REPORT_ERROR(error_reporter, + "Quantization not yet supported for op: %", + EnumNameBuiltinOperator(op_code)); return kTfLiteError; } From 761d850ac6456aed93ab250ff49af3f0a6a62960 Mon Sep 17 00:00:00 2001 From: Elena Zhelezina Date: Wed, 3 Jun 2020 17:56:39 +0100 Subject: [PATCH 037/867] Renamed option with the prefix EXPERIMENTAL_. Change-Id: Idb84736507d5c07ebdf182b8a15d55906d0d7fc0 --- tensorflow/lite/python/convert.py | 2 +- tensorflow/lite/python/lite.py | 2 +- tensorflow/lite/python/lite_test.py | 2 +- 3 files changed, 3 insertions(+), 3 deletions(-) diff --git a/tensorflow/lite/python/convert.py b/tensorflow/lite/python/convert.py index c30987a5898..939de61c608 100644 --- a/tensorflow/lite/python/convert.py +++ b/tensorflow/lite/python/convert.py @@ -98,7 +98,7 @@ class OpsSet(enum.Enum): # and int16 activations. # Specifying this will throw an error for operations that do not yet have # quantized implementations. - TFLITE_BUILTINS_ACTIVATIONS_INT16_WEIGHTS_INT8 = "TFLITE_BUILTINS_ACTIVATIONS_INT16_WEIGHTS_INT8" + EXPERIMENTAL_TFLITE_BUILTINS_ACTIVATIONS_INT16_WEIGHTS_INT8 = "EXPERIMENTAL_TFLITE_BUILTINS_ACTIVATIONS_INT16_WEIGHTS_INT8" def __str__(self): return self.value diff --git a/tensorflow/lite/python/lite.py b/tensorflow/lite/python/lite.py index 010952820b9..781007241b4 100644 --- a/tensorflow/lite/python/lite.py +++ b/tensorflow/lite/python/lite.py @@ -356,7 +356,7 @@ class QuantizationMode(object): def _is_int16x8_target_required(self): return bool( set(self._target_spec.supported_ops).intersection([ - OpsSet.TFLITE_BUILTINS_ACTIVATIONS_INT16_WEIGHTS_INT8 + OpsSet.EXPERIMENTAL_TFLITE_BUILTINS_ACTIVATIONS_INT16_WEIGHTS_INT8 ])) def _is_allow_float(self): diff --git a/tensorflow/lite/python/lite_test.py b/tensorflow/lite/python/lite_test.py index 4075a887943..1d052b88c10 100644 --- a/tensorflow/lite/python/lite_test.py +++ b/tensorflow/lite/python/lite_test.py @@ -885,7 +885,7 @@ class FromSessionTest(TestModels, parameterized.TestCase): # Quantize model to Int8: with disable mlir ('UseTfliteBuiltinsIntDisableMLIR', [lite.OpsSet.TFLITE_BUILTINS_INT8], False), # Quantize model to Int16: with disable mlir - ('UseTfliteBuiltinsInt16DisableMLIR', [lite.OpsSet.TFLITE_BUILTINS_ACTIVATIONS_INT16_WEIGHTS_INT8], False)) + ('UseTfliteBuiltinsInt16DisableMLIR', [lite.OpsSet.EXPERIMENTAL_TFLITE_BUILTINS_ACTIVATIONS_INT16_WEIGHTS_INT8], False)) def testCalibrateAndQuantizeBuiltinInt(self, supported_ops, enable_mlir): with ops.Graph().as_default(): inp, output, calibration_gen = self._getCalibrationQuantizeModel() From ab6b2ffde37bf4443c7dadc39312f0429f417db6 Mon Sep 17 00:00:00 2001 From: Nathan Luehr Date: Fri, 15 May 2020 11:40:22 -0500 Subject: [PATCH 038/867] Removed TENSOR_OP disable env vars. * TF_DISABLE_CUBLAS_TENSOR_OP_MATH * TF_DISABLE_CUDNN_TENSOR_OP_MATH * TF_DISABLE_CUDNN_RNN_TENSOR_OP_MATH --- tensorflow/stream_executor/cuda/cuda_blas.cc | 21 ++------ tensorflow/stream_executor/cuda/cuda_dnn.cc | 55 +++++--------------- 2 files changed, 16 insertions(+), 60 deletions(-) diff --git a/tensorflow/stream_executor/cuda/cuda_blas.cc b/tensorflow/stream_executor/cuda/cuda_blas.cc index c9f0fc462c9..65c07e72154 100644 --- a/tensorflow/stream_executor/cuda/cuda_blas.cc +++ b/tensorflow/stream_executor/cuda/cuda_blas.cc @@ -101,18 +101,6 @@ static std::string ToString(cublasStatus_t status) { } } -// Decide whether to enable TENSOR_OP_MATH -static bool TensorOpMathEnabled() { - static bool is_enabled = [] { - bool is_disabled; - TF_CHECK_OK( - tensorflow::ReadBoolFromEnvVar("TF_DISABLE_CUBLAS_TENSOR_OP_MATH", - /*default_val=*/false, &is_disabled)); - return !is_disabled; - }(); - return is_enabled; -} - // cuBLAS has interfaces that permit pointers to be passed from either the host // memory space or the device memory space; however, you must instruct it as to // which address space those pointers are in with cublasSetPointerMode. @@ -1640,7 +1628,7 @@ bool CUDABlas::DoBlasGemm( &cc_minor); // GPUs < sm_70 don't support tensor ops. - if (cc_major >= 7 && TensorOpMathEnabled()) { + if (cc_major >= 7) { use_tensor_ops = true; } #endif @@ -1921,8 +1909,7 @@ static bool TensorOpsAvailable(int cc_major) { // strictly correct. We can't simply enable it, though, as that would change // clients' behavior significantly: Using tensor ops on fp32 inputs cause them // to be rounded to fp16. - if (cc_major >= 7 && TensorOpMathEnabled() && - std::is_same::value) { + if (cc_major >= 7 && std::is_same::value) { return true; } #endif @@ -2270,7 +2257,7 @@ port::Status CUDABlas::DoBlasGemmBatchedInternal( if (stream->parent()->GetDeviceDescription().cuda_compute_capability( &cc_major, &cc_minor) && cc_major >= 5) { - bool use_tensor_ops = TensorOpMathEnabled() && data_type == CUDA_R_16F; + bool use_tensor_ops = data_type == CUDA_R_16F; cublasGemmAlgo_t algo = (use_tensor_ops ? CUBLAS_GEMM_DFALT_TENSOR_OP : CUBLAS_GEMM_DFALT); cudaDataType_t compute_type = @@ -2425,7 +2412,7 @@ bool CUDABlas::DoBlasGemmStridedBatched( if (stream->parent()->GetDeviceDescription().cuda_compute_capability( &cc_major, &cc_minor)) { // GPUs < sm_70 don't support tensor ops. - if (cc_major >= 7 && TensorOpMathEnabled()) { + if (cc_major >= 7) { use_tensor_ops = true; } #if CUDA_VERSION >= 9010 diff --git a/tensorflow/stream_executor/cuda/cuda_dnn.cc b/tensorflow/stream_executor/cuda/cuda_dnn.cc index 6122877f91f..780f1475c2c 100755 --- a/tensorflow/stream_executor/cuda/cuda_dnn.cc +++ b/tensorflow/stream_executor/cuda/cuda_dnn.cc @@ -603,31 +603,6 @@ class CudnnFilterDescriptor { SE_DISALLOW_COPY_AND_ASSIGN(CudnnFilterDescriptor); }; -// A helper function to decide whether to enable the TENSOR_OP_MATH math type -bool TensorOpMathEnabled() { - static bool is_enabled = [] { - bool is_disabled = false; - TF_CHECK_OK( - tensorflow::ReadBoolFromEnvVar("TF_DISABLE_CUDNN_TENSOR_OP_MATH", - /*default_val=*/false, &is_disabled)); - return !is_disabled; - }(); - return is_enabled; -} - -// A helper function to decide whether to enable the TENSOR_OP_MATH math type -// for RNNs. -bool RnnTensorOpMathEnabled() { - static bool is_enabled = [] { - bool is_disabled = false; - TF_CHECK_OK( - tensorflow::ReadBoolFromEnvVar("TF_DISABLE_CUDNN_RNN_TENSOR_OP_MATH", - /*default_val=*/false, &is_disabled)); - return !is_disabled; - }(); - return is_enabled; -} - // A helper function to decide whether to use // CUDNN_BATCHNORM_SPATIAL_PERSISTENT in batchnorm. This mode can be faster in // some tasks because an optimized path may be selected for CUDNN_DATA_FLOAT @@ -751,9 +726,7 @@ class CudnnConvolutionDescriptor { #if CUDNN_VERSION >= 7000 cudnnMathType_t math_type = (use_tensor_op_math ? CUDNN_TENSOR_OP_MATH : CUDNN_DEFAULT_MATH); - if (TensorOpMathEnabled()) { - CHECK_CUDNN_OK(cudnnSetConvolutionMathType(handle_.get(), math_type)); - } + CHECK_CUDNN_OK(cudnnSetConvolutionMathType(handle_.get(), math_type)); #endif } @@ -1157,21 +1130,19 @@ class CudnnRnnDescriptor : public dnn::RnnDescriptor { // in profile mode, which is run with algorithms returned from // GetRnnAlgorithms() (which are non-default and explicitly set whether to // use tensor ops). CuDNN 7.2.1 fixed this issue - if (RnnTensorOpMathEnabled()) { - cudnnMathType_t math_type; - if (algorithm_config.algorithm().has_value()) { - math_type = algorithm_config.algorithm()->tensor_ops_enabled() - ? CUDNN_TENSOR_OP_MATH - : CUDNN_DEFAULT_MATH; - } else { + cudnnMathType_t math_type; + if (algorithm_config.algorithm().has_value()) { + math_type = algorithm_config.algorithm()->tensor_ops_enabled() + ? CUDNN_TENSOR_OP_MATH + : CUDNN_DEFAULT_MATH; + } else { #if CUDNN_VERSION >= 7201 - math_type = CUDNN_TENSOR_OP_MATH; + math_type = CUDNN_TENSOR_OP_MATH; #else - math_type = CUDNN_DEFAULT_MATH; + math_type = CUDNN_DEFAULT_MATH; #endif // CUDNN_VERSION >= 7201 - } - CHECK_CUDNN_OK(cudnnSetRNNMatrixMathType(rnn_desc.get(), math_type)); } + CHECK_CUDNN_OK(cudnnSetRNNMatrixMathType(rnn_desc.get(), math_type)); #endif // CUDNN_VERSION >= 7000 return CudnnRnnDescriptor(cudnn, std::move(rnn_desc), std::move(rnn_plan), @@ -2605,7 +2576,7 @@ AllocateCudnnConvolutionBackwardFilterWorkspace( } static bool TensorOpMathAvailable(int cc_major) { - return cc_major >= 7 && CUDNN_VERSION >= 7000 && TensorOpMathEnabled(); + return cc_major >= 7 && CUDNN_VERSION >= 7000; } port::StatusOr GetCudnnConvolutionForwardAlgorithm( @@ -3399,9 +3370,7 @@ bool CudnnSupport::GetRnnAlgorithms( for (auto i : algo_types) { out_algorithms->push_back({i, /*use_tensor_ops=*/false}); #if CUDNN_VERSION >= 7100 - if (RnnTensorOpMathEnabled()) { - out_algorithms->push_back({i, /*use_tensor_ops=*/true}); - } + out_algorithms->push_back({i, /*use_tensor_ops=*/true}); #endif } return true; From 3cad62e356b8c72d03af13ba29f7ace29a6f0772 Mon Sep 17 00:00:00 2001 From: Nathan Luehr Date: Fri, 15 May 2020 11:46:41 -0500 Subject: [PATCH 039/867] Add global setting control TF32 execution --- tensorflow/core/platform/BUILD | 7 +++++++ tensorflow/core/platform/tf32_utils.cc | 27 ++++++++++++++++++++++++++ tensorflow/core/platform/tf32_utils.h | 27 ++++++++++++++++++++++++++ 3 files changed, 61 insertions(+) create mode 100644 tensorflow/core/platform/tf32_utils.cc create mode 100644 tensorflow/core/platform/tf32_utils.h diff --git a/tensorflow/core/platform/BUILD b/tensorflow/core/platform/BUILD index c7ff378d2ac..f27d2f09208 100644 --- a/tensorflow/core/platform/BUILD +++ b/tensorflow/core/platform/BUILD @@ -937,6 +937,13 @@ cc_library( alwayslink = 1, ) +cc_library( + name = "tf32_utils", + srcs = ["tf32_utils.cc"], + hdrs = ["tf32_utils.h"], + copts = tf_copts(), +) + tf_cc_tests( name = "low_level_library_tests", size = "small", diff --git a/tensorflow/core/platform/tf32_utils.cc b/tensorflow/core/platform/tf32_utils.cc new file mode 100644 index 00000000000..715b5996dc3 --- /dev/null +++ b/tensorflow/core/platform/tf32_utils.cc @@ -0,0 +1,27 @@ +/* Copyright 2020 The TensorFlow Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. +==============================================================================*/ + +#include "tensorflow/core/platform/tf32_utils.h" + +namespace tensorflow { + +// TODO(nluehr): enable tf32 execution by default after TF32 Ampere testing. +static bool tf32_enabled = false; + +void allow_tf32_execution(bool allow) { tf32_enabled = allow; } + +bool tf32_execution_allowed() { return tf32_enabled; } + +} // namespace tensorflow diff --git a/tensorflow/core/platform/tf32_utils.h b/tensorflow/core/platform/tf32_utils.h new file mode 100644 index 00000000000..a0ce58f9bbd --- /dev/null +++ b/tensorflow/core/platform/tf32_utils.h @@ -0,0 +1,27 @@ +/* Copyright 2020 The TensorFlow Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. +==============================================================================*/ + +#ifndef TENSORFLOW_CORE_PLATFORM_TF32_UTILS_H_ +#define TENSORFLOW_CORE_PLATFORM_TF32_UTILS_H_ + +namespace tensorflow { + +void allow_tf32_execution(bool allow); + +bool tf32_execution_allowed(); + +} // namespace tensorflow + +#endif // TENSORFLOW_CORE_PLATFORM_TF32_UTILS_H_ From 8bfee17f5880eccdb759fb47ab11b782f201cf0f Mon Sep 17 00:00:00 2001 From: Nathan Luehr Date: Fri, 15 May 2020 13:33:02 -0500 Subject: [PATCH 040/867] Python tf.config tf32 interface --- tensorflow/python/BUILD | 11 +++++++++++ tensorflow/python/framework/config.py | 26 ++++++++++++++++++++++++++ tensorflow/python/util/tf32.cc | 22 ++++++++++++++++++++++ 3 files changed, 59 insertions(+) create mode 100644 tensorflow/python/util/tf32.cc diff --git a/tensorflow/python/BUILD b/tensorflow/python/BUILD index a49e4b74def..997ec6c924f 100644 --- a/tensorflow/python/BUILD +++ b/tensorflow/python/BUILD @@ -746,6 +746,16 @@ tf_python_pybind_extension( ], ) +tf_python_pybind_extension( + name = "_pywrap_tf32_execution", + srcs = ["util/tf32.cc"], + module_name = "_pywrap_tf32_execution", + deps = [ + "//tensorflow/core/platform:tf32_utils", + "@pybind11", + ], +) + tf_python_pybind_extension( name = "_pywrap_util_port", srcs = ["util/port_wrapper.cc"], @@ -5573,6 +5583,7 @@ py_library( "//tensorflow:composite_tensor_whitelist", ], deps = [ + ":_pywrap_tf32_execution", ":tf_decorator", ":tf_export", ":tf_stack", diff --git a/tensorflow/python/framework/config.py b/tensorflow/python/framework/config.py index 5361d7290e8..042af4d1023 100644 --- a/tensorflow/python/framework/config.py +++ b/tensorflow/python/framework/config.py @@ -18,10 +18,36 @@ from __future__ import absolute_import from __future__ import division from __future__ import print_function +from tensorflow.python import _pywrap_tf32_execution from tensorflow.python.eager import context from tensorflow.python.util import deprecation from tensorflow.python.util.tf_export import tf_export +def tensor_float32_execution_allowed(): + """Get if TensorFloat-32 operations are enabled on supported hardware. + + Returns: + True if TensorFloat-32 execution is enabled and False otherwise. + """ + return _pywrap_tf32_execution.is_allowed() + +def allow_tensor_float_32_execution(allow): + """Allow use of TensorFloat-32 with float32 ops on supported hardware. + + TensorFloat-32 is a math mode introduced with the NVIDIA Ampere architecture. + TensorFloat-32 kernels take float32 inputs and produce float32 outputs. + Internally, the inputs are cast to a custom representation with 10-bit + mantissa (similar to float16) and 8-bit exponent (similar to float32) and are + executed using TensorCores with float32 accumulation. For more information, + see https://blogs.nvidia.com/blog/2020/05/14/tensorfloat-32-precision-format/. + + TensorFloat-32 execution is disabled by default, but this may change in a + future version. + + Args: + allow: whether to allow TensorFloat-32 execution + """ + _pywrap_tf32_execution.allow(allow) @tf_export('config.threading.get_intra_op_parallelism_threads') def get_intra_op_parallelism_threads(): diff --git a/tensorflow/python/util/tf32.cc b/tensorflow/python/util/tf32.cc new file mode 100644 index 00000000000..7dece6ccdae --- /dev/null +++ b/tensorflow/python/util/tf32.cc @@ -0,0 +1,22 @@ +/* Copyright 2020 The TensorFlow Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. +==============================================================================*/ + +#include "pybind11/pybind11.h" +#include "tensorflow/core/platform/tf32_utils.h" + +PYBIND11_MODULE(_pywrap_tf32_execution, m) { + m.def("allow", &tensorflow::allow_tf32_execution); + m.def("is_allowed", &tensorflow::tf32_execution_allowed); +} From dedb51aec2a766bdeb8b4c2ab1700bfcf7687966 Mon Sep 17 00:00:00 2001 From: Nathan Luehr Date: Tue, 19 May 2020 14:58:30 -0500 Subject: [PATCH 041/867] Convolution TF32 Plumbing --- tensorflow/stream_executor/cuda/BUILD | 1 + tensorflow/stream_executor/cuda/cuda_dnn.cc | 200 +++++++++++++------- 2 files changed, 135 insertions(+), 66 deletions(-) mode change 100755 => 100644 tensorflow/stream_executor/cuda/cuda_dnn.cc diff --git a/tensorflow/stream_executor/cuda/BUILD b/tensorflow/stream_executor/cuda/BUILD index 1457a36beaf..2749281335e 100644 --- a/tensorflow/stream_executor/cuda/BUILD +++ b/tensorflow/stream_executor/cuda/BUILD @@ -353,6 +353,7 @@ cc_library( "@local_config_cuda//cuda:cudnn_header", "//tensorflow/core:lib", "//tensorflow/core:lib_internal", + "//tensorflow/core/platform:tf32_utils", "//tensorflow/stream_executor:dnn", "//tensorflow/stream_executor:event", "//tensorflow/stream_executor:plugin_registry", diff --git a/tensorflow/stream_executor/cuda/cuda_dnn.cc b/tensorflow/stream_executor/cuda/cuda_dnn.cc old mode 100755 new mode 100644 index 780f1475c2c..53296f4eea5 --- a/tensorflow/stream_executor/cuda/cuda_dnn.cc +++ b/tensorflow/stream_executor/cuda/cuda_dnn.cc @@ -20,8 +20,8 @@ limitations under the License. #include #include "absl/strings/str_cat.h" -#include "third_party/eigen3/Eigen/Core" #include "tensorflow/core/lib/core/errors.h" +#include "tensorflow/core/platform/tf32_utils.h" #include "tensorflow/core/util/env_var.h" #include "tensorflow/stream_executor/cuda/cuda_activation.h" #include "tensorflow/stream_executor/cuda/cuda_diagnostics.h" @@ -42,6 +42,7 @@ limitations under the License. #include "tensorflow/stream_executor/scratch_allocator.h" #include "tensorflow/stream_executor/stream.h" #include "tensorflow/stream_executor/stream_executor_pimpl.h" +#include "third_party/eigen3/Eigen/Core" // clang-format off #include "third_party/gpus/cudnn/cudnn.h" #include "absl/strings/string_view.h" @@ -707,10 +708,6 @@ class CudnnConvolutionDescriptor { : CUDNN_CROSS_CORRELATION, data_type)); - // NOTE(benbarsdell): This only applies if tensor op math is enabled - // and algo selection is set to Default. - this->set_use_tensor_op_math(true); - #if CUDNN_MAJOR >= 7 VLOG(2) << "Requesting grouped convolution: " << convolution_descriptor.group_count(); @@ -722,10 +719,14 @@ class CudnnConvolutionDescriptor { #endif } - void set_use_tensor_op_math(bool use_tensor_op_math) const { + void set_use_tensor_op_math(bool use_tensor_op_math) { #if CUDNN_VERSION >= 7000 cudnnMathType_t math_type = +#if CUDNN_VERSION >= 8000 + (use_tensor_op_math ? CUDNN_DEFAULT_MATH : CUDNN_FMA_MATH); +#else (use_tensor_op_math ? CUDNN_TENSOR_OP_MATH : CUDNN_DEFAULT_MATH); +#endif CHECK_CUDNN_OK(cudnnSetConvolutionMathType(handle_.get(), math_type)); #endif } @@ -738,6 +739,38 @@ class CudnnConvolutionDescriptor { SE_DISALLOW_COPY_AND_ASSIGN(CudnnConvolutionDescriptor); }; +// A helper function to query if a CudnnConvolutionDescriptor has tensor_op_math +// set +static bool IsTensorMathOpSet(const CudnnConvolutionDescriptor& conv) { + cudnnMathType_t math_type; + CHECK_CUDNN_OK(cudnnGetConvolutionMathType(conv.handle(), &math_type)); +#if CUDNN_VERSION >= 8000 + return math_type != CUDNN_FMA_MATH; +#else + return math_type == CUDNN_TENSOR_OP_MATH; +#endif +} + +static bool TensorOpMathAvailable(int cc_major) { + return cc_major >= 7 && CUDNN_VERSION >= 7000; +} + +static bool IsTensorMathAllowed(Stream* stream, dnn::DataType input_type) { + int cc_major, cc_minor; + std::tie(cc_major, cc_minor) = GetCcMajorMinor(stream); + if (!TensorOpMathAvailable(cc_major)) { + return false; + } + if (input_type == dnn::DataType::kFloat) { + if (CUDNN_VERSION < 8000) { + return false; + } else if (!tensorflow::tf32_execution_allowed()) { + return false; + } + } + return true; +} + // Turns a PoolingDescriptor structure into a cudnn pooling descriptor handle // within a scope. class CudnnPoolingDescriptor { @@ -2450,10 +2483,11 @@ port::StatusOr> AllocateCudnnConvolutionForwardWorkspace( const CudnnTensorDescriptor& output_nd, const dnn::AlgorithmDesc& algorithm_desc, ScratchAllocator* scratch_allocator) { - // TODO(csigg): This has side effects on the convolution descriptor. It is - // functionally correct because the convolution is run with the algorithm of - // the last call to this function, but should be fixed anyway. - conv.set_use_tensor_op_math(algorithm_desc.tensor_ops_enabled()); + if (IsTensorMathOpSet(conv) != algorithm_desc.tensor_ops_enabled()) { + return port::Status( + port::error::INTERNAL, + "Mismatch between cudnn conv and algorithm descriptors."); + } // Query the size of the workspace and allocate it. size_t size_in_bytes; @@ -2493,10 +2527,11 @@ AllocateCudnnConvolutionBackwardDataWorkspace( const CudnnTensorDescriptor& output_nd, const dnn::AlgorithmDesc& algorithm_desc, ScratchAllocator* scratch_allocator) { - // TODO(csigg): This has side effects on the convolution descriptor. It is - // functionally correct because the convolution is run with the algorithm of - // the last call to this function, but should be fixed anyway. - conv.set_use_tensor_op_math(algorithm_desc.tensor_ops_enabled()); + if (IsTensorMathOpSet(conv) != algorithm_desc.tensor_ops_enabled()) { + return port::Status( + port::error::INTERNAL, + "Mismatch between cudnn conv and algorithm descriptors."); + } // Query the size of the workspace and allocate it. size_t size_in_bytes; @@ -2538,10 +2573,11 @@ AllocateCudnnConvolutionBackwardFilterWorkspace( const CudnnTensorDescriptor& output_nd, const dnn::AlgorithmDesc& algorithm_desc, ScratchAllocator* scratch_allocator) { - // TODO(csigg): This has side effects on the convolution descriptor. It is - // functionally correct because the convolution is run with the algorithm of - // the last call to this function, but should be fixed anyway. - conv.set_use_tensor_op_math(algorithm_desc.tensor_ops_enabled()); + if (IsTensorMathOpSet(conv) != algorithm_desc.tensor_ops_enabled()) { + return port::Status( + port::error::INTERNAL, + "Mismatch between cudnn conv and algorithm descriptors."); + } // Query the size of the workspace and allocate it. size_t size_in_bytes; @@ -2575,18 +2611,39 @@ AllocateCudnnConvolutionBackwardFilterWorkspace( return scratch_allocator->AllocateBytes(size_in_bytes); } -static bool TensorOpMathAvailable(int cc_major) { - return cc_major >= 7 && CUDNN_VERSION >= 7000; +port::StatusOr UseTensorOps(Stream* stream, dnn::DataType type, + absl::optional desc) { + bool use_tensor_ops; + if (desc.has_value()) { + use_tensor_ops = desc->tensor_ops_enabled(); + if (use_tensor_ops && !IsTensorMathAllowed(stream, type)) { + return port::Status(port::error::INVALID_ARGUMENT, + "Algo requests disallowed tensor op evaluation."); + } + } else { + use_tensor_ops = IsTensorMathAllowed(stream, type); + } + return use_tensor_ops; } port::StatusOr GetCudnnConvolutionForwardAlgorithm( Stream* stream, const CudnnHandle& cudnn, const dnn::AlgorithmConfig& algorithm_config, const CudnnTensorDescriptor& input_nd, const CudnnFilterDescriptor& filter, - const CudnnConvolutionDescriptor& conv, + dnn::DataType element_type, + const dnn::ConvolutionDescriptor& convolution_descriptor, const CudnnTensorDescriptor& output_nd, ScratchAllocator* scratch_allocator, DeviceMemory* scratch) { absl::optional algo_desc = algorithm_config.algorithm(); + + CudnnConvolutionDescriptor conv( + convolution_descriptor, + ToCudnnDataType(GetConvAccumulatorType(element_type))); + bool use_tensor_ops; + SE_ASSIGN_OR_RETURN(use_tensor_ops, + UseTensorOps(stream, element_type, algo_desc)); + conv.set_use_tensor_op_math(use_tensor_ops); + if (!algo_desc.has_value()) { // Pick fastest algorithm within memory limit according to cuDNN's // heuristics. @@ -2599,10 +2656,7 @@ port::StatusOr GetCudnnConvolutionForwardAlgorithm( GetCudnnConvolutionForwardAlgo( cudnn, input_nd, filter, conv, output_nd, specify_workspace_limit, memory_limit_bytes)); - int cc_major, cc_minor; - std::tie(cc_major, cc_minor) = GetCcMajorMinor(stream); - algo_desc = dnn::AlgorithmDesc( - algo, /*use_tensor_ops=*/TensorOpMathAvailable(cc_major)); + algo_desc = dnn::AlgorithmDesc(algo, use_tensor_ops); } const auto scratch_or = AllocateCudnnConvolutionForwardWorkspace( @@ -2626,6 +2680,9 @@ port::StatusOr GetCudnnConvolutionForwardAlgorithm( "Returned status: ", scratch_or.status().ToString())); } + SE_ASSIGN_OR_RETURN(use_tensor_ops, + UseTensorOps(stream, element_type, algo_desc)); + conv.set_use_tensor_op_math(use_tensor_ops); SE_ASSIGN_OR_RETURN(*scratch, AllocateCudnnConvolutionForwardWorkspace( stream, cudnn, input_nd, filter, conv, output_nd, *algo_desc, scratch_allocator)); @@ -2636,10 +2693,19 @@ port::StatusOr GetCudnnConvolutionBackwardDataAlgorithm( Stream* stream, const CudnnHandle& cudnn, const dnn::AlgorithmConfig& algorithm_config, const CudnnTensorDescriptor& input_nd, const CudnnFilterDescriptor& filter, - const CudnnConvolutionDescriptor& conv, + dnn::DataType element_type, + const dnn::ConvolutionDescriptor& convolution_descriptor, const CudnnTensorDescriptor& output_nd, ScratchAllocator* scratch_allocator, DeviceMemory* scratch) { absl::optional algo_desc = algorithm_config.algorithm(); + CudnnConvolutionDescriptor conv( + convolution_descriptor, + ToCudnnDataType(GetConvAccumulatorType(element_type))); + bool use_tensor_ops; + SE_ASSIGN_OR_RETURN(use_tensor_ops, + UseTensorOps(stream, element_type, algo_desc)); + conv.set_use_tensor_op_math(use_tensor_ops); + if (!algo_desc.has_value()) { // Pick fastest algorithm within memory limit according to cuDNN's // heuristics. @@ -2652,10 +2718,7 @@ port::StatusOr GetCudnnConvolutionBackwardDataAlgorithm( GetCudnnConvolutionBackwardDataAlgo( cudnn, input_nd, filter, conv, output_nd, specify_workspace_limit, memory_limit_bytes)); - int cc_major, cc_minor; - std::tie(cc_major, cc_minor) = GetCcMajorMinor(stream); - algo_desc = dnn::AlgorithmDesc( - algo, /*use_tensor_ops=*/TensorOpMathAvailable(cc_major)); + algo_desc = dnn::AlgorithmDesc(algo, use_tensor_ops); } const auto scratch_or = AllocateCudnnConvolutionBackwardDataWorkspace( @@ -2678,6 +2741,9 @@ port::StatusOr GetCudnnConvolutionBackwardDataAlgorithm( "while a secondary algorithm is not provided."); } + SE_ASSIGN_OR_RETURN(use_tensor_ops, + UseTensorOps(stream, element_type, algo_desc)); + conv.set_use_tensor_op_math(use_tensor_ops); SE_ASSIGN_OR_RETURN(*scratch, AllocateCudnnConvolutionBackwardDataWorkspace( stream, cudnn, input_nd, filter, conv, output_nd, *algo_desc, scratch_allocator)); @@ -2688,10 +2754,19 @@ port::StatusOr GetCudnnConvolutionBackwardFilterAlgorithm( Stream* stream, const CudnnHandle& cudnn, const dnn::AlgorithmConfig& algorithm_config, const CudnnTensorDescriptor& input_nd, const CudnnFilterDescriptor& filter, - const CudnnConvolutionDescriptor& conv, + dnn::DataType element_type, + const dnn::ConvolutionDescriptor& convolution_descriptor, const CudnnTensorDescriptor& output_nd, ScratchAllocator* scratch_allocator, DeviceMemory* scratch) { absl::optional algo_desc = algorithm_config.algorithm(); + CudnnConvolutionDescriptor conv( + convolution_descriptor, + ToCudnnDataType(GetConvAccumulatorType(element_type))); + bool use_tensor_ops; + SE_ASSIGN_OR_RETURN(use_tensor_ops, + UseTensorOps(stream, element_type, algo_desc)); + conv.set_use_tensor_op_math(use_tensor_ops); + if (!algo_desc.has_value()) { // Pick fastest algorithm within memory limit according to cuDNN's // heuristics. @@ -2704,10 +2779,7 @@ port::StatusOr GetCudnnConvolutionBackwardFilterAlgorithm( GetCudnnConvolutionBackwardFilterAlgo( cudnn, input_nd, filter, conv, output_nd, specify_workspace_limit, memory_limit_bytes)); - int cc_major, cc_minor; - std::tie(cc_major, cc_minor) = GetCcMajorMinor(stream); - algo_desc = dnn::AlgorithmDesc( - algo, /*use_tensor_ops=*/TensorOpMathAvailable(cc_major)); + algo_desc = dnn::AlgorithmDesc(algo, use_tensor_ops); } auto scratch_or = AllocateCudnnConvolutionBackwardFilterWorkspace( @@ -2730,6 +2802,9 @@ port::StatusOr GetCudnnConvolutionBackwardFilterAlgorithm( "while a secondary algorithm is not provided."); } + SE_ASSIGN_OR_RETURN(use_tensor_ops, + UseTensorOps(stream, element_type, algo_desc)); + conv.set_use_tensor_op_math(use_tensor_ops); SE_ASSIGN_OR_RETURN(*scratch, AllocateCudnnConvolutionBackwardFilterWorkspace( stream, cudnn, input_nd, filter, conv, output_nd, *algo_desc, scratch_allocator)); @@ -2894,35 +2969,32 @@ port::Status CudnnSupport::DoPrepareForConvolution( CudnnTensorDescriptor output_nd( output_descriptor, ToCudnnDataType(element_type, output_descriptor.layout())); - CudnnConvolutionDescriptor conv( - convolution_descriptor, - ToCudnnDataType(GetConvAccumulatorType(element_type))); auto cudnn = cudnn_->GetHandle(parent_, stream); switch (kind) { case dnn::ConvolutionKind::FORWARD: { - SE_ASSIGN_OR_RETURN( - *algorithm_desc, - GetCudnnConvolutionForwardAlgorithm( - stream, cudnn, algorithm_config, input_nd, filter_nd, conv, - output_nd, scratch_allocator, scratch_memory)); + SE_ASSIGN_OR_RETURN(*algorithm_desc, + GetCudnnConvolutionForwardAlgorithm( + stream, cudnn, algorithm_config, input_nd, + filter_nd, element_type, convolution_descriptor, + output_nd, scratch_allocator, scratch_memory)); break; } case dnn::ConvolutionKind::BACKWARD_DATA: { - SE_ASSIGN_OR_RETURN( - *algorithm_desc, - GetCudnnConvolutionBackwardDataAlgorithm( - stream, cudnn, algorithm_config, input_nd, filter_nd, conv, - output_nd, scratch_allocator, scratch_memory)); + SE_ASSIGN_OR_RETURN(*algorithm_desc, + GetCudnnConvolutionBackwardDataAlgorithm( + stream, cudnn, algorithm_config, input_nd, + filter_nd, element_type, convolution_descriptor, + output_nd, scratch_allocator, scratch_memory)); break; } case dnn::ConvolutionKind::BACKWARD_FILTER: { - SE_ASSIGN_OR_RETURN( - *algorithm_desc, - GetCudnnConvolutionBackwardFilterAlgorithm( - stream, cudnn, algorithm_config, input_nd, filter_nd, conv, - output_nd, scratch_allocator, scratch_memory)); + SE_ASSIGN_OR_RETURN(*algorithm_desc, + GetCudnnConvolutionBackwardFilterAlgorithm( + stream, cudnn, algorithm_config, input_nd, + filter_nd, element_type, convolution_descriptor, + output_nd, scratch_allocator, scratch_memory)); break; } default: @@ -2951,8 +3023,9 @@ port::Status CudnnSupport::DoConvolve( auto accumulator_type = GetConvAccumulatorType(element_type); CudnnConvolutionDescriptor conv(convolution_descriptor, ToCudnnDataType(accumulator_type)); - // Set use_tensor_math param to correct value - conv.set_use_tensor_op_math(algorithm_desc.tensor_ops_enabled()); + SE_ASSIGN_OR_RETURN(bool use_tensor_ops, + UseTensorOps(stream, element_type, algorithm_desc)); + conv.set_use_tensor_op_math(use_tensor_ops); auto cudnn = cudnn_->GetHandle(parent_, stream); // Alpha is the scaling factor for input. @@ -3185,14 +3258,6 @@ port::Status CudnnSupport::DoConvolve( return port::Status::OK(); } -// A helper function to query if a CudnnConvolutionDescriptor has tensor_op_math -// set -static bool IsTensorMathOpSet(const CudnnConvolutionDescriptor& conv) { - cudnnMathType_t math_type; - CHECK_CUDNN_OK(cudnnGetConvolutionMathType(conv.handle(), &math_type)); - return math_type == CUDNN_TENSOR_OP_MATH; -} - template port::Status CudnnSupport::DoFusedConvolveImpl( @@ -3226,8 +3291,6 @@ port::Status CudnnSupport::DoFusedConvolveImpl( filter_descriptor, GetCudnnDataType(conv_input_descriptor.layout())); CudnnTensorDescriptor bias_nd(bias_descriptor, GetCudnnDataType()); - CudnnConvolutionDescriptor conv(convolution_descriptor, - ToCudnnDataType(accumulator_type)); auto cudnn = cudnn_->GetHandle(parent_, stream); @@ -3237,9 +3300,14 @@ port::Status CudnnSupport::DoFusedConvolveImpl( SE_ASSIGN_OR_RETURN( dnn::AlgorithmDesc algo_desc, GetCudnnConvolutionForwardAlgorithm( - stream, cudnn, algorithm_config, conv_input_nd, filter, conv, + stream, cudnn, algorithm_config, conv_input_nd, filter, + dnn::ToDataType::value, convolution_descriptor, output_nd, scratch_allocator, &scratch)); + CudnnConvolutionDescriptor conv(convolution_descriptor, + ToCudnnDataType(accumulator_type)); + conv.set_use_tensor_op_math(algo_desc.tensor_ops_enabled()); + std::unique_ptr timer; if (is_profiling) { timer.reset(new GpuTimer(parent_)); // NOLINT From 0f58bb63090222cef0eebe74630b2d4d9d886a2f Mon Sep 17 00:00:00 2001 From: Nathan Luehr Date: Tue, 19 May 2020 15:54:10 -0500 Subject: [PATCH 042/867] Plumb TF32 for RNN --- tensorflow/stream_executor/cuda/cuda_dnn.cc | 30 ++++++++++++++------- 1 file changed, 21 insertions(+), 9 deletions(-) diff --git a/tensorflow/stream_executor/cuda/cuda_dnn.cc b/tensorflow/stream_executor/cuda/cuda_dnn.cc index 53296f4eea5..fa06d410323 100644 --- a/tensorflow/stream_executor/cuda/cuda_dnn.cc +++ b/tensorflow/stream_executor/cuda/cuda_dnn.cc @@ -1163,17 +1163,26 @@ class CudnnRnnDescriptor : public dnn::RnnDescriptor { // in profile mode, which is run with algorithms returned from // GetRnnAlgorithms() (which are non-default and explicitly set whether to // use tensor ops). CuDNN 7.2.1 fixed this issue - cudnnMathType_t math_type; + bool allow_tensor_ops = + data_type != CUDNN_DATA_FLOAT || tensorflow::tf32_execution_allowed(); + bool use_tensor_ops; if (algorithm_config.algorithm().has_value()) { - math_type = algorithm_config.algorithm()->tensor_ops_enabled() - ? CUDNN_TENSOR_OP_MATH - : CUDNN_DEFAULT_MATH; + use_tensor_ops = algorithm_config.algorithm()->tensor_ops_enabled(); } else { -#if CUDNN_VERSION >= 7201 - math_type = CUDNN_TENSOR_OP_MATH; -#else - math_type = CUDNN_DEFAULT_MATH; -#endif // CUDNN_VERSION >= 7201 + use_tensor_ops = CUDNN_VERSION >= 7201 && allow_tensor_ops; + } + + if (use_tensor_ops && !allow_tensor_ops) { + return port::Status(port::error::INVALID_ARGUMENT, + "Algo requests disallowed tensor op evaluation."); + } + + cudnnMathType_t math_type; + if (use_tensor_ops) { + math_type = + CUDNN_VERSION >= 8000 ? CUDNN_DEFAULT_MATH : CUDNN_TENSOR_OP_MATH; + } else { + math_type = CUDNN_VERSION >= 8000 ? CUDNN_FMA_MATH : CUDNN_DEFAULT_MATH; } CHECK_CUDNN_OK(cudnnSetRNNMatrixMathType(rnn_desc.get(), math_type)); #endif // CUDNN_VERSION >= 7000 @@ -2626,6 +2635,9 @@ port::StatusOr UseTensorOps(Stream* stream, dnn::DataType type, return use_tensor_ops; } +cudnnDataType_t GetRnnComputeType(dnn::DataType data_type); +dnn::DataType GetConvAccumulatorType(dnn::DataType data_type); + port::StatusOr GetCudnnConvolutionForwardAlgorithm( Stream* stream, const CudnnHandle& cudnn, const dnn::AlgorithmConfig& algorithm_config, From b67608e66c54224fa52200095fba09df0f2b3c71 Mon Sep 17 00:00:00 2001 From: Nathan Luehr Date: Wed, 20 May 2020 10:06:35 -0500 Subject: [PATCH 043/867] Plumb TF32 for cublas gemm --- tensorflow/stream_executor/cuda/BUILD | 1 + tensorflow/stream_executor/cuda/cuda_blas.cc | 84 +++++++++----------- tensorflow/stream_executor/cuda/cuda_blas.h | 8 +- 3 files changed, 43 insertions(+), 50 deletions(-) diff --git a/tensorflow/stream_executor/cuda/BUILD b/tensorflow/stream_executor/cuda/BUILD index 2749281335e..519033a62d8 100644 --- a/tensorflow/stream_executor/cuda/BUILD +++ b/tensorflow/stream_executor/cuda/BUILD @@ -251,6 +251,7 @@ cc_library( "@local_config_cuda//cuda:cuda_headers", "//tensorflow/core:lib", "//tensorflow/core:lib_internal", + "//tensorflow/core/platform:tf32_utils", "//tensorflow/stream_executor", "//tensorflow/stream_executor:event", "//tensorflow/stream_executor:host_or_device_scalar", diff --git a/tensorflow/stream_executor/cuda/cuda_blas.cc b/tensorflow/stream_executor/cuda/cuda_blas.cc index 65c07e72154..e2cbb0b75df 100644 --- a/tensorflow/stream_executor/cuda/cuda_blas.cc +++ b/tensorflow/stream_executor/cuda/cuda_blas.cc @@ -48,7 +48,7 @@ limitations under the License. #include "absl/strings/str_cat.h" #include "absl/strings/str_format.h" -#include "third_party/eigen3/Eigen/Core" +#include "tensorflow/core/platform/tf32_utils.h" #include "tensorflow/core/util/env_var.h" #include "tensorflow/stream_executor/cuda/cuda_activation.h" #include "tensorflow/stream_executor/cuda/cuda_gpu_executor.h" @@ -66,6 +66,7 @@ limitations under the License. #include "tensorflow/stream_executor/plugin_registry.h" #include "tensorflow/stream_executor/scratch_allocator.h" #include "tensorflow/stream_executor/stream_executor.h" +#include "third_party/eigen3/Eigen/Core" namespace stream_executor { namespace gpu { @@ -225,6 +226,18 @@ bool CUDABlas::Init() { return false; } +#if CUDA_VERSION >= 9000 +#if CUBLAS_VER_MAJOR >= 11 + ret = cublasSetMathMode(blas_, CUBLAS_TF32_TENSOR_OP_MATH); +#else + ret = cublasSetMathMode(blas_, CUBLAS_TENSOR_OP_MATH); +#endif + if (ret != CUBLAS_STATUS_SUCCESS) { + LOG(ERROR) << "failed to set cublas default math mode: " << ToString(ret); + return false; + } +#endif + return true; } @@ -387,7 +400,7 @@ cudaDataType_t CUDAComputationType(blas::ComputationType ty) { template bool CUDABlas::DoBlasInternalImpl(FuncT cublas_func, Stream *stream, bool pointer_mode_host, bool err_on_failure, - bool use_tensor_op_math, Args... args) { + Args... args) { absl::MutexLock lock(&mu_); CHECK(blas_ != nullptr); @@ -401,10 +414,10 @@ bool CUDABlas::DoBlasInternalImpl(FuncT cublas_func, Stream *stream, : CUBLAS_POINTER_MODE_DEVICE)) { return false; } -#if CUDA_VERSION >= 9000 +#if CUBLAS_VER_MAJOR >= 11 ScopedCublasMathMode math_mode{blas_}; - if (use_tensor_op_math) { - if (!math_mode.Init(CUBLAS_TENSOR_OP_MATH)) { + if (!tensorflow::tf32_execution_allowed()) { + if (!math_mode.Init(CUBLAS_DEFAULT_MATH)) { return false; } } @@ -1621,21 +1634,9 @@ bool CUDABlas::DoBlasGemm( } } - bool use_tensor_ops = false; -#if CUDA_VERSION >= 9000 - int cc_major, cc_minor; - stream->parent()->GetDeviceDescription().cuda_compute_capability(&cc_major, - &cc_minor); - - // GPUs < sm_70 don't support tensor ops. - if (cc_major >= 7) { - use_tensor_ops = true; - } -#endif - return DoBlasInternalImpl( cublasSgemmEx, stream, true /* = pointer_mode_host */, - true /* = err_on_failure= */, use_tensor_ops, CUDABlasTranspose(transa), + true /* = err_on_failure= */, CUDABlasTranspose(transa), CUDABlasTranspose(transb), m, n, k, &alpha, GpuMemory(a), SE_CUDA_DATA_HALF, lda, GpuMemory(b), SE_CUDA_DATA_HALF, ldb, &beta, GpuMemoryMutable(c), SE_CUDA_DATA_HALF, ldc); @@ -2257,7 +2258,8 @@ port::Status CUDABlas::DoBlasGemmBatchedInternal( if (stream->parent()->GetDeviceDescription().cuda_compute_capability( &cc_major, &cc_minor) && cc_major >= 5) { - bool use_tensor_ops = data_type == CUDA_R_16F; + bool use_tensor_ops = + data_type == CUDA_R_16F || tensorflow::tf32_execution_allowed(); cublasGemmAlgo_t algo = (use_tensor_ops ? CUBLAS_GEMM_DFALT_TENSOR_OP : CUBLAS_GEMM_DFALT); cudaDataType_t compute_type = @@ -2271,7 +2273,7 @@ port::Status CUDABlas::DoBlasGemmBatchedInternal( bool ok; ok = DoBlasInternalImpl( AS_LAMBDA(cublasGemmBatchedEx), stream, true /* = pointer_mode_host */, - true /* = err_on_failure */, use_tensor_ops, CUDABlasTranspose(transa), + true /* = err_on_failure */, CUDABlasTranspose(transa), CUDABlasTranspose(transb), m, n, k, &alpha, a_void_ptrs, data_type, lda, b_void_ptrs, data_type, ldb, &beta, c_void_ptrs, data_type, ldc, batch_count, compute_type, algo); @@ -2406,33 +2408,25 @@ bool CUDABlas::DoBlasGemmStridedBatched( int lda, int64 stride_a, const DeviceMemory &b, int ldb, int64 stride_b, float beta, DeviceMemory *c, int ldc, int64 stride_c, int batch_count) { - bool use_tensor_ops = false; -#if CUDA_VERSION >= 9000 +#if CUDA_VERSION >= 9010 int cc_major, cc_minor; if (stream->parent()->GetDeviceDescription().cuda_compute_capability( - &cc_major, &cc_minor)) { - // GPUs < sm_70 don't support tensor ops. - if (cc_major >= 7) { - use_tensor_ops = true; + &cc_major, &cc_minor) && + cc_major >= 5) { + cublasGemmAlgo_t algo = + (cc_major >= 7 ? CUBLAS_GEMM_DFALT_TENSOR_OP : CUBLAS_GEMM_DFALT); + bool ok = DoBlasInternalImpl( + AS_LAMBDA(cublasGemmStridedBatchedEx), stream, + true /* = pointer_mode_host */, true /* = err_on_failure */, + CUDABlasTranspose(transa), CUDABlasTranspose(transb), m, n, k, &alpha, + GpuMemory(a), CUDA_R_16F, lda, stride_a, GpuMemory(b), CUDA_R_16F, ldb, + stride_b, &beta, GpuMemoryMutable(c), CUDA_R_16F, ldc, stride_c, + batch_count, CUDA_R_32F, algo); + if (ok) { + return true; } -#if CUDA_VERSION >= 9010 - if (cc_major >= 5) { - cublasGemmAlgo_t algo = - (use_tensor_ops ? CUBLAS_GEMM_DFALT_TENSOR_OP : CUBLAS_GEMM_DFALT); - bool ok = DoBlasInternalImpl( - AS_LAMBDA(cublasGemmStridedBatchedEx), stream, - true /* = pointer_mode_host */, true /* = err_on_failure */, - use_tensor_ops, CUDABlasTranspose(transa), CUDABlasTranspose(transb), - m, n, k, &alpha, GpuMemory(a), CUDA_R_16F, lda, stride_a, - GpuMemory(b), CUDA_R_16F, ldb, stride_b, &beta, GpuMemoryMutable(c), - CUDA_R_16F, ldc, stride_c, batch_count, CUDA_R_32F, algo); - if (ok) { - return true; - } - LOG(ERROR) << "failed BLAS call, see log for details"; - return false; - } -#endif + LOG(ERROR) << "failed BLAS call, see log for details"; + return false; } #endif // Either CUDA_VERSION < 9.1 or SM < 5.0. Fall back to a loop. @@ -2445,7 +2439,7 @@ bool CUDABlas::DoBlasGemmStridedBatched( reinterpret_cast<__half *>(GpuMemoryMutable(c) + batch * stride_c); bool ok = DoBlasInternalImpl( cublasSgemmEx, stream, true /* = pointer_mode_host */, - true /* = err_on_failure= */, use_tensor_ops, CUDABlasTranspose(transa), + true /* = err_on_failure= */, CUDABlasTranspose(transa), CUDABlasTranspose(transb), m, n, k, &alpha, a_matrix, SE_CUDA_DATA_HALF, lda, b_matrix, SE_CUDA_DATA_HALF, ldb, &beta, c_matrix, SE_CUDA_DATA_HALF, ldc); diff --git a/tensorflow/stream_executor/cuda/cuda_blas.h b/tensorflow/stream_executor/cuda/cuda_blas.h index 817bdb72777..556456c83db 100644 --- a/tensorflow/stream_executor/cuda/cuda_blas.h +++ b/tensorflow/stream_executor/cuda/cuda_blas.h @@ -83,7 +83,7 @@ class CUDABlas : public blas::BlasSupport { template bool DoBlasInternalImpl(FuncT cublas_func, Stream *stream, bool pointer_mode_host, bool err_on_failure, - bool use_tensor_op_math, Args... args); + Args... args); // Convenience functions that call DoBlasInternalImpl with different values // for err_on_failure. @@ -91,8 +91,7 @@ class CUDABlas : public blas::BlasSupport { bool DoBlasInternal(FuncT cublas_func, Stream *stream, bool pointer_mode_host, Args... args) { return DoBlasInternalImpl(cublas_func, stream, pointer_mode_host, - /*err_on_failure=*/true, /*use_tensor_ops=*/false, - args...); + /*err_on_failure=*/true, args...); } template bool DoBlasInternalFailureOK(FuncT cublas_func, Stream *stream, @@ -100,8 +99,7 @@ class CUDABlas : public blas::BlasSupport { // Tensor ops are hard-coded off in this path, but can still be enabled with // a specific algorithm choice as in DoBlasGemmWithAlgorithmImpl(). return DoBlasInternalImpl(cublas_func, stream, pointer_mode_host, - /*err_on_failure=*/false, - /*use_tensor_ops=*/false, args...); + /*err_on_failure=*/false, args...); } // A helper function to implement DoBlasGemmBatched interfaces for generic From 521b7595b7adba5627a2687befc3fb41bc5c2bec Mon Sep 17 00:00:00 2001 From: "902449@58880@bigcat_chen@ASIC" Date: Thu, 4 Jun 2020 17:19:31 +0800 Subject: [PATCH 044/867] TFLM: add HIMAX WE1 EVB to support TFLM example(hello word and person detection INT8) --- .../lite/micro/examples/hello_world/README.md | 132 +++- .../micro/examples/hello_world/README.md~ | 595 ++++++++++++++++++ .../hello_world/himax_we1_evb/constants.cc | 19 + .../himax_we1_evb/output_handler.cc | 35 ++ .../himax_we1_evb/output_handler.cc~ | 53 ++ .../person_detection_experimental/README.md | 129 +++- .../himax_we1_evb/detection_responder.cc | 34 + .../himax_we1_evb/image_provider.cc | 44 ++ .../himax_we1_evb/image_provider.cc~ | 44 ++ .../himax_we1_evb/main_functions.cc | 127 ++++ .../{we_i => himax_we1_evb}/debug_log.cc | 9 +- .../make/targets/himax_we1_evb_makefile.inc | 91 +++ .../make/targets/himax_we1_evb_makefile.inc~ | 93 +++ .../tools/make/third_party_downloads.inc | 4 + .../tools/make/third_party_downloads.inc~ | 86 +++ 15 files changed, 1476 insertions(+), 19 deletions(-) create mode 100644 tensorflow/lite/micro/examples/hello_world/README.md~ create mode 100644 tensorflow/lite/micro/examples/hello_world/himax_we1_evb/constants.cc create mode 100644 tensorflow/lite/micro/examples/hello_world/himax_we1_evb/output_handler.cc create mode 100644 tensorflow/lite/micro/examples/hello_world/himax_we1_evb/output_handler.cc~ create mode 100644 tensorflow/lite/micro/examples/person_detection_experimental/himax_we1_evb/detection_responder.cc create mode 100644 tensorflow/lite/micro/examples/person_detection_experimental/himax_we1_evb/image_provider.cc create mode 100644 tensorflow/lite/micro/examples/person_detection_experimental/himax_we1_evb/image_provider.cc~ create mode 100644 tensorflow/lite/micro/examples/person_detection_experimental/himax_we1_evb/main_functions.cc rename tensorflow/lite/micro/{we_i => himax_we1_evb}/debug_log.cc (90%) create mode 100644 tensorflow/lite/micro/tools/make/targets/himax_we1_evb_makefile.inc create mode 100644 tensorflow/lite/micro/tools/make/targets/himax_we1_evb_makefile.inc~ create mode 100644 tensorflow/lite/micro/tools/make/third_party_downloads.inc~ diff --git a/tensorflow/lite/micro/examples/hello_world/README.md b/tensorflow/lite/micro/examples/hello_world/README.md index 3b633890306..9c0a5e2306a 100644 --- a/tensorflow/lite/micro/examples/hello_world/README.md +++ b/tensorflow/lite/micro/examples/hello_world/README.md @@ -14,13 +14,34 @@ of the device. ## Table of contents -- [Deploy to ARC EM SDP](#deploy-to-arc-em-sdp) -- [Deploy to Arduino](#deploy-to-arduino) -- [Deploy to ESP32](#deploy-to-esp32) -- [Deploy to SparkFun Edge](#deploy-to-sparkfun-edge) -- [Deploy to STM32F746](#deploy-to-STM32F746) -- [Run the tests on a development machine](#run-the-tests-on-a-development-machine) -- [Train your own model](#train-your-own-model) +- [Hello World Example](#hello-world-example) + - [Table of contents](#table-of-contents) + - [Deploy to ARC EM SDP](#deploy-to-arc-em-sdp) + - [Initial Setup](#initial-setup) + - [Generate Example Project](#generate-example-project) + - [Build and Run Example](#build-and-run-example) + - [Deploy to Arduino](#deploy-to-arduino) + - [Install the Arduino_TensorFlowLite library](#install-the-arduinotensorflowlite-library) + - [Load and run the example](#load-and-run-the-example) + - [Deploy to ESP32](#deploy-to-esp32) + - [Install the ESP IDF](#install-the-esp-idf) + - [Generate the examples](#generate-the-examples) + - [Building the example](#building-the-example) + - [Load and run the example](#load-and-run-the-example-1) + - [Deploy to himax WE1 EVB](#deploy-to-himax-we1-evb) + - [Initial Setup](#initial-setup-1) + - [MetaWare Development Toolkit](#metaware-development-toolkit) + - [Make Tool version](#make-tool-version) + - [Serial Terminal Emulation Application](#serial-terminal-emulation-application) + - [Generate Example Project](#generate-example-project-1) + - [Build and Burn Example](#build-and-burn-example) + - [Deploy to SparkFun Edge](#deploy-to-sparkfun-edge) + - [Compile the binary](#compile-the-binary) + - [Sign the binary](#sign-the-binary) + - [Flash the binary](#flash-the-binary) + - [Deploy to STM32F746](#deploy-to-stm32f746) + - [Run the tests on a development machine](#run-the-tests-on-a-development-machine) + - [Train your own model](#train-your-own-model) ## Deploy to ARC EM SDP @@ -191,6 +212,103 @@ The previous two commands can be combined: idf.py --port /dev/ttyUSB0 flash monitor ``` +## Deploy to himax WE1 EVB + +The following instructions will help you build and deploy this example to +[HIMAX WE1 EVB](https://github.com/HimaxWiseEyePlus/bsp_tflu/tree/master/HIMAX_WE1_EVB_board_brief) +board. To undstand more about using this board, please check +[HIMAX WE1 EVB user guide](https://github.com/HimaxWiseEyePlus/bsp_tflu/tree/master/HIMAX_WE1_EVB_user_guide). + +### Initial Setup + +To use the HIMAX WE1 EVB, please make sure following software are installed: + +#### MetaWare Development Toolkit + +See +[Install the Synopsys DesignWare ARC MetaWare Development Toolkit](/tensorflow/lite/micro/tools/make/targets/arc/README.md#install-the-synopsys-designware-arc-metaware-development-toolkit) +section for instructions on toolchain installation. + +#### Make Tool version + +A `'make'` tool is required for deploying Tensorflow Lite Micro +applications on HIMAX WE1 EVB, See +[Check make tool version](/tensorflow/lite/micro/tools/make/targets/arc/README.md#make-tool) +section for proper environment. + +#### Serial Terminal Emulation Application + +There are 2 main purposes for HIMAX WE1 EVB Debug UART port + +- print application output +- burn application to flash by using xmodem send application binary + +You can use any terminal emulation program (like [PuTTY](https://www.putty.org/) or [minicom](https://linux.die.net/man/1/minicom)). + + +### Generate Example Project + +The example project for HIMAX WE1 EVB platform can be generated with the following +command: + +Download related third party data + +``` +make -f tensorflow/lite/micro/tools/make/Makefile TARGET=himax_we1_evb third_party_downloads +``` + +Generate hello world project + +``` +make -f tensorflow/lite/micro/tools/make/Makefile generate_hello_world_make_project TARGET=himax_we1_evb TAGS=no_arc_mli +``` + +### Build and Burn Example + +Following the Steps to run hello world example at HIMAX WE1 EVB platform. + +1. Go to the generated example project directory. + + ``` + cd tensorflow/lite/micro/tools/make/gen/himax_we1_evb_arc/prj/hello_world/make + ``` + +2. Build the example using + + ``` + make app + ``` + +3. After example build finish, copy ELF file and map file to image generate tool directory. + image generate tool directory located at `'tensorflow/lite/micro/tools/make/downloads/himax_we1_sdk/image_gen_linux_v3/'` + + ``` + cp hello_world.elf himax_we1_evb.map ../../../../../downloads/himax_we1_sdk/image_gen_linux_v3/ + ``` + +4. Go to flash image generate tool directory. + + ``` + cd ../../../../../downloads/himax_we1_sdk/image_gen_linux_v3/ + ``` + +5. run image generate tool, generate flash image file. + + * Before running image generate tool, by typing `sudo chmod +x image_gen` + and `sudo chmod +x sign_tool` to make sure it is executable. + + ``` + image_gen -e hello_world.elf -m himax_we1_evb.map -o out.img + ``` + + +6. Download flash image file to HIMAX WE1 EVB by UART: + + * more detail about download image through UART can be found at [HIMAX WE1 EVB update Flash image](https://github.com/HimaxWiseEyePlus/bsp_tflu/tree/master/HIMAX_WE1_EVB_user_guide#flash-image-update) + +After these steps, press reset button on the HIMAX WE1 EVB, you will see application output in the serial +terminal. + ## Deploy to SparkFun Edge The following instructions will help you build and deploy this sample on the diff --git a/tensorflow/lite/micro/examples/hello_world/README.md~ b/tensorflow/lite/micro/examples/hello_world/README.md~ new file mode 100644 index 00000000000..011711493d5 --- /dev/null +++ b/tensorflow/lite/micro/examples/hello_world/README.md~ @@ -0,0 +1,595 @@ +# Hello World Example + +This example is designed to demonstrate the absolute basics of using [TensorFlow +Lite for Microcontrollers](https://www.tensorflow.org/lite/microcontrollers). +It includes the full end-to-end workflow of training a model, converting it for +use with TensorFlow Lite for Microcontrollers for running inference on a +microcontroller. + +The model is trained to replicate a `sine` function and generates a pattern of +data to either blink LEDs or control an animation, depending on the capabilities +of the device. + +![Animation on STM32F746](images/animation_on_STM32F746.gif) + +## Table of contents + +- [Hello World Example](#hello-world-example) + - [Table of contents](#table-of-contents) + - [Deploy to ARC EM SDP](#deploy-to-arc-em-sdp) + - [Initial Setup](#initial-setup) + - [Generate Example Project](#generate-example-project) + - [Build and Run Example](#build-and-run-example) + - [Deploy to Arduino](#deploy-to-arduino) + - [Install the Arduino_TensorFlowLite library](#install-the-arduinotensorflowlite-library) + - [Load and run the example](#load-and-run-the-example) + - [Deploy to ESP32](#deploy-to-esp32) + - [Install the ESP IDF](#install-the-esp-idf) + - [Generate the examples](#generate-the-examples) + - [Building the example](#building-the-example) + - [Load and run the example](#load-and-run-the-example-1) + - [Deploy to himax WE1 EVB](#deploy-to-himax-we1-evb) + - [Initial Setup](#initial-setup-1) + - [MetaWare Development Toolkit](#metaware-development-toolkit) + - [Make Tool version](#make-tool-version) + - [Serial Terminal Emulation Application](#serial-terminal-emulation-application) + - [Generate Example Project](#generate-example-project-1) + - [Build and Burn Example](#build-and-burn-example) + - [Deploy to SparkFun Edge](#deploy-to-sparkfun-edge) + - [Compile the binary](#compile-the-binary) + - [Sign the binary](#sign-the-binary) + - [Flash the binary](#flash-the-binary) + - [Deploy to STM32F746](#deploy-to-stm32f746) + - [Run the tests on a development machine](#run-the-tests-on-a-development-machine) + - [Train your own model](#train-your-own-model) + +## Deploy to ARC EM SDP + +The following instructions will help you to build and deploy this example to +[ARC EM SDP](https://www.synopsys.com/dw/ipdir.php?ds=arc-em-software-development-platform) +board. General information and instructions on using the board with TensorFlow +Lite Micro can be found in the common +[ARC targets description](/tensorflow/lite/micro/tools/make/targets/arc/README.md). + +### Initial Setup + +Follow the instructions on the +[ARC EM SDP Initial Setup](/tensorflow/lite/micro/tools/make/targets/arc/README.md#ARC-EM-Software-Development-Platform-ARC-EM-SDP) +to get and install all required tools for work with ARC EM SDP. + +### Generate Example Project + +The example project for ARC EM SDP platform can be generated with the following +command: + +``` +make -f tensorflow/lite/micro/tools/make/Makefile TARGET=arc_emsdp TAGS=no_arc_mli generate_hello_world_make_project +``` + +### Build and Run Example + +For more detailed information on building and running examples see the +appropriate sections of general descriptions of the +[ARC EM SDP usage with TFLM](/tensorflow/lite/micro/tools/make/targets/arc/README.md#ARC-EM-Software-Development-Platform-ARC-EM-SDP). +In the directory with generated project you can also find a +*README_ARC_EMSDP.md* file with instructions and options on building and +running. Here we only briefly mention main steps which are typically enough to +get it started. + +1. You need to + [connect the board](/tensorflow/lite/micro/tools/make/targets/arc/README.md#connect-the-board) + and open an serial connection. + +2. Go to the generated example project director + + ``` + cd tensorflow/lite/micro/tools/make/gen/arc_emsdp_arc/prj/hello_world/make + ``` + +3. Build the example using + + ``` + make app + ``` + +4. To generate artefacts for self-boot of example from the board use + + ``` + make flash + ``` + +5. To run application from the board using microSD card: + + * Copy the content of the created /bin folder into the root of microSD + card. Note that the card must be formatted as FAT32 with default cluster + size (but less than 32 Kbytes) + * Plug in the microSD card into the J11 connector. + * Push the RST button. If a red LED is lit beside RST button, push the CFG + button. + +6. If you have the MetaWare Debugger installed in your environment: + + * To run application from the console using it type `make run`. + * To stop the execution type `Ctrl+C` in the console several times. + +In both cases (step 5 and 6) you will see the application output in the serial +terminal. + +## Deploy to Arduino + +The following instructions will help you build and deploy this sample +to [Arduino](https://www.arduino.cc/) devices. + +![Animation on Arduino MKRZERO](images/animation_on_arduino_mkrzero.gif) + +The sample has been tested with the following devices: + +- [Arduino Nano 33 BLE Sense](https://store.arduino.cc/usa/nano-33-ble-sense-with-headers) +- [Arduino MKRZERO](https://store.arduino.cc/usa/arduino-mkrzero) + +The sample will use PWM to fade an LED on and off according to the model's +output. In the code, the `LED_BUILTIN` constant is used to specify the board's +built-in LED as the one being controlled. However, on some boards, this built-in +LED is not attached to a pin with PWM capabilities. In this case, the LED will +blink instead of fading. + +### Install the Arduino_TensorFlowLite library + +This example application is included as part of the official TensorFlow Lite +Arduino library. To install it, open the Arduino library manager in +`Tools -> Manage Libraries...` and search for `Arduino_TensorFlowLite`. + +### Load and run the example + +Once the library has been added, go to `File -> Examples`. You should see an +example near the bottom of the list named `TensorFlowLite:hello_world`. Select +it and click `hello_world` to load the example. + +Use the Arduino IDE to build and upload the example. Once it is running, +you should see the built-in LED on your device flashing. + +The Arduino Desktop IDE includes a plotter that we can use to display the sine +wave graphically. To view it, go to `Tools -> Serial Plotter`. You will see one +datapoint being logged for each inference cycle, expressed as a number between 0 +and 255. + +## Deploy to ESP32 + +The following instructions will help you build and deploy this sample +to [ESP32](https://www.espressif.com/en/products/hardware/esp32/overview) +devices using the [ESP IDF](https://github.com/espressif/esp-idf). + +The sample has been tested on ESP-IDF version 4.0 with the following devices: +- [ESP32-DevKitC](http://esp-idf.readthedocs.io/en/latest/get-started/get-started-devkitc.html) +- [ESP-EYE](https://github.com/espressif/esp-who/blob/master/docs/en/get-started/ESP-EYE_Getting_Started_Guide.md) + +### Install the ESP IDF + +Follow the instructions of the +[ESP-IDF get started guide](https://docs.espressif.com/projects/esp-idf/en/latest/get-started/index.html) +to setup the toolchain and the ESP-IDF itself. + +The next steps assume that the +[IDF environment variables are set](https://docs.espressif.com/projects/esp-idf/en/latest/get-started/index.html#step-4-set-up-the-environment-variables) : + + * The `IDF_PATH` environment variable is set + * `idf.py` and Xtensa-esp32 tools (e.g. `xtensa-esp32-elf-gcc`) are in `$PATH` + +### Generate the examples +The example project can be generated with the following command: +``` +make -f tensorflow/lite/micro/tools/make/Makefile TARGET=esp generate_hello_world_esp_project +``` + +### Building the example + +Go the the example project directory +``` +cd tensorflow/lite/micro/tools/make/gen/esp_xtensa-esp32/prj/hello_world/esp-idf +``` + +Then build with `idf.py` +``` +idf.py build +``` + +### Load and run the example + +To flash (replace `/dev/ttyUSB0` with the device serial port): +``` +idf.py --port /dev/ttyUSB0 flash +``` + +Monitor the serial output: +``` +idf.py --port /dev/ttyUSB0 monitor +``` + +Use `Ctrl+]` to exit. + +The previous two commands can be combined: +``` +idf.py --port /dev/ttyUSB0 flash monitor +``` + +## Deploy to himax WE1 EVB + +The following instructions will help you build and deploy this example to +[HIMAX WE1 EVB](https://github.com/HimaxWiseEyePlus/bsp_tflu/tree/master/HIMAX_WE1_EVB_board_brief) +board. To undstand more about using this board, please check +[HIMAX WE1 EVB user guide](https://github.com/HimaxWiseEyePlus/bsp_tflu/tree/master/HIMAX_WE1_EVB_user_guide). + +### Initial Setup + +To use the HIMAX WE1 EVB, please make sure following software are installed: + +#### MetaWare Development Toolkit + +See +[Install the Synopsys DesignWare ARC MetaWare Development Toolkit](/tensorflow/lite/micro/tools/make/targets/arc/README.md#install-the-synopsys-designware-arc-metaware-development-toolkit) +section for instructions on toolchain installation. + +#### Make Tool version + +A `'make'` tool is required for deploying Tensorflow Lite Micro +applications on HIMAX WE1 EVB, See +[Check make tool version](/tensorflow/lite/micro/tools/make/targets/arc/README.md#make-tool) +section for proper environment. + +#### Serial Terminal Emulation Application + +There are 2 main purposes for HIMAX WE1 EVB Debug UART port + +- print application output +- burn application to flash by using xmodem send application binary + +You can use any terminal emulation program (like [PuTTY](https://www.putty.org/) or [minicom](https://linux.die.net/man/1/minicom)). + + +### Generate Example Project + +The example project for HIMAX WE1 EVB platform can be generated with the following +command: + +Download related third party data + +``` +make -f tensorflow/lite/micro/tools/make/Makefile TARGET=himax_we1_evb third_party_downloads +``` + +Generate hello world project + +``` +make -f tensorflow/lite/micro/tools/make/Makefile generate_hello_world_make_project TARGET=himax_we1_evb +``` + +### Build and Burn Example + +Following the Steps to run hello world example at HIMAX WE1 EVB platform. + +1. Go to the generated example project directory. + + ``` + cd tensorflow/lite/micro/tools/make/gen/himax_we1_evb_arc/prj/hello_world/make + ``` + +2. Build the example using + + ``` + make app + ``` + +3. After example build finish, copy ELF file and map file to image generate tool directory. + image generate tool directory located at `'tensorflow/lite/micro/tools/make/downloads/himax_we1_sdk/image_gen_linux_v3/'` + + ``` + cp hello_world.elf himax_we1_evb.map ../../../../../downloads/himax_we1_sdk/image_gen_linux_v3/ + ``` + +4. Go to flash image generate tool directory. + + ``` + cd ../../../../../downloads/himax_we1_sdk/image_gen_linux_v3/ + ``` + +5. run image generate tool, generate flash image file. + + * Before running image generate tool, by typing `sudo chmod +x image_gen` + and `sudo chmod +x sign_tool` to make sure it is executable. + + ``` + image_gen -e hello_world.elf -m himax_we1_evb.map -o out.img + ``` + + +6. Download flash image file to HIMAX WE1 EVB by UART: + + * more detail about download image through UART can be found at [HIMAX WE1 EVB update Flash image](https://github.com/HimaxWiseEyePlus/bsp_tflu/tree/master/HIMAX_WE1_EVB_user_guide#flash-image-update) + +After these steps, press reset button on the HIMAX WE1 EVB, you will see application output in the serial +terminal. + +## Deploy to SparkFun Edge + +The following instructions will help you build and deploy this sample on the +[SparkFun Edge development board](https://sparkfun.com/products/15170). + +![Animation on SparkFun Edge](images/animation_on_sparkfun_edge.gif) + +If you're new to using this board, we recommend walking through the +[AI on a microcontroller with TensorFlow Lite and SparkFun Edge](https://codelabs.developers.google.com/codelabs/sparkfun-tensorflow) +codelab to get an understanding of the workflow. + +### Compile the binary + +The following command will download the required dependencies and then compile a +binary for the SparkFun Edge: + +``` +make -f tensorflow/lite/micro/tools/make/Makefile TARGET=sparkfun_edge hello_world_bin +``` + +The binary will be created in the following location: + +``` +tensorflow/lite/micro/tools/make/gen/sparkfun_edge_cortex-m4/bin/hello_world.bin +``` + +### Sign the binary + +The binary must be signed with cryptographic keys to be deployed to the device. +We'll now run some commands that will sign our binary so it can be flashed to +the SparkFun Edge. The scripts we are using come from the Ambiq SDK, which is +downloaded when the `Makefile` is run. + +Enter the following command to set up some dummy cryptographic keys we can use +for development: + +``` +cp tensorflow/lite/micro/tools/make/downloads/AmbiqSuite-Rel2.2.0/tools/apollo3_scripts/keys_info0.py \ +tensorflow/lite/micro/tools/make/downloads/AmbiqSuite-Rel2.2.0/tools/apollo3_scripts/keys_info.py +``` + +Next, run the following command to create a signed binary: + +``` +python3 tensorflow/lite/micro/tools/make/downloads/AmbiqSuite-Rel2.2.0/tools/apollo3_scripts/create_cust_image_blob.py \ +--bin tensorflow/lite/micro/tools/make/gen/sparkfun_edge_cortex-m4/bin/hello_world.bin \ +--load-address 0xC000 \ +--magic-num 0xCB \ +-o main_nonsecure_ota \ +--version 0x0 +``` + +This will create the file `main_nonsecure_ota.bin`. We'll now run another +command to create a final version of the file that can be used to flash our +device with the bootloader script we will use in the next step: + +``` +python3 tensorflow/lite/micro/tools/make/downloads/AmbiqSuite-Rel2.2.0/tools/apollo3_scripts/create_cust_wireupdate_blob.py \ +--load-address 0x20000 \ +--bin main_nonsecure_ota.bin \ +-i 6 \ +-o main_nonsecure_wire \ +--options 0x1 +``` + +You should now have a file called `main_nonsecure_wire.bin` in the directory +where you ran the commands. This is the file we'll be flashing to the device. + +### Flash the binary + +Next, attach the board to your computer via a USB-to-serial adapter. + +**Note:** If you're using the [SparkFun Serial Basic Breakout](https://www.sparkfun.com/products/15096), +you should [install the latest drivers](https://learn.sparkfun.com/tutorials/sparkfun-serial-basic-ch340c-hookup-guide#drivers-if-you-need-them) +before you continue. + +Once connected, assign the USB device name to an environment variable: + +``` +export DEVICENAME=put your device name here +``` + +Set another variable with the baud rate: + +``` +export BAUD_RATE=921600 +``` + +Now, hold the button marked `14` on the device. While still holding the button, +hit the button marked `RST`. Continue holding the button marked `14` while +running the following command: + +``` +python3 tensorflow/lite/micro/tools/make/downloads/AmbiqSuite-Rel2.2.0/tools/apollo3_scripts/uart_wired_update.py \ +-b ${BAUD_RATE} ${DEVICENAME} \ +-r 1 \ +-f main_nonsecure_wire.bin \ +-i 6 +``` + +You should see a long stream of output as the binary is flashed to the device. +Once you see the following lines, flashing is complete: + +``` +Sending Reset Command. +Done. +``` + +If you don't see these lines, flashing may have failed. Try running through the +steps in [Flash the binary](#flash-the-binary) again (you can skip over setting +the environment variables). If you continue to run into problems, follow the +[AI on a microcontroller with TensorFlow Lite and SparkFun Edge](https://codelabs.developers.google.com/codelabs/sparkfun-tensorflow) +codelab, which includes more comprehensive instructions for the flashing +process. + +The binary should now be deployed to the device. Hit the button marked `RST` to +reboot the board. You should see the device's four LEDs flashing in sequence. + +Debug information is logged by the board while the program is running. To view +it, establish a serial connection to the board using a baud rate of `115200`. +On OSX and Linux, the following command should work: + +``` +screen ${DEVICENAME} 115200 +``` + +You will see a lot of output flying past! To stop the scrolling, hit `Ctrl+A`, +immediately followed by `Esc`. You can then use the arrow keys to explore the +output, which will contain the results of running inference on various `x` +values: + +``` +x_value: 1.1843798*2^2, y_value: -1.9542645*2^-1 +``` + +To stop viewing the debug output with `screen`, hit `Ctrl+A`, immediately +followed by the `K` key, then hit the `Y` key. + + +## Deploy to STM32F746 + +The following instructions will help you build and deploy the sample to the +[STM32F7 discovery kit](https://os.mbed.com/platforms/ST-Discovery-F746NG/) +using [ARM Mbed](https://github.com/ARMmbed/mbed-cli). + +![Animation on STM32F746](images/animation_on_STM32F746.gif) + +Before we begin, you'll need the following: + +- STM32F7 discovery kit board +- Mini-USB cable +- ARM Mbed CLI ([installation instructions](https://os.mbed.com/docs/mbed-os/v5.12/tools/installation-and-setup.html)) +- Python 2.7 and pip + +Since Mbed requires a special folder structure for projects, we'll first run a +command to generate a subfolder containing the required source files in this +structure: + +``` +make -f tensorflow/lite/micro/tools/make/Makefile TARGET=mbed TAGS="CMSIS disco_f746ng" generate_hello_world_mbed_project +``` + +This will result in the creation of a new folder: + +``` +tensorflow/lite/micro/tools/make/gen/mbed_cortex-m4/prj/hello_world/mbed +``` + +This folder contains all of the example's dependencies structured in the correct +way for Mbed to be able to build it. + +Change into the directory and run the following commands, making sure you are +using Python 2.7.15. + +First, tell Mbed that the current directory is the root of an Mbed project: + +``` +mbed config root . +``` + +Next, tell Mbed to download the dependencies and prepare to build: + +``` +mbed deploy +``` + +By default, Mbed will build the project using C++98. However, TensorFlow Lite +requires C++11. Run the following Python snippet to modify the Mbed +configuration files so that it uses C++11: + +``` +python -c 'import fileinput, glob; +for filename in glob.glob("mbed-os/tools/profiles/*.json"): + for line in fileinput.input(filename, inplace=True): + print line.replace("\"-std=gnu++98\"","\"-std=c++11\", \"-fpermissive\"")' + +``` + +Finally, run the following command to compile: + +``` +mbed compile -m DISCO_F746NG -t GCC_ARM +``` + +This should result in a binary at the following path: + +``` +./BUILD/DISCO_F746NG/GCC_ARM/mbed.bin +``` + +To deploy, plug in your STM board and copy the file to it. On MacOS, you can do +this with the following command: + +``` +cp ./BUILD/DISCO_F746NG/GCC_ARM/mbed.bin /Volumes/DIS_F746NG/ +``` + +Copying the file will initiate the flashing process. Once this is complete, you +should see an animation on the device's screen. + + +``` +screen /dev/tty.usbmodem14403 9600 +``` + +In addition to this animation, debug information is logged by the board while +the program is running. To view it, establish a serial connection to the board +using a baud rate of `9600`. On OSX and Linux, the following command should +work, replacing `/dev/tty.devicename` with the name of your device as it appears +in `/dev`: + +``` +screen /dev/tty.devicename 9600 +``` + +You will see a lot of output flying past! To stop the scrolling, hit `Ctrl+A`, +immediately followed by `Esc`. You can then use the arrow keys to explore the +output, which will contain the results of running inference on various `x` +values: + +``` +x_value: 1.1843798*2^2, y_value: -1.9542645*2^-1 +``` + +To stop viewing the debug output with `screen`, hit `Ctrl+A`, immediately +followed by the `K` key, then hit the `Y` key. + +### Run the tests on a development machine + +To compile and test this example on a desktop Linux or macOS machine, first +clone the TensorFlow repository from GitHub to a convenient place: + +```bash +git clone --depth 1 https://github.com/tensorflow/tensorflow.git +``` + +Next, `cd` into the source directory from a terminal, and then run the following +command: + +```bash +make -f tensorflow/lite/micro/tools/make/Makefile test_hello_world_test +``` + +This will take a few minutes, and downloads frameworks the code uses. Once the +process has finished, you should see a series of files get compiled, followed by +some logging output from a test, which should conclude with +`~~~ALL TESTS PASSED~~~`. + +If you see this, it means that a small program has been built and run that loads +the trained TensorFlow model, runs some example inputs through it, and got the +expected outputs. + +To understand how TensorFlow Lite does this, you can look at the source in +[hello_world_test.cc](hello_world_test.cc). +It's a fairly small amount of code that creates an interpreter, gets a handle to +a model that's been compiled into the program, and then invokes the interpreter +with the model and sample inputs. + +### Train your own model + +So far you have used an existing trained model to run inference on +microcontrollers. If you wish to train your own model, follow the instructions +given in the [train/](train/) directory. + diff --git a/tensorflow/lite/micro/examples/hello_world/himax_we1_evb/constants.cc b/tensorflow/lite/micro/examples/hello_world/himax_we1_evb/constants.cc new file mode 100644 index 00000000000..1816a2f3207 --- /dev/null +++ b/tensorflow/lite/micro/examples/hello_world/himax_we1_evb/constants.cc @@ -0,0 +1,19 @@ +/* Copyright 2019 The TensorFlow Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. +==============================================================================*/ + +#include "tensorflow/lite/micro/examples/hello_world/constants.h" + +// This is tuned so that a full cycle takes ~4 seconds on a SparkFun Edge. +const int kInferencesPerCycle = 1000; diff --git a/tensorflow/lite/micro/examples/hello_world/himax_we1_evb/output_handler.cc b/tensorflow/lite/micro/examples/hello_world/himax_we1_evb/output_handler.cc new file mode 100644 index 00000000000..8ca028acc55 --- /dev/null +++ b/tensorflow/lite/micro/examples/hello_world/himax_we1_evb/output_handler.cc @@ -0,0 +1,35 @@ +/* Copyright 2019 The TensorFlow Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. +==============================================================================*/ + +#include "tensorflow/lite/micro/examples/hello_world/output_handler.h" + + +/* +This function trigger different device's LEDthrough y value. +y value range -1 <= y <= 1. +| Range is from -1~1 | LEDs | +| 0 <= y <= 1 | [ 0 1 ] | +| -1 <= y < 0 | [ 1 0 ] | + +*/ +void HandleOutput(tflite::ErrorReporter* error_reporter, float x_value, + float y_value) { + // The first time this method runs, set up our LEDs correctly + + // Log the current X and Y values + TF_LITE_REPORT_ERROR(error_reporter, "x_value: %f, y_value: %f\n", + static_cast(x_value), + static_cast(y_value)); +} diff --git a/tensorflow/lite/micro/examples/hello_world/himax_we1_evb/output_handler.cc~ b/tensorflow/lite/micro/examples/hello_world/himax_we1_evb/output_handler.cc~ new file mode 100644 index 00000000000..b59242d0b6f --- /dev/null +++ b/tensorflow/lite/micro/examples/hello_world/himax_we1_evb/output_handler.cc~ @@ -0,0 +1,53 @@ +/* Copyright 2019 The TensorFlow Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. +==============================================================================*/ + +#include "tensorflow/lite/micro/examples/hello_world/output_handler.h" + + +/* +This function trigger different device's LEDthrough y value. +y value range -1 <= y <= 1. +| Range is from -1~1 | LEDs | +| 0 <= y <= 1 | [ 0 1 ] | +| -1 <= y < 0 | [ 1 0 ] | + +*/ +void HandleOutput(tflite::ErrorReporter* error_reporter, float x_value, + float y_value) { + // The first time this method runs, set up our LEDs correctly +/* static bool is_initialized = false; + if (!is_initialized) { + // TODO Setup LED's as outputs + + // end of setup + is_initialized = true; + } + + // Set the LEDs to represent negative values + if (y_value < 0) { + //enable LED1 + + //enable LED0 + } else if (y_value > 0) { + //enable LED0 + + //enable LED1 + } + */ + // Log the current X and Y values + TF_LITE_REPORT_ERROR(error_reporter, "x_value: %f, y_value: %f\n", + static_cast(x_value), + static_cast(y_value)); +} diff --git a/tensorflow/lite/micro/examples/person_detection_experimental/README.md b/tensorflow/lite/micro/examples/person_detection_experimental/README.md index bf99b40d776..4d53e551431 100644 --- a/tensorflow/lite/micro/examples/person_detection_experimental/README.md +++ b/tensorflow/lite/micro/examples/person_detection_experimental/README.md @@ -7,13 +7,31 @@ This uses the experimental int8 quantized version of the person detection model. ## Table of contents -- [Getting started](#getting-started) -- [Running on ARC EM SDP](#running-on-arc-em-sdp) -- [Running on Arduino](#running-on-arduino) -- [Running on SparkFun Edge](#running-on-sparkfun-edge) -- [Run the tests on a development machine](#run-the-tests-on-a-development-machine) -- [Debugging image capture](#debugging-image-capture) -- [Training your own model](#training-your-own-model) +- [Person detection example](#person-detection-example) + - [Table of contents](#table-of-contents) + - [Running on ARC EM SDP](#running-on-arc-em-sdp) + - [Initial setup](#initial-setup) + - [Generate Example Project](#generate-example-project) + - [Build and Run Example](#build-and-run-example) + - [Running on Arduino](#running-on-arduino) + - [Hardware](#hardware) + - [Install the Arduino_TensorFlowLite library](#install-the-arduinotensorflowlite-library) + - [Install other libraries](#install-other-libraries) + - [Load and run the example](#load-and-run-the-example) + - [Running on HIMAX WE1 EVB](#running-on-himax-we1-evb) + - [Initial Setup](#initial-setup) + - [MetaWare Development Toolkit](#metaware-development-toolkit) + - [Make Tool version](#make-tool-version) + - [Serial Terminal Emulation Application](#serial-terminal-emulation-application) + - [Generate Example Project](#generate-example-project-1) + - [Build and Burn Example](#build-and-burn-example) + - [Running on SparkFun Edge](#running-on-sparkfun-edge) + - [Compile the binary](#compile-the-binary) + - [Sign the binary](#sign-the-binary) + - [Flash the binary](#flash-the-binary) + - [Run the tests on a development machine](#run-the-tests-on-a-development-machine) + - [Debugging image capture](#debugging-image-capture) + - [Training your own model](#training-your-own-model) ## Running on ARC EM SDP @@ -260,6 +278,103 @@ From the log, we can see that it took around 170 ms to capture and read the image data from the camera module, 180 ms to decode the JPEG and convert it to greyscale, and 18.6 seconds to run inference. +## Running on HIMAX WE1 EVB + +The following instructions will help you build and deploy this example to +[HIMAX WE1 EVB](https://github.com/HimaxWiseEyePlus/bsp_tflu/tree/master/HIMAX_WE1_EVB_board_brief) +board. To undstand more about using this board, please check +[HIMAX WE1 EVB user guide](https://github.com/HimaxWiseEyePlus/bsp_tflu/tree/master/HIMAX_WE1_EVB_user_guide). + +### Initial Setup + +To use the HIMAX WE1 EVB, please make sure following software are installed: + +#### MetaWare Development Toolkit + +See +[Install the Synopsys DesignWare ARC MetaWare Development Toolkit](/tensorflow/lite/micro/tools/make/targets/arc/README.md#install-the-synopsys-designware-arc-metaware-development-toolkit) +section for instructions on toolchain installation. + +#### Make Tool version + +A `'make'` tool is required for deploying Tensorflow Lite Micro +applications on HIMAX WE1 EVB, See +[Check make tool version](/tensorflow/lite/micro/tools/make/targets/arc/README.md#make-tool) +section for proper environment. + +#### Serial Terminal Emulation Application + +There are 2 main purposes for HIMAX WE1 EVB Debug UART port + +- print application output +- burn application to flash by using xmodem send application binary + +You can use any terminal emulation program (like [PuTTY](https://www.putty.org/) or [minicom](https://linux.die.net/man/1/minicom)). + + +### Generate Example Project + +The example project for HIMAX WE1 EVB platform can be generated with the following +command: + +Download related third party data + +``` +make -f tensorflow/lite/micro/tools/make/Makefile TARGET=himax_we1_evb third_party_downloads +``` + +Generate person detection project + +``` +make -f tensorflow/lite/micro/tools/make/Makefile generate_person_detection_int8_make_project TARGET=himax_we1_evb +``` + +### Build and Burn Example + +Following the Steps to run person detection example at HIMAX WE1 EVB platform. + +1. Go to the generated example project directory. + + ``` + cd tensorflow/lite/micro/tools/make/gen/himax_we1_evb_arc/prj/person_detection_int8/make + ``` + +2. Build the example using + + ``` + make app + ``` + +3. After example build finish, copy ELF file and map file to image generate tool directory. + image generate tool directory located at `'tensorflow/lite/micro/tools/make/downloads/himax_we1_sdk/image_gen_linux_v3/'` + + ``` + cp person_detection_int8.elf himax_we1_evb.map ../../../../../downloads/himax_we1_sdk/image_gen_linux_v3/ + ``` + +4. Go to flash image generate tool directory. + + ``` + cd ../../../../../downloads/himax_we1_sdk/image_gen_linux_v3/ + ``` + +5. run image generate tool, generate flash image file. + + * Before running image generate tool, by typing `sudo chmod +x image_gen` + and `sudo chmod +x sign_tool` to make sure it is executable. + + ``` + image_gen -e person_detection_int8.elf -m himax_we1_evb.map -o out.img + ``` + + +6. Download flash image file to HIMAX WE1 EVB by UART: + + * more detail about download image through UART can be found at [HIMAX WE1 EVB update Flash image](https://github.com/HimaxWiseEyePlus/bsp_tflu/tree/master/HIMAX_WE1_EVB_user_guide#flash-image-update) + +After these steps, press reset button on the HIMAX WE1 EVB, you will see application output in the serial +terminal. + ## Running on SparkFun Edge The following instructions will help you build and deploy this sample on the diff --git a/tensorflow/lite/micro/examples/person_detection_experimental/himax_we1_evb/detection_responder.cc b/tensorflow/lite/micro/examples/person_detection_experimental/himax_we1_evb/detection_responder.cc new file mode 100644 index 00000000000..a353dc8a9b8 --- /dev/null +++ b/tensorflow/lite/micro/examples/person_detection_experimental/himax_we1_evb/detection_responder.cc @@ -0,0 +1,34 @@ +/* Copyright 2019 The TensorFlow Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. +==============================================================================*/ + +#include "tensorflow/lite/micro/examples/person_detection_experimental/detection_responder.h" + +#include "hx_drv_tflm.h" + +// This dummy implementation writes person and no person scores to the error +// console. Real applications will want to take some custom action instead, and +// should implement their own versions of this function. +void RespondToDetection(tflite::ErrorReporter* error_reporter, + int8_t person_score, int8_t no_person_score) { + + if (person_score > no_person_score) { + hx_drv_led_on(HX_DRV_LED_GREEN); + } else { + hx_drv_led_off(HX_DRV_LED_GREEN); + } + + TF_LITE_REPORT_ERROR(error_reporter, "person score:%d no person score %d", + person_score, no_person_score); +} diff --git a/tensorflow/lite/micro/examples/person_detection_experimental/himax_we1_evb/image_provider.cc b/tensorflow/lite/micro/examples/person_detection_experimental/himax_we1_evb/image_provider.cc new file mode 100644 index 00000000000..727d93c61d1 --- /dev/null +++ b/tensorflow/lite/micro/examples/person_detection_experimental/himax_we1_evb/image_provider.cc @@ -0,0 +1,44 @@ +/* Copyright 2019 The TensorFlow Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. +==============================================================================*/ + +#include "tensorflow/lite/micro/examples/person_detection_experimental/image_provider.h" + +#include "tensorflow/lite/micro/examples/person_detection_experimental/model_settings.h" + +#include "hx_drv_tflm.h" + +hx_drv_sensor_image_config_t g_pimg_config; + + +TfLiteStatus GetImage(tflite::ErrorReporter* error_reporter, int image_width, + int image_height, int channels, int8_t* image_data) { + static bool is_initialized = false; + + if (!is_initialized) { + if(hx_drv_sensor_initial(&g_pimg_config)!= HX_DRV_LIB_PASS) + { + return kTfLiteError; + } + is_initialized = true; + } + + hx_drv_sensor_capture(&g_pimg_config); + + hx_drv_image_rescale((uint8_t*)g_pimg_config.raw_address, g_pimg_config.img_width, g_pimg_config.img_height, + image_data, image_width, image_height); + + + return kTfLiteOk; +} diff --git a/tensorflow/lite/micro/examples/person_detection_experimental/himax_we1_evb/image_provider.cc~ b/tensorflow/lite/micro/examples/person_detection_experimental/himax_we1_evb/image_provider.cc~ new file mode 100644 index 00000000000..d5b4d136642 --- /dev/null +++ b/tensorflow/lite/micro/examples/person_detection_experimental/himax_we1_evb/image_provider.cc~ @@ -0,0 +1,44 @@ +/* Copyright 2019 The TensorFlow Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. +==============================================================================*/ + +#include "tensorflow/lite/micro/examples/person_detection_experimental/image_provider.h" + +#include "tensorflow/lite/micro/examples/person_detection_experimental/model_settings.h" + +#include "hx_drv_tflm.h" + +hx_drv_sensor_image_config_t g_pimg_config; + + +TfLiteStatus GetImage(tflite::ErrorReporter* error_reporter, int image_width, + int image_height, int channels, int8_t* image_data) { + static bool is_initialized = false; + + if (!is_initialized) { + if(hx_drv_sensor_initial(&g_pimg_config)!= HX_DRV_LIB_PASS) + { + return kTfLiteError; + } + is_initialized = true; + } + + hx_drv_sensor_capture(&g_pimg_config); + + hx_drv_image_rescale((uint8_t*)g_pimg_config.raw_address, g_pimg_config.img_width, g_pimg_config.img_height, + image_data, image_data, image_height); + + + return kTfLiteOk; +} diff --git a/tensorflow/lite/micro/examples/person_detection_experimental/himax_we1_evb/main_functions.cc b/tensorflow/lite/micro/examples/person_detection_experimental/himax_we1_evb/main_functions.cc new file mode 100644 index 00000000000..552b52c9c51 --- /dev/null +++ b/tensorflow/lite/micro/examples/person_detection_experimental/himax_we1_evb/main_functions.cc @@ -0,0 +1,127 @@ +/* Copyright 2019 The TensorFlow Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. +==============================================================================*/ + +#include "tensorflow/lite/micro/examples/person_detection_experimental/main_functions.h" + +#include "tensorflow/lite/micro/examples/person_detection_experimental/detection_responder.h" +#include "tensorflow/lite/micro/examples/person_detection_experimental/image_provider.h" +#include "tensorflow/lite/micro/examples/person_detection_experimental/model_settings.h" +#include "tensorflow/lite/micro/examples/person_detection_experimental/person_detect_model_data.h" +#include "tensorflow/lite/micro/kernels/micro_ops.h" +#include "tensorflow/lite/micro/micro_error_reporter.h" +#include "tensorflow/lite/micro/micro_interpreter.h" +#include "tensorflow/lite/micro/micro_mutable_op_resolver.h" +#include "tensorflow/lite/schema/schema_generated.h" +#include "tensorflow/lite/version.h" + +// Globals, used for compatibility with Arduino-style sketches. +namespace { +tflite::ErrorReporter* error_reporter = nullptr; +const tflite::Model* model = nullptr; +tflite::MicroInterpreter* interpreter = nullptr; +TfLiteTensor* input = nullptr; + +// In order to use optimized tensorflow lite kernels, a signed int8 quantized +// model is preferred over the legacy unsigned model format. This means that +// throughout this project, input images must be converted from unisgned to +// signed format. The easiest and quickest way to convert from unsigned to +// signed 8-bit integers is to subtract 128 from the unsigned value to get a +// signed value. + +// An area of memory to use for input, output, and intermediate arrays. +constexpr int kTensorArenaSize = 125 * 1024; +#pragma Bss(".tensor_arena") +static uint8_t tensor_arena[kTensorArenaSize]; +#pragma Bss() +} // namespace + +// The name of this function is important for Arduino compatibility. +void setup() { + // Set up logging. Google style is to avoid globals or statics because of + // lifetime uncertainty, but since this has a trivial destructor it's okay. + // NOLINTNEXTLINE(runtime-global-variables) + static tflite::MicroErrorReporter micro_error_reporter; + error_reporter = µ_error_reporter; + + // Map the model into a usable data structure. This doesn't involve any + // copying or parsing, it's a very lightweight operation. + model = tflite::GetModel(g_person_detect_model_data); + if (model->version() != TFLITE_SCHEMA_VERSION) { + TF_LITE_REPORT_ERROR(error_reporter, + "Model provided is schema version %d not equal " + "to supported version %d.", + model->version(), TFLITE_SCHEMA_VERSION); + return; + } + + // Pull in only the operation implementations we need. + // This relies on a complete list of all the ops needed by this graph. + // An easier approach is to just use the AllOpsResolver, but this will + // incur some penalty in code space for op implementations that are not + // needed by this graph. + // + // tflite::ops::micro::AllOpsResolver resolver; + // NOLINTNEXTLINE(runtime-global-variables) + static tflite::MicroOpResolver<12> micro_op_resolver; + micro_op_resolver.AddBuiltin(tflite::BuiltinOperator_DEPTHWISE_CONV_2D, + tflite::ops::micro::Register_DEPTHWISE_CONV_2D(), + 1, 3); + micro_op_resolver.AddBuiltin(tflite::BuiltinOperator_CONV_2D, + tflite::ops::micro::Register_CONV_2D(), 1, 3); + micro_op_resolver.AddBuiltin(tflite::BuiltinOperator_AVERAGE_POOL_2D, + tflite::ops::micro::Register_AVERAGE_POOL_2D(), + 1, 2); + micro_op_resolver.AddBuiltin(tflite::BuiltinOperator_RESHAPE, + tflite::ops::micro::Register_RESHAPE()); + micro_op_resolver.AddBuiltin(tflite::BuiltinOperator_SOFTMAX, + tflite::ops::micro::Register_SOFTMAX(), 1, 3); + + // Build an interpreter to run the model with. + // NOLINTNEXTLINE(runtime-global-variables) + static tflite::MicroInterpreter static_interpreter( + model, micro_op_resolver, tensor_arena, kTensorArenaSize, error_reporter); + interpreter = &static_interpreter; + + // Allocate memory from the tensor_arena for the model's tensors. + TfLiteStatus allocate_status = interpreter->AllocateTensors(); + if (allocate_status != kTfLiteOk) { + TF_LITE_REPORT_ERROR(error_reporter, "AllocateTensors() failed"); + return; + } + + // Get information about the memory area to use for the model's input. + input = interpreter->input(0); +} + +// The name of this function is important for Arduino compatibility. +void loop() { + // Get image from provider. + if (kTfLiteOk != GetImage(error_reporter, kNumCols, kNumRows, kNumChannels, + input->data.int8)) { + TF_LITE_REPORT_ERROR(error_reporter, "Image capture failed."); + } + + // Run the model on this input and make sure it succeeds. + if (kTfLiteOk != interpreter->Invoke()) { + TF_LITE_REPORT_ERROR(error_reporter, "Invoke failed."); + } + + TfLiteTensor* output = interpreter->output(0); + + // Process the inference results. + int8_t person_score = output->data.uint8[kPersonIndex]; + int8_t no_person_score = output->data.uint8[kNotAPersonIndex]; + RespondToDetection(error_reporter, person_score, no_person_score); +} diff --git a/tensorflow/lite/micro/we_i/debug_log.cc b/tensorflow/lite/micro/himax_we1_evb/debug_log.cc similarity index 90% rename from tensorflow/lite/micro/we_i/debug_log.cc rename to tensorflow/lite/micro/himax_we1_evb/debug_log.cc index a115d476aff..32af2625630 100644 --- a/tensorflow/lite/micro/we_i/debug_log.cc +++ b/tensorflow/lite/micro/himax_we1_evb/debug_log.cc @@ -18,16 +18,15 @@ limitations under the License. // the Ambiq Apollo 3. #include "tensorflow/lite/micro/debug_log.h" -#include "xprintf.h" -#include "console_io.h" -#include +#include "hx_drv_tflm.h" + extern "C" void DebugLog(const char* s) { static bool is_initialized = false; if (!is_initialized) { - xprintf_setup(); + hx_drv_uart_initial(); is_initialized = true; } - xprintf("%s", s); + hx_drv_uart_print("%s", s); } diff --git a/tensorflow/lite/micro/tools/make/targets/himax_we1_evb_makefile.inc b/tensorflow/lite/micro/tools/make/targets/himax_we1_evb_makefile.inc new file mode 100644 index 00000000000..60fc2e7cca1 --- /dev/null +++ b/tensorflow/lite/micro/tools/make/targets/himax_we1_evb_makefile.inc @@ -0,0 +1,91 @@ +# Settings for himax WE_1 evb. +ifeq ($(TARGET), himax_we1_evb) + + CC_TOOL = ccac + AR_TOOL = arac + CXX_TOOL = ccac + LD_TOOL := ccac + TARGET_ARCH := arc + #ARC_TOOLCHAIN := mwdt + + BUILD_ARC_MLI := false + ARC_MLI_PRE_COMPILED_TARGET := himax_arcem9d_r16 + + include $(MAKEFILE_DIR)/targets/arc/arc_common.inc + + #download SDK & MLI + HIMAX_WE1_SDK_NAME := himax_we1_sdk + $(eval $(call add_third_party_download,$(HIMAX_WE1_SDK_URL),$(HIMAX_WE1_SDK_MD5),$(HIMAX_WE1_SDK_NAME),)) + + #export path of toolchain + #export PATH := $(MAKEFILE_DIR)/downloads/$(HIMAX_WE1_SDK_NAME)/image_gen_linux_v3/:$(PATH) + + TCF_FILE := $(PWD)/$(MAKEFILE_DIR)/downloads/$(HIMAX_WE1_SDK_NAME)/arcem9d_wei_r16.tcf + LCF_FILE := $(PWD)/$(MAKEFILE_DIR)/downloads/$(HIMAX_WE1_SDK_NAME)/memory.lcf + ARCLIB_FILE := $(PWD)/$(MAKEFILE_DIR)/downloads/$(HIMAX_WE1_SDK_NAME)/libembarc.a + LIB_HEADER_FILE := $(PWD)/$(MAKEFILE_DIR)/downloads/$(HIMAX_WE1_SDK_NAME)/hx_drv_tflm.h + + + DEFAULT_HEAPSZ := 8192 + DEFAULT_STACKSZ := 8192 + + TCF_FILE_NAME = $(notdir $(TCF_FILE)) + ARC_TARGET_FILES_DIRS = $(dir $(TCF_FILE_NAME)) + MAKE_PROJECT_FILES += $(TCF_FILE_NAME) + + LCF_FILE_NAME = $(notdir $(LCF_FILE)) + ARC_TARGET_FILES_DIRS += $(dir $(LCF_FILE)) + MAKE_PROJECT_FILES += $(LCF_FILE_NAME) + + ARCLIB_FILE_NAME = $(notdir $(ARCLIB_FILE)) + ARC_TARGET_FILES_DIRS += $(dir $(ARCLIB_FILE)) + MAKE_PROJECT_FILES += $(ARCLIB_FILE_NAME) + + LIB_HEADER_FILE_NAME = $(notdir $(LIB_HEADER_FILE)) + ARC_TARGET_FILES_DIRS += $(dir $(LIB_HEADER_FILE)) + MAKE_PROJECT_FILES += $(LIB_HEADER_FILE_NAME) + + + + # Need a pointer to the TCF and lcf file + + PLATFORM_FLAGS = \ + -DNDEBUG \ + -g \ + -DCPU_ARC \ + -Hnosdata \ + -DTF_LITE_STATIC_MEMORY \ + -tcf=$(TCF_FILE_NAME) \ + -Hnocopyr \ + -Hpurge \ + -Hcl \ + -fslp-vectorize-aggressive \ + -ffunction-sections \ + -fdata-sections \ + -tcf_core_config \ + + CXXFLAGS += -fno-rtti -DSCRATCH_MEM_Z_SIZE=0x10000 $(PLATFORM_FLAGS) + CCFLAGS += $(PLATFORM_FLAGS) + + INCLUDES+= \ + -I $(MAKEFILE_DIR)/downloads/$(WEI_SDK_NAME) \ + -I $(MAKEFILE_DIR)/downloads/kissfft + + GENERATED_PROJECT_INCLUDES += \ + -I. \ + -I./third_party/kissfft + + LDFLAGS += \ + -Hheap=8192 \ + -tcf=$(TCF_FILE_NAME) \ + -Hnocopyr \ + -m \ + -Hldopt=-Coutput=$(TARGET).map \ + $(LCF_FILE_NAME) \ + -Hldopt=-Bgrouplib $(ARCLIB_FILE_NAME) + + CXXFLAGS := $(filter-out -std=c++11,$(CXXFLAGS)) + CCFLAGS := $(filter-out -std=c11,$(CCFLAGS)) + MICROLITE_LIBS := $(filter-out -lm,$(MICROLITE_LIBS)) + +endif diff --git a/tensorflow/lite/micro/tools/make/targets/himax_we1_evb_makefile.inc~ b/tensorflow/lite/micro/tools/make/targets/himax_we1_evb_makefile.inc~ new file mode 100644 index 00000000000..733f258fbbb --- /dev/null +++ b/tensorflow/lite/micro/tools/make/targets/himax_we1_evb_makefile.inc~ @@ -0,0 +1,93 @@ +# Settings for himax WE_1 evb. +ifeq ($(TARGET), himax_we1_evb) + + CC_TOOL = ccac + AR_TOOL = arac + CXX_TOOL = ccac + LD_TOOL := ccac + TARGET_ARCH := arc + #ARC_TOOLCHAIN := mwdt + + BUILD_ARC_MLI := false + ARC_MLI_PRE_COMPILED_TARGET := himax_arcem9d_r16 + +include $(MAKEFILE_DIR)/targets/arc/arc_common.inc + #download SDK & MLI + HIMAX_WE1_SDK_NAME := himax_we1_sdk + #MLI_LIB_DIR = arc_mli_package + #MLI_LIB_DIR = arc_mli_package + #$(eval $(call add_third_party_download,$(EMBARC_MLI_PRE_COMPILED_URL),$(EMBARC_MLI_PRE_COMPILED_MD5),$(MLI_LIB_DIR),)) + $(eval $(call add_third_party_download,$(HIMAX_WE1_SDK_URL),$(HIMAX_WE1_SDK_MD5),$(HIMAX_WE1_SDK_NAME),)) + + #export path of toolchain + #export PATH := $(MAKEFILE_DIR)/downloads/$(HIMAX_WE1_SDK_NAME)/image_gen_linux_v3/:$(PATH) + + TCF_FILE := $(PWD)/$(MAKEFILE_DIR)/downloads/$(HIMAX_WE1_SDK_NAME)/arcem9d_wei_r16.tcf + LCF_FILE := $(PWD)/$(MAKEFILE_DIR)/downloads/$(HIMAX_WE1_SDK_NAME)/memory.lcf + ARCLIB_FILE := $(PWD)/$(MAKEFILE_DIR)/downloads/$(HIMAX_WE1_SDK_NAME)/libembarc.a + LIB_HEADER_FILE := $(PWD)/$(MAKEFILE_DIR)/downloads/$(HIMAX_WE1_SDK_NAME)/hx_drv_tflm.h + + + DEFAULT_HEAPSZ := 8192 + DEFAULT_STACKSZ := 8192 + + TCF_FILE_NAME = $(notdir $(TCF_FILE)) + ARC_TARGET_FILES_DIRS = $(dir $(TCF_FILE_NAME)) + MAKE_PROJECT_FILES += $(TCF_FILE_NAME) + + LCF_FILE_NAME = $(notdir $(LCF_FILE)) + ARC_TARGET_FILES_DIRS += $(dir $(LCF_FILE)) + MAKE_PROJECT_FILES += $(LCF_FILE_NAME) + + ARCLIB_FILE_NAME = $(notdir $(ARCLIB_FILE)) + ARC_TARGET_FILES_DIRS += $(dir $(ARCLIB_FILE)) + MAKE_PROJECT_FILES += $(ARCLIB_FILE_NAME) + + LIB_HEADER_FILE_NAME = $(notdir $(LIB_HEADER_FILE)) + ARC_TARGET_FILES_DIRS += $(dir $(LIB_HEADER_FILE)) + MAKE_PROJECT_FILES += $(LIB_HEADER_FILE_NAME) + + + + # Need a pointer to the TCF and lcf file + + PLATFORM_FLAGS = \ + -DNDEBUG \ + -g \ + -DCPU_ARC \ + -Hnosdata \ + -DTF_LITE_STATIC_MEMORY \ + -tcf=$(TCF_FILE_NAME) \ + -Hnocopyr \ + -Hpurge \ + -Hcl \ + -fslp-vectorize-aggressive \ + -ffunction-sections \ + -fdata-sections \ + -tcf_core_config \ + + CXXFLAGS += -fno-rtti -DSCRATCH_MEM_Z_SIZE=0x10000 $(PLATFORM_FLAGS) + CCFLAGS += $(PLATFORM_FLAGS) + + INCLUDES+= \ + -I $(MAKEFILE_DIR)/downloads/$(WEI_SDK_NAME) \ + -I $(MAKEFILE_DIR)/downloads/kissfft + + GENERATED_PROJECT_INCLUDES += \ + -I. \ + -I./third_party/kissfft + + LDFLAGS += \ + -Hheap=8192 \ + -tcf=$(TCF_FILE_NAME) \ + -Hnocopyr \ + -m \ + -Hldopt=-Coutput=$(TARGET).map \ + $(LCF_FILE_NAME) \ + -Hldopt=-Bgrouplib $(ARCLIB_FILE_NAME) + + CXXFLAGS := $(filter-out -std=c++11,$(CXXFLAGS)) + CCFLAGS := $(filter-out -std=c11,$(CCFLAGS)) + MICROLITE_LIBS := $(filter-out -lm,$(MICROLITE_LIBS)) + +endif diff --git a/tensorflow/lite/micro/tools/make/third_party_downloads.inc b/tensorflow/lite/micro/tools/make/third_party_downloads.inc index 806501a004a..75a51e0df10 100644 --- a/tensorflow/lite/micro/tools/make/third_party_downloads.inc +++ b/tensorflow/lite/micro/tools/make/third_party_downloads.inc @@ -80,3 +80,7 @@ EMBARC_MLI_PRE_COMPILED_MD5 := "a95ff9e0370434484f14e7e4114327f6" XTENSA_HIFI4_URL :="https://github.com/foss-xtensa/nnlib-hifi4/raw/master/archive/xa_nnlib_04_07.zip" XTENSA_HIFI4_MD5 :="f234764928f9a42901df33a27e118c8b" +HIMAX_WE1_SDK_URL ="https://www.himax.com.tw/we-i/himax_we1_sdk_v02.zip" +HIMAX_WE1_SDK_MD5 ="9a4b2f29b16052764e437b64bdcba816" + + diff --git a/tensorflow/lite/micro/tools/make/third_party_downloads.inc~ b/tensorflow/lite/micro/tools/make/third_party_downloads.inc~ new file mode 100644 index 00000000000..3c7ee1b64d2 --- /dev/null +++ b/tensorflow/lite/micro/tools/make/third_party_downloads.inc~ @@ -0,0 +1,86 @@ +# Add URLs and MD5 checksums for third-party libraries here. + +GEMMLOWP_URL := "https://github.com/google/gemmlowp/archive/719139ce755a0f31cbf1c37f7f98adcc7fc9f425.zip" +GEMMLOWP_MD5 := "7e8191b24853d75de2af87622ad293ba" + +ifeq ($(HOST_OS),windows) + FLATBUFFERS_URL := "https://github.com/google/flatbuffers/archive/v1.12.0.zip" + FLATBUFFERS_MD5 := "a1afdbf114dec01a861c1b8c917d0fc7" +else + FLATBUFFERS_URL := "https://github.com/google/flatbuffers/archive/v1.12.0.tar.gz" + FLATBUFFERS_MD5 := "c62ffefb3d4548b127cca14ce047f16c" +endif + +ifeq ($(HOST_OS),osx) + GCC_EMBEDDED_URL := "https://developer.arm.com/-/media/Files/downloads/gnu-rm/7-2018q2/gcc-arm-none-eabi-7-2018-q2-update-mac.tar.bz2" + GCC_EMBEDDED_MD5 := "a66be9828cf3c57d7d21178e07cd8904" +else ifeq ($(HOST_OS),windows) + GCC_EMBEDDED_URL := "https://developer.arm.com/-/media/Files/downloads/gnu-rm/7-2018q2/gcc-arm-none-eabi-7-2018-q2-update-win32.zip" + GCC_EMBEDDED_MD5 := "bc8ae26d7c429f30d583a605a4bcf9bc" +else + GCC_EMBEDDED_URL := "https://developer.arm.com/-/media/Files/downloads/gnu-rm/7-2018q2/gcc-arm-none-eabi-7-2018-q2-update-linux.tar.bz2" + GCC_EMBEDDED_MD5 := "299ebd3f1c2c90930d28ab82e5d8d6c0" +endif + +LEON_BCC2_URL := "https://www.gaisler.com/anonftp/bcc2/bin/bcc-2.0.7-gcc-linux64.tar.xz" +LEON_BCC2_MD5 := "cdf78082be4882da2a92c9baa82fe765" + +TSIM_URL := "https://www.gaisler.com/anonftp/tsim/tsim-eval-2.0.63.tar.gz" +TSIM_MD5 := "afa0095d3ed989a949e1467f94e41d2f" + +CMSIS_URL := "https://github.com/ARM-software/CMSIS_5/archive/1150e71e07c79b538efd842aba5b210a31827ae5.zip" +CMSIS_MD5 := "e05f4222ef58825193910b41a0871dcb" + +AM_SDK_URL := "http://s3.asia.ambiqmicro.com/downloads/AmbiqSuite-Rel2.2.0.zip" +AM_SDK_MD5 := "7605fa2d4d97e6bb7a1190c92b66b597" +AM_SDK_DEST := AmbiqSuite-Rel2.2.0 + +SF_BSPS_URL := "https://github.com/sparkfun/SparkFun_Apollo3_AmbiqSuite_BSPs/archive/v0.0.7.zip" +SF_BSPS_MD5 := "34199f7e754735661d1c8a70a40ca7a3" +SF_BSPS_DEST := boards_sfe + +STM32_BARE_LIB_URL := "https://github.com/google/stm32_bare_lib/archive/c07d611fb0af58450c5a3e0ab4d52b47f99bc82d.zip" +STM32_BARE_LIB_MD5 := "282bff40d4d0b92278fd123a3b6e3123" + +ifeq ($(HOST_OS),osx) + RISCV_TOOLCHAIN_URL := "https://static.dev.sifive.com/dev-tools/riscv64-unknown-elf-gcc-8.1.0-2019.01.0-x86_64-apple-darwin.tar.gz" + RISCV_TOOLCHAIN_MD5 := "2ac2fa00618b9ab7fa0c7d0ec173de94" +else + RISCV_TOOLCHAIN_URL := "https://static.dev.sifive.com/dev-tools/riscv64-unknown-elf-gcc-20181030-x86_64-linux-ubuntu14.tar.gz" + RISCV_TOOLCHAIN_MD5="2366b7afe36a54dc94fb0ff8a0830934" +endif + +SIFIVE_FE310_LIB_URL := "https://github.com/sifive/freedom-e-sdk/archive/baeeb8fd497a99b3c141d7494309ec2e64f19bdf.zip" +SIFIVE_FE310_LIB_MD5 := "06ee24c4956f8e21670ab3395861fe64" + +KISSFFT_URL="https://github.com/mborgerding/kissfft/archive/v130.zip" +KISSFFT_MD5="438ba1fef5783cc5f5f201395cc477ca" + +RUY_URL="https://github.com/google/ruy/archive/1b313682ef8b8fc8ed08719c610d1c3503b016bf.zip" +RUY_MD5="2d54f058f8f7120dfc1ecee79dbf259e" + +CIFAR10_DATASET_URL="https://www.cs.toronto.edu/~kriz/cifar-10-binary.tar.gz" +CIFAR10_DATASET_MD5="c32a1d4ab5d03f1284b67883e8d87530" + +IMAGE_RECOGNITION_MODEL_URL := "https://storage.googleapis.com/download.tensorflow.org/models/tflite/cifar_image_recognition_model_2020_4_14.zip" +IMAGE_RECOGNITION_MODEL_MD5 := "2b886156e7ef4d6e53d0f1a4bc800e56" + +PERSON_MODEL_URL := "https://storage.googleapis.com/download.tensorflow.org/data/tf_lite_micro_person_data_grayscale_2019_11_21.zip" +PERSON_MODEL_MD5 := "fe2934bd0788f1dcc7af3f0a954542ab" + +PERSON_MODEL_INT8_URL := "https://storage.googleapis.com/download.tensorflow.org/data/tf_lite_micro_person_data_int8_grayscale_2020_01_13.zip" +PERSON_MODEL_INT8_MD5 := "8a7d2c70325f53136faea6dde517b8cc" + +EMBARC_MLI_URL := "https://github.com/foss-for-synopsys-dwc-arc-processors/embarc_mli/archive/58284867ca52d1f43b25045e8601999d7359d986.zip" +EMBARC_MLI_MD5 := "2bf4982a327fdaa9d475803ce014d1ef" + +EMBARC_MLI_PRE_COMPILED_URL := "https://github.com/foss-for-synopsys-dwc-arc-processors/embarc_mli/releases/download/Release_1.1_RC2/embARC_MLI_package.zip" +EMBARC_MLI_PRE_COMPILED_MD5 := "a95ff9e0370434484f14e7e4114327f6" + +XTENSA_HIFI4_URL :="https://github.com/foss-xtensa/nnlib-hifi4/raw/master/archive/xa_nnlib_04_07.zip" +XTENSA_HIFI4_MD5 :="f234764928f9a42901df33a27e118c8b" + +HIMAX_WE1_SDK_URL ="https://www.himax.com.tw/we-i/himax_we1_sdk_v02.zip" +HIMAX_WE1_SDK_MD5 ="5063c24d298fbcfe118163f3ccc43079" + + From 96eb311826d68179bf85a228e294ae55f39ef2e4 Mon Sep 17 00:00:00 2001 From: "902449@58880@bigcat_chen@ASIC" Date: Thu, 4 Jun 2020 18:08:43 +0800 Subject: [PATCH 045/867] TFLM: update example readme --- .../lite/micro/examples/hello_world/README.md | 36 +- .../micro/examples/hello_world/README.md~ | 2 +- .../hello_world/himax_we1_evb/constants.cc | 19 - .../himax_we1_evb/output_handler.cc | 35 -- .../himax_we1_evb/output_handler.cc~ | 53 -- .../person_detection_experimental/README.md | 33 +- .../person_detection_experimental/README.md~ | 568 ++++++++++++++++++ .../tools/make/third_party_downloads.inc~ | 86 --- 8 files changed, 585 insertions(+), 247 deletions(-) delete mode 100644 tensorflow/lite/micro/examples/hello_world/himax_we1_evb/constants.cc delete mode 100644 tensorflow/lite/micro/examples/hello_world/himax_we1_evb/output_handler.cc delete mode 100644 tensorflow/lite/micro/examples/hello_world/himax_we1_evb/output_handler.cc~ create mode 100644 tensorflow/lite/micro/examples/person_detection_experimental/README.md~ delete mode 100644 tensorflow/lite/micro/tools/make/third_party_downloads.inc~ diff --git a/tensorflow/lite/micro/examples/hello_world/README.md b/tensorflow/lite/micro/examples/hello_world/README.md index 9c0a5e2306a..d3762ada790 100644 --- a/tensorflow/lite/micro/examples/hello_world/README.md +++ b/tensorflow/lite/micro/examples/hello_world/README.md @@ -14,34 +14,14 @@ of the device. ## Table of contents -- [Hello World Example](#hello-world-example) - - [Table of contents](#table-of-contents) - - [Deploy to ARC EM SDP](#deploy-to-arc-em-sdp) - - [Initial Setup](#initial-setup) - - [Generate Example Project](#generate-example-project) - - [Build and Run Example](#build-and-run-example) - - [Deploy to Arduino](#deploy-to-arduino) - - [Install the Arduino_TensorFlowLite library](#install-the-arduinotensorflowlite-library) - - [Load and run the example](#load-and-run-the-example) - - [Deploy to ESP32](#deploy-to-esp32) - - [Install the ESP IDF](#install-the-esp-idf) - - [Generate the examples](#generate-the-examples) - - [Building the example](#building-the-example) - - [Load and run the example](#load-and-run-the-example-1) - - [Deploy to himax WE1 EVB](#deploy-to-himax-we1-evb) - - [Initial Setup](#initial-setup-1) - - [MetaWare Development Toolkit](#metaware-development-toolkit) - - [Make Tool version](#make-tool-version) - - [Serial Terminal Emulation Application](#serial-terminal-emulation-application) - - [Generate Example Project](#generate-example-project-1) - - [Build and Burn Example](#build-and-burn-example) - - [Deploy to SparkFun Edge](#deploy-to-sparkfun-edge) - - [Compile the binary](#compile-the-binary) - - [Sign the binary](#sign-the-binary) - - [Flash the binary](#flash-the-binary) - - [Deploy to STM32F746](#deploy-to-stm32f746) - - [Run the tests on a development machine](#run-the-tests-on-a-development-machine) - - [Train your own model](#train-your-own-model) +- [Deploy to ARC EM SDP](#deploy-to-arc-em-sdp) +- [Deploy to Arduino](#deploy-to-arduino) +- [Deploy to ESP32](#deploy-to-esp32) +- [Deploy to himax WE1 EVB](#deploy-to-himax-we1-evb) +- [Deploy to SparkFun Edge](#deploy-to-sparkfun-edge) +- [Deploy to STM32F746](#deploy-to-STM32F746) +- [Run the tests on a development machine](#run-the-tests-on-a-development-machine) +- [Train your own model](#train-your-own-model) ## Deploy to ARC EM SDP diff --git a/tensorflow/lite/micro/examples/hello_world/README.md~ b/tensorflow/lite/micro/examples/hello_world/README.md~ index 011711493d5..9c0a5e2306a 100644 --- a/tensorflow/lite/micro/examples/hello_world/README.md~ +++ b/tensorflow/lite/micro/examples/hello_world/README.md~ @@ -260,7 +260,7 @@ make -f tensorflow/lite/micro/tools/make/Makefile TARGET=himax_we1_evb third_par Generate hello world project ``` -make -f tensorflow/lite/micro/tools/make/Makefile generate_hello_world_make_project TARGET=himax_we1_evb +make -f tensorflow/lite/micro/tools/make/Makefile generate_hello_world_make_project TARGET=himax_we1_evb TAGS=no_arc_mli ``` ### Build and Burn Example diff --git a/tensorflow/lite/micro/examples/hello_world/himax_we1_evb/constants.cc b/tensorflow/lite/micro/examples/hello_world/himax_we1_evb/constants.cc deleted file mode 100644 index 1816a2f3207..00000000000 --- a/tensorflow/lite/micro/examples/hello_world/himax_we1_evb/constants.cc +++ /dev/null @@ -1,19 +0,0 @@ -/* Copyright 2019 The TensorFlow Authors. All Rights Reserved. - -Licensed under the Apache License, Version 2.0 (the "License"); -you may not use this file except in compliance with the License. -You may obtain a copy of the License at - - http://www.apache.org/licenses/LICENSE-2.0 - -Unless required by applicable law or agreed to in writing, software -distributed under the License is distributed on an "AS IS" BASIS, -WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -See the License for the specific language governing permissions and -limitations under the License. -==============================================================================*/ - -#include "tensorflow/lite/micro/examples/hello_world/constants.h" - -// This is tuned so that a full cycle takes ~4 seconds on a SparkFun Edge. -const int kInferencesPerCycle = 1000; diff --git a/tensorflow/lite/micro/examples/hello_world/himax_we1_evb/output_handler.cc b/tensorflow/lite/micro/examples/hello_world/himax_we1_evb/output_handler.cc deleted file mode 100644 index 8ca028acc55..00000000000 --- a/tensorflow/lite/micro/examples/hello_world/himax_we1_evb/output_handler.cc +++ /dev/null @@ -1,35 +0,0 @@ -/* Copyright 2019 The TensorFlow Authors. All Rights Reserved. - -Licensed under the Apache License, Version 2.0 (the "License"); -you may not use this file except in compliance with the License. -You may obtain a copy of the License at - - http://www.apache.org/licenses/LICENSE-2.0 - -Unless required by applicable law or agreed to in writing, software -distributed under the License is distributed on an "AS IS" BASIS, -WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -See the License for the specific language governing permissions and -limitations under the License. -==============================================================================*/ - -#include "tensorflow/lite/micro/examples/hello_world/output_handler.h" - - -/* -This function trigger different device's LEDthrough y value. -y value range -1 <= y <= 1. -| Range is from -1~1 | LEDs | -| 0 <= y <= 1 | [ 0 1 ] | -| -1 <= y < 0 | [ 1 0 ] | - -*/ -void HandleOutput(tflite::ErrorReporter* error_reporter, float x_value, - float y_value) { - // The first time this method runs, set up our LEDs correctly - - // Log the current X and Y values - TF_LITE_REPORT_ERROR(error_reporter, "x_value: %f, y_value: %f\n", - static_cast(x_value), - static_cast(y_value)); -} diff --git a/tensorflow/lite/micro/examples/hello_world/himax_we1_evb/output_handler.cc~ b/tensorflow/lite/micro/examples/hello_world/himax_we1_evb/output_handler.cc~ deleted file mode 100644 index b59242d0b6f..00000000000 --- a/tensorflow/lite/micro/examples/hello_world/himax_we1_evb/output_handler.cc~ +++ /dev/null @@ -1,53 +0,0 @@ -/* Copyright 2019 The TensorFlow Authors. All Rights Reserved. - -Licensed under the Apache License, Version 2.0 (the "License"); -you may not use this file except in compliance with the License. -You may obtain a copy of the License at - - http://www.apache.org/licenses/LICENSE-2.0 - -Unless required by applicable law or agreed to in writing, software -distributed under the License is distributed on an "AS IS" BASIS, -WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -See the License for the specific language governing permissions and -limitations under the License. -==============================================================================*/ - -#include "tensorflow/lite/micro/examples/hello_world/output_handler.h" - - -/* -This function trigger different device's LEDthrough y value. -y value range -1 <= y <= 1. -| Range is from -1~1 | LEDs | -| 0 <= y <= 1 | [ 0 1 ] | -| -1 <= y < 0 | [ 1 0 ] | - -*/ -void HandleOutput(tflite::ErrorReporter* error_reporter, float x_value, - float y_value) { - // The first time this method runs, set up our LEDs correctly -/* static bool is_initialized = false; - if (!is_initialized) { - // TODO Setup LED's as outputs - - // end of setup - is_initialized = true; - } - - // Set the LEDs to represent negative values - if (y_value < 0) { - //enable LED1 - - //enable LED0 - } else if (y_value > 0) { - //enable LED0 - - //enable LED1 - } - */ - // Log the current X and Y values - TF_LITE_REPORT_ERROR(error_reporter, "x_value: %f, y_value: %f\n", - static_cast(x_value), - static_cast(y_value)); -} diff --git a/tensorflow/lite/micro/examples/person_detection_experimental/README.md b/tensorflow/lite/micro/examples/person_detection_experimental/README.md index 4d53e551431..06f5640986f 100644 --- a/tensorflow/lite/micro/examples/person_detection_experimental/README.md +++ b/tensorflow/lite/micro/examples/person_detection_experimental/README.md @@ -7,31 +7,14 @@ This uses the experimental int8 quantized version of the person detection model. ## Table of contents -- [Person detection example](#person-detection-example) - - [Table of contents](#table-of-contents) - - [Running on ARC EM SDP](#running-on-arc-em-sdp) - - [Initial setup](#initial-setup) - - [Generate Example Project](#generate-example-project) - - [Build and Run Example](#build-and-run-example) - - [Running on Arduino](#running-on-arduino) - - [Hardware](#hardware) - - [Install the Arduino_TensorFlowLite library](#install-the-arduinotensorflowlite-library) - - [Install other libraries](#install-other-libraries) - - [Load and run the example](#load-and-run-the-example) - - [Running on HIMAX WE1 EVB](#running-on-himax-we1-evb) - - [Initial Setup](#initial-setup) - - [MetaWare Development Toolkit](#metaware-development-toolkit) - - [Make Tool version](#make-tool-version) - - [Serial Terminal Emulation Application](#serial-terminal-emulation-application) - - [Generate Example Project](#generate-example-project-1) - - [Build and Burn Example](#build-and-burn-example) - - [Running on SparkFun Edge](#running-on-sparkfun-edge) - - [Compile the binary](#compile-the-binary) - - [Sign the binary](#sign-the-binary) - - [Flash the binary](#flash-the-binary) - - [Run the tests on a development machine](#run-the-tests-on-a-development-machine) - - [Debugging image capture](#debugging-image-capture) - - [Training your own model](#training-your-own-model) +- [Getting started](#getting-started) +- [Running on ARC EM SDP](#running-on-arc-em-sdp) +- [Running on Arduino](#running-on-arduino) +- [Running on HIMAX WE1 EVB](#running-on-himax-we1-evb) +- [Running on SparkFun Edge](#running-on-sparkfun-edge) +- [Run the tests on a development machine](#run-the-tests-on-a-development-machine) +- [Debugging image capture](#debugging-image-capture) +- [Training your own model](#training-your-own-model) ## Running on ARC EM SDP diff --git a/tensorflow/lite/micro/examples/person_detection_experimental/README.md~ b/tensorflow/lite/micro/examples/person_detection_experimental/README.md~ new file mode 100644 index 00000000000..4d53e551431 --- /dev/null +++ b/tensorflow/lite/micro/examples/person_detection_experimental/README.md~ @@ -0,0 +1,568 @@ +# Person detection example + +This example shows how you can use Tensorflow Lite to run a 250 kilobyte neural +network to recognize people in images captured by a camera. It is designed to +run on systems with small amounts of memory such as microcontrollers and DSPs. +This uses the experimental int8 quantized version of the person detection model. + +## Table of contents + +- [Person detection example](#person-detection-example) + - [Table of contents](#table-of-contents) + - [Running on ARC EM SDP](#running-on-arc-em-sdp) + - [Initial setup](#initial-setup) + - [Generate Example Project](#generate-example-project) + - [Build and Run Example](#build-and-run-example) + - [Running on Arduino](#running-on-arduino) + - [Hardware](#hardware) + - [Install the Arduino_TensorFlowLite library](#install-the-arduinotensorflowlite-library) + - [Install other libraries](#install-other-libraries) + - [Load and run the example](#load-and-run-the-example) + - [Running on HIMAX WE1 EVB](#running-on-himax-we1-evb) + - [Initial Setup](#initial-setup) + - [MetaWare Development Toolkit](#metaware-development-toolkit) + - [Make Tool version](#make-tool-version) + - [Serial Terminal Emulation Application](#serial-terminal-emulation-application) + - [Generate Example Project](#generate-example-project-1) + - [Build and Burn Example](#build-and-burn-example) + - [Running on SparkFun Edge](#running-on-sparkfun-edge) + - [Compile the binary](#compile-the-binary) + - [Sign the binary](#sign-the-binary) + - [Flash the binary](#flash-the-binary) + - [Run the tests on a development machine](#run-the-tests-on-a-development-machine) + - [Debugging image capture](#debugging-image-capture) + - [Training your own model](#training-your-own-model) + +## Running on ARC EM SDP + +The following instructions will help you to build and deploy this example to +[ARC EM SDP](https://www.synopsys.com/dw/ipdir.php?ds=arc-em-software-development-platform) +board. General information and instructions on using the board with TensorFlow +Lite Micro can be found in the common +[ARC targets description](/tensorflow/lite/micro/tools/make/targets/arc/README.md). + +This example uses asymmetric int8 quantization and can therefore leverage +optimized int8 kernels from the embARC MLI library + +The ARC EM SDP board contains a rich set of extension interfaces. You can choose +any compatible camera and modify +[image_provider.cc](/tensorflow/lite/micro/examples/person_detection_experimental/image_provider.cc) +file accordingly to use input from your specific camera. By default, results of +running this example are printed to the console. If you would like to instead +implement some target-specific actions, you need to modify +[detection_responder.cc](/tensorflow/lite/micro/examples/person_detection_experimental/detection_responder.cc) +accordingly. + +The reference implementations of these files are used by default on the EM SDP. + +### Initial setup + +Follow the instructions on the +[ARC EM SDP Initial Setup](/tensorflow/lite/micro/tools/make/targets/arc/README.md#ARC-EM-Software-Development-Platform-ARC-EM-SDP) +to get and install all required tools for work with ARC EM SDP. + +### Generate Example Project + +The example project for ARC EM SDP platform can be generated with the following +command: + +``` +make -f tensorflow/lite/micro/tools/make/Makefile TARGET=arc_emsdp generate_person_detection_int8_make_project +``` + +### Build and Run Example + +For more detailed information on building and running examples see the +appropriate sections of general descriptions of the +[ARC EM SDP usage with TFLM](/tensorflow/lite/micro/tools/make/targets/arc/README.md#ARC-EM-Software-Development-Platform-ARC-EM-SDP). +In the directory with generated project you can also find a +*README_ARC_EMSDP.md* file with instructions and options on building and +running. Here we only briefly mention main steps which are typically enough to +get it started. + +1. You need to + [connect the board](/tensorflow/lite/micro/tools/make/targets/arc/README.md#connect-the-board) + and open an serial connection. + +2. Go to the generated example project director + + ``` + cd tensorflow/lite/micro/tools/make/gen/arc_emsdp_arc/prj/person_detection_int8/make + ``` + +3. Build the example using + + ``` + make app + ``` + +4. To generate artefacts for self-boot of example from the board use + + ``` + make flash + ``` + +5. To run application from the board using microSD card: + + * Copy the content of the created /bin folder into the root of microSD + card. Note that the card must be formatted as FAT32 with default cluster + size (but less than 32 Kbytes) + * Plug in the microSD card into the J11 connector. + * Push the RST button. If a red LED is lit beside RST button, push the CFG + button. + +6. If you have the MetaWare Debugger installed in your environment: + + * To run application from the console using it type `make run`. + * To stop the execution type `Ctrl+C` in the console several times. + +In both cases (step 5 and 6) you will see the application output in the serial +terminal. + +## Running on Arduino + +The following instructions will help you build and deploy this sample +to [Arduino](https://www.arduino.cc/) devices. + +The sample has been tested with the following device: + +- [Arduino Nano 33 BLE Sense](https://store.arduino.cc/usa/nano-33-ble-sense-with-headers) + +You will also need the following camera module: + +- [Arducam Mini 2MP Plus](https://www.amazon.com/Arducam-Module-Megapixels-Arduino-Mega2560/dp/B012UXNDOY) + +### Hardware + +Connect the Arducam pins as follows: + +|Arducam pin name|Arduino pin name| +|----------------|----------------| +|CS|D7 (unlabelled, immediately to the right of D6)| +|MOSI|D11| +|MISO|D12| +|SCK|D13| +|GND|GND (either pin marked GND is fine)| +|VCC|3.3 V| +|SDA|A4| +|SCL|A5| + +### Install the Arduino_TensorFlowLite library + +Download the current nightly build of the library: +[person_detection.zip](https://storage.googleapis.com/download.tensorflow.org/data/tf_lite_micro_person_data_int8_grayscale_2020_01_13.zip) + +This example application is included as part of the official TensorFlow Lite +Arduino library. To install it, open the Arduino library manager in +`Tools -> Manage Libraries...` and search for `Arduino_TensorFlowLite`. + +### Install other libraries + +In addition to the TensorFlow library, you'll also need to install two +libraries: + +* The Arducam library, so our code can interface with the hardware +* The JPEGDecoder library, so we can decode JPEG-encoded images + +The Arducam Arduino library is available from GitHub at +[https://github.com/ArduCAM/Arduino](https://github.com/ArduCAM/Arduino). +To install it, download or clone the repository. Next, copy its `ArduCAM` +subdirectory into your `Arduino/libraries` directory. To find this directory on +your machine, check the *Sketchbook location* in the Arduino IDE's +*Preferences* window. + +After downloading the library, you'll need to edit one of its files to make sure +it is configured for the Arducam Mini 2MP Plus. To do so, open the following +file: + +``` +Arduino/libraries/ArduCAM/memorysaver.h +``` + +You'll see a bunch of `#define` statements listed. Make sure that they are all +commented out, except for `#define OV2640_MINI_2MP_PLUS`, as so: + +``` +//Step 1: select the hardware platform, only one at a time +//#define OV2640_MINI_2MP +//#define OV3640_MINI_3MP +//#define OV5642_MINI_5MP +//#define OV5642_MINI_5MP_BIT_ROTATION_FIXED +#define OV2640_MINI_2MP_PLUS +//#define OV5642_MINI_5MP_PLUS +//#define OV5640_MINI_5MP_PLUS +``` + +Once you save the file, we're done configuring the Arducam library. + +Our next step is to install the JPEGDecoder library. We can do this from within +the Arduino IDE. First, go to the *Manage Libraries...* option in the *Tools* +menu and search for `JPEGDecoder`. You should install version _1.8.0_ of the +library. + +Once the library has installed, we'll need to configure it to disable some +optional components that are not compatible with the Arduino Nano 33 BLE Sense. +Open the following file: + +``` +Arduino/libraries/JPEGDecoder/src/User_Config.h +``` + +Make sure that both `#define LOAD_SD_LIBRARY` and `#define LOAD_SDFAT_LIBRARY` +are commented out, as shown in this excerpt from the file: + +```c++ +// Comment out the next #defines if you are not using an SD Card to store the JPEGs +// Commenting out the line is NOT essential but will save some FLASH space if +// SD Card access is not needed. Note: use of SdFat is currently untested! + +//#define LOAD_SD_LIBRARY // Default SD Card library +//#define LOAD_SDFAT_LIBRARY // Use SdFat library instead, so SD Card SPI can be bit bashed +``` + +Once you've saved the file, you are done installing libraries. + +### Load and run the example + +Go to `File -> Examples`. You should see an +example near the bottom of the list named `TensorFlowLite`. Select +it and click `person_detection` to load the example. Connect your device, then +build and upload the example. + +To test the camera, start by pointing the device's camera at something that is +definitely not a person, or just covering it up. The next time the blue LED +flashes, the device will capture a frame from the camera and begin to run +inference. Since the vision model we are using for person detection is +relatively large, it takes a long time to run inference—around 19 seconds at the +time of writing, though it's possible TensorFlow Lite has gotten faster since +then. + +After 19 seconds or so, the inference result will be translated into another LED +being lit. Since you pointed the camera at something that isn't a person, the +red LED should light up. + +Now, try pointing the device's camera at yourself! The next time the blue LED +flashes, the device will capture another image and begin to run inference. After +19 seconds, the green LED should light up! + +Remember, image data is captured as a snapshot before each inference, whenever +the blue LED flashes. Whatever the camera is pointed at during that moment is +what will be fed into the model. It doesn't matter where the camera is pointed +until the next time an image is captured, when the blue LED will flash again. + +If you're getting seemingly incorrect results, make sure you are in an +environment with good lighting. You should also make sure that the camera is +oriented correctly, with the pins pointing downwards, so that the images it +captures are the right way up—the model was not trained to recognize upside-down +people! In addition, it's good to remember that this is a tiny model, which +trades accuracy for small size. It works very well, but it isn't accurate 100% +of the time. + +We can also see the results of inference via the Arduino Serial Monitor. To do +this, open the *Serial Monitor* from the *Tools* menu. You'll see a detailed +log of what is happening while our application runs. It's also interesting to +check the *Show timestamp* box, so you can see how long each part of the process +takes: + +``` +14:17:50.714 -> Starting capture +14:17:50.714 -> Image captured +14:17:50.784 -> Reading 3080 bytes from ArduCAM +14:17:50.887 -> Finished reading +14:17:50.887 -> Decoding JPEG and converting to greyscale +14:17:51.074 -> Image decoded and processed +14:18:09.710 -> Person score: 246 No person score: 66 +``` + +From the log, we can see that it took around 170 ms to capture and read the +image data from the camera module, 180 ms to decode the JPEG and convert it to +greyscale, and 18.6 seconds to run inference. + +## Running on HIMAX WE1 EVB + +The following instructions will help you build and deploy this example to +[HIMAX WE1 EVB](https://github.com/HimaxWiseEyePlus/bsp_tflu/tree/master/HIMAX_WE1_EVB_board_brief) +board. To undstand more about using this board, please check +[HIMAX WE1 EVB user guide](https://github.com/HimaxWiseEyePlus/bsp_tflu/tree/master/HIMAX_WE1_EVB_user_guide). + +### Initial Setup + +To use the HIMAX WE1 EVB, please make sure following software are installed: + +#### MetaWare Development Toolkit + +See +[Install the Synopsys DesignWare ARC MetaWare Development Toolkit](/tensorflow/lite/micro/tools/make/targets/arc/README.md#install-the-synopsys-designware-arc-metaware-development-toolkit) +section for instructions on toolchain installation. + +#### Make Tool version + +A `'make'` tool is required for deploying Tensorflow Lite Micro +applications on HIMAX WE1 EVB, See +[Check make tool version](/tensorflow/lite/micro/tools/make/targets/arc/README.md#make-tool) +section for proper environment. + +#### Serial Terminal Emulation Application + +There are 2 main purposes for HIMAX WE1 EVB Debug UART port + +- print application output +- burn application to flash by using xmodem send application binary + +You can use any terminal emulation program (like [PuTTY](https://www.putty.org/) or [minicom](https://linux.die.net/man/1/minicom)). + + +### Generate Example Project + +The example project for HIMAX WE1 EVB platform can be generated with the following +command: + +Download related third party data + +``` +make -f tensorflow/lite/micro/tools/make/Makefile TARGET=himax_we1_evb third_party_downloads +``` + +Generate person detection project + +``` +make -f tensorflow/lite/micro/tools/make/Makefile generate_person_detection_int8_make_project TARGET=himax_we1_evb +``` + +### Build and Burn Example + +Following the Steps to run person detection example at HIMAX WE1 EVB platform. + +1. Go to the generated example project directory. + + ``` + cd tensorflow/lite/micro/tools/make/gen/himax_we1_evb_arc/prj/person_detection_int8/make + ``` + +2. Build the example using + + ``` + make app + ``` + +3. After example build finish, copy ELF file and map file to image generate tool directory. + image generate tool directory located at `'tensorflow/lite/micro/tools/make/downloads/himax_we1_sdk/image_gen_linux_v3/'` + + ``` + cp person_detection_int8.elf himax_we1_evb.map ../../../../../downloads/himax_we1_sdk/image_gen_linux_v3/ + ``` + +4. Go to flash image generate tool directory. + + ``` + cd ../../../../../downloads/himax_we1_sdk/image_gen_linux_v3/ + ``` + +5. run image generate tool, generate flash image file. + + * Before running image generate tool, by typing `sudo chmod +x image_gen` + and `sudo chmod +x sign_tool` to make sure it is executable. + + ``` + image_gen -e person_detection_int8.elf -m himax_we1_evb.map -o out.img + ``` + + +6. Download flash image file to HIMAX WE1 EVB by UART: + + * more detail about download image through UART can be found at [HIMAX WE1 EVB update Flash image](https://github.com/HimaxWiseEyePlus/bsp_tflu/tree/master/HIMAX_WE1_EVB_user_guide#flash-image-update) + +After these steps, press reset button on the HIMAX WE1 EVB, you will see application output in the serial +terminal. + +## Running on SparkFun Edge + +The following instructions will help you build and deploy this sample on the +[SparkFun Edge development board](https://sparkfun.com/products/15170). This +sample requires the Sparkfun Himax camera for the Sparkfun Edge board. It is +not available for purchase yet. + +If you're new to using this board, we recommend walking through the +[AI on a microcontroller with TensorFlow Lite and SparkFun Edge](https://codelabs.developers.google.com/codelabs/sparkfun-tensorflow) +codelab to get an understanding of the workflow. + +### Compile the binary + +The following command will download the required dependencies and then compile a +binary for the SparkFun Edge: + +``` +make -f tensorflow/lite/micro/tools/make/Makefile TARGET=sparkfun_edge person_detection_bin +``` + +The binary will be created in the following location: + +``` +tensorflow/lite/micro/tools/make/gen/sparkfun_edge_cortex-m4/bin/person_detection.bin +``` + +### Sign the binary + +The binary must be signed with cryptographic keys to be deployed to the device. +We'll now run some commands that will sign our binary so it can be flashed to +the SparkFun Edge. The scripts we are using come from the Ambiq SDK, which is +downloaded when the `Makefile` is run. + +Enter the following command to set up some dummy cryptographic keys we can use +for development: + +``` +cp tensorflow/lite/micro/tools/make/downloads/AmbiqSuite-Rel2.0.0/tools/apollo3_scripts/keys_info0.py \ +tensorflow/lite/micro/tools/make/downloads/AmbiqSuite-Rel2.0.0/tools/apollo3_scripts/keys_info.py +``` + +Next, run the following command to create a signed binary: + +``` +python3 tensorflow/lite/micro/tools/make/downloads/AmbiqSuite-Rel2.0.0/tools/apollo3_scripts/create_cust_image_blob.py \ +--bin tensorflow/lite/micro/tools/make/gen/sparkfun_edge_cortex-m4/bin/person_detection.bin \ +--load-address 0xC000 \ +--magic-num 0xCB \ +-o main_nonsecure_ota \ +--version 0x0 +``` + +This will create the file `main_nonsecure_ota.bin`. We'll now run another +command to create a final version of the file that can be used to flash our +device with the bootloader script we will use in the next step: + +``` +python3 tensorflow/lite/micro/tools/make/downloads/AmbiqSuite-Rel2.0.0/tools/apollo3_scripts/create_cust_wireupdate_blob.py \ +--load-address 0x20000 \ +--bin main_nonsecure_ota.bin \ +-i 6 \ +-o main_nonsecure_wire \ +--options 0x1 +``` + +You should now have a file called `main_nonsecure_wire.bin` in the directory +where you ran the commands. This is the file we'll be flashing to the device. + +### Flash the binary + +Next, attach the board to your computer via a USB-to-serial adapter. + +**Note:** If you're using the [SparkFun Serial Basic Breakout](https://www.sparkfun.com/products/15096), +you should [install the latest drivers](https://learn.sparkfun.com/tutorials/sparkfun-serial-basic-ch340c-hookup-guide#drivers-if-you-need-them) +before you continue. + +Once connected, assign the USB device name to an environment variable: + +``` +export DEVICENAME=put your device name here +``` + +Set another variable with the baud rate: + +``` +export BAUD_RATE=921600 +``` + +Now, hold the button marked `14` on the device. While still holding the button, +hit the button marked `RST`. Continue holding the button marked `14` while +running the following command: + +``` +python3 tensorflow/lite/micro/tools/make/downloads/AmbiqSuite-Rel2.0.0/tools/apollo3_scripts/uart_wired_update.py \ +-b ${BAUD_RATE} ${DEVICENAME} \ +-r 1 \ +-f main_nonsecure_wire.bin \ +-i 6 +``` + +You should see a long stream of output as the binary is flashed to the device. +Once you see the following lines, flashing is complete: + +``` +Sending Reset Command. +Done. +``` + +If you don't see these lines, flashing may have failed. Try running through the +steps in [Flash the binary](#flash-the-binary) again (you can skip over setting +the environment variables). If you continue to run into problems, follow the +[AI on a microcontroller with TensorFlow Lite and SparkFun Edge](https://codelabs.developers.google.com/codelabs/sparkfun-tensorflow) +codelab, which includes more comprehensive instructions for the flashing +process. + +The binary should now be deployed to the device. Hit the button marked `RST` to +reboot the board. You should see the device's four LEDs flashing in sequence. + +Debug information is logged by the board while the program is running. To view +it, establish a serial connection to the board using a baud rate of `115200`. +On OSX and Linux, the following command should work: + +``` +screen ${DEVICENAME} 115200 +``` + +To stop viewing the debug output with `screen`, hit `Ctrl+A`, immediately +followed by the `K` key, then hit the `Y` key. + +## Run the tests on a development machine + +To compile and test this example on a desktop Linux or MacOS machine, download +[the TensorFlow source code](https://github.com/tensorflow/tensorflow), `cd` +into the source directory from a terminal, and then run the following command: + +``` +make -f tensorflow/lite/micro/tools/make/Makefile +``` + +This will take a few minutes, and downloads frameworks the code uses like +[CMSIS](https://developer.arm.com/embedded/cmsis) and +[flatbuffers](https://google.github.io/flatbuffers/). Once that process has +finished, run: + +``` +make -f tensorflow/lite/micro/tools/make/Makefile test_person_detection_test +``` + +You should see a series of files get compiled, followed by some logging output +from a test, which should conclude with `~~~ALL TESTS PASSED~~~`. If you see +this, it means that a small program has been built and run that loads a trained +TensorFlow model, runs some example images through it, and got the expected +outputs. This particular test runs images with a and without a person in them, +and checks that the network correctly identifies them. + +To understand how TensorFlow Lite does this, you can look at the `TestInvoke()` +function in +[person_detection_test.cc](https://github.com/tensorflow/tensorflow/tree/master/tensorflow/lite/micro/examples/person_detection/person_detection_test.cc). +It's a fairly small amount of code, creating an interpreter, getting a handle to +a model that's been compiled into the program, and then invoking the interpreter +with the model and sample inputs. + +## Debugging image capture +When the sample is running, check the LEDs to determine whether the inference is +running correctly. If the red light is stuck on, it means there was an error +communicating with the camera. This is likely due to an incorrectly connected +or broken camera. + +During inference, the blue LED will toggle every time inference is complete. The +orange LED indicates that no person was found, and the green LED indicates a +person was found. The red LED should never turn on, since it indicates an error. + +In order to view the captured image, set the DUMP_IMAGE define in main.cc.  This +causes the board to log raw image info to the console. After the board has been +flashed and reset, dump the log to a text file: + + +``` +screen -L -Logfile ${DEVICENAME} 115200 +``` + +Next, run the raw to bitmap converter to view captured images: + +``` +python3 raw_to_bitmap.py -r GRAY -i +``` + +## Training your own model + +You can train your own model with some easy-to-use scripts. See +[training_a_model.md](training_a_model.md) for instructions. diff --git a/tensorflow/lite/micro/tools/make/third_party_downloads.inc~ b/tensorflow/lite/micro/tools/make/third_party_downloads.inc~ deleted file mode 100644 index 3c7ee1b64d2..00000000000 --- a/tensorflow/lite/micro/tools/make/third_party_downloads.inc~ +++ /dev/null @@ -1,86 +0,0 @@ -# Add URLs and MD5 checksums for third-party libraries here. - -GEMMLOWP_URL := "https://github.com/google/gemmlowp/archive/719139ce755a0f31cbf1c37f7f98adcc7fc9f425.zip" -GEMMLOWP_MD5 := "7e8191b24853d75de2af87622ad293ba" - -ifeq ($(HOST_OS),windows) - FLATBUFFERS_URL := "https://github.com/google/flatbuffers/archive/v1.12.0.zip" - FLATBUFFERS_MD5 := "a1afdbf114dec01a861c1b8c917d0fc7" -else - FLATBUFFERS_URL := "https://github.com/google/flatbuffers/archive/v1.12.0.tar.gz" - FLATBUFFERS_MD5 := "c62ffefb3d4548b127cca14ce047f16c" -endif - -ifeq ($(HOST_OS),osx) - GCC_EMBEDDED_URL := "https://developer.arm.com/-/media/Files/downloads/gnu-rm/7-2018q2/gcc-arm-none-eabi-7-2018-q2-update-mac.tar.bz2" - GCC_EMBEDDED_MD5 := "a66be9828cf3c57d7d21178e07cd8904" -else ifeq ($(HOST_OS),windows) - GCC_EMBEDDED_URL := "https://developer.arm.com/-/media/Files/downloads/gnu-rm/7-2018q2/gcc-arm-none-eabi-7-2018-q2-update-win32.zip" - GCC_EMBEDDED_MD5 := "bc8ae26d7c429f30d583a605a4bcf9bc" -else - GCC_EMBEDDED_URL := "https://developer.arm.com/-/media/Files/downloads/gnu-rm/7-2018q2/gcc-arm-none-eabi-7-2018-q2-update-linux.tar.bz2" - GCC_EMBEDDED_MD5 := "299ebd3f1c2c90930d28ab82e5d8d6c0" -endif - -LEON_BCC2_URL := "https://www.gaisler.com/anonftp/bcc2/bin/bcc-2.0.7-gcc-linux64.tar.xz" -LEON_BCC2_MD5 := "cdf78082be4882da2a92c9baa82fe765" - -TSIM_URL := "https://www.gaisler.com/anonftp/tsim/tsim-eval-2.0.63.tar.gz" -TSIM_MD5 := "afa0095d3ed989a949e1467f94e41d2f" - -CMSIS_URL := "https://github.com/ARM-software/CMSIS_5/archive/1150e71e07c79b538efd842aba5b210a31827ae5.zip" -CMSIS_MD5 := "e05f4222ef58825193910b41a0871dcb" - -AM_SDK_URL := "http://s3.asia.ambiqmicro.com/downloads/AmbiqSuite-Rel2.2.0.zip" -AM_SDK_MD5 := "7605fa2d4d97e6bb7a1190c92b66b597" -AM_SDK_DEST := AmbiqSuite-Rel2.2.0 - -SF_BSPS_URL := "https://github.com/sparkfun/SparkFun_Apollo3_AmbiqSuite_BSPs/archive/v0.0.7.zip" -SF_BSPS_MD5 := "34199f7e754735661d1c8a70a40ca7a3" -SF_BSPS_DEST := boards_sfe - -STM32_BARE_LIB_URL := "https://github.com/google/stm32_bare_lib/archive/c07d611fb0af58450c5a3e0ab4d52b47f99bc82d.zip" -STM32_BARE_LIB_MD5 := "282bff40d4d0b92278fd123a3b6e3123" - -ifeq ($(HOST_OS),osx) - RISCV_TOOLCHAIN_URL := "https://static.dev.sifive.com/dev-tools/riscv64-unknown-elf-gcc-8.1.0-2019.01.0-x86_64-apple-darwin.tar.gz" - RISCV_TOOLCHAIN_MD5 := "2ac2fa00618b9ab7fa0c7d0ec173de94" -else - RISCV_TOOLCHAIN_URL := "https://static.dev.sifive.com/dev-tools/riscv64-unknown-elf-gcc-20181030-x86_64-linux-ubuntu14.tar.gz" - RISCV_TOOLCHAIN_MD5="2366b7afe36a54dc94fb0ff8a0830934" -endif - -SIFIVE_FE310_LIB_URL := "https://github.com/sifive/freedom-e-sdk/archive/baeeb8fd497a99b3c141d7494309ec2e64f19bdf.zip" -SIFIVE_FE310_LIB_MD5 := "06ee24c4956f8e21670ab3395861fe64" - -KISSFFT_URL="https://github.com/mborgerding/kissfft/archive/v130.zip" -KISSFFT_MD5="438ba1fef5783cc5f5f201395cc477ca" - -RUY_URL="https://github.com/google/ruy/archive/1b313682ef8b8fc8ed08719c610d1c3503b016bf.zip" -RUY_MD5="2d54f058f8f7120dfc1ecee79dbf259e" - -CIFAR10_DATASET_URL="https://www.cs.toronto.edu/~kriz/cifar-10-binary.tar.gz" -CIFAR10_DATASET_MD5="c32a1d4ab5d03f1284b67883e8d87530" - -IMAGE_RECOGNITION_MODEL_URL := "https://storage.googleapis.com/download.tensorflow.org/models/tflite/cifar_image_recognition_model_2020_4_14.zip" -IMAGE_RECOGNITION_MODEL_MD5 := "2b886156e7ef4d6e53d0f1a4bc800e56" - -PERSON_MODEL_URL := "https://storage.googleapis.com/download.tensorflow.org/data/tf_lite_micro_person_data_grayscale_2019_11_21.zip" -PERSON_MODEL_MD5 := "fe2934bd0788f1dcc7af3f0a954542ab" - -PERSON_MODEL_INT8_URL := "https://storage.googleapis.com/download.tensorflow.org/data/tf_lite_micro_person_data_int8_grayscale_2020_01_13.zip" -PERSON_MODEL_INT8_MD5 := "8a7d2c70325f53136faea6dde517b8cc" - -EMBARC_MLI_URL := "https://github.com/foss-for-synopsys-dwc-arc-processors/embarc_mli/archive/58284867ca52d1f43b25045e8601999d7359d986.zip" -EMBARC_MLI_MD5 := "2bf4982a327fdaa9d475803ce014d1ef" - -EMBARC_MLI_PRE_COMPILED_URL := "https://github.com/foss-for-synopsys-dwc-arc-processors/embarc_mli/releases/download/Release_1.1_RC2/embARC_MLI_package.zip" -EMBARC_MLI_PRE_COMPILED_MD5 := "a95ff9e0370434484f14e7e4114327f6" - -XTENSA_HIFI4_URL :="https://github.com/foss-xtensa/nnlib-hifi4/raw/master/archive/xa_nnlib_04_07.zip" -XTENSA_HIFI4_MD5 :="f234764928f9a42901df33a27e118c8b" - -HIMAX_WE1_SDK_URL ="https://www.himax.com.tw/we-i/himax_we1_sdk_v02.zip" -HIMAX_WE1_SDK_MD5 ="5063c24d298fbcfe118163f3ccc43079" - - From 233f1d53f829f157788ed687c07698b12cc8e091 Mon Sep 17 00:00:00 2001 From: "902449@58880@bigcat_chen@ASIC" Date: Thu, 4 Jun 2020 18:16:32 +0800 Subject: [PATCH 046/867] remove temp readme in example directory --- .../micro/examples/hello_world/README.md~ | 595 ------------------ .../person_detection_experimental/README.md~ | 568 ----------------- 2 files changed, 1163 deletions(-) delete mode 100644 tensorflow/lite/micro/examples/hello_world/README.md~ delete mode 100644 tensorflow/lite/micro/examples/person_detection_experimental/README.md~ diff --git a/tensorflow/lite/micro/examples/hello_world/README.md~ b/tensorflow/lite/micro/examples/hello_world/README.md~ deleted file mode 100644 index 9c0a5e2306a..00000000000 --- a/tensorflow/lite/micro/examples/hello_world/README.md~ +++ /dev/null @@ -1,595 +0,0 @@ -# Hello World Example - -This example is designed to demonstrate the absolute basics of using [TensorFlow -Lite for Microcontrollers](https://www.tensorflow.org/lite/microcontrollers). -It includes the full end-to-end workflow of training a model, converting it for -use with TensorFlow Lite for Microcontrollers for running inference on a -microcontroller. - -The model is trained to replicate a `sine` function and generates a pattern of -data to either blink LEDs or control an animation, depending on the capabilities -of the device. - -![Animation on STM32F746](images/animation_on_STM32F746.gif) - -## Table of contents - -- [Hello World Example](#hello-world-example) - - [Table of contents](#table-of-contents) - - [Deploy to ARC EM SDP](#deploy-to-arc-em-sdp) - - [Initial Setup](#initial-setup) - - [Generate Example Project](#generate-example-project) - - [Build and Run Example](#build-and-run-example) - - [Deploy to Arduino](#deploy-to-arduino) - - [Install the Arduino_TensorFlowLite library](#install-the-arduinotensorflowlite-library) - - [Load and run the example](#load-and-run-the-example) - - [Deploy to ESP32](#deploy-to-esp32) - - [Install the ESP IDF](#install-the-esp-idf) - - [Generate the examples](#generate-the-examples) - - [Building the example](#building-the-example) - - [Load and run the example](#load-and-run-the-example-1) - - [Deploy to himax WE1 EVB](#deploy-to-himax-we1-evb) - - [Initial Setup](#initial-setup-1) - - [MetaWare Development Toolkit](#metaware-development-toolkit) - - [Make Tool version](#make-tool-version) - - [Serial Terminal Emulation Application](#serial-terminal-emulation-application) - - [Generate Example Project](#generate-example-project-1) - - [Build and Burn Example](#build-and-burn-example) - - [Deploy to SparkFun Edge](#deploy-to-sparkfun-edge) - - [Compile the binary](#compile-the-binary) - - [Sign the binary](#sign-the-binary) - - [Flash the binary](#flash-the-binary) - - [Deploy to STM32F746](#deploy-to-stm32f746) - - [Run the tests on a development machine](#run-the-tests-on-a-development-machine) - - [Train your own model](#train-your-own-model) - -## Deploy to ARC EM SDP - -The following instructions will help you to build and deploy this example to -[ARC EM SDP](https://www.synopsys.com/dw/ipdir.php?ds=arc-em-software-development-platform) -board. General information and instructions on using the board with TensorFlow -Lite Micro can be found in the common -[ARC targets description](/tensorflow/lite/micro/tools/make/targets/arc/README.md). - -### Initial Setup - -Follow the instructions on the -[ARC EM SDP Initial Setup](/tensorflow/lite/micro/tools/make/targets/arc/README.md#ARC-EM-Software-Development-Platform-ARC-EM-SDP) -to get and install all required tools for work with ARC EM SDP. - -### Generate Example Project - -The example project for ARC EM SDP platform can be generated with the following -command: - -``` -make -f tensorflow/lite/micro/tools/make/Makefile TARGET=arc_emsdp TAGS=no_arc_mli generate_hello_world_make_project -``` - -### Build and Run Example - -For more detailed information on building and running examples see the -appropriate sections of general descriptions of the -[ARC EM SDP usage with TFLM](/tensorflow/lite/micro/tools/make/targets/arc/README.md#ARC-EM-Software-Development-Platform-ARC-EM-SDP). -In the directory with generated project you can also find a -*README_ARC_EMSDP.md* file with instructions and options on building and -running. Here we only briefly mention main steps which are typically enough to -get it started. - -1. You need to - [connect the board](/tensorflow/lite/micro/tools/make/targets/arc/README.md#connect-the-board) - and open an serial connection. - -2. Go to the generated example project director - - ``` - cd tensorflow/lite/micro/tools/make/gen/arc_emsdp_arc/prj/hello_world/make - ``` - -3. Build the example using - - ``` - make app - ``` - -4. To generate artefacts for self-boot of example from the board use - - ``` - make flash - ``` - -5. To run application from the board using microSD card: - - * Copy the content of the created /bin folder into the root of microSD - card. Note that the card must be formatted as FAT32 with default cluster - size (but less than 32 Kbytes) - * Plug in the microSD card into the J11 connector. - * Push the RST button. If a red LED is lit beside RST button, push the CFG - button. - -6. If you have the MetaWare Debugger installed in your environment: - - * To run application from the console using it type `make run`. - * To stop the execution type `Ctrl+C` in the console several times. - -In both cases (step 5 and 6) you will see the application output in the serial -terminal. - -## Deploy to Arduino - -The following instructions will help you build and deploy this sample -to [Arduino](https://www.arduino.cc/) devices. - -![Animation on Arduino MKRZERO](images/animation_on_arduino_mkrzero.gif) - -The sample has been tested with the following devices: - -- [Arduino Nano 33 BLE Sense](https://store.arduino.cc/usa/nano-33-ble-sense-with-headers) -- [Arduino MKRZERO](https://store.arduino.cc/usa/arduino-mkrzero) - -The sample will use PWM to fade an LED on and off according to the model's -output. In the code, the `LED_BUILTIN` constant is used to specify the board's -built-in LED as the one being controlled. However, on some boards, this built-in -LED is not attached to a pin with PWM capabilities. In this case, the LED will -blink instead of fading. - -### Install the Arduino_TensorFlowLite library - -This example application is included as part of the official TensorFlow Lite -Arduino library. To install it, open the Arduino library manager in -`Tools -> Manage Libraries...` and search for `Arduino_TensorFlowLite`. - -### Load and run the example - -Once the library has been added, go to `File -> Examples`. You should see an -example near the bottom of the list named `TensorFlowLite:hello_world`. Select -it and click `hello_world` to load the example. - -Use the Arduino IDE to build and upload the example. Once it is running, -you should see the built-in LED on your device flashing. - -The Arduino Desktop IDE includes a plotter that we can use to display the sine -wave graphically. To view it, go to `Tools -> Serial Plotter`. You will see one -datapoint being logged for each inference cycle, expressed as a number between 0 -and 255. - -## Deploy to ESP32 - -The following instructions will help you build and deploy this sample -to [ESP32](https://www.espressif.com/en/products/hardware/esp32/overview) -devices using the [ESP IDF](https://github.com/espressif/esp-idf). - -The sample has been tested on ESP-IDF version 4.0 with the following devices: -- [ESP32-DevKitC](http://esp-idf.readthedocs.io/en/latest/get-started/get-started-devkitc.html) -- [ESP-EYE](https://github.com/espressif/esp-who/blob/master/docs/en/get-started/ESP-EYE_Getting_Started_Guide.md) - -### Install the ESP IDF - -Follow the instructions of the -[ESP-IDF get started guide](https://docs.espressif.com/projects/esp-idf/en/latest/get-started/index.html) -to setup the toolchain and the ESP-IDF itself. - -The next steps assume that the -[IDF environment variables are set](https://docs.espressif.com/projects/esp-idf/en/latest/get-started/index.html#step-4-set-up-the-environment-variables) : - - * The `IDF_PATH` environment variable is set - * `idf.py` and Xtensa-esp32 tools (e.g. `xtensa-esp32-elf-gcc`) are in `$PATH` - -### Generate the examples -The example project can be generated with the following command: -``` -make -f tensorflow/lite/micro/tools/make/Makefile TARGET=esp generate_hello_world_esp_project -``` - -### Building the example - -Go the the example project directory -``` -cd tensorflow/lite/micro/tools/make/gen/esp_xtensa-esp32/prj/hello_world/esp-idf -``` - -Then build with `idf.py` -``` -idf.py build -``` - -### Load and run the example - -To flash (replace `/dev/ttyUSB0` with the device serial port): -``` -idf.py --port /dev/ttyUSB0 flash -``` - -Monitor the serial output: -``` -idf.py --port /dev/ttyUSB0 monitor -``` - -Use `Ctrl+]` to exit. - -The previous two commands can be combined: -``` -idf.py --port /dev/ttyUSB0 flash monitor -``` - -## Deploy to himax WE1 EVB - -The following instructions will help you build and deploy this example to -[HIMAX WE1 EVB](https://github.com/HimaxWiseEyePlus/bsp_tflu/tree/master/HIMAX_WE1_EVB_board_brief) -board. To undstand more about using this board, please check -[HIMAX WE1 EVB user guide](https://github.com/HimaxWiseEyePlus/bsp_tflu/tree/master/HIMAX_WE1_EVB_user_guide). - -### Initial Setup - -To use the HIMAX WE1 EVB, please make sure following software are installed: - -#### MetaWare Development Toolkit - -See -[Install the Synopsys DesignWare ARC MetaWare Development Toolkit](/tensorflow/lite/micro/tools/make/targets/arc/README.md#install-the-synopsys-designware-arc-metaware-development-toolkit) -section for instructions on toolchain installation. - -#### Make Tool version - -A `'make'` tool is required for deploying Tensorflow Lite Micro -applications on HIMAX WE1 EVB, See -[Check make tool version](/tensorflow/lite/micro/tools/make/targets/arc/README.md#make-tool) -section for proper environment. - -#### Serial Terminal Emulation Application - -There are 2 main purposes for HIMAX WE1 EVB Debug UART port - -- print application output -- burn application to flash by using xmodem send application binary - -You can use any terminal emulation program (like [PuTTY](https://www.putty.org/) or [minicom](https://linux.die.net/man/1/minicom)). - - -### Generate Example Project - -The example project for HIMAX WE1 EVB platform can be generated with the following -command: - -Download related third party data - -``` -make -f tensorflow/lite/micro/tools/make/Makefile TARGET=himax_we1_evb third_party_downloads -``` - -Generate hello world project - -``` -make -f tensorflow/lite/micro/tools/make/Makefile generate_hello_world_make_project TARGET=himax_we1_evb TAGS=no_arc_mli -``` - -### Build and Burn Example - -Following the Steps to run hello world example at HIMAX WE1 EVB platform. - -1. Go to the generated example project directory. - - ``` - cd tensorflow/lite/micro/tools/make/gen/himax_we1_evb_arc/prj/hello_world/make - ``` - -2. Build the example using - - ``` - make app - ``` - -3. After example build finish, copy ELF file and map file to image generate tool directory. - image generate tool directory located at `'tensorflow/lite/micro/tools/make/downloads/himax_we1_sdk/image_gen_linux_v3/'` - - ``` - cp hello_world.elf himax_we1_evb.map ../../../../../downloads/himax_we1_sdk/image_gen_linux_v3/ - ``` - -4. Go to flash image generate tool directory. - - ``` - cd ../../../../../downloads/himax_we1_sdk/image_gen_linux_v3/ - ``` - -5. run image generate tool, generate flash image file. - - * Before running image generate tool, by typing `sudo chmod +x image_gen` - and `sudo chmod +x sign_tool` to make sure it is executable. - - ``` - image_gen -e hello_world.elf -m himax_we1_evb.map -o out.img - ``` - - -6. Download flash image file to HIMAX WE1 EVB by UART: - - * more detail about download image through UART can be found at [HIMAX WE1 EVB update Flash image](https://github.com/HimaxWiseEyePlus/bsp_tflu/tree/master/HIMAX_WE1_EVB_user_guide#flash-image-update) - -After these steps, press reset button on the HIMAX WE1 EVB, you will see application output in the serial -terminal. - -## Deploy to SparkFun Edge - -The following instructions will help you build and deploy this sample on the -[SparkFun Edge development board](https://sparkfun.com/products/15170). - -![Animation on SparkFun Edge](images/animation_on_sparkfun_edge.gif) - -If you're new to using this board, we recommend walking through the -[AI on a microcontroller with TensorFlow Lite and SparkFun Edge](https://codelabs.developers.google.com/codelabs/sparkfun-tensorflow) -codelab to get an understanding of the workflow. - -### Compile the binary - -The following command will download the required dependencies and then compile a -binary for the SparkFun Edge: - -``` -make -f tensorflow/lite/micro/tools/make/Makefile TARGET=sparkfun_edge hello_world_bin -``` - -The binary will be created in the following location: - -``` -tensorflow/lite/micro/tools/make/gen/sparkfun_edge_cortex-m4/bin/hello_world.bin -``` - -### Sign the binary - -The binary must be signed with cryptographic keys to be deployed to the device. -We'll now run some commands that will sign our binary so it can be flashed to -the SparkFun Edge. The scripts we are using come from the Ambiq SDK, which is -downloaded when the `Makefile` is run. - -Enter the following command to set up some dummy cryptographic keys we can use -for development: - -``` -cp tensorflow/lite/micro/tools/make/downloads/AmbiqSuite-Rel2.2.0/tools/apollo3_scripts/keys_info0.py \ -tensorflow/lite/micro/tools/make/downloads/AmbiqSuite-Rel2.2.0/tools/apollo3_scripts/keys_info.py -``` - -Next, run the following command to create a signed binary: - -``` -python3 tensorflow/lite/micro/tools/make/downloads/AmbiqSuite-Rel2.2.0/tools/apollo3_scripts/create_cust_image_blob.py \ ---bin tensorflow/lite/micro/tools/make/gen/sparkfun_edge_cortex-m4/bin/hello_world.bin \ ---load-address 0xC000 \ ---magic-num 0xCB \ --o main_nonsecure_ota \ ---version 0x0 -``` - -This will create the file `main_nonsecure_ota.bin`. We'll now run another -command to create a final version of the file that can be used to flash our -device with the bootloader script we will use in the next step: - -``` -python3 tensorflow/lite/micro/tools/make/downloads/AmbiqSuite-Rel2.2.0/tools/apollo3_scripts/create_cust_wireupdate_blob.py \ ---load-address 0x20000 \ ---bin main_nonsecure_ota.bin \ --i 6 \ --o main_nonsecure_wire \ ---options 0x1 -``` - -You should now have a file called `main_nonsecure_wire.bin` in the directory -where you ran the commands. This is the file we'll be flashing to the device. - -### Flash the binary - -Next, attach the board to your computer via a USB-to-serial adapter. - -**Note:** If you're using the [SparkFun Serial Basic Breakout](https://www.sparkfun.com/products/15096), -you should [install the latest drivers](https://learn.sparkfun.com/tutorials/sparkfun-serial-basic-ch340c-hookup-guide#drivers-if-you-need-them) -before you continue. - -Once connected, assign the USB device name to an environment variable: - -``` -export DEVICENAME=put your device name here -``` - -Set another variable with the baud rate: - -``` -export BAUD_RATE=921600 -``` - -Now, hold the button marked `14` on the device. While still holding the button, -hit the button marked `RST`. Continue holding the button marked `14` while -running the following command: - -``` -python3 tensorflow/lite/micro/tools/make/downloads/AmbiqSuite-Rel2.2.0/tools/apollo3_scripts/uart_wired_update.py \ --b ${BAUD_RATE} ${DEVICENAME} \ --r 1 \ --f main_nonsecure_wire.bin \ --i 6 -``` - -You should see a long stream of output as the binary is flashed to the device. -Once you see the following lines, flashing is complete: - -``` -Sending Reset Command. -Done. -``` - -If you don't see these lines, flashing may have failed. Try running through the -steps in [Flash the binary](#flash-the-binary) again (you can skip over setting -the environment variables). If you continue to run into problems, follow the -[AI on a microcontroller with TensorFlow Lite and SparkFun Edge](https://codelabs.developers.google.com/codelabs/sparkfun-tensorflow) -codelab, which includes more comprehensive instructions for the flashing -process. - -The binary should now be deployed to the device. Hit the button marked `RST` to -reboot the board. You should see the device's four LEDs flashing in sequence. - -Debug information is logged by the board while the program is running. To view -it, establish a serial connection to the board using a baud rate of `115200`. -On OSX and Linux, the following command should work: - -``` -screen ${DEVICENAME} 115200 -``` - -You will see a lot of output flying past! To stop the scrolling, hit `Ctrl+A`, -immediately followed by `Esc`. You can then use the arrow keys to explore the -output, which will contain the results of running inference on various `x` -values: - -``` -x_value: 1.1843798*2^2, y_value: -1.9542645*2^-1 -``` - -To stop viewing the debug output with `screen`, hit `Ctrl+A`, immediately -followed by the `K` key, then hit the `Y` key. - - -## Deploy to STM32F746 - -The following instructions will help you build and deploy the sample to the -[STM32F7 discovery kit](https://os.mbed.com/platforms/ST-Discovery-F746NG/) -using [ARM Mbed](https://github.com/ARMmbed/mbed-cli). - -![Animation on STM32F746](images/animation_on_STM32F746.gif) - -Before we begin, you'll need the following: - -- STM32F7 discovery kit board -- Mini-USB cable -- ARM Mbed CLI ([installation instructions](https://os.mbed.com/docs/mbed-os/v5.12/tools/installation-and-setup.html)) -- Python 2.7 and pip - -Since Mbed requires a special folder structure for projects, we'll first run a -command to generate a subfolder containing the required source files in this -structure: - -``` -make -f tensorflow/lite/micro/tools/make/Makefile TARGET=mbed TAGS="CMSIS disco_f746ng" generate_hello_world_mbed_project -``` - -This will result in the creation of a new folder: - -``` -tensorflow/lite/micro/tools/make/gen/mbed_cortex-m4/prj/hello_world/mbed -``` - -This folder contains all of the example's dependencies structured in the correct -way for Mbed to be able to build it. - -Change into the directory and run the following commands, making sure you are -using Python 2.7.15. - -First, tell Mbed that the current directory is the root of an Mbed project: - -``` -mbed config root . -``` - -Next, tell Mbed to download the dependencies and prepare to build: - -``` -mbed deploy -``` - -By default, Mbed will build the project using C++98. However, TensorFlow Lite -requires C++11. Run the following Python snippet to modify the Mbed -configuration files so that it uses C++11: - -``` -python -c 'import fileinput, glob; -for filename in glob.glob("mbed-os/tools/profiles/*.json"): - for line in fileinput.input(filename, inplace=True): - print line.replace("\"-std=gnu++98\"","\"-std=c++11\", \"-fpermissive\"")' - -``` - -Finally, run the following command to compile: - -``` -mbed compile -m DISCO_F746NG -t GCC_ARM -``` - -This should result in a binary at the following path: - -``` -./BUILD/DISCO_F746NG/GCC_ARM/mbed.bin -``` - -To deploy, plug in your STM board and copy the file to it. On MacOS, you can do -this with the following command: - -``` -cp ./BUILD/DISCO_F746NG/GCC_ARM/mbed.bin /Volumes/DIS_F746NG/ -``` - -Copying the file will initiate the flashing process. Once this is complete, you -should see an animation on the device's screen. - - -``` -screen /dev/tty.usbmodem14403 9600 -``` - -In addition to this animation, debug information is logged by the board while -the program is running. To view it, establish a serial connection to the board -using a baud rate of `9600`. On OSX and Linux, the following command should -work, replacing `/dev/tty.devicename` with the name of your device as it appears -in `/dev`: - -``` -screen /dev/tty.devicename 9600 -``` - -You will see a lot of output flying past! To stop the scrolling, hit `Ctrl+A`, -immediately followed by `Esc`. You can then use the arrow keys to explore the -output, which will contain the results of running inference on various `x` -values: - -``` -x_value: 1.1843798*2^2, y_value: -1.9542645*2^-1 -``` - -To stop viewing the debug output with `screen`, hit `Ctrl+A`, immediately -followed by the `K` key, then hit the `Y` key. - -### Run the tests on a development machine - -To compile and test this example on a desktop Linux or macOS machine, first -clone the TensorFlow repository from GitHub to a convenient place: - -```bash -git clone --depth 1 https://github.com/tensorflow/tensorflow.git -``` - -Next, `cd` into the source directory from a terminal, and then run the following -command: - -```bash -make -f tensorflow/lite/micro/tools/make/Makefile test_hello_world_test -``` - -This will take a few minutes, and downloads frameworks the code uses. Once the -process has finished, you should see a series of files get compiled, followed by -some logging output from a test, which should conclude with -`~~~ALL TESTS PASSED~~~`. - -If you see this, it means that a small program has been built and run that loads -the trained TensorFlow model, runs some example inputs through it, and got the -expected outputs. - -To understand how TensorFlow Lite does this, you can look at the source in -[hello_world_test.cc](hello_world_test.cc). -It's a fairly small amount of code that creates an interpreter, gets a handle to -a model that's been compiled into the program, and then invokes the interpreter -with the model and sample inputs. - -### Train your own model - -So far you have used an existing trained model to run inference on -microcontrollers. If you wish to train your own model, follow the instructions -given in the [train/](train/) directory. - diff --git a/tensorflow/lite/micro/examples/person_detection_experimental/README.md~ b/tensorflow/lite/micro/examples/person_detection_experimental/README.md~ deleted file mode 100644 index 4d53e551431..00000000000 --- a/tensorflow/lite/micro/examples/person_detection_experimental/README.md~ +++ /dev/null @@ -1,568 +0,0 @@ -# Person detection example - -This example shows how you can use Tensorflow Lite to run a 250 kilobyte neural -network to recognize people in images captured by a camera. It is designed to -run on systems with small amounts of memory such as microcontrollers and DSPs. -This uses the experimental int8 quantized version of the person detection model. - -## Table of contents - -- [Person detection example](#person-detection-example) - - [Table of contents](#table-of-contents) - - [Running on ARC EM SDP](#running-on-arc-em-sdp) - - [Initial setup](#initial-setup) - - [Generate Example Project](#generate-example-project) - - [Build and Run Example](#build-and-run-example) - - [Running on Arduino](#running-on-arduino) - - [Hardware](#hardware) - - [Install the Arduino_TensorFlowLite library](#install-the-arduinotensorflowlite-library) - - [Install other libraries](#install-other-libraries) - - [Load and run the example](#load-and-run-the-example) - - [Running on HIMAX WE1 EVB](#running-on-himax-we1-evb) - - [Initial Setup](#initial-setup) - - [MetaWare Development Toolkit](#metaware-development-toolkit) - - [Make Tool version](#make-tool-version) - - [Serial Terminal Emulation Application](#serial-terminal-emulation-application) - - [Generate Example Project](#generate-example-project-1) - - [Build and Burn Example](#build-and-burn-example) - - [Running on SparkFun Edge](#running-on-sparkfun-edge) - - [Compile the binary](#compile-the-binary) - - [Sign the binary](#sign-the-binary) - - [Flash the binary](#flash-the-binary) - - [Run the tests on a development machine](#run-the-tests-on-a-development-machine) - - [Debugging image capture](#debugging-image-capture) - - [Training your own model](#training-your-own-model) - -## Running on ARC EM SDP - -The following instructions will help you to build and deploy this example to -[ARC EM SDP](https://www.synopsys.com/dw/ipdir.php?ds=arc-em-software-development-platform) -board. General information and instructions on using the board with TensorFlow -Lite Micro can be found in the common -[ARC targets description](/tensorflow/lite/micro/tools/make/targets/arc/README.md). - -This example uses asymmetric int8 quantization and can therefore leverage -optimized int8 kernels from the embARC MLI library - -The ARC EM SDP board contains a rich set of extension interfaces. You can choose -any compatible camera and modify -[image_provider.cc](/tensorflow/lite/micro/examples/person_detection_experimental/image_provider.cc) -file accordingly to use input from your specific camera. By default, results of -running this example are printed to the console. If you would like to instead -implement some target-specific actions, you need to modify -[detection_responder.cc](/tensorflow/lite/micro/examples/person_detection_experimental/detection_responder.cc) -accordingly. - -The reference implementations of these files are used by default on the EM SDP. - -### Initial setup - -Follow the instructions on the -[ARC EM SDP Initial Setup](/tensorflow/lite/micro/tools/make/targets/arc/README.md#ARC-EM-Software-Development-Platform-ARC-EM-SDP) -to get and install all required tools for work with ARC EM SDP. - -### Generate Example Project - -The example project for ARC EM SDP platform can be generated with the following -command: - -``` -make -f tensorflow/lite/micro/tools/make/Makefile TARGET=arc_emsdp generate_person_detection_int8_make_project -``` - -### Build and Run Example - -For more detailed information on building and running examples see the -appropriate sections of general descriptions of the -[ARC EM SDP usage with TFLM](/tensorflow/lite/micro/tools/make/targets/arc/README.md#ARC-EM-Software-Development-Platform-ARC-EM-SDP). -In the directory with generated project you can also find a -*README_ARC_EMSDP.md* file with instructions and options on building and -running. Here we only briefly mention main steps which are typically enough to -get it started. - -1. You need to - [connect the board](/tensorflow/lite/micro/tools/make/targets/arc/README.md#connect-the-board) - and open an serial connection. - -2. Go to the generated example project director - - ``` - cd tensorflow/lite/micro/tools/make/gen/arc_emsdp_arc/prj/person_detection_int8/make - ``` - -3. Build the example using - - ``` - make app - ``` - -4. To generate artefacts for self-boot of example from the board use - - ``` - make flash - ``` - -5. To run application from the board using microSD card: - - * Copy the content of the created /bin folder into the root of microSD - card. Note that the card must be formatted as FAT32 with default cluster - size (but less than 32 Kbytes) - * Plug in the microSD card into the J11 connector. - * Push the RST button. If a red LED is lit beside RST button, push the CFG - button. - -6. If you have the MetaWare Debugger installed in your environment: - - * To run application from the console using it type `make run`. - * To stop the execution type `Ctrl+C` in the console several times. - -In both cases (step 5 and 6) you will see the application output in the serial -terminal. - -## Running on Arduino - -The following instructions will help you build and deploy this sample -to [Arduino](https://www.arduino.cc/) devices. - -The sample has been tested with the following device: - -- [Arduino Nano 33 BLE Sense](https://store.arduino.cc/usa/nano-33-ble-sense-with-headers) - -You will also need the following camera module: - -- [Arducam Mini 2MP Plus](https://www.amazon.com/Arducam-Module-Megapixels-Arduino-Mega2560/dp/B012UXNDOY) - -### Hardware - -Connect the Arducam pins as follows: - -|Arducam pin name|Arduino pin name| -|----------------|----------------| -|CS|D7 (unlabelled, immediately to the right of D6)| -|MOSI|D11| -|MISO|D12| -|SCK|D13| -|GND|GND (either pin marked GND is fine)| -|VCC|3.3 V| -|SDA|A4| -|SCL|A5| - -### Install the Arduino_TensorFlowLite library - -Download the current nightly build of the library: -[person_detection.zip](https://storage.googleapis.com/download.tensorflow.org/data/tf_lite_micro_person_data_int8_grayscale_2020_01_13.zip) - -This example application is included as part of the official TensorFlow Lite -Arduino library. To install it, open the Arduino library manager in -`Tools -> Manage Libraries...` and search for `Arduino_TensorFlowLite`. - -### Install other libraries - -In addition to the TensorFlow library, you'll also need to install two -libraries: - -* The Arducam library, so our code can interface with the hardware -* The JPEGDecoder library, so we can decode JPEG-encoded images - -The Arducam Arduino library is available from GitHub at -[https://github.com/ArduCAM/Arduino](https://github.com/ArduCAM/Arduino). -To install it, download or clone the repository. Next, copy its `ArduCAM` -subdirectory into your `Arduino/libraries` directory. To find this directory on -your machine, check the *Sketchbook location* in the Arduino IDE's -*Preferences* window. - -After downloading the library, you'll need to edit one of its files to make sure -it is configured for the Arducam Mini 2MP Plus. To do so, open the following -file: - -``` -Arduino/libraries/ArduCAM/memorysaver.h -``` - -You'll see a bunch of `#define` statements listed. Make sure that they are all -commented out, except for `#define OV2640_MINI_2MP_PLUS`, as so: - -``` -//Step 1: select the hardware platform, only one at a time -//#define OV2640_MINI_2MP -//#define OV3640_MINI_3MP -//#define OV5642_MINI_5MP -//#define OV5642_MINI_5MP_BIT_ROTATION_FIXED -#define OV2640_MINI_2MP_PLUS -//#define OV5642_MINI_5MP_PLUS -//#define OV5640_MINI_5MP_PLUS -``` - -Once you save the file, we're done configuring the Arducam library. - -Our next step is to install the JPEGDecoder library. We can do this from within -the Arduino IDE. First, go to the *Manage Libraries...* option in the *Tools* -menu and search for `JPEGDecoder`. You should install version _1.8.0_ of the -library. - -Once the library has installed, we'll need to configure it to disable some -optional components that are not compatible with the Arduino Nano 33 BLE Sense. -Open the following file: - -``` -Arduino/libraries/JPEGDecoder/src/User_Config.h -``` - -Make sure that both `#define LOAD_SD_LIBRARY` and `#define LOAD_SDFAT_LIBRARY` -are commented out, as shown in this excerpt from the file: - -```c++ -// Comment out the next #defines if you are not using an SD Card to store the JPEGs -// Commenting out the line is NOT essential but will save some FLASH space if -// SD Card access is not needed. Note: use of SdFat is currently untested! - -//#define LOAD_SD_LIBRARY // Default SD Card library -//#define LOAD_SDFAT_LIBRARY // Use SdFat library instead, so SD Card SPI can be bit bashed -``` - -Once you've saved the file, you are done installing libraries. - -### Load and run the example - -Go to `File -> Examples`. You should see an -example near the bottom of the list named `TensorFlowLite`. Select -it and click `person_detection` to load the example. Connect your device, then -build and upload the example. - -To test the camera, start by pointing the device's camera at something that is -definitely not a person, or just covering it up. The next time the blue LED -flashes, the device will capture a frame from the camera and begin to run -inference. Since the vision model we are using for person detection is -relatively large, it takes a long time to run inference—around 19 seconds at the -time of writing, though it's possible TensorFlow Lite has gotten faster since -then. - -After 19 seconds or so, the inference result will be translated into another LED -being lit. Since you pointed the camera at something that isn't a person, the -red LED should light up. - -Now, try pointing the device's camera at yourself! The next time the blue LED -flashes, the device will capture another image and begin to run inference. After -19 seconds, the green LED should light up! - -Remember, image data is captured as a snapshot before each inference, whenever -the blue LED flashes. Whatever the camera is pointed at during that moment is -what will be fed into the model. It doesn't matter where the camera is pointed -until the next time an image is captured, when the blue LED will flash again. - -If you're getting seemingly incorrect results, make sure you are in an -environment with good lighting. You should also make sure that the camera is -oriented correctly, with the pins pointing downwards, so that the images it -captures are the right way up—the model was not trained to recognize upside-down -people! In addition, it's good to remember that this is a tiny model, which -trades accuracy for small size. It works very well, but it isn't accurate 100% -of the time. - -We can also see the results of inference via the Arduino Serial Monitor. To do -this, open the *Serial Monitor* from the *Tools* menu. You'll see a detailed -log of what is happening while our application runs. It's also interesting to -check the *Show timestamp* box, so you can see how long each part of the process -takes: - -``` -14:17:50.714 -> Starting capture -14:17:50.714 -> Image captured -14:17:50.784 -> Reading 3080 bytes from ArduCAM -14:17:50.887 -> Finished reading -14:17:50.887 -> Decoding JPEG and converting to greyscale -14:17:51.074 -> Image decoded and processed -14:18:09.710 -> Person score: 246 No person score: 66 -``` - -From the log, we can see that it took around 170 ms to capture and read the -image data from the camera module, 180 ms to decode the JPEG and convert it to -greyscale, and 18.6 seconds to run inference. - -## Running on HIMAX WE1 EVB - -The following instructions will help you build and deploy this example to -[HIMAX WE1 EVB](https://github.com/HimaxWiseEyePlus/bsp_tflu/tree/master/HIMAX_WE1_EVB_board_brief) -board. To undstand more about using this board, please check -[HIMAX WE1 EVB user guide](https://github.com/HimaxWiseEyePlus/bsp_tflu/tree/master/HIMAX_WE1_EVB_user_guide). - -### Initial Setup - -To use the HIMAX WE1 EVB, please make sure following software are installed: - -#### MetaWare Development Toolkit - -See -[Install the Synopsys DesignWare ARC MetaWare Development Toolkit](/tensorflow/lite/micro/tools/make/targets/arc/README.md#install-the-synopsys-designware-arc-metaware-development-toolkit) -section for instructions on toolchain installation. - -#### Make Tool version - -A `'make'` tool is required for deploying Tensorflow Lite Micro -applications on HIMAX WE1 EVB, See -[Check make tool version](/tensorflow/lite/micro/tools/make/targets/arc/README.md#make-tool) -section for proper environment. - -#### Serial Terminal Emulation Application - -There are 2 main purposes for HIMAX WE1 EVB Debug UART port - -- print application output -- burn application to flash by using xmodem send application binary - -You can use any terminal emulation program (like [PuTTY](https://www.putty.org/) or [minicom](https://linux.die.net/man/1/minicom)). - - -### Generate Example Project - -The example project for HIMAX WE1 EVB platform can be generated with the following -command: - -Download related third party data - -``` -make -f tensorflow/lite/micro/tools/make/Makefile TARGET=himax_we1_evb third_party_downloads -``` - -Generate person detection project - -``` -make -f tensorflow/lite/micro/tools/make/Makefile generate_person_detection_int8_make_project TARGET=himax_we1_evb -``` - -### Build and Burn Example - -Following the Steps to run person detection example at HIMAX WE1 EVB platform. - -1. Go to the generated example project directory. - - ``` - cd tensorflow/lite/micro/tools/make/gen/himax_we1_evb_arc/prj/person_detection_int8/make - ``` - -2. Build the example using - - ``` - make app - ``` - -3. After example build finish, copy ELF file and map file to image generate tool directory. - image generate tool directory located at `'tensorflow/lite/micro/tools/make/downloads/himax_we1_sdk/image_gen_linux_v3/'` - - ``` - cp person_detection_int8.elf himax_we1_evb.map ../../../../../downloads/himax_we1_sdk/image_gen_linux_v3/ - ``` - -4. Go to flash image generate tool directory. - - ``` - cd ../../../../../downloads/himax_we1_sdk/image_gen_linux_v3/ - ``` - -5. run image generate tool, generate flash image file. - - * Before running image generate tool, by typing `sudo chmod +x image_gen` - and `sudo chmod +x sign_tool` to make sure it is executable. - - ``` - image_gen -e person_detection_int8.elf -m himax_we1_evb.map -o out.img - ``` - - -6. Download flash image file to HIMAX WE1 EVB by UART: - - * more detail about download image through UART can be found at [HIMAX WE1 EVB update Flash image](https://github.com/HimaxWiseEyePlus/bsp_tflu/tree/master/HIMAX_WE1_EVB_user_guide#flash-image-update) - -After these steps, press reset button on the HIMAX WE1 EVB, you will see application output in the serial -terminal. - -## Running on SparkFun Edge - -The following instructions will help you build and deploy this sample on the -[SparkFun Edge development board](https://sparkfun.com/products/15170). This -sample requires the Sparkfun Himax camera for the Sparkfun Edge board. It is -not available for purchase yet. - -If you're new to using this board, we recommend walking through the -[AI on a microcontroller with TensorFlow Lite and SparkFun Edge](https://codelabs.developers.google.com/codelabs/sparkfun-tensorflow) -codelab to get an understanding of the workflow. - -### Compile the binary - -The following command will download the required dependencies and then compile a -binary for the SparkFun Edge: - -``` -make -f tensorflow/lite/micro/tools/make/Makefile TARGET=sparkfun_edge person_detection_bin -``` - -The binary will be created in the following location: - -``` -tensorflow/lite/micro/tools/make/gen/sparkfun_edge_cortex-m4/bin/person_detection.bin -``` - -### Sign the binary - -The binary must be signed with cryptographic keys to be deployed to the device. -We'll now run some commands that will sign our binary so it can be flashed to -the SparkFun Edge. The scripts we are using come from the Ambiq SDK, which is -downloaded when the `Makefile` is run. - -Enter the following command to set up some dummy cryptographic keys we can use -for development: - -``` -cp tensorflow/lite/micro/tools/make/downloads/AmbiqSuite-Rel2.0.0/tools/apollo3_scripts/keys_info0.py \ -tensorflow/lite/micro/tools/make/downloads/AmbiqSuite-Rel2.0.0/tools/apollo3_scripts/keys_info.py -``` - -Next, run the following command to create a signed binary: - -``` -python3 tensorflow/lite/micro/tools/make/downloads/AmbiqSuite-Rel2.0.0/tools/apollo3_scripts/create_cust_image_blob.py \ ---bin tensorflow/lite/micro/tools/make/gen/sparkfun_edge_cortex-m4/bin/person_detection.bin \ ---load-address 0xC000 \ ---magic-num 0xCB \ --o main_nonsecure_ota \ ---version 0x0 -``` - -This will create the file `main_nonsecure_ota.bin`. We'll now run another -command to create a final version of the file that can be used to flash our -device with the bootloader script we will use in the next step: - -``` -python3 tensorflow/lite/micro/tools/make/downloads/AmbiqSuite-Rel2.0.0/tools/apollo3_scripts/create_cust_wireupdate_blob.py \ ---load-address 0x20000 \ ---bin main_nonsecure_ota.bin \ --i 6 \ --o main_nonsecure_wire \ ---options 0x1 -``` - -You should now have a file called `main_nonsecure_wire.bin` in the directory -where you ran the commands. This is the file we'll be flashing to the device. - -### Flash the binary - -Next, attach the board to your computer via a USB-to-serial adapter. - -**Note:** If you're using the [SparkFun Serial Basic Breakout](https://www.sparkfun.com/products/15096), -you should [install the latest drivers](https://learn.sparkfun.com/tutorials/sparkfun-serial-basic-ch340c-hookup-guide#drivers-if-you-need-them) -before you continue. - -Once connected, assign the USB device name to an environment variable: - -``` -export DEVICENAME=put your device name here -``` - -Set another variable with the baud rate: - -``` -export BAUD_RATE=921600 -``` - -Now, hold the button marked `14` on the device. While still holding the button, -hit the button marked `RST`. Continue holding the button marked `14` while -running the following command: - -``` -python3 tensorflow/lite/micro/tools/make/downloads/AmbiqSuite-Rel2.0.0/tools/apollo3_scripts/uart_wired_update.py \ --b ${BAUD_RATE} ${DEVICENAME} \ --r 1 \ --f main_nonsecure_wire.bin \ --i 6 -``` - -You should see a long stream of output as the binary is flashed to the device. -Once you see the following lines, flashing is complete: - -``` -Sending Reset Command. -Done. -``` - -If you don't see these lines, flashing may have failed. Try running through the -steps in [Flash the binary](#flash-the-binary) again (you can skip over setting -the environment variables). If you continue to run into problems, follow the -[AI on a microcontroller with TensorFlow Lite and SparkFun Edge](https://codelabs.developers.google.com/codelabs/sparkfun-tensorflow) -codelab, which includes more comprehensive instructions for the flashing -process. - -The binary should now be deployed to the device. Hit the button marked `RST` to -reboot the board. You should see the device's four LEDs flashing in sequence. - -Debug information is logged by the board while the program is running. To view -it, establish a serial connection to the board using a baud rate of `115200`. -On OSX and Linux, the following command should work: - -``` -screen ${DEVICENAME} 115200 -``` - -To stop viewing the debug output with `screen`, hit `Ctrl+A`, immediately -followed by the `K` key, then hit the `Y` key. - -## Run the tests on a development machine - -To compile and test this example on a desktop Linux or MacOS machine, download -[the TensorFlow source code](https://github.com/tensorflow/tensorflow), `cd` -into the source directory from a terminal, and then run the following command: - -``` -make -f tensorflow/lite/micro/tools/make/Makefile -``` - -This will take a few minutes, and downloads frameworks the code uses like -[CMSIS](https://developer.arm.com/embedded/cmsis) and -[flatbuffers](https://google.github.io/flatbuffers/). Once that process has -finished, run: - -``` -make -f tensorflow/lite/micro/tools/make/Makefile test_person_detection_test -``` - -You should see a series of files get compiled, followed by some logging output -from a test, which should conclude with `~~~ALL TESTS PASSED~~~`. If you see -this, it means that a small program has been built and run that loads a trained -TensorFlow model, runs some example images through it, and got the expected -outputs. This particular test runs images with a and without a person in them, -and checks that the network correctly identifies them. - -To understand how TensorFlow Lite does this, you can look at the `TestInvoke()` -function in -[person_detection_test.cc](https://github.com/tensorflow/tensorflow/tree/master/tensorflow/lite/micro/examples/person_detection/person_detection_test.cc). -It's a fairly small amount of code, creating an interpreter, getting a handle to -a model that's been compiled into the program, and then invoking the interpreter -with the model and sample inputs. - -## Debugging image capture -When the sample is running, check the LEDs to determine whether the inference is -running correctly. If the red light is stuck on, it means there was an error -communicating with the camera. This is likely due to an incorrectly connected -or broken camera. - -During inference, the blue LED will toggle every time inference is complete. The -orange LED indicates that no person was found, and the green LED indicates a -person was found. The red LED should never turn on, since it indicates an error. - -In order to view the captured image, set the DUMP_IMAGE define in main.cc.  This -causes the board to log raw image info to the console. After the board has been -flashed and reset, dump the log to a text file: - - -``` -screen -L -Logfile ${DEVICENAME} 115200 -``` - -Next, run the raw to bitmap converter to view captured images: - -``` -python3 raw_to_bitmap.py -r GRAY -i -``` - -## Training your own model - -You can train your own model with some easy-to-use scripts. See -[training_a_model.md](training_a_model.md) for instructions. From 39e65d52e400f8c343195e2f8ac34f286648a415 Mon Sep 17 00:00:00 2001 From: "902449@58880@bigcat_chen@ASIC" Date: Thu, 4 Jun 2020 18:19:58 +0800 Subject: [PATCH 047/867] remove temp makefile in target --- .../make/targets/himax_we1_evb_makefile.inc~ | 93 ------------------- 1 file changed, 93 deletions(-) delete mode 100644 tensorflow/lite/micro/tools/make/targets/himax_we1_evb_makefile.inc~ diff --git a/tensorflow/lite/micro/tools/make/targets/himax_we1_evb_makefile.inc~ b/tensorflow/lite/micro/tools/make/targets/himax_we1_evb_makefile.inc~ deleted file mode 100644 index 733f258fbbb..00000000000 --- a/tensorflow/lite/micro/tools/make/targets/himax_we1_evb_makefile.inc~ +++ /dev/null @@ -1,93 +0,0 @@ -# Settings for himax WE_1 evb. -ifeq ($(TARGET), himax_we1_evb) - - CC_TOOL = ccac - AR_TOOL = arac - CXX_TOOL = ccac - LD_TOOL := ccac - TARGET_ARCH := arc - #ARC_TOOLCHAIN := mwdt - - BUILD_ARC_MLI := false - ARC_MLI_PRE_COMPILED_TARGET := himax_arcem9d_r16 - -include $(MAKEFILE_DIR)/targets/arc/arc_common.inc - #download SDK & MLI - HIMAX_WE1_SDK_NAME := himax_we1_sdk - #MLI_LIB_DIR = arc_mli_package - #MLI_LIB_DIR = arc_mli_package - #$(eval $(call add_third_party_download,$(EMBARC_MLI_PRE_COMPILED_URL),$(EMBARC_MLI_PRE_COMPILED_MD5),$(MLI_LIB_DIR),)) - $(eval $(call add_third_party_download,$(HIMAX_WE1_SDK_URL),$(HIMAX_WE1_SDK_MD5),$(HIMAX_WE1_SDK_NAME),)) - - #export path of toolchain - #export PATH := $(MAKEFILE_DIR)/downloads/$(HIMAX_WE1_SDK_NAME)/image_gen_linux_v3/:$(PATH) - - TCF_FILE := $(PWD)/$(MAKEFILE_DIR)/downloads/$(HIMAX_WE1_SDK_NAME)/arcem9d_wei_r16.tcf - LCF_FILE := $(PWD)/$(MAKEFILE_DIR)/downloads/$(HIMAX_WE1_SDK_NAME)/memory.lcf - ARCLIB_FILE := $(PWD)/$(MAKEFILE_DIR)/downloads/$(HIMAX_WE1_SDK_NAME)/libembarc.a - LIB_HEADER_FILE := $(PWD)/$(MAKEFILE_DIR)/downloads/$(HIMAX_WE1_SDK_NAME)/hx_drv_tflm.h - - - DEFAULT_HEAPSZ := 8192 - DEFAULT_STACKSZ := 8192 - - TCF_FILE_NAME = $(notdir $(TCF_FILE)) - ARC_TARGET_FILES_DIRS = $(dir $(TCF_FILE_NAME)) - MAKE_PROJECT_FILES += $(TCF_FILE_NAME) - - LCF_FILE_NAME = $(notdir $(LCF_FILE)) - ARC_TARGET_FILES_DIRS += $(dir $(LCF_FILE)) - MAKE_PROJECT_FILES += $(LCF_FILE_NAME) - - ARCLIB_FILE_NAME = $(notdir $(ARCLIB_FILE)) - ARC_TARGET_FILES_DIRS += $(dir $(ARCLIB_FILE)) - MAKE_PROJECT_FILES += $(ARCLIB_FILE_NAME) - - LIB_HEADER_FILE_NAME = $(notdir $(LIB_HEADER_FILE)) - ARC_TARGET_FILES_DIRS += $(dir $(LIB_HEADER_FILE)) - MAKE_PROJECT_FILES += $(LIB_HEADER_FILE_NAME) - - - - # Need a pointer to the TCF and lcf file - - PLATFORM_FLAGS = \ - -DNDEBUG \ - -g \ - -DCPU_ARC \ - -Hnosdata \ - -DTF_LITE_STATIC_MEMORY \ - -tcf=$(TCF_FILE_NAME) \ - -Hnocopyr \ - -Hpurge \ - -Hcl \ - -fslp-vectorize-aggressive \ - -ffunction-sections \ - -fdata-sections \ - -tcf_core_config \ - - CXXFLAGS += -fno-rtti -DSCRATCH_MEM_Z_SIZE=0x10000 $(PLATFORM_FLAGS) - CCFLAGS += $(PLATFORM_FLAGS) - - INCLUDES+= \ - -I $(MAKEFILE_DIR)/downloads/$(WEI_SDK_NAME) \ - -I $(MAKEFILE_DIR)/downloads/kissfft - - GENERATED_PROJECT_INCLUDES += \ - -I. \ - -I./third_party/kissfft - - LDFLAGS += \ - -Hheap=8192 \ - -tcf=$(TCF_FILE_NAME) \ - -Hnocopyr \ - -m \ - -Hldopt=-Coutput=$(TARGET).map \ - $(LCF_FILE_NAME) \ - -Hldopt=-Bgrouplib $(ARCLIB_FILE_NAME) - - CXXFLAGS := $(filter-out -std=c++11,$(CXXFLAGS)) - CCFLAGS := $(filter-out -std=c11,$(CCFLAGS)) - MICROLITE_LIBS := $(filter-out -lm,$(MICROLITE_LIBS)) - -endif From bdf6adfc0e78b5d1e21df527200b9cfaad5830c2 Mon Sep 17 00:00:00 2001 From: Nishidha Panpaliya Date: Thu, 4 Jun 2020 14:52:03 +0000 Subject: [PATCH 048/867] Review comments addressed --- tensorflow/compiler/xla/service/cpu/BUILD | 10 +++++++ .../service/cpu/test_target_triple_helper.h | 27 +++++++++++++++++++ .../compiler/xla/service/cpu/tests/BUILD | 4 +++ .../service/cpu/tests/cpu_dyn_shape_test.cc | 3 ++- .../cpu/tests/cpu_eigen_dot_operation_test.cc | 3 ++- .../cpu/tests/cpu_key_value_sort_test.cc | 3 ++- .../cpu/tests/cpu_literal_caching_test.cc | 5 ++-- .../xla/service/cpu/tests/cpu_outfeed_test.cc | 5 ++-- ...ed_reduce_with_no_vector_registers_test.cc | 3 ++- .../xla/tests/local_client_aot_test_helper.cc | 2 ++ 10 files changed, 57 insertions(+), 8 deletions(-) create mode 100644 tensorflow/compiler/xla/service/cpu/test_target_triple_helper.h diff --git a/tensorflow/compiler/xla/service/cpu/BUILD b/tensorflow/compiler/xla/service/cpu/BUILD index 3460e65b0a2..7be4d3e724a 100644 --- a/tensorflow/compiler/xla/service/cpu/BUILD +++ b/tensorflow/compiler/xla/service/cpu/BUILD @@ -30,6 +30,15 @@ filegroup( ]), ) +cc_library( + name = "test_header_helper", + testonly = True, + hdrs = ["test_target_triple_helper.h"], + deps = [ + "//tensorflow/core:test", + ], +) + filegroup( name = "single_threaded_runtime_srcs", srcs = [ @@ -1071,6 +1080,7 @@ tf_cc_test( deps = [ ":cpu_compiler", ":cpu_transfer_manager", + ":test_header_helper", "//tensorflow/compiler/xla:test", "//tensorflow/compiler/xla/tests:hlo_test_base", "//tensorflow/compiler/xla/tests:xla_internal_test_main", diff --git a/tensorflow/compiler/xla/service/cpu/test_target_triple_helper.h b/tensorflow/compiler/xla/service/cpu/test_target_triple_helper.h new file mode 100644 index 00000000000..e248f6de8bd --- /dev/null +++ b/tensorflow/compiler/xla/service/cpu/test_target_triple_helper.h @@ -0,0 +1,27 @@ +/* Copyright 2017 The TensorFlow Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. +==============================================================================*/ + +#ifndef TENSORFLOW_TEST_TARGET_TRIPLE_HELPER_H_ +#define TENSORFLOW_TEST_TARGET_TRIPLE_HELPER_H_ + +#if (defined(__powerpc__) || defined(__ppc__) && (__BYTE_ORDER__ == __ORDER_LITTLE_ENDIAN__)) +static const std::string kTargetCpuForHost="ppc"; +static const std::string kTargetTripleForHost="ppc64le-ibm-linux-gnu"; +#else +static const std::string kTargetCpuForHost=""; +static const std::string kTargetTripleForHost="x86_64-pc-linux"; +#endif + +#endif diff --git a/tensorflow/compiler/xla/service/cpu/tests/BUILD b/tensorflow/compiler/xla/service/cpu/tests/BUILD index 1ac8509cdb1..18624330a26 100644 --- a/tensorflow/compiler/xla/service/cpu/tests/BUILD +++ b/tensorflow/compiler/xla/service/cpu/tests/BUILD @@ -42,6 +42,7 @@ tf_cc_test( "//tensorflow/compiler/xla/service:hlo", "//tensorflow/compiler/xla/service/cpu:cpu_compiler", "//tensorflow/compiler/xla/service/cpu/tests:cpu_codegen_test", + "//tensorflow/compiler/xla/service/cpu:test_header_helper", "//tensorflow/core:lib", "//tensorflow/core:test", "//tensorflow/core:test_main", @@ -216,6 +217,7 @@ tf_cc_test( "//tensorflow/compiler/xla/service:hlo_parser", "//tensorflow/compiler/xla/service/cpu:cpu_compiler", "//tensorflow/compiler/xla/service/cpu/tests:cpu_codegen_test", + "//tensorflow/compiler/xla/service/cpu:test_header_helper", "//tensorflow/core:lib", "//tensorflow/core:test", "//tensorflow/core:test_main", @@ -229,6 +231,7 @@ tf_cc_test( "//tensorflow/compiler/xla/service:hlo", "//tensorflow/compiler/xla/service/cpu:cpu_compiler", "//tensorflow/compiler/xla/service/cpu/tests:cpu_codegen_test", + "//tensorflow/compiler/xla/service/cpu:test_header_helper", "//tensorflow/core:lib", "//tensorflow/core:test", "//tensorflow/core:test_main", @@ -242,6 +245,7 @@ tf_cc_test( "//tensorflow/compiler/xla/service:hlo", "//tensorflow/compiler/xla/service/cpu:cpu_compiler", "//tensorflow/compiler/xla/service/cpu/tests:cpu_codegen_test", + "//tensorflow/compiler/xla/service/cpu:test_header_helper", "//tensorflow/core:lib", "//tensorflow/core:test", "//tensorflow/core:test_main", diff --git a/tensorflow/compiler/xla/service/cpu/tests/cpu_dyn_shape_test.cc b/tensorflow/compiler/xla/service/cpu/tests/cpu_dyn_shape_test.cc index 46249caa0c7..7f9fab5dab0 100644 --- a/tensorflow/compiler/xla/service/cpu/tests/cpu_dyn_shape_test.cc +++ b/tensorflow/compiler/xla/service/cpu/tests/cpu_dyn_shape_test.cc @@ -17,6 +17,7 @@ limitations under the License. #include "tensorflow/compiler/xla/service/cpu/cpu_compiler.h" #include "tensorflow/compiler/xla/service/cpu/tests/cpu_codegen_test.h" +#include "tensorflow/compiler/xla/service/cpu/test_target_triple_helper.h" namespace xla { namespace cpu { @@ -46,7 +47,7 @@ TEST_F(CpuDynamicShapeTest, DynamicShapeR2) { )"; CpuAotCompilationOptions options{ - /*triple=*/"x86_64", /*cpu_name=*/"", /*features=*/"", + /*triple=*/kTargetTripleForHost, /*cpu_name=*/kTargetCpuForHost, /*features=*/"", /*entry_point_name=*/"entry", /*relocation_model=*/CpuAotCompilationOptions::RelocationModel::Static}; diff --git a/tensorflow/compiler/xla/service/cpu/tests/cpu_eigen_dot_operation_test.cc b/tensorflow/compiler/xla/service/cpu/tests/cpu_eigen_dot_operation_test.cc index 8b7f843582b..40314ae5158 100644 --- a/tensorflow/compiler/xla/service/cpu/tests/cpu_eigen_dot_operation_test.cc +++ b/tensorflow/compiler/xla/service/cpu/tests/cpu_eigen_dot_operation_test.cc @@ -23,6 +23,7 @@ limitations under the License. #include "tensorflow/compiler/xla/service/cpu/tests/cpu_codegen_test.h" #include "tensorflow/compiler/xla/service/hlo_computation.h" #include "tensorflow/compiler/xla/tests/test_utils.h" +#include "tensorflow/compiler/xla/service/cpu/test_target_triple_helper.h" #include "tensorflow/core/platform/test.h" namespace xla { @@ -45,7 +46,7 @@ class CpuEigenDotOperationTest void CompileAndCheck(std::unique_ptr entry_computation, const string& filecheck_lines) { CpuAotCompilationOptions options{ - /*triple=*/"x86_64", /*cpu_name=*/"", /*features=*/"", + /*triple=*/kTargetTripleForHost, /*cpu_name=*/kTargetCpuForHost, /*features=*/"", /*entry_point_name=*/"entry", /*relocation_model=*/CpuAotCompilationOptions::RelocationModel::Static}; diff --git a/tensorflow/compiler/xla/service/cpu/tests/cpu_key_value_sort_test.cc b/tensorflow/compiler/xla/service/cpu/tests/cpu_key_value_sort_test.cc index f3b7b91b2b5..8bb8acb557d 100644 --- a/tensorflow/compiler/xla/service/cpu/tests/cpu_key_value_sort_test.cc +++ b/tensorflow/compiler/xla/service/cpu/tests/cpu_key_value_sort_test.cc @@ -17,6 +17,7 @@ limitations under the License. #include "tensorflow/compiler/xla/service/cpu/cpu_compiler.h" #include "tensorflow/compiler/xla/service/cpu/tests/cpu_codegen_test.h" +#include "tensorflow/compiler/xla/service/cpu/test_target_triple_helper.h" namespace xla { namespace cpu { @@ -48,7 +49,7 @@ CHECK: call void @__xla_cpu_runtime_KeyValueSort TF_ASSERT_OK_AND_ASSIGN(auto module, ParseAndReturnVerifiedModule(hlo_text)); CpuAotCompilationOptions options{ - /*triple=*/"x86_64", /*cpu_name=*/"", /*features=*/"", + /*triple=*/kTargetTripleForHost, /*cpu_name=*/kTargetCpuForHost, /*features=*/"", /*entry_point_name=*/"entry", /*relocation_model=*/CpuAotCompilationOptions::RelocationModel::Static}; diff --git a/tensorflow/compiler/xla/service/cpu/tests/cpu_literal_caching_test.cc b/tensorflow/compiler/xla/service/cpu/tests/cpu_literal_caching_test.cc index fc670201125..b86e23fc7b6 100644 --- a/tensorflow/compiler/xla/service/cpu/tests/cpu_literal_caching_test.cc +++ b/tensorflow/compiler/xla/service/cpu/tests/cpu_literal_caching_test.cc @@ -16,6 +16,7 @@ limitations under the License. #include "tensorflow/compiler/xla/service/cpu/cpu_compiler.h" #include "tensorflow/compiler/xla/service/cpu/tests/cpu_codegen_test.h" #include "tensorflow/compiler/xla/service/hlo_parser.h" +#include "tensorflow/compiler/xla/service/cpu/test_target_triple_helper.h" namespace xla { namespace cpu { @@ -64,7 +65,7 @@ CHECK-NOT: private unnamed_addr constant [48 x i8] ParseAndReturnVerifiedModule(hlo_text)); CpuAotCompilationOptions options{ - /*triple=*/"x86_64-pc-linux", /*cpu_name=*/"", /*features=*/"", + /*triple=*/kTargetTripleForHost, /*cpu_name=*/kTargetCpuForHost, /*features=*/"", /*entry_point_name=*/"entry", /*relocation_model=*/CpuAotCompilationOptions::RelocationModel::Static}; @@ -112,7 +113,7 @@ CHECK-NOT: private unnamed_addr constant [8 x i8] ParseAndReturnVerifiedModule(hlo_text)); CpuAotCompilationOptions options{ - /*triple=*/"x86_64-pc-linux", /*cpu_name=*/"", /*features=*/"", + /*triple=*/kTargetTripleForHost, /*cpu_name=*/kTargetCpuForHost, /*features=*/"", /*entry_point_name=*/"entry", /*relocation_model=*/CpuAotCompilationOptions::RelocationModel::Static}; diff --git a/tensorflow/compiler/xla/service/cpu/tests/cpu_outfeed_test.cc b/tensorflow/compiler/xla/service/cpu/tests/cpu_outfeed_test.cc index ad83c485998..ea5f282fb2b 100644 --- a/tensorflow/compiler/xla/service/cpu/tests/cpu_outfeed_test.cc +++ b/tensorflow/compiler/xla/service/cpu/tests/cpu_outfeed_test.cc @@ -17,6 +17,7 @@ limitations under the License. #include "tensorflow/compiler/xla/service/cpu/cpu_compiler.h" #include "tensorflow/compiler/xla/service/cpu/tests/cpu_codegen_test.h" +#include "tensorflow/compiler/xla/service/cpu/test_target_triple_helper.h" namespace xla { namespace cpu { @@ -46,7 +47,7 @@ CHECK: private unnamed_addr constant [48 x i8] TF_ASSERT_OK_AND_ASSIGN(auto module, ParseAndReturnVerifiedModule(hlo_text)); CpuAotCompilationOptions options{ - /*triple=*/"x86_64-pc-linux", /*cpu_name=*/"", /*features=*/"", + /*triple=*/kTargetTripleForHost, /*cpu_name=*/kTargetCpuForHost, /*features=*/"", /*entry_point_name=*/"entry", /*relocation_model=*/CpuAotCompilationOptions::RelocationModel::Static}; @@ -73,7 +74,7 @@ CHECK: Outfeed TF_ASSERT_OK_AND_ASSIGN(auto module, ParseAndReturnVerifiedModule(hlo_text)); CpuAotCompilationOptions options{ - /*triple=*/"x86_64-pc-linux", /*cpu_name=*/"", /*features=*/"", + /*triple=*/kTargetTripleForHost, /*cpu_name=*/kTargetCpuForHost, /*features=*/"", /*entry_point_name=*/"entry", /*relocation_model=*/CpuAotCompilationOptions::RelocationModel::Static}; diff --git a/tensorflow/compiler/xla/service/cpu/vectorized_reduce_with_no_vector_registers_test.cc b/tensorflow/compiler/xla/service/cpu/vectorized_reduce_with_no_vector_registers_test.cc index 754885d8744..e59a531b114 100644 --- a/tensorflow/compiler/xla/service/cpu/vectorized_reduce_with_no_vector_registers_test.cc +++ b/tensorflow/compiler/xla/service/cpu/vectorized_reduce_with_no_vector_registers_test.cc @@ -20,6 +20,7 @@ limitations under the License. #include "tensorflow/compiler/xla/service/cpu/cpu_compiler.h" #include "tensorflow/compiler/xla/test.h" #include "tensorflow/compiler/xla/tests/hlo_test_base.h" +#include "tensorflow/compiler/xla/service/cpu/test_target_triple_helper.h" namespace xla { namespace { @@ -75,7 +76,7 @@ ENTRY main { // Check that the GetTargetVectorRegisterByteSize is itself working. TF_ASSERT_OK_AND_ASSIGN(unsigned vector_register_byte_size_for_x86_64, - GetTargetVectorRegisterByteSize("x86_64-pc-linux")); + GetTargetVectorRegisterByteSize(kTargetTripleForHost)); ASSERT_EQ(vector_register_byte_size_for_x86_64, 16); std::string triple = "i686-none-android"; diff --git a/tensorflow/compiler/xla/tests/local_client_aot_test_helper.cc b/tensorflow/compiler/xla/tests/local_client_aot_test_helper.cc index 53c0d84854e..3e9a3ec2314 100644 --- a/tensorflow/compiler/xla/tests/local_client_aot_test_helper.cc +++ b/tensorflow/compiler/xla/tests/local_client_aot_test_helper.cc @@ -71,6 +71,8 @@ int main(int argc, char** argv) { triple_string = "aarch64-none-linux-gnu"; } else if (target_cpu == "x64_windows") { triple_string = "x86_64-pc-windows-msvc19"; + } else if (target_cpu == "ppc") { + triple_string = "ppc64le-ibm-linux-gnu"; } else if (target_cpu == "local") { triple_string = llvm::sys::getDefaultTargetTriple(); } else { From ac123654efc63ffa17240479a5b926ca6357c766 Mon Sep 17 00:00:00 2001 From: "902449@58880@bigcat_chen@ASIC" Date: Fri, 5 Jun 2020 10:03:47 +0800 Subject: [PATCH 049/867] TFLM: update hello world example readme --- .../lite/micro/examples/hello_world/README.md | 4 +- .../micro/examples/hello_world/README.md~ | 575 ++++++++++++++++++ 2 files changed, 577 insertions(+), 2 deletions(-) create mode 100644 tensorflow/lite/micro/examples/hello_world/README.md~ diff --git a/tensorflow/lite/micro/examples/hello_world/README.md b/tensorflow/lite/micro/examples/hello_world/README.md index d3762ada790..26b0f12c83a 100644 --- a/tensorflow/lite/micro/examples/hello_world/README.md +++ b/tensorflow/lite/micro/examples/hello_world/README.md @@ -17,7 +17,7 @@ of the device. - [Deploy to ARC EM SDP](#deploy-to-arc-em-sdp) - [Deploy to Arduino](#deploy-to-arduino) - [Deploy to ESP32](#deploy-to-esp32) -- [Deploy to himax WE1 EVB](#deploy-to-himax-we1-evb) +- [Deploy to Himax WE1 EVB](#deploy-to-himax-we1-evb) - [Deploy to SparkFun Edge](#deploy-to-sparkfun-edge) - [Deploy to STM32F746](#deploy-to-STM32F746) - [Run the tests on a development machine](#run-the-tests-on-a-development-machine) @@ -192,7 +192,7 @@ The previous two commands can be combined: idf.py --port /dev/ttyUSB0 flash monitor ``` -## Deploy to himax WE1 EVB +## Deploy to Himax WE1 EVB The following instructions will help you build and deploy this example to [HIMAX WE1 EVB](https://github.com/HimaxWiseEyePlus/bsp_tflu/tree/master/HIMAX_WE1_EVB_board_brief) diff --git a/tensorflow/lite/micro/examples/hello_world/README.md~ b/tensorflow/lite/micro/examples/hello_world/README.md~ new file mode 100644 index 00000000000..d3762ada790 --- /dev/null +++ b/tensorflow/lite/micro/examples/hello_world/README.md~ @@ -0,0 +1,575 @@ +# Hello World Example + +This example is designed to demonstrate the absolute basics of using [TensorFlow +Lite for Microcontrollers](https://www.tensorflow.org/lite/microcontrollers). +It includes the full end-to-end workflow of training a model, converting it for +use with TensorFlow Lite for Microcontrollers for running inference on a +microcontroller. + +The model is trained to replicate a `sine` function and generates a pattern of +data to either blink LEDs or control an animation, depending on the capabilities +of the device. + +![Animation on STM32F746](images/animation_on_STM32F746.gif) + +## Table of contents + +- [Deploy to ARC EM SDP](#deploy-to-arc-em-sdp) +- [Deploy to Arduino](#deploy-to-arduino) +- [Deploy to ESP32](#deploy-to-esp32) +- [Deploy to himax WE1 EVB](#deploy-to-himax-we1-evb) +- [Deploy to SparkFun Edge](#deploy-to-sparkfun-edge) +- [Deploy to STM32F746](#deploy-to-STM32F746) +- [Run the tests on a development machine](#run-the-tests-on-a-development-machine) +- [Train your own model](#train-your-own-model) + +## Deploy to ARC EM SDP + +The following instructions will help you to build and deploy this example to +[ARC EM SDP](https://www.synopsys.com/dw/ipdir.php?ds=arc-em-software-development-platform) +board. General information and instructions on using the board with TensorFlow +Lite Micro can be found in the common +[ARC targets description](/tensorflow/lite/micro/tools/make/targets/arc/README.md). + +### Initial Setup + +Follow the instructions on the +[ARC EM SDP Initial Setup](/tensorflow/lite/micro/tools/make/targets/arc/README.md#ARC-EM-Software-Development-Platform-ARC-EM-SDP) +to get and install all required tools for work with ARC EM SDP. + +### Generate Example Project + +The example project for ARC EM SDP platform can be generated with the following +command: + +``` +make -f tensorflow/lite/micro/tools/make/Makefile TARGET=arc_emsdp TAGS=no_arc_mli generate_hello_world_make_project +``` + +### Build and Run Example + +For more detailed information on building and running examples see the +appropriate sections of general descriptions of the +[ARC EM SDP usage with TFLM](/tensorflow/lite/micro/tools/make/targets/arc/README.md#ARC-EM-Software-Development-Platform-ARC-EM-SDP). +In the directory with generated project you can also find a +*README_ARC_EMSDP.md* file with instructions and options on building and +running. Here we only briefly mention main steps which are typically enough to +get it started. + +1. You need to + [connect the board](/tensorflow/lite/micro/tools/make/targets/arc/README.md#connect-the-board) + and open an serial connection. + +2. Go to the generated example project director + + ``` + cd tensorflow/lite/micro/tools/make/gen/arc_emsdp_arc/prj/hello_world/make + ``` + +3. Build the example using + + ``` + make app + ``` + +4. To generate artefacts for self-boot of example from the board use + + ``` + make flash + ``` + +5. To run application from the board using microSD card: + + * Copy the content of the created /bin folder into the root of microSD + card. Note that the card must be formatted as FAT32 with default cluster + size (but less than 32 Kbytes) + * Plug in the microSD card into the J11 connector. + * Push the RST button. If a red LED is lit beside RST button, push the CFG + button. + +6. If you have the MetaWare Debugger installed in your environment: + + * To run application from the console using it type `make run`. + * To stop the execution type `Ctrl+C` in the console several times. + +In both cases (step 5 and 6) you will see the application output in the serial +terminal. + +## Deploy to Arduino + +The following instructions will help you build and deploy this sample +to [Arduino](https://www.arduino.cc/) devices. + +![Animation on Arduino MKRZERO](images/animation_on_arduino_mkrzero.gif) + +The sample has been tested with the following devices: + +- [Arduino Nano 33 BLE Sense](https://store.arduino.cc/usa/nano-33-ble-sense-with-headers) +- [Arduino MKRZERO](https://store.arduino.cc/usa/arduino-mkrzero) + +The sample will use PWM to fade an LED on and off according to the model's +output. In the code, the `LED_BUILTIN` constant is used to specify the board's +built-in LED as the one being controlled. However, on some boards, this built-in +LED is not attached to a pin with PWM capabilities. In this case, the LED will +blink instead of fading. + +### Install the Arduino_TensorFlowLite library + +This example application is included as part of the official TensorFlow Lite +Arduino library. To install it, open the Arduino library manager in +`Tools -> Manage Libraries...` and search for `Arduino_TensorFlowLite`. + +### Load and run the example + +Once the library has been added, go to `File -> Examples`. You should see an +example near the bottom of the list named `TensorFlowLite:hello_world`. Select +it and click `hello_world` to load the example. + +Use the Arduino IDE to build and upload the example. Once it is running, +you should see the built-in LED on your device flashing. + +The Arduino Desktop IDE includes a plotter that we can use to display the sine +wave graphically. To view it, go to `Tools -> Serial Plotter`. You will see one +datapoint being logged for each inference cycle, expressed as a number between 0 +and 255. + +## Deploy to ESP32 + +The following instructions will help you build and deploy this sample +to [ESP32](https://www.espressif.com/en/products/hardware/esp32/overview) +devices using the [ESP IDF](https://github.com/espressif/esp-idf). + +The sample has been tested on ESP-IDF version 4.0 with the following devices: +- [ESP32-DevKitC](http://esp-idf.readthedocs.io/en/latest/get-started/get-started-devkitc.html) +- [ESP-EYE](https://github.com/espressif/esp-who/blob/master/docs/en/get-started/ESP-EYE_Getting_Started_Guide.md) + +### Install the ESP IDF + +Follow the instructions of the +[ESP-IDF get started guide](https://docs.espressif.com/projects/esp-idf/en/latest/get-started/index.html) +to setup the toolchain and the ESP-IDF itself. + +The next steps assume that the +[IDF environment variables are set](https://docs.espressif.com/projects/esp-idf/en/latest/get-started/index.html#step-4-set-up-the-environment-variables) : + + * The `IDF_PATH` environment variable is set + * `idf.py` and Xtensa-esp32 tools (e.g. `xtensa-esp32-elf-gcc`) are in `$PATH` + +### Generate the examples +The example project can be generated with the following command: +``` +make -f tensorflow/lite/micro/tools/make/Makefile TARGET=esp generate_hello_world_esp_project +``` + +### Building the example + +Go the the example project directory +``` +cd tensorflow/lite/micro/tools/make/gen/esp_xtensa-esp32/prj/hello_world/esp-idf +``` + +Then build with `idf.py` +``` +idf.py build +``` + +### Load and run the example + +To flash (replace `/dev/ttyUSB0` with the device serial port): +``` +idf.py --port /dev/ttyUSB0 flash +``` + +Monitor the serial output: +``` +idf.py --port /dev/ttyUSB0 monitor +``` + +Use `Ctrl+]` to exit. + +The previous two commands can be combined: +``` +idf.py --port /dev/ttyUSB0 flash monitor +``` + +## Deploy to himax WE1 EVB + +The following instructions will help you build and deploy this example to +[HIMAX WE1 EVB](https://github.com/HimaxWiseEyePlus/bsp_tflu/tree/master/HIMAX_WE1_EVB_board_brief) +board. To undstand more about using this board, please check +[HIMAX WE1 EVB user guide](https://github.com/HimaxWiseEyePlus/bsp_tflu/tree/master/HIMAX_WE1_EVB_user_guide). + +### Initial Setup + +To use the HIMAX WE1 EVB, please make sure following software are installed: + +#### MetaWare Development Toolkit + +See +[Install the Synopsys DesignWare ARC MetaWare Development Toolkit](/tensorflow/lite/micro/tools/make/targets/arc/README.md#install-the-synopsys-designware-arc-metaware-development-toolkit) +section for instructions on toolchain installation. + +#### Make Tool version + +A `'make'` tool is required for deploying Tensorflow Lite Micro +applications on HIMAX WE1 EVB, See +[Check make tool version](/tensorflow/lite/micro/tools/make/targets/arc/README.md#make-tool) +section for proper environment. + +#### Serial Terminal Emulation Application + +There are 2 main purposes for HIMAX WE1 EVB Debug UART port + +- print application output +- burn application to flash by using xmodem send application binary + +You can use any terminal emulation program (like [PuTTY](https://www.putty.org/) or [minicom](https://linux.die.net/man/1/minicom)). + + +### Generate Example Project + +The example project for HIMAX WE1 EVB platform can be generated with the following +command: + +Download related third party data + +``` +make -f tensorflow/lite/micro/tools/make/Makefile TARGET=himax_we1_evb third_party_downloads +``` + +Generate hello world project + +``` +make -f tensorflow/lite/micro/tools/make/Makefile generate_hello_world_make_project TARGET=himax_we1_evb TAGS=no_arc_mli +``` + +### Build and Burn Example + +Following the Steps to run hello world example at HIMAX WE1 EVB platform. + +1. Go to the generated example project directory. + + ``` + cd tensorflow/lite/micro/tools/make/gen/himax_we1_evb_arc/prj/hello_world/make + ``` + +2. Build the example using + + ``` + make app + ``` + +3. After example build finish, copy ELF file and map file to image generate tool directory. + image generate tool directory located at `'tensorflow/lite/micro/tools/make/downloads/himax_we1_sdk/image_gen_linux_v3/'` + + ``` + cp hello_world.elf himax_we1_evb.map ../../../../../downloads/himax_we1_sdk/image_gen_linux_v3/ + ``` + +4. Go to flash image generate tool directory. + + ``` + cd ../../../../../downloads/himax_we1_sdk/image_gen_linux_v3/ + ``` + +5. run image generate tool, generate flash image file. + + * Before running image generate tool, by typing `sudo chmod +x image_gen` + and `sudo chmod +x sign_tool` to make sure it is executable. + + ``` + image_gen -e hello_world.elf -m himax_we1_evb.map -o out.img + ``` + + +6. Download flash image file to HIMAX WE1 EVB by UART: + + * more detail about download image through UART can be found at [HIMAX WE1 EVB update Flash image](https://github.com/HimaxWiseEyePlus/bsp_tflu/tree/master/HIMAX_WE1_EVB_user_guide#flash-image-update) + +After these steps, press reset button on the HIMAX WE1 EVB, you will see application output in the serial +terminal. + +## Deploy to SparkFun Edge + +The following instructions will help you build and deploy this sample on the +[SparkFun Edge development board](https://sparkfun.com/products/15170). + +![Animation on SparkFun Edge](images/animation_on_sparkfun_edge.gif) + +If you're new to using this board, we recommend walking through the +[AI on a microcontroller with TensorFlow Lite and SparkFun Edge](https://codelabs.developers.google.com/codelabs/sparkfun-tensorflow) +codelab to get an understanding of the workflow. + +### Compile the binary + +The following command will download the required dependencies and then compile a +binary for the SparkFun Edge: + +``` +make -f tensorflow/lite/micro/tools/make/Makefile TARGET=sparkfun_edge hello_world_bin +``` + +The binary will be created in the following location: + +``` +tensorflow/lite/micro/tools/make/gen/sparkfun_edge_cortex-m4/bin/hello_world.bin +``` + +### Sign the binary + +The binary must be signed with cryptographic keys to be deployed to the device. +We'll now run some commands that will sign our binary so it can be flashed to +the SparkFun Edge. The scripts we are using come from the Ambiq SDK, which is +downloaded when the `Makefile` is run. + +Enter the following command to set up some dummy cryptographic keys we can use +for development: + +``` +cp tensorflow/lite/micro/tools/make/downloads/AmbiqSuite-Rel2.2.0/tools/apollo3_scripts/keys_info0.py \ +tensorflow/lite/micro/tools/make/downloads/AmbiqSuite-Rel2.2.0/tools/apollo3_scripts/keys_info.py +``` + +Next, run the following command to create a signed binary: + +``` +python3 tensorflow/lite/micro/tools/make/downloads/AmbiqSuite-Rel2.2.0/tools/apollo3_scripts/create_cust_image_blob.py \ +--bin tensorflow/lite/micro/tools/make/gen/sparkfun_edge_cortex-m4/bin/hello_world.bin \ +--load-address 0xC000 \ +--magic-num 0xCB \ +-o main_nonsecure_ota \ +--version 0x0 +``` + +This will create the file `main_nonsecure_ota.bin`. We'll now run another +command to create a final version of the file that can be used to flash our +device with the bootloader script we will use in the next step: + +``` +python3 tensorflow/lite/micro/tools/make/downloads/AmbiqSuite-Rel2.2.0/tools/apollo3_scripts/create_cust_wireupdate_blob.py \ +--load-address 0x20000 \ +--bin main_nonsecure_ota.bin \ +-i 6 \ +-o main_nonsecure_wire \ +--options 0x1 +``` + +You should now have a file called `main_nonsecure_wire.bin` in the directory +where you ran the commands. This is the file we'll be flashing to the device. + +### Flash the binary + +Next, attach the board to your computer via a USB-to-serial adapter. + +**Note:** If you're using the [SparkFun Serial Basic Breakout](https://www.sparkfun.com/products/15096), +you should [install the latest drivers](https://learn.sparkfun.com/tutorials/sparkfun-serial-basic-ch340c-hookup-guide#drivers-if-you-need-them) +before you continue. + +Once connected, assign the USB device name to an environment variable: + +``` +export DEVICENAME=put your device name here +``` + +Set another variable with the baud rate: + +``` +export BAUD_RATE=921600 +``` + +Now, hold the button marked `14` on the device. While still holding the button, +hit the button marked `RST`. Continue holding the button marked `14` while +running the following command: + +``` +python3 tensorflow/lite/micro/tools/make/downloads/AmbiqSuite-Rel2.2.0/tools/apollo3_scripts/uart_wired_update.py \ +-b ${BAUD_RATE} ${DEVICENAME} \ +-r 1 \ +-f main_nonsecure_wire.bin \ +-i 6 +``` + +You should see a long stream of output as the binary is flashed to the device. +Once you see the following lines, flashing is complete: + +``` +Sending Reset Command. +Done. +``` + +If you don't see these lines, flashing may have failed. Try running through the +steps in [Flash the binary](#flash-the-binary) again (you can skip over setting +the environment variables). If you continue to run into problems, follow the +[AI on a microcontroller with TensorFlow Lite and SparkFun Edge](https://codelabs.developers.google.com/codelabs/sparkfun-tensorflow) +codelab, which includes more comprehensive instructions for the flashing +process. + +The binary should now be deployed to the device. Hit the button marked `RST` to +reboot the board. You should see the device's four LEDs flashing in sequence. + +Debug information is logged by the board while the program is running. To view +it, establish a serial connection to the board using a baud rate of `115200`. +On OSX and Linux, the following command should work: + +``` +screen ${DEVICENAME} 115200 +``` + +You will see a lot of output flying past! To stop the scrolling, hit `Ctrl+A`, +immediately followed by `Esc`. You can then use the arrow keys to explore the +output, which will contain the results of running inference on various `x` +values: + +``` +x_value: 1.1843798*2^2, y_value: -1.9542645*2^-1 +``` + +To stop viewing the debug output with `screen`, hit `Ctrl+A`, immediately +followed by the `K` key, then hit the `Y` key. + + +## Deploy to STM32F746 + +The following instructions will help you build and deploy the sample to the +[STM32F7 discovery kit](https://os.mbed.com/platforms/ST-Discovery-F746NG/) +using [ARM Mbed](https://github.com/ARMmbed/mbed-cli). + +![Animation on STM32F746](images/animation_on_STM32F746.gif) + +Before we begin, you'll need the following: + +- STM32F7 discovery kit board +- Mini-USB cable +- ARM Mbed CLI ([installation instructions](https://os.mbed.com/docs/mbed-os/v5.12/tools/installation-and-setup.html)) +- Python 2.7 and pip + +Since Mbed requires a special folder structure for projects, we'll first run a +command to generate a subfolder containing the required source files in this +structure: + +``` +make -f tensorflow/lite/micro/tools/make/Makefile TARGET=mbed TAGS="CMSIS disco_f746ng" generate_hello_world_mbed_project +``` + +This will result in the creation of a new folder: + +``` +tensorflow/lite/micro/tools/make/gen/mbed_cortex-m4/prj/hello_world/mbed +``` + +This folder contains all of the example's dependencies structured in the correct +way for Mbed to be able to build it. + +Change into the directory and run the following commands, making sure you are +using Python 2.7.15. + +First, tell Mbed that the current directory is the root of an Mbed project: + +``` +mbed config root . +``` + +Next, tell Mbed to download the dependencies and prepare to build: + +``` +mbed deploy +``` + +By default, Mbed will build the project using C++98. However, TensorFlow Lite +requires C++11. Run the following Python snippet to modify the Mbed +configuration files so that it uses C++11: + +``` +python -c 'import fileinput, glob; +for filename in glob.glob("mbed-os/tools/profiles/*.json"): + for line in fileinput.input(filename, inplace=True): + print line.replace("\"-std=gnu++98\"","\"-std=c++11\", \"-fpermissive\"")' + +``` + +Finally, run the following command to compile: + +``` +mbed compile -m DISCO_F746NG -t GCC_ARM +``` + +This should result in a binary at the following path: + +``` +./BUILD/DISCO_F746NG/GCC_ARM/mbed.bin +``` + +To deploy, plug in your STM board and copy the file to it. On MacOS, you can do +this with the following command: + +``` +cp ./BUILD/DISCO_F746NG/GCC_ARM/mbed.bin /Volumes/DIS_F746NG/ +``` + +Copying the file will initiate the flashing process. Once this is complete, you +should see an animation on the device's screen. + + +``` +screen /dev/tty.usbmodem14403 9600 +``` + +In addition to this animation, debug information is logged by the board while +the program is running. To view it, establish a serial connection to the board +using a baud rate of `9600`. On OSX and Linux, the following command should +work, replacing `/dev/tty.devicename` with the name of your device as it appears +in `/dev`: + +``` +screen /dev/tty.devicename 9600 +``` + +You will see a lot of output flying past! To stop the scrolling, hit `Ctrl+A`, +immediately followed by `Esc`. You can then use the arrow keys to explore the +output, which will contain the results of running inference on various `x` +values: + +``` +x_value: 1.1843798*2^2, y_value: -1.9542645*2^-1 +``` + +To stop viewing the debug output with `screen`, hit `Ctrl+A`, immediately +followed by the `K` key, then hit the `Y` key. + +### Run the tests on a development machine + +To compile and test this example on a desktop Linux or macOS machine, first +clone the TensorFlow repository from GitHub to a convenient place: + +```bash +git clone --depth 1 https://github.com/tensorflow/tensorflow.git +``` + +Next, `cd` into the source directory from a terminal, and then run the following +command: + +```bash +make -f tensorflow/lite/micro/tools/make/Makefile test_hello_world_test +``` + +This will take a few minutes, and downloads frameworks the code uses. Once the +process has finished, you should see a series of files get compiled, followed by +some logging output from a test, which should conclude with +`~~~ALL TESTS PASSED~~~`. + +If you see this, it means that a small program has been built and run that loads +the trained TensorFlow model, runs some example inputs through it, and got the +expected outputs. + +To understand how TensorFlow Lite does this, you can look at the source in +[hello_world_test.cc](hello_world_test.cc). +It's a fairly small amount of code that creates an interpreter, gets a handle to +a model that's been compiled into the program, and then invokes the interpreter +with the model and sample inputs. + +### Train your own model + +So far you have used an existing trained model to run inference on +microcontrollers. If you wish to train your own model, follow the instructions +given in the [train/](train/) directory. + From 1c26e6abd76fe700ecf87d892ceed1dc5bfa90d3 Mon Sep 17 00:00:00 2001 From: "902449@58880@bigcat_chen@ASIC" Date: Fri, 5 Jun 2020 10:12:46 +0800 Subject: [PATCH 050/867] TFLM: delete temp readme file in hello world example --- .../micro/examples/hello_world/README.md~ | 575 ------------------ 1 file changed, 575 deletions(-) delete mode 100644 tensorflow/lite/micro/examples/hello_world/README.md~ diff --git a/tensorflow/lite/micro/examples/hello_world/README.md~ b/tensorflow/lite/micro/examples/hello_world/README.md~ deleted file mode 100644 index d3762ada790..00000000000 --- a/tensorflow/lite/micro/examples/hello_world/README.md~ +++ /dev/null @@ -1,575 +0,0 @@ -# Hello World Example - -This example is designed to demonstrate the absolute basics of using [TensorFlow -Lite for Microcontrollers](https://www.tensorflow.org/lite/microcontrollers). -It includes the full end-to-end workflow of training a model, converting it for -use with TensorFlow Lite for Microcontrollers for running inference on a -microcontroller. - -The model is trained to replicate a `sine` function and generates a pattern of -data to either blink LEDs or control an animation, depending on the capabilities -of the device. - -![Animation on STM32F746](images/animation_on_STM32F746.gif) - -## Table of contents - -- [Deploy to ARC EM SDP](#deploy-to-arc-em-sdp) -- [Deploy to Arduino](#deploy-to-arduino) -- [Deploy to ESP32](#deploy-to-esp32) -- [Deploy to himax WE1 EVB](#deploy-to-himax-we1-evb) -- [Deploy to SparkFun Edge](#deploy-to-sparkfun-edge) -- [Deploy to STM32F746](#deploy-to-STM32F746) -- [Run the tests on a development machine](#run-the-tests-on-a-development-machine) -- [Train your own model](#train-your-own-model) - -## Deploy to ARC EM SDP - -The following instructions will help you to build and deploy this example to -[ARC EM SDP](https://www.synopsys.com/dw/ipdir.php?ds=arc-em-software-development-platform) -board. General information and instructions on using the board with TensorFlow -Lite Micro can be found in the common -[ARC targets description](/tensorflow/lite/micro/tools/make/targets/arc/README.md). - -### Initial Setup - -Follow the instructions on the -[ARC EM SDP Initial Setup](/tensorflow/lite/micro/tools/make/targets/arc/README.md#ARC-EM-Software-Development-Platform-ARC-EM-SDP) -to get and install all required tools for work with ARC EM SDP. - -### Generate Example Project - -The example project for ARC EM SDP platform can be generated with the following -command: - -``` -make -f tensorflow/lite/micro/tools/make/Makefile TARGET=arc_emsdp TAGS=no_arc_mli generate_hello_world_make_project -``` - -### Build and Run Example - -For more detailed information on building and running examples see the -appropriate sections of general descriptions of the -[ARC EM SDP usage with TFLM](/tensorflow/lite/micro/tools/make/targets/arc/README.md#ARC-EM-Software-Development-Platform-ARC-EM-SDP). -In the directory with generated project you can also find a -*README_ARC_EMSDP.md* file with instructions and options on building and -running. Here we only briefly mention main steps which are typically enough to -get it started. - -1. You need to - [connect the board](/tensorflow/lite/micro/tools/make/targets/arc/README.md#connect-the-board) - and open an serial connection. - -2. Go to the generated example project director - - ``` - cd tensorflow/lite/micro/tools/make/gen/arc_emsdp_arc/prj/hello_world/make - ``` - -3. Build the example using - - ``` - make app - ``` - -4. To generate artefacts for self-boot of example from the board use - - ``` - make flash - ``` - -5. To run application from the board using microSD card: - - * Copy the content of the created /bin folder into the root of microSD - card. Note that the card must be formatted as FAT32 with default cluster - size (but less than 32 Kbytes) - * Plug in the microSD card into the J11 connector. - * Push the RST button. If a red LED is lit beside RST button, push the CFG - button. - -6. If you have the MetaWare Debugger installed in your environment: - - * To run application from the console using it type `make run`. - * To stop the execution type `Ctrl+C` in the console several times. - -In both cases (step 5 and 6) you will see the application output in the serial -terminal. - -## Deploy to Arduino - -The following instructions will help you build and deploy this sample -to [Arduino](https://www.arduino.cc/) devices. - -![Animation on Arduino MKRZERO](images/animation_on_arduino_mkrzero.gif) - -The sample has been tested with the following devices: - -- [Arduino Nano 33 BLE Sense](https://store.arduino.cc/usa/nano-33-ble-sense-with-headers) -- [Arduino MKRZERO](https://store.arduino.cc/usa/arduino-mkrzero) - -The sample will use PWM to fade an LED on and off according to the model's -output. In the code, the `LED_BUILTIN` constant is used to specify the board's -built-in LED as the one being controlled. However, on some boards, this built-in -LED is not attached to a pin with PWM capabilities. In this case, the LED will -blink instead of fading. - -### Install the Arduino_TensorFlowLite library - -This example application is included as part of the official TensorFlow Lite -Arduino library. To install it, open the Arduino library manager in -`Tools -> Manage Libraries...` and search for `Arduino_TensorFlowLite`. - -### Load and run the example - -Once the library has been added, go to `File -> Examples`. You should see an -example near the bottom of the list named `TensorFlowLite:hello_world`. Select -it and click `hello_world` to load the example. - -Use the Arduino IDE to build and upload the example. Once it is running, -you should see the built-in LED on your device flashing. - -The Arduino Desktop IDE includes a plotter that we can use to display the sine -wave graphically. To view it, go to `Tools -> Serial Plotter`. You will see one -datapoint being logged for each inference cycle, expressed as a number between 0 -and 255. - -## Deploy to ESP32 - -The following instructions will help you build and deploy this sample -to [ESP32](https://www.espressif.com/en/products/hardware/esp32/overview) -devices using the [ESP IDF](https://github.com/espressif/esp-idf). - -The sample has been tested on ESP-IDF version 4.0 with the following devices: -- [ESP32-DevKitC](http://esp-idf.readthedocs.io/en/latest/get-started/get-started-devkitc.html) -- [ESP-EYE](https://github.com/espressif/esp-who/blob/master/docs/en/get-started/ESP-EYE_Getting_Started_Guide.md) - -### Install the ESP IDF - -Follow the instructions of the -[ESP-IDF get started guide](https://docs.espressif.com/projects/esp-idf/en/latest/get-started/index.html) -to setup the toolchain and the ESP-IDF itself. - -The next steps assume that the -[IDF environment variables are set](https://docs.espressif.com/projects/esp-idf/en/latest/get-started/index.html#step-4-set-up-the-environment-variables) : - - * The `IDF_PATH` environment variable is set - * `idf.py` and Xtensa-esp32 tools (e.g. `xtensa-esp32-elf-gcc`) are in `$PATH` - -### Generate the examples -The example project can be generated with the following command: -``` -make -f tensorflow/lite/micro/tools/make/Makefile TARGET=esp generate_hello_world_esp_project -``` - -### Building the example - -Go the the example project directory -``` -cd tensorflow/lite/micro/tools/make/gen/esp_xtensa-esp32/prj/hello_world/esp-idf -``` - -Then build with `idf.py` -``` -idf.py build -``` - -### Load and run the example - -To flash (replace `/dev/ttyUSB0` with the device serial port): -``` -idf.py --port /dev/ttyUSB0 flash -``` - -Monitor the serial output: -``` -idf.py --port /dev/ttyUSB0 monitor -``` - -Use `Ctrl+]` to exit. - -The previous two commands can be combined: -``` -idf.py --port /dev/ttyUSB0 flash monitor -``` - -## Deploy to himax WE1 EVB - -The following instructions will help you build and deploy this example to -[HIMAX WE1 EVB](https://github.com/HimaxWiseEyePlus/bsp_tflu/tree/master/HIMAX_WE1_EVB_board_brief) -board. To undstand more about using this board, please check -[HIMAX WE1 EVB user guide](https://github.com/HimaxWiseEyePlus/bsp_tflu/tree/master/HIMAX_WE1_EVB_user_guide). - -### Initial Setup - -To use the HIMAX WE1 EVB, please make sure following software are installed: - -#### MetaWare Development Toolkit - -See -[Install the Synopsys DesignWare ARC MetaWare Development Toolkit](/tensorflow/lite/micro/tools/make/targets/arc/README.md#install-the-synopsys-designware-arc-metaware-development-toolkit) -section for instructions on toolchain installation. - -#### Make Tool version - -A `'make'` tool is required for deploying Tensorflow Lite Micro -applications on HIMAX WE1 EVB, See -[Check make tool version](/tensorflow/lite/micro/tools/make/targets/arc/README.md#make-tool) -section for proper environment. - -#### Serial Terminal Emulation Application - -There are 2 main purposes for HIMAX WE1 EVB Debug UART port - -- print application output -- burn application to flash by using xmodem send application binary - -You can use any terminal emulation program (like [PuTTY](https://www.putty.org/) or [minicom](https://linux.die.net/man/1/minicom)). - - -### Generate Example Project - -The example project for HIMAX WE1 EVB platform can be generated with the following -command: - -Download related third party data - -``` -make -f tensorflow/lite/micro/tools/make/Makefile TARGET=himax_we1_evb third_party_downloads -``` - -Generate hello world project - -``` -make -f tensorflow/lite/micro/tools/make/Makefile generate_hello_world_make_project TARGET=himax_we1_evb TAGS=no_arc_mli -``` - -### Build and Burn Example - -Following the Steps to run hello world example at HIMAX WE1 EVB platform. - -1. Go to the generated example project directory. - - ``` - cd tensorflow/lite/micro/tools/make/gen/himax_we1_evb_arc/prj/hello_world/make - ``` - -2. Build the example using - - ``` - make app - ``` - -3. After example build finish, copy ELF file and map file to image generate tool directory. - image generate tool directory located at `'tensorflow/lite/micro/tools/make/downloads/himax_we1_sdk/image_gen_linux_v3/'` - - ``` - cp hello_world.elf himax_we1_evb.map ../../../../../downloads/himax_we1_sdk/image_gen_linux_v3/ - ``` - -4. Go to flash image generate tool directory. - - ``` - cd ../../../../../downloads/himax_we1_sdk/image_gen_linux_v3/ - ``` - -5. run image generate tool, generate flash image file. - - * Before running image generate tool, by typing `sudo chmod +x image_gen` - and `sudo chmod +x sign_tool` to make sure it is executable. - - ``` - image_gen -e hello_world.elf -m himax_we1_evb.map -o out.img - ``` - - -6. Download flash image file to HIMAX WE1 EVB by UART: - - * more detail about download image through UART can be found at [HIMAX WE1 EVB update Flash image](https://github.com/HimaxWiseEyePlus/bsp_tflu/tree/master/HIMAX_WE1_EVB_user_guide#flash-image-update) - -After these steps, press reset button on the HIMAX WE1 EVB, you will see application output in the serial -terminal. - -## Deploy to SparkFun Edge - -The following instructions will help you build and deploy this sample on the -[SparkFun Edge development board](https://sparkfun.com/products/15170). - -![Animation on SparkFun Edge](images/animation_on_sparkfun_edge.gif) - -If you're new to using this board, we recommend walking through the -[AI on a microcontroller with TensorFlow Lite and SparkFun Edge](https://codelabs.developers.google.com/codelabs/sparkfun-tensorflow) -codelab to get an understanding of the workflow. - -### Compile the binary - -The following command will download the required dependencies and then compile a -binary for the SparkFun Edge: - -``` -make -f tensorflow/lite/micro/tools/make/Makefile TARGET=sparkfun_edge hello_world_bin -``` - -The binary will be created in the following location: - -``` -tensorflow/lite/micro/tools/make/gen/sparkfun_edge_cortex-m4/bin/hello_world.bin -``` - -### Sign the binary - -The binary must be signed with cryptographic keys to be deployed to the device. -We'll now run some commands that will sign our binary so it can be flashed to -the SparkFun Edge. The scripts we are using come from the Ambiq SDK, which is -downloaded when the `Makefile` is run. - -Enter the following command to set up some dummy cryptographic keys we can use -for development: - -``` -cp tensorflow/lite/micro/tools/make/downloads/AmbiqSuite-Rel2.2.0/tools/apollo3_scripts/keys_info0.py \ -tensorflow/lite/micro/tools/make/downloads/AmbiqSuite-Rel2.2.0/tools/apollo3_scripts/keys_info.py -``` - -Next, run the following command to create a signed binary: - -``` -python3 tensorflow/lite/micro/tools/make/downloads/AmbiqSuite-Rel2.2.0/tools/apollo3_scripts/create_cust_image_blob.py \ ---bin tensorflow/lite/micro/tools/make/gen/sparkfun_edge_cortex-m4/bin/hello_world.bin \ ---load-address 0xC000 \ ---magic-num 0xCB \ --o main_nonsecure_ota \ ---version 0x0 -``` - -This will create the file `main_nonsecure_ota.bin`. We'll now run another -command to create a final version of the file that can be used to flash our -device with the bootloader script we will use in the next step: - -``` -python3 tensorflow/lite/micro/tools/make/downloads/AmbiqSuite-Rel2.2.0/tools/apollo3_scripts/create_cust_wireupdate_blob.py \ ---load-address 0x20000 \ ---bin main_nonsecure_ota.bin \ --i 6 \ --o main_nonsecure_wire \ ---options 0x1 -``` - -You should now have a file called `main_nonsecure_wire.bin` in the directory -where you ran the commands. This is the file we'll be flashing to the device. - -### Flash the binary - -Next, attach the board to your computer via a USB-to-serial adapter. - -**Note:** If you're using the [SparkFun Serial Basic Breakout](https://www.sparkfun.com/products/15096), -you should [install the latest drivers](https://learn.sparkfun.com/tutorials/sparkfun-serial-basic-ch340c-hookup-guide#drivers-if-you-need-them) -before you continue. - -Once connected, assign the USB device name to an environment variable: - -``` -export DEVICENAME=put your device name here -``` - -Set another variable with the baud rate: - -``` -export BAUD_RATE=921600 -``` - -Now, hold the button marked `14` on the device. While still holding the button, -hit the button marked `RST`. Continue holding the button marked `14` while -running the following command: - -``` -python3 tensorflow/lite/micro/tools/make/downloads/AmbiqSuite-Rel2.2.0/tools/apollo3_scripts/uart_wired_update.py \ --b ${BAUD_RATE} ${DEVICENAME} \ --r 1 \ --f main_nonsecure_wire.bin \ --i 6 -``` - -You should see a long stream of output as the binary is flashed to the device. -Once you see the following lines, flashing is complete: - -``` -Sending Reset Command. -Done. -``` - -If you don't see these lines, flashing may have failed. Try running through the -steps in [Flash the binary](#flash-the-binary) again (you can skip over setting -the environment variables). If you continue to run into problems, follow the -[AI on a microcontroller with TensorFlow Lite and SparkFun Edge](https://codelabs.developers.google.com/codelabs/sparkfun-tensorflow) -codelab, which includes more comprehensive instructions for the flashing -process. - -The binary should now be deployed to the device. Hit the button marked `RST` to -reboot the board. You should see the device's four LEDs flashing in sequence. - -Debug information is logged by the board while the program is running. To view -it, establish a serial connection to the board using a baud rate of `115200`. -On OSX and Linux, the following command should work: - -``` -screen ${DEVICENAME} 115200 -``` - -You will see a lot of output flying past! To stop the scrolling, hit `Ctrl+A`, -immediately followed by `Esc`. You can then use the arrow keys to explore the -output, which will contain the results of running inference on various `x` -values: - -``` -x_value: 1.1843798*2^2, y_value: -1.9542645*2^-1 -``` - -To stop viewing the debug output with `screen`, hit `Ctrl+A`, immediately -followed by the `K` key, then hit the `Y` key. - - -## Deploy to STM32F746 - -The following instructions will help you build and deploy the sample to the -[STM32F7 discovery kit](https://os.mbed.com/platforms/ST-Discovery-F746NG/) -using [ARM Mbed](https://github.com/ARMmbed/mbed-cli). - -![Animation on STM32F746](images/animation_on_STM32F746.gif) - -Before we begin, you'll need the following: - -- STM32F7 discovery kit board -- Mini-USB cable -- ARM Mbed CLI ([installation instructions](https://os.mbed.com/docs/mbed-os/v5.12/tools/installation-and-setup.html)) -- Python 2.7 and pip - -Since Mbed requires a special folder structure for projects, we'll first run a -command to generate a subfolder containing the required source files in this -structure: - -``` -make -f tensorflow/lite/micro/tools/make/Makefile TARGET=mbed TAGS="CMSIS disco_f746ng" generate_hello_world_mbed_project -``` - -This will result in the creation of a new folder: - -``` -tensorflow/lite/micro/tools/make/gen/mbed_cortex-m4/prj/hello_world/mbed -``` - -This folder contains all of the example's dependencies structured in the correct -way for Mbed to be able to build it. - -Change into the directory and run the following commands, making sure you are -using Python 2.7.15. - -First, tell Mbed that the current directory is the root of an Mbed project: - -``` -mbed config root . -``` - -Next, tell Mbed to download the dependencies and prepare to build: - -``` -mbed deploy -``` - -By default, Mbed will build the project using C++98. However, TensorFlow Lite -requires C++11. Run the following Python snippet to modify the Mbed -configuration files so that it uses C++11: - -``` -python -c 'import fileinput, glob; -for filename in glob.glob("mbed-os/tools/profiles/*.json"): - for line in fileinput.input(filename, inplace=True): - print line.replace("\"-std=gnu++98\"","\"-std=c++11\", \"-fpermissive\"")' - -``` - -Finally, run the following command to compile: - -``` -mbed compile -m DISCO_F746NG -t GCC_ARM -``` - -This should result in a binary at the following path: - -``` -./BUILD/DISCO_F746NG/GCC_ARM/mbed.bin -``` - -To deploy, plug in your STM board and copy the file to it. On MacOS, you can do -this with the following command: - -``` -cp ./BUILD/DISCO_F746NG/GCC_ARM/mbed.bin /Volumes/DIS_F746NG/ -``` - -Copying the file will initiate the flashing process. Once this is complete, you -should see an animation on the device's screen. - - -``` -screen /dev/tty.usbmodem14403 9600 -``` - -In addition to this animation, debug information is logged by the board while -the program is running. To view it, establish a serial connection to the board -using a baud rate of `9600`. On OSX and Linux, the following command should -work, replacing `/dev/tty.devicename` with the name of your device as it appears -in `/dev`: - -``` -screen /dev/tty.devicename 9600 -``` - -You will see a lot of output flying past! To stop the scrolling, hit `Ctrl+A`, -immediately followed by `Esc`. You can then use the arrow keys to explore the -output, which will contain the results of running inference on various `x` -values: - -``` -x_value: 1.1843798*2^2, y_value: -1.9542645*2^-1 -``` - -To stop viewing the debug output with `screen`, hit `Ctrl+A`, immediately -followed by the `K` key, then hit the `Y` key. - -### Run the tests on a development machine - -To compile and test this example on a desktop Linux or macOS machine, first -clone the TensorFlow repository from GitHub to a convenient place: - -```bash -git clone --depth 1 https://github.com/tensorflow/tensorflow.git -``` - -Next, `cd` into the source directory from a terminal, and then run the following -command: - -```bash -make -f tensorflow/lite/micro/tools/make/Makefile test_hello_world_test -``` - -This will take a few minutes, and downloads frameworks the code uses. Once the -process has finished, you should see a series of files get compiled, followed by -some logging output from a test, which should conclude with -`~~~ALL TESTS PASSED~~~`. - -If you see this, it means that a small program has been built and run that loads -the trained TensorFlow model, runs some example inputs through it, and got the -expected outputs. - -To understand how TensorFlow Lite does this, you can look at the source in -[hello_world_test.cc](hello_world_test.cc). -It's a fairly small amount of code that creates an interpreter, gets a handle to -a model that's been compiled into the program, and then invokes the interpreter -with the model and sample inputs. - -### Train your own model - -So far you have used an existing trained model to run inference on -microcontrollers. If you wish to train your own model, follow the instructions -given in the [train/](train/) directory. - From d6676205f20e6a9476f6e0eca8f5b00367f9c623 Mon Sep 17 00:00:00 2001 From: "902449@58880@bigcat_chen@ASIC" Date: Fri, 5 Jun 2020 10:16:04 +0800 Subject: [PATCH 051/867] TFLM: remove temp file in person detection example --- .../himax_we1_evb/image_provider.cc~ | 44 ------------------- 1 file changed, 44 deletions(-) delete mode 100644 tensorflow/lite/micro/examples/person_detection_experimental/himax_we1_evb/image_provider.cc~ diff --git a/tensorflow/lite/micro/examples/person_detection_experimental/himax_we1_evb/image_provider.cc~ b/tensorflow/lite/micro/examples/person_detection_experimental/himax_we1_evb/image_provider.cc~ deleted file mode 100644 index d5b4d136642..00000000000 --- a/tensorflow/lite/micro/examples/person_detection_experimental/himax_we1_evb/image_provider.cc~ +++ /dev/null @@ -1,44 +0,0 @@ -/* Copyright 2019 The TensorFlow Authors. All Rights Reserved. - -Licensed under the Apache License, Version 2.0 (the "License"); -you may not use this file except in compliance with the License. -You may obtain a copy of the License at - - http://www.apache.org/licenses/LICENSE-2.0 - -Unless required by applicable law or agreed to in writing, software -distributed under the License is distributed on an "AS IS" BASIS, -WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -See the License for the specific language governing permissions and -limitations under the License. -==============================================================================*/ - -#include "tensorflow/lite/micro/examples/person_detection_experimental/image_provider.h" - -#include "tensorflow/lite/micro/examples/person_detection_experimental/model_settings.h" - -#include "hx_drv_tflm.h" - -hx_drv_sensor_image_config_t g_pimg_config; - - -TfLiteStatus GetImage(tflite::ErrorReporter* error_reporter, int image_width, - int image_height, int channels, int8_t* image_data) { - static bool is_initialized = false; - - if (!is_initialized) { - if(hx_drv_sensor_initial(&g_pimg_config)!= HX_DRV_LIB_PASS) - { - return kTfLiteError; - } - is_initialized = true; - } - - hx_drv_sensor_capture(&g_pimg_config); - - hx_drv_image_rescale((uint8_t*)g_pimg_config.raw_address, g_pimg_config.img_width, g_pimg_config.img_height, - image_data, image_data, image_height); - - - return kTfLiteOk; -} From 4f222bed159eeb2743fb8c97fa5e36d5bea4e5da Mon Sep 17 00:00:00 2001 From: Nishidha Panpaliya Date: Fri, 5 Jun 2020 07:49:19 +0000 Subject: [PATCH 052/867] Fixed copyright --- tensorflow/compiler/xla/service/cpu/test_target_triple_helper.h | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/tensorflow/compiler/xla/service/cpu/test_target_triple_helper.h b/tensorflow/compiler/xla/service/cpu/test_target_triple_helper.h index e248f6de8bd..a28d48b7eb0 100644 --- a/tensorflow/compiler/xla/service/cpu/test_target_triple_helper.h +++ b/tensorflow/compiler/xla/service/cpu/test_target_triple_helper.h @@ -1,4 +1,4 @@ -/* Copyright 2017 The TensorFlow Authors. All Rights Reserved. +/* Copyright 2020 The TensorFlow Authors. All Rights Reserved. Licensed under the Apache License, Version 2.0 (the "License"); you may not use this file except in compliance with the License. From 66eb52eaee54576c7c1b5fe887d5d0400b557f86 Mon Sep 17 00:00:00 2001 From: Nishidha Panpaliya Date: Fri, 5 Jun 2020 08:19:55 +0000 Subject: [PATCH 053/867] Fixed build error in one of the xla tests --- tensorflow/compiler/xla/service/cpu/tests/BUILD | 1 + 1 file changed, 1 insertion(+) diff --git a/tensorflow/compiler/xla/service/cpu/tests/BUILD b/tensorflow/compiler/xla/service/cpu/tests/BUILD index 18624330a26..9036c5c9024 100644 --- a/tensorflow/compiler/xla/service/cpu/tests/BUILD +++ b/tensorflow/compiler/xla/service/cpu/tests/BUILD @@ -138,6 +138,7 @@ tf_cc_test( "//tensorflow/compiler/xla/service/cpu:cpu_compiler", "//tensorflow/compiler/xla/service/cpu/tests:cpu_codegen_test", "//tensorflow/compiler/xla/tests:test_utils", + "//tensorflow/compiler/xla/service/cpu:test_header_helper", "//tensorflow/core:lib", "//tensorflow/core:test", "//tensorflow/core:test_main", From d8bb6569bc6c820b098fc0b56c3c1a7a318422a9 Mon Sep 17 00:00:00 2001 From: Yixing Fu Date: Fri, 5 Jun 2020 18:55:32 -0400 Subject: [PATCH 054/867] correct summing total blocks --- tensorflow/python/keras/applications/efficientnet.py | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/tensorflow/python/keras/applications/efficientnet.py b/tensorflow/python/keras/applications/efficientnet.py index e1413b08533..e3c1a261e80 100644 --- a/tensorflow/python/keras/applications/efficientnet.py +++ b/tensorflow/python/keras/applications/efficientnet.py @@ -334,7 +334,7 @@ def EfficientNet( blocks_args = copy.deepcopy(blocks_args) b = 0 - blocks = float(sum(args['repeats'] for args in blocks_args)) + blocks = float(sum(round_repeats(args['repeats']) for args in blocks_args)) for (i, args) in enumerate(blocks_args): assert args['repeats'] > 0 # Update block input and output filters based on depth multiplier. From b8f57874ba13cf24a54878c3f4f6cd3f387a0c44 Mon Sep 17 00:00:00 2001 From: cclauss Date: Sun, 7 Jun 2020 15:56:59 +0200 Subject: [PATCH 055/867] Fix SyntaxWarnings on Python >= 3.8 --- tensorflow/python/kernel_tests/matrix_band_part_op_test.py | 2 +- tensorflow/python/kernel_tests/matrix_solve_ls_op_test.py | 2 +- .../python/ops/ragged/ragged_batch_gather_with_default_op.py | 3 +-- tensorflow/python/ops/random_ops.py | 4 ++-- tensorflow/tools/docs/doc_generator_visitor.py | 5 ++--- 5 files changed, 7 insertions(+), 9 deletions(-) diff --git a/tensorflow/python/kernel_tests/matrix_band_part_op_test.py b/tensorflow/python/kernel_tests/matrix_band_part_op_test.py index fdb7e4a1a4e..25b502cf814 100644 --- a/tensorflow/python/kernel_tests/matrix_band_part_op_test.py +++ b/tensorflow/python/kernel_tests/matrix_band_part_op_test.py @@ -56,7 +56,7 @@ def _GetMatrixBandPartTest(dtype_, batch_shape_, shape_): band_np = np.triu(band_np, -lower) if upper >= 0: band_np = np.tril(band_np, upper) - if batch_shape_ is not (): + if batch_shape_ != (): band_np = np.tile(band_np, batch_shape_ + (1, 1)) for index_dtype in [dtypes_lib.int32, dtypes_lib.int64]: with self.cached_session(use_gpu=False): diff --git a/tensorflow/python/kernel_tests/matrix_solve_ls_op_test.py b/tensorflow/python/kernel_tests/matrix_solve_ls_op_test.py index b7a159e2eff..889ea0dbd6c 100644 --- a/tensorflow/python/kernel_tests/matrix_solve_ls_op_test.py +++ b/tensorflow/python/kernel_tests/matrix_solve_ls_op_test.py @@ -107,7 +107,7 @@ class MatrixSolveLsOpTest(test_lib.TestCase): np_ans = _SolveWithNumpy(x, y, l2_regularizer=l2_regularizer) np_r = np.dot(np.conj(a.T), b - np.dot(a, np_ans)) np_r_norm = np.sqrt(np.sum(np.conj(np_r) * np_r)) - if batch_shape is not (): + if batch_shape != (): a = np.tile(a, batch_shape + (1, 1)) b = np.tile(b, batch_shape + (1, 1)) np_ans = np.tile(np_ans, batch_shape + (1, 1)) diff --git a/tensorflow/python/ops/ragged/ragged_batch_gather_with_default_op.py b/tensorflow/python/ops/ragged/ragged_batch_gather_with_default_op.py index 377fd84f96e..06690f86a50 100644 --- a/tensorflow/python/ops/ragged/ragged_batch_gather_with_default_op.py +++ b/tensorflow/python/ops/ragged/ragged_batch_gather_with_default_op.py @@ -81,8 +81,7 @@ def batch_gather_with_default(params, return_dtype=True)) # TODO(hterry): lift this restriction and support default_values of # of rank > 1 - if (default_value.shape.ndims is not 0 - and default_value.shape.ndims is not 1): + if default_value.shape.ndims not in (0, 1): raise ValueError('"default_value" must be a scalar or vector') upper_bounds = None if indices.shape.ndims is None: diff --git a/tensorflow/python/ops/random_ops.py b/tensorflow/python/ops/random_ops.py index 1af91ed0dd3..9932a76b678 100644 --- a/tensorflow/python/ops/random_ops.py +++ b/tensorflow/python/ops/random_ops.py @@ -288,8 +288,8 @@ def random_uniform(shape, shape = tensor_util.shape_tensor(shape) # In case of [0,1) floating results, minval and maxval is unused. We do an # `is` comparison here since this is cheaper than isinstance or __eq__. - minval_is_zero = minval is 0 # pylint: disable=literal-comparison - maxval_is_one = maxval is 1 # pylint: disable=literal-comparison + minval_is_zero = minval == 0 + maxval_is_one = maxval == 1 if not minval_is_zero or not maxval_is_one or dtype.is_integer: minval = ops.convert_to_tensor(minval, dtype=dtype, name="min") maxval = ops.convert_to_tensor(maxval, dtype=dtype, name="max") diff --git a/tensorflow/tools/docs/doc_generator_visitor.py b/tensorflow/tools/docs/doc_generator_visitor.py index ec2102a5935..ac5b09346ec 100644 --- a/tensorflow/tools/docs/doc_generator_visitor.py +++ b/tensorflow/tools/docs/doc_generator_visitor.py @@ -240,10 +240,9 @@ class DocGeneratorVisitor(object): # We cannot use the duplicate mechanism for some constants, since e.g., # id(c1) == id(c2) with c1=1, c2=1. This is unproblematic since constants # have no usable docstring and won't be documented automatically. - if (py_object is not None and + if (py_object not in (None, ()) not isinstance(py_object, six.integer_types + six.string_types + - (six.binary_type, six.text_type, float, complex, bool)) - and py_object is not ()): # pylint: disable=literal-comparison + (six.binary_type, six.text_type, float, complex, bool))): object_id = id(py_object) if object_id in reverse_index: master_name = reverse_index[object_id] From b9db1ee4174f26a94ac332ff8f60c9e0152403a8 Mon Sep 17 00:00:00 2001 From: "902449@58880@bigcat_chen@ASIC" Date: Mon, 8 Jun 2020 16:53:21 +0800 Subject: [PATCH 056/867] sync third_party_downloads to avoid conflict --- .../micro/tools/make/third_party_downloads.inc | 18 ++++++++++++------ 1 file changed, 12 insertions(+), 6 deletions(-) diff --git a/tensorflow/lite/micro/tools/make/third_party_downloads.inc b/tensorflow/lite/micro/tools/make/third_party_downloads.inc index 75a51e0df10..85016dc49b6 100644 --- a/tensorflow/lite/micro/tools/make/third_party_downloads.inc +++ b/tensorflow/lite/micro/tools/make/third_party_downloads.inc @@ -62,14 +62,14 @@ RUY_MD5="2d54f058f8f7120dfc1ecee79dbf259e" CIFAR10_DATASET_URL="https://www.cs.toronto.edu/~kriz/cifar-10-binary.tar.gz" CIFAR10_DATASET_MD5="c32a1d4ab5d03f1284b67883e8d87530" -IMAGE_RECOGNITION_MODEL_URL := "https://storage.googleapis.com/download.tensorflow.org/models/tflite/cifar_image_recognition_model_2020_4_14.zip" -IMAGE_RECOGNITION_MODEL_MD5 := "2b886156e7ef4d6e53d0f1a4bc800e56" +IMAGE_RECOGNITION_MODEL_URL := "https://storage.googleapis.com/download.tensorflow.org/models/tflite/cifar_image_recognition_model_2020_05_27.zip" +IMAGE_RECOGNITION_MODEL_MD5 := "1f4607b05ac45b8a6146fb883dbc2d7b" -PERSON_MODEL_URL := "https://storage.googleapis.com/download.tensorflow.org/data/tf_lite_micro_person_data_grayscale_2019_11_21.zip" -PERSON_MODEL_MD5 := "fe2934bd0788f1dcc7af3f0a954542ab" +PERSON_MODEL_URL := "https://storage.googleapis.com/download.tensorflow.org/data/tf_lite_micro_person_data_grayscale_2020_05_27.zip" +PERSON_MODEL_MD5 := "55b85f76e2995153e660391d4a209ef1" -PERSON_MODEL_INT8_URL := "https://storage.googleapis.com/download.tensorflow.org/data/tf_lite_micro_person_data_int8_grayscale_2020_01_13.zip" -PERSON_MODEL_INT8_MD5 := "8a7d2c70325f53136faea6dde517b8cc" +PERSON_MODEL_INT8_URL := "https://storage.googleapis.com/download.tensorflow.org/data/tf_lite_micro_person_data_int8_grayscale_2020_05_27.zip" +PERSON_MODEL_INT8_MD5 := "a0ede2d058aa2a1d413893455dd55352" EMBARC_MLI_URL := "https://github.com/foss-for-synopsys-dwc-arc-processors/embarc_mli/archive/58284867ca52d1f43b25045e8601999d7359d986.zip" EMBARC_MLI_MD5 := "2bf4982a327fdaa9d475803ce014d1ef" @@ -77,9 +77,15 @@ EMBARC_MLI_MD5 := "2bf4982a327fdaa9d475803ce014d1ef" EMBARC_MLI_PRE_COMPILED_URL := "https://github.com/foss-for-synopsys-dwc-arc-processors/embarc_mli/releases/download/Release_1.1_RC2/embARC_MLI_package.zip" EMBARC_MLI_PRE_COMPILED_MD5 := "a95ff9e0370434484f14e7e4114327f6" +ZEPHYR_URL := "https://github.com/antmicro/zephyr/archive/55e36b9.zip" +ZEPHYR_MD5 := "755622eb4812fde918a6382b65d50c3b" + XTENSA_HIFI4_URL :="https://github.com/foss-xtensa/nnlib-hifi4/raw/master/archive/xa_nnlib_04_07.zip" XTENSA_HIFI4_MD5 :="f234764928f9a42901df33a27e118c8b" +ETHOSU_URL := "https://git.mlplatform.org/ml/ethos-u/ethos-u-core-driver.git/snapshot/ethos-u-core-driver-bcb5aaa99756f1b5c1295b079ebdd60996bc75a5.tar.gz" +ETHOSU_MD5 := "d2073c8d88fc167fd5c46b5dcda58ea1" + HIMAX_WE1_SDK_URL ="https://www.himax.com.tw/we-i/himax_we1_sdk_v02.zip" HIMAX_WE1_SDK_MD5 ="9a4b2f29b16052764e437b64bdcba816" From 0c8343f7f3a8066c507b97ff84d8b298655cc5f4 Mon Sep 17 00:00:00 2001 From: Nathan Luehr Date: Mon, 8 Jun 2020 11:21:49 -0500 Subject: [PATCH 057/867] Address review comments --- tensorflow/core/platform/tf32_utils.cc | 10 ++++++---- tensorflow/core/platform/tf32_utils.h | 2 +- tensorflow/python/framework/config.py | 7 +++++-- 3 files changed, 12 insertions(+), 7 deletions(-) diff --git a/tensorflow/core/platform/tf32_utils.cc b/tensorflow/core/platform/tf32_utils.cc index 715b5996dc3..4456e768c0a 100644 --- a/tensorflow/core/platform/tf32_utils.cc +++ b/tensorflow/core/platform/tf32_utils.cc @@ -14,14 +14,16 @@ limitations under the License. ==============================================================================*/ #include "tensorflow/core/platform/tf32_utils.h" +#include namespace tensorflow { -// TODO(nluehr): enable tf32 execution by default after TF32 Ampere testing. -static bool tf32_enabled = false; +// Whether TensorFloat-32 should be used where supported. +// TODO(nluehr): Maybe enable by default after TF32 Ampere testing. +static std::atomic tf32_allowed{false}; -void allow_tf32_execution(bool allow) { tf32_enabled = allow; } +void allow_tf32_execution(bool allowed) { tf32_allowed = allowed; } -bool tf32_execution_allowed() { return tf32_enabled; } +bool tf32_execution_allowed() { return tf32_allowed; } } // namespace tensorflow diff --git a/tensorflow/core/platform/tf32_utils.h b/tensorflow/core/platform/tf32_utils.h index a0ce58f9bbd..7a158d00ad3 100644 --- a/tensorflow/core/platform/tf32_utils.h +++ b/tensorflow/core/platform/tf32_utils.h @@ -18,7 +18,7 @@ limitations under the License. namespace tensorflow { -void allow_tf32_execution(bool allow); +void allow_tf32_execution(bool allowed); bool tf32_execution_allowed(); diff --git a/tensorflow/python/framework/config.py b/tensorflow/python/framework/config.py index 042af4d1023..a356e6d9a16 100644 --- a/tensorflow/python/framework/config.py +++ b/tensorflow/python/framework/config.py @@ -23,6 +23,8 @@ from tensorflow.python.eager import context from tensorflow.python.util import deprecation from tensorflow.python.util.tf_export import tf_export + +# No tf_export until TF is built against CUDA11 which is required for TF32. def tensor_float32_execution_allowed(): """Get if TensorFloat-32 operations are enabled on supported hardware. @@ -31,7 +33,8 @@ def tensor_float32_execution_allowed(): """ return _pywrap_tf32_execution.is_allowed() -def allow_tensor_float_32_execution(allow): +# No tf_export until TF is built against CUDA11 which is required for TF32. +def allow_tensor_float_32_execution(allowed): """Allow use of TensorFloat-32 with float32 ops on supported hardware. TensorFloat-32 is a math mode introduced with the NVIDIA Ampere architecture. @@ -47,7 +50,7 @@ def allow_tensor_float_32_execution(allow): Args: allow: whether to allow TensorFloat-32 execution """ - _pywrap_tf32_execution.allow(allow) + _pywrap_tf32_execution.allow(allowed) @tf_export('config.threading.get_intra_op_parallelism_threads') def get_intra_op_parallelism_threads(): From ed41bb08250cad9f3ddbb6c7fb83e1216ee06031 Mon Sep 17 00:00:00 2001 From: Elena Zhelezina Date: Mon, 8 Jun 2020 17:57:27 +0100 Subject: [PATCH 058/867] Fix for CI failure. Change-Id: I66a5b5ab559207071ea62619e9e612fda9a73202 --- tensorflow/tools/api/golden/v1/tensorflow.lite.-ops-set.pbtxt | 4 ++++ .../tools/api/golden/v1/tensorflow.lite.constants.pbtxt | 4 ++++ tensorflow/tools/api/golden/v2/tensorflow.lite.-ops-set.pbtxt | 4 ++++ 3 files changed, 12 insertions(+) diff --git a/tensorflow/tools/api/golden/v1/tensorflow.lite.-ops-set.pbtxt b/tensorflow/tools/api/golden/v1/tensorflow.lite.-ops-set.pbtxt index c3199b24d98..9538fe382a0 100644 --- a/tensorflow/tools/api/golden/v1/tensorflow.lite.-ops-set.pbtxt +++ b/tensorflow/tools/api/golden/v1/tensorflow.lite.-ops-set.pbtxt @@ -1,6 +1,10 @@ path: "tensorflow.lite.OpsSet" tf_class { is_instance: "" + member { + name: "EXPERIMENTAL_TFLITE_BUILTINS_ACTIVATIONS_INT16_WEIGHTS_INT8" + mtype: "" + } member { name: "SELECT_TF_OPS" mtype: "" diff --git a/tensorflow/tools/api/golden/v1/tensorflow.lite.constants.pbtxt b/tensorflow/tools/api/golden/v1/tensorflow.lite.constants.pbtxt index 27c227dac64..7f62da6662a 100644 --- a/tensorflow/tools/api/golden/v1/tensorflow.lite.constants.pbtxt +++ b/tensorflow/tools/api/golden/v1/tensorflow.lite.constants.pbtxt @@ -12,6 +12,10 @@ tf_module { name: "GRAPHVIZ_DOT" mtype: "" } + member { + name: "INT16" + mtype: "" + } member { name: "INT32" mtype: "" diff --git a/tensorflow/tools/api/golden/v2/tensorflow.lite.-ops-set.pbtxt b/tensorflow/tools/api/golden/v2/tensorflow.lite.-ops-set.pbtxt index c3199b24d98..9538fe382a0 100644 --- a/tensorflow/tools/api/golden/v2/tensorflow.lite.-ops-set.pbtxt +++ b/tensorflow/tools/api/golden/v2/tensorflow.lite.-ops-set.pbtxt @@ -1,6 +1,10 @@ path: "tensorflow.lite.OpsSet" tf_class { is_instance: "" + member { + name: "EXPERIMENTAL_TFLITE_BUILTINS_ACTIVATIONS_INT16_WEIGHTS_INT8" + mtype: "" + } member { name: "SELECT_TF_OPS" mtype: "" From dbc7faeecdbdc223c67b03284ba5dc7d25668d3c Mon Sep 17 00:00:00 2001 From: Elena Zhelezina Date: Mon, 8 Jun 2020 20:20:22 +0100 Subject: [PATCH 059/867] Addressed reviewer's comment. Change-Id: I5bda332514d8070731b807b750ee7a423d6b4d78 --- tensorflow/lite/python/convert.py | 9 +++++++-- 1 file changed, 7 insertions(+), 2 deletions(-) diff --git a/tensorflow/lite/python/convert.py b/tensorflow/lite/python/convert.py index b1095a469f6..52edb700195 100644 --- a/tensorflow/lite/python/convert.py +++ b/tensorflow/lite/python/convert.py @@ -94,10 +94,15 @@ class OpsSet(enum.Enum): # quantized implementations. TFLITE_BUILTINS_INT8 = "TFLITE_BUILTINS_INT8" - # Convert model using only TensorFlow Lite operations with quantized int8 weights - # and int16 activations. + # Convert model using only TensorFlow Lite operations with quantized int8 weights, + # int16 activations and int64 bias. # Specifying this will throw an error for operations that do not yet have # quantized implementations. + # This quantization mode should be used in models for super-resolution, + # audio signal processing or image de-noising. It improves accuracy + # significantly, but only slightly increases the model size. + # WARNING: These ops are currently experimental and have not yet been finalized. + # They are only compatible with CPU execution, and have not been optimized for production. EXPERIMENTAL_TFLITE_BUILTINS_ACTIVATIONS_INT16_WEIGHTS_INT8 = "EXPERIMENTAL_TFLITE_BUILTINS_ACTIVATIONS_INT16_WEIGHTS_INT8" def __str__(self): From 13d3b343498d499f87230f2e596b738be5cf1109 Mon Sep 17 00:00:00 2001 From: "902449@58880@bigcat_chen@ASIC" Date: Tue, 9 Jun 2020 09:28:23 +0800 Subject: [PATCH 060/867] modify example main API usage --- .../himax_we1_evb/main_functions.cc | 17 ++++++++--------- 1 file changed, 8 insertions(+), 9 deletions(-) diff --git a/tensorflow/lite/micro/examples/person_detection_experimental/himax_we1_evb/main_functions.cc b/tensorflow/lite/micro/examples/person_detection_experimental/himax_we1_evb/main_functions.cc index 552b52c9c51..f0c7a405974 100644 --- a/tensorflow/lite/micro/examples/person_detection_experimental/himax_we1_evb/main_functions.cc +++ b/tensorflow/lite/micro/examples/person_detection_experimental/himax_we1_evb/main_functions.cc @@ -72,21 +72,20 @@ void setup() { // incur some penalty in code space for op implementations that are not // needed by this graph. // - // tflite::ops::micro::AllOpsResolver resolver; + // tflite::AllOpsResolver resolver; // NOLINTNEXTLINE(runtime-global-variables) - static tflite::MicroOpResolver<12> micro_op_resolver; - micro_op_resolver.AddBuiltin(tflite::BuiltinOperator_DEPTHWISE_CONV_2D, - tflite::ops::micro::Register_DEPTHWISE_CONV_2D(), - 1, 3); + static tflite::MicroMutableOpResolver<5> micro_op_resolver; + micro_op_resolver.AddBuiltin( + tflite::BuiltinOperator_DEPTHWISE_CONV_2D, + tflite::ops::micro::Register_DEPTHWISE_CONV_2D()); micro_op_resolver.AddBuiltin(tflite::BuiltinOperator_CONV_2D, - tflite::ops::micro::Register_CONV_2D(), 1, 3); + tflite::ops::micro::Register_CONV_2D()); micro_op_resolver.AddBuiltin(tflite::BuiltinOperator_AVERAGE_POOL_2D, - tflite::ops::micro::Register_AVERAGE_POOL_2D(), - 1, 2); + tflite::ops::micro::Register_AVERAGE_POOL_2D()); micro_op_resolver.AddBuiltin(tflite::BuiltinOperator_RESHAPE, tflite::ops::micro::Register_RESHAPE()); micro_op_resolver.AddBuiltin(tflite::BuiltinOperator_SOFTMAX, - tflite::ops::micro::Register_SOFTMAX(), 1, 3); + tflite::ops::micro::Register_SOFTMAX()); // Build an interpreter to run the model with. // NOLINTNEXTLINE(runtime-global-variables) From 4644b47db199ccda856c81ca88d6a3c58c41890f Mon Sep 17 00:00:00 2001 From: Eugene Kuznetsov Date: Tue, 9 Jun 2020 05:13:35 +0000 Subject: [PATCH 061/867] Remove duplicate macros --- .../core/kernels/non_max_suppression_op.cu.cc | 19 ------------------- 1 file changed, 19 deletions(-) diff --git a/tensorflow/core/kernels/non_max_suppression_op.cu.cc b/tensorflow/core/kernels/non_max_suppression_op.cu.cc index 53559b20419..c2cae2ab212 100644 --- a/tensorflow/core/kernels/non_max_suppression_op.cu.cc +++ b/tensorflow/core/kernels/non_max_suppression_op.cu.cc @@ -28,25 +28,6 @@ limitations under the License. #include "tensorflow/core/util/gpu_launch_config.h" #include "tensorflow/stream_executor/stream_executor.h" -#define TF_RETURN_IF_CUDA_ERROR(result) \ - do { \ - cudaError_t error(result); \ - if (!SE_PREDICT_TRUE(error == cudaSuccess)) { \ - return errors::Internal("Cuda call failed with ", \ - cudaGetErrorString(error)); \ - } \ - } while (0) - -#define TF_OP_REQUIRES_CUDA_SUCCESS(context, result) \ - do { \ - cudaError_t error(result); \ - if (!SE_PREDICT_TRUE(error == cudaSuccess)) { \ - context->SetStatus(errors::Internal("Cuda call failed with", \ - cudaGetErrorString(error))); \ - return; \ - } \ - } while (0) - struct __align__(16) Box { float x1, y1, x2, y2; }; From 8681b1bf543a33c62a9bec29625594d1fd9b921a Mon Sep 17 00:00:00 2001 From: Eugene Kuznetsov Date: Tue, 9 Jun 2020 15:11:45 +0000 Subject: [PATCH 062/867] Adding 3d Pooling using latest MIOpen API --- tensorflow/cc/gradients/nn_grad_test.cc | 8 - tensorflow/core/kernels/cudnn_pooling_gpu.cc | 31 ++ tensorflow/python/eager/backprop_test.py | 3 - tensorflow/python/kernel_tests/BUILD | 1 - tensorflow/python/kernel_tests/pool_test.py | 4 - .../ops/parallel_for/control_flow_ops_test.py | 2 - tensorflow/stream_executor/rocm/rocm_dnn.cc | 443 +++++++++++------- tensorflow/stream_executor/rocm/rocm_dnn.h | 68 +++ 8 files changed, 360 insertions(+), 200 deletions(-) diff --git a/tensorflow/cc/gradients/nn_grad_test.cc b/tensorflow/cc/gradients/nn_grad_test.cc index 942ec08f451..f5a09e09dcd 100644 --- a/tensorflow/cc/gradients/nn_grad_test.cc +++ b/tensorflow/cc/gradients/nn_grad_test.cc @@ -259,9 +259,6 @@ TEST_F(NNGradTest, MaxPoolGradV2Helper) { RunTest(x, x_init_value, y, y_shape); } -// TODO(rocm): -// Re-enable this test once 3D pooling is supported on ROCm platform -#ifndef TENSORFLOW_USE_ROCM TEST_F(NNGradTest, MaxPool3DGradHelper) { TensorShape x_shape({1, 3, 3, 3, 1}); TensorShape y_shape({1, 1, 1, 1, 1}); @@ -274,7 +271,6 @@ TEST_F(NNGradTest, MaxPool3DGradHelper) { SetRandomValuesForMaxPooling(&x_init_value); RunTest(x, x_init_value, y, y_shape); } -#endif TEST_F(NNGradTest, AvgPoolGradHelper) { TensorShape x_shape({1, 2, 2, 1}); @@ -287,9 +283,6 @@ TEST_F(NNGradTest, AvgPoolGradHelper) { RunTest(x, x_shape, y, y_shape); } -// TODO(rocm): -// Re-enable this test once 3D pooling is supported on ROCm platform -#ifndef TENSORFLOW_USE_ROCM TEST_F(NNGradTest, AvgPool3DGradHelper) { TensorShape x_shape({1, 3, 3, 3, 1}); TensorShape y_shape({1, 1, 1, 1, 1}); @@ -300,7 +293,6 @@ TEST_F(NNGradTest, AvgPool3DGradHelper) { auto y = AvgPool3D(scope_, x, ksize, strides, "SAME"); RunTest(x, x_shape, y, y_shape); } -#endif TEST_F(NNGradTest, LRN) { TensorShape x_shape({1, 1, 2, 1}); diff --git a/tensorflow/core/kernels/cudnn_pooling_gpu.cc b/tensorflow/core/kernels/cudnn_pooling_gpu.cc index eb7d16e3074..60088133d5d 100644 --- a/tensorflow/core/kernels/cudnn_pooling_gpu.cc +++ b/tensorflow/core/kernels/cudnn_pooling_gpu.cc @@ -98,10 +98,25 @@ void DnnPooling3dOp::Compute(OpKernelContext* context, auto* stream = context->op_device_context()->stream(); OP_REQUIRES(context, stream, errors::Internal("No GPU stream available.")); +#if TENSORFLOW_USE_ROCM + static int64 PoolingScratchSize = GetDnnWorkspaceLimit( + // default value is in bytes despite the name of the environment variable + "TF_CUDNN_WORKSPACE_LIMIT_IN_MB", 1LL << 32 // 4GB + ); + + DnnScratchAllocator scratch_allocator(PoolingScratchSize, context); + bool status = + stream + ->ThenPoolForward(pooling_desc, input_desc, input_data, output_desc, + &output_data, &scratch_allocator) + .ok(); +#else bool status = stream ->ThenPoolForward(pooling_desc, input_desc, input_data, output_desc, &output_data) .ok(); +#endif + OP_REQUIRES(context, status, errors::Internal("dnn PoolForward launch failed")); @@ -225,12 +240,28 @@ void DnnPooling3dGradOp::Compute( auto* stream = context->op_device_context()->stream(); OP_REQUIRES(context, stream, errors::Internal("No GPU stream available.")); +#if TENSORFLOW_USE_ROCM + static int64 PoolingScratchSize = GetDnnWorkspaceLimit( + // default value is in bytes despite the name of the environment variable + "TF_CUDNN_WORKSPACE_LIMIT_IN_MB", 1LL << 32 // 4GB + ); + + DnnScratchAllocator scratch_allocator(PoolingScratchSize, context); + bool status = stream + ->ThenPoolBackward(pooling_desc, orig_input_desc, + orig_input_data, orig_output_desc, + orig_output_data, output_backprop_data, + &input_backprop_data, &scratch_allocator) + .ok(); +#else bool status = stream ->ThenPoolBackward(pooling_desc, orig_input_desc, orig_input_data, orig_output_desc, orig_output_data, output_backprop_data, &input_backprop_data) .ok(); +#endif + OP_REQUIRES(context, status, errors::Internal("dnn PoolBackward launch failed")); diff --git a/tensorflow/python/eager/backprop_test.py b/tensorflow/python/eager/backprop_test.py index b28aaa3a626..bb909433fa7 100644 --- a/tensorflow/python/eager/backprop_test.py +++ b/tensorflow/python/eager/backprop_test.py @@ -1455,9 +1455,6 @@ class BackpropTest(test.TestCase, parameterized.TestCase): @test_util.run_in_graph_and_eager_modes def testMaxPooling3DGradient(self): - if test.is_built_with_rocm(): - self.skipTest('Pooling with 3D tensors is not supported in ROCm') - def forward(a): r = max_pooling3d(a, pool_size=pool_size, strides=strides, padding='SAME') return r diff --git a/tensorflow/python/kernel_tests/BUILD b/tensorflow/python/kernel_tests/BUILD index a04c874c9d6..ae4e0244357 100644 --- a/tensorflow/python/kernel_tests/BUILD +++ b/tensorflow/python/kernel_tests/BUILD @@ -2995,7 +2995,6 @@ cuda_py_test( name = "pooling_ops_3d_test", size = "medium", srcs = ["pooling_ops_3d_test.py"], - tags = ["no_rocm"], deps = [ "//tensorflow/python:client_testlib", "//tensorflow/python:framework_for_generated_wrappers", diff --git a/tensorflow/python/kernel_tests/pool_test.py b/tensorflow/python/kernel_tests/pool_test.py index 0f0eaa25402..01680f5c1f0 100644 --- a/tensorflow/python/kernel_tests/pool_test.py +++ b/tensorflow/python/kernel_tests/pool_test.py @@ -219,8 +219,6 @@ class PoolingTest(test.TestCase): strides=strides) def testPool3D(self): - if test.is_built_with_rocm(): - self.skipTest("Pooling with 3D tensors is not supported in ROCm") with self.session(use_gpu=test.is_gpu_available()): for padding in ["SAME", "VALID"]: for pooling_type in ["MAX", "AVG"]: @@ -363,8 +361,6 @@ class PoolingTest(test.TestCase): @test_util.run_deprecated_v1 def testGradient3D(self): - if test.is_built_with_rocm(): - self.skipTest("Pooling with 3D tensors is not supported in ROCm") with self.session(use_gpu=test.is_gpu_available()): for padding in ["SAME", "VALID"]: for pooling_type in ["AVG", "MAX"]: diff --git a/tensorflow/python/ops/parallel_for/control_flow_ops_test.py b/tensorflow/python/ops/parallel_for/control_flow_ops_test.py index 243471553d9..a6546fee742 100644 --- a/tensorflow/python/ops/parallel_for/control_flow_ops_test.py +++ b/tensorflow/python/ops/parallel_for/control_flow_ops_test.py @@ -488,8 +488,6 @@ class NNTest(PForTestCase): self._test_loop_fn(loop_fn, 3) def test_max_pool3d(self): - if test.is_built_with_rocm(): - self.skipTest("Pooling with 3D tensors is not supported in ROCm") with backprop.GradientTape(persistent=True) as g: x = random_ops.random_uniform([3, 3, 2, 12, 12, 3]) g.watch(x) diff --git a/tensorflow/stream_executor/rocm/rocm_dnn.cc b/tensorflow/stream_executor/rocm/rocm_dnn.cc index e0ead6d57e8..9c09784b3f1 100644 --- a/tensorflow/stream_executor/rocm/rocm_dnn.cc +++ b/tensorflow/stream_executor/rocm/rocm_dnn.cc @@ -263,7 +263,8 @@ namespace wrap { __macro(miopenFindConvolutionForwardAlgorithm) \ __macro(miopenCreateTensorDescriptor) \ __macro(miopenDestroyTensorDescriptor) \ - __macro(miopenSet2dPoolingDescriptor) \ + __macro(miopenSetNdPoolingDescriptor) \ + __macro(miopenSetPoolingIndexType) \ __macro(miopenSetLRNDescriptor) \ __macro(miopenLRNGetWorkSpaceSize) \ __macro(miopenCreateConvolutionDescriptor) \ @@ -290,7 +291,7 @@ namespace wrap { __macro(miopenSetTensorDescriptor) \ __macro(miopenGetTensorDescriptorSize) \ __macro(miopenPoolingForward) \ - __macro(miopenPoolingGetWorkSpaceSize) \ + __macro(miopenPoolingGetWorkSpaceSizeV2 \ __macro(miopenPoolingBackward) \ __macro(miopenLRNForward) \ __macro(miopenLRNBackward) \ @@ -605,6 +606,11 @@ MIOpenSupport::MIOpenSupport(GpuExecutor* parent) : parent_(parent) { // swich to Find Mode if env var TF_ROCM_USE_IMMEDIATE_MODE is set tensorflow::ReadBoolFromEnvVar("TF_ROCM_USE_IMMEDIATE_MODE", false, &use_immediate_mode_); + + bool enable_pooling_cache = false; + tensorflow::ReadBoolFromEnvVar("TF_ROCM_BW_POOL_CACHE", false, + &enable_pooling_cache); + if (enable_pooling_cache) m_pooling_cache_allowed = true; } port::Status MIOpenSupport::Init() { @@ -844,17 +850,19 @@ class ScopedPoolingDescriptor { std::transform(shape64.cbegin(), shape64.cend(), shape.begin(), &CheckedNarrowing); - if (nd != 2) { - LOG(FATAL) << "miopen requires pooling dimensions be 2" - << ToString(status); - } - - status = wrap::miopenSet2dPoolingDescriptor( + status = wrap::miopenSetNdPoolingDescriptor( handle_, (pooling_descriptor.mode() == dnn::PoolingMode::kMaximum ? miopenPoolingMax : miopenPoolingAverage), - shape[0], shape[1], padding[0], padding[1], strides[0], strides[1]); + nd, shape.data(), padding.data(), strides.data()); + + // Note: The index type has to be uint32 type for now because MIOpen + // API assumes all input indexes to be the same type. Since a tensor + // descriptor can only use int32 type, the index type here need to be + // aligned with the tensor index type of the (input) tensor descritptor + status = wrap::miopenSetPoolingIndexType(handle_, miopenIndexUint32); + if (status != miopenStatusSuccess) { LOG(FATAL) << "could not set miopen pooling descriptor: " << ToString(status); @@ -4009,10 +4017,94 @@ bool MIOpenSupport::DoPoolForward( const DeviceMemory& input_data, const dnn::BatchDescriptor& output_dimensions, DeviceMemory* output_data, ScratchAllocator* workspace_allocator) { - LOG(ERROR) << "miopen does not support pooling for dobule type yet"; + LOG(ERROR) << "miopen does not support pooling for double type yet"; return false; } +bool PoolingWorkspaceDescriptor::IsSame( + const dnn::BatchDescriptor& input_dimensions, + const dnn::BatchDescriptor& output_dimensions, + const dnn::PoolingDescriptor& pooling_dimensions, int _type) { + return dtype == _type && + input_dims == + input_dimensions.full_dims(dnn::DataLayout::kBatchDepthYX) && + output_dims == + output_dimensions.full_dims(dnn::DataLayout::kBatchDepthYX) && + op.mode() == pooling_dimensions.mode() && + op.window() == pooling_dimensions.window() && + op.padding() == pooling_dimensions.padding() && + op.strides() == pooling_dimensions.strides(); +} + +bool PoolingWorkspaceCache::find( + const void* p, const dnn::BatchDescriptor& input_dimensions, + const dnn::BatchDescriptor& output_dimensions, + const dnn::PoolingDescriptor& pooling_dimensions, int _type, + PoolingWorkspaceDescriptor*& pdesc) { + pdesc = 0; + auto it = cache.find(p); + if (it == cache.end()) { + return false; + } + if (!it->second.IsSame(input_dimensions, output_dimensions, + pooling_dimensions, _type)) { + return false; + } + pdesc = &it->second; + return true; +} + +void PoolingWorkspaceCache::insert( + const void* p, const dnn::BatchDescriptor& input_dimensions, + const dnn::BatchDescriptor& output_dimensions, + const dnn::PoolingDescriptor& pooling_dimensions, int _type, + std::unique_ptr>& workspace, size_t wsp_size, + hipStream_t hip_stream) { + PoolingWorkspaceDescriptor* desc = 0; + auto it = cache.find(p); + if (it != cache.end()) { + // replacing an entry with the same pointer but different attributes + // (if everything matches, the caller is expected to reuse the entry) + desc = &it->second; + hipStreamSynchronize(hip_stream); + memory_used -= desc->workspace_size; + } else { + cache[p] = PoolingWorkspaceDescriptor(); + desc = &cache[p]; + } + desc->input_dims = input_dimensions.full_dims(dnn::DataLayout::kBatchDepthYX); + desc->output_dims = + output_dimensions.full_dims(dnn::DataLayout::kBatchDepthYX); + desc->op = pooling_dimensions; + desc->dtype = _type; + desc->timestamp = timestamp; + timestamp++; + desc->workspace = std::move(workspace); + desc->workspace_size = wsp_size; + memory_used += wsp_size; + trim(hip_stream); +} + +void PoolingWorkspaceCache::trim(hipStream_t hip_stream) { + if (memory_used < memory_budget && cache.size() < trim_size) return; + bool must_sync = true; + while (true) { + int new_size = cache.size() - (cache.size() >> 2); + std::vector old_entries; + for (auto& x : cache) + if (x.second.timestamp + new_size < timestamp) + old_entries.push_back(x.first); + if (old_entries.empty()) break; + if (must_sync) hipStreamSynchronize(hip_stream); + must_sync = true; + for (auto x : old_entries) { + memory_used -= cache[x].workspace_size; + cache.erase(x); + } + if (memory_used < memory_budget || cache.size() < 10) break; + } +} + bool MIOpenSupport::DoPoolForward( Stream* stream, const dnn::PoolingDescriptor& pooling_dimensions, const dnn::BatchDescriptor& input_dimensions, @@ -4020,7 +4112,6 @@ bool MIOpenSupport::DoPoolForward( const dnn::BatchDescriptor& output_dimensions, DeviceMemory* output_data, ScratchAllocator* workspace_allocator) { auto miopen = miopen_->GetHandle(parent_, stream); - // Alpha is the scaling factor for input. float alpha = 1.0; // Beta is the scaling factor for output. @@ -4030,10 +4121,48 @@ bool MIOpenSupport::DoPoolForward( ScopedTensorDescriptor dest_desc{output_dimensions, miopenFloat}; ScopedPoolingDescriptor pooling_desc{pooling_dimensions}; + bool do_backward = false; + uint8* workspace = 0; + size_t workspace_size = 0; + std::unique_ptr> wsp_mem; + if (m_pooling_cache_enabled) { + do_backward = true; + auto status = wrap::miopenPoolingGetWorkSpaceSizeV2( + pooling_desc.handle(), dest_desc.handle(), &workspace_size); + if (status != miopenStatusSuccess) { + LOG(ERROR) + << "failed to obtain workspace size for backward pooling on stream: " + << ToString(status); + return false; + } + if (workspace_size != 0) { + PoolingWorkspaceDescriptor* pdesc = 0; + bool cache_hit = + m_pooling_cache_allowed && + m_pooling_cache.find(input_data.opaque(), input_dimensions, + output_dimensions, pooling_dimensions, + miopenFloat, pdesc); + if (cache_hit) { + // reusing the same buffer + workspace = reinterpret_cast( + pdesc->workspace->mutable_device_memory()->opaque()); + } else { + wsp_mem = stream->AllocateTemporaryArray(workspace_size) + .ConsumeValueOrDie(); + workspace = reinterpret_cast( + wsp_mem->mutable_device_memory()->opaque()); + m_pooling_cache.insert(input_data.opaque(), input_dimensions, + output_dimensions, pooling_dimensions, + miopenFloat, wsp_mem, workspace_size, + AsGpuStreamValue(stream)); + } + } + } + auto status = wrap::miopenPoolingForward( miopen.handle(), pooling_desc.handle(), &alpha, src_desc.handle(), input_data.opaque(), &beta, dest_desc.handle(), output_data->opaque(), - false, nullptr, 0); + do_backward, workspace, workspace_size); if (status != miopenStatusSuccess) { LOG(ERROR) << "failed to enqueue forward pooling on stream: " << ToString(status); @@ -4072,6 +4201,118 @@ bool MIOpenSupport::DoPoolForward( return true; } +template +bool MIOpenSupport::DoPoolBackwardImpl( + Stream* stream, const dnn::PoolingDescriptor& pooling_dimensions, + const dnn::BatchDescriptor& input_dimensions, + const DeviceMemory& input_data, + const dnn::BatchDescriptor& output_dimensions, + const DeviceMemory& output_data, const DeviceMemory& input_diff_data, + DeviceMemory* output_diff_data, ScratchAllocator* workspace_allocator) { + auto miopen = miopen_->GetHandle(parent_, stream); + if (m_pooling_cache_allowed) m_pooling_cache_enabled = true; + // Alpha is the scaling factor for input. + float alpha = 1.0; + // Beta is the scaling factor for output. + float beta = 0.0; + + auto type = + std::is_same::value + ? miopenFloat + : (std::is_same::value ? miopenHalf + : (miopenDataType_t)-1); + + ScopedTensorDescriptor src_desc{input_dimensions, type}; + ScopedTensorDescriptor dest_desc{output_dimensions, type}; + ScopedPoolingDescriptor pooling_desc{pooling_dimensions}; + + uint8* workspace_ptr = 0; + DeviceMemory workspace; + PoolingWorkspaceDescriptor* pdesc = 0; + + size_t workspace_size_in_bytes = 0; + auto status = wrap::miopenPoolingGetWorkSpaceSizeV2( + pooling_desc.handle(), dest_desc.handle(), &workspace_size_in_bytes); + if (status != miopenStatusSuccess) { + LOG(ERROR) + << "failed to obtain workspace size for backward pooling on stream: " + << ToString(status); + return false; + } + + // Allocate the workspace. + if (workspace_size_in_bytes > 0) { + bool cache_hit = m_pooling_cache_allowed && + m_pooling_cache.find(input_data.opaque(), input_dimensions, + output_dimensions, pooling_dimensions, + type, pdesc); + if (cache_hit) { + assert(pdesc != 0); + workspace_ptr = reinterpret_cast( + pdesc->workspace->mutable_device_memory()->opaque()); + VLOG(1) << "Pooling cache hit"; + } else { + VLOG(1) << "Pooling cache miss"; + assert(workspace_allocator); + auto allocated = + workspace_allocator->AllocateBytes(workspace_size_in_bytes); + if (!allocated.ok() || (workspace = allocated.ValueOrDie()) == nullptr) { + LOG(ERROR) << "Failed to allocate backward pooling workspace"; + return false; + } + DeviceMemory dest2; // duplicated dest from forward: + int64 dest2_size = 0; + + // miopen requires the strides and dims to be ordered as BDYX. + std::vector dims64 = + output_dimensions.full_dims(dnn::DataLayout::kBatchDepthYX); + // miopen does not use strides and must have 4D tensor. + // std::vector dims(pooling_dimensions.ndims() + 2); + + dest2_size = sizeof(T); + for (auto& x : dims64) dest2_size *= x; + + if (dest2_size > 0) { + assert(workspace_allocator); + auto allocated = workspace_allocator->AllocateBytes(dest2_size); + if (!allocated.ok() || (dest2 = allocated.ValueOrDie()) == nullptr) { + LOG(ERROR) << "Failed to allocate backward pooling workspace"; + return false; + } + } else { + LOG(ERROR) << "Failed to calculate tensor size to chain forward and " + "backward pooling"; + } + + status = wrap::miopenPoolingForward( + miopen.handle(), pooling_desc.handle(), &alpha, src_desc.handle(), + input_data.opaque(), &beta, dest_desc.handle(), dest2.opaque(), true, + workspace.opaque(), workspace_size_in_bytes); + + if (status != miopenStatusSuccess) { + LOG(ERROR) + << "failed to enqueue forward pooling (before backward) on stream: " + << ToString(status); + return false; + } + workspace_ptr = reinterpret_cast(workspace.opaque()); + } + } + status = wrap::miopenPoolingBackward( + miopen.handle(), pooling_desc.handle(), &alpha, dest_desc.handle(), + output_data.opaque(), dest_desc.handle(), input_diff_data.opaque(), + src_desc.handle(), input_data.opaque(), &beta, src_desc.handle(), + output_diff_data->opaque(), workspace_ptr); + + if (status != miopenStatusSuccess) { + LOG(ERROR) << "failed to enqueue backward pooling on stream: " + << ToString(status); + return false; + } + + return true; +} + bool MIOpenSupport::DoPoolBackward( Stream* stream, const dnn::PoolingDescriptor& pooling_dimensions, const dnn::BatchDescriptor& input_dimensions, @@ -4094,91 +4335,10 @@ bool MIOpenSupport::DoPoolBackward( const DeviceMemory& input_diff_data, DeviceMemory* output_diff_data, ScratchAllocator* workspace_allocator) { - auto miopen = miopen_->GetHandle(parent_, stream); - - // Alpha is the scaling factor for input. - float alpha = 1.0; - // Beta is the scaling factor for output. - float beta = 0.0; - - ScopedTensorDescriptor src_desc{input_dimensions, miopenFloat}; - ScopedTensorDescriptor dest_desc{output_dimensions, miopenFloat}; - ScopedPoolingDescriptor pooling_desc{pooling_dimensions}; - - DeviceMemory workspace; - size_t workspace_size_in_bytes = 0; - auto status = wrap::miopenPoolingGetWorkSpaceSize(dest_desc.handle(), - &workspace_size_in_bytes); - - if (status != miopenStatusSuccess) { - LOG(ERROR) - << "failed to obtain workspace size for backward pooling on stream: " - << ToString(status); - return false; - } - - // Allocate the workspace. - if (workspace_size_in_bytes > 0) { - assert(workspace_allocator); - auto allocated = - workspace_allocator->AllocateBytes(workspace_size_in_bytes); - if (!allocated.ok() || (workspace = allocated.ValueOrDie()) == nullptr) { - LOG(ERROR) << "Failed to allocate backward pooling workspace"; - return false; - } - } - - DeviceMemory dest2; // duplicated dest from forward: - int dest2_size = 0; - - // miopen requires the strides and dims to be ordered as BDYX. - std::vector dims64 = - output_dimensions.full_dims(dnn::DataLayout::kBatchDepthYX); - - // miopen does not use strides and must have 4D tensor. - std::vector dims(4); - - std::transform(dims64.cbegin(), dims64.cend(), dims.begin(), - &CheckedNarrowing); - - dest2_size = dims[0] * dims[1] * dims[2] * dims[3] * sizeof(float); - - if (dest2_size > 0) { - assert(workspace_allocator); - auto allocated = workspace_allocator->AllocateBytes(dest2_size); - if (!allocated.ok() || (dest2 = allocated.ValueOrDie()) == nullptr) { - LOG(ERROR) << "Failed to allocate backward pooling workspace"; - return false; - } - } else { - LOG(ERROR) << "Failed to calculate tensor size to chain forward and " - "backward pooling"; - } - - status = wrap::miopenPoolingForward( - miopen.handle(), pooling_desc.handle(), &alpha, src_desc.handle(), - input_data.opaque(), &beta, dest_desc.handle(), dest2.opaque(), true, - workspace.opaque(), workspace_size_in_bytes); - - if (status != miopenStatusSuccess) { - LOG(ERROR) - << "failed to enqueue forward pooling (before backward) on stream: " - << ToString(status); - return false; - } - - status = wrap::miopenPoolingBackward( - miopen.handle(), pooling_desc.handle(), &alpha, dest_desc.handle(), - dest2.opaque(), dest_desc.handle(), input_diff_data.opaque(), - src_desc.handle(), input_data.opaque(), &beta, src_desc.handle(), - output_diff_data->opaque(), workspace.opaque()); - - if (status != miopenStatusSuccess) { - LOG(ERROR) << "failed to enqueue backward pooling on stream: " - << ToString(status); - return false; - } - return true; + return DoPoolBackwardImpl(stream, pooling_dimensions, input_dimensions, + input_data, output_dimensions, output_data, + input_diff_data, output_diff_data, + workspace_allocator); } bool MIOpenSupport::DoPoolBackward( @@ -4190,91 +4350,10 @@ bool MIOpenSupport::DoPoolBackward( const DeviceMemory& input_diff_data, DeviceMemory* output_diff_data, ScratchAllocator* workspace_allocator) { - auto miopen = miopen_->GetHandle(parent_, stream); - - // Alpha is the scaling factor for input. - float alpha = 1.0; - // Beta is the scaling factor for output. - float beta = 0.0; - - ScopedTensorDescriptor src_desc{input_dimensions, miopenHalf}; - ScopedTensorDescriptor dest_desc{output_dimensions, miopenHalf}; - ScopedPoolingDescriptor pooling_desc{pooling_dimensions}; - - DeviceMemory workspace; - size_t workspace_size_in_bytes = 0; - auto status = wrap::miopenPoolingGetWorkSpaceSize(dest_desc.handle(), - &workspace_size_in_bytes); - - if (status != miopenStatusSuccess) { - LOG(ERROR) - << "failed to obtain workspace size for backward pooling on stream: " - << ToString(status); - return false; - } - - // Allocate the workspace. - if (workspace_size_in_bytes > 0) { - assert(workspace_allocator); - auto allocated = - workspace_allocator->AllocateBytes(workspace_size_in_bytes); - if (!allocated.ok() || (workspace = allocated.ValueOrDie()) == nullptr) { - LOG(ERROR) << "Failed to allocate backward pooling workspace"; - return false; - } - } - - DeviceMemory dest2; // duplicated dest from forward: - int dest2_size = 0; - - // miopen requires the strides and dims to be ordered as BDYX. - std::vector dims64 = - output_dimensions.full_dims(dnn::DataLayout::kBatchDepthYX); - - // miopen does not use strides and must have 4D tensor. - std::vector dims(4); - - std::transform(dims64.cbegin(), dims64.cend(), dims.begin(), - &CheckedNarrowing); - - dest2_size = dims[0] * dims[1] * dims[2] * dims[3] * sizeof(float); - - if (dest2_size > 0) { - assert(workspace_allocator); - auto allocated = workspace_allocator->AllocateBytes(dest2_size); - if (!allocated.ok() || (dest2 = allocated.ValueOrDie()) == nullptr) { - LOG(ERROR) << "Failed to allocate backward pooling workspace"; - return false; - } - } else { - LOG(ERROR) << "Failed to calculate tensor size to chain forward and " - "backward pooling"; - } - - status = wrap::miopenPoolingForward( - miopen.handle(), pooling_desc.handle(), &alpha, src_desc.handle(), - input_data.opaque(), &beta, dest_desc.handle(), dest2.opaque(), true, - workspace.opaque(), workspace_size_in_bytes); - - if (status != miopenStatusSuccess) { - LOG(ERROR) - << "failed to enqueue forward pooling (before backward) on stream: " - << ToString(status); - return false; - } - - status = wrap::miopenPoolingBackward( - miopen.handle(), pooling_desc.handle(), &alpha, dest_desc.handle(), - dest2.opaque(), dest_desc.handle(), input_diff_data.opaque(), - src_desc.handle(), input_data.opaque(), &beta, src_desc.handle(), - output_diff_data->opaque(), workspace.opaque()); - - if (status != miopenStatusSuccess) { - LOG(ERROR) << "failed to enqueue backward pooling on stream: " - << ToString(status); - return false; - } - return true; + return DoPoolBackwardImpl(stream, pooling_dimensions, input_dimensions, + input_data, output_dimensions, output_data, + input_diff_data, output_diff_data, + workspace_allocator); } bool MIOpenSupport::DoNormalizeWithDimensions( diff --git a/tensorflow/stream_executor/rocm/rocm_dnn.h b/tensorflow/stream_executor/rocm/rocm_dnn.h index 40e156b5f74..4f568702d96 100644 --- a/tensorflow/stream_executor/rocm/rocm_dnn.h +++ b/tensorflow/stream_executor/rocm/rocm_dnn.h @@ -20,6 +20,7 @@ limitations under the License. #define TENSORFLOW_STREAM_EXECUTOR_ROCM_ROCM_DNN_H_ #include "absl/synchronization/mutex.h" +#include "rocm/include/miopen/miopen.h" #include "tensorflow/core/platform/thread_annotations.h" #include "tensorflow/stream_executor/dnn.h" #include "tensorflow/stream_executor/lib/status.h" @@ -38,6 +39,39 @@ class MIOpenCTCLossDescriptor; // Opaque and unique identifier for the MIOpen plugin. extern const PluginId kMIOpenPlugin; +struct PoolingWorkspaceDescriptor { + std::vector input_dims; + std::vector output_dims; + dnn::PoolingDescriptor op; + int dtype; + uint64_t timestamp; + std::unique_ptr> workspace; + size_t workspace_size; + bool IsSame(const dnn::BatchDescriptor& input_dimensions, + const dnn::BatchDescriptor& output_dimensions, + const dnn::PoolingDescriptor& pooling_dimensions, int _type); +}; + +struct PoolingWorkspaceCache { + std::map cache; + const int trim_size = 1000; + const uint64_t memory_budget = 2e7; + uint64_t timestamp = 0; + uint64_t memory_used = 0; + bool find(const void* p, const dnn::BatchDescriptor& input_dimensions, + const dnn::BatchDescriptor& output_dimensions, + const dnn::PoolingDescriptor& pooling_dimensions, int _type, + PoolingWorkspaceDescriptor*& pdesc); + void insert(const void* p, const dnn::BatchDescriptor& input_dimensions, + const dnn::BatchDescriptor& output_dimensions, + const dnn::PoolingDescriptor& pooling_dimensions, int _type, + std::unique_ptr>& workspace, + size_t wsp_size, hipStream_t hip_stream); + + private: + void trim(hipStream_t hip_stream); +}; + // miopen-library based DNN support. For details on overridden interface // functions, see dnn.h. class MIOpenSupport : public dnn::DnnSupport { @@ -664,6 +698,10 @@ class MIOpenSupport : public dnn::DnnSupport { // Provide access to the MIOpen handle. std::unique_ptr miopen_; + PoolingWorkspaceCache m_pooling_cache; + bool m_pooling_cache_allowed = false; + bool m_pooling_cache_enabled = false; + template bool DoBatchNormalizationForwardImpl( Stream* stream, dnn::DataType input_data_type, @@ -847,6 +885,36 @@ class MIOpenSupport : public dnn::DnnSupport { ScratchAllocator* scratch_allocator, std::vector* out_algorithms); + port::Status DoCtcLossImpl( + Stream* stream, const MIOpenRnnStateTensorDescriptor& probs_desc, + const DeviceMemoryBase probs_data, absl::Span labels_data, + absl::Span labels_lengths_data, + absl::Span input_lengths_data, DeviceMemoryBase costs_data, + const MIOpenRnnStateTensorDescriptor& grads_desc, + DeviceMemoryBase grads_data, const MIOpenCTCLossDescriptor& ctc_loss_desc, + DeviceMemory scratch_memory); + + port::Status DoPrepareForCtcLoss( + Stream* stream, dnn::DataType element_type, + const dnn::RnnStateTensorDescriptor& probs_desc, + const dnn::RnnStateTensorDescriptor& grads_desc, + absl::Span labels_data, + absl::Span labels_lengths_data, + absl::Span input_lengths_data, + ScratchAllocator* scratch_allocator, + DeviceMemory* scratch_memory) override; + + template + bool DoPoolBackwardImpl(Stream* stream, + const dnn::PoolingDescriptor& pooling_dimensions, + const dnn::BatchDescriptor& input_dimensions, + const DeviceMemory& input_data, + const dnn::BatchDescriptor& output_dimensions, + const DeviceMemory& output_data, + const DeviceMemory& input_diff_data, + DeviceMemory* output_diff_data, + ScratchAllocator* workspace_allocator = nullptr); + SE_DISALLOW_COPY_AND_ASSIGN(MIOpenSupport); }; From 507c7549317221bcf5b418a66fd0212cd4a7443b Mon Sep 17 00:00:00 2001 From: Elena Zhelezina Date: Tue, 9 Jun 2020 17:47:16 +0100 Subject: [PATCH 063/867] Fix for pylint errors. Change-Id: Idd96d7a41fd459c86ab0f6fbb63e5d543509145d --- tensorflow/lite/python/convert.py | 3 ++- tensorflow/lite/python/lite.py | 27 ++++++++++--------- tensorflow/lite/python/lite_test.py | 10 ++++--- tensorflow/lite/python/optimize/calibrator.py | 3 ++- .../lite/python/optimize/calibrator_test.py | 3 ++- 5 files changed, 28 insertions(+), 18 deletions(-) diff --git a/tensorflow/lite/python/convert.py b/tensorflow/lite/python/convert.py index 52edb700195..68e23634b2e 100644 --- a/tensorflow/lite/python/convert.py +++ b/tensorflow/lite/python/convert.py @@ -103,7 +103,8 @@ class OpsSet(enum.Enum): # significantly, but only slightly increases the model size. # WARNING: These ops are currently experimental and have not yet been finalized. # They are only compatible with CPU execution, and have not been optimized for production. - EXPERIMENTAL_TFLITE_BUILTINS_ACTIVATIONS_INT16_WEIGHTS_INT8 = "EXPERIMENTAL_TFLITE_BUILTINS_ACTIVATIONS_INT16_WEIGHTS_INT8" + EXPERIMENTAL_TFLITE_BUILTINS_ACTIVATIONS_INT16_WEIGHTS_INT8 = \ + "EXPERIMENTAL_TFLITE_BUILTINS_ACTIVATIONS_INT16_WEIGHTS_INT8" def __str__(self): return self.value diff --git a/tensorflow/lite/python/lite.py b/tensorflow/lite/python/lite.py index 26c6f0855af..bed48860b00 100644 --- a/tensorflow/lite/python/lite.py +++ b/tensorflow/lite/python/lite.py @@ -251,7 +251,8 @@ class QuantizationMode(object): self.post_training_fp16()) def activations_type(self): - return constants.INT16 if self._is_int16x8_target_required() else constants.INT8 + return constants.INT16 if self._is_int16x8_target_required() \ + else constants.INT8 def converter_flags(self, inference_ty=None, inference_input_ty=None): """Flags to the converter.""" @@ -262,7 +263,8 @@ class QuantizationMode(object): if self.training_time_int8_allow_float(): return { - "inference_type": inference_ty if inference_ty else self.activations_type(), + "inference_type": inference_ty if inference_ty else \ + self.activations_type(), "inference_input_type": inference_input_ty if inference_input_ty else constants.FLOAT, "post_training_quantize": False, # disable dynamic range quantization @@ -359,15 +361,15 @@ class QuantizationMode(object): def _is_int16x8_target_required(self): return bool( - set(self._target_spec.supported_ops).intersection([ - OpsSet.EXPERIMENTAL_TFLITE_BUILTINS_ACTIVATIONS_INT16_WEIGHTS_INT8 - ])) + set(self._target_spec.supported_ops).intersection([ + OpsSet.EXPERIMENTAL_TFLITE_BUILTINS_ACTIVATIONS_INT16_WEIGHTS_INT8 + ])) def _is_allow_float(self): return bool( - set(self._target_spec.supported_ops).intersection([ - OpsSet.TFLITE_BUILTINS - ])) + set(self._target_spec.supported_ops).intersection([ + OpsSet.TFLITE_BUILTINS + ])) def _any_optimization_enabled(self): return bool( @@ -441,7 +443,8 @@ class TFLiteConverterBase(object): return _get_grappler_config(optimizers) def _calibrate_quantize_model(self, result, inference_input_type, - inference_output_type, activations_type, allow_float): + inference_output_type, activations_type, + allow_float): """Calibrate and quantize the model.""" if not isinstance(self.representative_dataset, RepresentativeDataset): self.representative_dataset = RepresentativeDataset( @@ -458,8 +461,8 @@ class TFLiteConverterBase(object): return _mlir_quantize(calibrated) else: return calibrate_quantize.calibrate_and_quantize( - self.representative_dataset.input_gen, inference_input_type, - inference_output_type, allow_float, activations_type) + self.representative_dataset.input_gen, inference_input_type, + inference_output_type, allow_float, activations_type) def _is_unknown_shapes_allowed(self): # Unknown dimensions are only allowed with the new converter. @@ -1992,7 +1995,7 @@ class TocoConverter(object): @classmethod @_deprecation.deprecated( - None, "Use `lite.TFLiteConverter.from_keras_model_file` instead.") + None, "Use `lite.TFLiteConverter.from_keras_model_file` instead.") def from_keras_model_file(cls, model_file, input_arrays=None, diff --git a/tensorflow/lite/python/lite_test.py b/tensorflow/lite/python/lite_test.py index 044b1211e17..cae49cb147f 100644 --- a/tensorflow/lite/python/lite_test.py +++ b/tensorflow/lite/python/lite_test.py @@ -882,11 +882,15 @@ class FromSessionTest(TestModels, parameterized.TestCase): @parameterized.named_parameters( # Quantize model to Int8: with enable mlir - ('UseTfliteBuiltinsIntEnableMLIR', [lite.OpsSet.TFLITE_BUILTINS_INT8], True), + ('UseTfliteBuiltinsIntEnableMLIR', + [lite.OpsSet.TFLITE_BUILTINS_INT8], True), # Quantize model to Int8: with disable mlir - ('UseTfliteBuiltinsIntDisableMLIR', [lite.OpsSet.TFLITE_BUILTINS_INT8], False), + ('UseTfliteBuiltinsIntDisableMLIR', + [lite.OpsSet.TFLITE_BUILTINS_INT8], False), # Quantize model to Int16: with disable mlir - ('UseTfliteBuiltinsInt16DisableMLIR', [lite.OpsSet.EXPERIMENTAL_TFLITE_BUILTINS_ACTIVATIONS_INT16_WEIGHTS_INT8], False)) + ('UseTfliteBuiltinsInt16DisableMLIR', + [lite.OpsSet.EXPERIMENTAL_TFLITE_BUILTINS_ACTIVATIONS_INT16_WEIGHTS_INT8], + False)) def testCalibrateAndQuantizeBuiltinInt(self, supported_ops, enable_mlir): with ops.Graph().as_default(): inp, output, calibration_gen = self._getCalibrationQuantizeModel() diff --git a/tensorflow/lite/python/optimize/calibrator.py b/tensorflow/lite/python/optimize/calibrator.py index 90c43fcddfa..2b08ec690ff 100644 --- a/tensorflow/lite/python/optimize/calibrator.py +++ b/tensorflow/lite/python/optimize/calibrator.py @@ -78,7 +78,8 @@ class Calibrator(object): computation, useful when targeting an integer-only backend. If False, an error will be thrown if an operation cannot be quantized, otherwise the model will fallback to float ops. - activations_type: A tf.dtype representing the desired type for activations. + activations_type: A tf.dtype representing the desired type for + activations. resize_input: A boolean. True if the shape of the sample data is different from the input. """ diff --git a/tensorflow/lite/python/optimize/calibrator_test.py b/tensorflow/lite/python/optimize/calibrator_test.py index f778c8a555d..d79d76b09ed 100644 --- a/tensorflow/lite/python/optimize/calibrator_test.py +++ b/tensorflow/lite/python/optimize/calibrator_test.py @@ -96,7 +96,8 @@ class CalibratorTest(test_util.TensorFlowTestCase, parameterized.TestCase): ('UseActivationTypeInt8 - EnableMlirQuantizer', constants.INT8), # Activation type Int16 ('UseActivationTypeInt16 - DisableEnableMlirQuantizer', constants.INT16)) - def test_calibration_with_quantization_multiple_inputs(self, activations_type): + def test_calibration_with_quantization_multiple_inputs(self, + activations_type): # Load multi add model from test data. # This model has 4 inputs of size (1, 8, 8, 3). model_path = resource_loader.get_path_to_datafile( From 3fc256a0dc31eae6711d2f7680493925f0fa4091 Mon Sep 17 00:00:00 2001 From: Elena Zhelezina Date: Tue, 9 Jun 2020 17:51:37 +0100 Subject: [PATCH 064/867] Fix for pylint. Change-Id: If2674380c25eb8973e73a407b75660088098e6da --- tensorflow/lite/python/lite_test.py | 9 +++++---- 1 file changed, 5 insertions(+), 4 deletions(-) diff --git a/tensorflow/lite/python/lite_test.py b/tensorflow/lite/python/lite_test.py index cae49cb147f..e6661c82894 100644 --- a/tensorflow/lite/python/lite_test.py +++ b/tensorflow/lite/python/lite_test.py @@ -883,14 +883,15 @@ class FromSessionTest(TestModels, parameterized.TestCase): @parameterized.named_parameters( # Quantize model to Int8: with enable mlir ('UseTfliteBuiltinsIntEnableMLIR', - [lite.OpsSet.TFLITE_BUILTINS_INT8], True), + [lite.OpsSet.TFLITE_BUILTINS_INT8], True), # Quantize model to Int8: with disable mlir ('UseTfliteBuiltinsIntDisableMLIR', - [lite.OpsSet.TFLITE_BUILTINS_INT8], False), + [lite.OpsSet.TFLITE_BUILTINS_INT8], False), # Quantize model to Int16: with disable mlir ('UseTfliteBuiltinsInt16DisableMLIR', - [lite.OpsSet.EXPERIMENTAL_TFLITE_BUILTINS_ACTIVATIONS_INT16_WEIGHTS_INT8], - False)) + [lite.OpsSet.\ + EXPERIMENTAL_TFLITE_BUILTINS_ACTIVATIONS_INT16_WEIGHTS_INT8], + False)) def testCalibrateAndQuantizeBuiltinInt(self, supported_ops, enable_mlir): with ops.Graph().as_default(): inp, output, calibration_gen = self._getCalibrationQuantizeModel() From 15b97dd44a42ae1ab4cc4192a5f7c820ac90b9ac Mon Sep 17 00:00:00 2001 From: nammbash Date: Tue, 9 Jun 2020 11:25:45 -0700 Subject: [PATCH 065/867] CPU Feature Guard MKL brand needs to change to oneDNNchange to one dnn --- tensorflow/core/platform/cpu_feature_guard.cc | 12 ++++++++++-- 1 file changed, 10 insertions(+), 2 deletions(-) diff --git a/tensorflow/core/platform/cpu_feature_guard.cc b/tensorflow/core/platform/cpu_feature_guard.cc index c5a5c287283..a020d3fd70e 100644 --- a/tensorflow/core/platform/cpu_feature_guard.cc +++ b/tensorflow/core/platform/cpu_feature_guard.cc @@ -138,8 +138,16 @@ void InfoAboutUnusedCPUFeatures() { CheckIfFeatureUnused(CPUFeature::FMA, "FMA", missing_instructions); #endif // __FMA__ #endif // else of if defined(_MSC_VER) && !defined(__clang__) - if (!missing_instructions.empty()) { - LOG(INFO) << "This TensorFlow binary is optimized with Intel(R) MKL-DNN " + + string intel_library_official_name( + "Intel(R) oneAPI Deep Neural Network Library (oneDNN) "); +#ifndef INTEL_MKL + intel_library_official_name = "oneAPI Deep Neural Network Library (oneDNN) "; +#endif + + if (!missing_instructions.empty()) { + LOG(INFO) << "This TensorFlow binary is optimized with " + << intel_library_official_name << "to use the following CPU instructions in performance-" << "critical operations: " << missing_instructions << std::endl << "To enable them in other operations, rebuild TensorFlow " From 47674cac85b61f7a438c6970a6a7ca49946a2622 Mon Sep 17 00:00:00 2001 From: Elena Zhelezina Date: Tue, 9 Jun 2020 22:27:57 +0100 Subject: [PATCH 066/867] Fix for pylint Change-Id: If03f60a3eebc7aed61c10870c545fe6035bcb2a3 --- tensorflow/lite/python/lite.py | 12 ++++++------ 1 file changed, 6 insertions(+), 6 deletions(-) diff --git a/tensorflow/lite/python/lite.py b/tensorflow/lite/python/lite.py index bed48860b00..06796ba820b 100644 --- a/tensorflow/lite/python/lite.py +++ b/tensorflow/lite/python/lite.py @@ -361,15 +361,15 @@ class QuantizationMode(object): def _is_int16x8_target_required(self): return bool( - set(self._target_spec.supported_ops).intersection([ - OpsSet.EXPERIMENTAL_TFLITE_BUILTINS_ACTIVATIONS_INT16_WEIGHTS_INT8 - ])) + set(self._target_spec.supported_ops).intersection([ + OpsSet.EXPERIMENTAL_TFLITE_BUILTINS_ACTIVATIONS_INT16_WEIGHTS_INT8 + ])) def _is_allow_float(self): return bool( - set(self._target_spec.supported_ops).intersection([ - OpsSet.TFLITE_BUILTINS - ])) + set(self._target_spec.supported_ops).intersection([ + OpsSet.TFLITE_BUILTINS + ])) def _any_optimization_enabled(self): return bool( From 84afc268a77f543fe64ecb45832701278a9eb129 Mon Sep 17 00:00:00 2001 From: Elena Zhelezina Date: Fri, 5 Jun 2020 11:47:09 +0100 Subject: [PATCH 067/867] Documentation on the new experimental option for 16x8. --- .../performance/post_training_quantization.md | 43 +++++++++++++++++++ 1 file changed, 43 insertions(+) diff --git a/tensorflow/lite/g3doc/performance/post_training_quantization.md b/tensorflow/lite/g3doc/performance/post_training_quantization.md index af7d9dbf02d..c48a2820d2f 100644 --- a/tensorflow/lite/g3doc/performance/post_training_quantization.md +++ b/tensorflow/lite/g3doc/performance/post_training_quantization.md @@ -151,6 +151,49 @@ The disadvantages of float16 quantization are as follows: to float32 when run on the CPU. (Note that the GPU delegate will not perform this dequantization, since it can operate on float16 data.) +### Integer only: 16-bit activations with 8-bit weights (experimental) + +This is an experimental quantization scheme. It is similar to the "integer only" +scheme, but activations are quantized based on their range to 16-bits, weights are +quantized in 8-bit integer and bias is quantized into 64-bit integer. +This is referred to as 16x8 quantization further. + +The main advantage of this quantization is that it can improve accuracy +significantly, but only slightly increase model size. + +
+import tensorflow as tf
+converter = tf.lite.TFLiteConverter.from_saved_model(saved_model_dir)
+converter.optimizations = [tf.lite.Optimize.DEFAULT]
+converter.target_spec.supported_types = [tf.lite.constants.EXPERIMENTAL_TFLITE_BUILTINS_ACTIVATIONS_INT16_WEIGHTS_INT8]
+tflite_quant_model = converter.convert()
+
+ +If 16x8 quantization is not supported for some operators in the model, +then the model still can be quantized, but unsupported operators kept in float. +The following option should be added to the target_spec to allow this. +
+import tensorflow as tf
+converter = tf.lite.TFLiteConverter.from_saved_model(saved_model_dir)
+converter.optimizations = [tf.lite.Optimize.DEFAULT]
+converter.target_spec.supported_types = [tf.lite.constants.EXPERIMENTAL_TFLITE_BUILTINS_ACTIVATIONS_INT16_WEIGHTS_INT8,
+tf.lite.OpsSet.TFLITE_BUILTINS]
+tflite_quant_model = converter.convert()
+
+ +Examples of the use cases where accuracy improvements provided by this quantization scheme include: +* super-resolution, +* audio signal processing such as noise cancelling and beamforming, +* image de-noising, +* HDR reconstruction from a single image. + +The disadvantage of this quantization is: + +* Currently inference is noticeably slower than 8-bit full integer due to the lack of optimized kernel implementation. +* Currently it is incompatible with the existing hardware accelerated TFLite delegates. + +Note: This is an experimental feature. + ### Model accuracy Since weights are quantized post training, there could be an accuracy loss, From dcfc2175c79ee6c610770b597c8d637daa1649bc Mon Sep 17 00:00:00 2001 From: Elena Zhelezina Date: Tue, 9 Jun 2020 22:35:57 +0100 Subject: [PATCH 068/867] Small change of comment per reviewer's note. Change-Id: I1233b95282befebfa0e6c06173f5e928aef60b22 --- tensorflow/lite/python/convert.py | 4 ++-- 1 file changed, 2 insertions(+), 2 deletions(-) diff --git a/tensorflow/lite/python/convert.py b/tensorflow/lite/python/convert.py index 68e23634b2e..ec70f793f21 100644 --- a/tensorflow/lite/python/convert.py +++ b/tensorflow/lite/python/convert.py @@ -98,9 +98,9 @@ class OpsSet(enum.Enum): # int16 activations and int64 bias. # Specifying this will throw an error for operations that do not yet have # quantized implementations. - # This quantization mode should be used in models for super-resolution, + # This quantization mode may be used in models for super-resolution, # audio signal processing or image de-noising. It improves accuracy - # significantly, but only slightly increases the model size. + # significantly, but only slightly increases the model size. # WARNING: These ops are currently experimental and have not yet been finalized. # They are only compatible with CPU execution, and have not been optimized for production. EXPERIMENTAL_TFLITE_BUILTINS_ACTIVATIONS_INT16_WEIGHTS_INT8 = \ From aaf693d29bb5a52d6dfd8c106e45b2ff513e6d84 Mon Sep 17 00:00:00 2001 From: shwetaoj Date: Tue, 2 Jun 2020 08:07:53 -0700 Subject: [PATCH 069/867] Added code to build multiinstance/multinode container --- .../ci_build/linux/mkl/Dockerfile.devel-mkl | 16 ++++-- .../ci_build/linux/mkl/build-dev-container.sh | 26 ++++++++++ .../linux/mkl/install_openmpi_horovod.sh | 49 +++++++++++++++++++ 3 files changed, 88 insertions(+), 3 deletions(-) create mode 100755 tensorflow/tools/ci_build/linux/mkl/install_openmpi_horovod.sh diff --git a/tensorflow/tools/ci_build/linux/mkl/Dockerfile.devel-mkl b/tensorflow/tools/ci_build/linux/mkl/Dockerfile.devel-mkl index 45ccf67d707..3893f61d940 100755 --- a/tensorflow/tools/ci_build/linux/mkl/Dockerfile.devel-mkl +++ b/tensorflow/tools/ci_build/linux/mkl/Dockerfile.devel-mkl @@ -15,6 +15,11 @@ ARG CONFIG_BFLOAT16_BUILD="" ARG ENABLE_SECURE_BUILD ARG BAZEL_VERSION="" ARG ENABLE_DNNL1="" +ARG ENABLE_HOROVOD="" +ARG OPENMPI_VERSION="" +ARG OPENMPI_DOWNLOAD_URL="" + +ENV DEBIAN_FRONTEND=noninteractive # Upgrade Bazel version if argument is passed RUN if [ "${BAZEL_VERSION}" != "" ]; then \ @@ -45,9 +50,6 @@ RUN ${PYTHON} set-build-env.py -p ${TARGET_PLATFORM} -f /root/.mkl.bazelrc \ # Pull the compiler flags we just wrote into root user's .bazelrc file RUN echo "import /root/.mkl.bazelrc" >>/root/.bazelrc -# Install futures>=0.17.1 for Python2.7 compatibility mode -RUN ${PIP} install future>=0.17.1 - RUN bazel --bazelrc=/root/.bazelrc build -c opt \ tensorflow/tools/pip_package:build_pip_package && \ bazel-bin/tensorflow/tools/pip_package/build_pip_package "${WHL_DIR}" && \ @@ -55,6 +57,14 @@ RUN bazel --bazelrc=/root/.bazelrc build -c opt \ rm -rf /root/.cache # Clean up Bazel cache when done. +#Install OpenMPI/Horovod +COPY install_openmpi_horovod.sh . +RUN if [ "${ENABLE_HOROVOD}" = "yes" ]; then \ + chmod +x install_openmpi_horovod.sh && \ + ${OPENMPI_VERSION} ${OPENMPI_DOWNLOAD_URL} install_openmpi_horovod.sh && \ + rm -rf install_openmpi_horovod.sh; \ + fi + # TensorBoard EXPOSE 6006 # IPython diff --git a/tensorflow/tools/ci_build/linux/mkl/build-dev-container.sh b/tensorflow/tools/ci_build/linux/mkl/build-dev-container.sh index eceef65aa38..da647153cdb 100755 --- a/tensorflow/tools/ci_build/linux/mkl/build-dev-container.sh +++ b/tensorflow/tools/ci_build/linux/mkl/build-dev-container.sh @@ -64,6 +64,9 @@ ENABLE_SECURE_BUILD=${ENABLE_SECURE_BUILD:-no} BAZEL_VERSION=${BAZEL_VERSION} BUILD_PY2_CONTAINERS=${BUILD_PY2_CONTAINERS:-no} ENABLE_DNNL1=${ENABLE_DNNL1:-no} +ENABLE_HOROVOD=${ENABLE_HOROVOD:-no} +OPENMPI_VERSION=${OPENMPI_VERSION} +OPENMPI_DOWNLOAD_URL=${OPENMPI_DOWNLOAD_URL} debug "ROOT_CONTAINER=${ROOT_CONTAINER}" debug "TF_ROOT_CONTAINER_TAG=${TF_ROOT_CONTAINER_TAG}" @@ -82,6 +85,9 @@ debug "TMP_DIR=${TMP_DIR}" debug "BAZEL_VERSION=${BAZEL_VERSION}" debug "BUILD_PY2_CONTAINERS=${BUILD_PY2_CONTAINERS}" debug "ENABLE_DNNL1=${ENABLE_DNNL1}" +debug "ENABLE_HOROVOD=${ENABLE_HOROVOD}" +debug "OPENMPI_VERSION=${OPENMPI_VERSION}" +debug "OPENMPI_DOWNLOAD_URL=${OPENMPI_DOWNLOAD_URL}" function build_container() { @@ -131,6 +137,13 @@ function build_container() TF_DOCKER_BUILD_ARGS+=("--build-arg BAZEL_VERSION=${BAZEL_VERSION}") fi + # Add build arg for installing OpenMPI/Horovod + if [[ ${ENABLE_HOROVOD} == "yes" ]]; then + TF_DOCKER_BUILD_ARGS+=("--build-arg ENABLE_HOROVOD=${ENABLE_HOROVOD}") + TF_DOCKER_BUILD_ARGS+=("--build-arg OPENMPI_VERSION=${OPENMPI_VERSION}") + TF_DOCKER_BUILD_ARGS+=("--build-arg OPENMPI_DOWNLOAD_URL=${OPENMPI_DOWNLOAD_URL}") + fi + # Perform docker build debug "Building docker image with image name and tag: ${TEMP_IMAGE_NAME}" CMD="${DOCKER_BINARY} build ${TF_DOCKER_BUILD_ARGS[@]} --no-cache --pull -t ${TEMP_IMAGE_NAME} -f Dockerfile.devel-mkl ." @@ -188,6 +201,19 @@ function test_container() die "FAIL: MKL enabled test in ${TEMP_IMAGE_NAME}" fi + # Test to check if horovod is installed successfully + debug "Test horovod in the container..." + if [[ ${ENABLE_HOROVOD} == "yes" ]]; then + HOROVOD_TEST_CMD=$(${DOCKER_BINARY} exec ${CONTAINER_ID} bash -c "${PYTHON} -c 'import horovod.tensorflow as hvd;'") + ${HOROVOD_TEST_CMD} + if [[ $? == "0" ]]; then + echo "PASS: HOROVOD installation test in ${TEMP_IMAGE_NAME}" + else + die "FAIL: HOROVOD installation test in ${TEMP_IMAGE_NAME}" + fi + fi + + # Stop the running docker container sleep 1 "${DOCKER_BINARY}" stop --time=0 ${CONTAINER_ID} diff --git a/tensorflow/tools/ci_build/linux/mkl/install_openmpi_horovod.sh b/tensorflow/tools/ci_build/linux/mkl/install_openmpi_horovod.sh new file mode 100755 index 00000000000..d1b297726ed --- /dev/null +++ b/tensorflow/tools/ci_build/linux/mkl/install_openmpi_horovod.sh @@ -0,0 +1,49 @@ +#!/usr/bin/env bash +# install OpenMPI, OpenSSH and Horovod + +set -e + +apt-get clean && apt-get update -y + +# Install Open MPI +OPENMPI_VERSION=${OPENMPI_VERSION:-openmpi-2.1.1} +OPENMPI_DOWNLOAD_URL=${OPENMPI_DOWNLOAD_URL:-https://www.open-mpi.org/software/ompi/v2.1/downloads/${OPENMPI_VERSION}.tar.gz} +echo "Installing OpenMPI version ${OPENMPI_VERSION}..." +echo "OpenMPI Download url ${OPENMPI_DOWNLOAD_URL}..." + +mkdir /tmp/openmpi +cd /tmp/openmpi +curl -fSsL -O ${OPENMPI_DOWNLOAD_URL} +tar zxf ${OPENMPI_VERSION}.tar.gz +cd ${OPENMPI_VERSION} +./configure --enable-mpirun-prefix-by-default +make -j $(nproc) all +make install +ldconfig +cd / +rm -rf /tmp/openmpi + +# Create a wrapper for OpenMPI to allow running as root by default +mv /usr/local/bin/mpirun /usr/local/bin/mpirun.real +echo '#!/bin/bash' > /usr/local/bin/mpirun +echo 'mpirun.real --allow-run-as-root "$@"' >> /usr/local/bin/mpirun +chmod a+x /usr/local/bin/mpirun + +# Configure OpenMPI to run good defaults: +echo "btl_tcp_if_exclude = lo,docker0" >> /usr/local/etc/openmpi-mca-params.conf + +#Check mpi version +echo 'OpenMPI version:' +mpirun --version + +# Install OpenSSH for MPI to communicate between containers +apt-get install -y --no-install-recommends --fix-missing openssh-client openssh-server libnuma-dev +mkdir -p /var/run/sshd +# Allow OpenSSH to talk to containers without asking for confirmation +cat /etc/ssh/ssh_config | grep -v StrictHostKeyChecking > /etc/ssh/ssh_config.new +echo " StrictHostKeyChecking no" >> /etc/ssh/ssh_config.new +mv /etc/ssh/ssh_config.new /etc/ssh/ssh_config + +#Install Horovod +HOROVOD_WITH_TENSORFLOW=1 +python3 -m pip install --no-cache-dir horovod==0.19.1 From ff359d4a48aeb1905f767d32e7da1a2d01d4ce6a Mon Sep 17 00:00:00 2001 From: shwetaoj Date: Tue, 2 Jun 2020 13:46:16 -0700 Subject: [PATCH 070/867] Setting default values for OpenMPI versions --- .../tools/ci_build/linux/mkl/Dockerfile.devel-mkl | 2 +- .../ci_build/linux/mkl/install_openmpi_horovod.sh | 11 +++++++++-- 2 files changed, 10 insertions(+), 3 deletions(-) diff --git a/tensorflow/tools/ci_build/linux/mkl/Dockerfile.devel-mkl b/tensorflow/tools/ci_build/linux/mkl/Dockerfile.devel-mkl index 3893f61d940..1fd54ff703f 100755 --- a/tensorflow/tools/ci_build/linux/mkl/Dockerfile.devel-mkl +++ b/tensorflow/tools/ci_build/linux/mkl/Dockerfile.devel-mkl @@ -61,7 +61,7 @@ RUN bazel --bazelrc=/root/.bazelrc build -c opt \ COPY install_openmpi_horovod.sh . RUN if [ "${ENABLE_HOROVOD}" = "yes" ]; then \ chmod +x install_openmpi_horovod.sh && \ - ${OPENMPI_VERSION} ${OPENMPI_DOWNLOAD_URL} install_openmpi_horovod.sh && \ + ./install_openmpi_horovod.sh ${OPENMPI_VERSION} ${OPENMPI_DOWNLOAD_URL} && \ rm -rf install_openmpi_horovod.sh; \ fi diff --git a/tensorflow/tools/ci_build/linux/mkl/install_openmpi_horovod.sh b/tensorflow/tools/ci_build/linux/mkl/install_openmpi_horovod.sh index d1b297726ed..4c8b04f6024 100755 --- a/tensorflow/tools/ci_build/linux/mkl/install_openmpi_horovod.sh +++ b/tensorflow/tools/ci_build/linux/mkl/install_openmpi_horovod.sh @@ -5,9 +5,16 @@ set -e apt-get clean && apt-get update -y +# Set default +if [[ $# -gt 1 ]]; then + OPENMPI_VERSION="${1}" + OPENMPI_DOWNLOAD_URL="${2}" +else + OPENMPI_VERSION=openmpi-2.1.1 + OPENMPI_DOWNLOAD_URL=https://www.open-mpi.org/software/ompi/v2.1/downloads/${OPENMPI_VERSION}.tar.gz +fi + # Install Open MPI -OPENMPI_VERSION=${OPENMPI_VERSION:-openmpi-2.1.1} -OPENMPI_DOWNLOAD_URL=${OPENMPI_DOWNLOAD_URL:-https://www.open-mpi.org/software/ompi/v2.1/downloads/${OPENMPI_VERSION}.tar.gz} echo "Installing OpenMPI version ${OPENMPI_VERSION}..." echo "OpenMPI Download url ${OPENMPI_DOWNLOAD_URL}..." From aa5bfd35fa5292d820493483da8540f8a6386c5f Mon Sep 17 00:00:00 2001 From: shwetaoj Date: Wed, 3 Jun 2020 12:15:24 -0700 Subject: [PATCH 071/867] Added license to the shell script --- .../linux/mkl/install_openmpi_horovod.sh | 17 ++++++++++++++++- 1 file changed, 16 insertions(+), 1 deletion(-) diff --git a/tensorflow/tools/ci_build/linux/mkl/install_openmpi_horovod.sh b/tensorflow/tools/ci_build/linux/mkl/install_openmpi_horovod.sh index 4c8b04f6024..0f5a670f0f2 100755 --- a/tensorflow/tools/ci_build/linux/mkl/install_openmpi_horovod.sh +++ b/tensorflow/tools/ci_build/linux/mkl/install_openmpi_horovod.sh @@ -1,5 +1,20 @@ #!/usr/bin/env bash -# install OpenMPI, OpenSSH and Horovod +# Copyright 2016 The TensorFlow Authors. All Rights Reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. +# ============================================================================== +# Install OpenMPI, OpenSSH and Horovod in Intel(R) MKL support +# Usage: install_openmpi_horovod.sh [openmpi version] [openmpi download url] set -e From 615d3ce1af92614a3285807caaf42f50acd66fae Mon Sep 17 00:00:00 2001 From: shwetaoj Date: Wed, 3 Jun 2020 13:14:39 -0700 Subject: [PATCH 072/867] Added install futures for backward compatibility --- tensorflow/tools/ci_build/linux/mkl/Dockerfile.devel-mkl | 3 +++ 1 file changed, 3 insertions(+) diff --git a/tensorflow/tools/ci_build/linux/mkl/Dockerfile.devel-mkl b/tensorflow/tools/ci_build/linux/mkl/Dockerfile.devel-mkl index 1fd54ff703f..f4ab7ba21c4 100755 --- a/tensorflow/tools/ci_build/linux/mkl/Dockerfile.devel-mkl +++ b/tensorflow/tools/ci_build/linux/mkl/Dockerfile.devel-mkl @@ -50,6 +50,9 @@ RUN ${PYTHON} set-build-env.py -p ${TARGET_PLATFORM} -f /root/.mkl.bazelrc \ # Pull the compiler flags we just wrote into root user's .bazelrc file RUN echo "import /root/.mkl.bazelrc" >>/root/.bazelrc +# Install futures>=0.17.1 for Python2.7 compatibility mode +RUN ${PIP} install future>=0.17.1 + RUN bazel --bazelrc=/root/.bazelrc build -c opt \ tensorflow/tools/pip_package:build_pip_package && \ bazel-bin/tensorflow/tools/pip_package/build_pip_package "${WHL_DIR}" && \ From 7a048082c1c7aa9057c8448b77ee4cde069ec3a7 Mon Sep 17 00:00:00 2001 From: shwetaoj Date: Wed, 3 Jun 2020 13:19:53 -0700 Subject: [PATCH 073/867] Removed extra line --- tensorflow/tools/ci_build/linux/mkl/build-dev-container.sh | 3 +-- 1 file changed, 1 insertion(+), 2 deletions(-) diff --git a/tensorflow/tools/ci_build/linux/mkl/build-dev-container.sh b/tensorflow/tools/ci_build/linux/mkl/build-dev-container.sh index da647153cdb..7278724ff64 100755 --- a/tensorflow/tools/ci_build/linux/mkl/build-dev-container.sh +++ b/tensorflow/tools/ci_build/linux/mkl/build-dev-container.sh @@ -202,8 +202,8 @@ function test_container() fi # Test to check if horovod is installed successfully - debug "Test horovod in the container..." if [[ ${ENABLE_HOROVOD} == "yes" ]]; then + debug "Test horovod in the container..." HOROVOD_TEST_CMD=$(${DOCKER_BINARY} exec ${CONTAINER_ID} bash -c "${PYTHON} -c 'import horovod.tensorflow as hvd;'") ${HOROVOD_TEST_CMD} if [[ $? == "0" ]]; then @@ -213,7 +213,6 @@ function test_container() fi fi - # Stop the running docker container sleep 1 "${DOCKER_BINARY}" stop --time=0 ${CONTAINER_ID} From 50f0ba885bc112b24b437c8a974c6a8deaace96b Mon Sep 17 00:00:00 2001 From: shwetaoj Date: Wed, 3 Jun 2020 20:05:47 -0700 Subject: [PATCH 074/867] Added parameter to pass horovod version and fixed comments --- .../ci_build/linux/mkl/Dockerfile.devel-mkl | 5 ++-- .../ci_build/linux/mkl/build-dev-container.sh | 3 +++ .../linux/mkl/install_openmpi_horovod.sh | 26 ++++++++----------- 3 files changed, 17 insertions(+), 17 deletions(-) diff --git a/tensorflow/tools/ci_build/linux/mkl/Dockerfile.devel-mkl b/tensorflow/tools/ci_build/linux/mkl/Dockerfile.devel-mkl index f4ab7ba21c4..8a5a0a42050 100755 --- a/tensorflow/tools/ci_build/linux/mkl/Dockerfile.devel-mkl +++ b/tensorflow/tools/ci_build/linux/mkl/Dockerfile.devel-mkl @@ -18,6 +18,7 @@ ARG ENABLE_DNNL1="" ARG ENABLE_HOROVOD="" ARG OPENMPI_VERSION="" ARG OPENMPI_DOWNLOAD_URL="" +ARG HOROVOD_VERSION="" ENV DEBIAN_FRONTEND=noninteractive @@ -60,11 +61,11 @@ RUN bazel --bazelrc=/root/.bazelrc build -c opt \ rm -rf /root/.cache # Clean up Bazel cache when done. -#Install OpenMPI/Horovod +# Install OpenMPI/Horovod COPY install_openmpi_horovod.sh . RUN if [ "${ENABLE_HOROVOD}" = "yes" ]; then \ chmod +x install_openmpi_horovod.sh && \ - ./install_openmpi_horovod.sh ${OPENMPI_VERSION} ${OPENMPI_DOWNLOAD_URL} && \ + ./install_openmpi_horovod.sh OPENMPI_VERSION=${OPENMPI_VERSION} OPENMPI_DOWNLOAD_URL=${OPENMPI_DOWNLOAD_URL} HOROVOD_VERSION=${HOROVOD_VERSION} && \ rm -rf install_openmpi_horovod.sh; \ fi diff --git a/tensorflow/tools/ci_build/linux/mkl/build-dev-container.sh b/tensorflow/tools/ci_build/linux/mkl/build-dev-container.sh index 7278724ff64..e9d7f1ff388 100755 --- a/tensorflow/tools/ci_build/linux/mkl/build-dev-container.sh +++ b/tensorflow/tools/ci_build/linux/mkl/build-dev-container.sh @@ -67,6 +67,7 @@ ENABLE_DNNL1=${ENABLE_DNNL1:-no} ENABLE_HOROVOD=${ENABLE_HOROVOD:-no} OPENMPI_VERSION=${OPENMPI_VERSION} OPENMPI_DOWNLOAD_URL=${OPENMPI_DOWNLOAD_URL} +HOROVOD_VERSION=${HOROVOD_VERSION} debug "ROOT_CONTAINER=${ROOT_CONTAINER}" debug "TF_ROOT_CONTAINER_TAG=${TF_ROOT_CONTAINER_TAG}" @@ -88,6 +89,7 @@ debug "ENABLE_DNNL1=${ENABLE_DNNL1}" debug "ENABLE_HOROVOD=${ENABLE_HOROVOD}" debug "OPENMPI_VERSION=${OPENMPI_VERSION}" debug "OPENMPI_DOWNLOAD_URL=${OPENMPI_DOWNLOAD_URL}" +debug "HOROVOD_VERSION=${HOROVOD_VERSION}" function build_container() { @@ -142,6 +144,7 @@ function build_container() TF_DOCKER_BUILD_ARGS+=("--build-arg ENABLE_HOROVOD=${ENABLE_HOROVOD}") TF_DOCKER_BUILD_ARGS+=("--build-arg OPENMPI_VERSION=${OPENMPI_VERSION}") TF_DOCKER_BUILD_ARGS+=("--build-arg OPENMPI_DOWNLOAD_URL=${OPENMPI_DOWNLOAD_URL}") + TF_DOCKER_BUILD_ARGS+=("--build-arg HOROVOD_VERSION=${HOROVOD_VERSION}") fi # Perform docker build diff --git a/tensorflow/tools/ci_build/linux/mkl/install_openmpi_horovod.sh b/tensorflow/tools/ci_build/linux/mkl/install_openmpi_horovod.sh index 0f5a670f0f2..b8d9739ceb6 100755 --- a/tensorflow/tools/ci_build/linux/mkl/install_openmpi_horovod.sh +++ b/tensorflow/tools/ci_build/linux/mkl/install_openmpi_horovod.sh @@ -1,5 +1,5 @@ #!/usr/bin/env bash -# Copyright 2016 The TensorFlow Authors. All Rights Reserved. +# Copyright 2020 The TensorFlow Authors. All Rights Reserved. # # Licensed under the Apache License, Version 2.0 (the "License"); # you may not use this file except in compliance with the License. @@ -13,25 +13,21 @@ # See the License for the specific language governing permissions and # limitations under the License. # ============================================================================== -# Install OpenMPI, OpenSSH and Horovod in Intel(R) MKL support -# Usage: install_openmpi_horovod.sh [openmpi version] [openmpi download url] +# Install OpenMPI, OpenSSH and Horovod during Intel(R) MKL container build +# Usage: install_openmpi_horovod.sh [OPENMPI_VERSION=] [OPENMPI_DOWNLOAD_URL=] [HOROVOD_VERSION=] set -e apt-get clean && apt-get update -y # Set default -if [[ $# -gt 1 ]]; then - OPENMPI_VERSION="${1}" - OPENMPI_DOWNLOAD_URL="${2}" -else - OPENMPI_VERSION=openmpi-2.1.1 - OPENMPI_DOWNLOAD_URL=https://www.open-mpi.org/software/ompi/v2.1/downloads/${OPENMPI_VERSION}.tar.gz -fi +OPENMPI_VERSION=${OPENMPI_VERSION:-openmpi-2.1.1} +OPENMPI_DOWNLOAD_URL=${OPENMPI_DOWNLOAD_URL:-https://www.open-mpi.org/software/ompi/v2.1/downloads/${OPENMPI_VERSION}.tar.gz} +HOROVOD_VERSION=${HOROVOD_VERSION:-0.19.1} # Install Open MPI -echo "Installing OpenMPI version ${OPENMPI_VERSION}..." -echo "OpenMPI Download url ${OPENMPI_DOWNLOAD_URL}..." +echo "Installing OpenMPI version ${OPENMPI_VERSION} ..." +echo "OpenMPI Download url ${OPENMPI_DOWNLOAD_URL} ..." mkdir /tmp/openmpi cd /tmp/openmpi @@ -54,7 +50,7 @@ chmod a+x /usr/local/bin/mpirun # Configure OpenMPI to run good defaults: echo "btl_tcp_if_exclude = lo,docker0" >> /usr/local/etc/openmpi-mca-params.conf -#Check mpi version +# Check mpi version echo 'OpenMPI version:' mpirun --version @@ -66,6 +62,6 @@ cat /etc/ssh/ssh_config | grep -v StrictHostKeyChecking > /etc/ssh/ssh_config.ne echo " StrictHostKeyChecking no" >> /etc/ssh/ssh_config.new mv /etc/ssh/ssh_config.new /etc/ssh/ssh_config -#Install Horovod +# Install Horovod HOROVOD_WITH_TENSORFLOW=1 -python3 -m pip install --no-cache-dir horovod==0.19.1 +python3 -m pip install --no-cache-dir horovod==${HOROVOD_VERSION} From 1ed2ab4638c56b3cce6d0f85efeaea3600b75214 Mon Sep 17 00:00:00 2001 From: justkw Date: Thu, 4 Jun 2020 09:22:07 -0700 Subject: [PATCH 075/867] Adding parameter to use --nightly_flag to install specific packages if building the nightly build --- tensorflow/tools/ci_build/linux/mkl/Dockerfile.devel-mkl | 5 +++-- tensorflow/tools/ci_build/linux/mkl/build-dev-container.sh | 7 +++++++ 2 files changed, 10 insertions(+), 2 deletions(-) diff --git a/tensorflow/tools/ci_build/linux/mkl/Dockerfile.devel-mkl b/tensorflow/tools/ci_build/linux/mkl/Dockerfile.devel-mkl index 8a5a0a42050..a78d13c7755 100755 --- a/tensorflow/tools/ci_build/linux/mkl/Dockerfile.devel-mkl +++ b/tensorflow/tools/ci_build/linux/mkl/Dockerfile.devel-mkl @@ -19,6 +19,7 @@ ARG ENABLE_HOROVOD="" ARG OPENMPI_VERSION="" ARG OPENMPI_DOWNLOAD_URL="" ARG HOROVOD_VERSION="" +ARG TF_NIGHTLY_FLAG="" ENV DEBIAN_FRONTEND=noninteractive @@ -56,8 +57,8 @@ RUN ${PIP} install future>=0.17.1 RUN bazel --bazelrc=/root/.bazelrc build -c opt \ tensorflow/tools/pip_package:build_pip_package && \ - bazel-bin/tensorflow/tools/pip_package/build_pip_package "${WHL_DIR}" && \ - ${PIP} --no-cache-dir install --upgrade "${WHL_DIR}"/tensorflow-*.whl && \ + bazel-bin/tensorflow/tools/pip_package/build_pip_package "${TF_NIGHTLY_FLAG}" "${WHL_DIR}" && \ + ${PIP} --no-cache-dir install --upgrade "${WHL_DIR}"/*.whl && \ rm -rf /root/.cache # Clean up Bazel cache when done. diff --git a/tensorflow/tools/ci_build/linux/mkl/build-dev-container.sh b/tensorflow/tools/ci_build/linux/mkl/build-dev-container.sh index e9d7f1ff388..83b3ebaf9c9 100755 --- a/tensorflow/tools/ci_build/linux/mkl/build-dev-container.sh +++ b/tensorflow/tools/ci_build/linux/mkl/build-dev-container.sh @@ -68,6 +68,7 @@ ENABLE_HOROVOD=${ENABLE_HOROVOD:-no} OPENMPI_VERSION=${OPENMPI_VERSION} OPENMPI_DOWNLOAD_URL=${OPENMPI_DOWNLOAD_URL} HOROVOD_VERSION=${HOROVOD_VERSION} +IS_NIGHTLY=${IS_NIGHTLY:-no} debug "ROOT_CONTAINER=${ROOT_CONTAINER}" debug "TF_ROOT_CONTAINER_TAG=${TF_ROOT_CONTAINER_TAG}" @@ -90,6 +91,7 @@ debug "ENABLE_HOROVOD=${ENABLE_HOROVOD}" debug "OPENMPI_VERSION=${OPENMPI_VERSION}" debug "OPENMPI_DOWNLOAD_URL=${OPENMPI_DOWNLOAD_URL}" debug "HOROVOD_VERSION=${HOROVOD_VERSION}" +debug "IS_NIGHTLY=${IS_NIGHTLY}" function build_container() { @@ -147,6 +149,11 @@ function build_container() TF_DOCKER_BUILD_ARGS+=("--build-arg HOROVOD_VERSION=${HOROVOD_VERSION}") fi + # Add build arg --nightly_flag for the nightly build + if [[ ${IS_NIGHTLY} == "yes" ]]; then + TF_DOCKER_BUILD_ARGS+=("--build-arg TF_NIGHTLY_FLAG=--nightly_flag") + fi + # Perform docker build debug "Building docker image with image name and tag: ${TEMP_IMAGE_NAME}" CMD="${DOCKER_BINARY} build ${TF_DOCKER_BUILD_ARGS[@]} --no-cache --pull -t ${TEMP_IMAGE_NAME} -f Dockerfile.devel-mkl ." From 83decf0d382b76771e2b1ad4fa43d208d5e40eb0 Mon Sep 17 00:00:00 2001 From: shwetaoj Date: Mon, 8 Jun 2020 07:50:14 -0700 Subject: [PATCH 076/867] Support multiple OS --- .../ci_build/linux/mkl/install_openmpi_horovod.sh | 13 ++++++++++++- 1 file changed, 12 insertions(+), 1 deletion(-) diff --git a/tensorflow/tools/ci_build/linux/mkl/install_openmpi_horovod.sh b/tensorflow/tools/ci_build/linux/mkl/install_openmpi_horovod.sh index b8d9739ceb6..6044927d2ce 100755 --- a/tensorflow/tools/ci_build/linux/mkl/install_openmpi_horovod.sh +++ b/tensorflow/tools/ci_build/linux/mkl/install_openmpi_horovod.sh @@ -55,7 +55,18 @@ echo 'OpenMPI version:' mpirun --version # Install OpenSSH for MPI to communicate between containers -apt-get install -y --no-install-recommends --fix-missing openssh-client openssh-server libnuma-dev +( apt-get update && apt-get install -y --no-install-recommends --fix-missing \ + libnuma-dev \ + openssh-server \ + openssh-client \ + apt-get clean && \ + rm -rf /var/lib/apt/lists/* ) || \ + ( yum -y update && yum -y install \ + numactl-devel \ + openssh-server \ + openssh-clients \ + yum clean all ) || \ + ( echo "Unsupported Linux distribution. Aborting!" && exit 1 ) mkdir -p /var/run/sshd # Allow OpenSSH to talk to containers without asking for confirmation cat /etc/ssh/ssh_config | grep -v StrictHostKeyChecking > /etc/ssh/ssh_config.new From 371c2e6f4f3f233041eda2d292a13824d98d769f Mon Sep 17 00:00:00 2001 From: shwetaoj Date: Mon, 8 Jun 2020 08:52:49 -0700 Subject: [PATCH 077/867] Bug fix --- .../tools/ci_build/linux/mkl/install_openmpi_horovod.sh | 4 ++-- 1 file changed, 2 insertions(+), 2 deletions(-) diff --git a/tensorflow/tools/ci_build/linux/mkl/install_openmpi_horovod.sh b/tensorflow/tools/ci_build/linux/mkl/install_openmpi_horovod.sh index 6044927d2ce..276d9945ab6 100755 --- a/tensorflow/tools/ci_build/linux/mkl/install_openmpi_horovod.sh +++ b/tensorflow/tools/ci_build/linux/mkl/install_openmpi_horovod.sh @@ -58,13 +58,13 @@ mpirun --version ( apt-get update && apt-get install -y --no-install-recommends --fix-missing \ libnuma-dev \ openssh-server \ - openssh-client \ + openssh-client && \ apt-get clean && \ rm -rf /var/lib/apt/lists/* ) || \ ( yum -y update && yum -y install \ numactl-devel \ openssh-server \ - openssh-clients \ + openssh-clients && \ yum clean all ) || \ ( echo "Unsupported Linux distribution. Aborting!" && exit 1 ) mkdir -p /var/run/sshd From da18384ad585b2d88a08119268c9a7134ee36bf5 Mon Sep 17 00:00:00 2001 From: shwetaoj Date: Mon, 8 Jun 2020 09:51:25 -0700 Subject: [PATCH 078/867] Bug fix --- tensorflow/tools/ci_build/linux/mkl/install_openmpi_horovod.sh | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/tensorflow/tools/ci_build/linux/mkl/install_openmpi_horovod.sh b/tensorflow/tools/ci_build/linux/mkl/install_openmpi_horovod.sh index 276d9945ab6..b765dbd70a6 100755 --- a/tensorflow/tools/ci_build/linux/mkl/install_openmpi_horovod.sh +++ b/tensorflow/tools/ci_build/linux/mkl/install_openmpi_horovod.sh @@ -58,7 +58,7 @@ mpirun --version ( apt-get update && apt-get install -y --no-install-recommends --fix-missing \ libnuma-dev \ openssh-server \ - openssh-client && \ + openssh-clients && \ apt-get clean && \ rm -rf /var/lib/apt/lists/* ) || \ ( yum -y update && yum -y install \ From ab86bb82faabec7b1d29c61df1cae0b45d0b0e8e Mon Sep 17 00:00:00 2001 From: shwetaoj Date: Tue, 9 Jun 2020 07:53:05 -0700 Subject: [PATCH 079/867] Bug fix --- .../linux/mkl/install_openmpi_horovod.sh | 33 ++++++++++--------- 1 file changed, 17 insertions(+), 16 deletions(-) diff --git a/tensorflow/tools/ci_build/linux/mkl/install_openmpi_horovod.sh b/tensorflow/tools/ci_build/linux/mkl/install_openmpi_horovod.sh index b765dbd70a6..aec40543a17 100755 --- a/tensorflow/tools/ci_build/linux/mkl/install_openmpi_horovod.sh +++ b/tensorflow/tools/ci_build/linux/mkl/install_openmpi_horovod.sh @@ -14,15 +14,14 @@ # limitations under the License. # ============================================================================== # Install OpenMPI, OpenSSH and Horovod during Intel(R) MKL container build -# Usage: install_openmpi_horovod.sh [OPENMPI_VERSION=] [OPENMPI_DOWNLOAD_URL=] [HOROVOD_VERSION=] +# Usage: install_openmpi_horovod.sh [OPENMPI_VERSION=] [OPENMPI_DOWNLOAD_URL=] +# [HOROVOD_VERSION=] set -e -apt-get clean && apt-get update -y - # Set default OPENMPI_VERSION=${OPENMPI_VERSION:-openmpi-2.1.1} -OPENMPI_DOWNLOAD_URL=${OPENMPI_DOWNLOAD_URL:-https://www.open-mpi.org/software/ompi/v2.1/downloads/${OPENMPI_VERSION}.tar.gz} +OPENMPI_DOWNLOAD_URL=${OPENMPI_DOWNLOAD_URL:-https://www.open-mpi.org/software/ompi/v2.1/downloads/openmpi-2.1.1.tar.gz} HOROVOD_VERSION=${HOROVOD_VERSION:-0.19.1} # Install Open MPI @@ -55,18 +54,20 @@ echo 'OpenMPI version:' mpirun --version # Install OpenSSH for MPI to communicate between containers -( apt-get update && apt-get install -y --no-install-recommends --fix-missing \ - libnuma-dev \ - openssh-server \ - openssh-clients && \ - apt-get clean && \ - rm -rf /var/lib/apt/lists/* ) || \ - ( yum -y update && yum -y install \ - numactl-devel \ - openssh-server \ - openssh-clients && \ - yum clean all ) || \ - ( echo "Unsupported Linux distribution. Aborting!" && exit 1 ) +apt-get clean && apt-get update && apt-get install -y --no-install-recommends --fix-missing \ + openssh-client openssh-server libnuma-dev && \ + rm -rf /var/lib/apt/lists/* +if [[ $? == "0" ]]; then + echo "PASS: OpenSSH installation" +else + yum -y update && yum -y install numactl-devel openssh-server openssh-clients && \ + yum clean all + if [[ $? == "0" ]]; then + echo "PASS: OpenSSH installation" + else + echo "Unsupported Linux distribution. Aborting!" && exit 1 + fi +fi mkdir -p /var/run/sshd # Allow OpenSSH to talk to containers without asking for confirmation cat /etc/ssh/ssh_config | grep -v StrictHostKeyChecking > /etc/ssh/ssh_config.new From ae408bb512e614469e24ccf0db6c031f6aeac030 Mon Sep 17 00:00:00 2001 From: shwetaoj Date: Tue, 9 Jun 2020 18:00:47 -0700 Subject: [PATCH 080/867] remvoe trailing white space --- tensorflow/tools/ci_build/linux/mkl/build-dev-container.sh | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/tensorflow/tools/ci_build/linux/mkl/build-dev-container.sh b/tensorflow/tools/ci_build/linux/mkl/build-dev-container.sh index 83b3ebaf9c9..6e789a54e87 100755 --- a/tensorflow/tools/ci_build/linux/mkl/build-dev-container.sh +++ b/tensorflow/tools/ci_build/linux/mkl/build-dev-container.sh @@ -214,7 +214,7 @@ function test_container() # Test to check if horovod is installed successfully if [[ ${ENABLE_HOROVOD} == "yes" ]]; then debug "Test horovod in the container..." - HOROVOD_TEST_CMD=$(${DOCKER_BINARY} exec ${CONTAINER_ID} bash -c "${PYTHON} -c 'import horovod.tensorflow as hvd;'") + HOROVOD_TEST_CMD=$(${DOCKER_BINARY} exec ${CONTAINER_ID} bash -c "${PYTHON} -c 'import horovod.tensorflow as hvd;'") ${HOROVOD_TEST_CMD} if [[ $? == "0" ]]; then echo "PASS: HOROVOD installation test in ${TEMP_IMAGE_NAME}" From 02c6b9edaf14fb7ca35c2b4f0bc80b74cc8a551a Mon Sep 17 00:00:00 2001 From: jerryyin Date: Wed, 10 Jun 2020 21:11:38 +0000 Subject: [PATCH 081/867] [ROCm][mlir] Disable mlir saved model test --- .../mlir/tensorflow/tests/tf_saved_model/build_defs.bzl | 4 +--- 1 file changed, 1 insertion(+), 3 deletions(-) diff --git a/tensorflow/compiler/mlir/tensorflow/tests/tf_saved_model/build_defs.bzl b/tensorflow/compiler/mlir/tensorflow/tests/tf_saved_model/build_defs.bzl index 594afa10453..95ad05aa1e6 100644 --- a/tensorflow/compiler/mlir/tensorflow/tests/tf_saved_model/build_defs.bzl +++ b/tensorflow/compiler/mlir/tensorflow/tests/tf_saved_model/build_defs.bzl @@ -4,8 +4,6 @@ load("//tensorflow/compiler/mlir:glob_lit_test.bzl", "lit_test") def tf_saved_model_test(name, data, tags = None): """Create a SavedModel test.""" - if tags == None: - tags = ["no_rocm"] native.py_binary( name = name, testonly = 1, @@ -26,5 +24,5 @@ def tf_saved_model_test(name, data, tags = None): name = name + ".py", data = [name] + data, driver = "@llvm-project//mlir:run_lit.sh", - tags = tags, + tags = tags + ["no_rocm"], ) From 3b95c2c54df8a7bc3641871197262841b803f8cd Mon Sep 17 00:00:00 2001 From: "902449@58880@bigcat_chen@ASIC" Date: Thu, 11 Jun 2020 13:54:16 +0800 Subject: [PATCH 082/867] Correcting for PR comments --- .../himax_we1_evb/detection_responder.cc | 1 - .../himax_we1_evb/image_provider.cc | 10 +- .../himax_we1_evb/main_functions.cc | 126 ------------------ .../lite/micro/himax_we1_evb/debug_log.cc | 5 +- 4 files changed, 6 insertions(+), 136 deletions(-) delete mode 100644 tensorflow/lite/micro/examples/person_detection_experimental/himax_we1_evb/main_functions.cc diff --git a/tensorflow/lite/micro/examples/person_detection_experimental/himax_we1_evb/detection_responder.cc b/tensorflow/lite/micro/examples/person_detection_experimental/himax_we1_evb/detection_responder.cc index a353dc8a9b8..ae5de962fd3 100644 --- a/tensorflow/lite/micro/examples/person_detection_experimental/himax_we1_evb/detection_responder.cc +++ b/tensorflow/lite/micro/examples/person_detection_experimental/himax_we1_evb/detection_responder.cc @@ -22,7 +22,6 @@ limitations under the License. // should implement their own versions of this function. void RespondToDetection(tflite::ErrorReporter* error_reporter, int8_t person_score, int8_t no_person_score) { - if (person_score > no_person_score) { hx_drv_led_on(HX_DRV_LED_GREEN); } else { diff --git a/tensorflow/lite/micro/examples/person_detection_experimental/himax_we1_evb/image_provider.cc b/tensorflow/lite/micro/examples/person_detection_experimental/himax_we1_evb/image_provider.cc index 727d93c61d1..4a3ab5775be 100644 --- a/tensorflow/lite/micro/examples/person_detection_experimental/himax_we1_evb/image_provider.cc +++ b/tensorflow/lite/micro/examples/person_detection_experimental/himax_we1_evb/image_provider.cc @@ -21,14 +21,12 @@ limitations under the License. hx_drv_sensor_image_config_t g_pimg_config; - TfLiteStatus GetImage(tflite::ErrorReporter* error_reporter, int image_width, int image_height, int channels, int8_t* image_data) { static bool is_initialized = false; if (!is_initialized) { - if(hx_drv_sensor_initial(&g_pimg_config)!= HX_DRV_LIB_PASS) - { + if (hx_drv_sensor_initial(&g_pimg_config) != HX_DRV_LIB_PASS) { return kTfLiteError; } is_initialized = true; @@ -36,9 +34,9 @@ TfLiteStatus GetImage(tflite::ErrorReporter* error_reporter, int image_width, hx_drv_sensor_capture(&g_pimg_config); - hx_drv_image_rescale((uint8_t*)g_pimg_config.raw_address, g_pimg_config.img_width, g_pimg_config.img_height, - image_data, image_width, image_height); - + hx_drv_image_rescale((uint8_t*)g_pimg_config.raw_address, + g_pimg_config.img_width, g_pimg_config.img_height, + image_data, image_width, image_height); return kTfLiteOk; } diff --git a/tensorflow/lite/micro/examples/person_detection_experimental/himax_we1_evb/main_functions.cc b/tensorflow/lite/micro/examples/person_detection_experimental/himax_we1_evb/main_functions.cc deleted file mode 100644 index f0c7a405974..00000000000 --- a/tensorflow/lite/micro/examples/person_detection_experimental/himax_we1_evb/main_functions.cc +++ /dev/null @@ -1,126 +0,0 @@ -/* Copyright 2019 The TensorFlow Authors. All Rights Reserved. - -Licensed under the Apache License, Version 2.0 (the "License"); -you may not use this file except in compliance with the License. -You may obtain a copy of the License at - - http://www.apache.org/licenses/LICENSE-2.0 - -Unless required by applicable law or agreed to in writing, software -distributed under the License is distributed on an "AS IS" BASIS, -WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -See the License for the specific language governing permissions and -limitations under the License. -==============================================================================*/ - -#include "tensorflow/lite/micro/examples/person_detection_experimental/main_functions.h" - -#include "tensorflow/lite/micro/examples/person_detection_experimental/detection_responder.h" -#include "tensorflow/lite/micro/examples/person_detection_experimental/image_provider.h" -#include "tensorflow/lite/micro/examples/person_detection_experimental/model_settings.h" -#include "tensorflow/lite/micro/examples/person_detection_experimental/person_detect_model_data.h" -#include "tensorflow/lite/micro/kernels/micro_ops.h" -#include "tensorflow/lite/micro/micro_error_reporter.h" -#include "tensorflow/lite/micro/micro_interpreter.h" -#include "tensorflow/lite/micro/micro_mutable_op_resolver.h" -#include "tensorflow/lite/schema/schema_generated.h" -#include "tensorflow/lite/version.h" - -// Globals, used for compatibility with Arduino-style sketches. -namespace { -tflite::ErrorReporter* error_reporter = nullptr; -const tflite::Model* model = nullptr; -tflite::MicroInterpreter* interpreter = nullptr; -TfLiteTensor* input = nullptr; - -// In order to use optimized tensorflow lite kernels, a signed int8 quantized -// model is preferred over the legacy unsigned model format. This means that -// throughout this project, input images must be converted from unisgned to -// signed format. The easiest and quickest way to convert from unsigned to -// signed 8-bit integers is to subtract 128 from the unsigned value to get a -// signed value. - -// An area of memory to use for input, output, and intermediate arrays. -constexpr int kTensorArenaSize = 125 * 1024; -#pragma Bss(".tensor_arena") -static uint8_t tensor_arena[kTensorArenaSize]; -#pragma Bss() -} // namespace - -// The name of this function is important for Arduino compatibility. -void setup() { - // Set up logging. Google style is to avoid globals or statics because of - // lifetime uncertainty, but since this has a trivial destructor it's okay. - // NOLINTNEXTLINE(runtime-global-variables) - static tflite::MicroErrorReporter micro_error_reporter; - error_reporter = µ_error_reporter; - - // Map the model into a usable data structure. This doesn't involve any - // copying or parsing, it's a very lightweight operation. - model = tflite::GetModel(g_person_detect_model_data); - if (model->version() != TFLITE_SCHEMA_VERSION) { - TF_LITE_REPORT_ERROR(error_reporter, - "Model provided is schema version %d not equal " - "to supported version %d.", - model->version(), TFLITE_SCHEMA_VERSION); - return; - } - - // Pull in only the operation implementations we need. - // This relies on a complete list of all the ops needed by this graph. - // An easier approach is to just use the AllOpsResolver, but this will - // incur some penalty in code space for op implementations that are not - // needed by this graph. - // - // tflite::AllOpsResolver resolver; - // NOLINTNEXTLINE(runtime-global-variables) - static tflite::MicroMutableOpResolver<5> micro_op_resolver; - micro_op_resolver.AddBuiltin( - tflite::BuiltinOperator_DEPTHWISE_CONV_2D, - tflite::ops::micro::Register_DEPTHWISE_CONV_2D()); - micro_op_resolver.AddBuiltin(tflite::BuiltinOperator_CONV_2D, - tflite::ops::micro::Register_CONV_2D()); - micro_op_resolver.AddBuiltin(tflite::BuiltinOperator_AVERAGE_POOL_2D, - tflite::ops::micro::Register_AVERAGE_POOL_2D()); - micro_op_resolver.AddBuiltin(tflite::BuiltinOperator_RESHAPE, - tflite::ops::micro::Register_RESHAPE()); - micro_op_resolver.AddBuiltin(tflite::BuiltinOperator_SOFTMAX, - tflite::ops::micro::Register_SOFTMAX()); - - // Build an interpreter to run the model with. - // NOLINTNEXTLINE(runtime-global-variables) - static tflite::MicroInterpreter static_interpreter( - model, micro_op_resolver, tensor_arena, kTensorArenaSize, error_reporter); - interpreter = &static_interpreter; - - // Allocate memory from the tensor_arena for the model's tensors. - TfLiteStatus allocate_status = interpreter->AllocateTensors(); - if (allocate_status != kTfLiteOk) { - TF_LITE_REPORT_ERROR(error_reporter, "AllocateTensors() failed"); - return; - } - - // Get information about the memory area to use for the model's input. - input = interpreter->input(0); -} - -// The name of this function is important for Arduino compatibility. -void loop() { - // Get image from provider. - if (kTfLiteOk != GetImage(error_reporter, kNumCols, kNumRows, kNumChannels, - input->data.int8)) { - TF_LITE_REPORT_ERROR(error_reporter, "Image capture failed."); - } - - // Run the model on this input and make sure it succeeds. - if (kTfLiteOk != interpreter->Invoke()) { - TF_LITE_REPORT_ERROR(error_reporter, "Invoke failed."); - } - - TfLiteTensor* output = interpreter->output(0); - - // Process the inference results. - int8_t person_score = output->data.uint8[kPersonIndex]; - int8_t no_person_score = output->data.uint8[kNotAPersonIndex]; - RespondToDetection(error_reporter, person_score, no_person_score); -} diff --git a/tensorflow/lite/micro/himax_we1_evb/debug_log.cc b/tensorflow/lite/micro/himax_we1_evb/debug_log.cc index 32af2625630..36ac3f3fa03 100644 --- a/tensorflow/lite/micro/himax_we1_evb/debug_log.cc +++ b/tensorflow/lite/micro/himax_we1_evb/debug_log.cc @@ -20,12 +20,11 @@ limitations under the License. #include "tensorflow/lite/micro/debug_log.h" #include "hx_drv_tflm.h" - extern "C" void DebugLog(const char* s) { static bool is_initialized = false; if (!is_initialized) { - hx_drv_uart_initial(); - is_initialized = true; + hx_drv_uart_initial(); + is_initialized = true; } hx_drv_uart_print("%s", s); From 013ddd96d0ce111ca5ec1422b2899b66ec41a036 Mon Sep 17 00:00:00 2001 From: Nathan Luehr Date: Tue, 9 Jun 2020 15:35:47 -0500 Subject: [PATCH 083/867] Use CUDNN_TENSOR_OP_MATH to enable tensor cores. --- tensorflow/stream_executor/cuda/cuda_dnn.cc | 5 ++--- 1 file changed, 2 insertions(+), 3 deletions(-) diff --git a/tensorflow/stream_executor/cuda/cuda_dnn.cc b/tensorflow/stream_executor/cuda/cuda_dnn.cc index fa06d410323..28ec6a842bb 100644 --- a/tensorflow/stream_executor/cuda/cuda_dnn.cc +++ b/tensorflow/stream_executor/cuda/cuda_dnn.cc @@ -723,7 +723,7 @@ class CudnnConvolutionDescriptor { #if CUDNN_VERSION >= 7000 cudnnMathType_t math_type = #if CUDNN_VERSION >= 8000 - (use_tensor_op_math ? CUDNN_DEFAULT_MATH : CUDNN_FMA_MATH); + (use_tensor_op_math ? CUDNN_TENSOR_OP_MATH : CUDNN_FMA_MATH); #else (use_tensor_op_math ? CUDNN_TENSOR_OP_MATH : CUDNN_DEFAULT_MATH); #endif @@ -1179,8 +1179,7 @@ class CudnnRnnDescriptor : public dnn::RnnDescriptor { cudnnMathType_t math_type; if (use_tensor_ops) { - math_type = - CUDNN_VERSION >= 8000 ? CUDNN_DEFAULT_MATH : CUDNN_TENSOR_OP_MATH; + math_type = CUDNN_TENSOR_OP_MATH; } else { math_type = CUDNN_VERSION >= 8000 ? CUDNN_FMA_MATH : CUDNN_DEFAULT_MATH; } From da0d85808e77aa62287ba22e822a9e83866a43a4 Mon Sep 17 00:00:00 2001 From: Nathan Luehr Date: Thu, 11 Jun 2020 15:35:04 -0500 Subject: [PATCH 084/867] Make python names consistent --- tensorflow/python/framework/config.py | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/tensorflow/python/framework/config.py b/tensorflow/python/framework/config.py index a356e6d9a16..1ff2fa613da 100644 --- a/tensorflow/python/framework/config.py +++ b/tensorflow/python/framework/config.py @@ -34,7 +34,7 @@ def tensor_float32_execution_allowed(): return _pywrap_tf32_execution.is_allowed() # No tf_export until TF is built against CUDA11 which is required for TF32. -def allow_tensor_float_32_execution(allowed): +def allow_tensor_float32_execution(allowed): """Allow use of TensorFloat-32 with float32 ops on supported hardware. TensorFloat-32 is a math mode introduced with the NVIDIA Ampere architecture. From d03b86fe4462ebad1f73d460c2aceab47372b239 Mon Sep 17 00:00:00 2001 From: Elena Zhelezina Date: Thu, 11 Jun 2020 23:05:40 +0100 Subject: [PATCH 085/867] Extended test to the case when new converter is enabled. Change-Id: I83f20d025027ad1266f99f9d79932cab4f1a9ed5 --- tensorflow/lite/python/lite_test.py | 6 +++++- 1 file changed, 5 insertions(+), 1 deletion(-) diff --git a/tensorflow/lite/python/lite_test.py b/tensorflow/lite/python/lite_test.py index e6661c82894..478840c5549 100644 --- a/tensorflow/lite/python/lite_test.py +++ b/tensorflow/lite/python/lite_test.py @@ -891,7 +891,11 @@ class FromSessionTest(TestModels, parameterized.TestCase): ('UseTfliteBuiltinsInt16DisableMLIR', [lite.OpsSet.\ EXPERIMENTAL_TFLITE_BUILTINS_ACTIVATIONS_INT16_WEIGHTS_INT8], - False)) + False), + ('UseTfliteBuiltinsInt16EnableMLIR', + [lite.OpsSet.\ + EXPERIMENTAL_TFLITE_BUILTINS_ACTIVATIONS_INT16_WEIGHTS_INT8], + True)) def testCalibrateAndQuantizeBuiltinInt(self, supported_ops, enable_mlir): with ops.Graph().as_default(): inp, output, calibration_gen = self._getCalibrationQuantizeModel() From 4bd42cdd0e68cebe8b280b323cde4d01a9d2bf3a Mon Sep 17 00:00:00 2001 From: Reed Date: Thu, 11 Jun 2020 17:16:28 -0700 Subject: [PATCH 086/867] Use float_32 instead of float32 in function names --- tensorflow/python/framework/config.py | 4 ++-- 1 file changed, 2 insertions(+), 2 deletions(-) diff --git a/tensorflow/python/framework/config.py b/tensorflow/python/framework/config.py index 1ff2fa613da..bbaa2ca8248 100644 --- a/tensorflow/python/framework/config.py +++ b/tensorflow/python/framework/config.py @@ -25,7 +25,7 @@ from tensorflow.python.util.tf_export import tf_export # No tf_export until TF is built against CUDA11 which is required for TF32. -def tensor_float32_execution_allowed(): +def tensor_float_32_execution_allowed(): """Get if TensorFloat-32 operations are enabled on supported hardware. Returns: @@ -34,7 +34,7 @@ def tensor_float32_execution_allowed(): return _pywrap_tf32_execution.is_allowed() # No tf_export until TF is built against CUDA11 which is required for TF32. -def allow_tensor_float32_execution(allowed): +def allow_tensor_float_32_execution(allowed): """Allow use of TensorFloat-32 with float32 ops on supported hardware. TensorFloat-32 is a math mode introduced with the NVIDIA Ampere architecture. From 77e46ebf9745d798863b5c7ed26d6bf077700008 Mon Sep 17 00:00:00 2001 From: Kaixi Hou Date: Mon, 30 Mar 2020 17:25:38 -0700 Subject: [PATCH 087/867] conv3d ndhwc plumbing --- tensorflow/core/kernels/conv_grad_ops_3d.cc | 148 ++++++++++++++++---- tensorflow/core/kernels/conv_ops_3d.cc | 88 +++++++++--- 2 files changed, 193 insertions(+), 43 deletions(-) diff --git a/tensorflow/core/kernels/conv_grad_ops_3d.cc b/tensorflow/core/kernels/conv_grad_ops_3d.cc index 2183d0e0885..fc2d58ec94f 100644 --- a/tensorflow/core/kernels/conv_grad_ops_3d.cc +++ b/tensorflow/core/kernels/conv_grad_ops_3d.cc @@ -50,6 +50,7 @@ using stream_executor::dnn::DimIndex; #include "tensorflow/stream_executor/gpu/gpu_asm_opts.h" #include "tensorflow/stream_executor/gpu/redzone_allocator.h" #include "tensorflow/stream_executor/tf_allocator_adapter.h" +#include "third_party/gpus/cudnn/cudnn.h" #endif // GOOGLE_CUDA namespace { @@ -1264,26 +1265,56 @@ class Conv3DBackpropInputOp : public OpKernel { CHECK(padding_rows >= 0 && padding_cols >= 0 && padding_planes >= 0) << "Negative paddings: (" << padding_rows << ", " << padding_cols << ", " << padding_planes << ")"; + +#if GOOGLE_CUDA + const bool compute_in_nhwc = CUDNN_VERSION >= 8000 && + DataTypeToEnum::value == DT_HALF; +#else + // fast NDHWC implementation is a CUDA only feature + const bool compute_in_nhwc = false; +#endif + const TensorFormat compute_data_format = + (compute_in_nhwc && data_format_ == FORMAT_NHWC) ? FORMAT_NHWC + : FORMAT_NCHW; + + VLOG(3) << "Compute Conv3DBackpropInput with cuDNN:" + << " data_format=" << ToString(data_format_) + << " compute_data_format=" << ToString(compute_data_format); + + constexpr auto kComputeInNHWC = + std::make_tuple(se::dnn::DataLayout::kBatchYXDepth, + se::dnn::FilterLayout::kOutputYXInput); + constexpr auto kComputeInNCHW = + std::make_tuple(se::dnn::DataLayout::kBatchDepthYX, + se::dnn::FilterLayout::kOutputInputYX); + + se::dnn::DataLayout compute_data_layout; + se::dnn::FilterLayout filter_layout; + + std::tie(compute_data_layout, filter_layout) = + compute_data_format == FORMAT_NHWC ? kComputeInNHWC : kComputeInNCHW; + se::dnn::BatchDescriptor input_desc(3); input_desc.set_count(dims.batch_size) .set_spatial_dim(DimIndex::X, compatible_input_shape.dim_size(4)) .set_spatial_dim(DimIndex::Y, compatible_input_shape.dim_size(3)) .set_spatial_dim(DimIndex::Z, compatible_input_shape.dim_size(2)) .set_feature_map_count(dims.in_depth) - .set_layout(se::dnn::DataLayout::kBatchDepthYX); + .set_layout(compute_data_layout); se::dnn::BatchDescriptor output_desc(3); output_desc.set_count(dims.batch_size) .set_spatial_dim(DimIndex::X, dims.output_size(2)) .set_spatial_dim(DimIndex::Y, dims.output_size(1)) .set_spatial_dim(DimIndex::Z, dims.output_size(0)) .set_feature_map_count(dims.out_depth) - .set_layout(se::dnn::DataLayout::kBatchDepthYX); + .set_layout(compute_data_layout); se::dnn::FilterDescriptor filter_desc(3); filter_desc.set_spatial_dim(DimIndex::X, dims.filter_size(2)) .set_spatial_dim(DimIndex::Y, dims.filter_size(1)) .set_spatial_dim(DimIndex::Z, dims.filter_size(0)) .set_input_feature_map_count(filter_shape.dim_size(3)) - .set_output_feature_map_count(filter_shape.dim_size(4)); + .set_output_feature_map_count(filter_shape.dim_size(4)) + .set_layout(filter_layout); se::dnn::ConvolutionDescriptor conv_desc(3); conv_desc.set_dilation_rate(DimIndex::X, dims.dilation(2)) .set_dilation_rate(DimIndex::Y, dims.dilation(1)) @@ -1298,21 +1329,33 @@ class Conv3DBackpropInputOp : public OpKernel { // Shape: out, in, z, y, x. Tensor transformed_filter; + auto dst_format = + compute_data_format == FORMAT_NCHW ? FORMAT_OIHW: FORMAT_OHWI; + TensorShape dst_shape = + dst_format == FORMAT_OIHW + ? TensorShape({filter_shape.dim_size(4), filter_shape.dim_size(3), + dims.filter_size(0), + dims.filter_size(1), + dims.filter_size(2)}) + : TensorShape({filter_shape.dim_size(4), + dims.filter_size(0), + dims.filter_size(1), + dims.filter_size(2), + filter_shape.dim_size(3)}); OP_REQUIRES_OK( context, context->allocate_temp( DataTypeToEnum::value, - TensorShape({filter_shape.dim_size(4), - filter_shape.dim_size(3), dims.filter_size(0), - dims.filter_size(1), dims.filter_size(2)}), + dst_shape, &transformed_filter)); + functor::TransformFilter()( - context->eigen_device(), FORMAT_OIHW, + context->eigen_device(), dst_format, To32Bit(filter.tensor()), To32Bit(transformed_filter.tensor())); // Shape: batch, filters, z, y, x. Tensor transformed_out_backprop; - if (data_format_ == FORMAT_NHWC) { + if (data_format_ == FORMAT_NHWC && compute_data_format == FORMAT_NCHW) { TensorShape nchw_shape = {dims.batch_size, dims.out_depth, dims.output_size(0), dims.output_size(1), dims.output_size(2)}; @@ -1333,8 +1376,15 @@ class Conv3DBackpropInputOp : public OpKernel { Tensor pre_transformed_in_backprop; OP_REQUIRES_OK( context, - context->allocate_temp(DataTypeToEnum::value, compatible_input_shape, - &pre_transformed_in_backprop)); + context->allocate_temp( + DataTypeToEnum::value, + ShapeFromFormat( + compute_data_format, compatible_input_shape.dim_size(0), + {{compatible_input_shape.dim_size(2), + compatible_input_shape.dim_size(3), + compatible_input_shape.dim_size(4)}}, + compatible_input_shape.dim_size(1)), + &pre_transformed_in_backprop)); auto out_backprop_ptr = AsDeviceMemory(transformed_out_backprop.template flat().data(), @@ -1355,7 +1405,7 @@ class Conv3DBackpropInputOp : public OpKernel { dims.batch_size, dims.in_depth, {{dims.input_size(0), dims.input_size(1), dims.input_size(2)}}, - FORMAT_NCHW, + compute_data_format, dims.out_depth, {{dims.filter_size(0), dims.filter_size(1), dims.filter_size(2)}}, {{dims.dilation(0), dims.dilation(1), dims.dilation(2)}}, @@ -1500,8 +1550,11 @@ class Conv3DBackpropInputOp : public OpKernel { OP_REQUIRES_OK(context, context->allocate_temp( DataTypeToEnum::value, - {dims.batch_size, dims.in_depth, dims.input_size(0), - dims.input_size(1), dims.input_size(2)}, + ShapeFromFormat( + compute_data_format, dims.batch_size, + {{dims.input_size(0), dims.input_size(1), + dims.input_size(2)}}, + dims.in_depth), &in_backprop_remove_padding)); // Remove the padding for odd spatial dimensions. @@ -1510,12 +1563,13 @@ class Conv3DBackpropInputOp : public OpKernel { To32Bit(const_cast(pre_transformed_in_backprop) .tensor()), {{0, 0, 0}}, {{-planes_odd, -rows_odd, -cols_odd}}, - To32Bit(in_backprop_remove_padding.tensor()), FORMAT_NCHW); + To32Bit(in_backprop_remove_padding.tensor()), + compute_data_format); pre_transformed_in_backprop = in_backprop_remove_padding; } - if (data_format_ == FORMAT_NHWC) { + if (data_format_ == FORMAT_NHWC && compute_data_format == FORMAT_NCHW) { auto toConstTensor = [](const Tensor& x) -> const Tensor { return x; }; functor::NCHWToNHWC()( context->eigen_device(), @@ -1723,6 +1777,35 @@ class Conv3DBackpropFilterOp : public OpKernel { CHECK(padding_rows >= 0 && padding_cols >= 0 && padding_planes >= 0) << "Negative paddings: (" << padding_rows << ", " << padding_cols << ", " << padding_planes << ")"; + +#if GOOGLE_CUDA + const bool compute_in_nhwc = CUDNN_VERSION >= 8000 && + DataTypeToEnum::value == DT_HALF; +#else + // fast NDHWC implementation is a CUDA only feature + const bool compute_in_nhwc = false; +#endif + const TensorFormat compute_data_format = + (compute_in_nhwc && data_format_ == FORMAT_NHWC) ? FORMAT_NHWC + : FORMAT_NCHW; + + VLOG(3) << "Compute Conv3DBackpropFilter with cuDNN:" + << " data_format=" << ToString(data_format_) + << " compute_data_format=" << ToString(compute_data_format); + + constexpr auto kComputeInNHWC = + std::make_tuple(se::dnn::DataLayout::kBatchYXDepth, + se::dnn::FilterLayout::kOutputYXInput); + constexpr auto kComputeInNCHW = + std::make_tuple(se::dnn::DataLayout::kBatchDepthYX, + se::dnn::FilterLayout::kOutputInputYX); + + se::dnn::DataLayout compute_data_layout; + se::dnn::FilterLayout filter_layout; + + std::tie(compute_data_layout, filter_layout) = + compute_data_format == FORMAT_NHWC ? kComputeInNHWC : kComputeInNCHW; + se::dnn::BatchDescriptor input_desc(3); input_desc.set_count(dims.batch_size) .set_spatial_dim(DimIndex::X, @@ -1732,20 +1815,21 @@ class Conv3DBackpropFilterOp : public OpKernel { .set_spatial_dim(DimIndex::Z, GetTensorDim(compatible_input, data_format_, '0')) .set_feature_map_count(dims.in_depth) - .set_layout(se::dnn::DataLayout::kBatchDepthYX); + .set_layout(compute_data_layout); se::dnn::BatchDescriptor output_desc(3); output_desc.set_count(dims.batch_size) .set_spatial_dim(DimIndex::X, dims.output_size(2)) .set_spatial_dim(DimIndex::Y, dims.output_size(1)) .set_spatial_dim(DimIndex::Z, dims.output_size(0)) .set_feature_map_count(dims.out_depth) - .set_layout(se::dnn::DataLayout::kBatchDepthYX); + .set_layout(compute_data_layout); se::dnn::FilterDescriptor filter_desc(3); filter_desc.set_spatial_dim(DimIndex::X, dims.filter_size(2)) .set_spatial_dim(DimIndex::Y, dims.filter_size(1)) .set_spatial_dim(DimIndex::Z, dims.filter_size(0)) .set_input_feature_map_count(filter_shape.dim_size(3)) - .set_output_feature_map_count(filter_shape.dim_size(4)); + .set_output_feature_map_count(filter_shape.dim_size(4)) + .set_layout(filter_layout); se::dnn::ConvolutionDescriptor conv_desc(3); conv_desc.set_dilation_rate(DimIndex::X, dims.dilation(2)) .set_dilation_rate(DimIndex::Y, dims.dilation(1)) @@ -1757,17 +1841,30 @@ class Conv3DBackpropFilterOp : public OpKernel { .set_zero_padding(DimIndex::Y, padding_rows / 2) .set_zero_padding(DimIndex::Z, padding_planes / 2) .set_group_count(dims.in_depth / filter_shape.dim_size(3)); + Tensor pre_transformed_filter_backprop; + auto dst_format = + compute_data_format == FORMAT_NCHW ? FORMAT_OIHW: FORMAT_OHWI; + TensorShape dst_shape = + dst_format == FORMAT_OIHW + ? TensorShape({filter_shape.dim_size(4), filter_shape.dim_size(3), + dims.filter_size(0), + dims.filter_size(1), + dims.filter_size(2)}) + : TensorShape({filter_shape.dim_size(4), + dims.filter_size(0), + dims.filter_size(1), + dims.filter_size(2), + filter_shape.dim_size(3)}); OP_REQUIRES_OK( context, context->allocate_temp( DataTypeToEnum::value, - TensorShape({filter_shape.dim_size(4), - filter_shape.dim_size(3), dims.filter_size(0), - dims.filter_size(1), dims.filter_size(2)}), + dst_shape, &pre_transformed_filter_backprop)); Tensor transformed_out_backprop; - if (data_format_ == FORMAT_NHWC) { + if (data_format_ == FORMAT_NHWC && compute_data_format == FORMAT_NCHW) { + VLOG(4) << "Convert the `out_backprop` tensor from NDHWC to NCDHW."; TensorShape nchw_shape = {dims.batch_size, dims.out_depth, dims.output_size(0), dims.output_size(1), dims.output_size(2)}; @@ -1785,7 +1882,8 @@ class Conv3DBackpropFilterOp : public OpKernel { transformed_out_backprop = out_backprop; } Tensor transformed_input; - if (data_format_ == FORMAT_NHWC) { + if (data_format_ == FORMAT_NHWC && compute_data_format == FORMAT_NCHW) { + VLOG(4) << "Convert the `input` tensor from NDHWC to NCDHW."; TensorShape nchw_shape = { dims.batch_size, dims.in_depth, compatible_input.dim_size(1), compatible_input.dim_size(2), compatible_input.dim_size(3)}; @@ -1823,7 +1921,7 @@ class Conv3DBackpropFilterOp : public OpKernel { dims.batch_size, dims.in_depth, {{dims.input_size(0), dims.input_size(1), dims.input_size(2)}}, - FORMAT_NCHW, + compute_data_format, dims.out_depth, {{dims.filter_size(0), dims.filter_size(1), dims.filter_size(2)}}, {{dims.dilation(0), dims.dilation(1), dims.dilation(2)}}, @@ -1947,7 +2045,7 @@ class Conv3DBackpropFilterOp : public OpKernel { auto toConstTensor = [](const Tensor& x) -> const Tensor { return x; }; functor::ReverseTransformFilter()( - context->eigen_device(), /*src_filter_format=*/FORMAT_OIHW, + context->eigen_device(), /*src_filter_format=*/dst_format, toConstTensor(pre_transformed_filter_backprop).template tensor(), filter_backprop->tensor()); } diff --git a/tensorflow/core/kernels/conv_ops_3d.cc b/tensorflow/core/kernels/conv_ops_3d.cc index 69e6fba4192..c1fe6c690cd 100644 --- a/tensorflow/core/kernels/conv_ops_3d.cc +++ b/tensorflow/core/kernels/conv_ops_3d.cc @@ -43,6 +43,7 @@ using stream_executor::dnn::DimIndex; #include "tensorflow/stream_executor/gpu/asm_compiler.h" #include "tensorflow/stream_executor/gpu/redzone_allocator.h" #include "tensorflow/stream_executor/tf_allocator_adapter.h" +#include "third_party/gpus/cudnn/cudnn.h" #endif // GOOGLE_CUDA namespace tensorflow { @@ -201,7 +202,23 @@ struct LaunchConvOp { } } - if (data_format == FORMAT_NHWC) { +#if GOOGLE_CUDA + const bool compute_in_nhwc = CUDNN_VERSION >= 8000 && + DataTypeToEnum::value == DT_HALF; +#else + // fast NHWC implementation is a CUDA only feature + const bool compute_in_nhwc = false; +#endif + const TensorFormat compute_data_format = + (compute_in_nhwc && data_format == FORMAT_NHWC) ? FORMAT_NHWC + : FORMAT_NCHW; + + VLOG(3) << "Compute Conv3D with cuDNN:" + << " data_format=" << ToString(data_format) + << " compute_data_format=" << ToString(compute_data_format); + + if (data_format == FORMAT_NHWC && compute_data_format == FORMAT_NCHW) { + VLOG(4) << "Convert the input tensor from NDHWC to NCDHW."; const TensorShape nchw_shape = ShapeFromFormat( FORMAT_NCHW, in_batch, {{in_planes, in_rows, in_cols}}, in_depth); if (in_depth > 1) { @@ -219,8 +236,26 @@ struct LaunchConvOp { } else { CHECK(input.CopyFrom(input, nchw_shape)); } + } else { + CHECK(data_format == compute_data_format) // Crash OK + << "Illegal data and compute format pair:" + << " data_format=" << ToString(data_format) + << " compute_data_format=" << ToString(compute_data_format); } + constexpr auto kComputeInNHWC = + std::make_tuple(se::dnn::DataLayout::kBatchYXDepth, + se::dnn::FilterLayout::kOutputYXInput); + constexpr auto kComputeInNCHW = + std::make_tuple(se::dnn::DataLayout::kBatchDepthYX, + se::dnn::FilterLayout::kOutputInputYX); + + se::dnn::DataLayout compute_data_layout; + se::dnn::FilterLayout filter_layout; + + std::tie(compute_data_layout, filter_layout) = + compute_data_format == FORMAT_NHWC ? kComputeInNHWC : kComputeInNCHW; + CHECK(pad_rows >= 0 && pad_cols >= 0 && pad_planes >= 0) << "Negative paddings: (" << pad_rows << ", " << pad_cols << ", " << pad_planes << ")"; @@ -230,20 +265,21 @@ struct LaunchConvOp { .set_spatial_dim(DimIndex::X, in_cols) .set_spatial_dim(DimIndex::Y, in_rows) .set_spatial_dim(DimIndex::Z, in_planes) - .set_layout(se::dnn::DataLayout::kBatchDepthYX); + .set_layout(compute_data_layout); se::dnn::BatchDescriptor output_desc(3); output_desc.set_count(in_batch) .set_spatial_dim(DimIndex::X, out_cols) .set_spatial_dim(DimIndex::Y, out_rows) .set_spatial_dim(DimIndex::Z, out_planes) .set_feature_map_count(out_depth) - .set_layout(se::dnn::DataLayout::kBatchDepthYX); + .set_layout(compute_data_layout); se::dnn::FilterDescriptor filter_desc(3); filter_desc.set_spatial_dim(DimIndex::X, filter_cols) .set_spatial_dim(DimIndex::Y, filter_rows) .set_spatial_dim(DimIndex::Z, filter_planes) .set_input_feature_map_count(filter_depth) - .set_output_feature_map_count(out_depth); + .set_output_feature_map_count(out_depth) + .set_layout(filter_layout); se::dnn::ConvolutionDescriptor conv_desc(3); conv_desc.set_dilation_rate(DimIndex::X, dilations[2]) .set_dilation_rate(DimIndex::Y, dilations[1]) @@ -257,25 +293,42 @@ struct LaunchConvOp { .set_group_count(in_depth / filter_depth); Tensor transformed_filter; + auto dst_format = + compute_data_format == FORMAT_NCHW ? FORMAT_OIHW: FORMAT_OHWI; + VLOG(4) << "Transform filter tensor from " << ToString(FORMAT_HWIO) + << " to " << ToString(dst_format); + TensorShape dst_shape = + dst_format == FORMAT_OIHW + ? TensorShape({filter.dim_size(4), filter.dim_size(3), + filter.dim_size(0), filter.dim_size(1), + filter.dim_size(2)}) + : TensorShape({filter.dim_size(4), filter.dim_size(0), + filter.dim_size(1), filter.dim_size(2), + filter.dim_size(3)}); OP_REQUIRES_OK( ctx, ctx->allocate_temp(DataTypeToEnum::value, - TensorShape({out_depth, in_depth, filter_planes, - filter_rows, filter_cols}), + dst_shape, &transformed_filter)); // filter: [x, y, z, in, out] - // t_filter: [out, in, x, y, z] + // t_filter: [out, in, x, y, z] (NCDHW) or + // t_filter: [out, x, y, z, in] (NDHWC) functor::TransformFilter()( - ctx->eigen_device(), FORMAT_OIHW, + ctx->eigen_device(), dst_format, To32Bit(filter.tensor()), To32Bit(transformed_filter.tensor())); Tensor transformed_output; - OP_REQUIRES_OK( - ctx, ctx->allocate_temp( - DataTypeToEnum::value, - ShapeFromFormat(FORMAT_NCHW, in_batch, - {{out_planes, out_rows, out_cols}}, out_depth), - &transformed_output)); + if (data_format != compute_data_format) { + VLOG(4) << "Allocate temporary memory for output in compute data format"; + OP_REQUIRES_OK( + ctx, ctx->allocate_temp( + DataTypeToEnum::value, + ShapeFromFormat(FORMAT_NCHW, in_batch, + {{out_planes, out_rows, out_cols}}, out_depth), + &transformed_output)); + } else { + transformed_output = *output; + } auto input_ptr = AsDeviceMemory(input.template flat().data(), input.template flat().size()); @@ -295,7 +348,7 @@ struct LaunchConvOp { in_batch, in_depth, {{in_planes, in_rows, in_cols}}, - FORMAT_NCHW, + compute_data_format, out_depth, {{filter_planes, filter_rows, filter_cols}}, {{dilations[0], dilations[1], dilations[2]}}, @@ -455,15 +508,14 @@ struct LaunchConvOp { ") filter shape(", filter.shape().DebugString(), ")")); } - if (data_format == FORMAT_NHWC) { + if (data_format == FORMAT_NHWC && compute_data_format == FORMAT_NCHW) { + VLOG(4) << "Convert the output tensor back from NCDHW to NDHWC."; // t_output: [b, out, x, y, z] // output: [b, x, y, z, out] functor::NCHWToNHWC()( ctx->eigen_device(), const_cast(transformed_output).tensor(), output->tensor()); - } else { - *output = transformed_output; } } }; From 6bb481c58b27b3fc99c0d8fc71b9d58f13e8b0ba Mon Sep 17 00:00:00 2001 From: Amedeo Cavallo Date: Mon, 15 Jun 2020 11:56:50 +0200 Subject: [PATCH 088/867] C linkage for stm32l4HAL target C linkage for output retargeting on stm32l4HAL target --- tensorflow/lite/micro/stm32f4HAL/debug_log.cc | 10 +++++++++- 1 file changed, 9 insertions(+), 1 deletion(-) diff --git a/tensorflow/lite/micro/stm32f4HAL/debug_log.cc b/tensorflow/lite/micro/stm32f4HAL/debug_log.cc index 4be3b40e782..90dd7cfd787 100644 --- a/tensorflow/lite/micro/stm32f4HAL/debug_log.cc +++ b/tensorflow/lite/micro/stm32f4HAL/debug_log.cc @@ -22,6 +22,10 @@ limitations under the License. extern UART_HandleTypeDef DEBUG_UART_HANDLE; +#ifdef __cplusplus + extern "C" { +#endif + #ifdef __GNUC__ int __io_putchar(int ch) { HAL_UART_Transmit(&DEBUG_UART_HANDLE, (uint8_t *)&ch, 1, HAL_MAX_DELAY); @@ -36,4 +40,8 @@ int fputc(int ch, FILE *f) { } #endif /* __GNUC__ */ -extern "C" void DebugLog(const char *s) { fprintf(stderr, "%s", s); } +void DebugLog(const char *s) { fprintf(stderr, "%s", s); } + +#ifdef __cplusplus +} +#endif From 01b30fa03f636396f977a76628123199b772463a Mon Sep 17 00:00:00 2001 From: Elena Zhelezina Date: Mon, 15 Jun 2020 17:30:29 +0100 Subject: [PATCH 089/867] Fix for the linter. Change-Id: Ie1185cc2cca9157655b22e1d3bb49ddc017a8f0e --- tensorflow/lite/tools/optimize/quantization_utils.cc | 1 - 1 file changed, 1 deletion(-) diff --git a/tensorflow/lite/tools/optimize/quantization_utils.cc b/tensorflow/lite/tools/optimize/quantization_utils.cc index cdf2743585e..cdc794c20c4 100644 --- a/tensorflow/lite/tools/optimize/quantization_utils.cc +++ b/tensorflow/lite/tools/optimize/quantization_utils.cc @@ -92,7 +92,6 @@ void GetSymmetricQuantizationParams( min = std::min(min, 0.0f); max = std::max(max, 0.0f); const float scale = std::max(std::abs(max), std::abs(min)) / half_quant_range; - int64_t zero_point = 0; quantization_params->min = std::vector(1, min); quantization_params->max = std::vector(1, max); quantization_params->scale = std::vector(1, scale); From 2fe65568b1475f23afc2ee9387e8f447d5eda4d8 Mon Sep 17 00:00:00 2001 From: sshiddib Date: Mon, 15 Jun 2020 13:14:03 -0700 Subject: [PATCH 090/867] [Intel MKL] Adding DNNL ops (part 1) supporting threadpool work --- .../core/kernels/mkl_conv_grad_filter_ops.cc | 17 ++++++- .../core/kernels/mkl_conv_grad_input_ops.cc | 11 ++++- tensorflow/core/kernels/mkl_conv_ops.cc | 14 ++++++ .../core/kernels/mkl_fused_batch_norm_op.cc | 46 ++++++++++++++++++- .../core/kernels/mkl_pooling_ops_common.cc | 25 +++++++++- tensorflow/core/kernels/mkl_quantize_op.cc | 5 ++ tensorflow/core/kernels/mkl_relu_op.cc | 18 +++++++- tensorflow/core/kernels/mkl_slice_op.cc | 8 +++- tensorflow/core/kernels/mkl_softmax_op.cc | 8 +++- 9 files changed, 143 insertions(+), 9 deletions(-) diff --git a/tensorflow/core/kernels/mkl_conv_grad_filter_ops.cc b/tensorflow/core/kernels/mkl_conv_grad_filter_ops.cc index 4c3cea4b6ff..12581d0bfa5 100644 --- a/tensorflow/core/kernels/mkl_conv_grad_filter_ops.cc +++ b/tensorflow/core/kernels/mkl_conv_grad_filter_ops.cc @@ -114,6 +114,21 @@ class MklConvBwdFilterPrimitive : public MklPrimitive { void Execute(const T* src_data, const T* diff_filter_data, const T* diff_bias_data, const T* diff_dst_data, std::shared_ptr bwd_filter_stream) { + // TODO: Create a common function and avoid the duplicate code +#ifdef ENABLE_MKLDNN_THREADPOOL + context_.src_mem->set_data_handle( + static_cast(const_cast(src_data)), *bwd_filter_stream); + context_.diff_filter_mem->set_data_handle( + static_cast(const_cast(diff_filter_data)), + *bwd_filter_stream); + if (diff_bias_data != nullptr) { + context_.diff_bias_mem->set_data_handle( + static_cast(const_cast(diff_bias_data)), + *bwd_filter_stream); + } + context_.diff_dst_mem->set_data_handle( + static_cast(const_cast(diff_dst_data)), *bwd_filter_stream); +#else context_.src_mem->set_data_handle( static_cast(const_cast(src_data))); context_.diff_filter_mem->set_data_handle( @@ -124,7 +139,7 @@ class MklConvBwdFilterPrimitive : public MklPrimitive { } context_.diff_dst_mem->set_data_handle( static_cast(const_cast(diff_dst_data))); - +#endif // ENABLE_MKLDNN_THREADPOOL #ifdef ENABLE_MKLDNN_V1 execute_primitives(context_.bwd_filter_primitives, bwd_filter_stream, context_.bwd_filter_primitives_args); diff --git a/tensorflow/core/kernels/mkl_conv_grad_input_ops.cc b/tensorflow/core/kernels/mkl_conv_grad_input_ops.cc index f9c8d11c67c..7177431029a 100644 --- a/tensorflow/core/kernels/mkl_conv_grad_input_ops.cc +++ b/tensorflow/core/kernels/mkl_conv_grad_input_ops.cc @@ -116,13 +116,22 @@ class MklConvBwdInputPrimitive : public MklPrimitive { void Execute(const T* diff_src_data, const T* filter_data, const T* diff_dst_data, std::shared_ptr bwd_input_stream) { + // TODO: Create a common function and avoid the duplicate code +#ifdef ENABLE_MKLDNN_THREADPOOL + context_.diff_src_mem->set_data_handle( + static_cast(const_cast(diff_src_data)), *bwd_input_stream); + context_.filter_mem->set_data_handle( + static_cast(const_cast(filter_data)), *bwd_input_stream); + context_.diff_dst_mem->set_data_handle( + static_cast(const_cast(diff_dst_data)), *bwd_input_stream); +#else context_.diff_src_mem->set_data_handle( static_cast(const_cast(diff_src_data))); context_.filter_mem->set_data_handle( static_cast(const_cast(filter_data))); context_.diff_dst_mem->set_data_handle( static_cast(const_cast(diff_dst_data))); - +#endif // ENABLE_MKLDNN_THREADPOOL #ifdef ENABLE_MKLDNN_V1 execute_primitives(context_.bwd_input_primitives, bwd_input_stream, context_.bwd_input_primitives_args); diff --git a/tensorflow/core/kernels/mkl_conv_ops.cc b/tensorflow/core/kernels/mkl_conv_ops.cc index 7d0510d03ac..210044436aa 100644 --- a/tensorflow/core/kernels/mkl_conv_ops.cc +++ b/tensorflow/core/kernels/mkl_conv_ops.cc @@ -110,6 +110,19 @@ class MklConvFwdPrimitive : public MklPrimitive { void Execute(const Tinput* src_data, const Tfilter* filter_data, const Tbias* bias_data, const Toutput* dst_data, std::shared_ptr fwd_stream) { + // TODO: Create a common function and avoid the duplicate code +#ifdef ENABLE_MKLDNN_THREADPOOL + context_.src_mem->set_data_handle( + static_cast(const_cast(src_data)), *fwd_stream); + context_.filter_mem->set_data_handle( + static_cast(const_cast(filter_data)), *fwd_stream); + if (bias_data != nullptr) { + context_.bias_mem->set_data_handle( + static_cast(const_cast(bias_data)), *fwd_stream); + } + context_.dst_mem->set_data_handle( + static_cast(const_cast(dst_data)), *fwd_stream); +#else context_.src_mem->set_data_handle( static_cast(const_cast(src_data))); context_.filter_mem->set_data_handle( @@ -120,6 +133,7 @@ class MklConvFwdPrimitive : public MklPrimitive { } context_.dst_mem->set_data_handle( static_cast(const_cast(dst_data))); +#endif // ENABLE_MKLDNN_THREADPOOL #ifdef ENABLE_MKLDNN_V1 DCHECK_EQ(context_.fwd_primitives.size(), context_.fwd_primitives_args.size()); diff --git a/tensorflow/core/kernels/mkl_fused_batch_norm_op.cc b/tensorflow/core/kernels/mkl_fused_batch_norm_op.cc index 954ae0492df..3b2c4f84039 100644 --- a/tensorflow/core/kernels/mkl_fused_batch_norm_op.cc +++ b/tensorflow/core/kernels/mkl_fused_batch_norm_op.cc @@ -94,6 +94,28 @@ class MklFusedBatchNormFwdPrimitive : public MklPrimitive { void Execute(const T* src_data, const U* weights_data, T* dst_data, U* mean_data, U* variance_data, std::shared_ptr fwd_stream, U* workspace_data) { + // TODO: Create a common function and avoid the duplicate code +#ifdef ENABLE_MKLDNN_THREADPOOL + context_.src_mem->set_data_handle( + static_cast(const_cast(src_data)), *fwd_stream); + context_.dst_mem->set_data_handle(static_cast(dst_data), + *fwd_stream); + + if (IS_SET(use_scale_shift)) + context_.weights_mem->set_data_handle( + static_cast(const_cast(weights_data)), *fwd_stream); + + if ((context_.pkind == prop_kind::forward_training) || + (IS_SET(use_global_stats))) { + context_.mean_mem->set_data_handle(static_cast(mean_data), + *fwd_stream); + context_.variance_mem->set_data_handle(static_cast(variance_data), + *fwd_stream); + } + if (workspace_data != nullptr) { + context_.ws_mem->set_data_handle(workspace_data, *fwd_stream); + } +#else context_.src_mem->set_data_handle( static_cast(const_cast(src_data))); context_.dst_mem->set_data_handle(static_cast(dst_data)); @@ -110,6 +132,7 @@ class MklFusedBatchNormFwdPrimitive : public MklPrimitive { if (workspace_data != nullptr) { context_.ws_mem->set_data_handle(workspace_data); } +#endif // ENABLE_MKLDNN_THREADPOOL #ifdef ENABLE_MKLDNN_V1 // Execute batch-normalization forward primitives. execute_primitives(context_.fwd_primitives, fwd_stream, context_.net_args); @@ -503,6 +526,27 @@ class MklFusedBatchNormBwdPrimitive : public MklPrimitive { const T* diff_dst_data, const U* weights_data, T* diff_src_data, U* diff_weights_data, U* res_space_data, std::shared_ptr bwd_stream) { + // TODO: Create a common function and avoid the duplicate code +#ifdef ENABLE_MKLDNN_THREADPOOL + context_.src_mem->set_data_handle( + static_cast(const_cast(src_data)), *bwd_stream); + context_.mean_mem->set_data_handle( + static_cast(const_cast(mean_data)), *bwd_stream); + context_.variance_mem->set_data_handle( + static_cast(const_cast(variance_data)), *bwd_stream); + context_.diff_dst_mem->set_data_handle( + static_cast(const_cast(diff_dst_data)), *bwd_stream); + + if (IS_SET(use_scale_shift)) { + context_.weights_mem->set_data_handle( + static_cast(const_cast(weights_data)), *bwd_stream); + context_.diff_weights_mem->set_data_handle( + static_cast(diff_weights_data), *bwd_stream); + } + + context_.diff_src_mem->set_data_handle(static_cast(diff_src_data), + *bwd_stream); +#else context_.src_mem->set_data_handle( static_cast(const_cast(src_data))); context_.mean_mem->set_data_handle( @@ -520,7 +564,7 @@ class MklFusedBatchNormBwdPrimitive : public MklPrimitive { } context_.diff_src_mem->set_data_handle(static_cast(diff_src_data)); - +#endif // ENABLE_MKLDNN_THREADPOOL #ifdef ENABLE_MKLDNN_V1 // Execute backward batch-normalization primitives. DCHECK_EQ(context_.bwd_primitives.size(), context_.net_args.size()); diff --git a/tensorflow/core/kernels/mkl_pooling_ops_common.cc b/tensorflow/core/kernels/mkl_pooling_ops_common.cc index 2dfc6db0075..5f1c9129ec3 100644 --- a/tensorflow/core/kernels/mkl_pooling_ops_common.cc +++ b/tensorflow/core/kernels/mkl_pooling_ops_common.cc @@ -127,6 +127,17 @@ template void MklPoolingFwdPrimitive::Execute(const T* src_data, T* dst_data, void* ws_data, std::shared_ptr fwd_stream) { +#ifdef ENABLE_MKLDNN_THREADPOOL + context_.src_mem->set_data_handle( + static_cast(const_cast(src_data)), *fwd_stream); + context_.dst_mem->set_data_handle(static_cast(dst_data), *fwd_stream); + if (context_.alg_kind == ALGORITHM::pooling_max && + context_.prop_kind == + prop_kind::forward_training) { // Max pooling must have workspace. + DCHECK(ws_data != nullptr); + context_.ws_mem->set_data_handle(ws_data, *fwd_stream); + } +#else context_.src_mem->set_data_handle( static_cast(const_cast(src_data))); context_.dst_mem->set_data_handle(static_cast(dst_data)); @@ -136,7 +147,7 @@ void MklPoolingFwdPrimitive::Execute(const T* src_data, T* dst_data, DCHECK(ws_data != nullptr); context_.ws_mem->set_data_handle(ws_data); } - +#endif // ENABLE_MKLDNN_THREADPOOL #ifdef ENABLE_MKLDNN_V1 execute_primitives(context_.fwd_primitives, fwd_stream, context_.net_args); #else @@ -269,6 +280,16 @@ template void MklPoolingBwdPrimitive::Execute(const T* diff_dst_data, T* diff_src_data, const void* ws_data, std::shared_ptr bwd_stream) { +#ifdef ENABLE_MKLDNN_THREADPOOL + context_.diff_dst_mem->set_data_handle( + static_cast(const_cast(diff_dst_data)), *bwd_stream); + context_.diff_src_mem->set_data_handle(static_cast(diff_src_data), + *bwd_stream); + if (context_.alg_kind == ALGORITHM::pooling_max) { + DCHECK(ws_data != nullptr); + context_.ws_mem->set_data_handle(const_cast(ws_data), *bwd_stream); + } +#else context_.diff_dst_mem->set_data_handle( static_cast(const_cast(diff_dst_data))); context_.diff_src_mem->set_data_handle(static_cast(diff_src_data)); @@ -276,7 +297,7 @@ void MklPoolingBwdPrimitive::Execute(const T* diff_dst_data, DCHECK(ws_data != nullptr); context_.ws_mem->set_data_handle(const_cast(ws_data)); } - +#endif // ENABLE_MKLDNN_THREADPOOL #ifdef ENABLE_MKLDNN_V1 execute_primitives(context_.bwd_primitives, bwd_stream, context_.net_args); #else diff --git a/tensorflow/core/kernels/mkl_quantize_op.cc b/tensorflow/core/kernels/mkl_quantize_op.cc index 5adb9862250..177cbb43d0b 100644 --- a/tensorflow/core/kernels/mkl_quantize_op.cc +++ b/tensorflow/core/kernels/mkl_quantize_op.cc @@ -88,8 +88,13 @@ class MklReorderWithScalePrimitive : public MklPrimitive { void Execute(void* src_data, void* dst_data, std::shared_ptr reorder_stream) { +#ifdef ENABLE_MKLDNN_THREADPOOL + context_.src_mem->set_data_handle(src_data, *reorder_stream); + context_.dst_mem->set_data_handle(dst_data, *reorder_stream); +#else context_.src_mem->set_data_handle(src_data); context_.dst_mem->set_data_handle(dst_data); +#endif // ENABLE_MKLDNN_THREADPOOL #ifndef ENABLE_MKLDNN_V1 reorder_stream->submit(context_.net); #else diff --git a/tensorflow/core/kernels/mkl_relu_op.cc b/tensorflow/core/kernels/mkl_relu_op.cc index 784bbc682dc..9af580de777 100644 --- a/tensorflow/core/kernels/mkl_relu_op.cc +++ b/tensorflow/core/kernels/mkl_relu_op.cc @@ -79,10 +79,16 @@ class MklEltwiseFwdPrimitive : public MklPrimitive { // dst_data: output data buffer of dst void Execute(const T* src_data, T* dst_data, std::shared_ptr fwd_stream) { +#ifdef ENABLE_MKLDNN_THREADPOOL + context_.src_mem->set_data_handle( + static_cast(const_cast(src_data)), *fwd_stream); + context_.dst_mem->set_data_handle(static_cast(dst_data), + *fwd_stream); +#else context_.src_mem->set_data_handle( static_cast(const_cast(src_data))); context_.dst_mem->set_data_handle(static_cast(dst_data)); - +#endif // ENABLE_MKLDNN_THREADPOOL #ifdef ENABLE_MKLDNN_V1 DCHECK_EQ(context_.fwd_primitives.size(), context_.fwd_primitives_args.size()); @@ -293,12 +299,20 @@ class MklEltwiseBwdPrimitive : public MklPrimitive { // diff_src_data: output data buffer of diff_src void Execute(const T* src_data, const T* diff_dst_data, T* diff_src_data, std::shared_ptr bwd_stream) { +#ifdef ENABLE_MKLDNN_THREADPOOL + context_.src_mem->set_data_handle( + static_cast(const_cast(src_data)), *bwd_stream); + context_.diff_dst_mem->set_data_handle( + static_cast(const_cast(diff_dst_data)), *bwd_stream); + context_.diff_src_mem->set_data_handle(static_cast(diff_src_data), + *bwd_stream); +#else context_.src_mem->set_data_handle( static_cast(const_cast(src_data))); context_.diff_dst_mem->set_data_handle( static_cast(const_cast(diff_dst_data))); context_.diff_src_mem->set_data_handle(static_cast(diff_src_data)); - +#endif // ENABLE_MKLDNN_THREADPOOL #ifdef ENABLE_MKLDNN_V1 DCHECK_EQ(context_.bwd_primitives.size(), context_.bwd_primitives_args.size()); diff --git a/tensorflow/core/kernels/mkl_slice_op.cc b/tensorflow/core/kernels/mkl_slice_op.cc index 4115691c79d..7e293e14d98 100644 --- a/tensorflow/core/kernels/mkl_slice_op.cc +++ b/tensorflow/core/kernels/mkl_slice_op.cc @@ -189,9 +189,15 @@ class MklSlicePrimitive : public MklPrimitive { void Execute(const MklSliceParams& sliceParams, std::shared_ptr slice_stream) { +#ifdef ENABLE_MKLDNN_THREADPOOL + context_.src_mem->set_data_handle(sliceParams.from->get_data_handle(), + *slice_stream); + context_.dst_mem->set_data_handle(sliceParams.to->get_data_handle(), + *slice_stream); +#else context_.src_mem->set_data_handle(sliceParams.from->get_data_handle()); context_.dst_mem->set_data_handle(sliceParams.to->get_data_handle()); - +#endif // ENABLE_MKLDNN_THREADPOOL #ifdef ENABLE_MKLDNN_V1 execute_primitives(context_.slice_primitives, slice_stream, context_.slice_primitives_args); diff --git a/tensorflow/core/kernels/mkl_softmax_op.cc b/tensorflow/core/kernels/mkl_softmax_op.cc index 4d1cf90f28d..2f51573fe13 100644 --- a/tensorflow/core/kernels/mkl_softmax_op.cc +++ b/tensorflow/core/kernels/mkl_softmax_op.cc @@ -59,10 +59,16 @@ class MklSoftmaxPrimitive : public MklPrimitive { // dst_data: output data buffer of dst void Execute(const T* src_data, T* dst_data, std::shared_ptr fwd_cpu_stream) { +#ifdef ENABLE_MKLDNN_THREADPOOL + context_.src_mem->set_data_handle( + static_cast(const_cast(src_data)), *fwd_cpu_stream); + context_.dst_mem->set_data_handle(static_cast(dst_data), + *fwd_cpu_stream); +#else context_.src_mem->set_data_handle( static_cast(const_cast(src_data))); context_.dst_mem->set_data_handle(static_cast(dst_data)); - +#endif // ENABLE_MKLDNN_THREADPOOL #ifdef ENABLE_MKLDNN_V1 DCHECK_EQ(context_.fwd_primitives.size(), context_.fwd_net_args.size()); execute_primitives(context_.fwd_primitives, fwd_cpu_stream, From 5016da312802f8372672ccbfa7a4207b8683a8e5 Mon Sep 17 00:00:00 2001 From: sshiddib Date: Mon, 15 Jun 2020 13:21:21 -0700 Subject: [PATCH 091/867] [Intel MKL] Adding DNNL ops (part 2) supporting threadpool work --- tensorflow/core/kernels/mkl_aggregate_ops.cc | 7 +++++-- tensorflow/core/kernels/mkl_concat_op.cc | 17 ++++++++++++++--- tensorflow/core/kernels/mkl_dequantize_op.cc | 7 +++++-- tensorflow/core/kernels/mkl_lrn_op.cc | 6 ++++-- tensorflow/core/kernels/mkl_transpose_op.cc | 5 +++-- tensorflow/core/util/mkl_util.h | 14 ++++++++++++-- 6 files changed, 43 insertions(+), 13 deletions(-) diff --git a/tensorflow/core/kernels/mkl_aggregate_ops.cc b/tensorflow/core/kernels/mkl_aggregate_ops.cc index ec5f80cb3fa..90e0ea9aa95 100644 --- a/tensorflow/core/kernels/mkl_aggregate_ops.cc +++ b/tensorflow/core/kernels/mkl_aggregate_ops.cc @@ -178,6 +178,9 @@ class MklAddNOp : public OpKernel { dnn_fmt = MklTensorFormatToMklDnnDataFormat(mkl_data_format); } + std::shared_ptr fwd_cpu_stream; + fwd_cpu_stream.reset(CreateStream(ctx, cpu_engine)); + // Create memory descriptor for MKL-DNN. // If all input in Tensorflow format, create block memory descriptor, // else convert TF format to MKL memory descriptor @@ -215,6 +218,7 @@ class MklAddNOp : public OpKernel { srcs_pd.push_back(memory::primitive_desc(md, cpu_engine)); #endif src.SetUsrMem(md, &src_tensor); + src.SetUsrMemDataHandle(&src_tensor, fwd_cpu_stream); inputs.push_back(src.GetOpMem()); } @@ -240,11 +244,10 @@ class MklAddNOp : public OpKernel { } AllocateOutputSetMklShape(ctx, kOutputIdx, &dst_tensor, output_tf_shape, output_mkl_shape); - dst.SetUsrMemDataHandle(dst_tensor); + dst.SetUsrMemDataHandle(dst_tensor, fwd_cpu_stream); // Create Sum op, and submit net for execution. std::vector net; - stream* fwd_cpu_stream = CreateStream(ctx, cpu_engine); #ifdef ENABLE_MKLDNN_V1 mkldnn::sum sum_op(sum_pd); std::unordered_map net_args = { diff --git a/tensorflow/core/kernels/mkl_concat_op.cc b/tensorflow/core/kernels/mkl_concat_op.cc index 976f778424e..4a5cb0a0d4f 100644 --- a/tensorflow/core/kernels/mkl_concat_op.cc +++ b/tensorflow/core/kernels/mkl_concat_op.cc @@ -281,11 +281,19 @@ class MklConcatFwdPrimitive : public MklPrimitive { std::shared_ptr fwd_stream) { DCHECK_EQ(in_data.size(), context_.data_mem.size()); for (size_t i = 0; i < concat_fwd_dims.num_inputs; i++) { +#ifdef ENABLE_MKLDNN_THREADPOOL + context_.data_mem_shdptr[i]->set_data_handle( + static_cast(in_data[i].get_data_handle()), *fwd_stream); + } + context_.dst_mem->set_data_handle( + static_cast(dst_data.get_data_handle()), *fwd_stream); +#else context_.data_mem_shdptr[i]->set_data_handle( static_cast(in_data[i].get_data_handle())); } context_.dst_mem->set_data_handle( static_cast(dst_data.get_data_handle())); +#endif // ENABLE_MKLDNN_THREADPOOL for (size_t i = 0; i < concat_fwd_dims.num_inputs; i++) { context_.data_mem[i] = *context_.data_mem_shdptr[i]; @@ -788,11 +796,13 @@ class MklConcatOp : public OpKernel { dnn_shape_dst); DCHECK(dst_tensor != nullptr) << "Output tensor pointer is NULL"; + std::shared_ptr fwd_cpu_stream; + fwd_cpu_stream.reset(CreateStream(context, cpu_engine)); + if (dnn_shape_dst.IsMklTensor()) dst_md = dnn_shape_dst.GetMklLayout(); dst.SetUsrMem(dst_md, dst_tensor); - std::shared_ptr fwd_cpu_stream; - fwd_cpu_stream.reset(CreateStream(context, cpu_engine)); + dst.SetUsrMemDataHandle(dst_tensor, fwd_cpu_stream); #ifdef ENABLE_MKLDNN_V1 auto concat_op = concat(concat_pd); std::unordered_map net_args = { @@ -830,9 +840,10 @@ class MklConcatOp : public OpKernel { dst_md = dnn_shape_dst.IsMklTensor() ? dnn_shape_dst.GetMklLayout() : dst_md; - dst.SetUsrMem(dst_md, dst_tensor); std::shared_ptr fwd_cpu_stream; fwd_cpu_stream.reset(CreateStream(context, concat_fwd->GetEngine())); + dst.SetUsrMem(dst_md, dst_tensor); + dst.SetUsrMemDataHandle(dst_tensor, fwd_cpu_stream); // Execute concat concat_fwd->Execute(srcs_mem, dst.GetOpMem(), concat_fwd_dims, fwd_cpu_stream); diff --git a/tensorflow/core/kernels/mkl_dequantize_op.cc b/tensorflow/core/kernels/mkl_dequantize_op.cc index 06570c1db1c..82d78250576 100644 --- a/tensorflow/core/kernels/mkl_dequantize_op.cc +++ b/tensorflow/core/kernels/mkl_dequantize_op.cc @@ -75,6 +75,9 @@ class MklDequantizeOp : public OpKernel { MklDnnData src(&cpu_engine); MklDnnData dst(&cpu_engine); + std::shared_ptr reorder_stream; + reorder_stream.reset(CreateStream(ctx, cpu_engine)); + // If input is in MKL layout, then simply grab input layout; otherwise, // construct input TF layout. For TF layout, although input shape // (src_dims) required is in MKL-DNN order, the layout is Tensorflow's @@ -85,6 +88,7 @@ class MklDequantizeOp : public OpKernel { : memory::desc(src_dims, MklDnnType(), MEMORY_FORMAT::nhwc); src.SetUsrMem(src_md, &src_tensor); + src.SetUsrMemDataHandle(&src_tensor, reorder_stream); Tensor* output_tensor = nullptr; MklDnnShape output_mkl_shape; @@ -129,6 +133,7 @@ class MklDequantizeOp : public OpKernel { AllocateOutputSetMklShape(ctx, 0, &output_tensor, output_tf_shape, output_mkl_shape); dst.SetUsrMem(dst_md, output_tensor); + dst.SetUsrMemDataHandle(output_tensor, reorder_stream); // The quantization logic here for mode SCALED is similar to the logic // in QuantizeAndDequantizeV2 and QuantizeAndDequantizeV3. @@ -155,8 +160,6 @@ class MklDequantizeOp : public OpKernel { // Also it does not define round_nearest (enum). attr.set_int_output_round_mode(mkldnn::round_mode::round_nearest); #endif // !ENABLE_MKLDNN_V1 - std::shared_ptr reorder_stream; - reorder_stream.reset(CreateStream(ctx, cpu_engine)); std::vector net; // Create reorder primitive and then execute. diff --git a/tensorflow/core/kernels/mkl_lrn_op.cc b/tensorflow/core/kernels/mkl_lrn_op.cc index a11e7ebcbf5..3e512d0792b 100644 --- a/tensorflow/core/kernels/mkl_lrn_op.cc +++ b/tensorflow/core/kernels/mkl_lrn_op.cc @@ -137,6 +137,7 @@ class MklLRNOp : public OpKernel { // that input is in NHWC layout with Channel being the last dimension. src_dnn_data.SetUsrMem(src_md, &src_tensor); src_dnn_data.SetOpMemDesc(input_dims, MEMORY_FORMAT::nhwc); + src_dnn_data.SetUsrMemDataHandle(&src_tensor, fwd_stream_); // dst_dnn_data has the same shape as input. dst_dnn_data.SetUsrMem(src_md); @@ -157,7 +158,7 @@ class MklLRNOp : public OpKernel { &output_tensor); OP_REQUIRES_OK(context, context->status()); DCHECK(output_tensor != nullptr); - dst_dnn_data.SetUsrMemDataHandle(output_tensor); + dst_dnn_data.SetUsrMemDataHandle(output_tensor, fwd_stream_); // Handle workspace required for MKL-DNN. AllocateWorkspaceTensor(context, lrn_prim_desc, &workspace_dnn_data); @@ -393,6 +394,7 @@ class MklLRNGradOp : public OpKernel { orig_input_dnn_shape.GetSizesAsMklDnnDims(); orig_input_dnn_data.SetUsrMem(orig_input_md, &orig_input_tensor); orig_input_dnn_data.SetOpMemDesc(orig_input_dims, MEMORY_FORMAT::nhwc); + orig_input_dnn_data.SetUsrMemDataHandle(&orig_input_tensor, bwd_stream_); // output_dnn_data has the same shape as original input output_dnn_data.SetUsrMem(orig_input_md); @@ -421,7 +423,7 @@ class MklLRNGradOp : public OpKernel { orig_input_format, &output_tensor); OP_REQUIRES_OK(context, context->status()); DCHECK(output_tensor != nullptr); - output_dnn_data.SetUsrMemDataHandle(output_tensor); + output_dnn_data.SetUsrMemDataHandle(output_tensor, bwd_stream_); // Create LRN primitive and add it to the net // At this point, workspace is enabled, so we don't need diff --git a/tensorflow/core/kernels/mkl_transpose_op.cc b/tensorflow/core/kernels/mkl_transpose_op.cc index 77a68afa752..2e5c6d2719b 100644 --- a/tensorflow/core/kernels/mkl_transpose_op.cc +++ b/tensorflow/core/kernels/mkl_transpose_op.cc @@ -137,6 +137,7 @@ Status MKLTransposeND(OpKernelContext* context, const Tensor& in_tensor, memory::dims out_strides = ReorderStrides(CalculateTFStrides(out_dims), perm); + std::shared_ptr transpose_stream; in.SetUsrMem(in_dims, in_strides, &in_tensor); // Output dimensions are same as input dimensions. We adjust the layout // using strides. @@ -144,16 +145,16 @@ Status MKLTransposeND(OpKernelContext* context, const Tensor& in_tensor, std::vector net; #ifdef ENABLE_MKLDNN_V1 - std::shared_ptr transpose_stream; auto* prim = FindOrCreateReorder(in.GetUsrMem(), out.GetUsrMem()); transpose_stream.reset(CreateStream(context, prim->GetEngine())); + in.SetUsrMemDataHandle(&in_tensor, transpose_stream); + out.SetUsrMemDataHandle(out_tensor, transpose_stream); net.push_back(*(prim->GetPrimitive())); std::vector net_args; net_args.push_back({{MKLDNN_ARG_FROM, *in.GetUsrMem()}, {MKLDNN_ARG_TO, *out.GetUsrMem()}}); execute_primitives(net, transpose_stream, net_args); #else - std::shared_ptr transpose_stream; transpose_stream.reset(new CPU_STREAM(cpu_engine)); net.push_back(FindOrCreateReorder(in.GetUsrMem(), out.GetUsrMem())); transpose_stream->submit(net).wait(); diff --git a/tensorflow/core/util/mkl_util.h b/tensorflow/core/util/mkl_util.h index 7f6272b09c1..996984eebc0 100644 --- a/tensorflow/core/util/mkl_util.h +++ b/tensorflow/core/util/mkl_util.h @@ -1524,17 +1524,27 @@ class MklDnnData { } /// Set function for data buffer of user memory primitive. - inline void SetUsrMemDataHandle(void* data_buffer) { + inline void SetUsrMemDataHandle(void* data_buffer, + std::shared_ptr t_stream = nullptr) { CHECK_NOTNULL(user_memory_); CHECK_NOTNULL(data_buffer); +#ifdef ENABLE_MKLDNN_THREADPOOL + user_memory_->set_data_handle(data_buffer, *t_stream); +#else user_memory_->set_data_handle(data_buffer); +#endif // ENABLE_MKLDNN_THREADPOOL } /// Set function for data buffer of user memory primitive. - inline void SetUsrMemDataHandle(const Tensor* tensor) { + inline void SetUsrMemDataHandle(const Tensor* tensor, + std::shared_ptr t_stream = nullptr) { CHECK_NOTNULL(user_memory_); CHECK_NOTNULL(tensor); +#ifdef ENABLE_MKLDNN_THREADPOOL + user_memory_->set_data_handle(GetTensorBuffer(tensor), *t_stream); +#else user_memory_->set_data_handle(GetTensorBuffer(tensor)); +#endif // ENABLE_MKLDNN_THREADPOOL } /// allocate function for data buffer From 609a60b44bbf934b31a1dce4f0aa84e731b83c35 Mon Sep 17 00:00:00 2001 From: Lukas Geiger Date: Tue, 16 Jun 2020 00:41:09 +0100 Subject: [PATCH 092/867] Refactor AutoCastVariable tests to rely on strategy_combinations --- .../keras/mixed_precision/experimental/BUILD | 3 +- .../experimental/autocast_variable_test.py | 130 +++++++++--------- 2 files changed, 64 insertions(+), 69 deletions(-) diff --git a/tensorflow/python/keras/mixed_precision/experimental/BUILD b/tensorflow/python/keras/mixed_precision/experimental/BUILD index 024b093c469..4060e455f84 100644 --- a/tensorflow/python/keras/mixed_precision/experimental/BUILD +++ b/tensorflow/python/keras/mixed_precision/experimental/BUILD @@ -144,9 +144,10 @@ py_test( "//tensorflow/python:client_testlib", "//tensorflow/python:framework", "//tensorflow/python:platform_test", + "//tensorflow/python/distribute:combinations", "//tensorflow/python/distribute:mirrored_strategy", + "//tensorflow/python/distribute:strategy_combinations", "//tensorflow/python/eager:context", - "//tensorflow/python/keras:combinations", "//tensorflow/python/keras/optimizer_v2", "@absl_py//absl/testing:parameterized", ], diff --git a/tensorflow/python/keras/mixed_precision/experimental/autocast_variable_test.py b/tensorflow/python/keras/mixed_precision/experimental/autocast_variable_test.py index 78041973cc1..95957f5634e 100644 --- a/tensorflow/python/keras/mixed_precision/experimental/autocast_variable_test.py +++ b/tensorflow/python/keras/mixed_precision/experimental/autocast_variable_test.py @@ -17,20 +17,23 @@ from __future__ import absolute_import from __future__ import division from __future__ import print_function +import contextlib import os from absl.testing import parameterized import numpy as np from tensorflow.python import tf2 +from tensorflow.python.distribute import combinations +from tensorflow.python.distribute import distribution_strategy_context as ds_context from tensorflow.python.distribute import mirrored_strategy +from tensorflow.python.distribute import strategy_combinations from tensorflow.python.eager import context from tensorflow.python.eager import def_function from tensorflow.python.framework import constant_op from tensorflow.python.framework import dtypes from tensorflow.python.framework import indexed_slices from tensorflow.python.framework import ops -from tensorflow.python.keras import combinations from tensorflow.python.keras.mixed_precision.experimental import autocast_variable from tensorflow.python.keras.optimizer_v2 import gradient_descent as gradient_descent_v2 from tensorflow.python.ops import array_ops @@ -40,30 +43,17 @@ from tensorflow.python.platform import test from tensorflow.python.training import gradient_descent as gradient_descent_v1 from tensorflow.python.training.tracking import util as trackable_utils -TESTCASES = ({ - 'testcase_name': 'base', - 'distribute': False -}, { - 'testcase_name': 'distribute', - 'distribute': True -}) - - -def get_distribute_scope(distribute): - - class DummyContextManager(object): - - def __enter__(self): - pass - - def __exit__(self, *args): - pass - - if distribute: - return mirrored_strategy.MirroredStrategy(['cpu:0']).scope() - else: - return DummyContextManager() +class DummyStrategy(object): + @contextlib.contextmanager + def scope(self): + yield +maybe_distribute = combinations.combine( + distribution=[ + combinations.NamedDistribution( + "Dummy", lambda: DummyStrategy(), required_gpus=None), + strategy_combinations.mirrored_strategy_with_cpu_1_and_2 + ]) def get_var(val, dtype, name=None): return variables.VariableV1(val, use_resource=True, dtype=dtype, name=name) @@ -71,10 +61,13 @@ def get_var(val, dtype, name=None): @combinations.generate(combinations.combine(mode=['graph', 'eager'])) class AutoCastVariableTest(test.TestCase, parameterized.TestCase): + def setUp(self): + strategy_combinations.set_virtual_cpus_to_at_least(3) + super(AutoCastVariableTest, self).setUp() - @parameterized.named_parameters(*TESTCASES) - def test_read(self, distribute): - with get_distribute_scope(distribute): + @combinations.generate(maybe_distribute) + def test_read(self, distribution): + with distribution.scope(): x = get_var(1., dtypes.float32) x = autocast_variable.create_autocast_variable(x) self.evaluate(x.initializer) @@ -116,9 +109,9 @@ class AutoCastVariableTest(test.TestCase, parameterized.TestCase): self.assertEqual(x.sparse_read([0]).dtype, dtypes.float16) self.assertEqual(x.gather_nd([0]).dtype, dtypes.float16) - @parameterized.named_parameters(*TESTCASES) - def test_read_nested_scopes(self, distribute): - with get_distribute_scope(distribute): + @combinations.generate(maybe_distribute) + def test_read_nested_scopes(self, distribution): + with distribution.scope(): x = get_var(1., dtypes.float32) x = autocast_variable.create_autocast_variable(x) self.evaluate(x.initializer) @@ -136,9 +129,9 @@ class AutoCastVariableTest(test.TestCase, parameterized.TestCase): self.assertEqual(x.dtype, dtypes.float16) self.assertEqual(x.read_value().dtype, dtypes.float16) - @parameterized.named_parameters(*TESTCASES) - def test_dtype_is_not_string(self, distribute): - with get_distribute_scope(distribute): + @combinations.generate(maybe_distribute) + def test_dtype_is_not_string(self, distribution): + with distribution.scope(): x = get_var(1., dtypes.float32) x = autocast_variable.create_autocast_variable(x) self.assertEqual(x.dtype, dtypes.float32) @@ -153,13 +146,13 @@ class AutoCastVariableTest(test.TestCase, parameterized.TestCase): self.assertEqual(x.true_dtype, dtypes.float32) self.assertIsInstance(x.true_dtype, dtypes.DType) - @parameterized.named_parameters(*TESTCASES) - def test_method_delegations(self, distribute): + @combinations.generate(maybe_distribute) + def test_method_delegations(self, distribution): # Test AutoCastVariable correctly delegates Variable methods to the # underlying variable. - with self.test_session(), get_distribute_scope(distribute): + with self.test_session(), distribution.scope(): for read_dtype in (dtypes.float32, dtypes.float16): - if distribute: + if ds_context.has_strategy(): # MirroredVariable.assign will (incorrectly) return a Mirrored value # instead of a MirroredVariable. So we cannot properly wrap it in an # AutoCastVariable. @@ -183,14 +176,14 @@ class AutoCastVariableTest(test.TestCase, parameterized.TestCase): self.assertEqual(x.aggregation, x._variable.aggregation) self.assertEqual(self.evaluate(x.initialized_value()), 7) if not context.executing_eagerly(): - if not distribute: + if not ds_context.has_strategy(): # These functions are not supported for DistributedVariables x.load(9) self.assertEqual(x.eval(), 9) self.assertEqual(self.evaluate(x.initial_value), 7) self.assertEqual(x.op, x._variable.op) self.assertEqual(x.graph, x._variable.graph) - if not distribute: + if not ds_context.has_strategy(): # These attributes are not supported for DistributedVariables self.assertIsNone(x.constraint) self.assertEqual(x.initializer, x._variable.initializer) @@ -202,7 +195,7 @@ class AutoCastVariableTest(test.TestCase, parameterized.TestCase): self.assertEqual(x.shape, ()) self.assertEqual(x.get_shape(), ()) - if not distribute: + if not ds_context.has_strategy(): # Test scatter_* methods. These are not supported for # DistributedVariables x = get_var([7, 8], dtypes.float32) @@ -233,9 +226,9 @@ class AutoCastVariableTest(test.TestCase, parameterized.TestCase): self.assertAllEqual( evaluate(x.scatter_nd_update([[0], [1]], [1., 2.])), [1, 2]) - @parameterized.named_parameters(*TESTCASES) - def test_operator_overloads(self, distribute): - with get_distribute_scope(distribute): + @combinations.generate(maybe_distribute) + def test_operator_overloads(self, distribution): + with distribution.scope(): for read_dtype in (dtypes.float32, dtypes.float16): x = get_var(7., dtypes.float32) x = autocast_variable.create_autocast_variable(x) @@ -280,9 +273,9 @@ class AutoCastVariableTest(test.TestCase, parameterized.TestCase): self.assertAllEqual(x == [7., 8., 10.], [True, True, False]) self.assertAllEqual(x != [7., 8., 10.], [False, False, True]) - @parameterized.named_parameters(*TESTCASES) - def test_assign(self, distribute): - with get_distribute_scope(distribute): + @combinations.generate(maybe_distribute) + def test_assign(self, distribution): + with distribution.scope(): x = get_var(0., dtypes.float32) x = autocast_variable.create_autocast_variable(x) self.evaluate(x.initializer) @@ -318,18 +311,19 @@ class AutoCastVariableTest(test.TestCase, parameterized.TestCase): self.assertAllClose(3., self.evaluate(x.assign_sub(3.))) # Assign multiple times - assign = x.assign(1.) - self.assertAllClose(1., self.evaluate(assign)) - self.assertAllClose(0., self.evaluate(assign.assign(0.))) - assign_add = x.assign_add(3.) - self.assertAllClose(3., self.evaluate(assign_add)) - self.assertAllClose(3. * 3, - self.evaluate(x.assign_add(3.).assign_add(3.))) - self.assertAllClose(3. * 3, x) - assign_sub = x.assign_sub(3.) - self.assertAllClose(3. * 2, self.evaluate(assign_sub)) - self.assertAllClose(0., - self.evaluate(x.assign_sub(3.).assign_sub(3.))) + if not ds_context.has_strategy(): + assign = x.assign(1.) + self.assertAllClose(1., self.evaluate(assign)) + self.assertAllClose(0., self.evaluate(assign.assign(0.))) + assign_add = x.assign_add(3.) + self.assertAllClose(3., self.evaluate(assign_add)) + self.assertAllClose(3. * 3, + self.evaluate(x.assign_add(3.).assign_add(3.))) + self.assertAllClose(3. * 3, x) + assign_sub = x.assign_sub(3.) + self.assertAllClose(3. * 2, self.evaluate(assign_sub)) + self.assertAllClose(0., + self.evaluate(x.assign_sub(3.).assign_sub(3.))) # Assign with read_value=False self.assertIsNone(self.evaluate(x.assign(1., read_value=False))) @@ -355,9 +349,9 @@ class AutoCastVariableTest(test.TestCase, parameterized.TestCase): # assign still expect float32 value even if in float16 scope run_and_check() - @parameterized.named_parameters(*TESTCASES) - def test_assign_stays_in_true_dtype(self, distribute): - with get_distribute_scope(distribute): + @combinations.generate(maybe_distribute) + def test_assign_stays_in_true_dtype(self, distribution): + with distribution.scope(): x = get_var(1., dtypes.float32) x = autocast_variable.create_autocast_variable(x) self.evaluate(x.initializer) @@ -382,10 +376,10 @@ class AutoCastVariableTest(test.TestCase, parameterized.TestCase): self.assertEqual(1., self.evaluate(x.value())) self.assertEqual(1. + small_val, self.evaluate(x.value())) - @parameterized.named_parameters(*TESTCASES) - def test_checkpoint(self, distribute): + @combinations.generate(maybe_distribute) + def test_checkpoint(self, distribution): with self.test_session(): - with get_distribute_scope(distribute): + with distribution.scope(): x = get_var(1., dtypes.float32) x = autocast_variable.create_autocast_variable(x) self.evaluate(x.initializer) @@ -398,9 +392,9 @@ class AutoCastVariableTest(test.TestCase, parameterized.TestCase): checkpoint.restore(save_path).assert_consumed().run_restore_ops() self.assertEqual(self.evaluate(x), 123.) - @parameterized.named_parameters(*TESTCASES) - def test_invalid_wrapped_variable(self, distribute): - with get_distribute_scope(distribute): + @combinations.generate(maybe_distribute) + def test_invalid_wrapped_variable(self, distribution): + with distribution.scope(): # Wrap a non-variable with self.assertRaisesRegexp(ValueError, 'variable must be of type'): x = constant_op.constant([1.], dtype=dtypes.float32) @@ -443,7 +437,7 @@ class AutoCastVariableTest(test.TestCase, parameterized.TestCase): ) def test_repr_distributed(self): - with get_distribute_scope(distribute=True): + with mirrored_strategy.MirroredStrategy(["/cpu:1", "/cpu:2"]).scope(): x = get_var(1., dtypes.float32) x = autocast_variable.create_autocast_variable(x) self.assertRegexpMatches( From fe6580b4d85b12a2ce7b1a529b70fdbfeedd899e Mon Sep 17 00:00:00 2001 From: Lukas Geiger Date: Tue, 16 Jun 2020 02:16:09 +0100 Subject: [PATCH 093/867] Add comment about ignoring distributed multi assignment --- .../keras/mixed_precision/experimental/autocast_variable_test.py | 1 + 1 file changed, 1 insertion(+) diff --git a/tensorflow/python/keras/mixed_precision/experimental/autocast_variable_test.py b/tensorflow/python/keras/mixed_precision/experimental/autocast_variable_test.py index 95957f5634e..14f26cdf953 100644 --- a/tensorflow/python/keras/mixed_precision/experimental/autocast_variable_test.py +++ b/tensorflow/python/keras/mixed_precision/experimental/autocast_variable_test.py @@ -311,6 +311,7 @@ class AutoCastVariableTest(test.TestCase, parameterized.TestCase): self.assertAllClose(3., self.evaluate(x.assign_sub(3.))) # Assign multiple times + # This currently only works if no strategy is used if not ds_context.has_strategy(): assign = x.assign(1.) self.assertAllClose(1., self.evaluate(assign)) From da2046b8a3b1bb79c77bf258aa8a52887bc3703a Mon Sep 17 00:00:00 2001 From: Lukas Geiger Date: Tue, 16 Jun 2020 02:25:18 +0100 Subject: [PATCH 094/867] Use default_strategy instead of dummy scope --- .../experimental/autocast_variable_test.py | 9 +-------- 1 file changed, 1 insertion(+), 8 deletions(-) diff --git a/tensorflow/python/keras/mixed_precision/experimental/autocast_variable_test.py b/tensorflow/python/keras/mixed_precision/experimental/autocast_variable_test.py index 14f26cdf953..c45015b644e 100644 --- a/tensorflow/python/keras/mixed_precision/experimental/autocast_variable_test.py +++ b/tensorflow/python/keras/mixed_precision/experimental/autocast_variable_test.py @@ -17,7 +17,6 @@ from __future__ import absolute_import from __future__ import division from __future__ import print_function -import contextlib import os from absl.testing import parameterized @@ -43,15 +42,9 @@ from tensorflow.python.platform import test from tensorflow.python.training import gradient_descent as gradient_descent_v1 from tensorflow.python.training.tracking import util as trackable_utils -class DummyStrategy(object): - @contextlib.contextmanager - def scope(self): - yield - maybe_distribute = combinations.combine( distribution=[ - combinations.NamedDistribution( - "Dummy", lambda: DummyStrategy(), required_gpus=None), + strategy_combinations.default_strategy, strategy_combinations.mirrored_strategy_with_cpu_1_and_2 ]) From 83f19c6a9e84fc6971ad0a7df5874603237a595f Mon Sep 17 00:00:00 2001 From: Yong Tang Date: Mon, 15 Jun 2020 17:16:56 +0000 Subject: [PATCH 095/867] Fix unknown output shape issue in autograph for tf.equal This PR tries to address the issue raised in 40471 where the output shape of an autograph consists of tf.equal could not inference correctly. Specifically `x.shape == [None, 10, 1]` and `y.shape == [None, 1, 4]` only yield `shape == None` (should be `shape == [None, 10, 4]`). The reason was that the shape inbference function for equal didn't capture the cases where both x and y's dim are None. This PR fixes the issue. This PR fixes 40471. Signed-off-by: Yong Tang --- tensorflow/core/framework/common_shape_fns.cc | 3 +++ tensorflow/core/ops/math_ops_test.cc | 2 +- .../python/autograph/operators/logical_test.py | 14 ++++++++++++++ 3 files changed, 18 insertions(+), 1 deletion(-) diff --git a/tensorflow/core/framework/common_shape_fns.cc b/tensorflow/core/framework/common_shape_fns.cc index 113adbdd432..7567db03c23 100644 --- a/tensorflow/core/framework/common_shape_fns.cc +++ b/tensorflow/core/framework/common_shape_fns.cc @@ -1936,6 +1936,7 @@ Status BroadcastBinaryOpOutputShapeFnHelper(InferenceContext* c, // in C++ op code, we must still assert that the unknown dim is either 1 // or the same as the known dim. // - If either dimension is 1, the other dimension is the output. + // - If both are unknown then dimension is unknown if (c->Value(dim_x) > 1) { if (!incompatible_shape_error) { *out = c->UnknownShape(); @@ -1954,6 +1955,8 @@ Status BroadcastBinaryOpOutputShapeFnHelper(InferenceContext* c, dims.push_back(dim_x); } else if (dim_y.SameHandle(dim_x)) { dims.push_back(dim_x); + } else if (!c->ValueKnown(dim_x) && !c->ValueKnown(dim_y)) { + dims.push_back(c->UnknownDim()); } else { if (!incompatible_shape_error) { *out = c->UnknownShape(); diff --git a/tensorflow/core/ops/math_ops_test.cc b/tensorflow/core/ops/math_ops_test.cc index 5c69a2a7f1c..a2837d88bde 100644 --- a/tensorflow/core/ops/math_ops_test.cc +++ b/tensorflow/core/ops/math_ops_test.cc @@ -120,7 +120,7 @@ TEST(MathOpsTest, BroadcastBinaryOps_ShapeFn) { INFER_OK(op, "[1];[?]", "[d1_0]"); INFER_OK(op, "[?];[2]", incompatible_shape_error ? "[d1_0]" : "?"); INFER_OK(op, "[2];[?]", incompatible_shape_error ? "[d0_0]" : "?"); - INFER_OK(op, "[?];[?]", incompatible_shape_error ? "[?]" : "?"); + INFER_OK(op, "[?];[?]", "[?]"); INFER_OK(op, "[];[?]", "[d1_0]"); INFER_OK(op, "[?];[]", "[d0_0]"); diff --git a/tensorflow/python/autograph/operators/logical_test.py b/tensorflow/python/autograph/operators/logical_test.py index e22f39932d1..0eab302a825 100644 --- a/tensorflow/python/autograph/operators/logical_test.py +++ b/tensorflow/python/autograph/operators/logical_test.py @@ -19,7 +19,9 @@ from __future__ import division from __future__ import print_function from tensorflow.python.autograph.operators import logical +from tensorflow.python.eager import def_function from tensorflow.python.framework import constant_op +from tensorflow.python.framework import tensor_spec from tensorflow.python.framework import test_util from tensorflow.python.platform import test @@ -83,6 +85,18 @@ class LogicalOperatorsTest(test.TestCase): t = logical.not_(self._tf_false()) self.assertEqual(self.evaluate(t), True) + # Test case for GitHub issue 40471 + def test_equal_output_shapes(self): + + @def_function.function(input_signature=[ + tensor_spec.TensorSpec([None, 10, 1]), + tensor_spec.TensorSpec([None, 1, 4])]) + def f(x, y): + z = x == y + return z + + self.assertAllEqual(f.get_concrete_function().output_shapes, [None, 10, 4]) + if __name__ == '__main__': test.main() From bedc750c7d06c79ca61b53c6eeaba727ba7e886a Mon Sep 17 00:00:00 2001 From: Meteorix Date: Tue, 16 Jun 2020 23:38:53 +0800 Subject: [PATCH 096/867] fix LaunchDepthwiseConvBackpropFilterOp --- tensorflow/core/kernels/depthwise_conv_op_gpu.h | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/tensorflow/core/kernels/depthwise_conv_op_gpu.h b/tensorflow/core/kernels/depthwise_conv_op_gpu.h index f712d9c48c0..f69878bea89 100644 --- a/tensorflow/core/kernels/depthwise_conv_op_gpu.h +++ b/tensorflow/core/kernels/depthwise_conv_op_gpu.h @@ -1763,7 +1763,7 @@ void LaunchDepthwiseConvBackpropFilterOp::operator()( int num_filter_backprop = args.filter_rows * args.filter_cols * args.out_depth; se::DeviceMemoryBase filter_bp_ptr(filter_backprop, num_filter_backprop); - stream->ThenMemset32(&filter_bp_ptr, 0, num_filter_backprop * sizeof(T)); + stream->ThenMemZero(&filter_bp_ptr, num_filter_backprop * sizeof(T)); if (args.filter_rows == 3 && args.filter_cols == 3) { OP_REQUIRES_OK( From 266387ecb45ca06a2bdf117a119cb65b9ab5dcd9 Mon Sep 17 00:00:00 2001 From: nammbash Date: Tue, 16 Jun 2020 12:02:14 -0700 Subject: [PATCH 097/867] create a single oneDNN string --- tensorflow/core/platform/cpu_feature_guard.cc | 11 ++--------- 1 file changed, 2 insertions(+), 9 deletions(-) diff --git a/tensorflow/core/platform/cpu_feature_guard.cc b/tensorflow/core/platform/cpu_feature_guard.cc index a020d3fd70e..13116de49dc 100644 --- a/tensorflow/core/platform/cpu_feature_guard.cc +++ b/tensorflow/core/platform/cpu_feature_guard.cc @@ -138,16 +138,9 @@ void InfoAboutUnusedCPUFeatures() { CheckIfFeatureUnused(CPUFeature::FMA, "FMA", missing_instructions); #endif // __FMA__ #endif // else of if defined(_MSC_VER) && !defined(__clang__) - - string intel_library_official_name( - "Intel(R) oneAPI Deep Neural Network Library (oneDNN) "); -#ifndef INTEL_MKL - intel_library_official_name = "oneAPI Deep Neural Network Library (oneDNN) "; -#endif - - if (!missing_instructions.empty()) { + if (!missing_instructions.empty()) { LOG(INFO) << "This TensorFlow binary is optimized with " - << intel_library_official_name + << "oneAPI Deep Neural Network Library (oneDNN)" << "to use the following CPU instructions in performance-" << "critical operations: " << missing_instructions << std::endl << "To enable them in other operations, rebuild TensorFlow " From 4334d8b06f333bee5dc17c942ea41beb5bd9c32d Mon Sep 17 00:00:00 2001 From: Mahmoud Abuzaina Date: Tue, 16 Jun 2020 14:26:13 -0700 Subject: [PATCH 098/867] Fixing build issue on windows --- third_party/mkl_dnn/mkldnn_v1.BUILD | 2 ++ 1 file changed, 2 insertions(+) diff --git a/third_party/mkl_dnn/mkldnn_v1.BUILD b/third_party/mkl_dnn/mkldnn_v1.BUILD index 47a7efecda3..7bdec138b99 100644 --- a/third_party/mkl_dnn/mkldnn_v1.BUILD +++ b/third_party/mkl_dnn/mkldnn_v1.BUILD @@ -71,6 +71,8 @@ cc_library( "src/cpu/**/*.cpp", "src/cpu/**/*.hpp", "src/cpu/xbyak/*.h", + "src/cpu/jit_utils/jitprofiling/*.c", + "src/cpu/jit_utils/jitprofiling/*.h", ]) + [ ":dnnl_config_h", ":dnnl_version_h", From f3d3480c811480c2b966df1086828e47131eb783 Mon Sep 17 00:00:00 2001 From: Koan-Sin Tan Date: Wed, 17 Jun 2020 09:25:28 +0800 Subject: [PATCH 099/867] [tflite] make label_image build on linux and macOS label_image doesn't build on Linux and macOS platforms ``` bazel build --config opt //tensorflow/lite/examples/label_image:label_image ``` shows something like ``` ERROR: /home/freedom/work/tensorflow/tensorflow/lite/examples/label_image/BUILD:15:1: undeclared inclusion(s) in rule '//tensorflow/lite/examples/label_image:label_image': this rule is missing dependency declarations for the following files included by 'tensorflow/lite/examples/label_image/label_image.cc': 'external/com_google_absl/absl/strings/string_view.h' 'external/com_google_absl/absl/base/internal/throw_delegate.h' ``` Add `"@com_google_absl//absl/strings"` to deps --- tensorflow/lite/examples/label_image/BUILD | 1 + 1 file changed, 1 insertion(+) diff --git a/tensorflow/lite/examples/label_image/BUILD b/tensorflow/lite/examples/label_image/BUILD index 01296b0b2a0..633f767c5e9 100644 --- a/tensorflow/lite/examples/label_image/BUILD +++ b/tensorflow/lite/examples/label_image/BUILD @@ -38,6 +38,7 @@ cc_binary( "//tensorflow/lite/profiling:profiler", "//tensorflow/lite/tools/evaluation:utils", "@com_google_absl//absl/memory", + "@com_google_absl//absl/strings", ] + select({ "//tensorflow:android": [ "//tensorflow/lite/delegates/gpu:delegate", From 2b9a7d5e40aa85e84f7d6196d3e7527675d33038 Mon Sep 17 00:00:00 2001 From: peng Date: Wed, 17 Jun 2020 11:21:12 +0800 Subject: [PATCH 100/867] skip if no cuda --- tensorflow/core/grappler/optimizers/remapper_test.cc | 3 +++ 1 file changed, 3 insertions(+) diff --git a/tensorflow/core/grappler/optimizers/remapper_test.cc b/tensorflow/core/grappler/optimizers/remapper_test.cc index 831aa51ee72..a947210d8a6 100644 --- a/tensorflow/core/grappler/optimizers/remapper_test.cc +++ b/tensorflow/core/grappler/optimizers/remapper_test.cc @@ -452,6 +452,9 @@ TEST_F(RemapperTest, FuseMatMulWithBias) { TEST_F(RemapperTest, FuseConv2DWithBiasAndActivationOnGPU) { using ::tensorflow::ops::Placeholder; +#if !(GOOGLE_CUDA) + GTEST_SKIP() << "No CUDA, skip FuseConv2DWithBiasAndActivation on GPU"; +#endif // !GOOGLE_CUDA tensorflow::Scope s = tensorflow::Scope::NewRootScope(); auto input_shape = Placeholder::Shape({8, 32, 32, 3}); From 9bf28d66871931c16ba8269873918a914bc3efe7 Mon Sep 17 00:00:00 2001 From: peng Date: Wed, 17 Jun 2020 11:24:34 +0800 Subject: [PATCH 101/867] delete empty lines --- tensorflow/core/grappler/optimizers/remapper.cc | 2 -- 1 file changed, 2 deletions(-) diff --git a/tensorflow/core/grappler/optimizers/remapper.cc b/tensorflow/core/grappler/optimizers/remapper.cc index b738a23de23..a6ddad8dfc8 100644 --- a/tensorflow/core/grappler/optimizers/remapper.cc +++ b/tensorflow/core/grappler/optimizers/remapper.cc @@ -1667,7 +1667,6 @@ bool RequiresInferredShapes(const RemapperContext& ctx, int node_index) { return false; }; - #ifdef INTEL_MKL return is_batch_norm_candidate() || is_batch_norm_fusion_candidate() || IsConv2DWithAdd(ctx, node_index); @@ -1817,7 +1816,6 @@ Status Remapper::Optimize(Cluster* cluster, const GrapplerItem& item, } #endif // !INTEL_MKL - // Remap FusedBatchNorm++ into the _FusedBatchNormEx. FusedBatchNormEx fused_batch_norm_ex; if (allow_non_differentiable_rewrites && From dd417ca4477794c1dc0bc3df5b818b420417b4bf Mon Sep 17 00:00:00 2001 From: peng Date: Wed, 17 Jun 2020 11:28:25 +0800 Subject: [PATCH 102/867] delete empty lines --- tensorflow/core/grappler/optimizers/remapper.cc | 1 - 1 file changed, 1 deletion(-) diff --git a/tensorflow/core/grappler/optimizers/remapper.cc b/tensorflow/core/grappler/optimizers/remapper.cc index a6ddad8dfc8..b1158813345 100644 --- a/tensorflow/core/grappler/optimizers/remapper.cc +++ b/tensorflow/core/grappler/optimizers/remapper.cc @@ -1783,7 +1783,6 @@ Status Remapper::Optimize(Cluster* cluster, const GrapplerItem& item, // Remove this once TF-MKL supports _FusedConv2D with these operations. #ifndef INTEL_MKL // Remap Conv2D+Squeeze+BiasAdd into the _FusedConv2D+Squeeze. - ContractionWithSqueezeAndBiasAdd contract_with_squeeze_and_bias; if (allow_non_differentiable_rewrites && FindConv2DWithSqueezeAndBias(ctx, i, &contract_with_squeeze_and_bias)) { From eaf10e9a279e270d8133a41cfc9690de10423a1d Mon Sep 17 00:00:00 2001 From: Lutz Roeder Date: Sun, 14 Jun 2020 23:02:59 -0700 Subject: [PATCH 103/867] Fix Keras documentation --- tensorflow/python/keras/layers/core.py | 8 ++++++-- tensorflow/python/keras/layers/wrappers.py | 9 +++++---- 2 files changed, 11 insertions(+), 6 deletions(-) diff --git a/tensorflow/python/keras/layers/core.py b/tensorflow/python/keras/layers/core.py index abfb025db30..e64a1c27bcf 100644 --- a/tensorflow/python/keras/layers/core.py +++ b/tensorflow/python/keras/layers/core.py @@ -825,10 +825,14 @@ class Lambda(Layer): returned as output mask regardless of what the input is. arguments: Optional dictionary of keyword arguments to be passed to the function. - Input shape: Arbitrary. Use the keyword argument input_shape (tuple of + + Input shape: + Arbitrary. Use the keyword argument input_shape (tuple of integers, does not include the samples axis) when using this layer as the first layer in a model. - Output shape: Specified by `output_shape` argument + + Output shape: + Specified by `output_shape` argument """ @trackable.no_automatic_dependency_tracking diff --git a/tensorflow/python/keras/layers/wrappers.py b/tensorflow/python/keras/layers/wrappers.py index 8fe3b3b20bb..23fef467cfe 100644 --- a/tensorflow/python/keras/layers/wrappers.py +++ b/tensorflow/python/keras/layers/wrappers.py @@ -341,10 +341,11 @@ class Bidirectional(Wrapper): combined. One of {'sum', 'mul', 'concat', 'ave', None}. If None, the outputs will not be combined, they will be returned as a list. Default value is 'concat'. - backward_layer: Optional `keras.layers.RNN`, or keras.layers.Layer` instance - to be used to handle backwards input processing. If `backward_layer` is - not provided, the layer instance passed as the `layer` argument will be - used to generate the backward layer automatically. + backward_layer: Optional `keras.layers.RNN`, or `keras.layers.Layer` + instance to be used to handle backwards input processing. + If `backward_layer` is not provided, the layer instance passed as the + `layer` argument will be used to generate the backward layer + automatically. Note that the provided `backward_layer` layer should have properties matching those of the `layer` argument, in particular it should have the same values for `stateful`, `return_states`, `return_sequence`, etc. From 2aa2332364e4b27bb4fa433f0c6bd1709fab7f34 Mon Sep 17 00:00:00 2001 From: Nishidha Panpaliya Date: Wed, 17 Jun 2020 09:13:57 +0000 Subject: [PATCH 104/867] Fixed build error --- tensorflow/compiler/xla/service/cpu/BUILD | 2 +- tensorflow/compiler/xla/service/cpu/tests/BUILD | 10 +++++----- 2 files changed, 6 insertions(+), 6 deletions(-) diff --git a/tensorflow/compiler/xla/service/cpu/BUILD b/tensorflow/compiler/xla/service/cpu/BUILD index 7be4d3e724a..8ea3672d6a8 100644 --- a/tensorflow/compiler/xla/service/cpu/BUILD +++ b/tensorflow/compiler/xla/service/cpu/BUILD @@ -1080,7 +1080,7 @@ tf_cc_test( deps = [ ":cpu_compiler", ":cpu_transfer_manager", - ":test_header_helper", + ":test_header_helper", "//tensorflow/compiler/xla:test", "//tensorflow/compiler/xla/tests:hlo_test_base", "//tensorflow/compiler/xla/tests:xla_internal_test_main", diff --git a/tensorflow/compiler/xla/service/cpu/tests/BUILD b/tensorflow/compiler/xla/service/cpu/tests/BUILD index 9036c5c9024..510e7b18fba 100644 --- a/tensorflow/compiler/xla/service/cpu/tests/BUILD +++ b/tensorflow/compiler/xla/service/cpu/tests/BUILD @@ -42,7 +42,7 @@ tf_cc_test( "//tensorflow/compiler/xla/service:hlo", "//tensorflow/compiler/xla/service/cpu:cpu_compiler", "//tensorflow/compiler/xla/service/cpu/tests:cpu_codegen_test", - "//tensorflow/compiler/xla/service/cpu:test_header_helper", + "//tensorflow/compiler/xla/service/cpu:test_header_helper", "//tensorflow/core:lib", "//tensorflow/core:test", "//tensorflow/core:test_main", @@ -138,7 +138,7 @@ tf_cc_test( "//tensorflow/compiler/xla/service/cpu:cpu_compiler", "//tensorflow/compiler/xla/service/cpu/tests:cpu_codegen_test", "//tensorflow/compiler/xla/tests:test_utils", - "//tensorflow/compiler/xla/service/cpu:test_header_helper", + "//tensorflow/compiler/xla/service/cpu:test_header_helper", "//tensorflow/core:lib", "//tensorflow/core:test", "//tensorflow/core:test_main", @@ -218,7 +218,7 @@ tf_cc_test( "//tensorflow/compiler/xla/service:hlo_parser", "//tensorflow/compiler/xla/service/cpu:cpu_compiler", "//tensorflow/compiler/xla/service/cpu/tests:cpu_codegen_test", - "//tensorflow/compiler/xla/service/cpu:test_header_helper", + "//tensorflow/compiler/xla/service/cpu:test_header_helper", "//tensorflow/core:lib", "//tensorflow/core:test", "//tensorflow/core:test_main", @@ -232,7 +232,7 @@ tf_cc_test( "//tensorflow/compiler/xla/service:hlo", "//tensorflow/compiler/xla/service/cpu:cpu_compiler", "//tensorflow/compiler/xla/service/cpu/tests:cpu_codegen_test", - "//tensorflow/compiler/xla/service/cpu:test_header_helper", + "//tensorflow/compiler/xla/service/cpu:test_header_helper", "//tensorflow/core:lib", "//tensorflow/core:test", "//tensorflow/core:test_main", @@ -246,7 +246,7 @@ tf_cc_test( "//tensorflow/compiler/xla/service:hlo", "//tensorflow/compiler/xla/service/cpu:cpu_compiler", "//tensorflow/compiler/xla/service/cpu/tests:cpu_codegen_test", - "//tensorflow/compiler/xla/service/cpu:test_header_helper", + "//tensorflow/compiler/xla/service/cpu:test_header_helper", "//tensorflow/core:lib", "//tensorflow/core:test", "//tensorflow/core:test_main", From 13b840aa7299c77e744a2eed5a366cbf98988fb0 Mon Sep 17 00:00:00 2001 From: Agoniii <815244047@qq.com> Date: Wed, 17 Jun 2020 19:10:45 +0800 Subject: [PATCH 105/867] modify Log to gry --- .../optimizers/auto_mixed_precision_test.cc | 14 +++++++------- 1 file changed, 7 insertions(+), 7 deletions(-) diff --git a/tensorflow/core/grappler/optimizers/auto_mixed_precision_test.cc b/tensorflow/core/grappler/optimizers/auto_mixed_precision_test.cc index 951279d37cd..3fa8260739c 100644 --- a/tensorflow/core/grappler/optimizers/auto_mixed_precision_test.cc +++ b/tensorflow/core/grappler/optimizers/auto_mixed_precision_test.cc @@ -287,10 +287,10 @@ TEST_F(AutoMixedPrecisionTest, Simple) { Output clr2 = ops::Relu(s.WithOpName("clr2"), gry1); Output wht1 = ops::MatMul(s.WithOpName("wht1"), clr2, clr2); Output clr3 = ops::Relu(s.WithOpName("clr3"), wht1); - Output blk2 = ops::Log(s.WithOpName("blk2"), clr3); - Output clr4 = ops::Relu(s.WithOpName("clr4"), blk2); - Output blk3 = ops::SparseMatMul(s.WithOpName("blk3"), clr4, clr4); - Output clr5 = ops::Relu(s.WithOpName("clr5"), blk3); + Output gry2 = ops::Log(s.WithOpName("gry2"), clr3); + Output clr4 = ops::Relu(s.WithOpName("clr4"), gry2); + Output blk2 = ops::SparseMatMul(s.WithOpName("blk2"), clr4, clr4); + Output clr5 = ops::Relu(s.WithOpName("clr5"), blk2); Output fetch = ops::Identity(s.WithOpName("fetch"), clr5); GrapplerItem item; @@ -313,10 +313,10 @@ TEST_F(AutoMixedPrecisionTest, Simple) { EXPECT_EQ(output_view.GetNode("clr2")->attr().at("T").type(), DT_HALF); EXPECT_EQ(output_view.GetNode("wht1")->attr().at("T").type(), DT_HALF); EXPECT_EQ(output_view.GetNode("clr3")->attr().at("T").type(), DT_HALF); - EXPECT_EQ(output_view.GetNode("blk2")->attr().at("T").type(), DT_FLOAT); + EXPECT_EQ(output_view.GetNode("gry2")->attr().at("T").type(), DT_FLOAT); EXPECT_EQ(output_view.GetNode("clr4")->attr().at("T").type(), DT_FLOAT); - EXPECT_EQ(output_view.GetNode("blk3")->attr().at("Ta").type(), DT_FLOAT); - EXPECT_EQ(output_view.GetNode("blk3")->attr().at("Tb").type(), DT_FLOAT); + EXPECT_EQ(output_view.GetNode("blk2")->attr().at("Ta").type(), DT_FLOAT); + EXPECT_EQ(output_view.GetNode("blk2")->attr().at("Tb").type(), DT_FLOAT); EXPECT_EQ(output_view.GetNode("clr5")->attr().at("T").type(), DT_FLOAT); auto tensors = EvaluateNodes(output, item.fetch); From 286cd7fc6839bb2fc999fd16fb1801f6b30656b8 Mon Sep 17 00:00:00 2001 From: Elena Zhelezina Date: Wed, 17 Jun 2020 12:22:53 +0100 Subject: [PATCH 106/867] Addressed reviewer's comments. Change-Id: I4b849e60540879ca89483ede675c63631bc9417b --- tensorflow/lite/python/lite.py | 25 +++++++------------------ 1 file changed, 7 insertions(+), 18 deletions(-) diff --git a/tensorflow/lite/python/lite.py b/tensorflow/lite/python/lite.py index 06796ba820b..cb2f1853619 100644 --- a/tensorflow/lite/python/lite.py +++ b/tensorflow/lite/python/lite.py @@ -299,32 +299,21 @@ class QuantizationMode(object): inference_input_type = input_ty if input_ty else constants.FLOAT inference_output_type = output_ty if output_ty else constants.FLOAT - if self.post_training_int8_no_float(): + + if self.post_training_int8_no_float() \ + or self.post_training_int16x8_no_float(): return True, { "inference_input_type": inference_input_type, "inference_output_type": inference_output_type, - "activations_type": constants.INT8, + "activations_type": self.activations_type(), "allow_float": False } - elif self.post_training_int8_allow_float(): + elif self.post_training_int8_allow_float() \ + or self.post_training_int16x8_allow_float(): return True, { "inference_input_type": inference_input_type, "inference_output_type": inference_output_type, - "activations_type": constants.INT8, - "allow_float": True - } - elif self.post_training_int16x8_no_float(): - return True, { - "inference_input_type": inference_input_type, - "inference_output_type": inference_output_type, - "activations_type": constants.INT16, - "allow_float": False - } - elif self.post_training_int16x8_allow_float(): - return True, { - "inference_input_type": inference_input_type, - "inference_output_type": inference_output_type, - "activations_type": constants.INT16, + "activations_type": self.activations_type(), "allow_float": True } else: From 3ebc53c394e34dfd780e6bf59ad3c96bd9b3fa79 Mon Sep 17 00:00:00 2001 From: Nathan Luehr Date: Thu, 11 Jun 2020 15:59:21 -0500 Subject: [PATCH 107/867] Relax stub include version checking. Remove upper bound on version check for latest inc files. --- tensorflow/stream_executor/cuda/cublas_stub.cc | 10 ++++------ tensorflow/stream_executor/cuda/cuda_stub.cc | 10 ++++------ tensorflow/stream_executor/cuda/cudart_stub.cc | 10 ++++------ tensorflow/stream_executor/cuda/cusolver_stub.cc | 10 ++++------ tensorflow/stream_executor/cuda/cusparse_stub.cc | 10 ++++------ 5 files changed, 20 insertions(+), 30 deletions(-) diff --git a/tensorflow/stream_executor/cuda/cublas_stub.cc b/tensorflow/stream_executor/cuda/cublas_stub.cc index 1cbfd51316c..76f3d9b134e 100644 --- a/tensorflow/stream_executor/cuda/cublas_stub.cc +++ b/tensorflow/stream_executor/cuda/cublas_stub.cc @@ -63,14 +63,12 @@ typedef enum {} cublasMath_t; #if CUDA_VERSION < 10000 #include "tensorflow/stream_executor/cuda/cublas_9_0.inc" -#elif CUDA_VERSION == 10000 +#elif CUDA_VERSION < 10010 #include "tensorflow/stream_executor/cuda/cublas_10_0.inc" -#elif CUDA_VERSION == 10010 +#elif CUDA_VERSION < 10020 #include "tensorflow/stream_executor/cuda/cublas_10_1.inc" -#elif CUDA_VERSION == 10020 +#elif CUDA_VERSION < 11000 #include "tensorflow/stream_executor/cuda/cublas_10_2.inc" -#elif CUBLAS_VER_MAJOR == 11 && CUBLAS_VER_MINOR == 0 -#include "tensorflow/stream_executor/cuda/cublas_11_0.inc" #else -#error "We have no wrapper for this version." +#include "tensorflow/stream_executor/cuda/cublas_11_0.inc" #endif diff --git a/tensorflow/stream_executor/cuda/cuda_stub.cc b/tensorflow/stream_executor/cuda/cuda_stub.cc index ce02be89c22..58c898a54ee 100644 --- a/tensorflow/stream_executor/cuda/cuda_stub.cc +++ b/tensorflow/stream_executor/cuda/cuda_stub.cc @@ -95,14 +95,12 @@ typedef void(CUDA_CB* CUhostFn)(void* userData); #if CUDA_VERSION < 10000 #include "tensorflow/stream_executor/cuda/cuda_9_0.inc" -#elif CUDA_VERSION == 10000 +#elif CUDA_VERSION < 10010 #include "tensorflow/stream_executor/cuda/cuda_10_0.inc" -#elif CUDA_VERSION <= 10010 +#elif CUDA_VERSION < 10020 #include "tensorflow/stream_executor/cuda/cuda_10_1.inc" -#elif CUDA_VERSION <= 10020 +#elif CUDA_VERSION < 11000 #include "tensorflow/stream_executor/cuda/cuda_10_2.inc" -#elif CUDA_VERSION <= 11000 -#include "tensorflow/stream_executor/cuda/cuda_11_0.inc" #else -#error "We have no wrapper for this version." +#include "tensorflow/stream_executor/cuda/cuda_11_0.inc" #endif diff --git a/tensorflow/stream_executor/cuda/cudart_stub.cc b/tensorflow/stream_executor/cuda/cudart_stub.cc index 3b9e0f2937b..2ab9d142e3c 100644 --- a/tensorflow/stream_executor/cuda/cudart_stub.cc +++ b/tensorflow/stream_executor/cuda/cudart_stub.cc @@ -53,16 +53,14 @@ cudaError_t GetSymbolNotFoundError() { // A bunch of new symbols were introduced in version 10 #if CUDART_VERSION < 10000 #include "tensorflow/stream_executor/cuda/cuda_runtime_9_0.inc" -#elif CUDART_VERSION == 10000 +#elif CUDART_VERSION < 10010 #include "tensorflow/stream_executor/cuda/cuda_runtime_10_0.inc" -#elif CUDART_VERSION == 10010 +#elif CUDART_VERSION < 10020 #include "tensorflow/stream_executor/cuda/cuda_runtime_10_1.inc" -#elif CUDART_VERSION == 10020 +#elif CUDART_VERSION < 11000 #include "tensorflow/stream_executor/cuda/cuda_runtime_10_2.inc" -#elif CUDART_VERSION == 11000 -#include "tensorflow/stream_executor/cuda/cuda_runtime_11_0.inc" #else -#error "We have no wrapper for this version." +#include "tensorflow/stream_executor/cuda/cuda_runtime_11_0.inc" #endif #undef __dv #undef __CUDA_DEPRECATED diff --git a/tensorflow/stream_executor/cuda/cusolver_stub.cc b/tensorflow/stream_executor/cuda/cusolver_stub.cc index a4b9cc37f9b..edf87c3dc0b 100644 --- a/tensorflow/stream_executor/cuda/cusolver_stub.cc +++ b/tensorflow/stream_executor/cuda/cusolver_stub.cc @@ -53,14 +53,12 @@ cusolverStatus_t GetSymbolNotFoundError() { #if CUDA_VERSION < 10000 #include "tensorflow/stream_executor/cuda/cusolver_dense_9_0.inc" -#elif CUDA_VERSION == 10000 +#elif CUDA_VERSION < 10010 #include "tensorflow/stream_executor/cuda/cusolver_dense_10_0.inc" -#elif CUDA_VERSION == 10010 +#elif CUDA_VERSION < 10020 #include "tensorflow/stream_executor/cuda/cusolver_dense_10_1.inc" -#elif CUDA_VERSION == 10020 +#elif CUDA_VERSION < 11000 #include "tensorflow/stream_executor/cuda/cusolver_dense_10_2.inc" -#elif CUDA_VERSION == 11000 -#include "tensorflow/stream_executor/cuda/cusolver_dense_11_0.inc" #else -#error "We don't have a wrapper for this version." +#include "tensorflow/stream_executor/cuda/cusolver_dense_11_0.inc" #endif diff --git a/tensorflow/stream_executor/cuda/cusparse_stub.cc b/tensorflow/stream_executor/cuda/cusparse_stub.cc index ae56402fbc3..caed4d1008e 100644 --- a/tensorflow/stream_executor/cuda/cusparse_stub.cc +++ b/tensorflow/stream_executor/cuda/cusparse_stub.cc @@ -53,14 +53,12 @@ cusparseStatus_t GetSymbolNotFoundError() { #if CUDA_VERSION < 10000 #include "tensorflow/stream_executor/cuda/cusparse_9_0.inc" -#elif CUDA_VERSION == 10000 +#elif CUDA_VERSION < 10010 #include "tensorflow/stream_executor/cuda/cusparse_10_0.inc" -#elif CUDA_VERSION == 10010 +#elif CUDA_VERSION < 10020 #include "tensorflow/stream_executor/cuda/cusparse_10_1.inc" -#elif CUDA_VERSION == 10020 +#elif CUDA_VERSION < 11000 #include "tensorflow/stream_executor/cuda/cusparse_10_2.inc" -#elif CUSPARSE_VER_MAJOR == 11 && CUSPARSE_VER_MINOR == 0 -#include "tensorflow/stream_executor/cuda/cusparse_11_0.inc" #else -#error "We don't have a wrapper for this version." +#include "tensorflow/stream_executor/cuda/cusparse_11_0.inc" #endif From ef55a40b374d7310e4ce3149d86395d403403d0d Mon Sep 17 00:00:00 2001 From: Yong Tang Date: Wed, 17 Jun 2020 18:38:24 +0000 Subject: [PATCH 108/867] Move tf.equal shape inference test to math_ops_test.cc also added additional shape inference test cases Signed-off-by: Yong Tang --- tensorflow/core/ops/math_ops_test.cc | 18 ++++++++++++++++++ .../python/autograph/operators/logical_test.py | 14 -------------- 2 files changed, 18 insertions(+), 14 deletions(-) diff --git a/tensorflow/core/ops/math_ops_test.cc b/tensorflow/core/ops/math_ops_test.cc index a2837d88bde..2b65f88042c 100644 --- a/tensorflow/core/ops/math_ops_test.cc +++ b/tensorflow/core/ops/math_ops_test.cc @@ -604,4 +604,22 @@ TEST(MathOpsTest, SobolSample) { INFER_OK(op, "[];[];[]", "[?,?]"); } + +TEST(MathOpsTest, EqualOp) { + ShapeInferenceTestOp op("Equal"); + AddNodeAttr("incompatible_shape_error", true, &op.node_def); + + INFER_OK(op, "?;?", "?"); + INFER_OK(op, "[1,2];?", "?"); + INFER_OK(op, "?;[1,2]", "?"); + + INFER_OK(op, "[1,2,3];[1]", "[d0_0,d0_1,d0_2]"); + INFER_OK(op, "[?,2,1];[1,3]", "[d0_0,d0_1,d1_1]"); + INFER_OK(op, "[1,?,3];[3,1]", "[d0_0,d1_0,d0_2]"); + INFER_OK(op, "[1,2,3];[2,1,3]", "[d1_0,d0_1,d0_2]"); + + // Note: Test case for GitHub issue 40471 + INFER_OK(op, "[?,10,1];[?,1,4]", "[?,d0_1,d1_2]"); + INFER_OK(op, "[10,?,1];[1,?,4]", "[d0_0,?,d1_2]"); +} } // end namespace tensorflow diff --git a/tensorflow/python/autograph/operators/logical_test.py b/tensorflow/python/autograph/operators/logical_test.py index 0eab302a825..e22f39932d1 100644 --- a/tensorflow/python/autograph/operators/logical_test.py +++ b/tensorflow/python/autograph/operators/logical_test.py @@ -19,9 +19,7 @@ from __future__ import division from __future__ import print_function from tensorflow.python.autograph.operators import logical -from tensorflow.python.eager import def_function from tensorflow.python.framework import constant_op -from tensorflow.python.framework import tensor_spec from tensorflow.python.framework import test_util from tensorflow.python.platform import test @@ -85,18 +83,6 @@ class LogicalOperatorsTest(test.TestCase): t = logical.not_(self._tf_false()) self.assertEqual(self.evaluate(t), True) - # Test case for GitHub issue 40471 - def test_equal_output_shapes(self): - - @def_function.function(input_signature=[ - tensor_spec.TensorSpec([None, 10, 1]), - tensor_spec.TensorSpec([None, 1, 4])]) - def f(x, y): - z = x == y - return z - - self.assertAllEqual(f.get_concrete_function().output_shapes, [None, 10, 4]) - if __name__ == '__main__': test.main() From cb61c42938010f3ff33b55ef4b41651c22fb7652 Mon Sep 17 00:00:00 2001 From: Matt Conley Date: Mon, 15 Jun 2020 11:22:30 -0700 Subject: [PATCH 109/867] [TFTRT] Add Dynamic Shape Tests for ConvertSquare Co-authored-by: Tamas Feher - Modify ConvertSquare tests to use newer TFTRT testing API - Add INT32 as a supported dtype for TFTRT ConvertSquare --- .../tf2tensorrt/convert/convert_nodes.cc | 5 ++ .../tf2tensorrt/convert/convert_nodes_test.cc | 64 +++++++------------ 2 files changed, 28 insertions(+), 41 deletions(-) diff --git a/tensorflow/compiler/tf2tensorrt/convert/convert_nodes.cc b/tensorflow/compiler/tf2tensorrt/convert/convert_nodes.cc index 20ee5ffd8f8..28b27959afc 100644 --- a/tensorflow/compiler/tf2tensorrt/convert/convert_nodes.cc +++ b/tensorflow/compiler/tf2tensorrt/convert/convert_nodes.cc @@ -4424,8 +4424,13 @@ Status ConvertSquare(OpConverterParams* params) { const auto& inputs = params->inputs; const auto& node_def = params->node_def; TF_RETURN_IF_ERROR(CheckInputsWeights(*params, {{"x", false}})); +#if IS_TRT_VERSION_GE(6, 0, 1, 0) + TF_RETURN_IF_ERROR(AllowDataTypes( + *params, {DataType::DT_FLOAT, DataType::DT_HALF, DataType::DT_INT32})); +#else TF_RETURN_IF_ERROR( AllowDataTypes(*params, {DataType::DT_FLOAT, DataType::DT_HALF})); +#endif if (params->validation_only) return Status::OK(); // Constant 2 with same rank as input diff --git a/tensorflow/compiler/tf2tensorrt/convert/convert_nodes_test.cc b/tensorflow/compiler/tf2tensorrt/convert/convert_nodes_test.cc index 450831910f6..1192b563e57 100644 --- a/tensorflow/compiler/tf2tensorrt/convert/convert_nodes_test.cc +++ b/tensorflow/compiler/tf2tensorrt/convert/convert_nodes_test.cc @@ -2754,58 +2754,40 @@ TEST_F(OpConverterTest, ConvertQuantize) { } } -template -void TestConvertSquare(OpConverterTest* test) { - test->Reset(); - typedef typename EnumToDataType::Type CType; - - Scope s = Scope::NewRootScope(); - auto input = ops::Placeholder(s.WithOpName("input"), dtype); - auto square = ops::Square(s.WithOpName("my_square"), input); - NodeDef node_def = square.operation.node()->def(); - - test->AddTestTensor("input", {1, 20}, /*batch_size=*/1, - TfDataTypeToTrt(dtype)); - test->RunValidationAndConversion(node_def); - TRT_TensorOrWeights output; - TF_EXPECT_OK(test->GetTensorOrWeights("my_square", &output)); - ASSERT_TRUE(output.is_tensor()); - ExpectTrtDimsEqualsArray({1, 20}, output.tensor()->getDimensions()); - - const int num_inputs = 20; - std::vector inputs(num_inputs); - std::vector expected_outputs(num_inputs); - for (int i = 0; i < num_inputs; ++i) { - const CType value = CType(i - 9); - inputs[i] = value; - expected_outputs[i] = value * value; - } - const DataVec input_data{{"input", test->AsTensor(inputs)}}; - // Engine outputs are converted to FP16 automatically if we set FP16 mode in - // the builder. - DataVec output_data{{"my_square", test->ConstructTensor(num_inputs)}}; - TF_EXPECT_OK(test->BuildAndRun(input_data, &output_data)); - ExpectArrayNear(expected_outputs, GetSpanForData(output_data[0])); -} - -TEST_F(OpConverterTest, ConvertSquare) { +TEST_P(OpConverterTest2, ConvertSquare) { { // Input is weights, should fail. Reset(); Scope s = Scope::NewRootScope(); - auto input = ops::Placeholder(s.WithOpName("input"), DT_FLOAT); + auto input = ops::Placeholder(s.WithOpName("input"), tf_type); auto square = ops::Square(s.WithOpName("my_square"), input); NodeDef node_def = square.operation.node()->def(); - AddTestWeights("input", {1, 2, 3}, {1, 2, 3, 4, -5, 6}); + AddTestWeights("input", {1, 2, 3}, {1, 2, 3, 4, -5, 6}, tf_type); RunValidationAndConversion( node_def, error::UNIMPLEMENTED, "The input \"x\" for Square must be a tensor, at my_square"); } - // OK. Note that kINT32 is not supported by IElementWiseLayer, so we don't - // test DT_INT32 type here. - TestConvertSquare(this); - TestConvertSquare(this); + Reset(); + + Scope s = Scope::NewRootScope(); + auto input = ops::Placeholder(s.WithOpName("input"), tf_type); + auto square = ops::Square(s.WithOpName("my_square"), input); + NodeDef node_def = square.operation.node()->def(); + + const int num_inputs = 20; + std::vector inputs(num_inputs); + std::vector expected_outputs(num_inputs); + + for (int i = 0; i < num_inputs; ++i) { + const float value = (i - 9); + inputs[i] = value; + expected_outputs[i] = value * value; + } + AddTestTensor("input", {1, 1, 20}, tf_type, inputs); + + TestOpConverter("my_square", node_def, {1, 1, 20}, Status::OK(), Status::OK(), + ArrayFloatNear(expected_outputs, 0)); } #if IS_TRT_VERSION_GE(5, 1, 0, 0) From d5ca984c5314d2e683eca87539d03ee4e35e3ee6 Mon Sep 17 00:00:00 2001 From: "A. Unique TensorFlower" Date: Wed, 17 Jun 2020 13:05:59 -0700 Subject: [PATCH 110/867] Allow the compiler to vectorize the loop. MIME-Version: 1.0 Content-Type: text/plain; charset=UTF-8 Content-Transfer-Encoding: 8bit ``` name old time/op new time/op delta BM_Conv2DBackpropInput_fp32_NHWC_cpu_in8x112x112x64_f2x2x64_s2x2_SAME 18.6ms ± 5% 18.5ms ±13% ~ (p=0.912 n=10+10) BM_Conv2DBackpropInput_fp32_NHWC_cpu_in8x56x56x128_f2x2x128_s2x2_SAME 12.7ms ±12% 12.7ms ±17% ~ (p=0.684 n=10+10) BM_Conv2DBackpropInput_fp32_NHWC_cpu_in8x28x28x256_f2x2x256_s2x2_SAME 12.2ms ± 8% 11.2ms ± 4% -8.21% (p=0.001 n=10+8) BM_Conv2DBackpropInput_fp32_NHWC_cpu_in8x112x112x64_f2x2x64_s2x2_VALID 18.7ms ±20% 18.6ms ±23% ~ (p=0.278 n=9+10) BM_Conv2DBackpropInput_fp32_NHWC_cpu_in8x56x56x128_f2x2x128_s2x2_VALID 12.5ms ±15% 11.4ms ± 2% -8.98% (p=0.000 n=10+9) BM_Conv2DBackpropInput_fp32_NHWC_cpu_in8x28x28x256_f2x2x256_s2x2_VALID 11.6ms ± 8% 11.1ms ± 2% -4.22% (p=0.011 n=9+9) BM_Conv2DBackpropInput_fp32_NHWC_cpu_in8x56x56x64_f1x1x64_s1x1_SAME 4.57ms ± 3% 4.34ms ± 1% -5.04% (p=0.000 n=8+9) BM_Conv2DBackpropInput_fp32_NHWC_cpu_in8x56x56x64_f1x1x256_s1x1_SAME 12.0ms ± 4% 11.5ms ± 2% -4.32% (p=0.000 n=8+9) BM_Conv2DBackpropInput_fp32_NHWC_cpu_in8x56x56x256_f1x1x64_s1x1_SAME 20.0ms ±31% 20.6ms ±17% ~ (p=0.912 n=10+10) BM_Conv2DBackpropInput_fp32_NHWC_cpu_in8x56x56x64_f3x3x64_s1x1_SAME 36.5ms ±21% 32.0ms ± 1% -12.30% (p=0.000 n=10+9) BM_Conv2DBackpropInput_fp32_NHWC_cpu_in8x28x28x128_f1x1x128_s1x1_SAME 3.71ms ±17% 3.33ms ± 1% -10.47% (p=0.000 n=10+9) BM_Conv2DBackpropInput_fp32_NHWC_cpu_in8x28x28x128_f1x1x512_s1x1_SAME 11.8ms ±16% 10.5ms ± 1% -11.37% (p=0.000 n=10+8) BM_Conv2DBackpropInput_fp32_NHWC_cpu_in8x28x28x512_f1x1x128_s1x1_SAME 13.1ms ±13% 11.4ms ± 2% -13.36% (p=0.000 n=10+8) BM_Conv2DBackpropInput_fp32_NHWC_cpu_in8x28x28x512_f3x3x128_s1x1_SAME 142ms ±12% 124ms ± 1% -13.22% (p=0.000 n=10+8) BM_Conv2DBackpropInput_fp32_NHWC_cpu_in8x14x14x256_f1x1x256_s1x1_SAME 3.51ms ±14% 3.18ms ±20% -9.43% (p=0.009 n=10+10) BM_Conv2DBackpropInput_fp32_NHWC_cpu_in8x14x14x256_f1x1x1024_s1x1_SAME 14.0ms ±18% 12.0ms ± 4% -13.80% (p=0.012 n=10+8) BM_Conv2DBackpropInput_fp32_NHWC_cpu_in8x14x14x1024_f1x1x256_s1x1_SAME 12.8ms ±18% 11.1ms ± 2% -13.57% (p=0.001 n=10+8) BM_Conv2DBackpropInput_fp32_NHWC_cpu_in8x14x14x256_f3x3x256_s1x1_SAME 23.0ms ±18% 19.9ms ± 4% -13.38% (p=0.004 n=10+8) BM_Conv2DBackpropInput_fp32_NHWC_cpu_in16x112x112x64_f2x2x64_s2x2_SAME 45.3ms ± 9% 40.5ms ± 4% -10.74% (p=0.000 n=10+9) BM_Conv2DBackpropInput_fp32_NHWC_cpu_in16x56x56x128_f2x2x128_s2x2_SAME 33.2ms ±13% 28.8ms ± 2% -13.11% (p=0.001 n=10+8) BM_Conv2DBackpropInput_fp32_NHWC_cpu_in16x28x28x256_f2x2x256_s2x2_SAME 31.5ms ±15% 26.7ms ± 2% -15.13% (p=0.000 n=10+8) BM_Conv2DBackpropInput_fp32_NHWC_cpu_in16x112x112x64_f2x2x64_s2x2_VALID 45.5ms ± 8% 41.3ms ± 9% -9.31% (p=0.000 n=10+10) BM_Conv2DBackpropInput_fp32_NHWC_cpu_in16x56x56x128_f2x2x128_s2x2_VALID 33.2ms ±12% 28.8ms ± 2% -13.38% (p=0.000 n=10+9) BM_Conv2DBackpropInput_fp32_NHWC_cpu_in16x28x28x256_f2x2x256_s2x2_VALID 31.6ms ±14% 26.7ms ± 1% -15.53% (p=0.000 n=10+8) BM_Conv2DBackpropInput_fp32_NHWC_cpu_in16x56x56x64_f1x1x64_s1x1_SAME 11.1ms ±15% 9.4ms ± 3% -15.29% (p=0.000 n=10+10) BM_Conv2DBackpropInput_fp32_NHWC_cpu_in16x56x56x64_f1x1x256_s1x1_SAME 27.3ms ±13% 23.5ms ± 2% -13.97% (p=0.000 n=10+10) BM_Conv2DBackpropInput_fp32_NHWC_cpu_in16x56x56x256_f1x1x64_s1x1_SAME 46.2ms ± 8% 40.9ms ± 5% -11.54% (p=0.000 n=10+10) BM_Conv2DBackpropInput_fp32_NHWC_cpu_in16x56x56x64_f3x3x64_s1x1_SAME 81.5ms ±17% 64.2ms ± 1% -21.17% (p=0.000 n=10+10) BM_Conv2DBackpropInput_fp32_NHWC_cpu_in16x28x28x128_f1x1x128_s1x1_SAME 8.30ms ±16% 6.80ms ± 1% -18.02% (p=0.000 n=10+10) BM_Conv2DBackpropInput_fp32_NHWC_cpu_in16x28x28x128_f1x1x512_s1x1_SAME 27.7ms ±14% 23.3ms ± 2% -15.87% (p=0.000 n=10+10) BM_Conv2DBackpropInput_fp32_NHWC_cpu_in16x28x28x512_f1x1x128_s1x1_SAME 34.9ms ±14% 28.8ms ± 2% -17.41% (p=0.000 n=10+10) BM_Conv2DBackpropInput_fp32_NHWC_cpu_in16x28x28x512_f3x3x128_s1x1_SAME 300ms ±12% 250ms ± 1% -16.80% (p=0.000 n=10+9) BM_Conv2DBackpropInput_fp32_NHWC_cpu_in16x14x14x256_f1x1x256_s1x1_SAME 7.91ms ±15% 7.17ms ±21% -9.36% (p=0.029 n=10+10) BM_Conv2DBackpropInput_fp32_NHWC_cpu_in16x14x14x256_f1x1x1024_s1x1_SAME 35.6ms ±15% 33.4ms ±13% -6.21% (p=0.023 n=10+10) BM_Conv2DBackpropInput_fp32_NHWC_cpu_in16x14x14x1024_f1x1x256_s1x1_SAME 28.0ms ± 3% 26.7ms ± 1% -4.81% (p=0.000 n=8+8) BM_Conv2DBackpropInput_fp32_NHWC_cpu_in16x14x14x256_f3x3x256_s1x1_SAME 41.4ms ± 3% 39.6ms ± 4% -4.43% (p=0.001 n=9+9) BM_Conv2DBackpropInput_fp32_NHWC_cpu_in32x112x112x64_f2x2x64_s2x2_SAME 90.0ms ± 3% 85.5ms ±11% -5.00% (p=0.010 n=9+10) BM_Conv2DBackpropInput_fp32_NHWC_cpu_in32x56x56x128_f2x2x128_s2x2_SAME 69.0ms ± 1% 65.8ms ± 3% -4.68% (p=0.000 n=9+9) BM_Conv2DBackpropInput_fp32_NHWC_cpu_in32x28x28x256_f2x2x256_s2x2_SAME 57.8ms ± 2% 55.8ms ± 2% -3.54% (p=0.000 n=9+9) BM_Conv2DBackpropInput_fp32_NHWC_cpu_in32x112x112x64_f2x2x64_s2x2_VALID 89.3ms ± 2% 86.1ms ± 5% -3.57% (p=0.006 n=9+10) BM_Conv2DBackpropInput_fp32_NHWC_cpu_in32x56x56x128_f2x2x128_s2x2_VALID 68.7ms ± 2% 65.7ms ± 3% -4.31% (p=0.001 n=8+9) BM_Conv2DBackpropInput_fp32_NHWC_cpu_in32x28x28x256_f2x2x256_s2x2_VALID 57.1ms ± 1% 55.6ms ± 1% -2.58% (p=0.000 n=8+10) BM_Conv2DBackpropInput_fp32_NHWC_cpu_in32x56x56x64_f1x1x64_s1x1_SAME 23.6ms ± 9% 21.9ms ± 1% -7.27% (p=0.000 n=10+9) BM_Conv2DBackpropInput_fp32_NHWC_cpu_in32x56x56x64_f1x1x256_s1x1_SAME 51.0ms ± 3% 48.4ms ± 5% -5.13% (p=0.001 n=8+10) BM_Conv2DBackpropInput_fp32_NHWC_cpu_in32x56x56x256_f1x1x64_s1x1_SAME 90.2ms ± 3% 85.4ms ± 5% -5.29% (p=0.000 n=10+9) BM_Conv2DBackpropInput_fp32_NHWC_cpu_in32x56x56x64_f3x3x64_s1x1_SAME 143ms ± 7% 133ms ± 3% -6.48% (p=0.000 n=9+9) BM_Conv2DBackpropInput_fp32_NHWC_cpu_in32x28x28x128_f1x1x128_s1x1_SAME 14.5ms ± 4% 14.0ms ± 2% -3.82% (p=0.000 n=10+8) BM_Conv2DBackpropInput_fp32_NHWC_cpu_in32x28x28x128_f1x1x512_s1x1_SAME 53.8ms ± 3% 51.5ms ± 2% -4.43% (p=0.000 n=10+9) BM_Conv2DBackpropInput_fp32_NHWC_cpu_in32x28x28x512_f1x1x128_s1x1_SAME 69.8ms ± 5% 66.9ms ± 8% -4.15% (p=0.010 n=10+9) BM_Conv2DBackpropInput_fp32_NHWC_cpu_in32x28x28x512_f3x3x128_s1x1_SAME 533ms ± 3% 508ms ± 0% -4.65% (p=0.000 n=10+8) BM_Conv2DBackpropInput_fp32_NHWC_cpu_in32x14x14x256_f1x1x256_s1x1_SAME 14.0ms ± 6% 14.9ms ±21% ~ (p=0.481 n=10+10) BM_Conv2DBackpropInput_fp32_NHWC_cpu_in32x14x14x256_f1x1x1024_s1x1_SAME 67.2ms ± 3% 71.0ms ±22% ~ (p=0.278 n=9+10) BM_Conv2DBackpropInput_fp32_NHWC_cpu_in32x14x14x1024_f1x1x256_s1x1_SAME 57.5ms ± 4% 56.6ms ± 5% ~ (p=0.055 n=10+8) BM_Conv2DBackpropInput_fp32_NHWC_cpu_in32x14x14x256_f3x3x256_s1x1_SAME 78.1ms ± 5% 75.5ms ± 3% -3.32% (p=0.011 n=9+8) ``` PiperOrigin-RevId: 316949219 Change-Id: I1a1fb517a5c28d489da9762b650577b61bf4e0de --- tensorflow/core/kernels/conv_grad_input_ops.cc | 3 +-- 1 file changed, 1 insertion(+), 2 deletions(-) diff --git a/tensorflow/core/kernels/conv_grad_input_ops.cc b/tensorflow/core/kernels/conv_grad_input_ops.cc index fd2f569a8b8..2dd63d1f4d0 100644 --- a/tensorflow/core/kernels/conv_grad_input_ops.cc +++ b/tensorflow/core/kernels/conv_grad_input_ops.cc @@ -76,7 +76,7 @@ template void Col2im(const T* col_data, const int depth, const int height, const int width, const int filter_h, const int filter_w, const int pad_t, const int pad_l, const int pad_b, const int pad_r, - const int stride_h, const int stride_w, T* im_data) { + const int stride_h, const int stride_w, T* __restrict im_data) { int height_col = (height + pad_t + pad_b - filter_h) / stride_h + 1; int width_col = (width + pad_l + pad_r - filter_w) / stride_w + 1; int h_pad = -pad_t; @@ -87,7 +87,6 @@ void Col2im(const T* col_data, const int depth, const int height, for (int ih = h_pad; ih < h_pad + filter_h; ++ih) { for (int iw = w_pad; iw < w_pad + filter_w; ++iw) { if (ih >= 0 && ih < height && iw >= 0 && iw < width) { - // TODO(andydavis) Vectorize this loop (if compiler does not). for (int i = 0; i < depth; ++i) { im_patch_data[i] += col_data[i]; } From 9426d35abc2ea3a5e790a928641749c5783e4b65 Mon Sep 17 00:00:00 2001 From: Lukasz Kaiser Date: Wed, 17 Jun 2020 13:07:01 -0700 Subject: [PATCH 111/867] Add the vdot op to tf-numpy. PiperOrigin-RevId: 316949446 Change-Id: I9b0718ec108486096e032729481ec9129863b429 --- tensorflow/python/ops/numpy_ops/np_math_ops.py | 10 ++++++++++ tensorflow/python/ops/numpy_ops/np_math_ops_test.py | 6 ++++++ 2 files changed, 16 insertions(+) diff --git a/tensorflow/python/ops/numpy_ops/np_math_ops.py b/tensorflow/python/ops/numpy_ops/np_math_ops.py index b32f78bee5a..abfd9087ffd 100644 --- a/tensorflow/python/ops/numpy_ops/np_math_ops.py +++ b/tensorflow/python/ops/numpy_ops/np_math_ops.py @@ -309,6 +309,16 @@ def cross(a, b, axisa=-1, axisb=-1, axisc=-1, axis=None): # pylint: disable=mis return _bin_op(f, a, b) +@np_utils.np_doc_only(np.vdot) +def vdot(a, b): # pylint: disable=missing-docstring + a, b = np_array_ops._promote_dtype(a, b) + a = np_array_ops.reshape(a, [-1]) + b = np_array_ops.reshape(b, [-1]) + if a.dtype == np_dtypes.complex128 or a.dtype == np_dtypes.complex64: + a = conj(a) + return dot(a, b) + + @np_utils.np_doc(np.power) def power(x1, x2): return _bin_op(math_ops.pow, x1, x2) diff --git a/tensorflow/python/ops/numpy_ops/np_math_ops_test.py b/tensorflow/python/ops/numpy_ops/np_math_ops_test.py index a3dfbb6c871..cb5326bcded 100644 --- a/tensorflow/python/ops/numpy_ops/np_math_ops_test.py +++ b/tensorflow/python/ops/numpy_ops/np_math_ops_test.py @@ -124,6 +124,12 @@ class MathTest(test.TestCase, parameterized.TestCase): np_math_ops.matmul( np_array_ops.ones([2, 3], np.int32), np_array_ops.ones([], np.int32)) + def testVDot(self): + operands = [([[1, 2], [3, 4]], [[3, 4], [6, 7]]), + ([[1, 2], [3, 4]], [3, 4, 6, 7])] + return self._testBinaryOp( + np_math_ops.vdot, np.vdot, 'vdot', operands=operands) + def _testUnaryOp(self, math_fun, np_fun, name): def run_test(a): From cca9b615b264651f5c70d604bb29146ddfbadede Mon Sep 17 00:00:00 2001 From: Scott Zhu Date: Wed, 17 Jun 2020 13:12:33 -0700 Subject: [PATCH 112/867] [tfdbg2] Fork local_cli_wrapper_test for keras related tests. PiperOrigin-RevId: 316950499 Change-Id: I428273592694426f72427e6236c68bdfb4e95eba --- tensorflow/python/debug/BUILD | 2 - .../debug/wrappers/local_cli_wrapper_test.py | 39 +------------------ 2 files changed, 1 insertion(+), 40 deletions(-) diff --git a/tensorflow/python/debug/BUILD b/tensorflow/python/debug/BUILD index 1ef0504ecb8..bb3bbbd87c4 100644 --- a/tensorflow/python/debug/BUILD +++ b/tensorflow/python/debug/BUILD @@ -788,7 +788,6 @@ cuda_py_test( "//tensorflow/python:training", "//tensorflow/python:variables", "//tensorflow/python/data/ops:dataset_ops", - "//tensorflow/python/keras", "//third_party/py/numpy", ], ) @@ -1400,7 +1399,6 @@ py_test( "//tensorflow/python:state_ops", "//tensorflow/python:training", "//tensorflow/python:variables", - "//tensorflow/python/keras", "//third_party/py/numpy", ], ) diff --git a/tensorflow/python/debug/wrappers/local_cli_wrapper_test.py b/tensorflow/python/debug/wrappers/local_cli_wrapper_test.py index 30bb99387b2..ab33a4af030 100644 --- a/tensorflow/python/debug/wrappers/local_cli_wrapper_test.py +++ b/tensorflow/python/debug/wrappers/local_cli_wrapper_test.py @@ -22,10 +22,10 @@ import tempfile import numpy as np -from tensorflow.python.debug.cli import cli_config from tensorflow.core.protobuf import config_pb2 from tensorflow.core.protobuf import rewriter_config_pb2 from tensorflow.python.client import session +from tensorflow.python.debug.cli import cli_config from tensorflow.python.debug.cli import cli_shared from tensorflow.python.debug.cli import debugger_cli_common from tensorflow.python.debug.cli import ui_factory @@ -36,9 +36,6 @@ from tensorflow.python.framework import errors from tensorflow.python.framework import ops from tensorflow.python.framework import test_util from tensorflow.python.lib.io import file_io -from tensorflow.python.keras import backend -from tensorflow.python.keras.engine import sequential -from tensorflow.python.keras.layers import core from tensorflow.python.ops import array_ops from tensorflow.python.ops import control_flow_ops from tensorflow.python.ops import math_ops @@ -832,40 +829,6 @@ class LocalCLIDebugWrapperSessionTest(test_util.TensorFlowTestCase): run_output = wrapped_sess.run([]) self.assertEqual([], run_output) - def testDebuggingKerasFitWithSkippedRunsWorks(self): - wrapped_sess = LocalCLIDebuggerWrapperSessionForTest( - [["run"], ["run"], ["run", "-t", "10"]], self.sess) - - backend.set_session(wrapped_sess) - - model = sequential.Sequential() - model.add(core.Dense(4, input_shape=[2], activation="relu")) - model.add(core.Dense(1)) - model.compile(loss="mse", optimizer="sgd") - - x = np.zeros([8, 2]) - y = np.zeros([8, 1]) - model.fit(x, y, epochs=2) - - self.assertEqual(2, len(wrapped_sess.observers["debug_dumps"])) - - def testDebuggingKerasFitWithProfilingWorks(self): - wrapped_sess = LocalCLIDebuggerWrapperSessionForTest( - [["run", "-p"]] * 10, self.sess) - - backend.set_session(wrapped_sess) - - model = sequential.Sequential() - model.add(core.Dense(4, input_shape=[2], activation="relu")) - model.add(core.Dense(1)) - model.compile(loss="mse", optimizer="sgd") - - x = np.zeros([8, 2]) - y = np.zeros([8, 1]) - model.fit(x, y, epochs=2) - - self.assertEqual(0, len(wrapped_sess.observers["debug_dumps"])) - def testRunsWithEmptyNestedFetchWorks(self): wrapped_sess = LocalCLIDebuggerWrapperSessionForTest( [["run"]], self.sess, dump_root="") From f8195170f88259829ab11538bd5d4b068dfa44a8 Mon Sep 17 00:00:00 2001 From: "A. Unique TensorFlower" Date: Wed, 17 Jun 2020 13:23:53 -0700 Subject: [PATCH 113/867] Enable a disabled testcase, since it is fixed. PiperOrigin-RevId: 316952692 Change-Id: I70d295634012685097b2324b156b2552218d62d7 --- .../keras/layers/preprocessing/text_vectorization_test.py | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/tensorflow/python/keras/layers/preprocessing/text_vectorization_test.py b/tensorflow/python/keras/layers/preprocessing/text_vectorization_test.py index 992f47efc85..e7f61e94724 100644 --- a/tensorflow/python/keras/layers/preprocessing/text_vectorization_test.py +++ b/tensorflow/python/keras/layers/preprocessing/text_vectorization_test.py @@ -1503,7 +1503,7 @@ class TextVectorizationSavingTest( loaded_model = keras.models.load_model(output_path) self.assertAllEqual(loaded_model.predict(input_array), expected_output) - def DISABLE_test_saving_with_tfidf(self): + def test_saving_with_tfidf(self): vocab_data = ["earth", "wind", "and", "fire"] tfidf_data = [.5, .25, .2, .125] input_array = np.array([["earth", "wind", "and", "earth"], From 56e71dd0e77507fc5200a650de5c736271df7f8a Mon Sep 17 00:00:00 2001 From: Pavithra Vijay Date: Wed, 17 Jun 2020 13:42:03 -0700 Subject: [PATCH 114/867] Removing identity backtracking from entropy losses. PiperOrigin-RevId: 316956157 Change-Id: I91130052e29e69ae131fe8aad0bbd1d4d42b00f1 --- tensorflow/python/keras/backend.py | 65 ++++++++----------- .../keras/tests/add_loss_correctness_test.py | 14 ++++ 2 files changed, 42 insertions(+), 37 deletions(-) diff --git a/tensorflow/python/keras/backend.py b/tensorflow/python/keras/backend.py index 391c695b18f..9330425272f 100644 --- a/tensorflow/python/keras/backend.py +++ b/tensorflow/python/keras/backend.py @@ -4637,12 +4637,6 @@ def softsign(x): return nn.softsign(x) -def _backtrack_identity(tensor): - while tensor.op.type == 'Identity': - tensor = tensor.op.inputs[0] - return tensor - - @keras_export('keras.backend.categorical_crossentropy') @dispatch.add_dispatch_support def categorical_crossentropy(target, output, from_logits=False, axis=-1): @@ -4695,17 +4689,16 @@ def categorical_crossentropy(target, output, from_logits=False, axis=-1): return nn.softmax_cross_entropy_with_logits_v2( labels=target, logits=output, axis=axis) - if not isinstance(output, (ops.EagerTensor, variables_module.Variable)): - output = _backtrack_identity(output) - if output.op.type == 'Softmax': - # When softmax activation function is used for output operation, we - # use logits from the softmax function directly to compute loss in order - # to prevent collapsing zero when training. - # See b/117284466 - assert len(output.op.inputs) == 1 - output = output.op.inputs[0] - return nn.softmax_cross_entropy_with_logits_v2( - labels=target, logits=output, axis=axis) + if (not isinstance(output, (ops.EagerTensor, variables_module.Variable)) and + output.op.type == 'Softmax'): + # When softmax activation function is used for output operation, we + # use logits from the softmax function directly to compute loss in order + # to prevent collapsing zero when training. + # See b/117284466 + assert len(output.op.inputs) == 1 + output = output.op.inputs[0] + return nn.softmax_cross_entropy_with_logits_v2( + labels=target, logits=output, axis=axis) # scale preds so that the class probas of each sample sum to 1 output = output / math_ops.reduce_sum(output, axis, True) @@ -4740,17 +4733,16 @@ def sparse_categorical_crossentropy(target, output, from_logits=False, axis=-1): target = ops.convert_to_tensor_v2(target) output = ops.convert_to_tensor_v2(output) - if not from_logits and not isinstance( - output, (ops.EagerTensor, variables_module.Variable)): - output = _backtrack_identity(output) - if output.op.type == 'Softmax': - # When softmax activation function is used for output operation, we - # use logits from the softmax function directly to compute loss in order - # to prevent collapsing zero when training. - # See b/117284466 - assert len(output.op.inputs) == 1 - output = output.op.inputs[0] - from_logits = True + if (not from_logits and + not isinstance(output, (ops.EagerTensor, variables_module.Variable)) and + output.op.type == 'Softmax'): + # When softmax activation function is used for output operation, we + # use logits from the softmax function directly to compute loss in order + # to prevent collapsing zero when training. + # See b/117284466 + assert len(output.op.inputs) == 1 + output = output.op.inputs[0] + from_logits = True if not from_logits: epsilon_ = _constant_to_tensor(epsilon(), output.dtype.base_dtype) @@ -4821,15 +4813,14 @@ def binary_crossentropy(target, output, from_logits=False): if from_logits: return nn.sigmoid_cross_entropy_with_logits(labels=target, logits=output) - if not isinstance(output, (ops.EagerTensor, variables_module.Variable)): - output = _backtrack_identity(output) - if output.op.type == 'Sigmoid': - # When sigmoid activation function is used for output operation, we - # use logits from the sigmoid function directly to compute loss in order - # to prevent collapsing zero when training. - assert len(output.op.inputs) == 1 - output = output.op.inputs[0] - return nn.sigmoid_cross_entropy_with_logits(labels=target, logits=output) + if (not isinstance(output, (ops.EagerTensor, variables_module.Variable)) and + output.op.type == 'Sigmoid'): + # When sigmoid activation function is used for output operation, we + # use logits from the sigmoid function directly to compute loss in order + # to prevent collapsing zero when training. + assert len(output.op.inputs) == 1 + output = output.op.inputs[0] + return nn.sigmoid_cross_entropy_with_logits(labels=target, logits=output) epsilon_ = _constant_to_tensor(epsilon(), output.dtype.base_dtype) output = clip_ops.clip_by_value(output, epsilon_, 1. - epsilon_) diff --git a/tensorflow/python/keras/tests/add_loss_correctness_test.py b/tensorflow/python/keras/tests/add_loss_correctness_test.py index a19eec75ffb..f99b285489d 100644 --- a/tensorflow/python/keras/tests/add_loss_correctness_test.py +++ b/tensorflow/python/keras/tests/add_loss_correctness_test.py @@ -34,6 +34,7 @@ from tensorflow.python.keras import testing_utils from tensorflow.python.ops import array_ops from tensorflow.python.ops import math_ops from tensorflow.python.platform import test +from tensorflow.python.platform import tf_logging as logging from tensorflow.python.training.rmsprop import RMSPropOptimizer MAE = losses.MeanAbsoluteError @@ -450,6 +451,19 @@ class TestAddLossCorrectness(keras_parameterized.TestCase): 'Expected a symbolic Tensors or a callable for the loss value'): model.add_loss(model.weights[0]) + @keras_parameterized.run_all_keras_modes + def test_add_entropy_loss_on_functional_model(self): + inputs = Input(shape=(1,)) + targets = Input(shape=(1,)) + outputs = testing_utils.Bias()(inputs) + model = Model([inputs, targets], outputs) + model.add_loss(losses.binary_crossentropy(targets, outputs)) + model.compile('sgd', run_eagerly=testing_utils.should_run_eagerly()) + with test.mock.patch.object(logging, 'warning') as mock_log: + model.fit([self.x, self.y], batch_size=3, epochs=5) + self.assertNotIn('Gradients do not exist for variables', + str(mock_log.call_args)) + if __name__ == '__main__': test.main() From 8e7be6f71aaa97f8c2ef806da8945a9d11f99830 Mon Sep 17 00:00:00 2001 From: Kuangyuan Chen Date: Wed, 17 Jun 2020 13:56:34 -0700 Subject: [PATCH 115/867] Populate side-effect decorators for HashTable ops. PiperOrigin-RevId: 316958908 Change-Id: I8e75fe538bb3dfc4b6d0742f63072fe87f133a4d --- tensorflow/compiler/mlir/tensorflow/ir/tf_generated_ops.td | 6 +++--- tensorflow/compiler/mlir/tensorflow/ir/tf_op_base.td | 3 +++ tensorflow/compiler/mlir/tensorflow/ir/tf_side_effects.h | 4 ++++ 3 files changed, 10 insertions(+), 3 deletions(-) diff --git a/tensorflow/compiler/mlir/tensorflow/ir/tf_generated_ops.td b/tensorflow/compiler/mlir/tensorflow/ir/tf_generated_ops.td index d403462e6a6..7ed5a215ab8 100644 --- a/tensorflow/compiler/mlir/tensorflow/ir/tf_generated_ops.td +++ b/tensorflow/compiler/mlir/tensorflow/ir/tf_generated_ops.td @@ -3862,7 +3862,7 @@ table will be immutable. ); let results = (outs - TF_ResourceTensor:$table_handle + Res:$table_handle ); } @@ -4814,7 +4814,7 @@ table. It must also be of the same type as the table values. }]; let arguments = (ins - TF_ResourceTensor:$table_handle, + Arg:$table_handle, TF_Tensor:$keys, TF_Tensor:$default_value ); @@ -4838,7 +4838,7 @@ The tensor `values` must be of the type of the table values. }]; let arguments = (ins - TF_ResourceTensor:$table_handle, + Arg:$table_handle, TF_Tensor:$keys, TF_Tensor:$values ); diff --git a/tensorflow/compiler/mlir/tensorflow/ir/tf_op_base.td b/tensorflow/compiler/mlir/tensorflow/ir/tf_op_base.td index 17424b54fc2..aac03061718 100644 --- a/tensorflow/compiler/mlir/tensorflow/ir/tf_op_base.td +++ b/tensorflow/compiler/mlir/tensorflow/ir/tf_op_base.td @@ -91,14 +91,17 @@ class TF_ResourceBase : def TF_VariableResource : TF_ResourceBase<"Variable">; def TF_StackResource : TF_ResourceBase<"Stack">; def TF_TensorArrayResource : TF_ResourceBase<"TensorArray">; +def TF_TableResource : TF_ResourceBase<"Table">; def TF_VariableRead : MemRead; def TF_StackRead : MemRead; def TF_TensorArrayRead : MemRead; +def TF_TableRead : MemRead; def TF_VariableWrite : MemWrite; def TF_StackWrite : MemWrite; def TF_TensorArrayWrite : MemWrite; +def TF_TableWrite: MemWrite; //===----------------------------------------------------------------------===// // TensorFlow op definitions diff --git a/tensorflow/compiler/mlir/tensorflow/ir/tf_side_effects.h b/tensorflow/compiler/mlir/tensorflow/ir/tf_side_effects.h index 9be61b1db39..ab1d7935bad 100644 --- a/tensorflow/compiler/mlir/tensorflow/ir/tf_side_effects.h +++ b/tensorflow/compiler/mlir/tensorflow/ir/tf_side_effects.h @@ -35,6 +35,10 @@ struct TensorArray : ::mlir::SideEffects::Resource::Base { StringRef getName() final { return "TensorArray"; } }; +struct Table : ::mlir::SideEffects::Resource::Base { + StringRef getName() final { return "Table"; } +}; + } // namespace ResourceEffects } // namespace TF } // namespace mlir From cbf8f57413d8a14c32515e88292012324a0990f0 Mon Sep 17 00:00:00 2001 From: Kuangyuan Chen Date: Wed, 17 Jun 2020 14:16:51 -0700 Subject: [PATCH 116/867] Import initialization graph in SignatureDef SavedModels as an MLIR function in TF saved model dialect. PiperOrigin-RevId: 316963067 Change-Id: I6c49ba4cfa3832f136bec36700aea695ad116737 --- tensorflow/compiler/mlir/tensorflow/BUILD | 4 +- .../mlir/tensorflow/ir/tf_saved_model.cc | 25 +++ .../mlir/tensorflow/ir/tf_saved_model_ops.td | 24 +++ .../tests/tf_saved_model/common_v1.py | 1 + .../tests/tf_saved_model/hash_table_v1.py | 92 +++++++++++ .../tensorflow/tests/tf_saved_model_ops.mlir | 5 + .../tests/tf_saved_model_ops_invalid.mlir | 33 ++++ .../mlir/tensorflow/translate/import_model.cc | 149 +++++++++++++----- 8 files changed, 290 insertions(+), 43 deletions(-) create mode 100644 tensorflow/compiler/mlir/tensorflow/tests/tf_saved_model/hash_table_v1.py diff --git a/tensorflow/compiler/mlir/tensorflow/BUILD b/tensorflow/compiler/mlir/tensorflow/BUILD index 904ccb7e820..17ed0e36a28 100644 --- a/tensorflow/compiler/mlir/tensorflow/BUILD +++ b/tensorflow/compiler/mlir/tensorflow/BUILD @@ -661,7 +661,9 @@ cc_library( ":tensorflow_types", ":translate_utils", "//tensorflow/cc/saved_model:bundle_v2", + "//tensorflow/cc/saved_model:constants", "//tensorflow/cc/saved_model:loader_lite", + "//tensorflow/cc/saved_model:loader_util", "//tensorflow/compiler/jit:shape_inference_helpers", "//tensorflow/compiler/mlir:op_or_arg_name_mapper", "//tensorflow/compiler/tf2xla:functionalize_control_flow", @@ -673,6 +675,7 @@ cc_library( "//tensorflow/core:lib", "//tensorflow/core:protos_all_cc", "//tensorflow/core/grappler/utils:transitive_fanin", + "//tensorflow/core/platform:protobuf_internal", "//tensorflow/core/platform:types", "//tensorflow/stream_executor/lib", "@com_google_absl//absl/algorithm:container", @@ -682,7 +685,6 @@ cc_library( "@com_google_absl//absl/strings", "@com_google_absl//absl/types:optional", "@llvm-project//llvm:Support", - "@llvm-project//mlir:Analysis", "@llvm-project//mlir:IR", "@llvm-project//mlir:Pass", "@llvm-project//mlir:StandardOps", diff --git a/tensorflow/compiler/mlir/tensorflow/ir/tf_saved_model.cc b/tensorflow/compiler/mlir/tensorflow/ir/tf_saved_model.cc index 140a778770c..6af70158e14 100644 --- a/tensorflow/compiler/mlir/tensorflow/ir/tf_saved_model.cc +++ b/tensorflow/compiler/mlir/tensorflow/ir/tf_saved_model.cc @@ -76,6 +76,23 @@ static LogicalResult Verify(GlobalTensorOp global_tensor) { return success(); } +static LogicalResult Verify(SessionInitializerOp session_initializer) { + mlir::SymbolTable symbol_table( + session_initializer.getParentOfType()); + + auto init_func_op = + symbol_table.lookup(session_initializer.initializer()); + if (!init_func_op) + return session_initializer.emitOpError() + << "the initializer function does not exist"; + + if (!init_func_op.getType().getResults().empty()) + return session_initializer.emitOpError() + << "the initializer function should have no output"; + + return success(); +} + #define GET_OP_CLASSES #include "tensorflow/compiler/mlir/tensorflow/ir/tf_saved_model.cc.inc" @@ -220,6 +237,14 @@ static LogicalResult VerifySavedModelModule( } } } + + auto session_initializers = module.getOps(); + if (std::distance(session_initializers.begin(), session_initializers.end()) > + 1) { + return (*++session_initializers.begin()).emitError() + << "there must be no more than one session_initializer op"; + } + SymbolTable symbol_table(module); auto symbol_uses = SymbolTable::getSymbolUses(&module.getBodyRegion()); if (!symbol_uses.hasValue()) { diff --git a/tensorflow/compiler/mlir/tensorflow/ir/tf_saved_model_ops.td b/tensorflow/compiler/mlir/tensorflow/ir/tf_saved_model_ops.td index 4431a160edf..497f4d90cb9 100644 --- a/tensorflow/compiler/mlir/tensorflow/ir/tf_saved_model_ops.td +++ b/tensorflow/compiler/mlir/tensorflow/ir/tf_saved_model_ops.td @@ -128,4 +128,28 @@ def TfSavedModel_GlobalTensorOp : TfSavedModel_Op<"global_tensor"> { let verifier = [{ return Verify(*this); }]; } +def TfSavedModel_SessionInitializerOp: TfSavedModel_Op<"session_initializer"> { + let summary = "Initializes TensorFlow session state."; + let description = [{ + Represents a session initializer function initializes TensorFlow session + state. It is used to initialize resources in the saved model before calling + any exported functions. There must be no more than one session initializer + in a saved model. + + The `initializer` represents the initialization function. The function have + no output and this function should be only called once. + + This is used, for example, to initialize hash tables stored in resources and + accessed by resource name (rather than as resource handles or bound inputs + which is how `global_tensor`s are referenced) + }]; + + let arguments = (ins + FlatSymbolRefAttr:$initializer + ); + + + let verifier = [{ return Verify(*this); }]; +} + #endif // SAVED_MODEL_DIALECT diff --git a/tensorflow/compiler/mlir/tensorflow/tests/tf_saved_model/common_v1.py b/tensorflow/compiler/mlir/tensorflow/tests/tf_saved_model/common_v1.py index 7171f63bb05..51ccbeb1fbd 100644 --- a/tensorflow/compiler/mlir/tensorflow/tests/tf_saved_model/common_v1.py +++ b/tensorflow/compiler/mlir/tensorflow/tests/tf_saved_model/common_v1.py @@ -84,6 +84,7 @@ def do_test(signature_def_map, show_debug_info=False): builder.add_meta_graph_and_variables( sess, [tf.saved_model.tag_constants.SERVING], signature_def_map, + main_op=tf.tables_initializer(), strip_default_attrs=True) builder.save() diff --git a/tensorflow/compiler/mlir/tensorflow/tests/tf_saved_model/hash_table_v1.py b/tensorflow/compiler/mlir/tensorflow/tests/tf_saved_model/hash_table_v1.py new file mode 100644 index 00000000000..64847434b82 --- /dev/null +++ b/tensorflow/compiler/mlir/tensorflow/tests/tf_saved_model/hash_table_v1.py @@ -0,0 +1,92 @@ +# Copyright 2019 The TensorFlow Authors. All Rights Reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. +# ============================================================================== + +# RUN: %p/hash_table_v1 | FileCheck %s + +# pylint: disable=missing-docstring,line-too-long +from __future__ import absolute_import +from __future__ import division +from __future__ import print_function + +import tensorflow.compat.v1 as tf +from tensorflow.compiler.mlir.tensorflow.tests.tf_saved_model import common_v1 + +# Verify that the tf.versions attribute exists. It is difficult to enforce +# contents, since the version numbers change over time. The conversion logic +# itself is verified in the common graphdef converter, so here just assert +# it is being invoked. +# CHECK: module +# CHECK-SAME: tf.versions +# CHECK-SAME: bad_consumers +# CHECK-SAME: min_consumer +# CHECK-SAME: producer + +# CHECK: "tf_saved_model.session_initializer"() {initializer = [[init:@.*]]} : () -> () +# CHECK: "tf_saved_model.global_tensor"() + +# CHECK: func {{@[a-zA-Z_0-9]+}}( +# CHECK-SAME: [[ARG0:%.*]]: tensor +# CHECK-SAME: [[ARG1:%.*]]: tensor () + // Representation for constants: (immutable) global tensor. // CHECK: tf_saved_model.global_tensor "tf_saved_model.global_tensor"() { diff --git a/tensorflow/compiler/mlir/tensorflow/tests/tf_saved_model_ops_invalid.mlir b/tensorflow/compiler/mlir/tensorflow/tests/tf_saved_model_ops_invalid.mlir index c055c6c9f56..544600cf6b8 100644 --- a/tensorflow/compiler/mlir/tensorflow/tests/tf_saved_model_ops_invalid.mlir +++ b/tensorflow/compiler/mlir/tensorflow/tests/tf_saved_model_ops_invalid.mlir @@ -258,3 +258,36 @@ module attributes {tf_saved_model.semantics} { // expected-error@+1 {{'type' attribute for immutable 'tf_saved_model.global_tensor' should have a static shape}} "tf_saved_model.global_tensor"() { sym_name = "v", type = tensor, value = dense<1.> : tensor<1xf32> } : () -> () } + +// ----- + +module attributes {tf_saved_model.semantics} { + + // expected-error@+1 {{the initializer function does not exist}} + "tf_saved_model.session_initializer"() { initializer = @init } : () -> () +} + +// ----- + +module attributes {tf_saved_model.semantics} { + + // expected-error@+1 {{the initializer function should have no output}} + "tf_saved_model.session_initializer"() { initializer = @init } : () -> () + func @init() -> tensor<1xf32> { + %0 = "tf.Const"() {value = dense<[1.0]> : tensor<1xf32> } : () -> tensor<1xf32> + return %0 : tensor<1xf32> + } +} + +// ----- + +module attributes {tf_saved_model.semantics} { + + "tf_saved_model.session_initializer"() { initializer = @init } : () -> () + // expected-error@+1 {{there must be no more than one session_initializer op}} + "tf_saved_model.session_initializer"() { initializer = @init } : () -> () + func @init() -> tensor<1xf32> { + %0 = "tf.Const"() {value = dense<[1.0]> : tensor<1xf32> } : () -> tensor<1xf32> + return %0 : tensor<1xf32> + } +} diff --git a/tensorflow/compiler/mlir/tensorflow/translate/import_model.cc b/tensorflow/compiler/mlir/tensorflow/translate/import_model.cc index 820d0ce31fb..3cff4217215 100644 --- a/tensorflow/compiler/mlir/tensorflow/translate/import_model.cc +++ b/tensorflow/compiler/mlir/tensorflow/translate/import_model.cc @@ -60,6 +60,8 @@ limitations under the License. #include "mlir/IR/Types.h" // from @llvm-project #include "mlir/IR/Verifier.h" // from @llvm-project #include "mlir/Pass/PassManager.h" // from @llvm-project +#include "tensorflow/cc/saved_model/constants.h" +#include "tensorflow/cc/saved_model/loader_util.h" #include "tensorflow/compiler/jit/shape_inference_helpers.h" #include "tensorflow/compiler/mlir/op_or_arg_name_mapper.h" #include "tensorflow/compiler/mlir/tensorflow/ir/tf_attributes.h" @@ -99,6 +101,7 @@ limitations under the License. #include "tensorflow/core/lib/core/errors.h" #include "tensorflow/core/lib/strings/str_util.h" #include "tensorflow/core/platform/protobuf.h" +#include "tensorflow/core/platform/protobuf_internal.h" #include "tensorflow/core/platform/types.h" #include "tensorflow/core/protobuf/graph_debug_info.pb.h" #include "tensorflow/core/protobuf/meta_graph.pb.h" @@ -116,6 +119,7 @@ using mlir::NamedAttrList; using mlir::TensorType; using mlir::TF::VarHandleOp; using mlir::tf_saved_model::GlobalTensorOp; +using mlir::tf_saved_model::SessionInitializerOp; using stream_executor::port::StatusOr; namespace { @@ -2955,6 +2959,13 @@ void SortSavedModelModule(mlir::ModuleOp module) { named_global_tensor.global_tensor.getOperation()->moveBefore( &module.getBody()->front()); } + + auto initializers = module.getOps(); + if (!initializers.empty()) { + (*initializers.begin()) + .getOperation() + ->moveBefore(&module.getBody()->front()); + } } Status CreateSavedModelIR( @@ -3241,17 +3252,29 @@ class SavedModelSignatureDefImporter { absl::Span exported_names, mlir::MLIRContext* context) : bundle_(bundle), + flib_def_(OpRegistry::Global(), graph_def().library()), + debug_info_(), exported_names_(exported_names), - module_(mlir::ModuleOp::create(mlir::UnknownLoc::get(context))) {} + module_(mlir::ModuleOp::create(mlir::UnknownLoc::get(context))) { + // debug_info might not be loaded with loader_lite. + if (bundle_.debug_info != nullptr) debug_info_ = *bundle_.debug_info; + } // Converts the SavedModel to the SavedModel dialect. Creates an MLIR function // for each signature. StatusOr ConvertSignatures(); - Status ConvertSignature(const GraphDef& graphdef, - const std::string& sig_def_key, - const SignatureDef& signature_def, - const GraphDebugInfo& debug_info, - const FunctionLibraryDefinition& flib_def); + Status ConvertSignature(const std::string& sig_def_key, + const SignatureDef& signature_def); + + // Converts the initialization graph in the SavedModel to an MLIR function. + Status ConvertInitializer(); + + // Converts a graph with feeds and fetches to an MLIR function. + StatusOr ConvertGraph( + const std::string& name, + const std::vector>& inputs, + const std::vector>& outputs, + const std::vector control_outputs); // Creates GlobalTensorOp for each variable and moves each VarHandle op to // the enclosing function's arguments. @@ -3273,18 +3296,62 @@ class SavedModelSignatureDefImporter { GraphImportConfig::InputArrays ParseInputArrays( const std::vector>& inputs); + const GraphDef& graph_def() const { + return bundle_.meta_graph_def.graph_def(); + } + const FunctionLibraryDefinition& flib_def() const { return flib_def_; } + const GraphDebugInfo& debug_info() const { return debug_info_; } + const SavedModelBundle& bundle_; + FunctionLibraryDefinition flib_def_; + GraphDebugInfo debug_info_; absl::Span exported_names_; mlir::OwningModuleRef module_; }; +Status SavedModelSignatureDefImporter::ConvertInitializer() { + std::vector asset_file_defs; + TF_RETURN_IF_ERROR( + internal::GetAssetFileDefs(bundle_.meta_graph_def, &asset_file_defs)); + + if (!asset_file_defs.empty()) + return errors::Unimplemented( + absl::StrCat("Assets are not supported in signaturedef importer")); + + std::string init_node_name; + TF_RETURN_IF_ERROR( + internal::GetInitOp("", bundle_.meta_graph_def, &init_node_name)); + + if (init_node_name.empty()) return Status::OK(); + + TF_ASSIGN_OR_RETURN(auto sub_module, + ConvertGraph(init_node_name, {}, {}, {init_node_name})); + + mlir::SymbolTable symbol_table(*sub_module); + + auto init_func_op = symbol_table.lookup(init_node_name); + + init_func_op.removeAttr("tf.entry_function"); + + mlir::OpBuilder builder(module_->getBodyRegion()); + + builder.create( + module_->getLoc(), builder.getSymbolRefAttr(init_func_op.getName())); + + // Move the converted functions to top level MLIR module. + auto* block = module_->getBody(); + auto* sub_block = sub_module->getBody(); + block->getOperations().splice( + mlir::Block::iterator(block->getTerminator()), sub_block->getOperations(), + sub_block->begin(), mlir::Block::iterator(sub_block->getTerminator())); + + return Status::OK(); +} + StatusOr SavedModelSignatureDefImporter::ConvertSignatures() { const auto& signatures = bundle_.GetSignatures(); - const auto& graphdef = bundle_.meta_graph_def.graph_def(); - PopulateTfVersions(module_.get(), graphdef.versions()); - - FunctionLibraryDefinition flib_def(OpRegistry::Global(), graphdef.library()); + PopulateTfVersions(module_.get(), graph_def().versions()); // debug_info might not be loaded with loader_lite. GraphDebugInfo debug_info; @@ -3307,9 +3374,10 @@ SavedModelSignatureDefImporter::ConvertSignatures() { continue; } - TF_RETURN_IF_ERROR(ConvertSignature(graphdef, sig_def_key, signature_def, - debug_info, flib_def)); + TF_RETURN_IF_ERROR(ConvertSignature(sig_def_key, signature_def)); } + + TF_RETURN_IF_ERROR(ConvertInitializer()); TF_RETURN_IF_ERROR(LiftVariables()); mlir::OpBuilder builder(module_->getBodyRegion()); @@ -3320,10 +3388,32 @@ SavedModelSignatureDefImporter::ConvertSignatures() { return std::move(module_); } +StatusOr SavedModelSignatureDefImporter::ConvertGraph( + const std::string& name, + const std::vector>& inputs, + const std::vector>& outputs, + const std::vector control_outputs) { + GraphImportConfig specs; + specs.prune_unused_nodes = true; + specs.inputs = ParseInputArrays(inputs); + for (auto& output : outputs) specs.outputs.push_back(output.second.name()); + specs.control_outputs = control_outputs; + + // Convert sub-graphdef to sub-graph. + GraphConstructorOptions options; + options.allow_internal_ops = true; + options.add_default_attributes = true; + Graph graph(OpRegistry::Global()); + + TF_RETURN_IF_ERROR(ConvertGraphDefToGraph(options, graph_def(), &graph)); + + // Convert sub-graph to MLIR module.true + return GraphDefImporter::Convert(module_->getContext(), graph, debug_info(), + flib_def(), specs, name); +} + Status SavedModelSignatureDefImporter::ConvertSignature( - const GraphDef& graphdef, const std::string& sig_def_key, - const SignatureDef& signature_def, const GraphDebugInfo& debug_info, - const FunctionLibraryDefinition& flib_def) { + const std::string& sig_def_key, const SignatureDef& signature_def) { // Create local vectors for the input and output and sort them to be // deterministic. We don't want anyone to really depend on the order, client // should lookup argument/result mapping by attribute name. @@ -3339,34 +3429,9 @@ Status SavedModelSignatureDefImporter::ConvertSignature( return lhs.first.size() < rhs.first.size() || lhs.first > rhs.first; }); - GraphImportConfig specs; - specs.prune_unused_nodes = true; - specs.inputs = ParseInputArrays(inputs); - for (auto& output : outputs) specs.outputs.push_back(output.second.name()); - - // Remove unused nodes and create sub-graphdef. - GraphDef sub_graph_def; - TF_RETURN_IF_ERROR(tensorflow::grappler::SetTransitiveFaninGraph( - graphdef, &sub_graph_def, - /*terminal_nodes=*/{specs.outputs.begin(), specs.outputs.end()})); - - // Set the function library definitions in the pruned graphdef. - *sub_graph_def.mutable_library() = flib_def.ToProto(); - - // Convert sub-graphdef to sub-graph. - GraphConstructorOptions options; - options.allow_internal_ops = true; - options.add_default_attributes = true; - Graph sub_graph(OpRegistry::Global()); - - TF_RETURN_IF_ERROR( - ConvertGraphDefToGraph(options, sub_graph_def, &sub_graph)); - // Convert sub-graph to MLIR module. - TF_ASSIGN_OR_RETURN( - auto sub_module, - GraphDefImporter::Convert(module_->getContext(), sub_graph, debug_info, - flib_def, specs, sig_def_key)); + TF_ASSIGN_OR_RETURN(auto sub_module, + ConvertGraph(sig_def_key, inputs, outputs, {})); mlir::OpBuilder builder(sub_module->getBodyRegion()); // Find the FuncOp which corresponds to current SignatureDef. From 5b5ab0034ae0f12731943177b275e35cb9f92bb1 Mon Sep 17 00:00:00 2001 From: Lukas Geiger Date: Wed, 17 Jun 2020 23:26:29 +0200 Subject: [PATCH 117/867] Fix distributed autocast variable assign --- .../experimental/autocast_variable.py | 104 +++++++++--------- .../experimental/autocast_variable_test.py | 30 +++-- 2 files changed, 77 insertions(+), 57 deletions(-) diff --git a/tensorflow/python/keras/mixed_precision/experimental/autocast_variable.py b/tensorflow/python/keras/mixed_precision/experimental/autocast_variable.py index 7d0abe30581..ca6420f0c0b 100644 --- a/tensorflow/python/keras/mixed_precision/experimental/autocast_variable.py +++ b/tensorflow/python/keras/mixed_precision/experimental/autocast_variable.py @@ -188,61 +188,88 @@ class AutoCastVariable(variables.Variable, core.Tensor): def constraint(self): return self._variable.constraint + def _apply_assign_update( + self, update_fn, value, use_locking=None, name=None, read_value=True): + if not read_value: + return update_fn(value, use_locking, name, read_value) + + if context.executing_eagerly() or ops.inside_function(): + assign_op = update_fn(value, use_locking, name, False) + with ops.control_dependencies([assign_op]): + return self + + # Fallback to wrapping the returned variable in graph mode if possible + assign_var = update_fn(value, use_locking, name, read_value) + if resource_variable_ops.is_resource_variable(assign_var): + return create_autocast_variable(assign_var) + return assign_var + + def _apply_update(self, update_fn, *args, **kwargs): + update_var = update_fn(*args, **kwargs) + if context.executing_eagerly() or ops.inside_function(): + with ops.control_dependencies([update_var]): + return self + + # Fallback to wrapping the returned variable in graph mode if possible + if resource_variable_ops.is_resource_variable(update_var): + return create_autocast_variable(update_var) + return update_var + def assign(self, value, use_locking=None, name=None, read_value=True): - assign_op = self._variable.assign(value, use_locking, name, read_value) - return _maybe_wrap(assign_op, wrap=read_value) + return self._apply_assign_update( + self._variable.assign, value, use_locking, name, read_value) def assign_add(self, delta, use_locking=None, name=None, read_value=True): - assign_op = self._variable.assign_add(delta, use_locking, name, read_value) - return _maybe_wrap(assign_op, wrap=read_value) + return self._apply_assign_update( + self._variable.assign_add, delta, use_locking, name, read_value) def assign_sub(self, delta, use_locking=None, name=None, read_value=True): - assign_op = self._variable.assign_sub(delta, use_locking, name, read_value) - return _maybe_wrap(assign_op, wrap=read_value) + return self._apply_assign_update( + self._variable.assign_sub, delta, use_locking, name, read_value) def scatter_sub(self, sparse_delta, use_locking=False, name=None): - var = self._variable.scatter_sub(sparse_delta, use_locking, name) - return _maybe_wrap(var) + return self._apply_update( + self._variable.scatter_sub, sparse_delta, use_locking, name) def scatter_add(self, sparse_delta, use_locking=False, name=None): - var = self._variable.scatter_add(sparse_delta, use_locking, name) - return _maybe_wrap(var) + return self._apply_update( + self._variable.scatter_add, sparse_delta, use_locking, name) def scatter_max(self, sparse_delta, use_locking=False, name=None): - var = self._variable.scatter_max(sparse_delta, use_locking, name) - return _maybe_wrap(var) + return self._apply_update( + self._variable.scatter_max, sparse_delta, use_locking, name) def scatter_min(self, sparse_delta, use_locking=False, name=None): - var = self._variable.scatter_min(sparse_delta, use_locking, name) - return _maybe_wrap(var) + return self._apply_update( + self._variable.scatter_min, sparse_delta, use_locking, name) def scatter_mul(self, sparse_delta, use_locking=False, name=None): - var = self._variable.scatter_mul(sparse_delta, use_locking, name) - return _maybe_wrap(var) + return self._apply_update( + self._variable.scatter_mul, sparse_delta, use_locking, name) def scatter_div(self, sparse_delta, use_locking=False, name=None): - var = self._variable.scatter_div(sparse_delta, use_locking, name) - return _maybe_wrap(var) + return self._apply_update( + self._variable.scatter_div, sparse_delta, use_locking, name) def scatter_update(self, sparse_delta, use_locking=False, name=None): - var = self._variable.scatter_update(sparse_delta, use_locking, name) - return _maybe_wrap(var) + return self._apply_update( + self._variable.scatter_update, sparse_delta, use_locking, name) def batch_scatter_update(self, sparse_delta, use_locking=False, name=None): - var = self._variable.batch_scatter_update(sparse_delta, use_locking, name) - return _maybe_wrap(var) + return self._apply_update( + self._variable.batch_scatter_update, sparse_delta, use_locking, name) def scatter_nd_sub(self, indices, updates, name=None): - var = self._variable.scatter_nd_sub(indices, updates, name) - return _maybe_wrap(var) + return self._apply_update( + self._variable.scatter_nd_sub, indices, updates, name) def scatter_nd_add(self, indices, updates, name=None): - var = self._variable.scatter_nd_add(indices, updates, name) - return _maybe_wrap(var) + return self._apply_update( + self._variable.scatter_nd_add, indices, updates, name) def scatter_nd_update(self, indices, updates, name=None): - var = self._variable.scatter_nd_update(indices, updates, name) - return _maybe_wrap(var) + return self._apply_update( + self._variable.scatter_nd_update, indices, updates, name) def load(self, value, session=None): return self._variable.load(value, session) @@ -462,24 +489,3 @@ def create_autocast_variable(variable): # pylint: enable=missing-format-attribute return AutoCastDistributedVariable(variable) - - -def _maybe_wrap(variable, wrap=True): - """Creates an AutoCastVariable that wraps another variable if applicable. - - This function is used to wrap the return value of AutoCastVariable.assign. - Unfortunately MirroredVariable.assign will (incorrectly) return a Mirrored - value instead of a MirroredVariable. So we cannot properly wrap it in an - AutoCastVariable. We return the original variable in that case. - - Args: - variable: A tf.Variable or op. - wrap: A boolean to define whether to wrap the variable in an - AutoCastVariable or not. - - Returns: - An AutoCastVariable if wrap is True and variable is a resource variable. - """ - if wrap and resource_variable_ops.is_resource_variable(variable): - return create_autocast_variable(variable) - return variable diff --git a/tensorflow/python/keras/mixed_precision/experimental/autocast_variable_test.py b/tensorflow/python/keras/mixed_precision/experimental/autocast_variable_test.py index c45015b644e..940bd07c813 100644 --- a/tensorflow/python/keras/mixed_precision/experimental/autocast_variable_test.py +++ b/tensorflow/python/keras/mixed_precision/experimental/autocast_variable_test.py @@ -304,8 +304,8 @@ class AutoCastVariableTest(test.TestCase, parameterized.TestCase): self.assertAllClose(3., self.evaluate(x.assign_sub(3.))) # Assign multiple times - # This currently only works if no strategy is used - if not ds_context.has_strategy(): + # This currently doesn't work in graph mode + if context.executing_eagerly() or ops.inside_function(): assign = x.assign(1.) self.assertAllClose(1., self.evaluate(assign)) self.assertAllClose(0., self.evaluate(assign.assign(0.))) @@ -343,6 +343,20 @@ class AutoCastVariableTest(test.TestCase, parameterized.TestCase): # assign still expect float32 value even if in float16 scope run_and_check() + @combinations.generate(maybe_distribute) + def test_assign_tf_function(self, distribution): + with distribution.scope(): + x = get_var(0., dtypes.float32) + x = autocast_variable.create_autocast_variable(x) + + @def_function.function + def run_assign(): + return x.assign(1.).assign_add(3.).assign_add(3.).assign_sub(2.) + + with ops.get_default_graph()._enable_auto_casting_variables( + dtypes.float16): + self.assertAllClose(5., self.evaluate(run_assign())) + @combinations.generate(maybe_distribute) def test_assign_stays_in_true_dtype(self, distribution): with distribution.scope(): @@ -357,18 +371,18 @@ class AutoCastVariableTest(test.TestCase, parameterized.TestCase): dtypes.float16): # Variable should be increased, despite it appearing to be the same # float16 value. - self.assertEqual(1. + small_val, - self.evaluate(x.assign(1. + small_tensor))) + self.evaluate(x.assign(1. + small_tensor)) + self.assertEqual(1. + small_val, self.evaluate(x._variable)) self.assertEqual(1., self.evaluate(x.value())) - self.assertEqual(1. + small_val, self.evaluate(x.value())) + self.assertEqual(1. + small_val, self.evaluate(x)) self.evaluate(x.assign(1.)) with ops.get_default_graph()._enable_auto_casting_variables( dtypes.float16): - self.assertEqual(1. + small_val, - self.evaluate(x.assign_add(small_tensor))) + self.evaluate(x.assign_add(small_tensor)) + self.assertEqual(1. + small_val, self.evaluate(x._variable)) self.assertEqual(1., self.evaluate(x.value())) - self.assertEqual(1. + small_val, self.evaluate(x.value())) + self.assertEqual(1. + small_val, self.evaluate(x)) @combinations.generate(maybe_distribute) def test_checkpoint(self, distribution): From 92d68dd5f10ce8ef40852f5bb207258ce4126edf Mon Sep 17 00:00:00 2001 From: Francois Chollet Date: Wed, 17 Jun 2020 14:30:33 -0700 Subject: [PATCH 118/867] Avoid deprecation warning related to the use of `collections.Sequence`. It will stop working in Python 3.8. PiperOrigin-RevId: 316965801 Change-Id: Ia44313b1920653a0dd0a94d404ac914b08239c43 --- tensorflow/python/keras/engine/data_adapter.py | 18 ------------------ tensorflow/python/keras/layers/recurrent.py | 9 ++++++--- 2 files changed, 6 insertions(+), 21 deletions(-) diff --git a/tensorflow/python/keras/engine/data_adapter.py b/tensorflow/python/keras/engine/data_adapter.py index 469355dd722..29a99137982 100644 --- a/tensorflow/python/keras/engine/data_adapter.py +++ b/tensorflow/python/keras/engine/data_adapter.py @@ -19,7 +19,6 @@ from __future__ import division from __future__ import print_function import abc -import collections import contextlib import functools import itertools @@ -57,7 +56,6 @@ try: from scipy import sparse as scipy_sparse # pylint: disable=g-import-not-at-top except ImportError: scipy_sparse = None - try: import pandas as pd # pylint: disable=g-import-not-at-top except ImportError: @@ -786,7 +784,6 @@ class GeneratorDataAdapter(DataAdapter): # Since we have to know the dtype of the python generator when we build the # dataset, we have to look at a batch to infer the structure. peek, x = self._peek_and_restore(x) - assert_not_namedtuple(peek) peek = self._standardize_batch(peek) peek = _process_tensorlike(peek) @@ -1070,21 +1067,6 @@ def broadcast_sample_weight_modes(target_structure, sample_weight_modes): return sample_weight_modes -def assert_not_namedtuple(x): - if (isinstance(x, tuple) and - # TODO(b/144192902): Use a namedtuple checking utility. - hasattr(x, "_fields") and - isinstance(x._fields, collections.Sequence) and - all(isinstance(f, six.string_types) for f in x._fields)): - raise ValueError( - "Received namedtuple ({}) with fields `{}` as input. namedtuples " - "cannot, in general, be unambiguously resolved into `x`, `y`, " - "and `sample_weight`. For this reason Keras has elected not to " - "support them. If you would like the value to be unpacked, " - "please explicitly convert it to a tuple before passing it to " - "Keras.".format(x.__class__, x._fields)) - - class DataHandler(object): """Handles iterating over epoch-level `tf.data.Iterator` objects.""" diff --git a/tensorflow/python/keras/layers/recurrent.py b/tensorflow/python/keras/layers/recurrent.py index 0ce17c6101e..78a4a33a533 100644 --- a/tensorflow/python/keras/layers/recurrent.py +++ b/tensorflow/python/keras/layers/recurrent.py @@ -19,8 +19,6 @@ from __future__ import absolute_import from __future__ import division from __future__ import print_function -import collections - import numpy as np from tensorflow.python.distribute import distribution_strategy_context as ds_context @@ -49,6 +47,11 @@ from tensorflow.python.util import nest from tensorflow.python.util.tf_export import keras_export from tensorflow.tools.docs import doc_controls +try: + from collections import abc as collections_abc # pylint: disable=g-import-not-at-top +except ImportError: # For Python 2 + import collections as collections_abc # pylint: disable=g-import-not-at-top + RECURRENT_DROPOUT_WARNING_MSG = ( 'RNN `implementation=2` is not supported when `recurrent_dropout` is set. ' @@ -828,7 +831,7 @@ class RNN(Layer): # input shape: `(samples, time (padded with zeros), input_dim)` # note that the .build() method of subclasses MUST define # self.input_spec and self.state_spec with complete input shapes. - if (isinstance(inputs, collections.Sequence) + if (isinstance(inputs, collections_abc.Sequence) and not isinstance(inputs, tuple)): # get initial_state from full input spec # as they could be copied to multiple GPU. From b920ae92628e58362a3ef5fd55b3ef6e1a2ff617 Mon Sep 17 00:00:00 2001 From: Hanhan Wang Date: Wed, 17 Jun 2020 14:31:35 -0700 Subject: [PATCH 119/867] Internal change PiperOrigin-RevId: 316965990 Change-Id: Ic5ee6b8b3fc04c7d5ca1548f8b638ba9c12a32d6 --- third_party/mlir/BUILD | 1 - third_party/mlir/test.BUILD | 1 - 2 files changed, 2 deletions(-) diff --git a/third_party/mlir/BUILD b/third_party/mlir/BUILD index cb0b2f9dc8e..476b8566265 100644 --- a/third_party/mlir/BUILD +++ b/third_party/mlir/BUILD @@ -3286,7 +3286,6 @@ cc_library( ":LinalgTransforms", ":Pass", ":StandardOps", - ":StandardOpsTransforms", ":Support", ":Transforms", ":VectorToLLVM", diff --git a/third_party/mlir/test.BUILD b/third_party/mlir/test.BUILD index 14c2ba7778e..23287ce28d6 100644 --- a/third_party/mlir/test.BUILD +++ b/third_party/mlir/test.BUILD @@ -166,7 +166,6 @@ cc_library( "@llvm-project//mlir:Pass", "@llvm-project//mlir:SCFDialect", "@llvm-project//mlir:StandardOps", - "@llvm-project//mlir:StandardOpsTransforms", "@llvm-project//mlir:Support", "@llvm-project//mlir:TargetNVVMIR", "@llvm-project//mlir:TargetROCDLIR", From 69e36409c0eb96379f8a4e8d5219f64382aaf75e Mon Sep 17 00:00:00 2001 From: Karim Nosir Date: Wed, 17 Jun 2020 14:32:39 -0700 Subject: [PATCH 120/867] Update documentation for FC op in hexagon delegate. FC now supports relu activation function PiperOrigin-RevId: 316966229 Change-Id: If5a42ecca8aa9b6e94474e75614153f50ca8ae3b --- tensorflow/lite/delegates/hexagon/README.md | 2 +- tensorflow/lite/delegates/hexagon/builders/matmul_builder.cc | 1 - 2 files changed, 1 insertion(+), 2 deletions(-) diff --git a/tensorflow/lite/delegates/hexagon/README.md b/tensorflow/lite/delegates/hexagon/README.md index 226f3a61077..deff36e80b8 100644 --- a/tensorflow/lite/delegates/hexagon/README.md +++ b/tensorflow/lite/delegates/hexagon/README.md @@ -74,7 +74,7 @@ are verified in `IsNodeSupportedByHexagon`: - depth_multiplier == 1 - dilation only supported when stride == 1 - Otherwise, stride height/width <= 3 -* FullyConnected (without any activation) +* FullyConnected * Hardswish * L2Normalization (without any activation) * Logistic (aka Sigmoid) diff --git a/tensorflow/lite/delegates/hexagon/builders/matmul_builder.cc b/tensorflow/lite/delegates/hexagon/builders/matmul_builder.cc index 0757ea6180e..6189294c3a1 100644 --- a/tensorflow/lite/delegates/hexagon/builders/matmul_builder.cc +++ b/tensorflow/lite/delegates/hexagon/builders/matmul_builder.cc @@ -151,7 +151,6 @@ TfLiteStatus AddFullyConnectedHelper(const TfLiteIntArray* inputs, // Data (8-bit), Weights (const, 8-bit) => MatMul => MatMul out (int32) // MatMul out (int32), Bias (int32) => QuantizedBiasAdd => BiasAdd out (int32) // BiasAdd out (int32) => Requantize_32to8 => Output (8-bit) -// TODO(b/129276536): Add activation support. TfLiteStatus MatMulWithConstWeightsOpBuilder::PopulateSubGraph( const TfLiteIntArray* inputs, const TfLiteIntArray* outputs, TfLiteContext* context) { From b6ff68822a59578f942e4fb8076757da8db278ae Mon Sep 17 00:00:00 2001 From: Mihai Maruseac Date: Wed, 17 Jun 2020 14:32:53 -0700 Subject: [PATCH 121/867] Manually define `GpuAtomic{Max,Min}` for `long long`. CUDA only defines `atomic{Max,Min}` for `long long` for compute capability 3.5 or higher. Hence, to prevent compile errors due to failed instantiation when compiling on `sm30` (or other similar ones), we manually define the corresponding wrappers. PiperOrigin-RevId: 316966290 Change-Id: I813e331309d12bdc1c06f7a74cb45c20c1833a41 --- tensorflow/core/util/gpu_device_functions.h | 10 ++++++++++ 1 file changed, 10 insertions(+) diff --git a/tensorflow/core/util/gpu_device_functions.h b/tensorflow/core/util/gpu_device_functions.h index de55b5c33c4..d4e09a7fc98 100644 --- a/tensorflow/core/util/gpu_device_functions.h +++ b/tensorflow/core/util/gpu_device_functions.h @@ -820,6 +820,11 @@ __device__ inline tensorflow::uint64 GpuAtomicMax(tensorflow::uint64* ptr, return detail::GpuAtomicCasHelper( ptr, [value](tensorflow::uint64 a) { return max(a, value); }); } + +__device__ inline int64 GpuAtomicMax(int64* ptr, int64 value) { + return detail::GpuAtomicCasHelper(ptr, + [value](int64 a) { return max(a, value); }); +} #endif CREATE_CUDA_DEVICE_FUNCTION_ALIAS(GpuAtomicMax, CudaAtomicMax); @@ -885,6 +890,11 @@ __device__ inline tensorflow::uint64 GpuAtomicMin(tensorflow::uint64* ptr, return detail::GpuAtomicCasHelper( ptr, [value](tensorflow::uint64 a) { return min(a, value); }); } + +__device__ inline int64 GpuAtomicMin(int64* ptr, int64 value) { + return detail::GpuAtomicCasHelper(ptr, + [value](int64 a) { return min(a, value); }); +} #endif CREATE_CUDA_DEVICE_FUNCTION_ALIAS(GpuAtomicMin, CudaAtomicMin); From ed2b3d6e1ec7999cd44e84c62d3c99a4b8f047b8 Mon Sep 17 00:00:00 2001 From: Jaesung Chung Date: Wed, 17 Jun 2020 14:47:39 -0700 Subject: [PATCH 122/867] Fix msan test in tf_saved_model_lift_variables test Now, all the tenors, generated from lift_variables_test.h, has initialized values. PiperOrigin-RevId: 316969327 Change-Id: If7f68aacf0c931430804dfaa1d73c0ea4be70c75 --- .../transforms/lift_variables_test_pass.h | 16 ++++++++++------ 1 file changed, 10 insertions(+), 6 deletions(-) diff --git a/tensorflow/compiler/mlir/tensorflow/transforms/lift_variables_test_pass.h b/tensorflow/compiler/mlir/tensorflow/transforms/lift_variables_test_pass.h index faecdf04368..0e6d844bed3 100644 --- a/tensorflow/compiler/mlir/tensorflow/transforms/lift_variables_test_pass.h +++ b/tensorflow/compiler/mlir/tensorflow/transforms/lift_variables_test_pass.h @@ -96,15 +96,19 @@ class FakeSession : public tensorflow::Session { for (const std::string& output_name : output_names) { Tensor output; if (output_name == "dense/bias") { - outputs->push_back( - Tensor(tensorflow::DT_FLOAT, tensorflow::TensorShape({50}))); + Tensor t = Tensor(tensorflow::DT_FLOAT, tensorflow::TensorShape({50})); + t.flat().setZero(); + outputs->push_back(t); } else if (output_name == "dense/kernel") { - outputs->push_back( - Tensor(tensorflow::DT_FLOAT, tensorflow::TensorShape({100, 50}))); + Tensor t = + Tensor(tensorflow::DT_FLOAT, tensorflow::TensorShape({100, 50})); + t.flat().setZero(); + outputs->push_back(t); } else { // Create a scalar float tensor. - outputs->push_back( - Tensor(tensorflow::DT_FLOAT, tensorflow::TensorShape({}))); + Tensor t = Tensor(tensorflow::DT_FLOAT, tensorflow::TensorShape({})); + t.flat()(0) = 1.0f; + outputs->push_back(t); } } return Status::OK(); From 7a928592460e51c453ebaaecf965ef9ec963890b Mon Sep 17 00:00:00 2001 From: Ken Franko Date: Wed, 17 Jun 2020 15:06:41 -0700 Subject: [PATCH 123/867] Add SaveOptions/CheckpointOptions to keras.Models.save_weights and keras_call_backs.ModelCheckpoint. PiperOrigin-RevId: 316973333 Change-Id: I43f5b59ece4b862db41ab0e99f3c8df0a0d3b901 --- tensorflow/python/keras/callbacks.py | 30 +++++++++++++--- tensorflow/python/keras/callbacks_test.py | 34 +++++++++++++++++++ tensorflow/python/keras/engine/training.py | 10 ++++-- .../golden/v1/tensorflow.keras.-model.pbtxt | 2 +- .../v1/tensorflow.keras.-sequential.pbtxt | 2 +- ...ow.keras.callbacks.-model-checkpoint.pbtxt | 2 +- ...low.keras.experimental.-linear-model.pbtxt | 2 +- ....keras.experimental.-wide-deep-model.pbtxt | 2 +- .../v1/tensorflow.keras.models.-model.pbtxt | 2 +- .../tensorflow.keras.models.-sequential.pbtxt | 2 +- .../golden/v2/tensorflow.keras.-model.pbtxt | 2 +- .../v2/tensorflow.keras.-sequential.pbtxt | 2 +- ...ow.keras.callbacks.-model-checkpoint.pbtxt | 2 +- ...low.keras.experimental.-linear-model.pbtxt | 2 +- ....keras.experimental.-wide-deep-model.pbtxt | 2 +- .../v2/tensorflow.keras.models.-model.pbtxt | 2 +- .../tensorflow.keras.models.-sequential.pbtxt | 2 +- 17 files changed, 82 insertions(+), 20 deletions(-) diff --git a/tensorflow/python/keras/callbacks.py b/tensorflow/python/keras/callbacks.py index 1bca5419774..1fae5abd84b 100644 --- a/tensorflow/python/keras/callbacks.py +++ b/tensorflow/python/keras/callbacks.py @@ -54,7 +54,9 @@ from tensorflow.python.ops import math_ops from tensorflow.python.ops import summary_ops_v2 from tensorflow.python.platform import tf_logging as logging from tensorflow.python.profiler import profiler_v2 as profiler +from tensorflow.python.saved_model import save_options as save_options_lib from tensorflow.python.training import checkpoint_management +from tensorflow.python.training.saving import checkpoint_options as checkpoint_options_lib from tensorflow.python.util import nest from tensorflow.python.util.compat import collections_abc from tensorflow.python.util.tf_export import keras_export @@ -1115,6 +1117,9 @@ class ModelCheckpoint(Callback): epochs, the monitored metric may potentially be less reliable (it could reflect as little as 1 batch, since the metrics get reset every epoch). Defaults to `'epoch'`. + options: Optional `tf.train.CheckpointOptions` object if + `save_weights_only` is true or optional `tf.saved_model.SavedOptions` + object if `save_weights_only` is false. **kwargs: Additional arguments for backwards compatibility. Possible key is `period`. """ @@ -1127,6 +1132,7 @@ class ModelCheckpoint(Callback): save_weights_only=False, mode='auto', save_freq='epoch', + options=None, **kwargs): super(ModelCheckpoint, self).__init__() self._supports_tf_logs = True @@ -1140,6 +1146,20 @@ class ModelCheckpoint(Callback): self._batches_seen_since_last_saving = 0 self._last_batch_seen = 0 + if save_weights_only: + if options is None or isinstance( + options, checkpoint_options_lib.CheckpointOptions): + self._options = options or checkpoint_options_lib.CheckpointOptions() + else: + raise TypeError('If save_weights_only is True, then `options` must be' + 'either None or a tf.train.CheckpointOptions') + else: + if options is None or isinstance(options, save_options_lib.SaveOptions): + self._options = options or save_options_lib.SaveOptions() + else: + raise TypeError('If save_weights_only is False, then `options` must be' + 'either None or a tf.saved_model.SaveOptions') + # Deprecated field `load_weights_on_restart` is for loading the checkpoint # file from `filepath` at the start of `model.fit()` # TODO(rchao): Remove the arg during next breaking release. @@ -1269,9 +1289,10 @@ class ModelCheckpoint(Callback): self.best, current, filepath)) self.best = current if self.save_weights_only: - self.model.save_weights(filepath, overwrite=True) + self.model.save_weights( + filepath, overwrite=True, options=self._options) else: - self.model.save(filepath, overwrite=True) + self.model.save(filepath, overwrite=True, options=self._options) else: if self.verbose > 0: print('\nEpoch %05d: %s did not improve from %0.5f' % @@ -1280,9 +1301,10 @@ class ModelCheckpoint(Callback): if self.verbose > 0: print('\nEpoch %05d: saving model to %s' % (epoch + 1, filepath)) if self.save_weights_only: - self.model.save_weights(filepath, overwrite=True) + self.model.save_weights( + filepath, overwrite=True, options=self._options) else: - self.model.save(filepath, overwrite=True) + self.model.save(filepath, overwrite=True, options=self._options) self._maybe_remove_file() except IOError as e: diff --git a/tensorflow/python/keras/callbacks_test.py b/tensorflow/python/keras/callbacks_test.py index 28f85304688..d180e85a1d9 100644 --- a/tensorflow/python/keras/callbacks_test.py +++ b/tensorflow/python/keras/callbacks_test.py @@ -49,9 +49,11 @@ from tensorflow.python.ops import math_ops from tensorflow.python.ops import summary_ops_v2 from tensorflow.python.platform import test from tensorflow.python.platform import tf_logging as logging +from tensorflow.python.saved_model import save_options as save_options_lib from tensorflow.python.summary import summary_iterator from tensorflow.python.training import adam from tensorflow.python.training import checkpoint_management +from tensorflow.python.training.saving import checkpoint_options as checkpoint_options_lib try: import h5py # pylint:disable=g-import-not-at-top @@ -666,6 +668,38 @@ class KerasCallbacksTest(keras_parameterized.TestCase): mode=mode, save_freq=3) + # Case 9: `ModelCheckpoint` with valid and invalid `options` argument. + with self.assertRaisesRegexp(TypeError, 'tf.train.CheckpointOptions'): + keras.callbacks.ModelCheckpoint( + filepath, + monitor=monitor, + save_best_only=save_best_only, + save_weights_only=True, + mode=mode, + options=save_options_lib.SaveOptions()) + with self.assertRaisesRegexp(TypeError, 'tf.saved_model.SaveOptions'): + keras.callbacks.ModelCheckpoint( + filepath, + monitor=monitor, + save_best_only=save_best_only, + save_weights_only=False, + mode=mode, + options=checkpoint_options_lib.CheckpointOptions()) + keras.callbacks.ModelCheckpoint( + filepath, + monitor=monitor, + save_best_only=save_best_only, + save_weights_only=True, + mode=mode, + options=checkpoint_options_lib.CheckpointOptions()) + keras.callbacks.ModelCheckpoint( + filepath, + monitor=monitor, + save_best_only=save_best_only, + save_weights_only=False, + mode=mode, + options=save_options_lib.SaveOptions()) + def _get_dummy_resource_for_model_checkpoint_testing(self): def get_input_datasets(): diff --git a/tensorflow/python/keras/engine/training.py b/tensorflow/python/keras/engine/training.py index 5567e1733a7..ccd184a8bc4 100644 --- a/tensorflow/python/keras/engine/training.py +++ b/tensorflow/python/keras/engine/training.py @@ -1979,7 +1979,11 @@ class Model(base_layer.Layer, version_utils.ModelVersionSelector): save.save_model(self, filepath, overwrite, include_optimizer, save_format, signatures, options) - def save_weights(self, filepath, overwrite=True, save_format=None): + def save_weights(self, + filepath, + overwrite=True, + save_format=None, + options=None): """Saves all layer weights. Either saves in HDF5 or in TensorFlow format based on the `save_format` @@ -2032,6 +2036,8 @@ class Model(base_layer.Layer, version_utils.ModelVersionSelector): save_format: Either 'tf' or 'h5'. A `filepath` ending in '.h5' or '.keras' will default to HDF5 if `save_format` is `None`. Otherwise `None` defaults to 'tf'. + options: Optional `tf.train.CheckpointOptions` object that specifies + options for saving weights. Raises: ImportError: If h5py is not available when attempting to save in HDF5 @@ -2093,7 +2099,7 @@ class Model(base_layer.Layer, version_utils.ModelVersionSelector): 'the TensorFlow format the optimizer\'s state will not be ' 'saved.\n\nConsider using a TensorFlow optimizer from `tf.train`.') % (optimizer,)) - self._trackable_saver.save(filepath, session=session) + self._trackable_saver.save(filepath, session=session, options=options) # Record this checkpoint so it's visible from tf.train.latest_checkpoint. checkpoint_management.update_checkpoint_state_internal( save_dir=os.path.dirname(filepath), diff --git a/tensorflow/tools/api/golden/v1/tensorflow.keras.-model.pbtxt b/tensorflow/tools/api/golden/v1/tensorflow.keras.-model.pbtxt index b62814e81cb..6318e577087 100644 --- a/tensorflow/tools/api/golden/v1/tensorflow.keras.-model.pbtxt +++ b/tensorflow/tools/api/golden/v1/tensorflow.keras.-model.pbtxt @@ -302,7 +302,7 @@ tf_class { } member_method { name: "save_weights" - argspec: "args=[\'self\', \'filepath\', \'overwrite\', \'save_format\'], varargs=None, keywords=None, defaults=[\'True\', \'None\'], " + argspec: "args=[\'self\', \'filepath\', \'overwrite\', \'save_format\', \'options\'], varargs=None, keywords=None, defaults=[\'True\', \'None\', \'None\'], " } member_method { name: "set_weights" diff --git a/tensorflow/tools/api/golden/v1/tensorflow.keras.-sequential.pbtxt b/tensorflow/tools/api/golden/v1/tensorflow.keras.-sequential.pbtxt index 7485a0b3c62..9b7b7736746 100644 --- a/tensorflow/tools/api/golden/v1/tensorflow.keras.-sequential.pbtxt +++ b/tensorflow/tools/api/golden/v1/tensorflow.keras.-sequential.pbtxt @@ -320,7 +320,7 @@ tf_class { } member_method { name: "save_weights" - argspec: "args=[\'self\', \'filepath\', \'overwrite\', \'save_format\'], varargs=None, keywords=None, defaults=[\'True\', \'None\'], " + argspec: "args=[\'self\', \'filepath\', \'overwrite\', \'save_format\', \'options\'], varargs=None, keywords=None, defaults=[\'True\', \'None\', \'None\'], " } member_method { name: "set_weights" diff --git a/tensorflow/tools/api/golden/v1/tensorflow.keras.callbacks.-model-checkpoint.pbtxt b/tensorflow/tools/api/golden/v1/tensorflow.keras.callbacks.-model-checkpoint.pbtxt index 5fb646e1c63..e6cc7aee5a8 100644 --- a/tensorflow/tools/api/golden/v1/tensorflow.keras.callbacks.-model-checkpoint.pbtxt +++ b/tensorflow/tools/api/golden/v1/tensorflow.keras.callbacks.-model-checkpoint.pbtxt @@ -5,7 +5,7 @@ tf_class { is_instance: "" member_method { name: "__init__" - argspec: "args=[\'self\', \'filepath\', \'monitor\', \'verbose\', \'save_best_only\', \'save_weights_only\', \'mode\', \'save_freq\'], varargs=None, keywords=kwargs, defaults=[\'val_loss\', \'0\', \'False\', \'False\', \'auto\', \'epoch\'], " + argspec: "args=[\'self\', \'filepath\', \'monitor\', \'verbose\', \'save_best_only\', \'save_weights_only\', \'mode\', \'save_freq\', \'options\'], varargs=None, keywords=kwargs, defaults=[\'val_loss\', \'0\', \'False\', \'False\', \'auto\', \'epoch\', \'None\'], " } member_method { name: "on_batch_begin" diff --git a/tensorflow/tools/api/golden/v1/tensorflow.keras.experimental.-linear-model.pbtxt b/tensorflow/tools/api/golden/v1/tensorflow.keras.experimental.-linear-model.pbtxt index bf980e5d116..976eb49d4c8 100644 --- a/tensorflow/tools/api/golden/v1/tensorflow.keras.experimental.-linear-model.pbtxt +++ b/tensorflow/tools/api/golden/v1/tensorflow.keras.experimental.-linear-model.pbtxt @@ -303,7 +303,7 @@ tf_class { } member_method { name: "save_weights" - argspec: "args=[\'self\', \'filepath\', \'overwrite\', \'save_format\'], varargs=None, keywords=None, defaults=[\'True\', \'None\'], " + argspec: "args=[\'self\', \'filepath\', \'overwrite\', \'save_format\', \'options\'], varargs=None, keywords=None, defaults=[\'True\', \'None\', \'None\'], " } member_method { name: "set_weights" diff --git a/tensorflow/tools/api/golden/v1/tensorflow.keras.experimental.-wide-deep-model.pbtxt b/tensorflow/tools/api/golden/v1/tensorflow.keras.experimental.-wide-deep-model.pbtxt index c214a5c3419..500aa28eae7 100644 --- a/tensorflow/tools/api/golden/v1/tensorflow.keras.experimental.-wide-deep-model.pbtxt +++ b/tensorflow/tools/api/golden/v1/tensorflow.keras.experimental.-wide-deep-model.pbtxt @@ -303,7 +303,7 @@ tf_class { } member_method { name: "save_weights" - argspec: "args=[\'self\', \'filepath\', \'overwrite\', \'save_format\'], varargs=None, keywords=None, defaults=[\'True\', \'None\'], " + argspec: "args=[\'self\', \'filepath\', \'overwrite\', \'save_format\', \'options\'], varargs=None, keywords=None, defaults=[\'True\', \'None\', \'None\'], " } member_method { name: "set_weights" diff --git a/tensorflow/tools/api/golden/v1/tensorflow.keras.models.-model.pbtxt b/tensorflow/tools/api/golden/v1/tensorflow.keras.models.-model.pbtxt index 86868c9d17f..ad0edc64606 100644 --- a/tensorflow/tools/api/golden/v1/tensorflow.keras.models.-model.pbtxt +++ b/tensorflow/tools/api/golden/v1/tensorflow.keras.models.-model.pbtxt @@ -302,7 +302,7 @@ tf_class { } member_method { name: "save_weights" - argspec: "args=[\'self\', \'filepath\', \'overwrite\', \'save_format\'], varargs=None, keywords=None, defaults=[\'True\', \'None\'], " + argspec: "args=[\'self\', \'filepath\', \'overwrite\', \'save_format\', \'options\'], varargs=None, keywords=None, defaults=[\'True\', \'None\', \'None\'], " } member_method { name: "set_weights" diff --git a/tensorflow/tools/api/golden/v1/tensorflow.keras.models.-sequential.pbtxt b/tensorflow/tools/api/golden/v1/tensorflow.keras.models.-sequential.pbtxt index 05aa19a915a..b38c669df0f 100644 --- a/tensorflow/tools/api/golden/v1/tensorflow.keras.models.-sequential.pbtxt +++ b/tensorflow/tools/api/golden/v1/tensorflow.keras.models.-sequential.pbtxt @@ -320,7 +320,7 @@ tf_class { } member_method { name: "save_weights" - argspec: "args=[\'self\', \'filepath\', \'overwrite\', \'save_format\'], varargs=None, keywords=None, defaults=[\'True\', \'None\'], " + argspec: "args=[\'self\', \'filepath\', \'overwrite\', \'save_format\', \'options\'], varargs=None, keywords=None, defaults=[\'True\', \'None\', \'None\'], " } member_method { name: "set_weights" diff --git a/tensorflow/tools/api/golden/v2/tensorflow.keras.-model.pbtxt b/tensorflow/tools/api/golden/v2/tensorflow.keras.-model.pbtxt index b62814e81cb..6318e577087 100644 --- a/tensorflow/tools/api/golden/v2/tensorflow.keras.-model.pbtxt +++ b/tensorflow/tools/api/golden/v2/tensorflow.keras.-model.pbtxt @@ -302,7 +302,7 @@ tf_class { } member_method { name: "save_weights" - argspec: "args=[\'self\', \'filepath\', \'overwrite\', \'save_format\'], varargs=None, keywords=None, defaults=[\'True\', \'None\'], " + argspec: "args=[\'self\', \'filepath\', \'overwrite\', \'save_format\', \'options\'], varargs=None, keywords=None, defaults=[\'True\', \'None\', \'None\'], " } member_method { name: "set_weights" diff --git a/tensorflow/tools/api/golden/v2/tensorflow.keras.-sequential.pbtxt b/tensorflow/tools/api/golden/v2/tensorflow.keras.-sequential.pbtxt index 7485a0b3c62..9b7b7736746 100644 --- a/tensorflow/tools/api/golden/v2/tensorflow.keras.-sequential.pbtxt +++ b/tensorflow/tools/api/golden/v2/tensorflow.keras.-sequential.pbtxt @@ -320,7 +320,7 @@ tf_class { } member_method { name: "save_weights" - argspec: "args=[\'self\', \'filepath\', \'overwrite\', \'save_format\'], varargs=None, keywords=None, defaults=[\'True\', \'None\'], " + argspec: "args=[\'self\', \'filepath\', \'overwrite\', \'save_format\', \'options\'], varargs=None, keywords=None, defaults=[\'True\', \'None\', \'None\'], " } member_method { name: "set_weights" diff --git a/tensorflow/tools/api/golden/v2/tensorflow.keras.callbacks.-model-checkpoint.pbtxt b/tensorflow/tools/api/golden/v2/tensorflow.keras.callbacks.-model-checkpoint.pbtxt index 5fb646e1c63..e6cc7aee5a8 100644 --- a/tensorflow/tools/api/golden/v2/tensorflow.keras.callbacks.-model-checkpoint.pbtxt +++ b/tensorflow/tools/api/golden/v2/tensorflow.keras.callbacks.-model-checkpoint.pbtxt @@ -5,7 +5,7 @@ tf_class { is_instance: "" member_method { name: "__init__" - argspec: "args=[\'self\', \'filepath\', \'monitor\', \'verbose\', \'save_best_only\', \'save_weights_only\', \'mode\', \'save_freq\'], varargs=None, keywords=kwargs, defaults=[\'val_loss\', \'0\', \'False\', \'False\', \'auto\', \'epoch\'], " + argspec: "args=[\'self\', \'filepath\', \'monitor\', \'verbose\', \'save_best_only\', \'save_weights_only\', \'mode\', \'save_freq\', \'options\'], varargs=None, keywords=kwargs, defaults=[\'val_loss\', \'0\', \'False\', \'False\', \'auto\', \'epoch\', \'None\'], " } member_method { name: "on_batch_begin" diff --git a/tensorflow/tools/api/golden/v2/tensorflow.keras.experimental.-linear-model.pbtxt b/tensorflow/tools/api/golden/v2/tensorflow.keras.experimental.-linear-model.pbtxt index bf980e5d116..976eb49d4c8 100644 --- a/tensorflow/tools/api/golden/v2/tensorflow.keras.experimental.-linear-model.pbtxt +++ b/tensorflow/tools/api/golden/v2/tensorflow.keras.experimental.-linear-model.pbtxt @@ -303,7 +303,7 @@ tf_class { } member_method { name: "save_weights" - argspec: "args=[\'self\', \'filepath\', \'overwrite\', \'save_format\'], varargs=None, keywords=None, defaults=[\'True\', \'None\'], " + argspec: "args=[\'self\', \'filepath\', \'overwrite\', \'save_format\', \'options\'], varargs=None, keywords=None, defaults=[\'True\', \'None\', \'None\'], " } member_method { name: "set_weights" diff --git a/tensorflow/tools/api/golden/v2/tensorflow.keras.experimental.-wide-deep-model.pbtxt b/tensorflow/tools/api/golden/v2/tensorflow.keras.experimental.-wide-deep-model.pbtxt index c214a5c3419..500aa28eae7 100644 --- a/tensorflow/tools/api/golden/v2/tensorflow.keras.experimental.-wide-deep-model.pbtxt +++ b/tensorflow/tools/api/golden/v2/tensorflow.keras.experimental.-wide-deep-model.pbtxt @@ -303,7 +303,7 @@ tf_class { } member_method { name: "save_weights" - argspec: "args=[\'self\', \'filepath\', \'overwrite\', \'save_format\'], varargs=None, keywords=None, defaults=[\'True\', \'None\'], " + argspec: "args=[\'self\', \'filepath\', \'overwrite\', \'save_format\', \'options\'], varargs=None, keywords=None, defaults=[\'True\', \'None\', \'None\'], " } member_method { name: "set_weights" diff --git a/tensorflow/tools/api/golden/v2/tensorflow.keras.models.-model.pbtxt b/tensorflow/tools/api/golden/v2/tensorflow.keras.models.-model.pbtxt index 86868c9d17f..ad0edc64606 100644 --- a/tensorflow/tools/api/golden/v2/tensorflow.keras.models.-model.pbtxt +++ b/tensorflow/tools/api/golden/v2/tensorflow.keras.models.-model.pbtxt @@ -302,7 +302,7 @@ tf_class { } member_method { name: "save_weights" - argspec: "args=[\'self\', \'filepath\', \'overwrite\', \'save_format\'], varargs=None, keywords=None, defaults=[\'True\', \'None\'], " + argspec: "args=[\'self\', \'filepath\', \'overwrite\', \'save_format\', \'options\'], varargs=None, keywords=None, defaults=[\'True\', \'None\', \'None\'], " } member_method { name: "set_weights" diff --git a/tensorflow/tools/api/golden/v2/tensorflow.keras.models.-sequential.pbtxt b/tensorflow/tools/api/golden/v2/tensorflow.keras.models.-sequential.pbtxt index 05aa19a915a..b38c669df0f 100644 --- a/tensorflow/tools/api/golden/v2/tensorflow.keras.models.-sequential.pbtxt +++ b/tensorflow/tools/api/golden/v2/tensorflow.keras.models.-sequential.pbtxt @@ -320,7 +320,7 @@ tf_class { } member_method { name: "save_weights" - argspec: "args=[\'self\', \'filepath\', \'overwrite\', \'save_format\'], varargs=None, keywords=None, defaults=[\'True\', \'None\'], " + argspec: "args=[\'self\', \'filepath\', \'overwrite\', \'save_format\', \'options\'], varargs=None, keywords=None, defaults=[\'True\', \'None\', \'None\'], " } member_method { name: "set_weights" From b186ba0334e98ebbee5eacf7fc1c66897eaa19bc Mon Sep 17 00:00:00 2001 From: Li Lao Date: Wed, 17 Jun 2020 15:21:23 -0700 Subject: [PATCH 124/867] Add TraceMeProducer/Consumer for SharedBatchScheduler. PiperOrigin-RevId: 316976310 Change-Id: I1cd2a03390aedd7e8e85b2826ab3aadd096bafdd --- tensorflow/core/kernels/batching_util/BUILD | 2 ++ .../kernels/batching_util/batch_scheduler.h | 21 ++++++++++++++++++- .../batching_util/shared_batch_scheduler.h | 18 +++++++++++----- .../core/profiler/lib/connected_traceme.h | 1 + 4 files changed, 36 insertions(+), 6 deletions(-) diff --git a/tensorflow/core/kernels/batching_util/BUILD b/tensorflow/core/kernels/batching_util/BUILD index 370c96c6e7f..803eb2e9048 100644 --- a/tensorflow/core/kernels/batching_util/BUILD +++ b/tensorflow/core/kernels/batching_util/BUILD @@ -70,6 +70,7 @@ cc_library( ":batch_scheduler_hdrs", ":periodic_function_dynamic", "//tensorflow/core:framework_headers_lib", + "//tensorflow/core/profiler/lib:connected_traceme", "//tensorflow/core/profiler/lib:traceme", ], ) @@ -81,6 +82,7 @@ cc_library( ":batch_scheduler", ":periodic_function_dynamic", "//tensorflow/core:lib", + "//tensorflow/core/profiler/lib:connected_traceme", "//tensorflow/core/profiler/lib:traceme", ], alwayslink = 1, diff --git a/tensorflow/core/kernels/batching_util/batch_scheduler.h b/tensorflow/core/kernels/batching_util/batch_scheduler.h index e418f8acbb1..d0e1d20bed4 100644 --- a/tensorflow/core/kernels/batching_util/batch_scheduler.h +++ b/tensorflow/core/kernels/batching_util/batch_scheduler.h @@ -77,7 +77,8 @@ class BatchTask { template class Batch { public: - Batch() = default; + Batch(); + explicit Batch(uint64 traceme_context_id); virtual ~Batch(); // Blocks until the batch is closed. // Appends 'task' to the batch. After calling AddTask(), the newly-added task @@ -113,6 +114,9 @@ class Batch { // Marks the batch as closed. Dies if called more than once. void Close(); + // Returns the TraceMe context id of this batch. + uint64 traceme_context_id() const; + private: mutable mutex mu_; @@ -125,6 +129,9 @@ class Batch { // Whether the batch has been closed. Notification closed_; + // The TracMe context id. + const uint64 traceme_context_id_; + TF_DISALLOW_COPY_AND_ASSIGN(Batch); }; @@ -187,6 +194,13 @@ class BatchScheduler { ////////// // Implementation details follow. API users need not read. +template +Batch::Batch() : Batch(0) {} + +template +Batch::Batch(uint64 traceme_context_id) + : traceme_context_id_(traceme_context_id) {} + template Batch::~Batch() { WaitUntilClosed(); @@ -275,6 +289,11 @@ void Batch::Close() { closed_.Notify(); } +template +uint64 Batch::traceme_context_id() const { + return traceme_context_id_; +} + } // namespace serving } // namespace tensorflow diff --git a/tensorflow/core/kernels/batching_util/shared_batch_scheduler.h b/tensorflow/core/kernels/batching_util/shared_batch_scheduler.h index 66bdff933d8..e47e069eff5 100644 --- a/tensorflow/core/kernels/batching_util/shared_batch_scheduler.h +++ b/tensorflow/core/kernels/batching_util/shared_batch_scheduler.h @@ -36,6 +36,7 @@ limitations under the License. #include "tensorflow/core/platform/env.h" #include "tensorflow/core/platform/thread_annotations.h" #include "tensorflow/core/platform/types.h" +#include "tensorflow/core/profiler/lib/connected_traceme.h" #include "tensorflow/core/profiler/lib/traceme.h" namespace tensorflow { @@ -311,6 +312,9 @@ class Queue { // The enqueued batches. See the invariants in the class comments above. std::deque>> batches_ TF_GUARDED_BY(mu_); + // The counter of the TraceMe context ids. + uint64 traceme_context_id_counter_ TF_GUARDED_BY(mu_) = 0; + // The time at which the first task was added to the open (back-most) batch // in 'batches_'. Valid iff that batch contains at least one task. uint64 open_batch_start_time_micros_ TF_GUARDED_BY(mu_); @@ -529,8 +533,6 @@ Queue::~Queue() { template Status Queue::Schedule(std::unique_ptr* task) { - profiler::TraceMe trace_me( - [task] { return strings::StrCat("Schedule:", (*task)->size()); }); if ((*task)->size() > options_.max_batch_size) { return errors::InvalidArgument("Task size ", (*task)->size(), " is larger than maximum batch size ", @@ -554,6 +556,10 @@ Status Queue::Schedule(std::unique_ptr* task) { if (batches_.back()->empty()) { open_batch_start_time_micros_ = env_->NowMicros(); } + profiler::TraceMeProducer trace_me( + [&] { return strings::StrCat("Schedule:", (*task)->size()); }, + profiler::ContextType::kSharedBatchScheduler, + batches_.back()->traceme_context_id()); batches_.back()->AddTask(std::move(*task)); if (!schedulable_batch_) { @@ -621,8 +627,10 @@ std::unique_ptr> Queue::ScheduleBatch() { template void Queue::ProcessBatch(std::unique_ptr> batch) { - profiler::TraceMe trace_me( - [&batch] { return strings::StrCat("ProcessBatch:", batch->size()); }); + profiler::TraceMeConsumer trace_me( + [&batch] { return strings::StrCat("ProcessBatch:", batch->size()); }, + profiler::ContextType::kSharedBatchScheduler, + batch->traceme_context_id()); process_batch_callback_(std::move(batch)); { @@ -665,7 +673,7 @@ bool Queue::IsEmptyInternal() const { template void Queue::StartNewBatch() { batches_.back()->Close(); - batches_.emplace_back(new Batch); + batches_.emplace_back(new Batch(++traceme_context_id_counter_)); } template diff --git a/tensorflow/core/profiler/lib/connected_traceme.h b/tensorflow/core/profiler/lib/connected_traceme.h index cbc610af407..ed8b4ac1ad2 100644 --- a/tensorflow/core/profiler/lib/connected_traceme.h +++ b/tensorflow/core/profiler/lib/connected_traceme.h @@ -28,6 +28,7 @@ namespace profiler { enum class ContextType : int { kGeneric, kTfExecutor, + kSharedBatchScheduler, }; /* From 19c51afbf1fd2f724cabc42313ce84e31a4defcc Mon Sep 17 00:00:00 2001 From: Jaesung Chung Date: Wed, 17 Jun 2020 15:42:14 -0700 Subject: [PATCH 125/867] Do not emit op errors on TF -> TFL legalization Applying legalization patterns will emit unwanted, transient errors when the replaced TFLite ops do not meet the sanity checks. In order to ignore the transient errors, the following lines override a diagnostic handler with an no-op handler only while this pass runs. PiperOrigin-RevId: 316980278 Change-Id: Idef14e13f36ff0ee3c4bb1a401f92ba217042dbb --- .../compiler/mlir/lite/converter_gen.cc | 18 +++++----------- .../mlir/lite/ir/tfl_op_interfaces.td | 3 +-- .../compiler/mlir/lite/tests/legalize-tf.mlir | 7 +++++++ .../mlir/lite/transforms/legalize_tf.cc | 21 ++++++++++++++++--- .../mlir/lite/transforms/runtime_verify.cc | 3 +-- 5 files changed, 32 insertions(+), 20 deletions(-) diff --git a/tensorflow/compiler/mlir/lite/converter_gen.cc b/tensorflow/compiler/mlir/lite/converter_gen.cc index 6df569a8031..edead2037a3 100644 --- a/tensorflow/compiler/mlir/lite/converter_gen.cc +++ b/tensorflow/compiler/mlir/lite/converter_gen.cc @@ -446,7 +446,7 @@ static void GenOperandResultVerifier(raw_ostream &os, auto desc = definit->getDef()->getValueAsString("tflRuntimeTypeDescription"); - // Emit a loop to check all the dynamic values in the pack. + // Emit a loop to check all operands. os << formatv(" for (Value v : top.getODS{0}{1}s({2})) {{\n", // Capitalize the first letter to match the function name valueKind.substr(0, 1).upper(), valueKind.substr(1), @@ -455,14 +455,10 @@ static void GenOperandResultVerifier(raw_ostream &os, os << " (void)v;\n" << " if (!(" << tgfmt(pred.getCondition(), &fctx.withSelf("v.getType()")) << ")) {\n" - << " if (failure_on_operand_type_mismatch) {\n" << formatv( - " return op->emitOpError(\"{0} #\") << index " + " return op->emitOpError(\"{0} #\") << index " "<< \" must be {1}, but got \" << v.getType();\n", valueKind, desc) - << " } else {\n" - << " return ::mlir::LogicalResult::Failure;\n" - << " }\n" << " }\n" // if << " ++index;\n" << " }\n"; // for @@ -487,8 +483,7 @@ static bool RuntimeVerifierWriterMain(raw_ostream &os, RecordKeeper &records) { mlir::tblgen::FmtContext verify_ctx; os << "::mlir::LogicalResult " << op.getCppClassName() - << "::VerifyTflRuntimeConstraints(::mlir::Operation *op, bool " - "failure_on_operand_type_mismatch) {\n"; + << "::VerifyTflRuntimeConstraints(::mlir::Operation *op) {\n"; os << " auto top = cast<" << op.getCppClassName() << ">(op); (void)top;\n"; verify_ctx.withOp("top"); @@ -529,11 +524,8 @@ static bool RuntimeVerifierWriterMain(raw_ostream &os, RecordKeeper &records) { mlir::tblgen::Pred pred(dyn_cast(val->getValue())); os << tgfmt( - " if (!($0)) {\n " - " if (failure_on_operand_type_mismatch) {\n" - " return top.emitOpError(\"failed to verify that $1\");\n" - " } else {\n" - " return ::mlir::LogicalResult::Failure;\n }\n }\n", + " if (!($0))\n" + " return top.emitOpError(\"failed to verify that $1\");\n", &verify_ctx, tgfmt(pred.getCondition(), &verify_ctx), desc); } os << " return top.verify();\n}\n"; diff --git a/tensorflow/compiler/mlir/lite/ir/tfl_op_interfaces.td b/tensorflow/compiler/mlir/lite/ir/tfl_op_interfaces.td index 23101113a6f..a79d79b5970 100644 --- a/tensorflow/compiler/mlir/lite/ir/tfl_op_interfaces.td +++ b/tensorflow/compiler/mlir/lite/ir/tfl_op_interfaces.td @@ -94,8 +94,7 @@ def TFL_RuntimeVerification : OpInterface<"TflRuntimeVerifyOpInterface"> { let methods = [ StaticInterfaceMethod< [{Returns whether the op's operands/results are supported by runtime.}], - "LogicalResult", "VerifyTflRuntimeConstraints", - (ins "Operation*":$op, "bool":$failure_on_operand_type_mismatch) + "LogicalResult", "VerifyTflRuntimeConstraints", (ins "Operation*":$op) >, ]; } diff --git a/tensorflow/compiler/mlir/lite/tests/legalize-tf.mlir b/tensorflow/compiler/mlir/lite/tests/legalize-tf.mlir index 1ae789f5468..5756fa6dec2 100644 --- a/tensorflow/compiler/mlir/lite/tests/legalize-tf.mlir +++ b/tensorflow/compiler/mlir/lite/tests/legalize-tf.mlir @@ -990,6 +990,13 @@ func @batch_to_space_nd(%arg0: tensor<4x2x2x3xf32>, %arg1: tensor<2xi32>, %arg2: // CHECK: "tfl.batch_to_space_nd"(%arg0, %arg1, %arg2) : (tensor<4x2x2x3xf32>, tensor<2xi32>, tensor<2x2xi32>) -> tensor } +func @batch_to_space_nd_unsupported(%arg0: tensor, %arg1: tensor<3xi32>, %arg2: tensor<3x2xi32>) -> tensor { + %0 = "tf.BatchToSpaceND"(%arg0, %arg1, %arg2) : (tensor, tensor<3xi32>, tensor<3x2xi32>) -> tensor + return %0 : tensor + // CHECK-LABEL: batch_to_space_nd_unsupported + // CHECK: "tf.BatchToSpaceND" +} + func @space_to_batch_nd(%arg0: tensor<1x4x4x3xf32>, %arg1: tensor<2xi32>, %arg2: tensor<2x2xi32>) -> tensor { %0 = "tf.SpaceToBatchND"(%arg0, %arg1, %arg2) : (tensor<1x4x4x3xf32>, tensor<2xi32>, tensor<2x2xi32>) -> tensor return %0 : tensor diff --git a/tensorflow/compiler/mlir/lite/transforms/legalize_tf.cc b/tensorflow/compiler/mlir/lite/transforms/legalize_tf.cc index 46ed134d7ee..1328a2baf5d 100644 --- a/tensorflow/compiler/mlir/lite/transforms/legalize_tf.cc +++ b/tensorflow/compiler/mlir/lite/transforms/legalize_tf.cc @@ -28,9 +28,11 @@ limitations under the License. #include "llvm/ADT/APInt.h" #include "llvm/ADT/ArrayRef.h" #include "llvm/ADT/StringSwitch.h" +#include "llvm/Support/Threading.h" #include "mlir/Dialect/Quant/FakeQuantSupport.h" // from @llvm-project #include "mlir/Dialect/Quant/UniformSupport.h" // from @llvm-project #include "mlir/IR/Attributes.h" // from @llvm-project +#include "mlir/IR/Diagnostics.h" // from @llvm-project #include "mlir/IR/MLIRContext.h" // from @llvm-project #include "mlir/IR/Operation.h" // from @llvm-project #include "mlir/IR/PatternMatch.h" // from @llvm-project @@ -767,13 +769,26 @@ void LegalizeTF::runOnFunction() { [](Operation* op) { auto tfl_op = dyn_cast_or_null(op); if (!tfl_op) return false; - return succeeded(tfl_op.VerifyTflRuntimeConstraints( - tfl_op.getOperation(), - /*failure_on_operand_type_mismatch=*/false)); + return succeeded(tfl_op.VerifyTflRuntimeConstraints(op)); })); } else { target.addLegalDialect(); } + + // Ignore transient errors by registering an no-op handler. + // Applying legalization patterns will emit unwanted, transient errors when + // the replaced TFLite ops do not meet the sanity checks. In order to ignore + // the transient errors, the following lines override a diagnostic handler + // with an no-op handler only while this pass runs. + uint64_t current_thread_id = llvm::get_threadid(); + ScopedDiagnosticHandler scoped_diag_handler( + context, [¤t_thread_id](Diagnostic&) -> LogicalResult { + // Consume only errors that are coming from the same thread in order not + // to ignore errors from other passes that are running. Things running + // in the pass manager can be multi-threaded. + return success(current_thread_id == llvm::get_threadid()); + }); + // Keep trying to convert. // TODO(karimnosseir): This is similar to what apply greedy patterns does. // Look if there is a function that tries until it converge. diff --git a/tensorflow/compiler/mlir/lite/transforms/runtime_verify.cc b/tensorflow/compiler/mlir/lite/transforms/runtime_verify.cc index 3268329b1c1..cc2e691180e 100644 --- a/tensorflow/compiler/mlir/lite/transforms/runtime_verify.cc +++ b/tensorflow/compiler/mlir/lite/transforms/runtime_verify.cc @@ -34,8 +34,7 @@ class RuntimeVerifyPass void RuntimeVerifyPass::runOnFunction() { getFunction().walk([&](TflRuntimeVerifyOpInterface op) { - if (failed(op.VerifyTflRuntimeConstraints( - op.getOperation(), /*failure_on_operand_type_mismatch=*/true))) + if (failed(op.VerifyTflRuntimeConstraints(op.getOperation()))) signalPassFailure(); }); } From db3640aaf1e1b520fa59cc134f822cbb7a2ed270 Mon Sep 17 00:00:00 2001 From: "A. Unique TensorFlower" Date: Wed, 17 Jun 2020 15:43:59 -0700 Subject: [PATCH 126/867] tf.numpy Add a README.md PiperOrigin-RevId: 316980572 Change-Id: I441a3d9dc0d36b663a47721f2335b4178a71db65 --- tensorflow/python/ops/numpy_ops/README.md | 94 +++++++++++++++++++++++ 1 file changed, 94 insertions(+) create mode 100644 tensorflow/python/ops/numpy_ops/README.md diff --git a/tensorflow/python/ops/numpy_ops/README.md b/tensorflow/python/ops/numpy_ops/README.md new file mode 100644 index 00000000000..3dc37423d26 --- /dev/null +++ b/tensorflow/python/ops/numpy_ops/README.md @@ -0,0 +1,94 @@ +# tf.experimental.numpy + +This module provides a subset of numpy API, built on top of TensorFlow +operations. APIs are based on numpy 1.16 version. + +The set of supported APIs may be expanded over time. Also future releases may +change the baseline version of numpy API being supported. A list of some +systematic differences with numpy are listed later in the "Differences with +Numpy" section. + +## Types + +The module provide an `ndarray` class which wraps an immutable `tf.Tensor`. +Additional functions are provided which accept array-like objects. Here +array-like objects includes `ndarrays` as defined by this module, as well as +`tf.Tensor`, in addition to types accepted by `numpy`. + +A subset of `numpy` dtypes are supported, along with `tf.bfloat16`. +Additionally, support is provided for selecting the default float type +(`np.float32` vs `np.float64`) given that some applications may prefer lower +precision. + +## Device Support + +Given that `ndarray` and functions wrap TensorFlow constructs, the code will +have GPU and TPU support on par with TensorFlow. Also the code can be wrapped +with `tf.function` and XLA compiled. Device placement can be controlled by using +`with tf.device` scopes. + +## Graph and Eager Modes + +Eager mode execution should typically match numpy semantics of executing +op-by-op. However the same code can be executed in graph mode, by putting it +inside a `tf.function`. This can change behavior of certain operations since +symbolic execution may not have information that is computed during runtime. + +Some differences are: + + * Shapes can be incomplete or unknown. This means that `ndarray.shape`, + `ndarray.size` and `ndarray.ndim` can return `ndarray` objects instead of + returning integer (or tuple of integer) values. + * Python control flow based on `ndarray` values may not work and may have to + be rewritten to use `tf.cond` or `tf.while_loop`. Note that autograph + conversion as part of `tf.function` should still work. + * `__len__`, `__iter__` and `__index__` properties of `ndarray` may similarly + not work in graph mode. + +## Mutation and Variables + +`ndarrays` currently wrap immutable `tf.Tensor`. Also currently mutation +operations like slice assigns are not supported. This may change in the future. + +There is currently no explict construct on par with `tf.Variable`. However one +can directly construct a `tf.Variable` and use that with the numpy APIs in this +module. See section on Interop. + +## Interop + +The numpy API calls can be interleaved with TensorFlow calls without incurring +Tensor data copies. This is true even if the `ndarray` or `tf.Tensor` is placed +on a non-CPU device. + +Additionally, one could put these calls in a `with tf.GradientTape()` context to +compute gradients through the numpy API calls. Similarly, code vectorization can +be done using `tf.vectorized_map()`. + +In general, the expected behavior should be on par with that of code involving +`tf.Tensor` and running stateless TensorFlow functions on them. + +## Array Interface + +The `ndarray` class implements the `__array__` interface. This should allow +these objects to be passed into contexts that expect a `numpy` or array-like +object (e.g. matplotlib). + + +## Differences with Numpy + +Here is a non-exhaustive list of differences: + + * Not all dtypes are currently supported. e.g. `np.float96`, `np.float128`. + `np.object`, `np.str`, `np.recarray` types are not supported. + * `ndarray` storage is in C order only. Fortran order, views, stride_tricks + are not supported. + * Only a subset of functions and modules are supported. This set would be + expanded over time. For supported functions, some arguments or argument + values may not be supported. This differences are listed in the function + comments. + * Buffer mutation is currently not supported. `ndarrays` wrap immutable + tensors. This means that output buffer arguments (e..g `out` in ufuncs) are + not supported + * full `ufunc` support is not provided. + * Numpy C API is not supported. Numpy's Cython and Swig integration are not + supported. From eb0d46affb1eccfb4de1463ef9c70fdedc227d22 Mon Sep 17 00:00:00 2001 From: "A. Unique TensorFlower" Date: Wed, 17 Jun 2020 15:44:12 -0700 Subject: [PATCH 127/867] Import initialization graph in SignatureDef SavedModels as an MLIR function in TF saved model dialect. PiperOrigin-RevId: 316980623 Change-Id: I3dafe9303ad65192ca0c137943be3cd1ea0e74e7 --- tensorflow/compiler/mlir/tensorflow/BUILD | 4 +- .../mlir/tensorflow/ir/tf_saved_model.cc | 25 --- .../mlir/tensorflow/ir/tf_saved_model_ops.td | 24 --- .../tests/tf_saved_model/common_v1.py | 1 - .../tests/tf_saved_model/hash_table_v1.py | 92 ----------- .../tensorflow/tests/tf_saved_model_ops.mlir | 5 - .../tests/tf_saved_model_ops_invalid.mlir | 33 ---- .../mlir/tensorflow/translate/import_model.cc | 149 +++++------------- 8 files changed, 43 insertions(+), 290 deletions(-) delete mode 100644 tensorflow/compiler/mlir/tensorflow/tests/tf_saved_model/hash_table_v1.py diff --git a/tensorflow/compiler/mlir/tensorflow/BUILD b/tensorflow/compiler/mlir/tensorflow/BUILD index 17ed0e36a28..904ccb7e820 100644 --- a/tensorflow/compiler/mlir/tensorflow/BUILD +++ b/tensorflow/compiler/mlir/tensorflow/BUILD @@ -661,9 +661,7 @@ cc_library( ":tensorflow_types", ":translate_utils", "//tensorflow/cc/saved_model:bundle_v2", - "//tensorflow/cc/saved_model:constants", "//tensorflow/cc/saved_model:loader_lite", - "//tensorflow/cc/saved_model:loader_util", "//tensorflow/compiler/jit:shape_inference_helpers", "//tensorflow/compiler/mlir:op_or_arg_name_mapper", "//tensorflow/compiler/tf2xla:functionalize_control_flow", @@ -675,7 +673,6 @@ cc_library( "//tensorflow/core:lib", "//tensorflow/core:protos_all_cc", "//tensorflow/core/grappler/utils:transitive_fanin", - "//tensorflow/core/platform:protobuf_internal", "//tensorflow/core/platform:types", "//tensorflow/stream_executor/lib", "@com_google_absl//absl/algorithm:container", @@ -685,6 +682,7 @@ cc_library( "@com_google_absl//absl/strings", "@com_google_absl//absl/types:optional", "@llvm-project//llvm:Support", + "@llvm-project//mlir:Analysis", "@llvm-project//mlir:IR", "@llvm-project//mlir:Pass", "@llvm-project//mlir:StandardOps", diff --git a/tensorflow/compiler/mlir/tensorflow/ir/tf_saved_model.cc b/tensorflow/compiler/mlir/tensorflow/ir/tf_saved_model.cc index 6af70158e14..140a778770c 100644 --- a/tensorflow/compiler/mlir/tensorflow/ir/tf_saved_model.cc +++ b/tensorflow/compiler/mlir/tensorflow/ir/tf_saved_model.cc @@ -76,23 +76,6 @@ static LogicalResult Verify(GlobalTensorOp global_tensor) { return success(); } -static LogicalResult Verify(SessionInitializerOp session_initializer) { - mlir::SymbolTable symbol_table( - session_initializer.getParentOfType()); - - auto init_func_op = - symbol_table.lookup(session_initializer.initializer()); - if (!init_func_op) - return session_initializer.emitOpError() - << "the initializer function does not exist"; - - if (!init_func_op.getType().getResults().empty()) - return session_initializer.emitOpError() - << "the initializer function should have no output"; - - return success(); -} - #define GET_OP_CLASSES #include "tensorflow/compiler/mlir/tensorflow/ir/tf_saved_model.cc.inc" @@ -237,14 +220,6 @@ static LogicalResult VerifySavedModelModule( } } } - - auto session_initializers = module.getOps(); - if (std::distance(session_initializers.begin(), session_initializers.end()) > - 1) { - return (*++session_initializers.begin()).emitError() - << "there must be no more than one session_initializer op"; - } - SymbolTable symbol_table(module); auto symbol_uses = SymbolTable::getSymbolUses(&module.getBodyRegion()); if (!symbol_uses.hasValue()) { diff --git a/tensorflow/compiler/mlir/tensorflow/ir/tf_saved_model_ops.td b/tensorflow/compiler/mlir/tensorflow/ir/tf_saved_model_ops.td index 497f4d90cb9..4431a160edf 100644 --- a/tensorflow/compiler/mlir/tensorflow/ir/tf_saved_model_ops.td +++ b/tensorflow/compiler/mlir/tensorflow/ir/tf_saved_model_ops.td @@ -128,28 +128,4 @@ def TfSavedModel_GlobalTensorOp : TfSavedModel_Op<"global_tensor"> { let verifier = [{ return Verify(*this); }]; } -def TfSavedModel_SessionInitializerOp: TfSavedModel_Op<"session_initializer"> { - let summary = "Initializes TensorFlow session state."; - let description = [{ - Represents a session initializer function initializes TensorFlow session - state. It is used to initialize resources in the saved model before calling - any exported functions. There must be no more than one session initializer - in a saved model. - - The `initializer` represents the initialization function. The function have - no output and this function should be only called once. - - This is used, for example, to initialize hash tables stored in resources and - accessed by resource name (rather than as resource handles or bound inputs - which is how `global_tensor`s are referenced) - }]; - - let arguments = (ins - FlatSymbolRefAttr:$initializer - ); - - - let verifier = [{ return Verify(*this); }]; -} - #endif // SAVED_MODEL_DIALECT diff --git a/tensorflow/compiler/mlir/tensorflow/tests/tf_saved_model/common_v1.py b/tensorflow/compiler/mlir/tensorflow/tests/tf_saved_model/common_v1.py index 51ccbeb1fbd..7171f63bb05 100644 --- a/tensorflow/compiler/mlir/tensorflow/tests/tf_saved_model/common_v1.py +++ b/tensorflow/compiler/mlir/tensorflow/tests/tf_saved_model/common_v1.py @@ -84,7 +84,6 @@ def do_test(signature_def_map, show_debug_info=False): builder.add_meta_graph_and_variables( sess, [tf.saved_model.tag_constants.SERVING], signature_def_map, - main_op=tf.tables_initializer(), strip_default_attrs=True) builder.save() diff --git a/tensorflow/compiler/mlir/tensorflow/tests/tf_saved_model/hash_table_v1.py b/tensorflow/compiler/mlir/tensorflow/tests/tf_saved_model/hash_table_v1.py deleted file mode 100644 index 64847434b82..00000000000 --- a/tensorflow/compiler/mlir/tensorflow/tests/tf_saved_model/hash_table_v1.py +++ /dev/null @@ -1,92 +0,0 @@ -# Copyright 2019 The TensorFlow Authors. All Rights Reserved. -# -# Licensed under the Apache License, Version 2.0 (the "License"); -# you may not use this file except in compliance with the License. -# You may obtain a copy of the License at -# -# http://www.apache.org/licenses/LICENSE-2.0 -# -# Unless required by applicable law or agreed to in writing, software -# distributed under the License is distributed on an "AS IS" BASIS, -# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -# See the License for the specific language governing permissions and -# limitations under the License. -# ============================================================================== - -# RUN: %p/hash_table_v1 | FileCheck %s - -# pylint: disable=missing-docstring,line-too-long -from __future__ import absolute_import -from __future__ import division -from __future__ import print_function - -import tensorflow.compat.v1 as tf -from tensorflow.compiler.mlir.tensorflow.tests.tf_saved_model import common_v1 - -# Verify that the tf.versions attribute exists. It is difficult to enforce -# contents, since the version numbers change over time. The conversion logic -# itself is verified in the common graphdef converter, so here just assert -# it is being invoked. -# CHECK: module -# CHECK-SAME: tf.versions -# CHECK-SAME: bad_consumers -# CHECK-SAME: min_consumer -# CHECK-SAME: producer - -# CHECK: "tf_saved_model.session_initializer"() {initializer = [[init:@.*]]} : () -> () -# CHECK: "tf_saved_model.global_tensor"() - -# CHECK: func {{@[a-zA-Z_0-9]+}}( -# CHECK-SAME: [[ARG0:%.*]]: tensor -# CHECK-SAME: [[ARG1:%.*]]: tensor () - // Representation for constants: (immutable) global tensor. // CHECK: tf_saved_model.global_tensor "tf_saved_model.global_tensor"() { diff --git a/tensorflow/compiler/mlir/tensorflow/tests/tf_saved_model_ops_invalid.mlir b/tensorflow/compiler/mlir/tensorflow/tests/tf_saved_model_ops_invalid.mlir index 544600cf6b8..c055c6c9f56 100644 --- a/tensorflow/compiler/mlir/tensorflow/tests/tf_saved_model_ops_invalid.mlir +++ b/tensorflow/compiler/mlir/tensorflow/tests/tf_saved_model_ops_invalid.mlir @@ -258,36 +258,3 @@ module attributes {tf_saved_model.semantics} { // expected-error@+1 {{'type' attribute for immutable 'tf_saved_model.global_tensor' should have a static shape}} "tf_saved_model.global_tensor"() { sym_name = "v", type = tensor, value = dense<1.> : tensor<1xf32> } : () -> () } - -// ----- - -module attributes {tf_saved_model.semantics} { - - // expected-error@+1 {{the initializer function does not exist}} - "tf_saved_model.session_initializer"() { initializer = @init } : () -> () -} - -// ----- - -module attributes {tf_saved_model.semantics} { - - // expected-error@+1 {{the initializer function should have no output}} - "tf_saved_model.session_initializer"() { initializer = @init } : () -> () - func @init() -> tensor<1xf32> { - %0 = "tf.Const"() {value = dense<[1.0]> : tensor<1xf32> } : () -> tensor<1xf32> - return %0 : tensor<1xf32> - } -} - -// ----- - -module attributes {tf_saved_model.semantics} { - - "tf_saved_model.session_initializer"() { initializer = @init } : () -> () - // expected-error@+1 {{there must be no more than one session_initializer op}} - "tf_saved_model.session_initializer"() { initializer = @init } : () -> () - func @init() -> tensor<1xf32> { - %0 = "tf.Const"() {value = dense<[1.0]> : tensor<1xf32> } : () -> tensor<1xf32> - return %0 : tensor<1xf32> - } -} diff --git a/tensorflow/compiler/mlir/tensorflow/translate/import_model.cc b/tensorflow/compiler/mlir/tensorflow/translate/import_model.cc index 3cff4217215..820d0ce31fb 100644 --- a/tensorflow/compiler/mlir/tensorflow/translate/import_model.cc +++ b/tensorflow/compiler/mlir/tensorflow/translate/import_model.cc @@ -60,8 +60,6 @@ limitations under the License. #include "mlir/IR/Types.h" // from @llvm-project #include "mlir/IR/Verifier.h" // from @llvm-project #include "mlir/Pass/PassManager.h" // from @llvm-project -#include "tensorflow/cc/saved_model/constants.h" -#include "tensorflow/cc/saved_model/loader_util.h" #include "tensorflow/compiler/jit/shape_inference_helpers.h" #include "tensorflow/compiler/mlir/op_or_arg_name_mapper.h" #include "tensorflow/compiler/mlir/tensorflow/ir/tf_attributes.h" @@ -101,7 +99,6 @@ limitations under the License. #include "tensorflow/core/lib/core/errors.h" #include "tensorflow/core/lib/strings/str_util.h" #include "tensorflow/core/platform/protobuf.h" -#include "tensorflow/core/platform/protobuf_internal.h" #include "tensorflow/core/platform/types.h" #include "tensorflow/core/protobuf/graph_debug_info.pb.h" #include "tensorflow/core/protobuf/meta_graph.pb.h" @@ -119,7 +116,6 @@ using mlir::NamedAttrList; using mlir::TensorType; using mlir::TF::VarHandleOp; using mlir::tf_saved_model::GlobalTensorOp; -using mlir::tf_saved_model::SessionInitializerOp; using stream_executor::port::StatusOr; namespace { @@ -2959,13 +2955,6 @@ void SortSavedModelModule(mlir::ModuleOp module) { named_global_tensor.global_tensor.getOperation()->moveBefore( &module.getBody()->front()); } - - auto initializers = module.getOps(); - if (!initializers.empty()) { - (*initializers.begin()) - .getOperation() - ->moveBefore(&module.getBody()->front()); - } } Status CreateSavedModelIR( @@ -3252,29 +3241,17 @@ class SavedModelSignatureDefImporter { absl::Span exported_names, mlir::MLIRContext* context) : bundle_(bundle), - flib_def_(OpRegistry::Global(), graph_def().library()), - debug_info_(), exported_names_(exported_names), - module_(mlir::ModuleOp::create(mlir::UnknownLoc::get(context))) { - // debug_info might not be loaded with loader_lite. - if (bundle_.debug_info != nullptr) debug_info_ = *bundle_.debug_info; - } + module_(mlir::ModuleOp::create(mlir::UnknownLoc::get(context))) {} // Converts the SavedModel to the SavedModel dialect. Creates an MLIR function // for each signature. StatusOr ConvertSignatures(); - Status ConvertSignature(const std::string& sig_def_key, - const SignatureDef& signature_def); - - // Converts the initialization graph in the SavedModel to an MLIR function. - Status ConvertInitializer(); - - // Converts a graph with feeds and fetches to an MLIR function. - StatusOr ConvertGraph( - const std::string& name, - const std::vector>& inputs, - const std::vector>& outputs, - const std::vector control_outputs); + Status ConvertSignature(const GraphDef& graphdef, + const std::string& sig_def_key, + const SignatureDef& signature_def, + const GraphDebugInfo& debug_info, + const FunctionLibraryDefinition& flib_def); // Creates GlobalTensorOp for each variable and moves each VarHandle op to // the enclosing function's arguments. @@ -3296,62 +3273,18 @@ class SavedModelSignatureDefImporter { GraphImportConfig::InputArrays ParseInputArrays( const std::vector>& inputs); - const GraphDef& graph_def() const { - return bundle_.meta_graph_def.graph_def(); - } - const FunctionLibraryDefinition& flib_def() const { return flib_def_; } - const GraphDebugInfo& debug_info() const { return debug_info_; } - const SavedModelBundle& bundle_; - FunctionLibraryDefinition flib_def_; - GraphDebugInfo debug_info_; absl::Span exported_names_; mlir::OwningModuleRef module_; }; -Status SavedModelSignatureDefImporter::ConvertInitializer() { - std::vector asset_file_defs; - TF_RETURN_IF_ERROR( - internal::GetAssetFileDefs(bundle_.meta_graph_def, &asset_file_defs)); - - if (!asset_file_defs.empty()) - return errors::Unimplemented( - absl::StrCat("Assets are not supported in signaturedef importer")); - - std::string init_node_name; - TF_RETURN_IF_ERROR( - internal::GetInitOp("", bundle_.meta_graph_def, &init_node_name)); - - if (init_node_name.empty()) return Status::OK(); - - TF_ASSIGN_OR_RETURN(auto sub_module, - ConvertGraph(init_node_name, {}, {}, {init_node_name})); - - mlir::SymbolTable symbol_table(*sub_module); - - auto init_func_op = symbol_table.lookup(init_node_name); - - init_func_op.removeAttr("tf.entry_function"); - - mlir::OpBuilder builder(module_->getBodyRegion()); - - builder.create( - module_->getLoc(), builder.getSymbolRefAttr(init_func_op.getName())); - - // Move the converted functions to top level MLIR module. - auto* block = module_->getBody(); - auto* sub_block = sub_module->getBody(); - block->getOperations().splice( - mlir::Block::iterator(block->getTerminator()), sub_block->getOperations(), - sub_block->begin(), mlir::Block::iterator(sub_block->getTerminator())); - - return Status::OK(); -} - StatusOr SavedModelSignatureDefImporter::ConvertSignatures() { const auto& signatures = bundle_.GetSignatures(); - PopulateTfVersions(module_.get(), graph_def().versions()); + const auto& graphdef = bundle_.meta_graph_def.graph_def(); + PopulateTfVersions(module_.get(), graphdef.versions()); + + FunctionLibraryDefinition flib_def(OpRegistry::Global(), graphdef.library()); // debug_info might not be loaded with loader_lite. GraphDebugInfo debug_info; @@ -3374,10 +3307,9 @@ SavedModelSignatureDefImporter::ConvertSignatures() { continue; } - TF_RETURN_IF_ERROR(ConvertSignature(sig_def_key, signature_def)); + TF_RETURN_IF_ERROR(ConvertSignature(graphdef, sig_def_key, signature_def, + debug_info, flib_def)); } - - TF_RETURN_IF_ERROR(ConvertInitializer()); TF_RETURN_IF_ERROR(LiftVariables()); mlir::OpBuilder builder(module_->getBodyRegion()); @@ -3388,32 +3320,10 @@ SavedModelSignatureDefImporter::ConvertSignatures() { return std::move(module_); } -StatusOr SavedModelSignatureDefImporter::ConvertGraph( - const std::string& name, - const std::vector>& inputs, - const std::vector>& outputs, - const std::vector control_outputs) { - GraphImportConfig specs; - specs.prune_unused_nodes = true; - specs.inputs = ParseInputArrays(inputs); - for (auto& output : outputs) specs.outputs.push_back(output.second.name()); - specs.control_outputs = control_outputs; - - // Convert sub-graphdef to sub-graph. - GraphConstructorOptions options; - options.allow_internal_ops = true; - options.add_default_attributes = true; - Graph graph(OpRegistry::Global()); - - TF_RETURN_IF_ERROR(ConvertGraphDefToGraph(options, graph_def(), &graph)); - - // Convert sub-graph to MLIR module.true - return GraphDefImporter::Convert(module_->getContext(), graph, debug_info(), - flib_def(), specs, name); -} - Status SavedModelSignatureDefImporter::ConvertSignature( - const std::string& sig_def_key, const SignatureDef& signature_def) { + const GraphDef& graphdef, const std::string& sig_def_key, + const SignatureDef& signature_def, const GraphDebugInfo& debug_info, + const FunctionLibraryDefinition& flib_def) { // Create local vectors for the input and output and sort them to be // deterministic. We don't want anyone to really depend on the order, client // should lookup argument/result mapping by attribute name. @@ -3429,9 +3339,34 @@ Status SavedModelSignatureDefImporter::ConvertSignature( return lhs.first.size() < rhs.first.size() || lhs.first > rhs.first; }); + GraphImportConfig specs; + specs.prune_unused_nodes = true; + specs.inputs = ParseInputArrays(inputs); + for (auto& output : outputs) specs.outputs.push_back(output.second.name()); + + // Remove unused nodes and create sub-graphdef. + GraphDef sub_graph_def; + TF_RETURN_IF_ERROR(tensorflow::grappler::SetTransitiveFaninGraph( + graphdef, &sub_graph_def, + /*terminal_nodes=*/{specs.outputs.begin(), specs.outputs.end()})); + + // Set the function library definitions in the pruned graphdef. + *sub_graph_def.mutable_library() = flib_def.ToProto(); + + // Convert sub-graphdef to sub-graph. + GraphConstructorOptions options; + options.allow_internal_ops = true; + options.add_default_attributes = true; + Graph sub_graph(OpRegistry::Global()); + + TF_RETURN_IF_ERROR( + ConvertGraphDefToGraph(options, sub_graph_def, &sub_graph)); + // Convert sub-graph to MLIR module. - TF_ASSIGN_OR_RETURN(auto sub_module, - ConvertGraph(sig_def_key, inputs, outputs, {})); + TF_ASSIGN_OR_RETURN( + auto sub_module, + GraphDefImporter::Convert(module_->getContext(), sub_graph, debug_info, + flib_def, specs, sig_def_key)); mlir::OpBuilder builder(sub_module->getBodyRegion()); // Find the FuncOp which corresponds to current SignatureDef. From 02be9b5ef8b3faac54feea8ee99ff14c62d40704 Mon Sep 17 00:00:00 2001 From: Robert David Date: Wed, 17 Jun 2020 15:46:30 -0700 Subject: [PATCH 128/867] Remove erroneous end-of-namespace comment at the end of a function. PiperOrigin-RevId: 316981028 Change-Id: I06c1063ba6206fe301f8fd6b89e3ac435ec56ba6 --- tensorflow/lite/kernels/lstm_eval.cc | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/tensorflow/lite/kernels/lstm_eval.cc b/tensorflow/lite/kernels/lstm_eval.cc index 4ac9e538317..2e7f300f9a9 100644 --- a/tensorflow/lite/kernels/lstm_eval.cc +++ b/tensorflow/lite/kernels/lstm_eval.cc @@ -893,7 +893,7 @@ inline void LstmStepHybrid( std::copy_n(output_state_ptr + b * n_output, n_output, output_ptr + b * output_batch_leading_dim); } -} // namespace +} // Fully quantized lstm kernel for 16 bit gate matmul output. // From 09bdeb13ebb20f12c76e3d0bd852f42b2f63ffef Mon Sep 17 00:00:00 2001 From: Andy Ly Date: Wed, 17 Jun 2020 15:46:55 -0700 Subject: [PATCH 129/867] Fix updating results of TPU cluster with parallel_execute results when TPU cluster results are not perfectly forwarded. After TPUExtractHeadTailOutsideCompilation some results left of the TPU cluster may not be used by an optional tf_device.replicate op if there is data parallelism. Instead, all results should be remapped if they are used outside of parallel_execute. PiperOrigin-RevId: 316981114 Change-Id: I5529074857e06cfe26a7141c262a6229fe848be6 --- .../tpu_extract_outside_compilation.mlir | 22 +++++++++++++++++- .../tpu_extract_outside_compilation.cc | 23 ++++++++----------- 2 files changed, 31 insertions(+), 14 deletions(-) diff --git a/tensorflow/compiler/mlir/tensorflow/tests/tpu_extract_outside_compilation.mlir b/tensorflow/compiler/mlir/tensorflow/tests/tpu_extract_outside_compilation.mlir index 6bb8e99d796..d88489f5da0 100644 --- a/tensorflow/compiler/mlir/tensorflow/tests/tpu_extract_outside_compilation.mlir +++ b/tensorflow/compiler/mlir/tensorflow/tests/tpu_extract_outside_compilation.mlir @@ -262,7 +262,6 @@ func @single_outside_compiled_input_output_single_outside_compilation(%arg0: ten return %1 : tensor } - // Tests extraction of a single outside compiled cluster with multiple input/output. // CHECK-LABEL: func @multiple_outside_compiled_input_output_single_outside_compilation @@ -439,3 +438,24 @@ func @multiple_outside_compiled_inputs_single_outside_compilation(%arg0: tensor< return %1 : tensor } + +// Tests only directly used results of tpu cluster are remapped with +// parallel_execute. + +// CHECK-LABEL: func @remapped_results +func @remapped_results(%arg0: tensor) -> tensor { + %0 = "tf.A"(%arg0) : (tensor) -> tensor + // CHECK: %[[REPLICATE:[0-9]*]]:2 = tf_device.replicate + // CHECK: %[[PARALLEL_EXECUTE_OUTPUT:[0-9]*]]:2 = "tf_device.parallel_execute" + // CHECK: tf_device.return %[[PARALLEL_EXECUTE_OUTPUT]]#1 : tensor + %1:2 = tf_device.replicate([%0, %arg0] as %ri_0: tensor) {n = 2 : i32} { + %2:2 = "tf_device.cluster"() ( { + %3 = "tf.A"() : () -> (tensor) + %4 = "tf.B"(%3) {_xla_outside_compilation = "cluster1"} : (tensor) -> (tensor) + %5:2 = "tf.C"(%4) : (tensor) -> (tensor, tensor) + tf_device.return %5#0, %5#1 : tensor, tensor + }) {cluster_attr = "cluster_attr"} : () -> (tensor, tensor) + tf_device.return %2#1 : tensor + } + return %1 : tensor +} diff --git a/tensorflow/compiler/mlir/tensorflow/transforms/tpu_extract_outside_compilation.cc b/tensorflow/compiler/mlir/tensorflow/transforms/tpu_extract_outside_compilation.cc index 54600faca4b..a2a19108326 100644 --- a/tensorflow/compiler/mlir/tensorflow/transforms/tpu_extract_outside_compilation.cc +++ b/tensorflow/compiler/mlir/tensorflow/transforms/tpu_extract_outside_compilation.cc @@ -108,18 +108,6 @@ tf_device::LaunchOp CreateLaunchOpForOutsideCluster( return launch_op; } -// Propagates the return from `parallel_execute_op` to parent replicate -// op if it exists. -void PropagateParallelExecuteReturnToReplicate( - tf_device::ParallelExecuteOp parallel_execute_op) { - // Update the return for the parallel_execute op parent. - auto replicate = llvm::dyn_cast_or_null( - parallel_execute_op.getParentOp()); - if (replicate) - replicate.GetBody().getTerminator()->setOperands( - parallel_execute_op.execute_outputs()); -} - // Extracts all externally provided operands of `cluster_ops`. llvm::SmallSetVector GetExternalOperands( llvm::ArrayRef cluster_ops) { @@ -305,7 +293,16 @@ void CreateParallelExecuteFromOutsideClusters( tpu_cluster.getOperation()->moveBefore( parallel_execute_tpu_block.getTerminator()); - PropagateParallelExecuteReturnToReplicate(parallel_execute_op); + // Remap cluster results with parallel_execute results if user is outside of + // parallel_execute. + for (auto result : + llvm::zip(tpu_cluster.getResults(), parallel_execute_op.getResults())) { + Value tpu_cluster_result = std::get<0>(result); + Value parallel_execute_result = std::get<1>(result); + for (auto& use : llvm::make_early_inc_range(tpu_cluster_result.getUses())) + if (!parallel_execute_op.getOperation()->isProperAncestor(use.getOwner())) + use.set(parallel_execute_result); + } } void TPUExtractOutsideCompilation::runOnFunction() { From 259ce6bd35c487ca39b375f16dc91d3636a9e838 Mon Sep 17 00:00:00 2001 From: "A. Unique TensorFlower" Date: Wed, 17 Jun 2020 15:47:01 -0700 Subject: [PATCH 130/867] Move the ScatterExpander pass after the DynamicPadder pass on cpu. DynamicPadder will pad the indices of scatter(operand 1) if indices is of dynamic size. If ScatterExpander pass is run first, scatter will be rewrite to a sequence of other ops and this pad will not be performed and give incorrect answer. PiperOrigin-RevId: 316981132 Change-Id: I2662d783e73a0467ebd4f1af21f7ac6cf93faef5 --- tensorflow/compiler/xla/service/cpu/cpu_compiler.cc | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/tensorflow/compiler/xla/service/cpu/cpu_compiler.cc b/tensorflow/compiler/xla/service/cpu/cpu_compiler.cc index b2416ac2799..31b9fe1c920 100644 --- a/tensorflow/compiler/xla/service/cpu/cpu_compiler.cc +++ b/tensorflow/compiler/xla/service/cpu/cpu_compiler.cc @@ -277,12 +277,12 @@ Status CpuCompiler::RunHloPassesThroughLayoutAssn( pipeline.AddPass( cost_model, /*convert_batch_groups_only=*/false); - pipeline.AddPass(); pipeline.AddPass( /*rewrite_training_op=*/true, /*rewrite_inference_op=*/true, /*rewrite_grad_op=*/true); pipeline.AddPass(); + pipeline.AddPass(); pipeline.AddPass(); pipeline.AddPass(target_machine_features); { From 2053b238d0d717aef4dfcf4be03daa43d7eeede2 Mon Sep 17 00:00:00 2001 From: "A. Unique TensorFlower" Date: Wed, 17 Jun 2020 15:50:24 -0700 Subject: [PATCH 131/867] add two internal stats. PiperOrigin-RevId: 316981769 Change-Id: Ia53c128b82fa750942acb5441ae6c8dc31478b07 --- tensorflow/core/profiler/utils/xplane_schema.cc | 5 ++++- tensorflow/core/profiler/utils/xplane_schema.h | 2 ++ 2 files changed, 6 insertions(+), 1 deletion(-) diff --git a/tensorflow/core/profiler/utils/xplane_schema.cc b/tensorflow/core/profiler/utils/xplane_schema.cc index 197dab75d3b..be53dcbdc01 100644 --- a/tensorflow/core/profiler/utils/xplane_schema.cc +++ b/tensorflow/core/profiler/utils/xplane_schema.cc @@ -166,6 +166,8 @@ const StatTypeMap& GetStatTypeMap() { {"is_eager", kIsEager}, {"tf_function_call", kTfFunctionCall}, {"tracing_count", kTfFunctionTracingCount}, + {"flops", kFlops}, + {"bytes_accessed", kBytesAccessed}, // Performance counter related. {"Raw Value", kRawValue}, {"Scaled Value", kScaledValue}, @@ -227,7 +229,8 @@ bool IsInternalStat(absl::optional stat_type) { StatType::kKernelDetails, StatType::kLevel0, StatType::kProducerType, StatType::kProducerId, StatType::kConsumerType, StatType::kConsumerId, - StatType::kIsRoot, StatType::kIsAsync}; + StatType::kIsRoot, StatType::kIsAsync, + StatType::kFlops, StatType::kBytesAccessed}; return stat_type.has_value() && kInternalStats->contains(*stat_type); } diff --git a/tensorflow/core/profiler/utils/xplane_schema.h b/tensorflow/core/profiler/utils/xplane_schema.h index 8b999dc6f9f..a31814cef06 100644 --- a/tensorflow/core/profiler/utils/xplane_schema.h +++ b/tensorflow/core/profiler/utils/xplane_schema.h @@ -153,6 +153,8 @@ enum StatType { kIsEager, kTfFunctionCall, kTfFunctionTracingCount, + kFlops, + kBytesAccessed, // Performance counter related. kRawValue, kScaledValue, From 97afd248868c7f28c197abde87cf610b550bdef9 Mon Sep 17 00:00:00 2001 From: Kaixi Hou Date: Wed, 17 Jun 2020 10:00:23 -0700 Subject: [PATCH 132/867] Relu grad GPU uses 8 float16 element vector --- tensorflow/core/kernels/relu_op_gpu.cu.cc | 89 +++++++++++++++++++++-- 1 file changed, 82 insertions(+), 7 deletions(-) diff --git a/tensorflow/core/kernels/relu_op_gpu.cu.cc b/tensorflow/core/kernels/relu_op_gpu.cu.cc index 27fd5f64249..983cc127863 100644 --- a/tensorflow/core/kernels/relu_op_gpu.cu.cc +++ b/tensorflow/core/kernels/relu_op_gpu.cu.cc @@ -35,6 +35,7 @@ namespace tensorflow { typedef Eigen::GpuDevice GPUDevice; +static constexpr int VectorSizeElements = 8; namespace functor { // This kernel computes ReluGrad by processing one half2, two fp16, at a time. @@ -93,6 +94,66 @@ __global__ void ReluGradHalfKernel(const Eigen::half* __restrict__ gradient, } } +__global__ void ReluGradHalfKernelVector( + const Eigen::half* __restrict__ gradient, + const Eigen::half* __restrict__ feature, + Eigen::half* __restrict__ backprop, int32 count) { + int32 half8_count = count / VectorSizeElements; + int32 index = blockIdx.x * blockDim.x + threadIdx.x; + + if (index < half8_count) { + // Cast to xx_h8 for vector load and store. + float4 gradient_h8 = reinterpret_cast(gradient)[index]; + float4 feature_h8 = reinterpret_cast(feature)[index]; + float4* p_backprop_h8 = reinterpret_cast(backprop) + index; + + half2 *gradient_h2 = reinterpret_cast(&gradient_h8); + half2 *feature_h2 = reinterpret_cast(&feature_h8); + float4 backprop_h8; + half2* p_backprop_h2 = reinterpret_cast(&backprop_h8); + + // Fast path, when half2 primitives are available. +#if __CUDA_ARCH__ >= 530 + const half2 kZeroH2 = __float2half2_rn(0.f); +#endif + for (int i = 0; i < VectorSizeElements / 2; i++) { +#if __CUDA_ARCH__ >= 530 + // mask = (feature > 0) + half2 mask_h2 = __hgt2(feature_h2[i], kZeroH2); + // backprop = mask * gradient + half2 backprop_h2 = __hmul2(mask_h2, gradient_h2[i]); +#else + // Fall back: convert half2 to float2 for processing. + float2 feature_f2 = __half22float2(feature_h2[i]); + float2 gradient_f2 = __half22float2(gradient_h2[i]); + float2 backprop_f2 = + make_float2((feature_f2.x > 0.0f) ? float(gradient_f2.x) : 0.0f, + (feature_f2.y > 0.0f) ? float(gradient_f2.y) : 0.0f); + // Convert back to half2. + half2 backprop_h2 = __float22half2_rn(backprop_f2); +#endif + p_backprop_h2[i] = backprop_h2; + } + // Write back the result. + *p_backprop_h8 = backprop_h8; + } + + int remaining_count = (count % VectorSizeElements); + + if (index < remaining_count) { + // Use first threads to process the remaining elements. + Eigen::half grad_h = gradient[half8_count * VectorSizeElements + index]; + Eigen::half feature_h = feature[half8_count * VectorSizeElements + index]; + + float grad_f = static_cast(grad_h); + float feature_f = static_cast(feature_h); + float backprop_f = (feature_f > 0) ? grad_f : 0; + + Eigen::half backprop_h(backprop_f); + backprop[half8_count * VectorSizeElements + index] = backprop_h; + } +} + template struct ReluGrad { // Computes ReluGrad backprop. @@ -108,15 +169,29 @@ struct ReluGrad { // NOTE: When the activation is exactly zero, we do not propagate the // associated gradient value. This allows the output of the Relu to be used, // as well as its input. + auto gradient_ptr = reinterpret_cast(gradient.data()); + auto feature_ptr = reinterpret_cast(feature.data()); + auto backprop_ptr = reinterpret_cast(backprop.data()); + bool aligned = gradient_ptr % 16 == 0 && feature_ptr % 16 == 0 && + backprop_ptr % 16 == 0; int32 count = gradient.size(); - if (count == 0) return; - int32 half2_count = Eigen::divup(count, 2); constexpr int32 kThreadInBlock = 512; - GpuLaunchConfig config = GetGpuLaunchConfigFixedBlockSize( - half2_count, d, ReluGradHalfKernel, 0, kThreadInBlock); - TF_CHECK_OK(GpuLaunchKernel( - ReluGradHalfKernel, config.block_count, config.thread_per_block, 0, - d.stream(), gradient.data(), feature.data(), backprop.data(), count)); + if (count == 0) return; + if (aligned) { + int32 half8_count = Eigen::divup(count, VectorSizeElements); + int32 kBlock = Eigen::divup(half8_count, kThreadInBlock); + TF_CHECK_OK(GpuLaunchKernel( + ReluGradHalfKernelVector, kBlock, kThreadInBlock, + 0, d.stream(), gradient.data(), feature.data(), backprop.data(), + count)); + } else { + int32 half2_count = Eigen::divup(count, 2); + GpuLaunchConfig config = GetGpuLaunchConfigFixedBlockSize( + half2_count, d, ReluGradHalfKernel, 0, kThreadInBlock); + TF_CHECK_OK(GpuLaunchKernel( + ReluGradHalfKernel, config.block_count, config.thread_per_block, 0, + d.stream(), gradient.data(), feature.data(), backprop.data(), count)); + } } }; From bae911298b8b79ffd24464adc8fc123566a79c24 Mon Sep 17 00:00:00 2001 From: Raman Sarokin Date: Wed, 17 Jun 2020 17:16:59 -0700 Subject: [PATCH 133/867] Additional check for Intel specialization. PiperOrigin-RevId: 316997503 Change-Id: Ibc97cd77c5e15bb021546aac754e2953b050f0f3 --- tensorflow/lite/delegates/gpu/cl/cl_device.cc | 27 +++++++++++++++++++ tensorflow/lite/delegates/gpu/cl/cl_device.h | 1 + .../delegates/gpu/cl/kernels/conv_powervr.cc | 2 +- 3 files changed, 29 insertions(+), 1 deletion(-) diff --git a/tensorflow/lite/delegates/gpu/cl/cl_device.cc b/tensorflow/lite/delegates/gpu/cl/cl_device.cc index 64e07428515..271fbce61ce 100644 --- a/tensorflow/lite/delegates/gpu/cl/cl_device.cc +++ b/tensorflow/lite/delegates/gpu/cl/cl_device.cc @@ -501,6 +501,33 @@ bool CLDevice::IsCL20OrHigher() const { info_.cl_version != OpenCLVersion::CL_1_2; } +bool CLDevice::SupportsSubGroupWithSize(int sub_group_size) const { + if (IsIntel()) { + if (SupportsExtension("cl_intel_required_subgroup_size")) { + size_t sub_groups_count; + cl_int error = + clGetDeviceInfo(id_, 0x4108 /*CL_DEVICE_SUB_GROUP_SIZES_INTEL*/, 0, + nullptr, &sub_groups_count); + if (error != CL_SUCCESS) { + return false; + } + std::vector sub_group_sizes(sub_groups_count); + error = clGetDeviceInfo(id_, 0x4108 /*CL_DEVICE_SUB_GROUP_SIZES_INTEL*/, + sizeof(size_t) * sub_groups_count, + sub_group_sizes.data(), nullptr); + if (error != CL_SUCCESS) { + return false; + } + for (int i = 0; i < sub_groups_count; ++i) { + if (sub_group_sizes[i] == sub_group_size) { + return true; + } + } + } + } + return false; +} + bool CLDevice::IsAdreno() const { return info_.vendor == Vendor::QUALCOMM; } bool CLDevice::IsAdreno3xx() const { diff --git a/tensorflow/lite/delegates/gpu/cl/cl_device.h b/tensorflow/lite/delegates/gpu/cl/cl_device.h index ae6a1d11af6..68abcf3e202 100644 --- a/tensorflow/lite/delegates/gpu/cl/cl_device.h +++ b/tensorflow/lite/delegates/gpu/cl/cl_device.h @@ -179,6 +179,7 @@ class CLDevice { bool SupportsFP32RTN() const; bool SupportsFP16RTN() const; bool IsCL20OrHigher() const; + bool SupportsSubGroupWithSize(int sub_group_size) const; bool IsAdreno() const; bool IsAdreno3xx() const; bool IsAdreno4xx() const; diff --git a/tensorflow/lite/delegates/gpu/cl/kernels/conv_powervr.cc b/tensorflow/lite/delegates/gpu/cl/kernels/conv_powervr.cc index bd694e7cc4f..c20cfdbeaa3 100644 --- a/tensorflow/lite/delegates/gpu/cl/kernels/conv_powervr.cc +++ b/tensorflow/lite/delegates/gpu/cl/kernels/conv_powervr.cc @@ -874,7 +874,7 @@ ConvPowerVR::ConvParams ConvPowerVR::GuessBestParams( if (definition.precision != CalculationsPrecision::F32_F16 && device.SupportsExtension("cl_khr_subgroups") && device.SupportsExtension("cl_intel_required_subgroup_size") && - device.IsCL20OrHigher()) { + device.IsCL20OrHigher() && device.SupportsSubGroupWithSize(16)) { conv_params.weights_upload_type = WeightsUploadType::PRIVATE_MEM_SIMD16_BROADCAST; } else { From 1f05cc5973c1182c35cd4b11ee8cf0f01ed707dc Mon Sep 17 00:00:00 2001 From: Raman Sarokin Date: Wed, 17 Jun 2020 17:18:45 -0700 Subject: [PATCH 134/867] Better FullyConnected/ConvTransposed selection for Intel. PiperOrigin-RevId: 316997745 Change-Id: I28befdd528917c3846ff6ae79b0f8427389dfc39 --- .../delegates/gpu/cl/kernels/convolution_transposed_3x3.cc | 2 +- .../delegates/gpu/cl/kernels/convolution_transposed_4x4.cc | 2 +- .../gpu/cl/selectors/convolution_transposed_selector.cc | 1 + .../delegates/gpu/cl/selectors/fully_connected_selector.cc | 3 +++ 4 files changed, 6 insertions(+), 2 deletions(-) diff --git a/tensorflow/lite/delegates/gpu/cl/kernels/convolution_transposed_3x3.cc b/tensorflow/lite/delegates/gpu/cl/kernels/convolution_transposed_3x3.cc index 4a68eda1d95..9b028721d2d 100644 --- a/tensorflow/lite/delegates/gpu/cl/kernels/convolution_transposed_3x3.cc +++ b/tensorflow/lite/delegates/gpu/cl/kernels/convolution_transposed_3x3.cc @@ -269,7 +269,7 @@ ConvolutionTransposed3x3::ConvolutionTransposed3x3( work_group_launch_order_(2, 0, 1) { if (device.IsPowerVR()) { weights_upload_type_ = WeightsUploadType::LOCAL_MEM_ASYNC; - } else if (device.IsNvidia()) { + } else if (device.IsNvidia() || device.IsIntel()) { weights_upload_type_ = WeightsUploadType::LOCAL_MEM_BY_THREADS; } else if (device.IsAMD()) { weights_upload_type_ = WeightsUploadType::CONSTANT_MEM; diff --git a/tensorflow/lite/delegates/gpu/cl/kernels/convolution_transposed_4x4.cc b/tensorflow/lite/delegates/gpu/cl/kernels/convolution_transposed_4x4.cc index 0f7f90989e8..209b675087e 100644 --- a/tensorflow/lite/delegates/gpu/cl/kernels/convolution_transposed_4x4.cc +++ b/tensorflow/lite/delegates/gpu/cl/kernels/convolution_transposed_4x4.cc @@ -270,7 +270,7 @@ ConvolutionTransposed4x4::ConvolutionTransposed4x4( : GPUOperation(definition) { if (device.IsPowerVR()) { weights_upload_type_ = WeightsUploadType::LOCAL_MEM_ASYNC; - } else if (device.IsNvidia()) { + } else if (device.IsNvidia() || device.IsIntel()) { weights_upload_type_ = WeightsUploadType::LOCAL_MEM_BY_THREADS; } else if (device.IsAMD()) { weights_upload_type_ = WeightsUploadType::CONSTANT_MEM; diff --git a/tensorflow/lite/delegates/gpu/cl/selectors/convolution_transposed_selector.cc b/tensorflow/lite/delegates/gpu/cl/selectors/convolution_transposed_selector.cc index 12e99b57aa7..5fdfdca073e 100644 --- a/tensorflow/lite/delegates/gpu/cl/selectors/convolution_transposed_selector.cc +++ b/tensorflow/lite/delegates/gpu/cl/selectors/convolution_transposed_selector.cc @@ -112,6 +112,7 @@ absl::Status SelectConvolutionTransposed( case Vendor::POWERVR: case Vendor::NVIDIA: case Vendor::AMD: + case Vendor::INTEL: return SelectConvolutionTransposedPowerVR(attr, creation_context, op_def, ptr); case Vendor::MALI: diff --git a/tensorflow/lite/delegates/gpu/cl/selectors/fully_connected_selector.cc b/tensorflow/lite/delegates/gpu/cl/selectors/fully_connected_selector.cc index 12a1d726368..eacbea8b586 100644 --- a/tensorflow/lite/delegates/gpu/cl/selectors/fully_connected_selector.cc +++ b/tensorflow/lite/delegates/gpu/cl/selectors/fully_connected_selector.cc @@ -109,6 +109,9 @@ absl::Status SelectFullyConnected(const FullyConnectedAttributes& attr, return SelectFullyConnectedAdreno(attr, creation_context, op_def, batch_size, ptr); case Vendor::POWERVR: + case Vendor::AMD: + case Vendor::NVIDIA: + case Vendor::INTEL: return SelectFullyConnectedPowerVR(attr, creation_context, op_def, batch_size, ptr); case Vendor::MALI: From 2a6d9a1e811f2c0d459839854597f8ca6b5c8750 Mon Sep 17 00:00:00 2001 From: Jose Baiocchi Date: Wed, 17 Jun 2020 17:18:58 -0700 Subject: [PATCH 135/867] Add test for op_metrics_db_utils IdleTimeRatio PiperOrigin-RevId: 316997793 Change-Id: I0b89e697d661b2ea1fb95e7a6d51800b407617d0 --- tensorflow/core/profiler/utils/BUILD | 11 +++++ .../utils/op_metrics_db_utils_test.cc | 46 +++++++++++++++++++ 2 files changed, 57 insertions(+) create mode 100644 tensorflow/core/profiler/utils/op_metrics_db_utils_test.cc diff --git a/tensorflow/core/profiler/utils/BUILD b/tensorflow/core/profiler/utils/BUILD index ece58802661..0262c5659b7 100644 --- a/tensorflow/core/profiler/utils/BUILD +++ b/tensorflow/core/profiler/utils/BUILD @@ -77,6 +77,17 @@ cc_library( ], ) +tf_cc_test( + name = "op_metrics_db_utils_test", + srcs = ["op_metrics_db_utils_test.cc"], + deps = [ + ":op_metrics_db_utils", + "//tensorflow/core:test", + "//tensorflow/core:test_main", + "//tensorflow/core/profiler/protobuf:op_metrics_proto_cc", + ], +) + cc_library( name = "op_utils", srcs = ["op_utils.cc"], diff --git a/tensorflow/core/profiler/utils/op_metrics_db_utils_test.cc b/tensorflow/core/profiler/utils/op_metrics_db_utils_test.cc new file mode 100644 index 00000000000..12c68426b2e --- /dev/null +++ b/tensorflow/core/profiler/utils/op_metrics_db_utils_test.cc @@ -0,0 +1,46 @@ +/* Copyright 2020 The TensorFlow Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. +==============================================================================*/ + +#include "tensorflow/core/profiler/utils/op_metrics_db_utils.h" + +#include "tensorflow/core/platform/test.h" +#include "tensorflow/core/profiler/protobuf/op_metrics.pb.h" + +namespace tensorflow { +namespace profiler { +namespace { + +constexpr double kMaxError = 1E-10; + +TEST(OpMetricsDbTest, IdleTimeRatio) { + OpMetricsDb metrics_db_0; + metrics_db_0.set_total_time_ps(100000000); + metrics_db_0.set_total_op_time_ps(60000000); + EXPECT_NEAR(0.4, IdleTimeRatio(metrics_db_0), kMaxError); + + OpMetricsDb metrics_db_1; + metrics_db_1.set_total_time_ps(200000000); + metrics_db_1.set_total_op_time_ps(150000000); + EXPECT_NEAR(0.25, IdleTimeRatio(metrics_db_1), kMaxError); + + OpMetricsDb metrics_db_2; + metrics_db_1.set_total_time_ps(0); + metrics_db_1.set_total_op_time_ps(0); + EXPECT_NEAR(1.0, IdleTimeRatio(metrics_db_2), kMaxError); +} + +} // namespace +} // namespace profiler +} // namespace tensorflow From 73cf8263c771f1813416af9310469c7041e7637e Mon Sep 17 00:00:00 2001 From: Raman Sarokin Date: Wed, 17 Jun 2020 17:21:24 -0700 Subject: [PATCH 136/867] Pooling converted to new style. Merged 2D and 3D versions into one. PiperOrigin-RevId: 316998142 Change-Id: I92c020476f085e6160a02282c1edafabdd72ca30 --- .../lite/delegates/gpu/cl/kernels/pooling.cc | 671 +++++++----------- .../lite/delegates/gpu/cl/kernels/pooling.h | 40 +- .../lite/delegates/gpu/cl/kernels/util.cc | 9 - .../lite/delegates/gpu/cl/kernels/util.h | 5 - .../lite/delegates/gpu/cl/tensor_type.cc | 36 +- .../lite/delegates/gpu/cl/tensor_type.h | 10 + 6 files changed, 313 insertions(+), 458 deletions(-) diff --git a/tensorflow/lite/delegates/gpu/cl/kernels/pooling.cc b/tensorflow/lite/delegates/gpu/cl/kernels/pooling.cc index e292f2dad7d..922d484c57d 100644 --- a/tensorflow/lite/delegates/gpu/cl/kernels/pooling.cc +++ b/tensorflow/lite/delegates/gpu/cl/kernels/pooling.cc @@ -25,366 +25,307 @@ namespace gpu { namespace cl { namespace { -std::string GetAveragePoolingKernelCode( - const OperationDef& op_def, bool stride_correction, const CLDevice& device, - const std::vector& linked_operations) { - TensorCodeGenerator src_tensor( - "src_data", WHSPoint{"src_size.x", "src_size.y", "src_size.z"}, - op_def.src_tensors[0]); - TensorCodeGenerator dst_tensor( - "dst_data", WHSPoint{"dst_size.x", "dst_size.y", "dst_size.z"}, - op_def.dst_tensors[0]); +std::string GetAveragePoolingKernelCode(const OperationDef& op_def, + bool stride_correction, + const CLDevice& device, + Arguments* args) { + auto src_desc = absl::make_unique(op_def.src_tensors[0]); + src_desc->SetTextureAddressMode(GetFastestZeroMode(device)); + if (op_def.IsBatchSupported()) { + src_desc->SetStateVar("BatchedWidth", "true"); + } + args->AddObjectRef("src_tensor", AccessType::READ, std::move(src_desc)); + auto dst_desc = absl::make_unique(op_def.dst_tensors[0]); + if (op_def.IsBatchSupported()) { + dst_desc->SetStateVar("BatchedWidth", "true"); + } + args->AddObjectRef("dst_tensor", AccessType::WRITE, std::move(dst_desc)); + if (op_def.dst_tensors[0].HasAxis(Axis::WIDTH)) { + args->AddInt("kernel_size_x"); + args->AddInt("padding_x"); + args->AddInt("stride_x"); + } + if (op_def.dst_tensors[0].HasAxis(Axis::HEIGHT)) { + args->AddInt("kernel_size_y"); + args->AddInt("padding_y"); + args->AddInt("stride_y"); + } + if (op_def.dst_tensors[0].HasAxis(Axis::DEPTH)) { + args->AddInt("kernel_size_z"); + args->AddInt("padding_z"); + args->AddInt("stride_z"); + } - const auto address_mode = GetFastestZeroMode(device); + std::map axis_to_src_coord = { + {Axis::WIDTH, "x_c"}, {Axis::HEIGHT, "y_c"}, {Axis::DEPTH, "d_c"}, + {Axis::CHANNELS, "Z"}, {Axis::BATCH, "B"}, + }; - std::string c = GetCommonDefines(op_def.precision); + std::map axis_to_dst_coord = { + {Axis::WIDTH, "X"}, {Axis::HEIGHT, "Y"}, {Axis::DEPTH, "D"}, + {Axis::CHANNELS, "Z"}, {Axis::BATCH, "B"}, + }; + + std::vector src_coords; + std::vector dst_coords; + for (auto axis : {Axis::WIDTH, Axis::HEIGHT, Axis::DEPTH, Axis::CHANNELS}) { + if (op_def.dst_tensors[0].HasAxis(axis)) { + dst_coords.push_back(axis_to_dst_coord[axis]); + } + if (op_def.src_tensors[0].HasAxis(axis)) { + src_coords.push_back(axis_to_src_coord[axis]); + } + } + std::string src_coord = src_coords[0]; + for (int i = 1; i < src_coords.size(); ++i) { + src_coord += ", " + src_coords[i]; + } + std::string dst_coord = dst_coords[0]; + for (int i = 1; i < dst_coords.size(); ++i) { + dst_coord += ", " + dst_coords[i]; + } const bool manual_clamp = op_def.src_tensors[0].storage_type == TensorStorageType::BUFFER || op_def.src_tensors[0].storage_type == TensorStorageType::IMAGE_BUFFER; + std::string c = GetCommonDefines(op_def.precision); c += "__kernel void main_function(\n"; - c += src_tensor.GetDeclaration(AccessType::READ); - c += GetArgsDeclaration(linked_operations); - c += dst_tensor.GetDeclaration(AccessType::WRITE) + ",\n"; - c += " int4 src_size, \n"; - c += " int4 dst_size, \n"; - c += " int2 kernel_size, \n"; - c += " int2 padding, \n"; - c += " int2 stride \n"; - c += ") {\n"; + c += "$0) {\n"; c += " int X = get_global_id(0);\n"; - c += " int Y = get_global_id(1);\n"; + if (op_def.dst_tensors[0].HasAxis(Axis::DEPTH)) { + c += " int linear_id_1 = get_global_id(1);\n"; + c += " int Y = linear_id_1 / args.dst_tensor.Depth();\n"; + c += " int D = linear_id_1 % args.dst_tensor.Depth();\n"; + } else { + c += " int Y = get_global_id(1);\n"; + } c += " int Z = get_global_id(2);\n"; - c += " if (X >= dst_size.x || Y >= dst_size.y || Z >= dst_size.z) return;\n"; + c += " if (X >= args.dst_tensor.Width() || Y >= args.dst_tensor.Height() || " + "Z >= args.dst_tensor.Slices()) { \n"; + c += " return; \n"; + c += " } \n"; c += " float4 r = (float4)(0.0f);\n"; c += " float window_size = 0.0;\n"; if (stride_correction) { c += " int xs = " + - GetXStrideCorrected("X", "src_size.w", "stride.x", "padding.x") + + GetXStrideCorrected("X", "args.src_tensor.Batch()", "args.stride_x", + "args.padding_x") + ";\n"; } else { - c += " int xs = X * stride.x + padding.x;\n"; + c += " int xs = X * args.stride_x + args.padding_x;\n"; } - c += " int ys = Y * stride.y + padding.y;\n"; - c += " for (int ky = 0; ky < kernel_size.y; ++ky) {\n"; + c += " int ys = Y * args.stride_y + args.padding_y;\n"; + if (op_def.dst_tensors[0].HasAxis(Axis::DEPTH)) { + c += " int ds = D * args.stride_z + args.padding_z;\n"; + c += " for (int kz = 0; kz < args.kernel_size_z; ++kz) {\n"; + c += " int d_c = ds + kz;\n"; + c += " if (d_c < 0 || d_c >= args.src_tensor.Depth()) continue;\n"; + } + c += " for (int ky = 0; ky < args.kernel_size_y; ++ky) {\n"; c += " int y_c = ys + ky;\n"; - c += " bool outside_y = y_c < 0 || y_c >= src_size.y;\n"; - c += " for (int kx = 0; kx < kernel_size.x; ++kx) {\n"; + c += " bool outside_y = y_c < 0 || y_c >= args.src_tensor.Height();\n"; + c += " for (int kx = 0; kx < args.kernel_size_x; ++kx) {\n"; if (op_def.IsBatchSupported()) { - c += " int x_c = xs + kx * src_size.w;\n"; + c += " int x_c = xs + kx * args.src_tensor.Batch();\n"; } else { c += " int x_c = xs + kx;\n"; } - c += " bool outside = outside_y || x_c < 0 || x_c >= src_size.x;\n"; + c += " bool outside = outside_y || x_c < 0 || x_c >= " + "args.src_tensor.Width();\n"; if (manual_clamp) { - c += " r += !outside ? " + - src_tensor.ReadAsFloatWHS("x_c", "y_c", "Z") + " : (float4)(0.0f);\n"; + c += " r += !outside ? args.src_tensor.Read(" + src_coord + + ") : " + "(float4)(0.0f);\n"; } else { - c += " r += " + - src_tensor.ReadAsFloatWHS("x_c", "y_c", "Z", address_mode) + ";\n"; + c += " r += args.src_tensor.Read(" + src_coord + ");\n"; } c += " window_size += !outside ? 1.0 : 0.0;\n"; c += " }\n"; c += " }\n"; + if (op_def.dst_tensors[0].HasAxis(Axis::DEPTH)) { + c += " } // Depth\n"; + } // If window_size==0, window covered nothing. This situation is a sign of // incorrectly constructed operation. NaNs are expected as output. c += " FLT4 result = TO_FLT4(r / window_size);\n"; - const LinkingContext context{"result", "X", "Y", "Z"}; - c += PostProcess(linked_operations, context); - c += " " + dst_tensor.WriteWHS("result", "X", "Y", "Z"); + c += " args.dst_tensor.Write(result, " + dst_coord + ");\n"; c += "}\n"; return c; } -std::string GetAveragePooling3DKernelCode( - const OperationDef& op_def, bool stride_correction, const CLDevice& device, - const std::vector& linked_operations) { - TensorCodeGenerator src_tensor( - "src_data", - WHDSPoint{"src_size.x", "src_size.y", "src_size.z", "src_size.w"}, - op_def.src_tensors[0]); - TensorCodeGenerator dst_tensor( - "dst_data", - WHDSPoint{"dst_size.x", "dst_size.y", "dst_size.z", "dst_size.w"}, - op_def.dst_tensors[0]); - - const auto address_mode = GetFastestZeroMode(device); - - std::string c = GetCommonDefines(op_def.precision); - - c += "__kernel void main_function(\n"; - c += src_tensor.GetDeclaration(AccessType::READ); - c += GetArgsDeclaration(linked_operations); - c += dst_tensor.GetDeclaration(AccessType::WRITE) + ",\n"; - c += " int4 src_size, \n"; - c += " int4 dst_size, \n"; +std::string GetMaxPoolingKernelCode(const OperationDef& op_def, + bool stride_correction, bool output_indices, + Arguments* args) { + auto src_desc = absl::make_unique(op_def.src_tensors[0]); if (op_def.IsBatchSupported()) { - c += " int batch_size, \n"; + src_desc->SetStateVar("BatchedWidth", "true"); } - c += " int4 kernel_size, \n"; - c += " int4 padding, \n"; - c += " int4 stride \n"; - c += ") {\n"; - c += " int X = get_global_id(0);\n"; - c += " int Y = get_global_id(1);\n"; - c += " int linear_id_z = get_global_id(2);\n"; - c += " int S = linear_id_z % dst_size.w;\n"; - c += " int Z = linear_id_z / dst_size.w;\n"; - c += " if (X >= dst_size.x || Y >= dst_size.y || Z >= dst_size.z) return;\n"; - c += " float4 r = (float4)(0.0f);\n"; - c += " float window_size = 0.0;\n"; - if (stride_correction) { - c += " int xs = " + - GetXStrideCorrected("X", "batch_size", "stride.x", "padding.x") + - ";\n"; - } else { - c += " int xs = X * stride.x + padding.x;\n"; - } - c += " int ys = Y * stride.y + padding.y;\n"; - c += " int zs = Z * stride.z + padding.z;\n"; - c += " for (int kz = 0; kz < kernel_size.z; ++kz) {\n"; - c += " int z_c = zs + kz;\n"; - c += " if (z_c < 0 || z_c >= src_size.z) continue;\n"; - c += " for (int ky = 0; ky < kernel_size.y; ++ky) {\n"; - c += " int y_c = ys + ky;\n"; - c += " if (y_c < 0 || y_c >= src_size.y) continue;\n"; - c += " for (int kx = 0; kx < kernel_size.x; ++kx) {\n"; + args->AddObjectRef("src_tensor", AccessType::READ, std::move(src_desc)); + auto dst_desc = absl::make_unique(op_def.dst_tensors[0]); if (op_def.IsBatchSupported()) { - c += " int x_c = xs + kx * batch_size;\n"; - } else { - c += " int x_c = xs + kx;\n"; + dst_desc->SetStateVar("BatchedWidth", "true"); } - c += " if(x_c < 0 || x_c >= src_size.x) continue;\n"; - c += " r += " + - src_tensor.ReadAsFloatWHDS("x_c", "y_c", "z_c", "S", address_mode) + - ";\n"; - c += " window_size += 1.0;\n"; - c += " }\n"; - c += " }\n"; - c += " }\n"; - // If window_size==0, window covered nothing. This situation is a sign of - // incorrectly constructed operation. NaNs are expected as output. - c += " FLT4 result = TO_FLT4(r / window_size);\n"; - const LinkingContext context{"result", "X", "Y", "Z"}; - c += PostProcess(linked_operations, context); - c += " " + dst_tensor.WriteWHDS("result", "X", "Y", "Z", "S"); - c += "}\n"; - - return c; -} - -std::string GetMaxPoolingKernelCode( - const OperationDef& op_def, bool stride_correction, - const std::vector& linked_operations, - bool output_indices) { - TensorCodeGenerator src_tensor( - "src_data", WHSPoint{"src_size.x", "src_size.y", "src_size.z"}, - op_def.src_tensors[0]); - TensorCodeGenerator dst_tensor( - "dst_data", WHSPoint{"dst_size.x", "dst_size.y", "dst_size.z"}, - op_def.dst_tensors[0]); - const auto dst_ind_def = - output_indices ? op_def.dst_tensors[1] : op_def.dst_tensors[0]; - TensorCodeGenerator indices_tensor( - "dst_indices", WHSPoint{"dst_size.x", "dst_size.y", "dst_size.z"}, - dst_ind_def); - - std::string c = GetCommonDefines(op_def.precision); - - c += "__kernel void main_function(\n"; - c += src_tensor.GetDeclaration(AccessType::READ); - c += GetArgsDeclaration(linked_operations); - c += dst_tensor.GetDeclaration(AccessType::WRITE) + ",\n"; + args->AddObjectRef("dst_tensor", AccessType::WRITE, std::move(dst_desc)); if (output_indices) { - c += indices_tensor.GetDeclaration(AccessType::WRITE) + ",\n"; + auto dst_ind_desc = + absl::make_unique(op_def.dst_tensors[1]); + if (op_def.IsBatchSupported()) { + dst_ind_desc->SetStateVar("BatchedWidth", "true"); + } + args->AddObjectRef("dst_indices", AccessType::WRITE, + std::move(dst_ind_desc)); } - c += " int4 src_size, \n"; - c += " int4 dst_size, \n"; - c += " int2 kernel_size, \n"; - c += " int2 padding, \n"; - c += " int2 stride \n"; - c += ") {\n"; + if (op_def.dst_tensors[0].HasAxis(Axis::WIDTH)) { + args->AddInt("kernel_size_x"); + args->AddInt("padding_x"); + args->AddInt("stride_x"); + } + if (op_def.dst_tensors[0].HasAxis(Axis::HEIGHT)) { + args->AddInt("kernel_size_y"); + args->AddInt("padding_y"); + args->AddInt("stride_y"); + } + if (op_def.dst_tensors[0].HasAxis(Axis::DEPTH)) { + args->AddInt("kernel_size_z"); + args->AddInt("padding_z"); + args->AddInt("stride_z"); + } + + std::map axis_to_src_coord = { + {Axis::WIDTH, "x_c"}, {Axis::HEIGHT, "y_c"}, {Axis::DEPTH, "d_c"}, + {Axis::CHANNELS, "Z"}, {Axis::BATCH, "B"}, + }; + + std::map axis_to_dst_coord = { + {Axis::WIDTH, "X"}, {Axis::HEIGHT, "Y"}, {Axis::DEPTH, "D"}, + {Axis::CHANNELS, "Z"}, {Axis::BATCH, "B"}, + }; + + std::vector src_coords; + std::vector dst_coords; + for (auto axis : {Axis::WIDTH, Axis::HEIGHT, Axis::DEPTH, Axis::CHANNELS}) { + if (op_def.dst_tensors[0].HasAxis(axis)) { + dst_coords.push_back(axis_to_dst_coord[axis]); + } + if (op_def.src_tensors[0].HasAxis(axis)) { + src_coords.push_back(axis_to_src_coord[axis]); + } + } + std::string src_coord = src_coords[0]; + for (int i = 1; i < src_coords.size(); ++i) { + src_coord += ", " + src_coords[i]; + } + std::string dst_coord = dst_coords[0]; + for (int i = 1; i < dst_coords.size(); ++i) { + dst_coord += ", " + dst_coords[i]; + } + + std::string c = GetCommonDefines(op_def.precision); + c += "__kernel void main_function(\n"; + c += "$0) {\n"; c += " int X = get_global_id(0);\n"; - c += " int Y = get_global_id(1);\n"; + if (op_def.dst_tensors[0].HasAxis(Axis::DEPTH)) { + c += " int linear_id_1 = get_global_id(1);\n"; + c += " int Y = linear_id_1 / args.dst_tensor.Depth();\n"; + c += " int D = linear_id_1 % args.dst_tensor.Depth();\n"; + } else { + c += " int Y = get_global_id(1);\n"; + } c += " int Z = get_global_id(2);\n"; - c += - " if (X >= dst_size.x || Y >= dst_size.y || Z >= dst_size.z) return; \n"; - c += " FLT4 maximum = (FLT4)(-10000.0f);\n"; - if (output_indices) { - c += " FLT4 indexes = (FLT4)(0.0f);\n"; - c += " FLT index_counter = (FLT)(0.1f);\n"; - } - if (stride_correction) { - c += " int xs = " + - GetXStrideCorrected("X", "src_size.w", "stride.x", "padding.x") + - ";\n"; - } else { - c += " int xs = X * stride.x + padding.x;\n"; - } - c += " int ys = Y * stride.y + padding.y;\n"; - c += " for (int ky = 0; ky < kernel_size.y; ++ky) {\n"; - c += " int y_c = ys + ky;\n"; - c += " bool outside_y = y_c < 0 || y_c >= src_size.y;\n"; - c += " for (int kx = 0; kx < kernel_size.x; ++kx) {\n"; - if (op_def.IsBatchSupported()) { - c += " int x_c = xs + kx * src_size.w;\n"; - } else { - c += " int x_c = xs + kx;\n"; - } - c += " bool outside_x = x_c < 0 || x_c >= src_size.x;\n"; - c += " if (!outside_x && !outside_y) {\n"; - c += " FLT4 src = " + src_tensor.ReadWHS("x_c", "y_c", "Z") + ";\n"; - if (output_indices) { - c += " if (src.x > maximum.x) {\n"; - c += " indexes.x = index_counter;\n"; - c += " maximum.x = src.x;\n"; - c += " }\n"; - c += " if (src.y > maximum.y) {\n"; - c += " indexes.y = index_counter;\n"; - c += " maximum.y = src.y;\n"; - c += " }\n"; - c += " if (src.z > maximum.z) {\n"; - c += " indexes.z = index_counter;\n"; - c += " maximum.z = src.z;\n"; - c += " }\n"; - c += " if (src.w > maximum.w) {\n"; - c += " indexes.w = index_counter;\n"; - c += " maximum.w = src.w;\n"; - c += " }\n"; - c += " index_counter += (FLT)(1.0f);\n"; - } else { - c += " maximum = max(src, maximum);\n"; - } - c += " }\n"; - c += " }\n"; - c += " }\n"; - const LinkingContext context{"maximum", "X", "Y", "Z"}; - c += PostProcess(linked_operations, context); - c += " " + dst_tensor.WriteWHS("maximum", "X", "Y", "Z"); - if (output_indices) { - c += " " + indices_tensor.WriteWHS("indexes", "X", "Y", "Z"); - } - c += "}\n"; - - return c; -} - -std::string GetMaxPooling3DKernelCode( - const OperationDef& op_def, bool stride_correction, - const std::vector& linked_operations, - bool output_indices) { - TensorCodeGenerator src_tensor( - "src_data", - WHDSPoint{"src_size.x", "src_size.y", "src_size.z", "src_size.w"}, - op_def.src_tensors[0]); - TensorCodeGenerator dst_tensor( - "dst_data", - WHDSPoint{"dst_size.x", "dst_size.y", "dst_size.z", "dst_size.w"}, - op_def.dst_tensors[0]); - const auto dst_ind_def = - output_indices ? op_def.dst_tensors[1] : op_def.dst_tensors[0]; - TensorCodeGenerator indices_tensor( - "dst_indices", - WHDSPoint{"dst_size.x", "dst_size.y", "dst_size.z", "dst_size.w"}, - dst_ind_def); - - std::string c = GetCommonDefines(op_def.precision); - - c += "__kernel void main_function(\n"; - c += src_tensor.GetDeclaration(AccessType::READ); - c += GetArgsDeclaration(linked_operations); - c += dst_tensor.GetDeclaration(AccessType::WRITE) + ",\n"; - if (output_indices) { - c += indices_tensor.GetDeclaration(AccessType::WRITE) + ",\n"; - } - c += " int4 src_size, \n"; - c += " int4 dst_size, \n"; - if (op_def.IsBatchSupported()) { - c += " int batch_size, \n"; - } - c += " int4 kernel_size, \n"; - c += " int4 padding, \n"; - c += " int4 stride \n"; - c += ") {\n"; - c += " int X = get_global_id(0);\n"; - c += " int Y = get_global_id(1);\n"; - c += " int linear_id_z = get_global_id(2);\n"; - c += " int S = linear_id_z % dst_size.w;\n"; - c += " int Z = linear_id_z / dst_size.w;\n"; - c += " if (X >= dst_size.x || Y >= dst_size.y || Z >= dst_size.z) return;\n"; + c += " if (X >= args.dst_tensor.Width() || Y >= args.dst_tensor.Height() || " + "Z >= args.dst_tensor.Slices()) { \n"; + c += " return; \n"; + c += " } \n"; c += " FLT4 maximum = (FLT4)(-10000.0f);\n"; if (output_indices) { c += " FLT4 indexes = (FLT4)(0.0f);\n"; } if (stride_correction) { c += " int xs = " + - GetXStrideCorrected("X", "batch_size", "stride.x", "padding.x") + + GetXStrideCorrected("X", "args.src_tensor.Batch()", "args.stride_x", + "args.padding_x") + ";\n"; } else { - c += " int xs = X * stride.x + padding.x;\n"; + c += " int xs = X * args.stride_x + args.padding_x;\n"; } - c += " int ys = Y * stride.y + padding.y;\n"; - c += " int zs = Z * stride.z + padding.z;\n"; - c += " for (int ky = 0; ky < kernel_size.y; ++ky) {\n"; + c += " int ys = Y * args.stride_y + args.padding_y;\n"; + c += " for (int ky = 0; ky < args.kernel_size_y; ++ky) {\n"; c += " int y_c = ys + ky;\n"; - c += " if (y_c < 0 || y_c >= src_size.y) continue;\n"; - c += " for (int kx = 0; kx < kernel_size.x; ++kx) {\n"; + c += " if (y_c < 0 || y_c >= args.src_tensor.Height()) continue;\n"; + c += " for (int kx = 0; kx < args.kernel_size_x; ++kx) {\n"; if (op_def.IsBatchSupported()) { - c += " int x_c = xs + kx * batch_size;\n"; + c += " int x_c = xs + kx * args.src_tensor.Batch();\n"; } else { c += " int x_c = xs + kx;\n"; } - c += " if (x_c < 0 || x_c >= src_size.x) continue;\n"; - c += " for (int kz = 0; kz < kernel_size.z; ++kz) {\n"; - c += " int z_c = zs + kz;\n"; - c += " if (z_c < 0 || z_c >= src_size.z) continue;\n"; - c += " FLT4 src = " + src_tensor.ReadWHDS("x_c", "y_c", "z_c", "S") + - ";\n"; - if (output_indices) { - c += " FLT index_counter = (FLT)((ky * kernel_size.x + kx) * " - "kernel_size.z + kz) + (FLT)(0.1f);\n"; - c += " if (src.x > maximum.x) {\n"; - c += " indexes.x = index_counter;\n"; - c += " maximum.x = src.x;\n"; - c += " }\n"; - c += " if (src.y > maximum.y) {\n"; - c += " indexes.y = index_counter;\n"; - c += " maximum.y = src.y;\n"; - c += " }\n"; - c += " if (src.z > maximum.z) {\n"; - c += " indexes.z = index_counter;\n"; - c += " maximum.z = src.z;\n"; - c += " }\n"; - c += " if (src.w > maximum.w) {\n"; - c += " indexes.w = index_counter;\n"; - c += " maximum.w = src.w;\n"; - c += " }\n"; - } else { - c += " maximum = max(src, maximum);\n"; + c += " if (x_c < 0 || x_c >= args.src_tensor.Width()) continue;\n"; + if (op_def.dst_tensors[0].HasAxis(Axis::DEPTH)) { + c += " int ds = D * args.stride_z + args.padding_z;\n"; + c += " for (int kz = 0; kz < args.kernel_size_z; ++kz) {\n"; + c += " int d_c = ds + kz;\n"; + c += " if (d_c < 0 || d_c >= args.src_tensor.Depth()) continue;\n"; + } + c += " FLT4 src = args.src_tensor.Read(" + src_coord + ");\n"; + if (output_indices) { + if (op_def.dst_tensors[0].HasAxis(Axis::DEPTH)) { + c += " FLT index_counter = (FLT)((ky * args.kernel_size_x + kx) * " + "args.kernel_size_z + kz) + (FLT)(0.1f);\n"; + } else { + c += " FLT index_counter = (FLT)(ky * args.kernel_size_x + kx) + " + "(FLT)(0.1f);\n"; + } + c += " if (src.x > maximum.x) {\n"; + c += " indexes.x = index_counter;\n"; + c += " maximum.x = src.x;\n"; + c += " }\n"; + c += " if (src.y > maximum.y) {\n"; + c += " indexes.y = index_counter;\n"; + c += " maximum.y = src.y;\n"; + c += " }\n"; + c += " if (src.z > maximum.z) {\n"; + c += " indexes.z = index_counter;\n"; + c += " maximum.z = src.z;\n"; + c += " }\n"; + c += " if (src.w > maximum.w) {\n"; + c += " indexes.w = index_counter;\n"; + c += " maximum.w = src.w;\n"; + c += " }\n"; + } else { + c += " maximum = max(src, maximum);\n"; + } + if (op_def.dst_tensors[0].HasAxis(Axis::DEPTH)) { + c += " } // Depth\n"; } - c += " };\n"; c += " }\n"; c += " }\n"; - const LinkingContext context{"maximum", "X", "Y", "Z"}; - c += PostProcess(linked_operations, context); - c += " " + dst_tensor.WriteWHDS("maximum", "X", "Y", "Z", "S"); + c += " args.dst_tensor.Write(maximum, " + dst_coord + ");\n"; if (output_indices) { - c += " " + indices_tensor.WriteWHDS("indexes", "X", "Y", "Z", "S"); + c += " args.dst_indices.Write(indexes, " + dst_coord + ");\n"; } c += "}\n"; + return c; } - } // namespace Pooling::Pooling(const OperationDef& definition, const Pooling2DAttributes& attr) : GPUOperation(definition), - stride_(attr.strides.w, attr.strides.h), - padding_(-attr.padding.prepended.w, -attr.padding.prepended.h), - kernel_size_(attr.kernel.w, attr.kernel.h), + stride_(attr.strides.w, attr.strides.h, 0, 0), + padding_(-attr.padding.prepended.w, -attr.padding.prepended.h, 0, 0), + kernel_size_(attr.kernel.w, attr.kernel.h, 0, 0), + type_(attr.type), + output_indices_(attr.output_indices) {} + +Pooling::Pooling(const OperationDef& definition, + const Pooling3DAttributes& attr) + : GPUOperation(definition), + stride_(attr.strides.w, attr.strides.h, attr.strides.d, 0), + padding_(-attr.padding.prepended.w, -attr.padding.prepended.h, + -attr.padding.prepended.d, 0), + kernel_size_(attr.kernel.w, attr.kernel.h, attr.kernel.d, 0), type_(attr.type), output_indices_(attr.output_indices) {} @@ -419,44 +360,56 @@ absl::Status Pooling::Compile(const CreationContext& creation_context) { switch (type_) { case PoolingType::AVERAGE: code = GetAveragePoolingKernelCode(definition_, stride_correction, - *creation_context.device, - linked_operations_); + *creation_context.device, &args_); break; case PoolingType::MAX: code = GetMaxPoolingKernelCode(definition_, stride_correction, - linked_operations_, output_indices_); + output_indices_, &args_); break; default: return absl::InvalidArgumentError( "You should create another kernel with this params"); break; } + std::string element_wise_code; + RETURN_IF_ERROR( + MergeOperations(linked_operations_, &args_, &element_wise_code)); + RETURN_IF_ERROR(args_.TransformToCLCode(creation_context.device->GetInfo(), + {{"dst_tensor", element_wise_code}}, + &code)); return creation_context.cache->GetOrCreateCLKernel( code, "main_function", *creation_context.context, *creation_context.device, &kernel_); } absl::Status Pooling::BindArguments() { - kernel_.ResetBindingCounter(); - RETURN_IF_ERROR(kernel_.SetMemoryAuto(src_[0]->GetMemoryPtr())); - RETURN_IF_ERROR(BindArgs(&kernel_, linked_operations_)); - RETURN_IF_ERROR(kernel_.SetMemoryAuto(dst_[0]->GetMemoryPtrForWriting())); - if (output_indices_) { - RETURN_IF_ERROR(kernel_.SetMemoryAuto(dst_[1]->GetMemoryPtrForWriting())); + RETURN_IF_ERROR(args_.SetObjectRef("src_tensor", src_[0])); + RETURN_IF_ERROR(args_.SetObjectRef("dst_tensor", dst_[0])); + if (definition_.dst_tensors[0].HasAxis(Axis::WIDTH)) { + RETURN_IF_ERROR(args_.SetInt("stride_x", stride_.x)); + RETURN_IF_ERROR(args_.SetInt("padding_x", padding_.x * src_[0]->Batch())); + RETURN_IF_ERROR(args_.SetInt("kernel_size_x", kernel_size_.x)); } - RETURN_IF_ERROR(kernel_.SetBytesAuto(src_[0]->GetWBatchedHSB())); - RETURN_IF_ERROR(kernel_.SetBytesAuto(dst_[0]->GetWBatchedHSB())); - RETURN_IF_ERROR(kernel_.SetBytesAuto(kernel_size_)); - RETURN_IF_ERROR( - kernel_.SetBytesAuto(int2(padding_.x * src_[0]->Batch(), padding_.y))); - RETURN_IF_ERROR(kernel_.SetBytesAuto(stride_)); - - return absl::OkStatus(); + if (definition_.dst_tensors[0].HasAxis(Axis::HEIGHT)) { + RETURN_IF_ERROR(args_.SetInt("stride_y", stride_.y)); + RETURN_IF_ERROR(args_.SetInt("padding_y", padding_.y)); + RETURN_IF_ERROR(args_.SetInt("kernel_size_y", kernel_size_.y)); + } + if (definition_.dst_tensors[0].HasAxis(Axis::DEPTH)) { + RETURN_IF_ERROR(args_.SetInt("stride_z", stride_.z)); + RETURN_IF_ERROR(args_.SetInt("padding_z", padding_.z)); + RETURN_IF_ERROR(args_.SetInt("kernel_size_z", kernel_size_.z)); + } + if (output_indices_) { + RETURN_IF_ERROR(args_.SetObjectRef("dst_indices", dst_[1])); + } + RETURN_IF_ERROR(SetArguments(linked_operations_, &args_)); + return args_.Bind(kernel_.kernel()); } int3 Pooling::GetGridSize() const { const int grid_x = dst_[0]->Width() * dst_[0]->Batch(); - const int grid_y = dst_[0]->Height(); + const int grid_y = dst_[0]->Height() * dst_[0]->Depth(); const int grid_z = dst_[0]->Slices(); return int3(grid_x, grid_y, grid_z); } @@ -476,107 +429,9 @@ Pooling CreatePooling(const OperationDef& definition, return Pooling(definition, attr); } -Pooling3D::Pooling3D(const OperationDef& definition, - const Pooling3DAttributes& attr) - : GPUOperation(definition), - stride_(attr.strides.w, attr.strides.h, attr.strides.d), - padding_(-attr.padding.prepended.w, -attr.padding.prepended.h, - -attr.padding.prepended.d), - kernel_size_(attr.kernel.w, attr.kernel.h, attr.kernel.d), - type_(attr.type), - output_indices_(attr.output_indices) {} - -Pooling3D::Pooling3D(Pooling3D&& kernel) - : GPUOperation(std::move(kernel)), - stride_(kernel.stride_), - padding_(kernel.padding_), - kernel_size_(kernel.kernel_size_), - type_(kernel.type_), - output_indices_(kernel.output_indices_), - kernel_(std::move(kernel.kernel_)), - work_group_size_(kernel.work_group_size_) {} - -Pooling3D& Pooling3D::operator=(Pooling3D&& kernel) { - if (this != &kernel) { - std::swap(stride_, kernel.stride_); - std::swap(padding_, kernel.padding_); - std::swap(kernel_size_, kernel.kernel_size_); - std::swap(type_, kernel.type_); - std::swap(output_indices_, kernel.output_indices_); - kernel_ = std::move(kernel.kernel_); - std::swap(work_group_size_, kernel.work_group_size_); - GPUOperation::operator=(std::move(kernel)); - } - return *this; -} - -absl::Status Pooling3D::Compile(const CreationContext& creation_context) { - std::string code; - const bool stride_correction = - definition_.IsBatchSupported() && stride_.x != 1; - switch (type_) { - case PoolingType::AVERAGE: - code = GetAveragePooling3DKernelCode(definition_, stride_correction, - *creation_context.device, - linked_operations_); - break; - case PoolingType::MAX: - code = GetMaxPooling3DKernelCode(definition_, stride_correction, - linked_operations_, output_indices_); - break; - default: - return absl::InvalidArgumentError( - "You should create another kernel with this params"); - break; - } - return creation_context.cache->GetOrCreateCLKernel( - code, "main_function", *creation_context.context, - *creation_context.device, &kernel_); -} - -absl::Status Pooling3D::BindArguments() { - kernel_.ResetBindingCounter(); - RETURN_IF_ERROR(kernel_.SetMemoryAuto(src_[0]->GetMemoryPtr())); - RETURN_IF_ERROR(BindArgs(&kernel_, linked_operations_)); - RETURN_IF_ERROR(kernel_.SetMemoryAuto(dst_[0]->GetMemoryPtrForWriting())); - if (output_indices_) { - RETURN_IF_ERROR(kernel_.SetMemoryAuto(dst_[1]->GetMemoryPtrForWriting())); - } - RETURN_IF_ERROR(kernel_.SetBytesAuto(src_[0]->GetWBatchedHDS())); - RETURN_IF_ERROR(kernel_.SetBytesAuto(dst_[0]->GetWBatchedHDS())); - if (definition_.IsBatchSupported()) { - RETURN_IF_ERROR(kernel_.SetBytesAuto(src_[0]->Batch())); - } - RETURN_IF_ERROR(kernel_.SetBytesAuto( - int4(kernel_size_.x, kernel_size_.y, kernel_size_.z, 1))); - RETURN_IF_ERROR(kernel_.SetBytesAuto( - int4(padding_.x * src_[0]->Batch(), padding_.y, padding_.z, 1))); - RETURN_IF_ERROR( - kernel_.SetBytesAuto(int4(stride_.x, stride_.y, stride_.z, 1))); - - return absl::OkStatus(); -} - -int3 Pooling3D::GetGridSize() const { - const int grid_x = dst_[0]->Width() * dst_[0]->Batch(); - const int grid_y = dst_[0]->Height(); - const int grid_z = dst_[0]->Slices() * dst_[0]->Depth(); - return int3(grid_x, grid_y, grid_z); -} - -absl::Status Pooling3D::Tune(const TuningParameters& params) { - RETURN_IF_ERROR(BindArguments()); - return GetBestWorkGroup(params, kernel_, GetGridSize(), &work_group_size_); -} - -absl::Status Pooling3D::AddToQueue(CLCommandQueue* queue) { - RETURN_IF_ERROR(BindArguments()); - return queue->DispatchImplicit(kernel_, GetGridSize(), work_group_size_); -} - -Pooling3D CreatePooling3D(const OperationDef& definition, - const Pooling3DAttributes& attr) { - return Pooling3D(definition, attr); +Pooling CreatePooling(const OperationDef& definition, + const Pooling3DAttributes& attr) { + return Pooling(definition, attr); } } // namespace cl diff --git a/tensorflow/lite/delegates/gpu/cl/kernels/pooling.h b/tensorflow/lite/delegates/gpu/cl/kernels/pooling.h index 09d2d5260f7..20719c90ae3 100644 --- a/tensorflow/lite/delegates/gpu/cl/kernels/pooling.h +++ b/tensorflow/lite/delegates/gpu/cl/kernels/pooling.h @@ -30,6 +30,7 @@ namespace cl { class Pooling : public GPUOperation { public: Pooling(const OperationDef& definition, const Pooling2DAttributes& attr); + Pooling(const OperationDef& definition, const Pooling3DAttributes& attr); absl::Status AddToQueue(CLCommandQueue* queue) override; absl::Status Tune(const TuningParameters& params) override; @@ -45,9 +46,9 @@ class Pooling : public GPUOperation { absl::Status BindArguments(); int3 GetGridSize() const; - int2 stride_; - int2 padding_; - int2 kernel_size_; + int4 stride_; + int4 padding_; + int4 kernel_size_; PoolingType type_; bool output_indices_; @@ -59,37 +60,8 @@ class Pooling : public GPUOperation { Pooling CreatePooling(const OperationDef& definition, const Pooling2DAttributes& attr); -class Pooling3D : public GPUOperation { - public: - Pooling3D(const OperationDef& definition, const Pooling3DAttributes& attr); - absl::Status AddToQueue(CLCommandQueue* queue) override; - absl::Status Tune(const TuningParameters& params) override; - - absl::Status Compile(const CreationContext& creation_context) override; - - // Move only - Pooling3D(Pooling3D&& kernel); - Pooling3D& operator=(Pooling3D&& kernel); - Pooling3D(const Pooling3D&) = delete; - Pooling3D& operator=(const Pooling3D&) = delete; - - private: - absl::Status BindArguments(); - int3 GetGridSize() const; - - int3 stride_; - int3 padding_; - int3 kernel_size_; - - PoolingType type_; - bool output_indices_; - - CLKernel kernel_; - int3 work_group_size_ = int3(8, 4, 1); -}; - -Pooling3D CreatePooling3D(const OperationDef& definition, - const Pooling3DAttributes& attr); +Pooling CreatePooling(const OperationDef& definition, + const Pooling3DAttributes& attr); } // namespace cl } // namespace gpu diff --git a/tensorflow/lite/delegates/gpu/cl/kernels/util.cc b/tensorflow/lite/delegates/gpu/cl/kernels/util.cc index 3161a73a18f..8cfff18b4ee 100644 --- a/tensorflow/lite/delegates/gpu/cl/kernels/util.cc +++ b/tensorflow/lite/delegates/gpu/cl/kernels/util.cc @@ -60,15 +60,6 @@ std::string GetImageModifier(AccessType access) { } } -std::string TextureAddressModeToString(TextureAddressMode address_mode) { - switch (address_mode) { - case TextureAddressMode::DONT_CARE: - return "smp_none"; - case TextureAddressMode::ZERO: - return "smp_zero"; - } -} - } // namespace std::string GetCommonDefines(CalculationsPrecision precision) { diff --git a/tensorflow/lite/delegates/gpu/cl/kernels/util.h b/tensorflow/lite/delegates/gpu/cl/kernels/util.h index 836a95f7407..3a51d064b40 100644 --- a/tensorflow/lite/delegates/gpu/cl/kernels/util.h +++ b/tensorflow/lite/delegates/gpu/cl/kernels/util.h @@ -36,11 +36,6 @@ namespace cl { std::string GetCommonDefines(CalculationsPrecision precision); -enum class TextureAddressMode { - DONT_CARE, // translated to CLK_ADDRESS_NONE - ZERO, // translated to CLK_ADDRESS_CLAMP -}; - struct WHSPoint { std::string w_name; std::string h_name; diff --git a/tensorflow/lite/delegates/gpu/cl/tensor_type.cc b/tensorflow/lite/delegates/gpu/cl/tensor_type.cc index 3b6c686a99a..8e048675697 100644 --- a/tensorflow/lite/delegates/gpu/cl/tensor_type.cc +++ b/tensorflow/lite/delegates/gpu/cl/tensor_type.cc @@ -45,6 +45,15 @@ std::string GetWriteImageFromDataType(DataType data_type) { } // namespace +std::string TextureAddressModeToString(TextureAddressMode address_mode) { + switch (address_mode) { + case TextureAddressMode::DONT_CARE: + return "smp_none"; + case TextureAddressMode::ZERO: + return "smp_zero"; + } +} + std::string ToString(TensorStorageType type) { switch (type) { case TensorStorageType::UNKNOWN: @@ -271,8 +280,10 @@ std::string TensorDescriptor::Read(DataType read_as_type, case TensorStorageType::TEXTURE_3D: case TensorStorageType::SINGLE_TEXTURE_2D: case TensorStorageType::TEXTURE_ARRAY: - return absl::StrCat(read_as, "(", image_type, ", smp_none, ", - global_address, ")"); + return absl::StrCat( + read_as, "(", image_type, + ", " + TextureAddressModeToString(ModeFromState()) + ", ", + global_address, ")"); case TensorStorageType::IMAGE_BUFFER: return absl::StrCat(read_as, "(image_buffer, ", global_address, ")"); case TensorStorageType::UNKNOWN: @@ -500,6 +511,14 @@ bool TensorDescriptor::HasAxis(Axis axis) const { return false; } +void TensorDescriptor::SetTextureAddressMode(TextureAddressMode mode) { + if (mode == TextureAddressMode::ZERO) { + state_vars_["TextureMode"] = "ZERO"; + } else { + state_vars_["TextureMode"] = "DONT_CARE"; + } +} + bool TensorDescriptor::ParseCoordsFromArgs(const std::vector& args, int offset, std::string* xc, std::string* yc, std::string* zc, @@ -549,6 +568,19 @@ bool TensorDescriptor::IsBatchedWidth() const { return it != state_vars_.end() && it->second == "true"; } +TextureAddressMode TensorDescriptor::ModeFromState() const { + auto it = state_vars_.find("TextureMode"); + if (it != state_vars_.end()) { + if (it->second == "ZERO") { + return TextureAddressMode::ZERO; + } else { + return TextureAddressMode::DONT_CARE; + } + } else { + return TextureAddressMode::DONT_CARE; + } +} + } // namespace cl } // namespace gpu } // namespace tflite diff --git a/tensorflow/lite/delegates/gpu/cl/tensor_type.h b/tensorflow/lite/delegates/gpu/cl/tensor_type.h index 58ebfc51ec4..2d4ae0c7335 100644 --- a/tensorflow/lite/delegates/gpu/cl/tensor_type.h +++ b/tensorflow/lite/delegates/gpu/cl/tensor_type.h @@ -27,6 +27,13 @@ namespace tflite { namespace gpu { namespace cl { +enum class TextureAddressMode { + DONT_CARE, // translated to CLK_ADDRESS_NONE + ZERO, // translated to CLK_ADDRESS_CLAMP +}; + +std::string TextureAddressModeToString(TextureAddressMode address_mode); + enum class TensorStorageType { UNKNOWN, BUFFER, @@ -71,6 +78,7 @@ struct TensorDescriptor : public GPUObjectDescriptor { GPUResources GetGPUResources(AccessType access_type) const override; bool HasAxis(Axis axis) const; + void SetTextureAddressMode(TextureAddressMode mode); absl::Status GetLinkingContextFromWriteSelector( const std::vector& args, std::string* value_name, @@ -106,6 +114,8 @@ struct TensorDescriptor : public GPUObjectDescriptor { bool IsBatchedWidth() const; + TextureAddressMode ModeFromState() const; + absl::Status GetDataTypeFromTemplateArgs(const std::string& template_arg, DataType* result) const; From b1933d67e5507bc1433c6de1700bcb0b8f3a6aec Mon Sep 17 00:00:00 2001 From: Raman Sarokin Date: Wed, 17 Jun 2020 17:22:03 -0700 Subject: [PATCH 137/867] Added support of different layouts for src/dst. For example src - HWC, dst - BHWC, or vice versa. PiperOrigin-RevId: 316998239 Change-Id: I89b07923020f185c356bb0b63926bbe81be55cb5 --- .../lite/delegates/gpu/cl/kernels/strided_slice.cc | 10 +++++----- tensorflow/lite/delegates/gpu/cl/kernels/transpose.cc | 9 +++++---- 2 files changed, 10 insertions(+), 9 deletions(-) diff --git a/tensorflow/lite/delegates/gpu/cl/kernels/strided_slice.cc b/tensorflow/lite/delegates/gpu/cl/kernels/strided_slice.cc index 2cf65f24447..d0c4e432f3a 100644 --- a/tensorflow/lite/delegates/gpu/cl/kernels/strided_slice.cc +++ b/tensorflow/lite/delegates/gpu/cl/kernels/strided_slice.cc @@ -42,11 +42,12 @@ std::string GetStridedSliceCode(const OperationDef& op_def, bool alignedx4, args->AddInt("stride_z"); args->AddInt("stride_b"); - const std::string dst_batch = op_def.IsBatchSupported() ? "B" : ""; + const std::string batch_id = + op_def.dst_tensors[0].HasAxis(Axis::BATCH) ? "B" : "0"; std::string c = GetCommonDefines(op_def.precision); c += "__kernel void main_function(\n"; c += "$0) {\n"; - if (op_def.IsBatchSupported()) { + if (op_def.dst_tensors[0].HasAxis(Axis::BATCH)) { c += " int linear_id = get_global_id(0);\n"; c += " int X = linear_id / args.dst_tensor.Batch();\n"; c += " int B = linear_id % args.dst_tensor.Batch();\n"; @@ -62,11 +63,10 @@ std::string GetStridedSliceCode(const OperationDef& op_def, bool alignedx4, c += " } \n"; c += " int s_x = X * args.stride_x + args.offset_x;\n"; c += " int s_y = Y * args.stride_y + args.offset_y;\n"; - if (op_def.IsBatchSupported()) { - c += " int s_b = B * args.stride_b + args.offset_b;\n"; + if (op_def.src_tensors[0].HasAxis(Axis::BATCH)) { + c += " int s_b = " + batch_id + " * args.stride_b + args.offset_b;\n"; c += " args.src_tensor.SetBatchRef(s_b);\n"; } - const std::string src_batch = op_def.IsBatchSupported() ? "s_b" : ""; if (alignedx4) { c += " int s_z = Z + args.offset_z;\n"; c += " FLT4 result = args.src_tensor.Read(s_x, s_y, s_z);\n"; diff --git a/tensorflow/lite/delegates/gpu/cl/kernels/transpose.cc b/tensorflow/lite/delegates/gpu/cl/kernels/transpose.cc index e12c44566b7..cacfd52542d 100644 --- a/tensorflow/lite/delegates/gpu/cl/kernels/transpose.cc +++ b/tensorflow/lite/delegates/gpu/cl/kernels/transpose.cc @@ -36,11 +36,12 @@ std::string GetTransposeCode( "dst_tensor", AccessType::WRITE, absl::make_unique(op_def.dst_tensors[0])); - const std::string batch_id = op_def.IsBatchSupported() ? "B" : ""; + const std::string batch_id = + op_def.dst_tensors[0].HasAxis(Axis::BATCH) ? "B" : "0"; std::string c = GetCommonDefines(op_def.precision); c += "__kernel void main_function(\n"; c += "$0) {\n"; - if (op_def.IsBatchSupported()) { + if (op_def.dst_tensors[0].HasAxis(Axis::BATCH)) { c += " int linear_id = get_global_id(0);\n"; c += " int X = linear_id / args.dst_tensor.Batch();\n"; c += " int B = linear_id % args.dst_tensor.Batch();\n"; @@ -65,7 +66,7 @@ std::string GetTransposeCode( remap[attr.perm.w] = 2; remap[attr.perm.c] = 3; if (attr.perm.c == 3) { // optimized reading when no channels permutation - const std::string bhw[] = {"B", "Y", "X"}; + const std::string bhw[] = {batch_id, "Y", "X"}; if (op_def.src_tensors[0].HasAxis(Axis::BATCH)) { c += " args.src_tensor.SetBatchRef(" + bhw[remap[0]] + ");\n"; } @@ -80,7 +81,7 @@ std::string GetTransposeCode( c += " for (int i = 0; i < 4; ++i) {\n"; c += " int dst_channel = Z * 4 + i;\n"; c += " if (dst_channel < args.dst_tensor.Channels()) {\n"; - const std::string bhwc[] = {"B", "Y", "X", "dst_channel"}; + const std::string bhwc[] = {batch_id, "Y", "X", "dst_channel"}; if (op_def.src_tensors[0].HasAxis(Axis::BATCH)) { c += " args.src_tensor.SetBatchRef(" + bhwc[remap[0]] + ");\n"; } From 274a0f944e50922759fe8c263f74beeb24884137 Mon Sep 17 00:00:00 2001 From: Raman Sarokin Date: Wed, 17 Jun 2020 17:29:06 -0700 Subject: [PATCH 138/867] ConcatXY converted to new style. Added support of concatenation in Batch and Depth axis. PiperOrigin-RevId: 316999295 Change-Id: I94f2168f2861790b3a30c79b2b3476aa44c55748 --- .../delegates/gpu/cl/kernels/concat_xy.cc | 160 +++++++++++------- .../gpu/cl/selectors/simple_selectors.cc | 6 +- 2 files changed, 99 insertions(+), 67 deletions(-) diff --git a/tensorflow/lite/delegates/gpu/cl/kernels/concat_xy.cc b/tensorflow/lite/delegates/gpu/cl/kernels/concat_xy.cc index ef7915afba5..5476cc22965 100644 --- a/tensorflow/lite/delegates/gpu/cl/kernels/concat_xy.cc +++ b/tensorflow/lite/delegates/gpu/cl/kernels/concat_xy.cc @@ -15,7 +15,9 @@ limitations under the License. #include "tensorflow/lite/delegates/gpu/cl/kernels/concat_xy.h" +#include #include +#include #include "tensorflow/lite/delegates/gpu/cl/kernels/util.h" #include "tensorflow/lite/delegates/gpu/cl/kernels/work_group_picking.h" @@ -27,51 +29,93 @@ namespace gpu { namespace cl { namespace { -std::string GetConcatKernelCode( - const OperationDef& op_def, int tensors_count, - const std::vector& linked_operations) { - std::vector srcs(tensors_count); - for (int i = 0; i < tensors_count; ++i) { - const std::string tensor_name = "src_data_" + std::to_string(i); - const std::string width = "src_size_" + std::to_string(i) + ".x"; - const std::string height = "src_size_" + std::to_string(i) + ".y"; - srcs[i] = - TensorCodeGenerator(tensor_name, WHSPoint{width, height, "dst_size.z"}, - op_def.src_tensors[i]); +std::string GetConcatKernelCode(const OperationDef& op_def, + const ConcatAttributes& attr, Arguments* args) { + std::vector tensor_names(op_def.src_tensors.size()); + for (int i = 0; i < op_def.src_tensors.size(); ++i) { + tensor_names[i] = "src_tensor_" + std::to_string(i); + args->AddObjectRef( + tensor_names[i], AccessType::READ, + absl::make_unique(op_def.src_tensors[0])); + } + args->AddObjectRef( + "dst_tensor", AccessType::WRITE, + absl::make_unique(op_def.dst_tensors[0])); + + std::map axis_to_selector = { + {Axis::WIDTH, "Width"}, {Axis::HEIGHT, "Height"}, + {Axis::DEPTH, "Depth"}, {Axis::CHANNELS, "Channels"}, + {Axis::BATCH, "Batch"}, + }; + std::map axis_to_coord = { + {Axis::WIDTH, "X"}, {Axis::HEIGHT, "Y"}, {Axis::DEPTH, "D"}, + {Axis::CHANNELS, "S"}, {Axis::BATCH, "B"}, + }; + + std::vector src_coords; + std::vector dst_coords; + for (auto axis : + {Axis::WIDTH, Axis::HEIGHT, Axis::DEPTH, Axis::CHANNELS, Axis::BATCH}) { + if (op_def.src_tensors[0].HasAxis(axis) && axis != Axis::BATCH) { + if (axis == attr.axis) { + src_coords.push_back("coord"); + } else { + src_coords.push_back(axis_to_coord[axis]); + } + } + if (op_def.dst_tensors[0].HasAxis(axis)) { + dst_coords.push_back(axis_to_coord[axis]); + } + } + std::string src_coord = src_coords[0]; + for (int i = 1; i < src_coords.size(); ++i) { + src_coord += ", " + src_coords[i]; + } + std::string dst_coord = dst_coords[0]; + for (int i = 1; i < dst_coords.size(); ++i) { + dst_coord += ", " + dst_coords[i]; } - TensorCodeGenerator dst("dst_data", - WHSPoint{"dst_size.x", "dst_size.y", "dst_size.z"}, - op_def.dst_tensors[0]); std::string c = GetCommonDefines(op_def.precision); - c += "__kernel void main_function(\n"; - for (const auto& src : srcs) { - c += src.GetDeclaration(AccessType::READ) + ",\n"; + c += "$0) {\n"; + if (op_def.dst_tensors[0].HasAxis(Axis::BATCH)) { + c += " int linear_id_0 = get_global_id(0);\n"; + c += " int X = linear_id_0 / args.dst_tensor.Batch();\n"; + c += " int B = linear_id_0 % args.dst_tensor.Batch();\n"; + } else { + c += " int X = get_global_id(0);\n"; } - c += dst.GetDeclaration(AccessType::WRITE); - c += GetArgsDeclaration(linked_operations); - for (int i = 0; i < tensors_count; ++i) { - const std::string uniform_name = "src_size_" + std::to_string(i); - c += " int4 " + uniform_name + ",\n"; + if (op_def.dst_tensors[0].HasAxis(Axis::DEPTH)) { + c += " int linear_id_1 = get_global_id(1);\n"; + c += " int Y = linear_id_1 / args.dst_tensor.Depth();\n"; + c += " int D = linear_id_1 % args.dst_tensor.Depth();\n"; + } else { + c += " int Y = get_global_id(1);\n"; } - c += " int4 dst_size \n"; - c += ") {\n"; - c += " int X = get_global_id(0);\n"; - c += " int Y = get_global_id(1);\n"; - c += " int Z = get_global_id(2);\n"; - c += " if (Z >= dst_size.z) return;\n"; - for (int i = 0; i < tensors_count; ++i) { - const std::string size_name = "src_size_" + std::to_string(i); - c += " if (X < " + size_name + ".x && Y < " + size_name + ".y) { \n"; - c += " FLT4 result = " + srcs[i].ReadWHS("X", "Y", "Z") + ";\n"; - c += " int dst_x = X + " + size_name + ".z;\n"; - c += " int dst_y = Y + " + size_name + ".w;\n"; - const LinkingContext context{"result", "dst_x", "dst_y", "Z"}; - c += PostProcess(linked_operations, context); - c += " " + dst.WriteWHS("result", "dst_x", "dst_y", "Z"); + c += " int S = get_global_id(2);\n"; + c += " if (X >= args.dst_tensor.Width() || Y >= args.dst_tensor.Height() || " + "S >= args.dst_tensor.Slices()) { \n"; + c += " return; \n"; + c += " } \n"; + c += " FLT4 result = (FLT4)(0.0f);\n"; + c += " int coord = " + axis_to_coord[attr.axis] + ";\n"; + for (int i = 0; i < op_def.src_tensors.size(); ++i) { + const std::string field = + "args." + tensor_names[i] + "." + axis_to_selector[attr.axis] + "()"; + c += " if (coord >= 0 && coord < " + field + ") { \n"; + if (op_def.src_tensors[i].HasAxis(Axis::BATCH)) { + if (attr.axis == Axis::BATCH) { + c += " args." + tensor_names[i] + ".SetBatchRef(coord);\n"; + } else { + c += " args." + tensor_names[i] + ".SetBatchRef(B);\n"; + } + } + c += " result = args." + tensor_names[i] + ".Read(" + src_coord + ");\n"; c += " } \n"; + c += " coord -= " + field + ";\n"; } + c += " args.dst_tensor.Write(result, " + dst_coord + ");\n"; c += "}\n"; return c; } @@ -97,46 +141,32 @@ ConcatXY& ConcatXY::operator=(ConcatXY&& operation) { } absl::Status ConcatXY::Compile(const CreationContext& creation_context) { - const auto code = - GetConcatKernelCode(definition_, tensors_count_, linked_operations_); + std::string code = GetConcatKernelCode(definition_, attr_, &args_); + std::string element_wise_code; + RETURN_IF_ERROR( + MergeOperations(linked_operations_, &args_, &element_wise_code)); + RETURN_IF_ERROR(args_.TransformToCLCode(creation_context.device->GetInfo(), + {{"dst_tensor", element_wise_code}}, + &code)); return creation_context.cache->GetOrCreateCLKernel( code, "main_function", *creation_context.context, *creation_context.device, &kernel_); } absl::Status ConcatXY::BindArguments() { - kernel_.ResetBindingCounter(); - for (int i = 0; i < tensors_count_; ++i) { - RETURN_IF_ERROR(kernel_.SetMemoryAuto(src_[i]->GetMemoryPtr())); - } - RETURN_IF_ERROR(kernel_.SetMemoryAuto(dst_[0]->GetMemoryPtrForWriting())); - RETURN_IF_ERROR(BindArgs(&kernel_, linked_operations_)); - int x_offset = 0; - int y_offset = 0; - for (int i = 0; i < tensors_count_; ++i) { - const int width = src_[i]->Width() * src_[i]->Batch(); - const int height = src_[i]->Height(); + for (int i = 0; i < definition_.src_tensors.size(); ++i) { RETURN_IF_ERROR( - kernel_.SetBytesAuto(int4(width, height, x_offset, y_offset))); - x_offset += attr_.axis == Axis::WIDTH ? width : 0; - y_offset += attr_.axis == Axis::HEIGHT ? height : 0; + args_.SetObjectRef("src_tensor_" + std::to_string(i), src_[i])); } - RETURN_IF_ERROR(kernel_.SetBytesAuto(dst_[0]->GetWBatchedHSB())); - return absl::OkStatus(); + RETURN_IF_ERROR(args_.SetObjectRef("dst_tensor", dst_[0])); + RETURN_IF_ERROR(SetArguments(linked_operations_, &args_)); + return args_.Bind(kernel_.kernel()); } int3 ConcatXY::GetGridSize() const { - int max_src_width = 0; - int max_src_height = 0; - for (int i = 0; i < tensors_count_; ++i) { - max_src_width = std::max(max_src_width, src_[i]->Width()); - max_src_height = std::max(max_src_height, src_[i]->Height()); - } - - const int grid_x = max_src_width * dst_[0]->Batch(); - const int grid_y = max_src_height; + const int grid_x = dst_[0]->Width() * dst_[0]->Batch(); + const int grid_y = dst_[0]->Height() * dst_[0]->Depth(); const int grid_z = dst_[0]->Slices(); - return int3(grid_x, grid_y, grid_z); } diff --git a/tensorflow/lite/delegates/gpu/cl/selectors/simple_selectors.cc b/tensorflow/lite/delegates/gpu/cl/selectors/simple_selectors.cc index 5fc04d12822..d9bd70cc6b3 100644 --- a/tensorflow/lite/delegates/gpu/cl/selectors/simple_selectors.cc +++ b/tensorflow/lite/delegates/gpu/cl/selectors/simple_selectors.cc @@ -105,8 +105,10 @@ absl::Status SelectConcat(const ConcatAttributes& attr, *ptr = absl::make_unique(std::move(operation)); return absl::OkStatus(); } - case Axis::WIDTH: - case Axis::HEIGHT: { + case Axis::BATCH: + case Axis::DEPTH: + case Axis::HEIGHT: + case Axis::WIDTH: { ConcatXY operation = CreateConcatXY(op_def, attr, channels.size()); *ptr = absl::make_unique(std::move(operation)); return absl::OkStatus(); From 70e2387ecc190d415cde234c28ce5f5ae6b90bea Mon Sep 17 00:00:00 2001 From: Raman Sarokin Date: Wed, 17 Jun 2020 17:29:11 -0700 Subject: [PATCH 139/867] Merged Reshape code generation for HWC/BHWC layouts. Added support of different layouts for src/dst. For example src - HWC, dst - BHWC, or vice versa. PiperOrigin-RevId: 316999310 Change-Id: I20bd9a12afba8bdcb832565f09350440349041bd --- .../lite/delegates/gpu/cl/kernels/reshape.cc | 79 ++++++------------- .../delegates/gpu/cl/kernels/reshapex4.cc | 69 ++++++---------- 2 files changed, 47 insertions(+), 101 deletions(-) diff --git a/tensorflow/lite/delegates/gpu/cl/kernels/reshape.cc b/tensorflow/lite/delegates/gpu/cl/kernels/reshape.cc index a99fff0a1da..5abfad60c1b 100644 --- a/tensorflow/lite/delegates/gpu/cl/kernels/reshape.cc +++ b/tensorflow/lite/delegates/gpu/cl/kernels/reshape.cc @@ -25,56 +25,6 @@ namespace gpu { namespace cl { namespace { -std::string GetReshapeBatchedCode(const OperationDef& op_def, Arguments* args) { - args->AddObjectRef( - "src_tensor", AccessType::READ, - absl::make_unique(op_def.src_tensors[0])); - args->AddObjectRef( - "dst_tensor", AccessType::WRITE, - absl::make_unique(op_def.dst_tensors[0])); - - std::string c = GetCommonDefines(op_def.precision); - c += "__kernel void main_function(\n"; - c += "$0) {\n"; - c += " int linear_id = get_global_id(0);\n"; - c += " int X = linear_id / args.dst_tensor.Batch();\n"; - c += " int B = linear_id % args.dst_tensor.Batch();\n"; - c += " int Y = get_global_id(1);\n"; - c += " int Z = get_global_id(2);\n"; - c += " if (X >= args.dst_tensor.Width() || Y >= args.dst_tensor.Height() || " - "Z >= args.dst_tensor.Slices()) { \n"; - c += " return; \n"; - c += " } \n"; - c += " FLT temps[4];\n"; - c += " temps[0] = (FLT)(0.0f);\n"; - c += " temps[1] = (FLT)(0.0f);\n"; - c += " temps[2] = (FLT)(0.0f);\n"; - c += " temps[3] = (FLT)(0.0f);\n"; - c += " int base = ((B * args.dst_tensor.Height() + Y) * " - "args.dst_tensor.Width() + X) * args.dst_tensor.Channels() + Z * 4;\n"; - c += " for (int i = 0; i < 4; ++i) {\n"; - c += " int dst_channel = Z * 4 + i;\n"; - c += " if (dst_channel < args.dst_tensor.Channels()) {;\n"; - c += " int p = base + i;\n"; - c += " int src_c = p % args.src_tensor.Channels();\n"; - c += " p = p / args.src_tensor.Channels();\n"; - c += " int src_x = p % args.src_tensor.Width();\n"; - c += " p = p / args.src_tensor.Width();\n"; - c += " int src_y = p % args.src_tensor.Height();\n"; - c += " int src_b = p / args.src_tensor.Height();\n"; - c += " int src_z = src_c / 4;\n"; - c += " int src_sub_ch = src_c % 4;\n"; - c += " FLT4 t = args.src_tensor.Read(src_x, src_y, src_z, src_b);\n"; - c += " FLT t_ar[4] = {t.x, t.y, t.z, t.w};\n"; - c += " temps[i] = t_ar[src_sub_ch];\n"; - c += " }\n"; - c += " }\n"; - c += " FLT4 result = (FLT4)(temps[0], temps[1], temps[2], temps[3]);\n"; - c += " args.dst_tensor.Write(result, X, Y, Z, B);\n"; - c += "}\n"; - return c; -} - std::string GetReshapeCode(const OperationDef& op_def, Arguments* args) { args->AddObjectRef( "src_tensor", AccessType::READ, @@ -86,7 +36,14 @@ std::string GetReshapeCode(const OperationDef& op_def, Arguments* args) { std::string c = GetCommonDefines(op_def.precision); c += "__kernel void main_function(\n"; c += "$0) {\n"; - c += " int X = get_global_id(0);\n"; + if (op_def.dst_tensors[0].HasAxis(Axis::BATCH)) { + c += " int linear_id = get_global_id(0);\n"; + c += " int X = linear_id / args.dst_tensor.Batch();\n"; + c += " int B = linear_id % args.dst_tensor.Batch();\n"; + c += " args.dst_tensor.SetBatchRef(B);\n"; + } else { + c += " int X = get_global_id(0);\n"; + } c += " int Y = get_global_id(1);\n"; c += " int Z = get_global_id(2);\n"; c += " if (X >= args.dst_tensor.Width() || Y >= args.dst_tensor.Height() || " @@ -98,8 +55,13 @@ std::string GetReshapeCode(const OperationDef& op_def, Arguments* args) { c += " temps[1] = (FLT)(0.0f);\n"; c += " temps[2] = (FLT)(0.0f);\n"; c += " temps[3] = (FLT)(0.0f);\n"; - c += " int base = (Y * args.dst_tensor.Width() + X) * " - "args.dst_tensor.Channels() + Z * 4;\n"; + if (op_def.dst_tensors[0].HasAxis(Axis::BATCH)) { + c += " int base = B;\n"; + } else { + c += " int base = 0;\n"; + } + c += " base = ((base * args.dst_tensor.Height() + Y) * " + "args.dst_tensor.Width() + X) * args.dst_tensor.Channels() + Z * 4;\n"; c += " for (int i = 0; i < 4; ++i) {\n"; c += " int dst_channel = Z * 4 + i;\n"; c += " if (dst_channel < args.dst_tensor.Channels()) {;\n"; @@ -107,7 +69,12 @@ std::string GetReshapeCode(const OperationDef& op_def, Arguments* args) { c += " int src_c = p % args.src_tensor.Channels();\n"; c += " p = p / args.src_tensor.Channels();\n"; c += " int src_x = p % args.src_tensor.Width();\n"; - c += " int src_y = p / args.src_tensor.Width();\n"; + c += " p = p / args.src_tensor.Width();\n"; + c += " int src_y = p % args.src_tensor.Height();\n"; + if (op_def.src_tensors[0].HasAxis(Axis::BATCH)) { + c += " int src_b = p / args.src_tensor.Height();\n"; + c += " args.src_tensor.SetBatchRef(src_b);\n"; + } c += " int src_z = src_c / 4;\n"; c += " int src_sub_ch = src_c % 4;\n"; c += " FLT4 t = args.src_tensor.Read(src_x, src_y, src_z);\n"; @@ -137,9 +104,7 @@ Reshape& Reshape::operator=(Reshape&& operation) { } absl::Status Reshape::Compile(const CreationContext& creation_context) { - std::string code = definition_.IsBatchSupported() - ? GetReshapeBatchedCode(definition_, &args_) - : GetReshapeCode(definition_, &args_); + std::string code = GetReshapeCode(definition_, &args_); std::string element_wise_code; RETURN_IF_ERROR( MergeOperations(linked_operations_, &args_, &element_wise_code)); diff --git a/tensorflow/lite/delegates/gpu/cl/kernels/reshapex4.cc b/tensorflow/lite/delegates/gpu/cl/kernels/reshapex4.cc index 0847fce5836..3edbe637aa2 100644 --- a/tensorflow/lite/delegates/gpu/cl/kernels/reshapex4.cc +++ b/tensorflow/lite/delegates/gpu/cl/kernels/reshapex4.cc @@ -25,40 +25,6 @@ namespace gpu { namespace cl { namespace { -std::string GetReshapeBatchedCode(const OperationDef& op_def, Arguments* args) { - args->AddObjectRef( - "src_tensor", AccessType::READ, - absl::make_unique(op_def.src_tensors[0])); - args->AddObjectRef( - "dst_tensor", AccessType::WRITE, - absl::make_unique(op_def.dst_tensors[0])); - - std::string c = GetCommonDefines(op_def.precision); - c += "__kernel void main_function(\n"; - c += "$0) {\n"; - c += " int linear_id = get_global_id(0);\n"; - c += " int X = linear_id / args.dst_tensor.Batch();\n"; - c += " int B = linear_id % args.dst_tensor.Batch();\n"; - c += " int Y = get_global_id(1);\n"; - c += " int Z = get_global_id(2);\n"; - c += " if (X >= args.dst_tensor.Width() || Y >= args.dst_tensor.Height() || " - "Z >= args.dst_tensor.Slices()) { \n"; - c += " return; \n"; - c += " } \n"; - c += " int dst_bhwc4 = ((B * args.dst_tensor.Height() + Y) * " - "args.dst_tensor.Width() + X) * args.dst_tensor.Slices() + Z;\n"; - c += " int src_z = dst_bhwc4 % args.src_tensor.Slices();\n"; - c += " dst_bhwc4 = dst_bhwc4 / args.src_tensor.Slices();\n"; - c += " int src_x = dst_bhwc4 % args.src_tensor.Width();\n"; - c += " dst_bhwc4 = dst_bhwc4 / args.src_tensor.Width();\n"; - c += " int src_y = dst_bhwc4 % args.src_tensor.Height();\n"; - c += " int src_b = dst_bhwc4 / args.src_tensor.Height();\n"; - c += " FLT4 result = args.src_tensor.Read(src_x, src_y, src_z, src_b);\n"; - c += " args.dst_tensor.Write(result, X, Y, Z, B);\n"; - c += "}\n"; - return c; -} - std::string GetReshapeCode(const OperationDef& op_def, Arguments* args) { args->AddObjectRef( "src_tensor", AccessType::READ, @@ -70,19 +36,36 @@ std::string GetReshapeCode(const OperationDef& op_def, Arguments* args) { std::string c = GetCommonDefines(op_def.precision); c += "__kernel void main_function(\n"; c += "$0) {\n"; - c += " int X = get_global_id(0);\n"; + if (op_def.dst_tensors[0].HasAxis(Axis::BATCH)) { + c += " int linear_id = get_global_id(0);\n"; + c += " int X = linear_id / args.dst_tensor.Batch();\n"; + c += " int B = linear_id % args.dst_tensor.Batch();\n"; + c += " args.dst_tensor.SetBatchRef(B);\n"; + } else { + c += " int X = get_global_id(0);\n"; + } c += " int Y = get_global_id(1);\n"; c += " int Z = get_global_id(2);\n"; c += " if (X >= args.dst_tensor.Width() || Y >= args.dst_tensor.Height() || " "Z >= args.dst_tensor.Slices()) { \n"; c += " return; \n"; c += " } \n"; - c += " int dst_hwc4 = (Y * args.dst_tensor.Width() + X) * " - "args.dst_tensor.Slices() + Z;\n"; - c += " int src_z = dst_hwc4 % args.src_tensor.Slices();\n"; - c += " dst_hwc4 = dst_hwc4 / args.src_tensor.Slices();\n"; - c += " int src_x = dst_hwc4 % args.src_tensor.Width();\n"; - c += " int src_y = dst_hwc4 / args.src_tensor.Width();\n"; + if (op_def.dst_tensors[0].HasAxis(Axis::BATCH)) { + c += " int dst_bhwc4 = B;\n"; + } else { + c += " int dst_bhwc4 = 0;\n"; + } + c += " dst_bhwc4 = ((dst_bhwc4 * args.dst_tensor.Height() + Y) * " + "args.dst_tensor.Width() + X) * args.dst_tensor.Slices() + Z;\n"; + c += " int src_z = dst_bhwc4 % args.src_tensor.Slices();\n"; + c += " dst_bhwc4 = dst_bhwc4 / args.src_tensor.Slices();\n"; + c += " int src_x = dst_bhwc4 % args.src_tensor.Width();\n"; + c += " dst_bhwc4 = dst_bhwc4 / args.src_tensor.Width();\n"; + c += " int src_y = dst_bhwc4 % args.src_tensor.Height();\n"; + if (op_def.src_tensors[0].HasAxis(Axis::BATCH)) { + c += " int src_b = dst_bhwc4 / args.src_tensor.Height();\n"; + c += " args.src_tensor.SetBatchRef(src_b);\n"; + } c += " FLT4 result = args.src_tensor.Read(src_x, src_y, src_z);\n"; c += " args.dst_tensor.Write(result, X, Y, Z);\n"; c += "}\n"; @@ -105,9 +88,7 @@ Reshapex4& Reshapex4::operator=(Reshapex4&& operation) { } absl::Status Reshapex4::Compile(const CreationContext& creation_context) { - std::string code = definition_.IsBatchSupported() - ? GetReshapeBatchedCode(definition_, &args_) - : GetReshapeCode(definition_, &args_); + std::string code = GetReshapeCode(definition_, &args_); std::string element_wise_code; RETURN_IF_ERROR( MergeOperations(linked_operations_, &args_, &element_wise_code)); From adf0573f9d874b568c9680a6d3cf310e555dcd62 Mon Sep 17 00:00:00 2001 From: "A. Unique TensorFlower" Date: Wed, 17 Jun 2020 17:37:35 -0700 Subject: [PATCH 140/867] Qualify uses of std::string PiperOrigin-RevId: 317000789 Change-Id: I6f847b235496d1dc8f8b3380e21ce890566a9a88 --- .../tests/fuse_binary_into_following_affine_test.cc | 5 +++-- .../tests/fuse_binary_into_preceding_affine_test.cc | 5 +++-- .../toco/graph_transformations/tests/lstm_utils_test.cc | 4 ++-- .../tests/resolve_constant_concatenation_test.cc | 8 ++++---- .../graph_transformations/tests/unpack_quantize_test.cc | 9 +++++---- 5 files changed, 17 insertions(+), 14 deletions(-) diff --git a/tensorflow/lite/toco/graph_transformations/tests/fuse_binary_into_following_affine_test.cc b/tensorflow/lite/toco/graph_transformations/tests/fuse_binary_into_following_affine_test.cc index 2cba6824cfb..d6cf31f4211 100644 --- a/tensorflow/lite/toco/graph_transformations/tests/fuse_binary_into_following_affine_test.cc +++ b/tensorflow/lite/toco/graph_transformations/tests/fuse_binary_into_following_affine_test.cc @@ -43,14 +43,15 @@ class FuseBinaryIntoFollowingAffineTest : public ::testing::Test { void SetUp() override { model_.reset(new Model); } - void CreateArray(const string& name, const std::vector& shape) { + void CreateArray(const std::string& name, const std::vector& shape) { Array& array = model_->GetOrCreateArray(name); array.data_type = ArrayDataType::kFloat; Shape* array_shape = array.mutable_shape(); *(array_shape->mutable_dims()) = shape; } - void CreateConstantArray(const string& name, const std::vector& shape, + void CreateConstantArray(const std::string& name, + const std::vector& shape, const std::vector& data) { CreateArray(name, shape); Array& array = model_->GetOrCreateArray(name); diff --git a/tensorflow/lite/toco/graph_transformations/tests/fuse_binary_into_preceding_affine_test.cc b/tensorflow/lite/toco/graph_transformations/tests/fuse_binary_into_preceding_affine_test.cc index b5c321c1a26..6c3dc7dc761 100644 --- a/tensorflow/lite/toco/graph_transformations/tests/fuse_binary_into_preceding_affine_test.cc +++ b/tensorflow/lite/toco/graph_transformations/tests/fuse_binary_into_preceding_affine_test.cc @@ -43,14 +43,15 @@ class FuseBinaryIntoPrecedingAffineTest : public ::testing::Test { void SetUp() override { model_.reset(new Model); } - void CreateArray(const string& name, const std::vector& shape) { + void CreateArray(const std::string& name, const std::vector& shape) { Array& array = model_->GetOrCreateArray(name); array.data_type = ArrayDataType::kFloat; Shape* array_shape = array.mutable_shape(); *(array_shape->mutable_dims()) = shape; } - void CreateConstantArray(const string& name, const std::vector& shape, + void CreateConstantArray(const std::string& name, + const std::vector& shape, const std::vector& data) { CreateArray(name, shape); Array& array = model_->GetOrCreateArray(name); diff --git a/tensorflow/lite/toco/graph_transformations/tests/lstm_utils_test.cc b/tensorflow/lite/toco/graph_transformations/tests/lstm_utils_test.cc index bdb27e8af2e..204e197e186 100644 --- a/tensorflow/lite/toco/graph_transformations/tests/lstm_utils_test.cc +++ b/tensorflow/lite/toco/graph_transformations/tests/lstm_utils_test.cc @@ -46,12 +46,12 @@ class CopyArrayDataTest : public ::testing::Test { int src_dim_1, int src_dim_2, std::initializer_list dst_data, int dst_dim_1, int dst_dim_2) { - string src_array = "src_array"; + std::string src_array = "src_array"; src_buffer_ = CreateFloatArrayBuffer( model, &src_array, src_dim_2 == 1 ? Shape({src_dim_1}) : Shape({src_dim_1, src_dim_2})); PopulateBuffer(src_buffer_, src_data); - string dst_array = "dst_array"; + std::string dst_array = "dst_array"; dst_buffer_ = CreateFloatArrayBuffer( model, &dst_array, dst_dim_2 == 1 ? Shape({dst_dim_1}) : Shape({dst_dim_1, dst_dim_2})); diff --git a/tensorflow/lite/toco/graph_transformations/tests/resolve_constant_concatenation_test.cc b/tensorflow/lite/toco/graph_transformations/tests/resolve_constant_concatenation_test.cc index bfed38ce7aa..5b0566fe074 100644 --- a/tensorflow/lite/toco/graph_transformations/tests/resolve_constant_concatenation_test.cc +++ b/tensorflow/lite/toco/graph_transformations/tests/resolve_constant_concatenation_test.cc @@ -107,10 +107,10 @@ class ResolveConstantConcatenationTest : public ::testing::Test { // together with 4 arrays as its inputs. // It receives the dimension of concatenation as input. void PrepareModel(Model* model, int axis) { - const string output_name("concat_op_output"); + const std::string output_name("concat_op_output"); model->flags.add_output_arrays(output_name); - std::vector concat_input_names = {"array0", "array1", "array2", - "array3"}; + std::vector concat_input_names = {"array0", "array1", "array2", + "array3"}; const int kDim = 3; const int kElementPerDim = 2; @@ -122,7 +122,7 @@ class ResolveConstantConcatenationTest : public ::testing::Test { {20., 21., 22., 23., 24., 25., 26., 27.}, {30., 31., 32., 33., 34., 35., 36., 37.}}; int cnt = 0; - for (const string& concat_input_name : concat_input_names) { + for (const std::string& concat_input_name : concat_input_names) { Array& in_array = model->GetOrCreateArray(concat_input_name); in_array.data_type = ArrayDataType::kFloat; diff --git a/tensorflow/lite/toco/graph_transformations/tests/unpack_quantize_test.cc b/tensorflow/lite/toco/graph_transformations/tests/unpack_quantize_test.cc index 2dc3fb35b0f..3cc4e725463 100755 --- a/tensorflow/lite/toco/graph_transformations/tests/unpack_quantize_test.cc +++ b/tensorflow/lite/toco/graph_transformations/tests/unpack_quantize_test.cc @@ -40,10 +40,11 @@ class UnpackQuantizeTest : public ::testing::Test { // 1. calculate min and max of the input. // 2. insert dequantization nodes after quantized outputs of Unpack operation. void PrepareModel(Model* model, int axis) { - std::vector unpack_output_names = {"unpack_out0", "unpack_out1"}; + std::vector unpack_output_names = {"unpack_out0", + "unpack_out1"}; model->flags.add_output_arrays(unpack_output_names[0]); model->flags.add_output_arrays(unpack_output_names[1]); - const string unpack_input_name("unpack_op_input"); + const std::string unpack_input_name("unpack_op_input"); const int kDim = 2; const int kElementPerDim = 2; @@ -75,7 +76,7 @@ class UnpackQuantizeTest : public ::testing::Test { // Configuring the necessary outputs. The outputs also happen to be in // kFloat. This is because during quantization transformation data types for // these arrays are going to be forced to be kUint8. - for (const string& unpack_output_name : unpack_output_names) { + for (const std::string& unpack_output_name : unpack_output_names) { Array& out_array = model->GetOrCreateArray(unpack_output_name); out_array.GetOrCreateMinMax(); out_array.data_type = ArrayDataType::kFloat; @@ -109,7 +110,7 @@ TEST_F(UnpackQuantizeTest, CheckUnpackPreservesQuantizationParameters) { ->Run(&model, /*op_index=*/0, &modified) .ok()); - const string output_name = model.flags.output_arrays(0); + const std::string output_name = model.flags.output_arrays(0); // Quantization transformation inserts NODE_NAME_DEQUANTIZE operations, // effectively making them the new outputs of the array. Old outputs of the From 56db128697fd0e9d3f689d472d53e33bf4ea1bb8 Mon Sep 17 00:00:00 2001 From: "A. Unique TensorFlower" Date: Wed, 17 Jun 2020 17:56:36 -0700 Subject: [PATCH 141/867] Qualify uses of std::string PiperOrigin-RevId: 317003622 Change-Id: Iae6a9a287ffd3b97dee8b9993c443db322936592 --- .../lite/toco/logging/conversion_log_util.cc | 67 ++++++++++--------- .../lite/toco/logging/conversion_log_util.h | 22 +++--- .../toco/logging/conversion_log_util_test.cc | 60 ++++++++--------- 3 files changed, 77 insertions(+), 72 deletions(-) diff --git a/tensorflow/lite/toco/logging/conversion_log_util.cc b/tensorflow/lite/toco/logging/conversion_log_util.cc index c23c305c750..55afa1370b3 100644 --- a/tensorflow/lite/toco/logging/conversion_log_util.cc +++ b/tensorflow/lite/toco/logging/conversion_log_util.cc @@ -34,8 +34,8 @@ namespace toco { namespace { -string TryGetOperatorName(const Operator& op) { - string op_name; +std::string TryGetOperatorName(const Operator& op) { + std::string op_name; if (!op.tensorflow_node_def.empty()) { // Parse op name from serialized NodeDef. tensorflow::NodeDef node_def; @@ -63,8 +63,8 @@ string TryGetOperatorName(const Operator& op) { return op_name; } -string GetOSVersion() { - string os_info; +std::string GetOSVersion() { + std::string os_info; #ifdef __linux__ utsname info; if (uname(&info)) { @@ -72,12 +72,13 @@ string GetOSVersion() { LOG(ERROR) << "Cannot get OS info."; return ""; } - os_info = string(info.sysname) + ";OSVer=" + string(info.release) + ";"; + os_info = + std::string(info.sysname) + ";OSVer=" + std::string(info.release) + ";"; #endif return os_info; } -string ShapeToStringNoSpace(const Shape& shape) { +std::string ShapeToStringNoSpace(const Shape& shape) { if (shape.dimensions_count() == 0) { return "[]"; } @@ -85,13 +86,13 @@ string ShapeToStringNoSpace(const Shape& shape) { return absl::StrCat("[", absl::StrJoin(shape.dims(), ","), "]"); } -string GetOperatorSignature( +std::string GetOperatorSignature( const Model& model, const Operator& op, const std::map>& op_types_map) { // The signature of an op has the following schema: // INPUT:SHAPE::TYPE::OUTPUT:SHAPE::TYPE::NAME:VERSION: - string op_signature; + std::string op_signature; constexpr char delimiter[] = "::"; // Get input shapes and types. @@ -137,8 +138,8 @@ string GetOperatorSignature( } // namespace -std::vector GetOperatorNames(const Model& model) { - std::vector op_names; +std::vector GetOperatorNames(const Model& model) { + std::vector op_names; for (const auto& op : model.operators) { op_names.push_back(TryGetOperatorName(*op)); } @@ -146,9 +147,9 @@ std::vector GetOperatorNames(const Model& model) { } void CountOperatorsByType(const Model& model, - std::map* built_in_ops, - std::map* custom_ops, - std::map* select_ops) { + std::map* built_in_ops, + std::map* custom_ops, + std::map* select_ops) { for (const auto& op : model.operators) { OperatorSignature op_signature = {op.get(), &model}; const auto ops_by_type = @@ -156,7 +157,7 @@ void CountOperatorsByType(const Model& model, tflite::details::OperatorKey op_key(op_signature, ops_by_type, true /*enable_select_tf_ops*/); - const string op_name = TryGetOperatorName(*op); + const std::string op_name = TryGetOperatorName(*op); if (op_key.is_custom_op()) { (*custom_ops)[op_name]++; } else if (op_key.is_flex_op()) { @@ -168,8 +169,9 @@ void CountOperatorsByType(const Model& model, } void GetInputAndOutputTypes( - const Model& model, TFLITE_PROTO_NS::RepeatedPtrField* input_types, - TFLITE_PROTO_NS::RepeatedPtrField* output_types) { + const Model& model, + TFLITE_PROTO_NS::RepeatedPtrField* input_types, + TFLITE_PROTO_NS::RepeatedPtrField* output_types) { for (const auto& input_array : model.flags.input_arrays()) { const Array& array = model.GetArray(input_array.name()); input_types->Add(ArrayDataTypeName(array.data_type)); @@ -180,15 +182,16 @@ void GetInputAndOutputTypes( } } -string GetTfLiteVersion() { return TFLITE_VERSION_STRING; } +std::string GetTfLiteVersion() { return TFLITE_VERSION_STRING; } -string GetCachedOSVersion() { - static string* version = new string(GetOSVersion()); +std::string GetCachedOSVersion() { + static std::string* version = new std::string(GetOSVersion()); return *version; } -void GetOpSignatures(const Model& model, - TFLITE_PROTO_NS::RepeatedPtrField* op_signatures) { +void GetOpSignatures( + const Model& model, + TFLITE_PROTO_NS::RepeatedPtrField* op_signatures) { const auto& op_types_map = tflite::BuildOperatorByTypeMap(true /*enable_select_tf_ops*/); for (const auto& op : model.operators) { @@ -196,7 +199,7 @@ void GetOpSignatures(const Model& model, } } -string GetModelHash(const Model& model) { +std::string GetModelHash(const Model& model) { // TODO(b/123519920): Implement the hash function for Model. // Need to consider different implementations for public/private models. return ""; @@ -204,18 +207,18 @@ string GetModelHash(const Model& model) { // This function scans through the error message string, extracts the part about // missing ops and prunes away all other information in the error info. -string SanitizeErrorMessage(const string& error_message) { - const string s1 = "Ops that can be supported by the flex runtime"; - const string s2 = "Ops that need custom implementation"; - string pruned_message; +std::string SanitizeErrorMessage(const std::string& error_message) { + const std::string s1 = "Ops that can be supported by the flex runtime"; + const std::string s2 = "Ops that need custom implementation"; + std::string pruned_message; size_t pos = error_message.find(s1); - if (pos != string::npos) { + if (pos != std::string::npos) { // Find the terminate point for flex op list. auto end = error_message.find(".", pos); pruned_message.append(error_message.substr(pos, end - pos + 1)); } pos = error_message.find(s2); - if (pos != string::npos) { + if (pos != std::string::npos) { // Find the terminate point for custom op list. auto end = error_message.find(".", pos); pruned_message.append(error_message.substr(pos, end - pos + 1)); @@ -225,18 +228,18 @@ string SanitizeErrorMessage(const string& error_message) { void PopulateConversionLog(const Model& model, TocoConversionLog* log) { // Get the list of ops after conversion. - const std::vector op_names = GetOperatorNames(model); + const std::vector op_names = GetOperatorNames(model); for (const auto& op_name : op_names) { log->add_op_list(op_name); } // Get op signatures. - TFLITE_PROTO_NS::RepeatedPtrField op_signatures; + TFLITE_PROTO_NS::RepeatedPtrField op_signatures; GetOpSignatures(model, &op_signatures); log->mutable_op_signatures()->CopyFrom(op_signatures); // Get op counts by category: custom, built-in or select. - std::map custom_ops, select_ops, built_in_ops; + std::map custom_ops, select_ops, built_in_ops; CountOperatorsByType(model, &built_in_ops, &custom_ops, &select_ops); log->mutable_custom_ops()->insert(custom_ops.cbegin(), custom_ops.cend()); log->mutable_built_in_ops()->insert(built_in_ops.cbegin(), @@ -244,7 +247,7 @@ void PopulateConversionLog(const Model& model, TocoConversionLog* log) { log->mutable_select_ops()->insert(select_ops.cbegin(), select_ops.cend()); // Get the model's input and output types. - TFLITE_PROTO_NS::RepeatedPtrField input_types, output_types; + TFLITE_PROTO_NS::RepeatedPtrField input_types, output_types; GetInputAndOutputTypes(model, &input_types, &output_types); log->mutable_input_tensor_types()->CopyFrom(input_types); log->mutable_output_tensor_types()->CopyFrom(output_types); diff --git a/tensorflow/lite/toco/logging/conversion_log_util.h b/tensorflow/lite/toco/logging/conversion_log_util.h index 2237615adbb..c21ec0792cc 100644 --- a/tensorflow/lite/toco/logging/conversion_log_util.h +++ b/tensorflow/lite/toco/logging/conversion_log_util.h @@ -25,37 +25,39 @@ namespace toco { // This function scans through the error message string, extracts the part about // missing ops and prunes away all other information in the error info. -string SanitizeErrorMessage(const string& error_message); +std::string SanitizeErrorMessage(const std::string& error_message); // Populates the TocoConversionLog proto after analyzing the model. void PopulateConversionLog(const Model& model, TocoConversionLog* log); // Returns the names of the operators in the model. -std::vector GetOperatorNames(const Model& model); +std::vector GetOperatorNames(const Model& model); // Counts the number of different types of operators in the model: // Built-in ops, custom ops and select ops. // Each map is mapping from the name of the operator (such as 'Conv') to its // total number of occurrences in the model. void CountOperatorsByType(const Model& model, - std::map* built_in_ops, - std::map* custom_ops, - std::map* select_ops); + std::map* built_in_ops, + std::map* custom_ops, + std::map* select_ops); // Gets the input and output types of the model. The input and output is // specified by model.flags.input_arrays and model.flags.output_arrays. void GetInputAndOutputTypes( - const Model& model, TFLITE_PROTO_NS::RepeatedPtrField* input_types, - TFLITE_PROTO_NS::RepeatedPtrField* output_types); + const Model& model, + TFLITE_PROTO_NS::RepeatedPtrField* input_types, + TFLITE_PROTO_NS::RepeatedPtrField* output_types); // Calculates signatures for all the ops in the model. An op signature is // defined by its input/output shapes and types, op name and its version. -void GetOpSignatures(const Model& model, - TFLITE_PROTO_NS::RepeatedPtrField* op_signatures); +void GetOpSignatures( + const Model& model, + TFLITE_PROTO_NS::RepeatedPtrField* op_signatures); // TODO(b/123519920): Implement this. // Calculates a unique hash for the model. -string GetModelHash(const Model& model); +std::string GetModelHash(const Model& model); } // namespace toco diff --git a/tensorflow/lite/toco/logging/conversion_log_util_test.cc b/tensorflow/lite/toco/logging/conversion_log_util_test.cc index c4960715f25..17111eca6d0 100644 --- a/tensorflow/lite/toco/logging/conversion_log_util_test.cc +++ b/tensorflow/lite/toco/logging/conversion_log_util_test.cc @@ -58,9 +58,9 @@ TEST(ConversionLogUtilTest, TestCountOperatorsByType) { Model model; // 1st Conv operator. std::unique_ptr conv1(new ConvOperator()); - const string conv1_input_name = "conv_input1"; - const string conv1_filter_name = "conv_filter1"; - const string conv1_output_name = "conv_output1"; + const std::string conv1_input_name = "conv_input1"; + const std::string conv1_filter_name = "conv_filter1"; + const std::string conv1_output_name = "conv_output1"; conv1->inputs.push_back(conv1_input_name); conv1->inputs.push_back(conv1_filter_name); conv1->outputs.push_back(conv1_output_name); @@ -71,9 +71,9 @@ TEST(ConversionLogUtilTest, TestCountOperatorsByType) { // 2nd Conv operator. std::unique_ptr conv2(new ConvOperator()); - const string conv2_input_name = "conv_input2"; - const string conv2_filter_name = "conv_filter2"; - const string conv2_output_name = "conv_output2"; + const std::string conv2_input_name = "conv_input2"; + const std::string conv2_filter_name = "conv_filter2"; + const std::string conv2_output_name = "conv_output2"; conv2->inputs.push_back(conv2_input_name); conv2->inputs.push_back(conv2_filter_name); conv2->outputs.push_back(conv2_output_name); @@ -83,7 +83,7 @@ TEST(ConversionLogUtilTest, TestCountOperatorsByType) { // Mean operator. std::unique_ptr mean(new MeanOperator()); - const string mean_input_name = "mean_input"; + const std::string mean_input_name = "mean_input"; mean->inputs.push_back(mean_input_name); array_map[mean_input_name] = std::unique_ptr(new Array); @@ -111,26 +111,26 @@ TEST(ConversionLogUtilTest, TestCountOperatorsByType) { model.operators.push_back(std::move(elu_grad)); model.operators.push_back(std::move(my_custom_op)); - std::map built_in_ops, select_ops, custom_ops; + std::map built_in_ops, select_ops, custom_ops; CountOperatorsByType(model, &built_in_ops, &custom_ops, &select_ops); EXPECT_THAT(built_in_ops, - UnorderedElementsAre(std::pair("Conv", 2), - std::pair("Mean", 1))); + UnorderedElementsAre(std::pair("Conv", 2), + std::pair("Mean", 1))); EXPECT_THAT(select_ops, - UnorderedElementsAre(std::pair("AvgPool3D", 1), - std::pair("EluGrad", 1))); - EXPECT_THAT(custom_ops, UnorderedElementsAre( - std::pair("MyAwesomeCustomOp", 1))); + UnorderedElementsAre(std::pair("AvgPool3D", 1), + std::pair("EluGrad", 1))); + EXPECT_THAT(custom_ops, UnorderedElementsAre(std::pair( + "MyAwesomeCustomOp", 1))); } TEST(ConversionLogUtilTest, TestGetInputAndOutputTypes) { Model model; auto& array_map = model.GetMutableArrayMap(); - const string input1 = "conv_input"; - const string input2 = "conv_filter"; - const string input3 = "feature"; - const string output = "softmax"; + const std::string input1 = "conv_input"; + const std::string input2 = "conv_filter"; + const std::string input3 = "feature"; + const std::string output = "softmax"; array_map[input1] = std::unique_ptr(new Array); array_map[input1]->data_type = ArrayDataType::kFloat; array_map[input2] = std::unique_ptr(new Array); @@ -149,7 +149,7 @@ TEST(ConversionLogUtilTest, TestGetInputAndOutputTypes) { *model.flags.add_input_arrays() = input_arrays[2]; model.flags.add_output_arrays(output); - TFLITE_PROTO_NS::RepeatedPtrField input_types, output_types; + TFLITE_PROTO_NS::RepeatedPtrField input_types, output_types; GetInputAndOutputTypes(model, &input_types, &output_types); EXPECT_THAT(input_types, ElementsAre("float", "float", "int16")); @@ -161,9 +161,9 @@ TEST(ConversionLogUtilTest, TestGetOpSignatures) { auto& array_map = model.GetMutableArrayMap(); std::unique_ptr conv(new ConvOperator()); - const string conv_input_name = "conv_input"; - const string conv_filter_name = "conv_filter"; - const string conv_output_name = "conv_output"; + const std::string conv_input_name = "conv_input"; + const std::string conv_filter_name = "conv_filter"; + const std::string conv_output_name = "conv_output"; conv->inputs.push_back(conv_input_name); conv->inputs.push_back(conv_filter_name); conv->outputs.push_back(conv_output_name); @@ -177,15 +177,15 @@ TEST(ConversionLogUtilTest, TestGetOpSignatures) { array_map[conv_output_name]->data_type = ArrayDataType::kFloat; array_map[conv_output_name]->copy_shape({4, 4, 2}); - const string mean_input_name = "mean_input"; - const string mean_output_name = "mean_output"; + const std::string mean_input_name = "mean_input"; + const std::string mean_output_name = "mean_output"; std::unique_ptr mean(new MeanOperator()); mean->inputs.push_back(mean_input_name); mean->outputs.push_back(mean_output_name); array_map[mean_input_name] = std::unique_ptr(new Array); array_map[mean_output_name] = std::unique_ptr(new Array); - const string avg_pool_3d_output_name = "avg_pool_output"; + const std::string avg_pool_3d_output_name = "avg_pool_output"; auto avg_pool_3d = absl::make_unique(); avg_pool_3d->tensorflow_op = "AvgPool3D"; tensorflow::NodeDef node_def; @@ -197,7 +197,7 @@ TEST(ConversionLogUtilTest, TestGetOpSignatures) { array_map[avg_pool_3d_output_name]->data_type = ArrayDataType::kInt32; array_map[avg_pool_3d_output_name]->copy_shape({2, 2}); - const string custom_op_output_name = "custom_op_output"; + const std::string custom_op_output_name = "custom_op_output"; auto my_custom_op = absl::make_unique(); my_custom_op->tensorflow_op = "MyAwesomeCustomOp"; my_custom_op->inputs.push_back(avg_pool_3d_output_name); @@ -211,7 +211,7 @@ TEST(ConversionLogUtilTest, TestGetOpSignatures) { model.operators.push_back(std::move(avg_pool_3d)); model.operators.push_back(std::move(my_custom_op)); - TFLITE_PROTO_NS::RepeatedPtrField op_signatures; + TFLITE_PROTO_NS::RepeatedPtrField op_signatures; GetOpSignatures(model, &op_signatures); EXPECT_THAT(op_signatures, UnorderedElementsAre( @@ -225,14 +225,14 @@ TEST(ConversionLogUtilTest, TestGetOpSignatures) { } TEST(ConversionLogUtilTest, TestSanitizeErrorMessage) { - const string error = + const std::string error = "error: failed while converting: 'main': Ops that can be supported by " "the flex runtime (enabled via setting the -emit-select-tf-ops flag): " "ResizeNearestNeighbor,ResizeNearestNeighbor. Ops that need custom " "implementation (enabled via setting the -emit-custom-ops flag): " "CombinedNonMaxSuppression.\nTraceback (most recent call last): File " "/usr/local/bin/toco_from_protos, line 8, in "; - const string pruned_error = + const std::string pruned_error = "Ops that can be supported by " "the flex runtime (enabled via setting the -emit-select-tf-ops flag): " "ResizeNearestNeighbor,ResizeNearestNeighbor.Ops that need custom " @@ -242,7 +242,7 @@ TEST(ConversionLogUtilTest, TestSanitizeErrorMessage) { } TEST(ConversionLogUtilTest, TestSanitizeErrorMessageNoMatching) { - const string error = + const std::string error = "error: failed while converting: 'main': Traceback (most recent call " "last): File " "/usr/local/bin/toco_from_protos, line 8, in "; From c31f2ca4a29d4469f29c57735d74cbd2748c0e03 Mon Sep 17 00:00:00 2001 From: Sam Holt Date: Tue, 16 Jun 2020 16:50:16 +0100 Subject: [PATCH 142/867] fix: convolutional padding argument valid and same explaination --- .../python/keras/layers/convolutional.py | 55 +++++++++++++++---- .../keras/layers/convolutional_recurrent.py | 6 ++ tensorflow/python/keras/layers/local.py | 2 + tensorflow/python/keras/layers/pooling.py | 22 ++++++-- .../keras/legacy_tf_layers/convolutional.py | 42 ++++++++++++++ tensorflow/python/keras/utils/conv_utils.py | 12 ++++ tensorflow/python/ops/nn_ops.py | 3 + 7 files changed, 128 insertions(+), 14 deletions(-) diff --git a/tensorflow/python/keras/layers/convolutional.py b/tensorflow/python/keras/layers/convolutional.py index 51f4e3b320a..471d94570a5 100644 --- a/tensorflow/python/keras/layers/convolutional.py +++ b/tensorflow/python/keras/layers/convolutional.py @@ -72,6 +72,10 @@ class Conv(Layer): Specifying any stride value != 1 is incompatible with specifying any `dilation_rate` value != 1. padding: One of `"valid"`, `"same"`, or `"causal"` (case-insensitive). + `"valid"` means no padding. `"same"` results in padding evenly to + the left/right or up/down of the input such that output has the same + height/width dimension as the input. `"causal"` results in causal + (dilated) convolutions, e.g. `output[t]` does not depend on `input[t+1:]`. data_format: A string, one of `channels_last` (default) or `channels_first`. The ordering of the dimensions in the inputs. `channels_last` corresponds to inputs with shape @@ -418,6 +422,9 @@ class Conv1D(Conv): Specifying any stride value != 1 is incompatible with specifying any `dilation_rate` value != 1. padding: One of `"valid"`, `"causal"` or `"same"` (case-insensitive). + `"valid"` means no padding. `"same"` results in padding evenly to + the left/right or up/down of the input such that output has the same + height/width dimension as the input. `"causal"` results in causal (dilated) convolutions, e.g. `output[t]` does not depend on `input[t+1:]`. Useful when modeling temporal data where the model should not violate the temporal order. @@ -571,6 +578,9 @@ class Conv2D(Conv): specify the same value for all spatial dimensions. Specifying any stride value != 1 is incompatible with specifying any `dilation_rate` value != 1. padding: one of `"valid"` or `"same"` (case-insensitive). + `"valid"` means no padding. `"same"` results in padding evenly to + the left/right or up/down of the input such that output has the same + height/width dimension as the input. data_format: A string, one of `channels_last` (default) or `channels_first`. The ordering of the dimensions in the inputs. `channels_last` corresponds to inputs with shape `(batch_size, height, width, channels)` while @@ -712,6 +722,9 @@ class Conv3D(Conv): specify the same value for all spatial dimensions. Specifying any stride value != 1 is incompatible with specifying any `dilation_rate` value != 1. padding: one of `"valid"` or `"same"` (case-insensitive). + `"valid"` means no padding. `"same"` results in padding evenly to + the left/right or up/down of the input such that output has the same + height/width dimension as the input. data_format: A string, one of `channels_last` (default) or `channels_first`. The ordering of the dimensions in the inputs. `channels_last` corresponds to inputs with shape `batch_shape + (spatial_dim1, spatial_dim2, @@ -833,6 +846,9 @@ class Conv1DTranspose(Conv1D): time dimension. Specifying a stride value != 1 is incompatible with specifying a `dilation_rate` value != 1. Defaults to 1. padding: one of `"valid"` or `"same"` (case-insensitive). + `"valid"` means no padding. `"same"` results in padding evenly to + the left/right or up/down of the input such that output has the same + height/width dimension as the input. output_padding: An integer specifying the amount of padding along the time dimension of the output tensor. The amount of output padding must be lower than the stride. @@ -1083,6 +1099,9 @@ class Conv2DTranspose(Conv2D): Specifying any stride value != 1 is incompatible with specifying any `dilation_rate` value != 1. padding: one of `"valid"` or `"same"` (case-insensitive). + `"valid"` means no padding. `"same"` results in padding evenly to + the left/right or up/down of the input such that output has the same + height/width dimension as the input. output_padding: An integer or tuple/list of 2 integers, specifying the amount of padding along the height and width of the output tensor. @@ -1371,19 +1390,22 @@ class Conv3DTranspose(Conv3D): Arguments: filters: Integer, the dimensionality of the output space - (i.e. the number of output filters in the convolution). + (i.e. the number of output filters in the convolution). kernel_size: An integer or tuple/list of 3 integers, specifying the - depth, height and width of the 3D convolution window. - Can be a single integer to specify the same value for - all spatial dimensions. + depth, height and width of the 3D convolution window. + Can be a single integer to specify the same value for + all spatial dimensions. strides: An integer or tuple/list of 3 integers, - specifying the strides of the convolution along the depth, height - and width. - Can be a single integer to specify the same value for - all spatial dimensions. - Specifying any stride value != 1 is incompatible with specifying - any `dilation_rate` value != 1. + specifying the strides of the convolution along the depth, height + and width. + Can be a single integer to specify the same value for + all spatial dimensions. + Specifying any stride value != 1 is incompatible with specifying + any `dilation_rate` value != 1. padding: one of `"valid"` or `"same"` (case-insensitive). + `"valid"` means no padding. `"same"` results in padding evenly to + the left/right or up/down of the input such that output has the same + height/width dimension as the input. output_padding: An integer or tuple/list of 3 integers, specifying the amount of padding along the depth, height, and width. @@ -1681,6 +1703,9 @@ class SeparableConv(Conv): Specifying any `stride` value != 1 is incompatible with specifying any `dilation_rate` value != 1. padding: One of `"valid"` or `"same"` (case-insensitive). + `"valid"` means no padding. `"same"` results in padding evenly to + the left/right or up/down of the input such that output has the same + height/width dimension as the input. data_format: A string, one of `channels_last` (default) or `channels_first`. The ordering of the dimensions in the inputs. `channels_last` corresponds to inputs with shape @@ -1885,6 +1910,10 @@ class SeparableConv1D(SeparableConv): Specifying any `stride` value != 1 is incompatible with specifying any `dilation_rate` value != 1. padding: One of `"valid"`, `"same"`, or `"causal"` (case-insensitive). + `"valid"` means no padding. `"same"` results in padding evenly to + the left/right or up/down of the input such that output has the same + height/width dimension as the input. `"causal"` results in causal + (dilated) convolutions, e.g. `output[t]` does not depend on `input[t+1:]`. data_format: A string, one of `channels_last` (default) or `channels_first`. The ordering of the dimensions in the inputs. `channels_last` corresponds to inputs with shape @@ -2070,6 +2099,9 @@ class SeparableConv2D(SeparableConv): Specifying any stride value != 1 is incompatible with specifying any `dilation_rate` value != 1. padding: one of `"valid"` or `"same"` (case-insensitive). + `"valid"` means no padding. `"same"` results in padding evenly to + the left/right or up/down of the input such that output has the same + height/width dimension as the input. data_format: A string, one of `channels_last` (default) or `channels_first`. The ordering of the dimensions in the inputs. @@ -2230,6 +2262,9 @@ class DepthwiseConv2D(Conv2D): Specifying any stride value != 1 is incompatible with specifying any `dilation_rate` value != 1. padding: one of `'valid'` or `'same'` (case-insensitive). + `"valid"` means no padding. `"same"` results in padding evenly to + the left/right or up/down of the input such that output has the same + height/width dimension as the input. depth_multiplier: The number of depthwise convolution output channels for each input channel. The total number of depthwise convolution output diff --git a/tensorflow/python/keras/layers/convolutional_recurrent.py b/tensorflow/python/keras/layers/convolutional_recurrent.py index 19831429b73..54196f8725c 100644 --- a/tensorflow/python/keras/layers/convolutional_recurrent.py +++ b/tensorflow/python/keras/layers/convolutional_recurrent.py @@ -434,6 +434,9 @@ class ConvLSTM2DCell(DropoutRNNCellMixin, Layer): Specifying any stride value != 1 is incompatible with specifying any `dilation_rate` value != 1. padding: One of `"valid"` or `"same"` (case-insensitive). + `"valid"` means no padding. `"same"` results in padding evenly to + the left/right or up/down of the input such that output has the same + height/width dimension as the input. data_format: A string, one of `channels_last` (default) or `channels_first`. It defaults to the `image_data_format` value found in your @@ -710,6 +713,9 @@ class ConvLSTM2D(ConvRNN2D): Specifying any stride value != 1 is incompatible with specifying any `dilation_rate` value != 1. padding: One of `"valid"` or `"same"` (case-insensitive). + `"valid"` means no padding. `"same"` results in padding evenly to + the left/right or up/down of the input such that output has the same + height/width dimension as the input. data_format: A string, one of `channels_last` (default) or `channels_first`. The ordering of the dimensions in the inputs. diff --git a/tensorflow/python/keras/layers/local.py b/tensorflow/python/keras/layers/local.py index 3e9c0f9c0a3..c33c88f3a3d 100644 --- a/tensorflow/python/keras/layers/local.py +++ b/tensorflow/python/keras/layers/local.py @@ -67,6 +67,7 @@ class LocallyConnected1D(Layer): any `dilation_rate` value != 1. padding: Currently only supports `"valid"` (case-insensitive). `"same"` may be supported in the future. + `"valid"` means no padding. data_format: A string, one of `channels_last` (default) or `channels_first`. The ordering of the dimensions in the inputs. @@ -375,6 +376,7 @@ class LocallyConnected2D(Layer): all spatial dimensions. padding: Currently only support `"valid"` (case-insensitive). `"same"` will be supported in future. + `"valid"` means no padding. data_format: A string, one of `channels_last` (default) or `channels_first`. The ordering of the dimensions in the inputs. diff --git a/tensorflow/python/keras/layers/pooling.py b/tensorflow/python/keras/layers/pooling.py index ff7d157acad..51dc5131a8a 100644 --- a/tensorflow/python/keras/layers/pooling.py +++ b/tensorflow/python/keras/layers/pooling.py @@ -164,8 +164,9 @@ class MaxPooling1D(Pooling1D): for each pooling step. If None, it will default to `pool_size`. padding: One of `"valid"` or `"same"` (case-insensitive). - "valid" adds no padding. "same" adds padding such that if the stride - is 1, the output shape is the same as the input shape. + `"valid"` means no padding. `"same"` results in padding evenly to + the left/right or up/down of the input such that output has the same + height/width dimension as the input. data_format: A string, one of `channels_last` (default) or `channels_first`. The ordering of the dimensions in the inputs. @@ -209,6 +210,9 @@ class AveragePooling1D(Pooling1D): E.g. 2 will halve the input. If None, it will default to `pool_size`. padding: One of `"valid"` or `"same"` (case-insensitive). + `"valid"` means no padding. `"same"` results in padding evenly to + the left/right or up/down of the input such that output has the same + height/width dimension as the input. data_format: A string, one of `channels_last` (default) or `channels_first`. The ordering of the dimensions in the inputs. @@ -419,8 +423,9 @@ class MaxPooling2D(Pooling2D): Strides values. Specifies how far the pooling window moves for each pooling step. If None, it will default to `pool_size`. padding: One of `"valid"` or `"same"` (case-insensitive). - "valid" adds no zero padding. "same" adds padding such that if the stride - is 1, the output shape is the same as input shape. + `"valid"` means no padding. `"same"` results in padding evenly to + the left/right or up/down of the input such that output has the same + height/width dimension as the input. data_format: A string, one of `channels_last` (default) or `channels_first`. The ordering of the dimensions in the inputs. @@ -475,6 +480,9 @@ class AveragePooling2D(Pooling2D): Strides values. If None, it will default to `pool_size`. padding: One of `"valid"` or `"same"` (case-insensitive). + `"valid"` means no padding. `"same"` results in padding evenly to + the left/right or up/down of the input such that output has the same + height/width dimension as the input. data_format: A string, one of `channels_last` (default) or `channels_first`. The ordering of the dimensions in the inputs. @@ -617,6 +625,9 @@ class MaxPooling3D(Pooling3D): `(2, 2, 2)` will halve the size of the 3D input in each dimension. strides: tuple of 3 integers, or None. Strides values. padding: One of `"valid"` or `"same"` (case-insensitive). + `"valid"` means no padding. `"same"` results in padding evenly to + the left/right or up/down of the input such that output has the same + height/width dimension as the input. data_format: A string, one of `channels_last` (default) or `channels_first`. The ordering of the dimensions in the inputs. @@ -667,6 +678,9 @@ class AveragePooling3D(Pooling3D): `(2, 2, 2)` will halve the size of the 3D input in each dimension. strides: tuple of 3 integers, or None. Strides values. padding: One of `"valid"` or `"same"` (case-insensitive). + `"valid"` means no padding. `"same"` results in padding evenly to + the left/right or up/down of the input such that output has the same + height/width dimension as the input. data_format: A string, one of `channels_last` (default) or `channels_first`. The ordering of the dimensions in the inputs. diff --git a/tensorflow/python/keras/legacy_tf_layers/convolutional.py b/tensorflow/python/keras/legacy_tf_layers/convolutional.py index 4c91251a0e7..4fd53531fd1 100644 --- a/tensorflow/python/keras/legacy_tf_layers/convolutional.py +++ b/tensorflow/python/keras/legacy_tf_layers/convolutional.py @@ -46,6 +46,9 @@ class Conv1D(keras_layers.Conv1D, base.Layer): Specifying any stride value != 1 is incompatible with specifying any `dilation_rate` value != 1. padding: One of `"valid"` or `"same"` (case-insensitive). + `"valid"` means no padding. `"same"` results in padding evenly to + the left/right or up/down of the input such that output has the same + height/width dimension as the input. data_format: A string, one of `channels_last` (default) or `channels_first`. The ordering of the dimensions in the inputs. `channels_last` corresponds to inputs with shape @@ -157,6 +160,9 @@ def conv1d(inputs, Specifying any stride value != 1 is incompatible with specifying any `dilation_rate` value != 1. padding: One of `"valid"` or `"same"` (case-insensitive). + `"valid"` means no padding. `"same"` results in padding evenly to + the left/right or up/down of the input such that output has the same + height/width dimension as the input. data_format: A string, one of `channels_last` (default) or `channels_first`. The ordering of the dimensions in the inputs. `channels_last` corresponds to inputs with shape @@ -242,6 +248,9 @@ class Conv2D(keras_layers.Conv2D, base.Layer): Specifying any stride value != 1 is incompatible with specifying any `dilation_rate` value != 1. padding: One of `"valid"` or `"same"` (case-insensitive). + `"valid"` means no padding. `"same"` results in padding evenly to + the left/right or up/down of the input such that output has the same + height/width dimension as the input. data_format: A string, one of `channels_last` (default) or `channels_first`. The ordering of the dimensions in the inputs. `channels_last` corresponds to inputs with shape @@ -360,6 +369,9 @@ def conv2d(inputs, Specifying any stride value != 1 is incompatible with specifying any `dilation_rate` value != 1. padding: One of `"valid"` or `"same"` (case-insensitive). + `"valid"` means no padding. `"same"` results in padding evenly to + the left/right or up/down of the input such that output has the same + height/width dimension as the input. data_format: A string, one of `channels_last` (default) or `channels_first`. The ordering of the dimensions in the inputs. `channels_last` corresponds to inputs with shape @@ -449,6 +461,9 @@ class Conv3D(keras_layers.Conv3D, base.Layer): Specifying any stride value != 1 is incompatible with specifying any `dilation_rate` value != 1. padding: One of `"valid"` or `"same"` (case-insensitive). + `"valid"` means no padding. `"same"` results in padding evenly to + the left/right or up/down of the input such that output has the same + height/width dimension as the input. data_format: A string, one of `channels_last` (default) or `channels_first`. The ordering of the dimensions in the inputs. `channels_last` corresponds to inputs with shape @@ -568,6 +583,9 @@ def conv3d(inputs, Specifying any stride value != 1 is incompatible with specifying any `dilation_rate` value != 1. padding: One of `"valid"` or `"same"` (case-insensitive). + `"valid"` means no padding. `"same"` results in padding evenly to + the left/right or up/down of the input such that output has the same + height/width dimension as the input. data_format: A string, one of `channels_last` (default) or `channels_first`. The ordering of the dimensions in the inputs. `channels_last` corresponds to inputs with shape @@ -652,6 +670,9 @@ class SeparableConv1D(keras_layers.SeparableConv1D, base.Layer): Specifying any `stride` value != 1 is incompatible with specifying any `dilation_rate` value != 1. padding: One of `"valid"` or `"same"` (case-insensitive). + `"valid"` means no padding. `"same"` results in padding evenly to + the left/right or up/down of the input such that output has the same + height/width dimension as the input. data_format: A string, one of `channels_last` (default) or `channels_first`. The ordering of the dimensions in the inputs. `channels_last` corresponds to inputs with shape @@ -761,6 +782,9 @@ class SeparableConv2D(keras_layers.SeparableConv2D, base.Layer): Specifying any `stride` value != 1 is incompatible with specifying any `dilation_rate` value != 1. padding: One of `"valid"` or `"same"` (case-insensitive). + `"valid"` means no padding. `"same"` results in padding evenly to + the left/right or up/down of the input such that output has the same + height/width dimension as the input. data_format: A string, one of `channels_last` (default) or `channels_first`. The ordering of the dimensions in the inputs. `channels_last` corresponds to inputs with shape @@ -897,6 +921,9 @@ def separable_conv1d(inputs, Specifying any `stride` value != 1 is incompatible with specifying any `dilation_rate` value != 1. padding: One of `"valid"` or `"same"` (case-insensitive). + `"valid"` means no padding. `"same"` results in padding evenly to + the left/right or up/down of the input such that output has the same + height/width dimension as the input. data_format: A string, one of `channels_last` (default) or `channels_first`. The ordering of the dimensions in the inputs. `channels_last` corresponds to inputs with shape @@ -1019,6 +1046,9 @@ def separable_conv2d(inputs, Specifying any `stride` value != 1 is incompatible with specifying any `dilation_rate` value != 1. padding: One of `"valid"` or `"same"` (case-insensitive). + `"valid"` means no padding. `"same"` results in padding evenly to + the left/right or up/down of the input such that output has the same + height/width dimension as the input. data_format: A string, one of `channels_last` (default) or `channels_first`. The ordering of the dimensions in the inputs. `channels_last` corresponds to inputs with shape @@ -1117,6 +1147,9 @@ class Conv2DTranspose(keras_layers.Conv2DTranspose, base.Layer): of the convolution. Can be a single integer to specify the same value for all spatial dimensions. padding: one of `"valid"` or `"same"` (case-insensitive). + `"valid"` means no padding. `"same"` results in padding evenly to + the left/right or up/down of the input such that output has the same + height/width dimension as the input. data_format: A string, one of `channels_last` (default) or `channels_first`. The ordering of the dimensions in the inputs. `channels_last` corresponds to inputs with shape @@ -1223,6 +1256,9 @@ def conv2d_transpose(inputs, of the convolution. Can be a single integer to specify the same value for all spatial dimensions. padding: one of `"valid"` or `"same"` (case-insensitive). + `"valid"` means no padding. `"same"` results in padding evenly to + the left/right or up/down of the input such that output has the same + height/width dimension as the input. data_format: A string, one of `channels_last` (default) or `channels_first`. The ordering of the dimensions in the inputs. `channels_last` corresponds to inputs with shape @@ -1295,6 +1331,9 @@ class Conv3DTranspose(keras_layers.Conv3DTranspose, base.Layer): Can be a single integer to specify the same value for all spatial dimensions. padding: One of `"valid"` or `"same"` (case-insensitive). + `"valid"` means no padding. `"same"` results in padding evenly to + the left/right or up/down of the input such that output has the same + height/width dimension as the input. data_format: A string, one of `channels_last` (default) or `channels_first`. The ordering of the dimensions in the inputs. `channels_last` corresponds to inputs with shape @@ -1396,6 +1435,9 @@ def conv3d_transpose(inputs, of the convolution. Can be a single integer to specify the same value for all spatial dimensions. padding: one of `"valid"` or `"same"` (case-insensitive). + `"valid"` means no padding. `"same"` results in padding evenly to + the left/right or up/down of the input such that output has the same + height/width dimension as the input. data_format: A string, one of `channels_last` (default) or `channels_first`. The ordering of the dimensions in the inputs. `channels_last` corresponds to inputs with shape diff --git a/tensorflow/python/keras/utils/conv_utils.py b/tensorflow/python/keras/utils/conv_utils.py index f38fdc18252..e8ee866d958 100644 --- a/tensorflow/python/keras/utils/conv_utils.py +++ b/tensorflow/python/keras/utils/conv_utils.py @@ -264,6 +264,9 @@ def conv_kernel_mask(input_shape, kernel_shape, strides, padding): receptive field. strides: tuple of size N, strides along each spatial dimension. padding: type of padding, string `"same"` or `"valid"`. + `"valid"` means no padding. `"same"` results in padding evenly to + the left/right or up/down of the input such that output has the same + height/width dimension as the input. Returns: A boolean 2N-D `np.ndarray` of shape @@ -338,6 +341,9 @@ def conv_kernel_idxs(input_shape, kernel_shape, strides, padding, filters_in, receptive field. strides: tuple of size N, strides along each spatial dimension. padding: type of padding, string `"same"` or `"valid"`. + `"valid"` means no padding. `"same"` results in padding evenly to + the left/right or up/down of the input such that output has the same + height/width dimension as the input. filters_in: `int`, number if filters in the input to the layer. filters_out: `int', number if filters in the output of the layer. data_format: string, "channels_first" or "channels_last". @@ -430,6 +436,9 @@ def conv_connected_inputs(input_shape, kernel_shape, output_position, strides, in the output of the convolution. strides: tuple of size N, strides along each spatial dimension. padding: type of padding, string `"same"` or `"valid"`. + `"valid"` means no padding. `"same"` results in padding evenly to + the left/right or up/down of the input such that output has the same + height/width dimension as the input. Returns: N ranges `[[p_in_left1, ..., p_in_right1], ..., @@ -468,6 +477,9 @@ def conv_output_shape(input_shape, kernel_shape, strides, padding): receptive field. strides: tuple of size N, strides along each spatial dimension. padding: type of padding, string `"same"` or `"valid"`. + `"valid"` means no padding. `"same"` results in padding evenly to + the left/right or up/down of the input such that output has the same + height/width dimension as the input. Returns: tuple of size N: `(d_out1, ..., d_outN)`, spatial shape of the output. diff --git a/tensorflow/python/ops/nn_ops.py b/tensorflow/python/ops/nn_ops.py index 1318f575737..5a9a63637f6 100644 --- a/tensorflow/python/ops/nn_ops.py +++ b/tensorflow/python/ops/nn_ops.py @@ -940,6 +940,9 @@ def convolution( filter: An (N+2)-D `Tensor` with the same type as `input` and shape `spatial_filter_shape + [in_channels, out_channels]`. padding: A string, either `"VALID"` or `"SAME"`. The padding algorithm. + `"valid"` means no padding. `"same"` results in padding evenly to + the left/right or up/down of the input such that output has the same + height/width dimension as the input. strides: Optional. Sequence of N ints >= 1. Specifies the output stride. Defaults to [1]*N. If any value of strides is > 1, then all values of dilation_rate must be 1. From c0ba8a09a7edf2b617d10024f7e0ec9abe21d73e Mon Sep 17 00:00:00 2001 From: Xinyi Wang Date: Wed, 17 Jun 2020 18:24:26 -0700 Subject: [PATCH 143/867] Export and Document DistributedDataset and DistributedIterator APIs PiperOrigin-RevId: 317007583 Change-Id: I7d7c4615a12a19fb4fd151a0457f176ffe2cd765 --- .../python/distribute/distribute_lib.py | 120 ++++--- tensorflow/python/distribute/input_lib.py | 324 +++++++++++++++++- tensorflow/python/distribute/values.py | 17 +- ...flow.distribute.-distributed-dataset.pbtxt | 16 + ...low.distribute.-distributed-iterator.pbtxt | 16 + .../api/golden/v2/tensorflow.distribute.pbtxt | 8 + 6 files changed, 444 insertions(+), 57 deletions(-) create mode 100644 tensorflow/tools/api/golden/v2/tensorflow.distribute.-distributed-dataset.pbtxt create mode 100644 tensorflow/tools/api/golden/v2/tensorflow.distribute.-distributed-iterator.pbtxt diff --git a/tensorflow/python/distribute/distribute_lib.py b/tensorflow/python/distribute/distribute_lib.py index a6dc35507e9..b6a89463426 100644 --- a/tensorflow/python/distribute/distribute_lib.py +++ b/tensorflow/python/distribute/distribute_lib.py @@ -684,7 +684,8 @@ class StrategyBase(object): instead. * Use `tf.distribute.Strategy.run` to run a function once per replica, taking values that may be "per-replica" (e.g. - from a distributed dataset) and returning "per-replica" values. + from a `tf.distribute.DistributedDataset` object) and returning + "per-replica" values. This function is executed in "replica context", which means each operation is performed separately on each replica. * Finally use a method (such as `tf.distribute.Strategy.reduce`) to @@ -720,7 +721,8 @@ class StrategyBase(object): distributed-specific behavior. You can use the `reduce` API to aggregate results across replicas and use - this as a return value from one iteration over the distributed dataset. Or + this as a return value from one iteration over a + `tf.distribute.DistributedDataset`. Or you can use `tf.keras.metrics` (such as loss, accuracy, etc.) to accumulate metrics across steps in a given epoch. @@ -859,12 +861,12 @@ class StrategyBase(object): return self.run(fn, args=args) def experimental_distribute_dataset(self, dataset, options=None): - """Distributes a tf.data.Dataset instance provided via `dataset`. + """Creates `tf.distribute.DistributedDataset` from `tf.data.Dataset`. - The returned distributed dataset can be iterated over similar to how - regular datasets can. - NOTE: Currently, the user cannot add any more transformations to a - distributed dataset. + The returned `tf.distribute.DistributedDataset` can be iterated over + similar to how regular datasets can. + NOTE: The user cannot add any more transformations to a + `tf.distribute.DistributedDataset`. The following is an example: @@ -878,48 +880,53 @@ class StrategyBase(object): # Distribute that dataset dist_dataset = strategy.experimental_distribute_dataset(dataset) - # Iterate over the distributed dataset + # Iterate over the `tf.distribute.DistributedDataset` for x in dist_dataset: # process dataset elements strategy.run(replica_fn, args=(x,)) ``` - In the code snippet above, the dataset `dist_dataset` is batched by - GLOBAL_BATCH_SIZE, and we iterate through it using `for x in dist_dataset`, - where x is one batch of data of GLOBAL_BATCH_SIZE containing N batches of - data of per-replica batch size, corresponding to N replicas. - `tf.distribute.Strategy.run` will take care of feeding - the right per-replica batch to the right `replica_fn` execution on each + In the code snippet above, the `tf.distribute.DistributedDataset` + `dist_dataset` is batched by `GLOBAL_BATCH_SIZE`, and we iterate through it + using `for x in dist_dataset`. `x` a `tf.distribute.DistributedValues` + containing data for all replicas, which aggregates to a batch of + `GLOBAL_BATCH_SIZE`. `tf.distribute.Strategy.run` will take care of feeding + the right per-replica data in `x` to the right `replica_fn` executed on each replica. - In a multi-worker setting, we will first attempt to distribute the dataset - by attempting to detect whether the dataset is being created out of - ReaderDatasets (e.g. TFRecordDataset, TextLineDataset, etc.) and if so, - attempting to shard the input files. Note that there has to be at least one - input file per worker. If you have less than one input file per worker, we - suggest that you should disable distributing your dataset using the method - below. + What's under the hood of this method, when we say the `tf.data.Dataset` + instance - `dataset` - gets distributed? It depends on how you set the + `tf.data.experimental.AutoShardPolicy` through + `tf.data.experimental.DistributeOptions`. By default, it is set to + `tf.data.experimental.AutoShardPolicy.AUTO`. In a multi-worker setting, we + will first attempt to distribute `dataset` by detecting whether `dataset` is + being created out of reader datasets (e.g. `tf.data.TFRecordDataset`, + `tf.data.TextLineDataset`, etc.) and if so, try to shard the input files. + Note that there has to be at least one input file per worker. If you have + less than one input file per worker, we suggest that you disable dataset + sharding across workers, by setting the + `tf.data.experimental.DistributeOptions.auto_shard_policy` to be + `tf.data.experimental.AutoShardPolicy.OFF`. - If that attempt is unsuccessful (e.g. the dataset is created from a - Dataset.range), we will shard the dataset evenly at the end by appending a - `.shard` operation to the end of the processing pipeline. This will cause - the entire preprocessing pipeline for all the data to be run on every - worker, and each worker will do redundant work. We will print a warning - if this method of sharding is selected. + If the attempt to shard by file is unsuccessful (i.e. the dataset is not + read from files), we will shard the dataset evenly at the end by + appending a `.shard` operation to the end of the processing pipeline. This + will cause the entire preprocessing pipeline for all the data to be run on + every worker, and each worker will do redundant work. We will print a + warning if this route is selected. - You can disable dataset sharding across workers using the - `auto_shard_policy` option in `tf.data.experimental.DistributeOptions`. - - Within each worker, we will also split the data among all the worker - devices (if more than one a present), and this will happen even if - multi-worker sharding is disabled using the method above. + As mentioned before, within each worker, we will also split the data among + all the worker devices (if more than one a present). This will happen + even if multi-worker sharding is disabled. If the above batch splitting and dataset sharding logic is undesirable, - please use `experimental_distribute_datasets_from_function` instead, which - does not do any automatic splitting or sharding. + please use + `tf.distribute.Strategy.experimental_distribute_datasets_from_function` + instead, which does not do any automatic splitting or sharding. - You can also use the `element_spec` property of the distributed dataset - returned by this API to query the `tf.TypeSpec` of the elements returned + You can also use the `element_spec` property of the + `tf.distribute.DistributedDataset` instance returned by this API to query + the `tf.TypeSpec` of the elements returned by the iterator. This can be used to set the `input_signature` property of a `tf.function`. @@ -938,12 +945,21 @@ class StrategyBase(object): # train model with inputs return - # Iterate over the distributed dataset + # Iterate over the `tf.distribute.DistributedDataset` for x in dist_dataset: # process dataset elements strategy.run(train_step, args=(x,)) ``` + Note: The order in which the data is processed by the workers when using + `tf.distribute.Strategy.experimental_distribute_dataset` or + `tf.distribute.Strategy.experimental_distribute_datasets_from_function` is + not guaranteed. This is typically required if you are using + `tf.distribute` to scale prediction. You can however insert an index for + each element in the batch and order outputs accordingly. Refer to [this + snippet](https://www.tensorflow.org/tutorials/distribute/input#caveats) + for an example of how to order outputs. + Args: dataset: `tf.data.Dataset` that will be sharded across all replicas using the rules stated above. @@ -951,8 +967,7 @@ class StrategyBase(object): dataset is distributed. Returns: - A "distributed `Dataset`", which acts like a `tf.data.Dataset` except - it produces "per-replica" values. + A `tf.distribute.DistributedDataset`. """ return self._extended._experimental_distribute_dataset(dataset, options) # pylint: disable=protected-access @@ -978,10 +993,10 @@ class StrategyBase(object): The `dataset_fn` should take an `tf.distribute.InputContext` instance where information about batching and input replication can be accessed. - You can also use the `element_spec` property of the distributed dataset - returned by this API to query the `tf.TypeSpec` of the elements returned - by the iterator. This can be used to set the `input_signature` property - of a `tf.function`. + You can also use the `element_spec` property of the + `tf.distribute.DistributedDataset` returned by this API to query the + `tf.TypeSpec` of the elements returned by the iterator. This can be used to + set the `input_signature` property of a `tf.function`. >>> global_batch_size = 8 >>> def dataset_fn(input_context): @@ -1010,6 +1025,16 @@ class StrategyBase(object): the global batch size. This may be computed using `input_context.get_per_replica_batch_size`. + + Note: The order in which the data is processed by the workers when using + `tf.distribute.Strategy.experimental_distribute_dataset` or + `tf.distribute.Strategy.experimental_distribute_datasets_from_function` is + not guaranteed. This is typically required if you are using + `tf.distribute` to scale prediction. You can however insert an index for + each element in the batch and order outputs accordingly. Refer to [this + snippet](https://www.tensorflow.org/tutorials/distribute/input#caveats) + for an example of how to order outputs. + Args: dataset_fn: A function taking a `tf.distribute.InputContext` instance and returning a `tf.data.Dataset`. @@ -1017,8 +1042,7 @@ class StrategyBase(object): dataset is distributed. Returns: - A "distributed `Dataset`", which acts like a `tf.data.Dataset` except - it produces "per-replica" values. + A `tf.distribute.DistributedDataset`. """ return self._extended._experimental_distribute_datasets_from_function( # pylint: disable=protected-access dataset_fn, options) @@ -1028,7 +1052,9 @@ class StrategyBase(object): Executes ops specified by `fn` on each replica. If `args` or `kwargs` have `tf.distribute.DistributedValues`, such as those produced by a - "distributed `Dataset`" or `experimental_distribute_values_from_function` + `tf.distribute.DistributedDataset` from + `tf.distribute.Strategy.experimental_distribute_dataset` or + `tf.distribute.Strategy.experimental_distribute_datasets_from_function`, when `fn` is executed on a particular replica, it will be executed with the component of `tf.distribute.DistributedValues` that correspond to that replica. diff --git a/tensorflow/python/distribute/input_lib.py b/tensorflow/python/distribute/input_lib.py index 85e2dac1c1d..ff468af7f87 100644 --- a/tensorflow/python/distribute/input_lib.py +++ b/tensorflow/python/distribute/input_lib.py @@ -18,6 +18,7 @@ from __future__ import absolute_import from __future__ import division from __future__ import print_function +import collections import functools import sys @@ -52,6 +53,8 @@ from tensorflow.python.ops.ragged import ragged_tensor from tensorflow.python.types import distribute as distribute_types from tensorflow.python.util import nest from tensorflow.python.util.deprecation import deprecated +from tensorflow.python.util.tf_export import tf_export +from tensorflow.tools.docs import doc_controls def get_distributed_dataset(dataset, @@ -138,6 +141,321 @@ def get_distributed_datasets_from_function(dataset_fn, strategy) +@tf_export("distribute.DistributedIterator", v1=[]) +class DistributedIteratorInterface(collections.Iterator, + distribute_types.Iterator): + """An iterator over `tf.distribute.DistributedDataset`. + + `tf.distribute.DistributedIterator` is the primary mechanism for enumerating + elements of a `tf.distribute.DistributedDataset`. It supports the Python + Iterator protocol, which means it can be iterated over using a for-loop or by + fetching individual elements explicitly via `get_next()`. + + You can create a `tf.distribute.DistributedIterator` by calling `iter` on + a `tf.distribute.DistributedDataset` or creating a python loop over a + `tf.distribute.DistributedDataset`. + + Visit the [tutorial](https://www.tensorflow.org/tutorials/distribute/input) + on distributed input for more examples and caveats. + """ + + def get_next(self): + """Returns the next input from the iterator for all replicas. + + Example use: + + >>> strategy = tf.distribute.MirroredStrategy() + >>> dataset = tf.data.Dataset.range(100).batch(2) + >>> dist_dataset = strategy.experimental_distribute_dataset(dataset) + >>> dist_dataset_iterator = iter(dist_dataset) + >>> @tf.function + ... def one_step(input): + ... return input + >>> step_num = 5 + >>> for _ in range(step_num): + ... strategy.run(one_step, args=(dist_dataset_iterator.get_next(),)) + >>> strategy.experimental_local_results(dist_dataset_iterator.get_next()) + (,) + + The above example corresponds to the case where you have only one device. If + you have two devices, for example, + ```python + strategy = tf.distribute.MirroredStrategy(['/gpu:0', '/gpu:1']) + ``` + Then the final line will print out: + ```python + (, + ) + ``` + + Returns: + A single `tf.Tensor` or a `tf.distribute.DistributedValues` which contains + the next input for all replicas. + + Raises: + `tf.errors.OutOfRangeError`: If the end of the iterator has been reached. + """ + raise NotImplementedError( + "DistributedIterator.get_next() must be implemented in descendants.") + + @property + def element_spec(self): + # pylint: disable=line-too-long + """The type specification of an element of `tf.distribute.DistributedIterator`. + + Example usage: + + >>> global_batch_size = 16 + >>> strategy = tf.distribute.MirroredStrategy() + >>> dataset = tf.data.Dataset.from_tensors(([1.],[2])).repeat(100).batch(global_batch_size) + >>> distributed_iterator = iter(strategy.experimental_distribute_dataset(dataset)) + >>> distributed_iterator.element_spec + (TensorSpec(shape=(None, 1), dtype=tf.float32, name=None), + TensorSpec(shape=(None, 1), dtype=tf.int32, name=None)) + + The above example corresponds to the case where you have only one device. If + you have two devices, for example, + ```python + strategy = tf.distribute.MirroredStrategy(['/gpu:0', '/gpu:1']) + ``` + Then the final line will print out: + ```python + (PerReplicaSpec(TensorSpec(shape=(None, 1), dtype=tf.float32, name=None), + TensorSpec(shape=(None, 1), dtype=tf.float32, name=None)), + PerReplicaSpec(TensorSpec(shape=(None, 1), dtype=tf.int32, name=None), + TensorSpec(shape=(None, 1), dtype=tf.int32, name=None))) + ``` + + Returns: + A nested structure of `tf.TypeSpec` objects matching the structure of an + element of this `tf.distribute.DistributedIterator`. This returned value + is typically a `tf.distribute.DistributedValues` object and specifies the + `tf.TensorSpec` of individual components. + """ + raise NotImplementedError( + "DistributedIterator.element_spec() must be implemented in descendants") + + +@tf_export("distribute.DistributedDataset", v1=[]) +class DistributedDatasetInterface(collections.Iterable, + distribute_types.Iterable): + # pylint: disable=line-too-long + """Represents a dataset distributed among devices and machines. + + A `tf.distribute.DistributedDataset` could be thought of as a "distributed" + dataset. When you use `tf.distribute` API to scale training to multiple + devices or machines, you also need to distribute the input data, which leads + to a `tf.distribute.DistributedDataset` instance, instead of a + `tf.data.Dataset` instance in the non-distributed case. In TF 2.x, + `tf.distribute.DistributedDataset` objects are Python iterables. + + Note: `tf.distribute.DistributedDataset` instances are *not* of type + `tf.data.Dataset`. It only supports two usages we will mention below: + iteration and `element_spec`. We don't support any other APIs to transform or + inspect the dataset. + + There are two APIs to create a `tf.distribute.DistributedDataset` object: + `tf.distribute.Strategy.experimental_distribute_dataset(dataset)`and + `tf.distribute.Strategy.experimental_distribute_datasets_from_function(dataset_fn)`. + *When to use which?* When you have a `tf.data.Dataset` instance, and the + regular batch splitting (i.e. re-batch the input `tf.data.Dataset` instance + with a new batch size that is equal to the global batch size divided by the + number of replicas in sync) and autosharding (i.e. the + `tf.data.experimental.AutoShardPolicy` options) work for you, use the former + API. Otherwise, if you are *not* using a canonical `tf.data.Dataset` instance, + or you would like to customize the batch splitting or sharding, you can wrap + these logic in a `dataset_fn` and use the latter API. Both API handles + prefetch to device for the user. For more details and examples, follow the + links to the APIs. + + + There are two main usages of a `DistributedDataset` object: + + 1. Iterate over it to generate the input for a single device or multiple + devices, which is a `tf.distribute.DistributedValues` instance. To do this, + you can: + + * use a pythonic for-loop construct: + + >>> global_batch_size = 2 + >>> strategy = tf.distribute.MirroredStrategy() + >>> dataset = tf.data.Dataset.from_tensors(([1.],[1.])).repeat(4).batch(global_batch_size) + >>> dist_dataset = strategy.experimental_distribute_dataset(dataset) + >>> @tf.function + ... def train_step(input): + ... features, labels = input + ... return labels - 0.3 * features + >>> for x in dist_dataset: + ... # train_step trains the model using the dataset elements + ... loss = strategy.run(train_step, args=(x,)) + ... print("Loss is", loss) + Loss is tf.Tensor( + [[0.7] + [0.7]], shape=(2, 1), dtype=float32) + Loss is tf.Tensor( + [[0.7] + [0.7]], shape=(2, 1), dtype=float32) + + Placing the loop inside a `tf.function` will give a performance boost. + However `break` and `return` are currently not supported if the loop is + placed inside a `tf.function`. We also don't support placing the loop + inside a `tf.function` when using + `tf.distribute.experimental.MultiWorkerMirroredStrategy` or + `tf.distribute.experimental.TPUStrategy` with multiple workers. + + * use `__iter__` to create an explicit iterator, which is of type + `tf.distribute.DistributedIterator` + + >>> global_batch_size = 4 + >>> strategy = tf.distribute.MirroredStrategy() + >>> train_dataset = tf.data.Dataset.from_tensors(([1.],[1.])).repeat(50).batch(global_batch_size) + >>> train_dist_dataset = strategy.experimental_distribute_dataset(train_dataset) + >>> @tf.function + ... def distributed_train_step(dataset_inputs): + ... def train_step(input): + ... loss = tf.constant(0.1) + ... return loss + ... per_replica_losses = strategy.run(train_step, args=(dataset_inputs,)) + ... return strategy.reduce(tf.distribute.ReduceOp.SUM, per_replica_losses,axis=None) + >>> EPOCHS = 2 + >>> STEPS = 3 + >>> for epoch in range(EPOCHS): + ... total_loss = 0.0 + ... num_batches = 0 + ... dist_dataset_iterator = iter(train_dist_dataset) + ... for _ in range(STEPS): + ... total_loss += distributed_train_step(next(dist_dataset_iterator)) + ... num_batches += 1 + ... average_train_loss = total_loss / num_batches + ... template = ("Epoch {}, Loss: {}") + ... print (template.format(epoch+1, average_train_loss)) + Epoch 1, Loss: 0.10000000894069672 + Epoch 2, Loss: 0.10000000894069672 + + + To achieve a performance improvement, you can also wrap the `strategy.run` + call with a `tf.range` inside a `tf.function`. This runs multiple steps in a + `tf.function`. Autograph will convert it to a `tf.while_loop` on the worker. + However, it is less flexible comparing with running a single step inside + `tf.function`. For example, you cannot run things eagerly or arbitrary + python code within the steps. + + + 2. Inspect the `tf.TypeSpec` of the data generated by `DistributedDataset`. + + `tf.distribute.DistributedDataset` generates + `tf.distribute.DistributedValues` as input to the devices. If you pass the + input to a `tf.function` and would like to specify the shape and type of + each Tensor argument to the function, you can pass a `tf.TypeSpec` object to + the `input_signature` argument of the `tf.function`. To get the + `tf.TypeSpec` of the input, you can use the `element_spec` property of the + `tf.distribute.DistributedDataset` or `tf.distribute.DistributedIterator` + object. + + For example: + + >>> global_batch_size = 2 + >>> epochs = 1 + >>> steps_per_epoch = 1 + >>> mirrored_strategy = tf.distribute.MirroredStrategy() + >>> dataset = tf.data.Dataset.from_tensors(([2.])).repeat(100).batch(global_batch_size) + >>> dist_dataset = mirrored_strategy.experimental_distribute_dataset(dataset) + >>> @tf.function(input_signature=[dist_dataset.element_spec]) + ... def train_step(per_replica_inputs): + ... def step_fn(inputs): + ... return tf.square(inputs) + ... return mirrored_strategy.run(step_fn, args=(per_replica_inputs,)) + >>> for _ in range(epochs): + ... iterator = iter(dist_dataset) + ... for _ in range(steps_per_epoch): + ... output = train_step(next(iterator)) + ... print(output) + tf.Tensor( + [[4.] + [4.]], shape=(2, 1), dtype=float32) + + + Visit the [tutorial](https://www.tensorflow.org/tutorials/distribute/input) + on distributed input for more examples and caveats. + """ + + def __iter__(self): + """Creates an iterator for the `tf.distribute.DistributedDataset`. + + The returned iterator implements the Python Iterator protocol. + + Example usage: + + >>> global_batch_size = 4 + >>> strategy = tf.distribute.MirroredStrategy() + >>> dataset = tf.data.Dataset.from_tensor_slices([1, 2, 3, 4]).repeat().batch(global_batch_size) + >>> distributed_iterator = iter(strategy.experimental_distribute_dataset(dataset)) + >>> print(next(distributed_iterator)) + tf.Tensor([1 2 3 4], shape=(4,), dtype=int32) + + + The above example corresponds to the case where you have only one device. If + you have two devices, for example, + ```python + strategy = tf.distribute.MirroredStrategy(['/gpu:0', '/gpu:1']) + ``` + Then the final line will print out: + ```python + PerReplica:{ + 0: tf.Tensor([1 2], shape=(2,), dtype=int32), + 1: tf.Tensor([3 4], shape=(2,), dtype=int32) + } + ``` + + Returns: + An `tf.distribute.DistributedIterator` instance for the given + `tf.distribute.DistributedDataset` object to enumerate over the + distributed data. + """ + raise NotImplementedError("Must be implemented in descendants") + + @property + def element_spec(self): + """The type specification of an element of this `tf.distribute.DistributedDataset`. + + Example usage: + + >>> global_batch_size = 16 + >>> strategy = tf.distribute.MirroredStrategy() + >>> dataset = tf.data.Dataset.from_tensors(([1.],[2])).repeat(100).batch(global_batch_size) + >>> dist_dataset = strategy.experimental_distribute_dataset(dataset) + >>> dist_dataset.element_spec + (TensorSpec(shape=(None, 1), dtype=tf.float32, name=None), + TensorSpec(shape=(None, 1), dtype=tf.int32, name=None)) + + The above example corresponds to the case where you have only one device. If + you have two devices, for example, + ```python + strategy = tf.distribute.MirroredStrategy(['/gpu:0', '/gpu:1']) + ``` + Then the final line will print out: + ```python + (PerReplicaSpec(TensorSpec(shape=(None, 1), dtype=tf.float32, name=None), + TensorSpec(shape=(None, 1), dtype=tf.float32, name=None)), + PerReplicaSpec(TensorSpec(shape=(None, 1), dtype=tf.int32, name=None), + TensorSpec(shape=(None, 1), dtype=tf.int32, name=None))) + ``` + + Returns: + A nested structure of `tf.TypeSpec` objects matching the structure of an + element of this `tf.distribute.DistributedDataset`. This returned value is + typically a `tf.distribute.DistributedValues` object and specifies the + `tf.TensorSpec` of individual components. + """ + raise NotImplementedError( + "DistributedDataset.element_spec must be implemented in descendants.") + + @doc_controls.do_not_generate_docs + def reduce(self, initial_state, reduce_func): + raise NotImplementedError( + "DistributedDataset.reduce must be implemented in descendants.") + + class InputWorkers(object): """A 1-to-many mapping from input worker devices to compute devices.""" @@ -259,9 +577,10 @@ def _get_static_shape(iterators): return static_shape -class DistributedIteratorBase(distribute_types.Iterator): +class DistributedIteratorBase(DistributedIteratorInterface): """Common implementation for all input iterators.""" + # pylint: disable=super-init-not-called def __init__(self, input_workers, iterators, strategy): static_shape = _get_static_shape(iterators) @@ -548,9 +867,10 @@ class DistributedIterator(DistributedIteratorBase, self._strategy) -class _IterableInput(distribute_types.Iterable): +class _IterableInput(DistributedDatasetInterface): """Base class for iterable inputs for distribution strategies.""" + # pylint: disable=super-init-not-called def __init__(self, input_workers): assert isinstance(input_workers, InputWorkers) self._input_workers = input_workers diff --git a/tensorflow/python/distribute/values.py b/tensorflow/python/distribute/values.py index c6e0eb34a7b..d0ed27c69de 100644 --- a/tensorflow/python/distribute/values.py +++ b/tensorflow/python/distribute/values.py @@ -75,20 +75,21 @@ def _on_write_update_replica(var, update_fn, value, **kwargs): class DistributedValues(object): """Base class for representing distributed values. - A subclass instance of DistributedValues is created when creating variables - within a distribution strategy, iterating a `tf.Dataset` or through - `strategy.run`. This base class should never be instantiated - directly. DistributedValues contains a value per replica. Depending on + A subclass instance of `tf.distribute.DistributedValues` is created when + creating variables within a distribution strategy, iterating a + `tf.distribute.DistributedDataset` or through `tf.distribute.Strategy.run`. + This base class should never be instantiated directly. + `tf.distribute.DistributedValues` contains a value per replica. Depending on the subclass, the values could either be synced on update, synced on demand, or never synced. - DistributedValues can be reduced to obtain single value across replicas, - as input into `run` or the per replica values inspected - using `experimental_local_results`. + `tf.distribute.DistributedValues` can be reduced to obtain single value across + replicas, as input into `tf.distribute.Strategy.run` or the per-replica values + inspected using `tf.distribute.Strategy.experimental_local_results`. Example usage: - 1. Created from Dataset: + 1. Created from a `tf.distribute.DistributedDataset`: >>> strategy = tf.distribute.MirroredStrategy() >>> dataset = tf.data.Dataset.from_tensor_slices([5., 6., 7., 8.]).batch(2) diff --git a/tensorflow/tools/api/golden/v2/tensorflow.distribute.-distributed-dataset.pbtxt b/tensorflow/tools/api/golden/v2/tensorflow.distribute.-distributed-dataset.pbtxt new file mode 100644 index 00000000000..a7b229c6c7c --- /dev/null +++ b/tensorflow/tools/api/golden/v2/tensorflow.distribute.-distributed-dataset.pbtxt @@ -0,0 +1,16 @@ +path: "tensorflow.distribute.DistributedDataset" +tf_class { + is_instance: "" + is_instance: "" + member { + name: "element_spec" + mtype: "" + } + member_method { + name: "__init__" + } + member_method { + name: "reduce" + argspec: "args=[\'self\', \'initial_state\', \'reduce_func\'], varargs=None, keywords=None, defaults=None" + } +} diff --git a/tensorflow/tools/api/golden/v2/tensorflow.distribute.-distributed-iterator.pbtxt b/tensorflow/tools/api/golden/v2/tensorflow.distribute.-distributed-iterator.pbtxt new file mode 100644 index 00000000000..f712d9058b9 --- /dev/null +++ b/tensorflow/tools/api/golden/v2/tensorflow.distribute.-distributed-iterator.pbtxt @@ -0,0 +1,16 @@ +path: "tensorflow.distribute.DistributedIterator" +tf_class { + is_instance: "" + is_instance: "" + member { + name: "element_spec" + mtype: "" + } + member_method { + name: "__init__" + } + member_method { + name: "get_next" + argspec: "args=[\'self\'], varargs=None, keywords=None, defaults=None" + } +} diff --git a/tensorflow/tools/api/golden/v2/tensorflow.distribute.pbtxt b/tensorflow/tools/api/golden/v2/tensorflow.distribute.pbtxt index 19d83909120..009cb7fe400 100644 --- a/tensorflow/tools/api/golden/v2/tensorflow.distribute.pbtxt +++ b/tensorflow/tools/api/golden/v2/tensorflow.distribute.pbtxt @@ -4,6 +4,14 @@ tf_module { name: "CrossDeviceOps" mtype: "" } + member { + name: "DistributedDataset" + mtype: "" + } + member { + name: "DistributedIterator" + mtype: "" + } member { name: "DistributedValues" mtype: "" From a50001edf9fea03069771934e94d3b3d32ff6a19 Mon Sep 17 00:00:00 2001 From: "A. Unique TensorFlower" Date: Wed, 17 Jun 2020 18:30:26 -0700 Subject: [PATCH 144/867] Internal change PiperOrigin-RevId: 317008433 Change-Id: I5146e28d2f77c7daab4bd023f2826ec8323cff02 --- tensorflow/tensorflow.bzl | 8 ++++---- 1 file changed, 4 insertions(+), 4 deletions(-) diff --git a/tensorflow/tensorflow.bzl b/tensorflow/tensorflow.bzl index f97363a919e..5da15b0a4d6 100644 --- a/tensorflow/tensorflow.bzl +++ b/tensorflow/tensorflow.bzl @@ -1817,13 +1817,13 @@ def tf_custom_op_library_additional_deps_impl(): # tf_collected_deps will be the union of the deps of the current target # and the tf_collected_deps of the dependencies of this target. def _collect_deps_aspect_impl(target, ctx): - alldeps = depset() + direct, transitive = [], [] if hasattr(ctx.rule.attr, "deps"): for dep in ctx.rule.attr.deps: - alldeps = depset([dep.label], transitive = [alldeps]) + direct.append(dep.label) if hasattr(dep, "tf_collected_deps"): - alldeps = depset(transitive = [alldeps, dep.tf_collected_deps]) - return struct(tf_collected_deps = alldeps) + transitive.append(dep.tf_collected_deps) + return struct(tf_collected_deps = depset(direct = direct, transitive = transitive)) collect_deps_aspect = aspect( attr_aspects = ["deps"], From ae20f08da9ce9e7336ab97cc9f77ce7a1c13ad12 Mon Sep 17 00:00:00 2001 From: Yunxing Dai Date: Wed, 17 Jun 2020 18:43:41 -0700 Subject: [PATCH 145/867] Properly support nest phi reduction in reverse order. If we replaced node B with node C, then replace node A with node B, we should redirect node A to node C instead. PiperOrigin-RevId: 317010443 Change-Id: I165496a3d1f6571815bfd61d096e26cbba39125a --- .../xla/service/hlo_dataflow_analysis.cc | 2 ++ .../compiler/xla/service/hlo_phi_graph.cc | 25 ++++++++++++++++++- .../compiler/xla/service/hlo_phi_graph.h | 2 +- .../xla/service/hlo_phi_graph_test.cc | 25 +++++++++++++++++++ 4 files changed, 52 insertions(+), 2 deletions(-) diff --git a/tensorflow/compiler/xla/service/hlo_dataflow_analysis.cc b/tensorflow/compiler/xla/service/hlo_dataflow_analysis.cc index f19882c9347..a46d20d5808 100644 --- a/tensorflow/compiler/xla/service/hlo_dataflow_analysis.cc +++ b/tensorflow/compiler/xla/service/hlo_dataflow_analysis.cc @@ -1007,6 +1007,8 @@ void HloDataflowAnalysis::OptimizePhiValues() { HloValue::Id phi_id = values[0]->id(); HloValue::Id new_id = phi_graph_.FindOptimizedValue(phi_id); if (new_id != phi_id) { + VLOG(1) << "Replacing " << values[0]->ToString() << " with " + << GetValue(new_id).ToString(); value_set->Clear(); const HloValue& new_value = GetValue(new_id); value_set->AddValue(&new_value); diff --git a/tensorflow/compiler/xla/service/hlo_phi_graph.cc b/tensorflow/compiler/xla/service/hlo_phi_graph.cc index 9b69771dab2..a2cba3d1bff 100644 --- a/tensorflow/compiler/xla/service/hlo_phi_graph.cc +++ b/tensorflow/compiler/xla/service/hlo_phi_graph.cc @@ -20,10 +20,11 @@ limitations under the License. namespace xla { HloValue::Id PhiGraph::GetOptimizedId(const HloValue& value) { Node* node = value_id_to_node_[value.id()]; + CHECK(!node->mark_as_dead); return node->value_id; } -// Returns true if the input to a hlo value is the same as `inputs`. +// Returns true if the inputs to a hlo value are the same as `inputs`. bool PhiGraph::InputsEqualTo(const HloValue& value, absl::Span inputs) { auto iter = value_id_to_node_.find(value.id()); @@ -42,6 +43,7 @@ bool PhiGraph::InputsEqualTo(const HloValue& value, HloValue::Id PhiGraph::FindOptimizedValue(const HloValue::Id id) { auto iter = value_id_to_node_.find(id); CHECK(iter != value_id_to_node_.end()); + CHECK(!iter->second->mark_as_dead); return iter->second->value_id; } @@ -66,6 +68,17 @@ PhiGraph::Node* PhiGraph::CreateOrReuseNode(const HloValue& value) { void PhiGraph::ReplaceNodeWith(PhiGraph::Node* node, PhiGraph::Node* replace) { // Update users. CHECK(node->is_phi); + if (node->mark_as_dead) { + // The node has already been replaced with another. + return; + } + if (replace->mark_as_dead) { + // The node we are placing with has already been replaced with another node. + auto iter = value_id_to_node_.find(replace->value_id); + CHECK(iter != value_id_to_node_.end()); + return ReplaceNodeWith(node, iter->second); + } + CHECK(!replace->mark_as_dead); for (Node* user : node->users) { absl::c_replace(user->operands, node, replace); } @@ -74,6 +87,7 @@ void PhiGraph::ReplaceNodeWith(PhiGraph::Node* node, PhiGraph::Node* replace) { for (Node* operand : node->operands) { absl::c_replace(operand->users, node, replace); } + for (HloValue::Id value_id : node_to_value_id_[node]) { CHECK(value_id_to_node_.contains(value_id)); value_id_to_node_[value_id] = replace; @@ -115,6 +129,8 @@ std::string PhiGraph::ToString() { } void PhiGraph::Optimize() { + VLOG(2) << "Optimizing phi graph:"; + XLA_VLOG_LINES(2, ToString()); // Set up users for each node. for (auto& node : node_storage_) { for (Node* input : node->operands) { @@ -141,6 +157,8 @@ void PhiGraph::Optimize() { Node* node_ptr = node.get(); + VLOG(2) << "Optimizing: " << node_ptr->value_id; + CHECK_GE(node_ptr->operands.size(), 1); // Remove self-referencing ids from users and operands. @@ -167,6 +185,9 @@ void PhiGraph::Optimize() { [&](Node* elem) { return elem == node_ptr->operands[0]; }); if (all_inputs_are_same) { + VLOG(1) << "All inputs to node " << node_ptr->value_id + << " are the same, replacing it with " + << node_ptr->operands[0]->value_id; ReplaceNodeWith(node_ptr, node_ptr->operands[0]); changed = true; continue; @@ -223,6 +244,8 @@ void PhiGraph::Optimize() { CHECK_EQ(node, non_phi); continue; } + VLOG(1) << "Replace node " << node->value_id + << " in the closure with node " << non_phi->value_id; ReplaceNodeWith(node, non_phi); changed = true; } diff --git a/tensorflow/compiler/xla/service/hlo_phi_graph.h b/tensorflow/compiler/xla/service/hlo_phi_graph.h index a0eb994438e..ca0d5c5009c 100644 --- a/tensorflow/compiler/xla/service/hlo_phi_graph.h +++ b/tensorflow/compiler/xla/service/hlo_phi_graph.h @@ -90,7 +90,7 @@ class PhiGraph { // to that phi. absl::flat_hash_map> node_to_value_id_; - // A mapping between a HloValue and node in the phi graph. + // A mapping from a HloValue to node in the phi graph. absl::flat_hash_map value_id_to_node_; std::vector> node_storage_; }; diff --git a/tensorflow/compiler/xla/service/hlo_phi_graph_test.cc b/tensorflow/compiler/xla/service/hlo_phi_graph_test.cc index 41f0454fe55..ee7300b160b 100644 --- a/tensorflow/compiler/xla/service/hlo_phi_graph_test.cc +++ b/tensorflow/compiler/xla/service/hlo_phi_graph_test.cc @@ -82,5 +82,30 @@ TEST_F(PhiGraphTest, CircularPhi) { EXPECT_EQ(D.id(), phi_graph.FindOptimizedValue(C.id())); } +TEST_F(PhiGraphTest, NestedPhiReduction) { + // def A = phi(B, C) + // def B = phi(C, E) + // def C = phi(A, B) + // def D = non-phi + // def E = Phi(D, D) + // 1. Replace E with D + // 2. Replace A B and C with E/D + PhiGraph phi_graph; + HloValue A = NewHloValue(true); + HloValue B = NewHloValue(true); + HloValue C = NewHloValue(true); + HloValue D = NewHloValue(false); + HloValue E = NewHloValue(true); + phi_graph.RegisterPhi(A, {&B, &C}); + phi_graph.RegisterPhi(B, {&E, &C}); + phi_graph.RegisterPhi(C, {&A, &B}); + phi_graph.RegisterPhi(E, {&D, &D}); + phi_graph.Optimize(); + EXPECT_EQ(D.id(), phi_graph.FindOptimizedValue(A.id())); + EXPECT_EQ(D.id(), phi_graph.FindOptimizedValue(B.id())); + EXPECT_EQ(D.id(), phi_graph.FindOptimizedValue(C.id())); + EXPECT_EQ(D.id(), phi_graph.FindOptimizedValue(E.id())); +} + } // namespace } // namespace xla From 7ae27e344cf9a501c89d785cfe4eb109a9848b47 Mon Sep 17 00:00:00 2001 From: Tiezhen WANG Date: Wed, 17 Jun 2020 20:04:47 -0700 Subject: [PATCH 146/867] TFL: selective registration: Disable linkopts not available on Windows. PiperOrigin-RevId: 317019750 Change-Id: Ibce23c534a0b8fe15ada085a57aec6f87db71a9b --- tensorflow/lite/build_def.bzl | 8 ++++---- 1 file changed, 4 insertions(+), 4 deletions(-) diff --git a/tensorflow/lite/build_def.bzl b/tensorflow/lite/build_def.bzl index 5e487395355..e6c92691b15 100644 --- a/tensorflow/lite/build_def.bzl +++ b/tensorflow/lite/build_def.bzl @@ -768,10 +768,10 @@ def tflite_custom_cc_library(name, models = [], srcs = [], deps = [], visibility name = name, srcs = real_srcs, copts = tflite_copts(), - linkopts = [ - "-lm", - "-ldl", - ], + linkopts = select({ + "//tensorflow:windows": [], + "//conditions:default": ["-lm", "-ldl"], + }), deps = depset([ "//tensorflow/lite:framework", "//tensorflow/lite/kernels:builtin_ops", From 697a0a4fe6d792b57a280e2f22fd23fe219e9d81 Mon Sep 17 00:00:00 2001 From: Thomas Raoux Date: Wed, 17 Jun 2020 20:43:17 -0700 Subject: [PATCH 147/867] Internal change PiperOrigin-RevId: 317023916 Change-Id: I81311283fae6c28cdee2d7c87c8d062d32725bef --- third_party/mlir/BUILD | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/third_party/mlir/BUILD b/third_party/mlir/BUILD index 476b8566265..db75b27e78b 100644 --- a/third_party/mlir/BUILD +++ b/third_party/mlir/BUILD @@ -2868,9 +2868,9 @@ cc_library( ":SideEffects", ":StandardOps", ":Support", - "//third_party/vulkan_loader", "@llvm-project//llvm:Support", "@vulkan_headers", + "@vulkan_sdk//:sdk", ], ) From 4f341bb742718721563ce6dccb965c85a1fbdcf5 Mon Sep 17 00:00:00 2001 From: Srinivas Vasudevan Date: Wed, 17 Jun 2020 21:01:24 -0700 Subject: [PATCH 148/867] Add Bessel functions to the public api: - tf.math.special.bessel_i0 - tf.math.special.bessel_i0e - tf.math.special.bessel_i1 - tf.math.special.bessel_i1e - tf.math.special.bessel_k0 - tf.math.special.bessel_k0e - tf.math.special.bessel_k1 - tf.math.special.bessel_k1e - tf.math.special.bessel_j0 - tf.math.special.bessel_j1 - tf.math.special.bessel_y0 - tf.math.special.bessel_y1 PiperOrigin-RevId: 317025879 Change-Id: I5c4407eda6bef0d1659b7a566979c7dbbad4ad83 --- tensorflow/compiler/tf2xla/python/xla.py | 5 +- .../api_def/base_api/api_def_BesselI0.pbtxt | 4 + .../api_def/base_api/api_def_BesselI0e.pbtxt | 8 +- .../api_def/base_api/api_def_BesselI1.pbtxt | 4 + .../api_def/base_api/api_def_BesselI1e.pbtxt | 8 +- .../api_def/base_api/api_def_BesselJ0.pbtxt | 4 + .../api_def/base_api/api_def_BesselJ1.pbtxt | 4 + .../api_def/base_api/api_def_BesselK0.pbtxt | 4 + .../api_def/base_api/api_def_BesselK0e.pbtxt | 4 + .../api_def/base_api/api_def_BesselK1.pbtxt | 4 + .../api_def/base_api/api_def_BesselK1e.pbtxt | 4 + .../api_def/base_api/api_def_BesselY0.pbtxt | 4 + .../api_def/base_api/api_def_BesselY1.pbtxt | 4 + .../python_api/api_def_BesselI0e.pbtxt | 6 - .../python_api/api_def_BesselI1e.pbtxt | 6 - tensorflow/core/kernels/cwise_op_bessel.cc | 29 -- tensorflow/core/kernels/cwise_ops.h | 6 - .../special_math/special_math_op_bessel.cc | 78 +++++ .../special_math_op_gpu_bessel.cu.cc} | 16 +- .../special_math/special_math_op_misc_impl.h | 38 +++ tensorflow/core/ops/math_ops.cc | 4 - tensorflow/core/ops/special_math_ops.cc | 49 ++- .../optimization/map_vectorization_test.py | 6 +- .../eager/pywrap_gradient_exclusions.cc | 6 +- .../kernel_tests/cwise_ops_unary_test.py | 20 +- tensorflow/python/ops/math_grad.py | 131 ++++++++- .../python/ops/parallel_for/math_test.py | 4 +- tensorflow/python/ops/parallel_for/pfor.py | 14 +- tensorflow/python/ops/signal/window_ops.py | 5 +- tensorflow/python/ops/sparse_ops.py | 6 +- tensorflow/python/ops/special_math_ops.py | 278 +++++++++++++++++- .../python/ops/special_math_ops_test.py | 234 ++++++++++++++- .../golden/v1/tensorflow.math.special.pbtxt | 48 +++ .../api/golden/v1/tensorflow.raw_ops.pbtxt | 40 +++ .../golden/v2/tensorflow.math.special.pbtxt | 48 +++ .../api/golden/v2/tensorflow.raw_ops.pbtxt | 40 +++ 36 files changed, 1028 insertions(+), 145 deletions(-) create mode 100644 tensorflow/core/api_def/base_api/api_def_BesselI0.pbtxt create mode 100644 tensorflow/core/api_def/base_api/api_def_BesselI1.pbtxt create mode 100644 tensorflow/core/api_def/base_api/api_def_BesselJ0.pbtxt create mode 100644 tensorflow/core/api_def/base_api/api_def_BesselJ1.pbtxt create mode 100644 tensorflow/core/api_def/base_api/api_def_BesselK0.pbtxt create mode 100644 tensorflow/core/api_def/base_api/api_def_BesselK0e.pbtxt create mode 100644 tensorflow/core/api_def/base_api/api_def_BesselK1.pbtxt create mode 100644 tensorflow/core/api_def/base_api/api_def_BesselK1e.pbtxt create mode 100644 tensorflow/core/api_def/base_api/api_def_BesselY0.pbtxt create mode 100644 tensorflow/core/api_def/base_api/api_def_BesselY1.pbtxt delete mode 100644 tensorflow/core/api_def/python_api/api_def_BesselI0e.pbtxt delete mode 100644 tensorflow/core/api_def/python_api/api_def_BesselI1e.pbtxt delete mode 100644 tensorflow/core/kernels/cwise_op_bessel.cc create mode 100644 tensorflow/core/kernels/special_math/special_math_op_bessel.cc rename tensorflow/core/kernels/{cwise_op_bessel.cu.cc => special_math/special_math_op_gpu_bessel.cu.cc} (58%) diff --git a/tensorflow/compiler/tf2xla/python/xla.py b/tensorflow/compiler/tf2xla/python/xla.py index c59c47e92fb..0ebca2d546f 100644 --- a/tensorflow/compiler/tf2xla/python/xla.py +++ b/tensorflow/compiler/tf2xla/python/xla.py @@ -37,6 +37,7 @@ from tensorflow.python.ops import gen_math_ops from tensorflow.python.ops import gen_random_ops from tensorflow.python.ops import math_ops from tensorflow.python.ops import random_ops +from tensorflow.python.ops import special_math_ops # TODO(phawkins): provide wrappers for all XLA operators. Currently the missing # ops include: @@ -103,8 +104,8 @@ sign = _unary_op(math_ops.sign) tanh = _unary_op(math_ops.tanh) # Bessel -bessel_i0e = _unary_op(math_ops.bessel_i0e) -bessel_i1e = _unary_op(math_ops.bessel_i1e) +bessel_i0e = _unary_op(special_math_ops.bessel_i0e) +bessel_i1e = _unary_op(special_math_ops.bessel_i1e) # Binary operators diff --git a/tensorflow/core/api_def/base_api/api_def_BesselI0.pbtxt b/tensorflow/core/api_def/base_api/api_def_BesselI0.pbtxt new file mode 100644 index 00000000000..2c47960429c --- /dev/null +++ b/tensorflow/core/api_def/base_api/api_def_BesselI0.pbtxt @@ -0,0 +1,4 @@ +op { + graph_op_name: "BesselI0" + visibility: HIDDEN +} diff --git a/tensorflow/core/api_def/base_api/api_def_BesselI0e.pbtxt b/tensorflow/core/api_def/base_api/api_def_BesselI0e.pbtxt index 08313cebb99..7965af4916e 100644 --- a/tensorflow/core/api_def/base_api/api_def_BesselI0e.pbtxt +++ b/tensorflow/core/api_def/base_api/api_def_BesselI0e.pbtxt @@ -1,10 +1,4 @@ op { graph_op_name: "BesselI0e" - summary: "Computes the Bessel i0e function of `x` element-wise." - description: <
{ - StringRef getName() final { return "Table"; } -}; - } // namespace ResourceEffects } // namespace TF } // namespace mlir From 64e1b489bb63c4b7ceb85bfa846ca699036561e8 Mon Sep 17 00:00:00 2001 From: Terry Heo Date: Wed, 17 Jun 2020 23:54:57 -0700 Subject: [PATCH 161/867] Enable flex delegate on tensorflow.lite.Interpreter Python package Usually, flex delegate is enabled by symbol override of AcquireFlexDelegate() function. But this approach doesn't work well with shared library. Since pywrap_tensorflow_internal.so is available for tensorflow PIP, I've made the following changes to enable flex delegate. - Included flex delegate module to the pywrap_tensorflow_internal.so. This file already contains most TF internal logic and having TFLite flex delegate impacts about 72K to the output. - Added new function of TF_AcquireFlexDelegate() in the delegate module. - Updated logic in AcquireFlexDelegate() of interpreter_builder.cc to check the availability of pywrap_tensorflow_internal.so and lookup the TF_AcquireFlexDelegate() symbol to enable flex delegate. Also updated python/lite_flex_test.py since flex delegate is supported with Python API PiperOrigin-RevId: 317044994 Change-Id: Ic5e953f4a675b3f5360a4c7d607568193103711a --- tensorflow/lite/delegates/flex/delegate.cc | 7 +++ tensorflow/lite/interpreter_builder.cc | 17 ++++++ tensorflow/lite/python/BUILD | 5 +- tensorflow/lite/python/lite_flex_test.py | 61 +++++++++++++--------- tensorflow/python/BUILD | 7 ++- 5 files changed, 68 insertions(+), 29 deletions(-) diff --git a/tensorflow/lite/delegates/flex/delegate.cc b/tensorflow/lite/delegates/flex/delegate.cc index 4741bddc2f5..b8b0d4e6d01 100644 --- a/tensorflow/lite/delegates/flex/delegate.cc +++ b/tensorflow/lite/delegates/flex/delegate.cc @@ -136,3 +136,10 @@ TfLiteStatus FlexDelegate::CopyFromBufferHandle( } } // namespace tflite + +// Exported C interface function which is used by AcquireFlexDelegate() at +// interpreter_build.cc. To export the function name globally, the function name +// must be matched with patterns in tf_version_script.lds +extern "C" tflite::TfLiteDelegateUniquePtr TF_AcquireFlexDelegate() { + return tflite::AcquireFlexDelegate(); +} diff --git a/tensorflow/lite/interpreter_builder.cc b/tensorflow/lite/interpreter_builder.cc index 43d81ef0770..d73b298e595 100644 --- a/tensorflow/lite/interpreter_builder.cc +++ b/tensorflow/lite/interpreter_builder.cc @@ -14,6 +14,9 @@ limitations under the License. ==============================================================================*/ #include "tensorflow/lite/interpreter_builder.h" +#if !defined(__ANDROID__) && !defined(__APPLE__) && !defined(_WIN32) +#include +#endif #include #include #include @@ -114,6 +117,20 @@ const char* kEmptyTensorName = ""; // For flex delegate, see also the strong override in // lite/delegates/flex/delegate.cc. TFLITE_ATTRIBUTE_WEAK Interpreter::TfLiteDelegatePtr AcquireFlexDelegate() { +#if !defined(__ANDROID__) && !defined(__APPLE__) && !defined(_WIN32) + // If _pywrap_tensorflow_internal.so is available, use + // TF_AcquireFlexDelegate() to initialize flex delegate. + void* lib_tf_internal = + dlopen("_pywrap_tensorflow_internal.so", RTLD_NOW | RTLD_LOCAL); + if (lib_tf_internal) { + auto TF_AcquireFlexDelegate = + reinterpret_cast( + dlsym(lib_tf_internal, "TF_AcquireFlexDelegate")); + if (TF_AcquireFlexDelegate) { + return TF_AcquireFlexDelegate(); + } + } +#endif return Interpreter::TfLiteDelegatePtr(nullptr, [](TfLiteDelegate*) {}); } diff --git a/tensorflow/lite/python/BUILD b/tensorflow/lite/python/BUILD index d25e7d5ef8d..1b64b7d1042 100644 --- a/tensorflow/lite/python/BUILD +++ b/tensorflow/lite/python/BUILD @@ -193,9 +193,8 @@ py_test( python_version = "PY3", srcs_version = "PY2AND3", tags = [ - # TODO(b/111881877): Enable in oss after resolving op registry issues. - "no_oss", - "no_windows", + "no_mac", # TODO(b/159077703): Enable Python API Flex support on MacOS. + "no_windows", # TODO(b/159077703): Enable Python API Flex support on Windows. ], deps = [ ":lite", diff --git a/tensorflow/lite/python/lite_flex_test.py b/tensorflow/lite/python/lite_flex_test.py index 26bee206d27..ffc157c2128 100644 --- a/tensorflow/lite/python/lite_flex_test.py +++ b/tensorflow/lite/python/lite_flex_test.py @@ -19,6 +19,7 @@ from __future__ import division from __future__ import print_function from absl.testing import parameterized +import numpy as np from tensorflow.lite.python import lite from tensorflow.lite.python.interpreter import Interpreter @@ -41,8 +42,7 @@ class FromSessionTest(test_util.TensorFlowTestCase, parameterized.TestCase): ('DisableMlirConverter', False)) # disable mlir def testFlexMode(self, enable_mlir): with ops.Graph().as_default(): - in_tensor = array_ops.placeholder( - shape=[1, 16, 16, 3], dtype=dtypes.float32) + in_tensor = array_ops.placeholder(shape=[1, 4], dtype=dtypes.float32) out_tensor = in_tensor + in_tensor sess = session.Session() @@ -54,19 +54,22 @@ class FromSessionTest(test_util.TensorFlowTestCase, parameterized.TestCase): tflite_model = converter.convert() self.assertTrue(tflite_model) - # Ensures the model contains TensorFlow ops. - # TODO(nupurgarg): Check values once there is a Python delegate interface. + # Check the model works with TensorFlow ops. interpreter = Interpreter(model_content=tflite_model) - with self.assertRaises(RuntimeError) as error: - interpreter.allocate_tensors() - self.assertIn( - 'Regular TensorFlow ops are not supported by this interpreter.', - str(error.exception)) + interpreter.allocate_tensors() + input_details = interpreter.get_input_details() + test_input = np.array([[1.0, 2.0, 3.0, 4.0]], dtype=np.float32) + interpreter.set_tensor(input_details[0]['index'], test_input) + interpreter.invoke() + + output_details = interpreter.get_output_details() + expected_output = np.array([[2.0, 4.0, 6.0, 8.0]], dtype=np.float32) + output_data = interpreter.get_tensor(output_details[0]['index']) + self.assertTrue((expected_output == output_data).all()) def testDeprecatedFlags(self): with ops.Graph().as_default(): - in_tensor = array_ops.placeholder( - shape=[1, 16, 16, 3], dtype=dtypes.float32) + in_tensor = array_ops.placeholder(shape=[1, 4], dtype=dtypes.float32) out_tensor = in_tensor + in_tensor sess = session.Session() @@ -83,14 +86,18 @@ class FromSessionTest(test_util.TensorFlowTestCase, parameterized.TestCase): tflite_model = converter.convert() self.assertTrue(tflite_model) - # Ensures the model contains TensorFlow ops. - # TODO(nupurgarg): Check values once there is a Python delegate interface. + # Check the model works with TensorFlow ops. interpreter = Interpreter(model_content=tflite_model) - with self.assertRaises(RuntimeError) as error: - interpreter.allocate_tensors() - self.assertIn( - 'Regular TensorFlow ops are not supported by this interpreter.', - str(error.exception)) + interpreter.allocate_tensors() + input_details = interpreter.get_input_details() + test_input = np.array([[1.0, 2.0, 3.0, 4.0]], dtype=np.float32) + interpreter.set_tensor(input_details[0]['index'], test_input) + interpreter.invoke() + + output_details = interpreter.get_output_details() + expected_output = np.array([[2.0, 4.0, 6.0, 8.0]], dtype=np.float32) + output_data = interpreter.get_tensor(output_details[0]['index']) + self.assertTrue((expected_output == output_data).all()) class FromConcreteFunctionTest(test_util.TensorFlowTestCase, @@ -114,14 +121,18 @@ class FromConcreteFunctionTest(test_util.TensorFlowTestCase, converter.experimental_new_converter = enable_mlir tflite_model = converter.convert() - # Ensures the model contains TensorFlow ops. - # TODO(nupurgarg): Check values once there is a Python delegate interface. + # Check the model works with TensorFlow ops. interpreter = Interpreter(model_content=tflite_model) - with self.assertRaises(RuntimeError) as error: - interpreter.allocate_tensors() - self.assertIn( - 'Regular TensorFlow ops are not supported by this interpreter.', - str(error.exception)) + interpreter.allocate_tensors() + input_details = interpreter.get_input_details() + test_input = np.array([4.0], dtype=np.float32) + interpreter.set_tensor(input_details[0]['index'], test_input) + interpreter.invoke() + + output_details = interpreter.get_output_details() + expected_output = np.array([24.0], dtype=np.float32) + output_data = interpreter.get_tensor(output_details[0]['index']) + self.assertTrue((expected_output == output_data).all()) if __name__ == '__main__': diff --git a/tensorflow/python/BUILD b/tensorflow/python/BUILD index d141b719aef..f53859b2915 100644 --- a/tensorflow/python/BUILD +++ b/tensorflow/python/BUILD @@ -6058,7 +6058,12 @@ pywrap_tensorflow_macro( "@ngraph_tf//:ngraph_tf", ]) + if_xla_available([ "//tensorflow/compiler/aot:tfcompile_lib", - ]), + ]) + select({ + "//tensorflow:windows": [], # TODO(b/159077703): Enable Flex on Windows + "//conditions:default": [ + "//tensorflow/lite/delegates/flex:delegate", + ], + }), ) # ** Targets for Windows build (start) ** From fabcd8f89cd5975331994049705e15cb75f32e0c Mon Sep 17 00:00:00 2001 From: Marat Dukhan Date: Thu, 18 Jun 2020 00:06:13 -0700 Subject: [PATCH 162/867] Update XNNPACK dependency Bring in fix for x86 builds on Android NDK r20 PiperOrigin-RevId: 317046250 Change-Id: I493cf294ea4e51c91a68b9bc8b062f6cecf4da7f --- tensorflow/workspace.bzl | 8 ++++---- 1 file changed, 4 insertions(+), 4 deletions(-) diff --git a/tensorflow/workspace.bzl b/tensorflow/workspace.bzl index fab6faf62d1..78f7e0ce03e 100755 --- a/tensorflow/workspace.bzl +++ b/tensorflow/workspace.bzl @@ -164,11 +164,11 @@ def tf_repositories(path_prefix = "", tf_repo_name = ""): tf_http_archive( name = "XNNPACK", - sha256 = "dfcc7b2894c5c6bc570d65ff22b371a28e0fcc672e75705cc3f1ccc1264c3f8b", - strip_prefix = "XNNPACK-cac25227b5d8046170f875ad80545696be908ee7", + sha256 = "4af883fea0a6ada106867f29670a6c0b7af74bee85d74a2e04356a670814a3d4", + strip_prefix = "XNNPACK-69a6a7667d96a84c596b0f4e00632b2037c17723", urls = [ - "https://storage.googleapis.com/mirror.tensorflow.org/github.com/google/XNNPACK/archive/cac25227b5d8046170f875ad80545696be908ee7.zip", - "https://github.com/google/XNNPACK/archive/cac25227b5d8046170f875ad80545696be908ee7.zip", + "https://storage.googleapis.com/mirror.tensorflow.org/github.com/google/XNNPACK/archive/69a6a7667d96a84c596b0f4e00632b2037c17723.zip", + "https://github.com/google/XNNPACK/archive/69a6a7667d96a84c596b0f4e00632b2037c17723.zip", ], ) From b38d5d1889b398ca45aa7f2e4f9f0184c77c6c55 Mon Sep 17 00:00:00 2001 From: Yunxing Dai Date: Thu, 18 Jun 2020 00:52:32 -0700 Subject: [PATCH 163/867] Remove dynamic dimension of strided slice grad if input to strided slice is static. If we slice a dynamic shaped tensor from a static tensor, the output of the gradient should still be static. Unfortunately this cannot be deduced alone by xla, so extra information is needed from the tf2xla bridge. PiperOrigin-RevId: 317051543 Change-Id: I7a8113c47a4aed145dfba7f7d12992ca35a9cf19 --- .../tf2xla/kernels/strided_slice_op.cc | 22 ++++++++++++++++ tensorflow/compiler/xla/client/xla_builder.cc | 23 ++++++++++++++++ tensorflow/compiler/xla/client/xla_builder.h | 6 +++++ .../compiler/xla/client/xla_builder_test.cc | 26 +++++++++++++++++++ 4 files changed, 77 insertions(+) diff --git a/tensorflow/compiler/tf2xla/kernels/strided_slice_op.cc b/tensorflow/compiler/tf2xla/kernels/strided_slice_op.cc index 2684c982600..51764018df1 100644 --- a/tensorflow/compiler/tf2xla/kernels/strided_slice_op.cc +++ b/tensorflow/compiler/tf2xla/kernels/strided_slice_op.cc @@ -350,6 +350,28 @@ class StridedSliceGradOp : public XlaOpKernel { grad = xla::Rev(grad, dimensions_to_reverse); } grad = xla::Pad(grad, zero, padding_config); + + xla::XlaOp dynamic_shape = ctx->Input(0); + xla::Shape grad_shape = ctx->builder()->GetShape(grad).ValueOrDie(); + ctx->set_dynamic_dimension_is_minus_one(true); + std::vector dynamic_size; + OP_REQUIRES_OK(ctx, ctx->ConstantInputAsIntVector(0, &dynamic_size)); + // Input of strided_slice_op has to have the same shape as output. + DCHECK_EQ(grad_shape.rank(), input_shape.dims()); + for (int64 dim = 0; dim < input_shape.dims(); ++dim) { + DCHECK_EQ(grad_shape.dimensions(dim), input_shape.dim_size(dim)); + if (dynamic_size[dim] == -1) { + // Input is a dynamic dimension, set the same dynamic dimension size in + // the output. + auto dim_size = xla::Slice(dynamic_shape, {dim}, {dim + 1}, {1}); + grad = xla::SetDimensionSize(grad, dim_size, dim); + } else if (grad_shape.is_dynamic_dimension(dim)) { + // Input is static but output is dynamic, respect input and remove any + // dynamic dim in the output. + grad = xla::RemoveDynamicDimension(grad, dim); + } + } + ctx->SetOutput(0, grad); } diff --git a/tensorflow/compiler/xla/client/xla_builder.cc b/tensorflow/compiler/xla/client/xla_builder.cc index bfba48862f6..c7b6a7f9491 100644 --- a/tensorflow/compiler/xla/client/xla_builder.cc +++ b/tensorflow/compiler/xla/client/xla_builder.cc @@ -2727,6 +2727,25 @@ XlaOp XlaBuilder::GetDimensionSize(XlaOp operand, int64 dimension) { }); } +XlaOp XlaBuilder::RemoveDynamicDimension(XlaOp operand, int64 dimension) { + return ReportErrorOrReturn([&]() -> StatusOr { + HloInstructionProto instr; + TF_ASSIGN_OR_RETURN(const Shape* operand_shape, GetShapePtr(operand)); + + Shape shape = *operand_shape; + shape.set_dynamic_dimension(dimension, false); + // Setting an op's dynamic dimension to its static size removes the dynamic + // dimension. + XlaOp static_size = + ConstantR0(this, operand_shape->dimensions(dimension)); + + *instr.mutable_shape() = shape.ToProto(); + instr.add_dimensions(dimension); + return AddInstruction(std::move(instr), HloOpcode::kSetDimensionSize, + {operand, static_size}); + }); +} + XlaOp XlaBuilder::SetDimensionSize(XlaOp operand, XlaOp val, int64 dimension) { return ReportErrorOrReturn([&]() -> StatusOr { HloInstructionProto instr; @@ -3827,4 +3846,8 @@ XlaOp SetDimensionSize(const XlaOp operand, const XlaOp val, int64 dimension) { return operand.builder()->SetDimensionSize(operand, val, dimension); } +XlaOp RemoveDynamicDimension(const XlaOp operand, int64 dimension) { + return operand.builder()->RemoveDynamicDimension(operand, dimension); +} + } // namespace xla diff --git a/tensorflow/compiler/xla/client/xla_builder.h b/tensorflow/compiler/xla/client/xla_builder.h index ffa6a7c3439..b8af180b83e 100644 --- a/tensorflow/compiler/xla/client/xla_builder.h +++ b/tensorflow/compiler/xla/client/xla_builder.h @@ -704,6 +704,8 @@ class XlaBuilder { XlaOp SetDimensionSize(XlaOp operand, XlaOp val, int64 dimension); + XlaOp RemoveDynamicDimension(XlaOp operand, int64 dimension); + StatusOr AddInstruction(HloInstructionProto&& instr, HloOpcode opcode, absl::Span operands = {}); @@ -1151,6 +1153,7 @@ class XlaBuilder { friend XlaOp GetDimensionSize(XlaOp operand, int64 dimension); friend XlaOp SetDimensionSize(XlaOp operand, XlaOp val, int64 dimension); + friend XlaOp RemoveDynamicDimension(XlaOp operand, int64 dimension); protected: // Returns OK status if the given op was built using this builder. Otherwise, @@ -2149,6 +2152,9 @@ XlaOp GetDimensionSize(XlaOp operand, int64 dimension); XlaOp SetDimensionSize(XlaOp operand, XlaOp val, int64 dimension); +// Returns the same op but with dynamic dimension removed. +XlaOp RemoveDynamicDimension(XlaOp operand, int64 dimension); + // Implementation details below this point. // diff --git a/tensorflow/compiler/xla/client/xla_builder_test.cc b/tensorflow/compiler/xla/client/xla_builder_test.cc index 4fa47077fca..7011c946203 100644 --- a/tensorflow/compiler/xla/client/xla_builder_test.cc +++ b/tensorflow/compiler/xla/client/xla_builder_test.cc @@ -556,6 +556,32 @@ TEST_F(XlaBuilderTest, DynamicParameter) { EXPECT_TRUE(param_shape.is_dynamic_dimension(0)); } +TEST_F(XlaBuilderTest, SetDimensionSize) { + XlaBuilder b(TestName()); + auto p0 = Parameter(&b, 0, ShapeUtil::MakeShape(F32, {10}), "p0"); + auto p1 = Parameter(&b, 1, ShapeUtil::MakeShape(S32, {}), "p1"); + auto set_dim_size = SetDimensionSize(p0, p1, 0); + TF_ASSERT_OK_AND_ASSIGN(auto module, + BuildHloModule(&b, /*root=*/set_dim_size)); + const Shape& root_shape = + module->entry_computation()->root_instruction()->shape(); + EXPECT_TRUE(root_shape.is_dynamic_dimension(0)); +} + +TEST_F(XlaBuilderTest, RemoveDimensionSize) { + XlaBuilder b(TestName()); + auto p0 = Parameter(&b, 0, ShapeUtil::MakeShape(F32, {10}), "p0"); + auto p1 = Parameter(&b, 1, ShapeUtil::MakeShape(S32, {}), "p1"); + auto set_dim_size = SetDimensionSize(p0, p1, 0); + auto remove_dim_size = RemoveDynamicDimension(set_dim_size, 0); + TF_ASSERT_OK_AND_ASSIGN(auto module, + BuildHloModule(&b, /*root=*/remove_dim_size)); + const Shape& root_shape = + module->entry_computation()->root_instruction()->shape(); + // Dynamic dimension has been removed. + EXPECT_FALSE(root_shape.is_dynamic_dimension(0)); +} + TEST_F(XlaBuilderTest, DynamicUnary) { XlaBuilder b(TestName()); Shape tuple_param_shape = ShapeUtil::MakeTupleShape( From e4af590df8ff7685de1679f2a6eb01a677d84775 Mon Sep 17 00:00:00 2001 From: Adrian Kuegel Date: Thu, 18 Jun 2020 01:35:40 -0700 Subject: [PATCH 164/867] Properly configure block and grid dimensions when launching generated kernels. Also prepare the possibility to specify unrolling. This is not enabled yet because there are some LLVM changes required. PiperOrigin-RevId: 317056534 Change-Id: I3de5dda52d80b528c4bd0026a5e160fda4296c32 --- .../mlir_generated_cwise_op_gpu_tanh.cu.cc | 71 ++++++++++++++----- 1 file changed, 53 insertions(+), 18 deletions(-) diff --git a/tensorflow/core/kernels/mlir_generated_cwise_op_gpu_tanh.cu.cc b/tensorflow/core/kernels/mlir_generated_cwise_op_gpu_tanh.cu.cc index 40dd7c7e49e..70de777239f 100644 --- a/tensorflow/core/kernels/mlir_generated_cwise_op_gpu_tanh.cu.cc +++ b/tensorflow/core/kernels/mlir_generated_cwise_op_gpu_tanh.cu.cc @@ -14,6 +14,7 @@ limitations under the License. ==============================================================================*/ #include +#include #include "absl/strings/string_view.h" #include "absl/types/span.h" @@ -45,9 +46,41 @@ Status CreateKernel(absl::string_view kernel_name, uint64_t num_args, return stream_exec->GetKernel(loader_spec, kernel_base.get()); } -class MlirGenerateTanhOp : public OpKernel { +struct LaunchConfig { + se::BlockDim blockDim; + se::ThreadDim threadDim; +}; + +LaunchConfig GetLaunchConfiguration(std::vector tile_sizes, + std::vector unrolling_factors, + std::vector shape) { + LaunchConfig result; + // Ensure the vectors are length 3 and pad with ones. + tile_sizes.resize(3, 1); + unrolling_factors.resize(3, 1); + shape.resize(3, 1); + // The number of threads is given by the tiling size. + result.threadDim = se::ThreadDim(tile_sizes[0], tile_sizes[1], tile_sizes[2]); + // We know that the kernel was generated by mapping the three outer-most + // dimensions to x,y,z dimensions. So we only need to compute those. + std::vector block_dims(3); + for (int i = 0; i < 3; ++i) { + // Compute the number of grids. We use ceildiv here as we have to allocate + // an extra thread/block if the division is not even. The kernel contains + // code to handle the boundaries. + int number_of_threads = + (shape[i] + unrolling_factors[i] - 1) / unrolling_factors[i]; + int number_of_grids = + (number_of_threads + tile_sizes[i] - 1) / tile_sizes[i]; + block_dims[i] = number_of_grids; + } + result.blockDim = se::BlockDim(block_dims[0], block_dims[1], block_dims[2]); + return result; +} + +class MlirGeneratedTanhOp : public OpKernel { public: - explicit MlirGenerateTanhOp(OpKernelConstruction* ctx) : OpKernel(ctx) {} + explicit MlirGeneratedTanhOp(OpKernelConstruction* ctx) : OpKernel(ctx) {} void Compute(OpKernelContext* ctx) override { auto* stream = ctx->op_device_context()->stream(); @@ -88,11 +121,13 @@ class MlirGenerateTanhOp : public OpKernel { args.add_argument(inp.NumElements()); args.add_argument(1); - // TODO(b/158649746): Choose block size and thread dim according to the - // number of input elements. For now, this supports at most 1024 elements. + // This has to be aligned with the configuration that was used when building + // the kernels. See the corresponding build rules in `cubin_headers/BUILD`. + LaunchConfig config = GetLaunchConfiguration( + {256}, {}, {static_cast(inp.NumElements())}); OP_REQUIRES_OK( - ctx, stream->parent()->Launch(stream, se::ThreadDim(inp.NumElements()), - se::BlockDim(1), *kernel, args)); + ctx, stream->parent()->Launch(stream, config.threadDim, config.blockDim, + *kernel, args)); } protected: @@ -103,26 +138,26 @@ class MlirGenerateTanhOp : public OpKernel { std::mutex mu_; }; -class MlirGenerateTanhF16Op : public MlirGenerateTanhOp { +class MlirGeneratedTanhF16Op : public MlirGeneratedTanhOp { public: - explicit MlirGenerateTanhF16Op(OpKernelConstruction* ctx) - : MlirGenerateTanhOp(ctx) { + explicit MlirGeneratedTanhF16Op(OpKernelConstruction* ctx) + : MlirGeneratedTanhOp(ctx) { cubin_data_ = kTanhF16Kernel; } }; -class MlirGenerateTanhF32Op : public MlirGenerateTanhOp { +class MlirGeneratedTanhF32Op : public MlirGeneratedTanhOp { public: - explicit MlirGenerateTanhF32Op(OpKernelConstruction* ctx) - : MlirGenerateTanhOp(ctx) { + explicit MlirGeneratedTanhF32Op(OpKernelConstruction* ctx) + : MlirGeneratedTanhOp(ctx) { cubin_data_ = kTanhF32Kernel; } }; -class MlirGenerateTanhF64Op : public MlirGenerateTanhOp { +class MlirGeneratedTanhF64Op : public MlirGeneratedTanhOp { public: - explicit MlirGenerateTanhF64Op(OpKernelConstruction* ctx) - : MlirGenerateTanhOp(ctx) { + explicit MlirGeneratedTanhF64Op(OpKernelConstruction* ctx) + : MlirGeneratedTanhOp(ctx) { cubin_data_ = kTanhF64Kernel; } }; @@ -130,11 +165,11 @@ class MlirGenerateTanhF64Op : public MlirGenerateTanhOp { REGISTER_KERNEL_BUILDER( Name("Tanh").Device(DEVICE_GPU).TypeConstraint("T"), - MlirGenerateTanhF16Op); + MlirGeneratedTanhF16Op); REGISTER_KERNEL_BUILDER( Name("Tanh").Device(DEVICE_GPU).TypeConstraint("T"), - MlirGenerateTanhF32Op); + MlirGeneratedTanhF32Op); REGISTER_KERNEL_BUILDER( Name("Tanh").Device(DEVICE_GPU).TypeConstraint("T"), - MlirGenerateTanhF64Op); + MlirGeneratedTanhF64Op); } // namespace tensorflow From ad1434444bda1e8321dcc09965a5dce8da847eed Mon Sep 17 00:00:00 2001 From: "A. Unique TensorFlower" Date: Thu, 18 Jun 2020 01:51:06 -0700 Subject: [PATCH 165/867] Remove dynamic dimension of strided slice grad if input to strided slice is static. If we slice a dynamic shaped tensor from a static tensor, the output of the gradient should still be static. Unfortunately this cannot be deduced alone by xla, so extra information is needed from the tf2xla bridge. PiperOrigin-RevId: 317058146 Change-Id: I33e4895e169c238ad3d73a57ada11c4984d11dfb --- .../tf2xla/kernels/strided_slice_op.cc | 22 ---------------- tensorflow/compiler/xla/client/xla_builder.cc | 23 ---------------- tensorflow/compiler/xla/client/xla_builder.h | 6 ----- .../compiler/xla/client/xla_builder_test.cc | 26 ------------------- 4 files changed, 77 deletions(-) diff --git a/tensorflow/compiler/tf2xla/kernels/strided_slice_op.cc b/tensorflow/compiler/tf2xla/kernels/strided_slice_op.cc index 51764018df1..2684c982600 100644 --- a/tensorflow/compiler/tf2xla/kernels/strided_slice_op.cc +++ b/tensorflow/compiler/tf2xla/kernels/strided_slice_op.cc @@ -350,28 +350,6 @@ class StridedSliceGradOp : public XlaOpKernel { grad = xla::Rev(grad, dimensions_to_reverse); } grad = xla::Pad(grad, zero, padding_config); - - xla::XlaOp dynamic_shape = ctx->Input(0); - xla::Shape grad_shape = ctx->builder()->GetShape(grad).ValueOrDie(); - ctx->set_dynamic_dimension_is_minus_one(true); - std::vector dynamic_size; - OP_REQUIRES_OK(ctx, ctx->ConstantInputAsIntVector(0, &dynamic_size)); - // Input of strided_slice_op has to have the same shape as output. - DCHECK_EQ(grad_shape.rank(), input_shape.dims()); - for (int64 dim = 0; dim < input_shape.dims(); ++dim) { - DCHECK_EQ(grad_shape.dimensions(dim), input_shape.dim_size(dim)); - if (dynamic_size[dim] == -1) { - // Input is a dynamic dimension, set the same dynamic dimension size in - // the output. - auto dim_size = xla::Slice(dynamic_shape, {dim}, {dim + 1}, {1}); - grad = xla::SetDimensionSize(grad, dim_size, dim); - } else if (grad_shape.is_dynamic_dimension(dim)) { - // Input is static but output is dynamic, respect input and remove any - // dynamic dim in the output. - grad = xla::RemoveDynamicDimension(grad, dim); - } - } - ctx->SetOutput(0, grad); } diff --git a/tensorflow/compiler/xla/client/xla_builder.cc b/tensorflow/compiler/xla/client/xla_builder.cc index c7b6a7f9491..bfba48862f6 100644 --- a/tensorflow/compiler/xla/client/xla_builder.cc +++ b/tensorflow/compiler/xla/client/xla_builder.cc @@ -2727,25 +2727,6 @@ XlaOp XlaBuilder::GetDimensionSize(XlaOp operand, int64 dimension) { }); } -XlaOp XlaBuilder::RemoveDynamicDimension(XlaOp operand, int64 dimension) { - return ReportErrorOrReturn([&]() -> StatusOr { - HloInstructionProto instr; - TF_ASSIGN_OR_RETURN(const Shape* operand_shape, GetShapePtr(operand)); - - Shape shape = *operand_shape; - shape.set_dynamic_dimension(dimension, false); - // Setting an op's dynamic dimension to its static size removes the dynamic - // dimension. - XlaOp static_size = - ConstantR0(this, operand_shape->dimensions(dimension)); - - *instr.mutable_shape() = shape.ToProto(); - instr.add_dimensions(dimension); - return AddInstruction(std::move(instr), HloOpcode::kSetDimensionSize, - {operand, static_size}); - }); -} - XlaOp XlaBuilder::SetDimensionSize(XlaOp operand, XlaOp val, int64 dimension) { return ReportErrorOrReturn([&]() -> StatusOr { HloInstructionProto instr; @@ -3846,8 +3827,4 @@ XlaOp SetDimensionSize(const XlaOp operand, const XlaOp val, int64 dimension) { return operand.builder()->SetDimensionSize(operand, val, dimension); } -XlaOp RemoveDynamicDimension(const XlaOp operand, int64 dimension) { - return operand.builder()->RemoveDynamicDimension(operand, dimension); -} - } // namespace xla diff --git a/tensorflow/compiler/xla/client/xla_builder.h b/tensorflow/compiler/xla/client/xla_builder.h index b8af180b83e..ffa6a7c3439 100644 --- a/tensorflow/compiler/xla/client/xla_builder.h +++ b/tensorflow/compiler/xla/client/xla_builder.h @@ -704,8 +704,6 @@ class XlaBuilder { XlaOp SetDimensionSize(XlaOp operand, XlaOp val, int64 dimension); - XlaOp RemoveDynamicDimension(XlaOp operand, int64 dimension); - StatusOr AddInstruction(HloInstructionProto&& instr, HloOpcode opcode, absl::Span operands = {}); @@ -1153,7 +1151,6 @@ class XlaBuilder { friend XlaOp GetDimensionSize(XlaOp operand, int64 dimension); friend XlaOp SetDimensionSize(XlaOp operand, XlaOp val, int64 dimension); - friend XlaOp RemoveDynamicDimension(XlaOp operand, int64 dimension); protected: // Returns OK status if the given op was built using this builder. Otherwise, @@ -2152,9 +2149,6 @@ XlaOp GetDimensionSize(XlaOp operand, int64 dimension); XlaOp SetDimensionSize(XlaOp operand, XlaOp val, int64 dimension); -// Returns the same op but with dynamic dimension removed. -XlaOp RemoveDynamicDimension(XlaOp operand, int64 dimension); - // Implementation details below this point. // diff --git a/tensorflow/compiler/xla/client/xla_builder_test.cc b/tensorflow/compiler/xla/client/xla_builder_test.cc index 7011c946203..4fa47077fca 100644 --- a/tensorflow/compiler/xla/client/xla_builder_test.cc +++ b/tensorflow/compiler/xla/client/xla_builder_test.cc @@ -556,32 +556,6 @@ TEST_F(XlaBuilderTest, DynamicParameter) { EXPECT_TRUE(param_shape.is_dynamic_dimension(0)); } -TEST_F(XlaBuilderTest, SetDimensionSize) { - XlaBuilder b(TestName()); - auto p0 = Parameter(&b, 0, ShapeUtil::MakeShape(F32, {10}), "p0"); - auto p1 = Parameter(&b, 1, ShapeUtil::MakeShape(S32, {}), "p1"); - auto set_dim_size = SetDimensionSize(p0, p1, 0); - TF_ASSERT_OK_AND_ASSIGN(auto module, - BuildHloModule(&b, /*root=*/set_dim_size)); - const Shape& root_shape = - module->entry_computation()->root_instruction()->shape(); - EXPECT_TRUE(root_shape.is_dynamic_dimension(0)); -} - -TEST_F(XlaBuilderTest, RemoveDimensionSize) { - XlaBuilder b(TestName()); - auto p0 = Parameter(&b, 0, ShapeUtil::MakeShape(F32, {10}), "p0"); - auto p1 = Parameter(&b, 1, ShapeUtil::MakeShape(S32, {}), "p1"); - auto set_dim_size = SetDimensionSize(p0, p1, 0); - auto remove_dim_size = RemoveDynamicDimension(set_dim_size, 0); - TF_ASSERT_OK_AND_ASSIGN(auto module, - BuildHloModule(&b, /*root=*/remove_dim_size)); - const Shape& root_shape = - module->entry_computation()->root_instruction()->shape(); - // Dynamic dimension has been removed. - EXPECT_FALSE(root_shape.is_dynamic_dimension(0)); -} - TEST_F(XlaBuilderTest, DynamicUnary) { XlaBuilder b(TestName()); Shape tuple_param_shape = ShapeUtil::MakeTupleShape( From 50f0fcc0005d3afec88b281a07563a603dbe0c7c Mon Sep 17 00:00:00 2001 From: "A. Unique TensorFlower" Date: Thu, 18 Jun 2020 02:01:45 -0700 Subject: [PATCH 166/867] compat: Update forward compatibility horizon to 2020-06-18 PiperOrigin-RevId: 317059247 Change-Id: I61e0e6659103a5b413a0013fb0782ba0d470d285 --- tensorflow/python/compat/compat.py | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/tensorflow/python/compat/compat.py b/tensorflow/python/compat/compat.py index 8a42b3dfdd3..32545ac8463 100644 --- a/tensorflow/python/compat/compat.py +++ b/tensorflow/python/compat/compat.py @@ -33,7 +33,7 @@ from tensorflow.python.util.tf_export import tf_export # This value changes every day with an automatic CL. It can be modified in code # via `forward_compatibility_horizon()` or with the environment variable # TF_FORWARD_COMPATIBILITY_DELTA_DAYS, which is added to the compatibility date. -_FORWARD_COMPATIBILITY_HORIZON = datetime.date(2020, 6, 17) +_FORWARD_COMPATIBILITY_HORIZON = datetime.date(2020, 6, 18) _FORWARD_COMPATIBILITY_DELTA_DAYS_VAR_NAME = "TF_FORWARD_COMPATIBILITY_DELTA_DAYS" _FORWARD_COMPATIBILITY_DATE_NUMBER = None From 39c6a3b2cd0193eae45a051344921a3c18a2e691 Mon Sep 17 00:00:00 2001 From: "A. Unique TensorFlower" Date: Thu, 18 Jun 2020 02:01:54 -0700 Subject: [PATCH 167/867] Update GraphDef version to 436. PiperOrigin-RevId: 317059270 Change-Id: I2eb1760fec06cae5cf58c41effa3263a0bdde142 --- tensorflow/core/public/version.h | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/tensorflow/core/public/version.h b/tensorflow/core/public/version.h index 3e4e3888d87..546d86e58fa 100644 --- a/tensorflow/core/public/version.h +++ b/tensorflow/core/public/version.h @@ -108,7 +108,7 @@ limitations under the License. #define TF_GRAPH_DEF_VERSION_MIN_PRODUCER 0 #define TF_GRAPH_DEF_VERSION_MIN_CONSUMER 0 -#define TF_GRAPH_DEF_VERSION 435 // Updated: 2020/6/17 +#define TF_GRAPH_DEF_VERSION 436 // Updated: 2020/6/18 // Checkpoint compatibility versions (the versions field in SavedSliceMeta). // From dd49e65c5b68c4b8113dfe5aadb988fdeb2abd57 Mon Sep 17 00:00:00 2001 From: Benjamin Kramer Date: Thu, 18 Jun 2020 02:38:26 -0700 Subject: [PATCH 168/867] [XLA:CPU] Fusion: Only check for reuse on expensive instructions The reuse condition walks over all instructions in the fusion, the fusion pass walks over all instructions, making this essentially quadratic. Moving the is_expensive check up doesn't completely avoid this behavior, but makes it much more unlikely. PiperOrigin-RevId: 317063582 Change-Id: I22459aa922e6d65c6c639ed81208d1d441a132bc --- .../compiler/xla/service/cpu/cpu_instruction_fusion.cc | 5 ++--- 1 file changed, 2 insertions(+), 3 deletions(-) diff --git a/tensorflow/compiler/xla/service/cpu/cpu_instruction_fusion.cc b/tensorflow/compiler/xla/service/cpu/cpu_instruction_fusion.cc index 97e0a518499..9460cc55e10 100644 --- a/tensorflow/compiler/xla/service/cpu/cpu_instruction_fusion.cc +++ b/tensorflow/compiler/xla/service/cpu/cpu_instruction_fusion.cc @@ -94,9 +94,8 @@ bool CpuInstructionFusion::ShouldFuse(HloInstruction* consumer, // Cost condition: not fuse (simple, expensive producers) and (consumers who // reuse operand elements). - if (producer->opcode() != HloOpcode::kFusion && - consumer->ReusesOperandElements(operand_index) && - is_expensive(*producer)) { + if (producer->opcode() != HloOpcode::kFusion && is_expensive(*producer) && + consumer->ReusesOperandElements(operand_index)) { VLOG(2) << "Fusion is not profitable."; return false; } From c21328bd992fe359b585452df92fc82976d27557 Mon Sep 17 00:00:00 2001 From: Lukas Geiger Date: Thu, 18 Jun 2020 12:06:40 +0200 Subject: [PATCH 169/867] Don't skip distributed test case in eager mode --- .../mixed_precision/experimental/autocast_variable_test.py | 4 ++-- 1 file changed, 2 insertions(+), 2 deletions(-) diff --git a/tensorflow/python/keras/mixed_precision/experimental/autocast_variable_test.py b/tensorflow/python/keras/mixed_precision/experimental/autocast_variable_test.py index 940bd07c813..2fa7c103258 100644 --- a/tensorflow/python/keras/mixed_precision/experimental/autocast_variable_test.py +++ b/tensorflow/python/keras/mixed_precision/experimental/autocast_variable_test.py @@ -304,8 +304,8 @@ class AutoCastVariableTest(test.TestCase, parameterized.TestCase): self.assertAllClose(3., self.evaluate(x.assign_sub(3.))) # Assign multiple times - # This currently doesn't work in graph mode - if context.executing_eagerly() or ops.inside_function(): + # This currently doesn't work in graph mode if a strategy is used + if not ds_context.has_strategy() or context.executing_eagerly(): assign = x.assign(1.) self.assertAllClose(1., self.evaluate(assign)) self.assertAllClose(0., self.evaluate(assign.assign(0.))) From de5620b74cb1688dcc81f3e50ad8fc24b18beb9b Mon Sep 17 00:00:00 2001 From: Stephan Herhut Date: Thu, 18 Jun 2020 04:33:29 -0700 Subject: [PATCH 170/867] Add option in kernel_lowering to use the tanh approximation. PiperOrigin-RevId: 317077338 Change-Id: I5cc7bfb84d6defba05439186377f90e422247d68 --- tensorflow/compiler/xla/service/mlir_gpu/BUILD | 1 + tensorflow/compiler/xla/service/mlir_gpu/kernel_lowering.cc | 5 +++++ tensorflow/compiler/xla/service/mlir_gpu/kernel_lowering.h | 1 + 3 files changed, 7 insertions(+) diff --git a/tensorflow/compiler/xla/service/mlir_gpu/BUILD b/tensorflow/compiler/xla/service/mlir_gpu/BUILD index ce45d937424..efe69450846 100644 --- a/tensorflow/compiler/xla/service/mlir_gpu/BUILD +++ b/tensorflow/compiler/xla/service/mlir_gpu/BUILD @@ -167,6 +167,7 @@ cc_library( "//tensorflow/compiler/mlir/xla:lhlo_legalize_to_affine", "//tensorflow/compiler/mlir/xla:lhlo_legalize_to_gpu", "//tensorflow/compiler/mlir/xla:xla_dialect_registration", + "//tensorflow/compiler/mlir/xla:xla_legalize_tanh_to_approximation", "//tensorflow/compiler/mlir/xla:xla_legalize_to_linalg", "//tensorflow/compiler/xla:status", "//tensorflow/compiler/xla:statusor", diff --git a/tensorflow/compiler/xla/service/mlir_gpu/kernel_lowering.cc b/tensorflow/compiler/xla/service/mlir_gpu/kernel_lowering.cc index b0cbddcdb92..9d5b52df010 100644 --- a/tensorflow/compiler/xla/service/mlir_gpu/kernel_lowering.cc +++ b/tensorflow/compiler/xla/service/mlir_gpu/kernel_lowering.cc @@ -505,6 +505,11 @@ Status LowerLHLOToGPU(mlir::ModuleOp module, LowerLHLOToGPUOptions options) { // Some basic cleanup. pm.addNestedPass<::mlir::FuncOp>(::mlir::createCanonicalizerPass()); pm.addNestedPass<::mlir::FuncOp>(::mlir::createCSEPass()); + // Approximate of requested. + if (options.use_approximations) { + pm.addNestedPass<::mlir::FuncOp>( + ::mlir::xla::createLegalizeTanhToApproximationPass()); + } // Move scalar operations into the launch to ensure smaller signatures. pm.addPass(absl::make_unique()); // Take launches to launches with kernels. diff --git a/tensorflow/compiler/xla/service/mlir_gpu/kernel_lowering.h b/tensorflow/compiler/xla/service/mlir_gpu/kernel_lowering.h index 77cf75b9e47..bd633bb06cb 100644 --- a/tensorflow/compiler/xla/service/mlir_gpu/kernel_lowering.h +++ b/tensorflow/compiler/xla/service/mlir_gpu/kernel_lowering.h @@ -28,6 +28,7 @@ struct LowerLHLOToGPUOptions { llvm::ArrayRef unroll_factors = {}; bool collapse_parallel_loops = true; bool rewrite_signature = true; + bool use_approximations = false; }; Status LowerLHLOToGPU(mlir::ModuleOp module, From 16e6f9e792741b98d27b4f1463057313d04acdd8 Mon Sep 17 00:00:00 2001 From: Stephan Herhut Date: Thu, 18 Jun 2020 05:00:40 -0700 Subject: [PATCH 171/867] Enable unrolling for the generated tanh kernel. PiperOrigin-RevId: 317080498 Change-Id: Idaac4b2efee78cd2cd44c68ede4ac56f1a4cde40 --- tensorflow/core/kernels/cubin_headers/BUILD | 1 + 1 file changed, 1 insertion(+) diff --git a/tensorflow/core/kernels/cubin_headers/BUILD b/tensorflow/core/kernels/cubin_headers/BUILD index b8ba164fbc3..ec8b44050db 100644 --- a/tensorflow/core/kernels/cubin_headers/BUILD +++ b/tensorflow/core/kernels/cubin_headers/BUILD @@ -37,4 +37,5 @@ gen_kernel_library( "f32", "f64", ], + unroll_factors = "4", ) From 0a541ad1cc89f1eead6a47dc1676bd86dc810937 Mon Sep 17 00:00:00 2001 From: Christian Sigg Date: Thu, 18 Jun 2020 05:07:41 -0700 Subject: [PATCH 172/867] Remove intermediate relocatable code stored in __nv_relfatbin sections, if objcopy is at least version 2.26 (which added support for --update-sections). The intermediate code is a result of separate compilation and linking, and removing it reduces TF's GPU wheel size. PiperOrigin-RevId: 317081343 Change-Id: I603477b4499344aeec653765be78de11f392eac6 --- third_party/nccl/build_defs.bzl.tpl | 103 ++++++++++++++++++++++++---- 1 file changed, 88 insertions(+), 15 deletions(-) diff --git a/third_party/nccl/build_defs.bzl.tpl b/third_party/nccl/build_defs.bzl.tpl index 9268af7c890..b520f71d0f1 100644 --- a/third_party/nccl/build_defs.bzl.tpl +++ b/third_party/nccl/build_defs.bzl.tpl @@ -169,35 +169,94 @@ _device_link = rule( ) """Links device code and generates source code for kernel registration.""" +def _prune_relocatable_code_impl(ctx): + """Clears __nv_relfatbin section containing relocatable device code.""" + empty_file = ctx.actions.declare_file(ctx.attr.name + "__nv_relfatbin") + ctx.actions.write(empty_file, "") + + # Parse 'objcopy --version' and update section if it's at least v2.26. + # Otherwise, simply copy the file without changing it. + # TODO(csigg): version parsing is brittle, can we do better? + command = r""" + objcopy=$1 \ + section=$2 \ + input=$3 \ + output=$4 \ + args="" \ + pattern='([0-9])\.([0-9]+)'; \ + if [[ $($objcopy --version) =~ $pattern ]] && { \ + [ ${BASH_REMATCH[1]} -gt 2 ] || \ + [ ${BASH_REMATCH[2]} -ge 26 ]; }; then \ + args="--update-section __nv_relfatbin=$section"; \ + fi; \ + $objcopy $args $input $output + """ + cc_toolchain = find_cpp_toolchain(ctx) + outputs = [] + for src in ctx.files.srcs: + out = ctx.actions.declare_file("pruned_" + src.basename, sibling = src) + ctx.actions.run_shell( + inputs = [empty_file] + ctx.files.srcs, # + ctx.files._crosstool, + outputs = [out], + arguments = [ + cc_toolchain.objcopy_executable, + empty_file.path, + src.path, + out.path, + ], + command = command, + ) + outputs.append(out) + return DefaultInfo(files = depset(outputs)) + +_prune_relocatable_code = rule( + implementation = _prune_relocatable_code_impl, + attrs = { + "srcs": attr.label_list(mandatory = True, allow_files = True), + "_cc_toolchain": attr.label( + default = "@bazel_tools//tools/cpp:current_cc_toolchain", + ), + # "_crosstool": attr.label_list( + # cfg = "host", + # default = ["@bazel_tools//tools/cpp:crosstool"] + # ), + }, +) + def _merge_archive_impl(ctx): # Generate an mri script to the merge archives in srcs and pass it to 'ar'. # See https://stackoverflow.com/a/23621751. files = _pic_only(ctx.files.srcs) mri_script = "create " + ctx.outputs.out.path for f in files: - mri_script += "\\naddlib " + f.path - mri_script += "\\nsave\\nend" + mri_script += r"\naddlib " + f.path + mri_script += r"\nsave\nend" cc_toolchain = find_cpp_toolchain(ctx) ctx.actions.run_shell( inputs = ctx.files.srcs, # + ctx.files._crosstool, outputs = [ctx.outputs.out], - command = "printf \"%s\" | %s -M" % (mri_script, cc_toolchain.ar_executable), + command = "echo -e \"%s\" | %s -M" % (mri_script, cc_toolchain.ar_executable), ) _merge_archive = rule( implementation = _merge_archive_impl, attrs = { "srcs": attr.label_list(mandatory = True, allow_files = True), - "_cc_toolchain": attr.label(default = "@bazel_tools//tools/cpp:current_cc_toolchain"), - # "_crosstool": attr.label_list(cfg = "host", default = ["@bazel_tools//tools/cpp:crosstool"]), + "_cc_toolchain": attr.label( + default = "@bazel_tools//tools/cpp:current_cc_toolchain", + ), + # "_crosstool": attr.label_list( + # cfg = "host", + # default = ["@bazel_tools//tools/cpp:crosstool"] + # ), }, outputs = {"out": "lib%{name}.a"}, ) """Merges srcs into a single archive.""" def cuda_rdc_library(name, hdrs = None, copts = None, linkstatic = True, **kwargs): - """Produces a cuda_library using separate compilation and linking. + r"""Produces a cuda_library using separate compilation and linking. CUDA separate compilation and linking allows device function calls across translation units. This is different from the normal whole program @@ -239,17 +298,24 @@ def cuda_rdc_library(name, hdrs = None, copts = None, linkstatic = True, **kwarg The steps marked with '*' are implemented in the _device_link rule. + The intermediate relocatable device code in xy.a is no longer needed at + this point and the corresponding section is replaced with an empty one using + objcopy. We do not remove the section completely because it is referenced by + relocations, and removing those as well breaks fatbin registration. + The object files in both xy.a and dlink.a reference symbols defined in the other archive. The separate archives are a side effect of using two cc_library targets to implement a single compilation trajectory. We could fix this once bazel supports C++ sandwich. For now, we just merge the two archives to avoid unresolved symbols: - xy.a dlink.a - \ / merge archive - xy_dlink.a - | cc_library (or alternatively, cc_import) - final target + xy.a + | objcopy --update-section __nv_relfatbin='' + dlink.a xy_pruned.a + \ / merge archive + xy_merged.a + | cc_library (or alternatively, cc_import) + final target Another complication is that cc_library produces (depending on the configuration) both PIC and non-PIC archives, but the distinction @@ -313,19 +379,26 @@ def cuda_rdc_library(name, hdrs = None, copts = None, linkstatic = True, **kwarg linkstatic = linkstatic, ) + # Remove intermediate relocatable device code. + pruned = name + "_pruned" + _prune_relocatable_code( + name = pruned, + srcs = [lib], + ) + # Repackage the two libs into a single archive. This is required because # both libs reference symbols defined in the other one. For details, see # https://eli.thegreenplace.net/2013/07/09/library-order-in-static-linking - archive = name + "_a" + merged = name + "_merged" _merge_archive( - name = archive, - srcs = [lib, dlink], + name = merged, + srcs = [pruned, dlink], ) # Create cc target from archive. native.cc_library( name = name, - srcs = [archive], + srcs = [merged], hdrs = hdrs, linkstatic = linkstatic, ) From 7378fabf90f489387e7dbfb111e20ddf03b1910a Mon Sep 17 00:00:00 2001 From: Adrian Kuegel Date: Thu, 18 Jun 2020 05:19:27 -0700 Subject: [PATCH 173/867] Emit tuple at the end of the Sort emitter. If sort has more than one operand, the result is a tuple. So far, we didn't emit the tuple at the end of the emitter. PiperOrigin-RevId: 317082573 Change-Id: I7bec31302ba2e40556b17654daa081428871a00e --- .../xla/service/gpu/ir_emitter_unnested.cc | 7 ++++ tensorflow/compiler/xla/tests/tuple_test.cc | 32 +++++++++++++++++++ 2 files changed, 39 insertions(+) diff --git a/tensorflow/compiler/xla/service/gpu/ir_emitter_unnested.cc b/tensorflow/compiler/xla/service/gpu/ir_emitter_unnested.cc index 937a0ea5bbc..74aad5f5bd5 100644 --- a/tensorflow/compiler/xla/service/gpu/ir_emitter_unnested.cc +++ b/tensorflow/compiler/xla/service/gpu/ir_emitter_unnested.cc @@ -1418,6 +1418,13 @@ Status IrEmitterUnnested::HandleSort(HloInstruction* sort) { AddThunkToThunkSequence( absl::make_unique(std::move(thunks), sort)); + if (sort->operand_count() > 1) { + // Emit the tuple as part of the last stage of sorting. + // We are currently in the block sorted.in_bounds.after. + b_.SetInsertPoint(b_.GetInsertBlock()->getTerminator()); + llvm_ir::EmitTuple(GetIrArray(*sort, *sort), + ConstructIrArrayForOutputs(*sort), &b_); + } return Status::OK(); } diff --git a/tensorflow/compiler/xla/tests/tuple_test.cc b/tensorflow/compiler/xla/tests/tuple_test.cc index 9ef589e5511..b6ad44497e6 100644 --- a/tensorflow/compiler/xla/tests/tuple_test.cc +++ b/tensorflow/compiler/xla/tests/tuple_test.cc @@ -577,5 +577,37 @@ XLA_TEST_F(TupleHloTest, EXPECT_TRUE(LiteralTestUtil::Equal(expected, literal)); } +XLA_TEST_F(TupleHloTest, TupleSelectOfSort) { + const char* testcase = R"( + HloModule sort + + compare { + p.1.lhs = s32[] parameter(2) + p.1.rhs = s32[] parameter(3) + p.0.lhs = f32[] parameter(0) + p.0.rhs = f32[] parameter(1) + ROOT lt = pred[] compare(p.0.lhs, p.0.rhs), direction=LT + } + + ENTRY Sort { + keys = f32[2]{0} iota(), iota_dimension=0 + values = s32[2]{0} iota(), iota_dimension=0 + preds = pred[] constant(true) + alt = (f32[2], s32[2]) parameter(0) + + sorted = (f32[2]{0}, s32[2]{0}) sort(keys, values), dimensions={0}, + to_apply=compare + ROOT selected = (f32[2], s32[2]) tuple-select(preds, sorted, alt) + } + )"; + auto module = ParseAndReturnVerifiedModule(testcase).ValueOrDie(); + auto param = LiteralUtil::MakeTupleOwned(LiteralUtil::CreateR1({2, 3}), + LiteralUtil::CreateR1({3, 4})); + auto expected = LiteralUtil::MakeTupleOwned( + LiteralUtil::CreateR1({0, 1}), LiteralUtil::CreateR1({0, 1})); + auto result = ExecuteAndTransfer(std::move(module), {¶m}); + EXPECT_TRUE(LiteralTestUtil::Equal(expected, result)); +} + } // namespace } // namespace xla From 78f6d51b8d9a2f3542918f004a29a6fbba232a40 Mon Sep 17 00:00:00 2001 From: Adrian Kuegel Date: Thu, 18 Jun 2020 05:26:02 -0700 Subject: [PATCH 174/867] Also set the unroll factor in the C++ integration code for tanh. PiperOrigin-RevId: 317083225 Change-Id: I7c2c26d664c15cbc967188da4b3012161edbcf49 --- tensorflow/core/kernels/mlir_generated_cwise_op_gpu_tanh.cu.cc | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/tensorflow/core/kernels/mlir_generated_cwise_op_gpu_tanh.cu.cc b/tensorflow/core/kernels/mlir_generated_cwise_op_gpu_tanh.cu.cc index 70de777239f..a122c5112e6 100644 --- a/tensorflow/core/kernels/mlir_generated_cwise_op_gpu_tanh.cu.cc +++ b/tensorflow/core/kernels/mlir_generated_cwise_op_gpu_tanh.cu.cc @@ -124,7 +124,7 @@ class MlirGeneratedTanhOp : public OpKernel { // This has to be aligned with the configuration that was used when building // the kernels. See the corresponding build rules in `cubin_headers/BUILD`. LaunchConfig config = GetLaunchConfiguration( - {256}, {}, {static_cast(inp.NumElements())}); + {256}, {4}, {static_cast(inp.NumElements())}); OP_REQUIRES_OK( ctx, stream->parent()->Launch(stream, config.threadDim, config.blockDim, *kernel, args)); From 2ad57928530e3e10ba89e91ff4df3c6fa9feafa0 Mon Sep 17 00:00:00 2001 From: Adrian Kuegel Date: Thu, 18 Jun 2020 05:29:17 -0700 Subject: [PATCH 175/867] Bump open source llvm revision to e3836fe1a5562875396705369353078ab07cf07a PiperOrigin-RevId: 317083604 Change-Id: Ifd810846a890ad137a31b4db5af5e049450765d3 --- tensorflow/workspace.bzl | 4 ++-- 1 file changed, 2 insertions(+), 2 deletions(-) diff --git a/tensorflow/workspace.bzl b/tensorflow/workspace.bzl index 78f7e0ce03e..1c165ba5aba 100755 --- a/tensorflow/workspace.bzl +++ b/tensorflow/workspace.bzl @@ -710,8 +710,8 @@ def tf_repositories(path_prefix = "", tf_repo_name = ""): ) # Check out LLVM and MLIR from llvm-project. - LLVM_COMMIT = "4799fb63b5513f655ca8e85416ec8fe35df49bae" - LLVM_SHA256 = "f401a61bd7f5b05bd8a3ffdfb1f32e9379cae2c8e988f3ae6772b588ad97c84a" + LLVM_COMMIT = "e3836fe1a5562875396705369353078ab07cf07a" + LLVM_SHA256 = "6a78815e2c71c560a11c8c1740d31b88a607d82b7ccc61dc142bef0f1f3fbde8" LLVM_URLS = [ "https://storage.googleapis.com/mirror.tensorflow.org/github.com/llvm/llvm-project/archive/{commit}.tar.gz".format(commit = LLVM_COMMIT), "https://github.com/llvm/llvm-project/archive/{commit}.tar.gz".format(commit = LLVM_COMMIT), From 4f5c65d4494b4e4831d016176d506227c011f01b Mon Sep 17 00:00:00 2001 From: "A. Unique TensorFlower" Date: Thu, 18 Jun 2020 06:50:27 -0700 Subject: [PATCH 176/867] Make linear layout more explicit. PiperOrigin-RevId: 317093123 Change-Id: I7af437c2c8afb31683bb659b1939eac2ce851da5 --- .../compiler/xla/service/layout_assignment.cc | 14 +++++++------ .../xla/service/layout_assignment_test.cc | 21 +++++++++++++++++++ 2 files changed, 29 insertions(+), 6 deletions(-) diff --git a/tensorflow/compiler/xla/service/layout_assignment.cc b/tensorflow/compiler/xla/service/layout_assignment.cc index 307fd82069e..a35ba140e86 100644 --- a/tensorflow/compiler/xla/service/layout_assignment.cc +++ b/tensorflow/compiler/xla/service/layout_assignment.cc @@ -951,12 +951,7 @@ Status LayoutAssignment::CheckLayouts(HloModule* module) { if (!Shape::Equal() .IgnoreDynamicDimension() .MinorToMajorOnlyInLayout()(instruction_subshape, - buffer->shape()) && - // TODO(mingyao): Use explicit linear layout tiling to - // detect and allow special bitcast. - instruction->opcode() != HloOpcode::kBitcast && - instruction->opcode() != HloOpcode::kGetTupleElement && - instruction->opcode() != HloOpcode::kTuple) { + buffer->shape())) { return InternalError( "Layout of instruction %s at index {%s} does not match " "source LogicalBuffer %s: %s vs %s", @@ -1803,6 +1798,13 @@ Status LayoutAssignment::ClearComputationLayouts(HloComputation* computation) { // potential bugs in the layout assignment pass that may accidentally use the // existing layout. for (HloInstruction* instruction : computation->instructions()) { + if (instruction->opcode() == HloOpcode::kBitcast) { + // bitcasts are inherently layout sensitive and so a bitcast instruction + // present in the IR before layout assignment is a bug. + return InternalError( + "Unexpected bitcast operation seen during layout assignment: %s.", + instruction->ToString()); + } // Some instructions carry mandatory layouts in their shape. if (instruction->opcode() != HloOpcode::kInfeed && !IsLayoutConstrainedCustomCall(instruction) && diff --git a/tensorflow/compiler/xla/service/layout_assignment_test.cc b/tensorflow/compiler/xla/service/layout_assignment_test.cc index 6e575247e6b..304a80c7a52 100644 --- a/tensorflow/compiler/xla/service/layout_assignment_test.cc +++ b/tensorflow/compiler/xla/service/layout_assignment_test.cc @@ -814,6 +814,27 @@ TEST_F(LayoutAssignmentTest, ConditionalAsymmetricLayout) { EXPECT_THAT(false_result->opcode(), HloOpcode::kCopy); } +TEST_F(LayoutAssignmentTest, InternalErrorOnBitcast) { + auto builder = HloComputation::Builder(TestName()); + auto constant0 = builder.AddInstruction( + HloInstruction::CreateConstant(LiteralUtil::CreateR2WithLayout( + {{1.0, 2.0}, {3.0, 4.0}}, LayoutUtil::MakeLayout({0, 1})))); + builder.AddInstruction( + HloInstruction::CreateBitcast(constant0->shape(), constant0)); + auto m = CreateNewVerifiedModule(); + m->AddEntryComputation(builder.Build()); + + ComputationLayout computation_layout( + m->entry_computation()->ComputeProgramShape()); + LayoutAssignment layout_assignment(&computation_layout); + Status error_status = layout_assignment.Run(m.get()).status(); + EXPECT_FALSE(error_status.ok()); + EXPECT_THAT( + error_status.error_message(), + ::testing::HasSubstr( + "Unexpected bitcast operation seen during layout assignment")); +} + TEST_F(LayoutAssignmentTest, ChannelLayoutMismatch) { // Pin non matching layouts to parameter and root. const char* module_str = R"( From c65fccf0a6671c90599d0d3426dd18597688ea3a Mon Sep 17 00:00:00 2001 From: shwetaoj Date: Wed, 17 Jun 2020 11:55:27 -0700 Subject: [PATCH 177/867] Format changes as per Google's feedback --- .../ci_build/linux/mkl/Dockerfile.devel-mkl | 4 ++-- .../ci_build/linux/mkl/build-dev-container.sh | 21 +++++++++---------- .../linux/mkl/install_openmpi_horovod.sh | 7 ++++--- 3 files changed, 16 insertions(+), 16 deletions(-) diff --git a/tensorflow/tools/ci_build/linux/mkl/Dockerfile.devel-mkl b/tensorflow/tools/ci_build/linux/mkl/Dockerfile.devel-mkl index a78d13c7755..1f80cba35f0 100755 --- a/tensorflow/tools/ci_build/linux/mkl/Dockerfile.devel-mkl +++ b/tensorflow/tools/ci_build/linux/mkl/Dockerfile.devel-mkl @@ -64,9 +64,9 @@ RUN bazel --bazelrc=/root/.bazelrc build -c opt \ # Install OpenMPI/Horovod COPY install_openmpi_horovod.sh . -RUN if [ "${ENABLE_HOROVOD}" = "yes" ]; then \ +RUN if [ "${ENABLE_HOROVOD}" == "yes" ]; then \ chmod +x install_openmpi_horovod.sh && \ - ./install_openmpi_horovod.sh OPENMPI_VERSION=${OPENMPI_VERSION} OPENMPI_DOWNLOAD_URL=${OPENMPI_DOWNLOAD_URL} HOROVOD_VERSION=${HOROVOD_VERSION} && \ + OPENMPI_VERSION=${OPENMPI_VERSION} OPENMPI_DOWNLOAD_URL=${OPENMPI_DOWNLOAD_URL} HOROVOD_VERSION=${HOROVOD_VERSION} ./install_openmpi_horovod.sh && \ rm -rf install_openmpi_horovod.sh; \ fi diff --git a/tensorflow/tools/ci_build/linux/mkl/build-dev-container.sh b/tensorflow/tools/ci_build/linux/mkl/build-dev-container.sh index 6e789a54e87..a0880b0e51c 100755 --- a/tensorflow/tools/ci_build/linux/mkl/build-dev-container.sh +++ b/tensorflow/tools/ci_build/linux/mkl/build-dev-container.sh @@ -201,21 +201,20 @@ function test_container() debug "ID of the running docker container: ${CONTAINER_ID}" debug "Performing basic sanity checks on the running container..." - TEST_CMD_1=$(${DOCKER_BINARY} exec ${CONTAINER_ID} bash -c "${PYTHON} -c 'from tensorflow.python import _pywrap_util_port; print(_pywrap_util_port.IsMklEnabled())'") - # Make TEST_CMD backward compatible with older code - TEST_CMD_2=$(${DOCKER_BINARY} exec ${CONTAINER_ID} bash -c "${PYTHON} -c 'from tensorflow.python import pywrap_tensorflow; print(pywrap_tensorflow.IsMklEnabled())'") - - if [ "${TEST_CMD_1}" = "True" -o "${TEST_CMD_2}" = "True" ] ; then - echo "PASS: MKL enabled test in ${TEMP_IMAGE_NAME}" - else - die "FAIL: MKL enabled test in ${TEMP_IMAGE_NAME}" - fi + { + ${DOCKER_BINARY} exec ${CONTAINER_ID} bash -c "${PYTHON} -c 'from tensorflow.python import _pywrap_util_port; print(_pywrap_util_port.IsMklEnabled())'" + echo "PASS: MKL enabled test in ${TEMP_IMAGE_NAME}" + } || { + ${DOCKER_BINARY} exec ${CONTAINER_ID} bash -c "${PYTHON} -c 'from tensorflow.python import pywrap_tensorflow; print(pywrap_tensorflow.IsMklEnabled())'" + echo "PASS: Old MKL enabled in ${TEMP_IMAGE_NAME}" + } || { + die "FAIL: MKL enabled test in ${TEMP_IMAGE_NAME}" + } # Test to check if horovod is installed successfully if [[ ${ENABLE_HOROVOD} == "yes" ]]; then debug "Test horovod in the container..." - HOROVOD_TEST_CMD=$(${DOCKER_BINARY} exec ${CONTAINER_ID} bash -c "${PYTHON} -c 'import horovod.tensorflow as hvd;'") - ${HOROVOD_TEST_CMD} + ${DOCKER_BINARY} exec ${CONTAINER_ID} bash -c "${PYTHON} -c 'import horovod.tensorflow as hvd;'" if [[ $? == "0" ]]; then echo "PASS: HOROVOD installation test in ${TEMP_IMAGE_NAME}" else diff --git a/tensorflow/tools/ci_build/linux/mkl/install_openmpi_horovod.sh b/tensorflow/tools/ci_build/linux/mkl/install_openmpi_horovod.sh index aec40543a17..9bc92ca4fef 100755 --- a/tensorflow/tools/ci_build/linux/mkl/install_openmpi_horovod.sh +++ b/tensorflow/tools/ci_build/linux/mkl/install_openmpi_horovod.sh @@ -54,8 +54,9 @@ echo 'OpenMPI version:' mpirun --version # Install OpenSSH for MPI to communicate between containers -apt-get clean && apt-get update && apt-get install -y --no-install-recommends --fix-missing \ - openssh-client openssh-server libnuma-dev && \ +apt-get clean && apt-get update && \ + apt-get install -y --no-install-recommends --fix-missing \ + openssh-client openssh-server libnuma-dev && \ rm -rf /var/lib/apt/lists/* if [[ $? == "0" ]]; then echo "PASS: OpenSSH installation" @@ -70,7 +71,7 @@ else fi mkdir -p /var/run/sshd # Allow OpenSSH to talk to containers without asking for confirmation -cat /etc/ssh/ssh_config | grep -v StrictHostKeyChecking > /etc/ssh/ssh_config.new +grep -v StrictHostKeyChecking /etc/ssh/ssh_config > /etc/ssh/ssh_config.new echo " StrictHostKeyChecking no" >> /etc/ssh/ssh_config.new mv /etc/ssh/ssh_config.new /etc/ssh/ssh_config From 55929ce6ae2bc9552526a038ef4d01d8bef4f4fd Mon Sep 17 00:00:00 2001 From: shwetaoj Date: Wed, 17 Jun 2020 12:52:57 -0700 Subject: [PATCH 178/867] Reverting == in Dockerfile --- tensorflow/tools/ci_build/linux/mkl/Dockerfile.devel-mkl | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/tensorflow/tools/ci_build/linux/mkl/Dockerfile.devel-mkl b/tensorflow/tools/ci_build/linux/mkl/Dockerfile.devel-mkl index 1f80cba35f0..80091e55a17 100755 --- a/tensorflow/tools/ci_build/linux/mkl/Dockerfile.devel-mkl +++ b/tensorflow/tools/ci_build/linux/mkl/Dockerfile.devel-mkl @@ -64,7 +64,7 @@ RUN bazel --bazelrc=/root/.bazelrc build -c opt \ # Install OpenMPI/Horovod COPY install_openmpi_horovod.sh . -RUN if [ "${ENABLE_HOROVOD}" == "yes" ]; then \ +RUN if [ "${ENABLE_HOROVOD}" = "yes" ]; then \ chmod +x install_openmpi_horovod.sh && \ OPENMPI_VERSION=${OPENMPI_VERSION} OPENMPI_DOWNLOAD_URL=${OPENMPI_DOWNLOAD_URL} HOROVOD_VERSION=${HOROVOD_VERSION} ./install_openmpi_horovod.sh && \ rm -rf install_openmpi_horovod.sh; \ From c054f40f66fa625f51085a20c48554c61d05c5fd Mon Sep 17 00:00:00 2001 From: Deven Desai Date: Thu, 18 Jun 2020 14:48:42 +0000 Subject: [PATCH 179/867] [ROCm] Fix ROCm CSB build failure - 200618 The folllowing commit introduces build error on the ROCm platform https://github.com/tensorflow/tensorflow/commit/b6ff68822a59578f942e4fb8076757da8db278ae build error ``` In file included from tensorflow/core/kernels/split_lib_gpu.cu.cc:27: In file included from ./tensorflow/core/util/gpu_kernel_helper.h:25: ./tensorflow/core/util/gpu_device_functions.h:824:25: error: redefinition of 'GpuAtomicMax' __device__ inline int64 GpuAtomicMax(int64* ptr, int64 value) { ^ ./tensorflow/core/util/gpu_device_functions.h:792:29: note: previous definition is here __device__ inline long long GpuAtomicMax(long long* ptr, long long value) { ^ ./tensorflow/core/util/gpu_device_functions.h:894:25: error: redefinition of 'GpuAtomicMin' __device__ inline int64 GpuAtomicMin(int64* ptr, int64 value) { ^ ./tensorflow/core/util/gpu_device_functions.h:862:29: note: previous definition is here __device__ inline long long GpuAtomicMin(long long* ptr, long long value) { ^ 2 errors generated. ... ... ``` The cause is a combination of two things * The condition `#if __CUDA_ARCH__ < 320` will hold true for ROCm too! * The issue being addressed by (the build breaking commit, for CUDA) was already fixed by this commit (https://github.com/tensorflow/tensorflow/commit/307485737f46a76c97aefb51b0fc3cd264c2bb94) within a `#if TENSORFLOW_USE_ROCM` block The fix being submitted in this PR, is to undo some of the changes introduced by the earlier ROCm commit, and combine that change with the change in the breaking commit. --- tensorflow/core/util/gpu_device_functions.h | 14 ++------------ 1 file changed, 2 insertions(+), 12 deletions(-) diff --git a/tensorflow/core/util/gpu_device_functions.h b/tensorflow/core/util/gpu_device_functions.h index d4e09a7fc98..a8158a1ab08 100644 --- a/tensorflow/core/util/gpu_device_functions.h +++ b/tensorflow/core/util/gpu_device_functions.h @@ -789,11 +789,6 @@ __device__ inline double GpuAtomicMax(double* ptr, double value) { ptr, [value](double a) { return fmax(a, value); }); } -__device__ inline long long GpuAtomicMax(long long* ptr, long long value) { - return detail::GpuAtomicCasHelper( - ptr, [value](long long a) { return max(a, value); }); -} - #else __device__ inline float GpuAtomicMax(float* ptr, float value) { @@ -814,7 +809,7 @@ __device__ inline Eigen::half GpuAtomicMax(Eigen::half* ptr, ptr, [value](Eigen::half a) { return max(a, value); }); } -#if __CUDA_ARCH__ < 320 +#if TENSORFLOW_USE_ROCM || (__CUDA_ARCH__ < 320) __device__ inline tensorflow::uint64 GpuAtomicMax(tensorflow::uint64* ptr, tensorflow::uint64 value) { return detail::GpuAtomicCasHelper( @@ -859,11 +854,6 @@ __device__ inline double GpuAtomicMin(double* ptr, double value) { ptr, [value](double a) { return fmin(a, value); }); } -__device__ inline long long GpuAtomicMin(long long* ptr, long long value) { - return detail::GpuAtomicCasHelper( - ptr, [value](long long a) { return min(a, value); }); -} - #else __device__ inline float GpuAtomicMin(float* ptr, float value) { @@ -884,7 +874,7 @@ __device__ inline Eigen::half GpuAtomicMin(Eigen::half* ptr, ptr, [value](Eigen::half a) { return min(a, value); }); } -#if __CUDA_ARCH__ < 320 +#if TENSORFLOW_USE_ROCM || (__CUDA_ARCH__ < 320) __device__ inline tensorflow::uint64 GpuAtomicMin(tensorflow::uint64* ptr, tensorflow::uint64 value) { return detail::GpuAtomicCasHelper( From 49e3e63dd7d6d7cf700013000efc8c8da80662ab Mon Sep 17 00:00:00 2001 From: "A. Unique TensorFlower" Date: Thu, 18 Jun 2020 08:20:28 -0700 Subject: [PATCH 180/867] Integrate LLVM at https://github.com/llvm/llvm-project/commit/92d8ad02e92f PiperOrigin-RevId: 317105768 Change-Id: I76b13f63a85dd4b8bbcaff3fd1ca624b82411079 --- tensorflow/workspace.bzl | 4 ++-- third_party/mlir/BUILD | 20 ++++++++++++++++++++ 2 files changed, 22 insertions(+), 2 deletions(-) diff --git a/tensorflow/workspace.bzl b/tensorflow/workspace.bzl index 1c165ba5aba..f5b0b7537dc 100755 --- a/tensorflow/workspace.bzl +++ b/tensorflow/workspace.bzl @@ -710,8 +710,8 @@ def tf_repositories(path_prefix = "", tf_repo_name = ""): ) # Check out LLVM and MLIR from llvm-project. - LLVM_COMMIT = "e3836fe1a5562875396705369353078ab07cf07a" - LLVM_SHA256 = "6a78815e2c71c560a11c8c1740d31b88a607d82b7ccc61dc142bef0f1f3fbde8" + LLVM_COMMIT = "92d8ad02e92fed3884169ba5d98056fe4fa5660d" + LLVM_SHA256 = "a4995ace7ddaef0c49293dc65771f58ef1fea96ebe1f39aa0a2d6d75d07f6cc7" LLVM_URLS = [ "https://storage.googleapis.com/mirror.tensorflow.org/github.com/llvm/llvm-project/archive/{commit}.tar.gz".format(commit = LLVM_COMMIT), "https://github.com/llvm/llvm-project/archive/{commit}.tar.gz".format(commit = LLVM_COMMIT), diff --git a/third_party/mlir/BUILD b/third_party/mlir/BUILD index db75b27e78b..a0d10667d23 100644 --- a/third_party/mlir/BUILD +++ b/third_party/mlir/BUILD @@ -738,12 +738,32 @@ cc_library( ":Pass", ":SCFDialect", ":Shape", + ":ShapeToStandardPatternsIncGen", ":StandardOps", ":Support", ":Transforms", ], ) +gentbl( + name = "ShapeToStandardPatternsIncGen", + strip_include_prefix = "include/mlir/Conversion/ShapeToStandard", + tbl_outs = [ + ( + "-gen-rewriters", + "include/mlir/Conversion/ShapeToStandard/ShapeToStandardPatterns.inc", + ), + ], + tblgen = ":mlir-tblgen", + td_file = "lib/Conversion/ShapeToStandard/ShapeToStandardPatterns.td", + td_srcs = [ + ":StdOpsTdFiles", + "include/mlir/Dialect/Shape/IR/ShapeBase.td", + "include/mlir/Dialect/Shape/IR/ShapeOps.td", + "include/mlir/Interfaces/InferTypeOpInterface.td", + ], +) + cc_library( name = "ShapeToSCF", srcs = glob([ From 3cea671a74251d62549280fbb6444ffc2cdc4f03 Mon Sep 17 00:00:00 2001 From: Zhenyu Tan Date: Thu, 18 Jun 2020 08:28:08 -0700 Subject: [PATCH 181/867] set `_batch_input_shape` before build for kpl. PiperOrigin-RevId: 317107061 Change-Id: Idd7f5dcbce9d15b2e3d708dd7ea3b2c3e5c1be7e --- .../python/keras/engine/base_preprocessing_layer.py | 2 ++ .../keras/layers/preprocessing/normalization_test.py | 12 ++++++++++++ 2 files changed, 14 insertions(+) diff --git a/tensorflow/python/keras/engine/base_preprocessing_layer.py b/tensorflow/python/keras/engine/base_preprocessing_layer.py index 08df07e33e3..b2ab0880422 100644 --- a/tensorflow/python/keras/engine/base_preprocessing_layer.py +++ b/tensorflow/python/keras/engine/base_preprocessing_layer.py @@ -190,6 +190,8 @@ class CombinerPreprocessingLayer(PreprocessingLayer): shape = data_element.shape except AttributeError: shape = None + # TODO (b/159261555): move this to base layer build. + self._batch_input_shape = shape self.build(shape) # Once we have built the Layer, we can process the input data. We do so diff --git a/tensorflow/python/keras/layers/preprocessing/normalization_test.py b/tensorflow/python/keras/layers/preprocessing/normalization_test.py index f5f68d9c51a..f97b8db50ec 100644 --- a/tensorflow/python/keras/layers/preprocessing/normalization_test.py +++ b/tensorflow/python/keras/layers/preprocessing/normalization_test.py @@ -318,6 +318,18 @@ class NormalizationTest(keras_parameterized.TestCase, layer.adapt(data) self.assertAllClose(expect, layer(data)) + def test_model_summary_after_layer_adapt(self): + data = np.array([[[0., 1., 2.], [0., 2., 6.]], + [[2., 3., 4.], [3., 6., 10.]]]) + cls = get_layer_class() + layer = cls(axis=-1) + layer.adapt(data) + model = keras.Sequential( + [layer, + keras.layers.Dense(64, activation="relu"), + keras.layers.Dense(1)]) + model.summary() + if __name__ == "__main__": test.main() From 06bc84b12edce4c4ce616d0dbad5e0d5178218b7 Mon Sep 17 00:00:00 2001 From: Jian Li Date: Thu, 18 Jun 2020 08:48:06 -0700 Subject: [PATCH 182/867] Create fully integer ResizeBilinear kernel. PiperOrigin-RevId: 317110515 Change-Id: I233b89083dda9f9e0150c97a88391a154fb32d50 --- .../internal/reference/reference_ops.h | 103 ++++++++++++++++++ 1 file changed, 103 insertions(+) diff --git a/tensorflow/lite/kernels/internal/reference/reference_ops.h b/tensorflow/lite/kernels/internal/reference/reference_ops.h index e991a21e3bd..5208b21eb4d 100644 --- a/tensorflow/lite/kernels/internal/reference/reference_ops.h +++ b/tensorflow/lite/kernels/internal/reference/reference_ops.h @@ -1645,6 +1645,109 @@ inline void ResizeBilinear(const tflite::ResizeBilinearParams& op_params, } } +inline void ComputeInterpolationValues(const int32 value, const int32 scale_10, + const bool half_pixel_centers, + int32 input_size, int32* scaled_value, + int32* lower_bound, int32* upper_bound) { + if (half_pixel_centers) { + *scaled_value = value * scale_10 + scale_10 / 2 - (1 << 9); + } else { + *scaled_value = value * scale_10; + } + *lower_bound = std::max(*scaled_value / (1 << 10), 0); + *upper_bound = std::min(*scaled_value / (1 << 10) + 1, input_size - 1); +} + +// Same as above but takes int8 as input and output. +inline void ResizeBilinear(const tflite::ResizeBilinearParams& op_params, + const RuntimeShape& unextended_input_shape, + const int8_t* input_data, + const RuntimeShape& unextended_output_size_shape, + const int32* output_size_data, + const RuntimeShape& unextended_output_shape, + int8_t* output_data) { + // If half_pixel_centers is True, align_corners must be False. + TFLITE_DCHECK(!op_params.half_pixel_centers || !op_params.align_corners); + TFLITE_DCHECK_LE(unextended_input_shape.DimensionsCount(), 4); + TFLITE_DCHECK_LE(unextended_output_size_shape.DimensionsCount(), 4); + TFLITE_DCHECK_LE(unextended_output_shape.DimensionsCount(), 4); + const RuntimeShape input_shape = + RuntimeShape::ExtendedShape(4, unextended_input_shape); + const RuntimeShape output_size_shape = + RuntimeShape::ExtendedShape(4, unextended_output_size_shape); + const RuntimeShape output_shape = + RuntimeShape::ExtendedShape(4, unextended_output_shape); + + const int32 batches = MatchingDim(input_shape, 0, output_shape, 0); + const int32 input_height = input_shape.Dims(1); + const int32 input_width = input_shape.Dims(2); + const int32 depth = MatchingDim(input_shape, 3, output_shape, 3); + + TFLITE_DCHECK_EQ(output_size_shape.Dims(0), 1); + TFLITE_DCHECK_EQ(output_size_shape.Dims(1), 1); + TFLITE_DCHECK_EQ(output_size_shape.Dims(2), 1); + TFLITE_DCHECK_EQ(output_size_shape.Dims(3), 2); + const int32 output_height = + output_size_data[Offset(output_size_shape, 0, 0, 0, 0)]; + const int32 output_width = + output_size_data[Offset(output_size_shape, 0, 0, 0, 1)]; + + int32 height_scale_10 = + ((1 << 10) * input_height + output_height / 2) / output_height; + int32 width_scale_10 = + ((1 << 10) * input_width + output_width / 2) / output_width; + if (op_params.align_corners && output_height > 1) { + height_scale_10 = + ((1 << 10) * (input_height - 1) + (output_height - 1) / 2) / + (output_height - 1); + } + if (op_params.align_corners && output_width > 1) { + width_scale_10 = ((1 << 10) * (input_width - 1) + (output_width - 1) / 2) / + (output_width - 1); + } + + for (int b = 0; b < batches; ++b) { + for (int y = 0; y < output_height; ++y) { + int32 input_y, y0, y1; + ComputeInterpolationValues(y, height_scale_10, + op_params.half_pixel_centers, input_height, + &input_y, &y0, &y1); + for (int x = 0; x < output_width; ++x) { + int32 input_x, x0, x1; + ComputeInterpolationValues(x, width_scale_10, + op_params.half_pixel_centers, input_width, + &input_x, &x0, &x1); + for (int c = 0; c < depth; ++c) { + const int64_t output_20_ll = + static_cast( + input_data[Offset(input_shape, b, y0, x0, c)]) * + ((1 << 10) - (input_y - (1 << 10) * y0)) * + ((1 << 10) - (input_x - (1 << 10) * x0)); + const int64_t output_20_lu = + static_cast( + input_data[Offset(input_shape, b, y1, x0, c)]) * + (input_y - (1 << 10) * y0) * + ((1 << 10) - (input_x - (1 << 10) * x0)); + const int64_t output_20_rl = + static_cast( + input_data[Offset(input_shape, b, y0, x1, c)]) * + ((1 << 10) - (input_y - (1 << 10) * y0)) * + (input_x - (1 << 10) * x0); + const int64_t output_20_ru = + static_cast( + input_data[Offset(input_shape, b, y1, x1, c)]) * + (input_y - (1 << 10) * y0) * (input_x - (1 << 10) * x0); + const int64_t output_20 = + output_20_ll + output_20_lu + output_20_rl + output_20_ru; + const int8_t interpolation = + static_cast((output_20 + (1 << 19)) / (1 << 20)); + output_data[Offset(output_shape, b, y, x, c)] = interpolation; + } + } + } + } +} + template inline void SpaceToBatchND( const SpaceToBatchParams& params, From eda7d05793ec75227069eb0c3f49e0377f33c963 Mon Sep 17 00:00:00 2001 From: Vo Van Nghia Date: Thu, 18 Jun 2020 23:02:21 +0700 Subject: [PATCH 183/867] Add NewWritableFile --- .../experimental/filesystem/plugins/gcs/BUILD | 1 + .../filesystem/plugins/gcs/gcs_filesystem.cc | 50 +++++++++++++++++++ 2 files changed, 51 insertions(+) diff --git a/tensorflow/c/experimental/filesystem/plugins/gcs/BUILD b/tensorflow/c/experimental/filesystem/plugins/gcs/BUILD index c9fee433589..05fd371088c 100644 --- a/tensorflow/c/experimental/filesystem/plugins/gcs/BUILD +++ b/tensorflow/c/experimental/filesystem/plugins/gcs/BUILD @@ -24,6 +24,7 @@ cc_library( "//tensorflow:windows": get_win_copts(), }), deps = [ + "//tensorflow/c:env", "//tensorflow/c:tf_status", "//tensorflow/c/experimental/filesystem:filesystem_interface", "@com_github_googlecloudplatform_google_cloud_cpp//:storage_client", diff --git a/tensorflow/c/experimental/filesystem/plugins/gcs/gcs_filesystem.cc b/tensorflow/c/experimental/filesystem/plugins/gcs/gcs_filesystem.cc index 8c54bc85439..4ddc8548486 100644 --- a/tensorflow/c/experimental/filesystem/plugins/gcs/gcs_filesystem.cc +++ b/tensorflow/c/experimental/filesystem/plugins/gcs/gcs_filesystem.cc @@ -15,8 +15,11 @@ limitations under the License. #include #include +#include + #include "absl/strings/string_view.h" #include "google/cloud/storage/client.h" +#include "tensorflow/c/env.h" #include "tensorflow/c/experimental/filesystem/filesystem_interface.h" #include "tensorflow/c/tf_status.h" @@ -75,6 +78,25 @@ static void ParseGCSPath(absl::string_view fname, bool object_empty_ok, strcpy(*object, object_view.data()); } +class TempFile : public std::fstream { + public: + // We should specify openmode each time we call TempFile. + TempFile(const char* temp_file_name, std::ios::openmode mode) + : std::fstream(temp_file_name, mode), name(temp_file_name) {} + TempFile(TempFile&& rhs) : std::fstream(std::move(rhs)), name(rhs.name) { + rhs.name = nullptr; + } + ~TempFile() { + std::fstream::close(); + std::remove(name); + plugin_memory_free(const_cast(name)); + } + const char* getName() { return name; } + + private: + const char* name; +}; + // SECTION 1. Implementation for `TF_RandomAccessFile` // ---------------------------------------------------------------------------- namespace tf_random_access_file { @@ -86,6 +108,20 @@ namespace tf_random_access_file { // SECTION 2. Implementation for `TF_WritableFile` // ---------------------------------------------------------------------------- namespace tf_writable_file { +typedef struct GCSFile { + const char* bucket; + const char* object; + gcs::Client* gcs_client; // not owned + TempFile outfile; + bool sync_need; +} GCSFile; + +static void Cleanup(TF_WritableFile* file) { + auto gcs_file = static_cast(file->plugin_file); + plugin_memory_free(const_cast(gcs_file->bucket)); + plugin_memory_free(const_cast(gcs_file->object)); + delete gcs_file; +} // TODO(vnvo2409): Implement later @@ -119,6 +155,20 @@ static void Init(TF_Filesystem* filesystem, TF_Status* status) { // TODO(vnvo2409): Implement later +static void NewWritableFile(const TF_Filesystem* filesystem, const char* path, + TF_WritableFile* file, TF_Status* status) { + char* bucket; + char* object; + ParseGCSPath(path, false, &bucket, &object, status); + if (TF_GetCode(status) != TF_OK) return; + + auto gcs_client = static_cast(filesystem->plugin_filesystem); + TempFile outfile(TF_GetTempFileName(""), std::ios::binary | std::ios::out); + file->plugin_file = new tf_writable_file::GCSFile( + {bucket, object, gcs_client, std::move(outfile), true}); + TF_SetStatus(status, TF_OK, ""); +} + } // namespace tf_gcs_filesystem static void ProvideFilesystemSupportFor(TF_FilesystemPluginOps* ops, From 2e3310031296f9232f7a58a5cfca2ee03b9a7c91 Mon Sep 17 00:00:00 2001 From: Jian Li Date: Thu, 18 Jun 2020 09:26:13 -0700 Subject: [PATCH 184/867] Add quantization test for transpose. PiperOrigin-RevId: 317117545 Change-Id: I81c4f9583f29205bcbdaae175eac59439cd19047 --- tensorflow/lite/tools/optimize/BUILD | 1 + .../tools/optimize/quantize_model_test.cc | 44 ++++++++++++++++++ tensorflow/lite/tools/optimize/test_util.cc | 2 + tensorflow/lite/tools/optimize/test_util.h | 3 ++ .../tools/optimize/testdata/transpose.bin | Bin 0 -> 544 bytes 5 files changed, 50 insertions(+) create mode 100644 tensorflow/lite/tools/optimize/testdata/transpose.bin diff --git a/tensorflow/lite/tools/optimize/BUILD b/tensorflow/lite/tools/optimize/BUILD index 3011c01cdeb..c10d4465e5c 100644 --- a/tensorflow/lite/tools/optimize/BUILD +++ b/tensorflow/lite/tools/optimize/BUILD @@ -296,6 +296,7 @@ tf_cc_test( "//tensorflow/lite/tools/optimize:testdata/split.bin", "//tensorflow/lite/tools/optimize:testdata/svdf_calibrated.bin", "//tensorflow/lite/tools/optimize:testdata/svdf_quantized.bin", + "//tensorflow/lite/tools/optimize:testdata/transpose.bin", "//tensorflow/lite/tools/optimize:testdata/unpack.bin", ], tags = [ diff --git a/tensorflow/lite/tools/optimize/quantize_model_test.cc b/tensorflow/lite/tools/optimize/quantize_model_test.cc index f8f1a9d4113..36b35af0065 100644 --- a/tensorflow/lite/tools/optimize/quantize_model_test.cc +++ b/tensorflow/lite/tools/optimize/quantize_model_test.cc @@ -1454,6 +1454,50 @@ TEST_F(QuantizeUnpackTest, VerifyUnpack) { unpack_output_1->quantization->zero_point[0]); } +class QuantizeTransposeTest : public QuantizeModelTest { + protected: + QuantizeTransposeTest() { + input_model_ = ReadModel(internal::kModelWithTranspose); + readonly_model_ = input_model_->GetModel(); + readonly_model_->UnPackTo(&model_); + } +}; + +TEST_F(QuantizeTransposeTest, VerifyTranspose) { + auto status = QuantizeModel(&builder_, &model_, &error_reporter_); + + ASSERT_EQ(kTfLiteOk, status); + + const auto subgraph = model_.subgraphs[0].get(); + auto op = subgraph->operators[1].get(); + + auto float_graph = readonly_model_->subgraphs()->Get(0); + + ASSERT_EQ(model_.operator_codes[op->opcode_index].get()->builtin_code, + BuiltinOperator_TRANSPOSE); + + // The model should only have one input and one outputs. + EXPECT_EQ(subgraph->inputs.size(), 1); + EXPECT_EQ(subgraph->outputs.size(), 1); + + // Get transpose input and output tensors + auto transpose_input = subgraph->tensors[op->inputs[0]].get(); + auto transpose_output = subgraph->tensors[op->outputs[0]].get(); + + // Verify transpose input is quantized. + ASSERT_EQ(float_graph->tensors()->Get(op->inputs[0])->type(), + TensorType_FLOAT32); + EXPECT_EQ(transpose_input->type, TensorType_INT8); + + // Ensure quantization parameters before and after transpose + // are preserved after quantization for all outputs of + // transpose. + EXPECT_FLOAT_EQ(transpose_input->quantization->scale[0], + transpose_output->quantization->scale[0]); + EXPECT_EQ(transpose_input->quantization->zero_point[0], + transpose_output->quantization->zero_point[0]); +} + } // namespace } // namespace optimize } // namespace tflite diff --git a/tensorflow/lite/tools/optimize/test_util.cc b/tensorflow/lite/tools/optimize/test_util.cc index 7d5e9d65f06..61e82ed3e34 100644 --- a/tensorflow/lite/tools/optimize/test_util.cc +++ b/tensorflow/lite/tools/optimize/test_util.cc @@ -61,6 +61,8 @@ const char* kModelWithMaximumOp = "maximum.bin"; const char* kLstmCalibrated2 = "lstm_calibrated2.bin"; const char* kLstmQuantized2 = "lstm_quantized2.bin"; +const char* kModelWithTranspose = "transpose.bin"; + const char* kSvdfCalibrated = "svdf_calibrated.bin"; const char* kSvdfQuantized = "svdf_quantized.bin"; diff --git a/tensorflow/lite/tools/optimize/test_util.h b/tensorflow/lite/tools/optimize/test_util.h index abcdbc21d36..4d2eadf283f 100644 --- a/tensorflow/lite/tools/optimize/test_util.h +++ b/tensorflow/lite/tools/optimize/test_util.h @@ -98,6 +98,9 @@ extern const char* kModelWithMaximumOp; extern const char* kLstmCalibrated2; extern const char* kLstmQuantized2; +// Test model with a transpose op. +extern const char* kModelWithTranspose; + // Test model with SVDF op. extern const char* kSvdfCalibrated; extern const char* kSvdfQuantized; diff --git a/tensorflow/lite/tools/optimize/testdata/transpose.bin b/tensorflow/lite/tools/optimize/testdata/transpose.bin new file mode 100644 index 0000000000000000000000000000000000000000..a76886e5b473b8de04253089e2acc931f5b6ec86 GIT binary patch literal 544 zcmZXRu}T9`5Jj(MjVXd!Y_Vku!A6q8#zJr*!77j@g&(jf1W7{J2!cgECtqN#f8$38 z>pk!JB8dZcXXnkGdH2l}lBN&i8|h4iGIi9|g>pyIvGH{zJ;O1wUyST;jl$@Jvv$Zy zZO56ve_UyE?{H_|P19uN(CgfoPKJ}-a4~TcvV{4>4R@Qpw81AJ(ugYiWOB$rIO!a73tn1~v9=!TD z*t_}63Qp>={nz7T*Mr%`9{U>E4-;YHY>54o2AZ!8-b#i^pkq(t=4});_C8OXdzxFK Q{oxGg5qEeRcIe>h2d$ Date: Thu, 18 Jun 2020 23:46:41 +0700 Subject: [PATCH 185/867] Move TempFile to gcs_helper --- .../experimental/filesystem/plugins/gcs/BUILD | 12 ++++++- .../filesystem/plugins/gcs/gcs_filesystem.cc | 20 +---------- .../filesystem/plugins/gcs/gcs_helper.cc | 19 +++++++++++ .../filesystem/plugins/gcs/gcs_helper.h | 33 +++++++++++++++++++ 4 files changed, 64 insertions(+), 20 deletions(-) create mode 100644 tensorflow/c/experimental/filesystem/plugins/gcs/gcs_helper.cc create mode 100644 tensorflow/c/experimental/filesystem/plugins/gcs/gcs_helper.h diff --git a/tensorflow/c/experimental/filesystem/plugins/gcs/BUILD b/tensorflow/c/experimental/filesystem/plugins/gcs/BUILD index 05fd371088c..d104181b264 100644 --- a/tensorflow/c/experimental/filesystem/plugins/gcs/BUILD +++ b/tensorflow/c/experimental/filesystem/plugins/gcs/BUILD @@ -24,10 +24,20 @@ cc_library( "//tensorflow:windows": get_win_copts(), }), deps = [ - "//tensorflow/c:env", + ":gcs_helper", "//tensorflow/c:tf_status", "//tensorflow/c/experimental/filesystem:filesystem_interface", "@com_github_googlecloudplatform_google_cloud_cpp//:storage_client", "@com_google_absl//absl/strings", ], ) + +cc_library( + name = "gcs_helper", + srcs = ["gcs_helper.cc"], + hdrs = ["gcs_helper.h"], + linkstatic = 1, + deps = [ + "//tensorflow/c:env", + ], +) diff --git a/tensorflow/c/experimental/filesystem/plugins/gcs/gcs_filesystem.cc b/tensorflow/c/experimental/filesystem/plugins/gcs/gcs_filesystem.cc index 4ddc8548486..2793194e0a8 100644 --- a/tensorflow/c/experimental/filesystem/plugins/gcs/gcs_filesystem.cc +++ b/tensorflow/c/experimental/filesystem/plugins/gcs/gcs_filesystem.cc @@ -21,6 +21,7 @@ limitations under the License. #include "google/cloud/storage/client.h" #include "tensorflow/c/env.h" #include "tensorflow/c/experimental/filesystem/filesystem_interface.h" +#include "tensorflow/c/experimental/filesystem/plugins/gcs/gcs_helper.h" #include "tensorflow/c/tf_status.h" // Implementation of a filesystem for GCS environments. @@ -78,25 +79,6 @@ static void ParseGCSPath(absl::string_view fname, bool object_empty_ok, strcpy(*object, object_view.data()); } -class TempFile : public std::fstream { - public: - // We should specify openmode each time we call TempFile. - TempFile(const char* temp_file_name, std::ios::openmode mode) - : std::fstream(temp_file_name, mode), name(temp_file_name) {} - TempFile(TempFile&& rhs) : std::fstream(std::move(rhs)), name(rhs.name) { - rhs.name = nullptr; - } - ~TempFile() { - std::fstream::close(); - std::remove(name); - plugin_memory_free(const_cast(name)); - } - const char* getName() { return name; } - - private: - const char* name; -}; - // SECTION 1. Implementation for `TF_RandomAccessFile` // ---------------------------------------------------------------------------- namespace tf_random_access_file { diff --git a/tensorflow/c/experimental/filesystem/plugins/gcs/gcs_helper.cc b/tensorflow/c/experimental/filesystem/plugins/gcs/gcs_helper.cc new file mode 100644 index 00000000000..139579c53ae --- /dev/null +++ b/tensorflow/c/experimental/filesystem/plugins/gcs/gcs_helper.cc @@ -0,0 +1,19 @@ +#include "tensorflow/c/experimental/filesystem/plugins/gcs/gcs_helper.h" + +#include + +#include +#include + +TempFile::TempFile(const char* temp_file_name, std::ios::openmode mode) + : std::fstream(temp_file_name, mode), name(temp_file_name) {} + +TempFile::TempFile(TempFile&& rhs) + : std::fstream(std::move(rhs)), name(std::move(rhs.name)) {} + +TempFile::~TempFile() { + std::fstream::close(); + std::remove(name.c_str()); +} + +const std::string TempFile::getName() const { return name; } \ No newline at end of file diff --git a/tensorflow/c/experimental/filesystem/plugins/gcs/gcs_helper.h b/tensorflow/c/experimental/filesystem/plugins/gcs/gcs_helper.h new file mode 100644 index 00000000000..437cbe560d6 --- /dev/null +++ b/tensorflow/c/experimental/filesystem/plugins/gcs/gcs_helper.h @@ -0,0 +1,33 @@ +/* Copyright 2020 The TensorFlow Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. +==============================================================================*/ +#ifndef TENSORFLOW_C_EXPERIMENTAL_FILESYSTEM_PLUGINS_GCS_GCS_HELPER_H_ +#define TENSORFLOW_C_EXPERIMENTAL_FILESYSTEM_PLUGINS_GCS_GCS_HELPER_H_ + +#include +#include + +class TempFile : public std::fstream { + public: + // We should specify openmode each time we call TempFile. + TempFile(const char* temp_file_name, std::ios::openmode mode); + TempFile(TempFile&& rhs); + ~TempFile(); + const std::string getName() const; + + private: + const std::string name; +}; + +#endif // TENSORFLOW_C_EXPERIMENTAL_FILESYSTEM_PLUGINS_GCS_GCS_HELPER_H_ From 18f54c42c62191b60edeff7e1308ed7b3305b0eb Mon Sep 17 00:00:00 2001 From: Srinivas Vasudevan Date: Thu, 18 Jun 2020 09:45:59 -0700 Subject: [PATCH 186/867] Disable special_math_ops_test on asan build. PiperOrigin-RevId: 317121277 Change-Id: I902f6c046c2766638b5dbbb90a0ec2942a3c536f --- tensorflow/python/BUILD | 1 + 1 file changed, 1 insertion(+) diff --git a/tensorflow/python/BUILD b/tensorflow/python/BUILD index f53859b2915..a4e72bf2460 100644 --- a/tensorflow/python/BUILD +++ b/tensorflow/python/BUILD @@ -5277,6 +5277,7 @@ cuda_py_test( shard_count = 10, tags = [ "no_windows_gpu", + "noasan", # b/159332048 "nomsan", # b/148630708 ], deps = [ From 89b80c5fb98ba57175626a9df490dbdd2bc8776b Mon Sep 17 00:00:00 2001 From: Srinivas Vasudevan Date: Thu, 18 Jun 2020 09:59:14 -0700 Subject: [PATCH 187/867] Add banded triangular solve op. PiperOrigin-RevId: 317124054 Change-Id: I54f090d7583b21fa18788a2deb02262d9c8231be --- .../api_def_BandedTriangularSolve.pbtxt | 4 + .../api_def_BandedTriangularSolve.pbtxt | 4 + tensorflow/core/kernels/BUILD | 27 ++ .../kernels/banded_triangular_solve_op.cc | 293 ++++++++++++++++++ .../banded_triangular_solve_op_test.cc | 180 +++++++++++ tensorflow/core/ops/linalg_ops.cc | 54 ++++ tensorflow/python/kernel_tests/BUILD | 11 + .../banded_triangular_solve_op_test.py | 232 ++++++++++++++ .../python/kernel_tests/linalg_grad_test.py | 52 ++++ tensorflow/python/ops/linalg/linalg_impl.py | 96 ++++++ tensorflow/python/ops/linalg_grad.py | 33 ++ .../api/golden/v1/tensorflow.raw_ops.pbtxt | 4 + .../api/golden/v2/tensorflow.linalg.pbtxt | 4 + .../api/golden/v2/tensorflow.raw_ops.pbtxt | 4 + 14 files changed, 998 insertions(+) create mode 100644 tensorflow/core/api_def/base_api/api_def_BandedTriangularSolve.pbtxt create mode 100644 tensorflow/core/api_def/python_api/api_def_BandedTriangularSolve.pbtxt create mode 100644 tensorflow/core/kernels/banded_triangular_solve_op.cc create mode 100644 tensorflow/core/kernels/banded_triangular_solve_op_test.cc create mode 100644 tensorflow/python/kernel_tests/banded_triangular_solve_op_test.py diff --git a/tensorflow/core/api_def/base_api/api_def_BandedTriangularSolve.pbtxt b/tensorflow/core/api_def/base_api/api_def_BandedTriangularSolve.pbtxt new file mode 100644 index 00000000000..ba5e1bdcaf2 --- /dev/null +++ b/tensorflow/core/api_def/base_api/api_def_BandedTriangularSolve.pbtxt @@ -0,0 +1,4 @@ +op { + graph_op_name: "BandedTriangularSolve" + visibility: HIDDEN +} diff --git a/tensorflow/core/api_def/python_api/api_def_BandedTriangularSolve.pbtxt b/tensorflow/core/api_def/python_api/api_def_BandedTriangularSolve.pbtxt new file mode 100644 index 00000000000..ba5e1bdcaf2 --- /dev/null +++ b/tensorflow/core/api_def/python_api/api_def_BandedTriangularSolve.pbtxt @@ -0,0 +1,4 @@ +op { + graph_op_name: "BandedTriangularSolve" + visibility: HIDDEN +} diff --git a/tensorflow/core/kernels/BUILD b/tensorflow/core/kernels/BUILD index b2b54adbcf9..b4730dad96c 100644 --- a/tensorflow/core/kernels/BUILD +++ b/tensorflow/core/kernels/BUILD @@ -3577,6 +3577,7 @@ tf_cc_tests( cc_library( name = "linalg", deps = [ + ":banded_triangular_solve_op", ":cholesky_grad", ":cholesky_op", ":determinant_op", @@ -3750,6 +3751,12 @@ tf_kernel_library( deps = LINALG_DEPS, ) +tf_kernel_library( + name = "banded_triangular_solve_op", + prefix = "banded_triangular_solve_op", + deps = LINALG_DEPS + [":fill_functor"], +) + tf_kernel_library( name = "matrix_triangular_solve_op", hdrs = ["matrix_triangular_solve_op_impl.h"], @@ -4425,6 +4432,26 @@ tf_cuda_cc_test( ], ) +tf_cuda_cc_test( + name = "banded_triangular_solve_op_test", + size = "small", + srcs = ["banded_triangular_solve_op_test.cc"], + deps = [ + ":banded_triangular_solve_op", + ":matrix_set_diag_op", + ":matrix_triangular_solve_op", + ":ops_testutil", + ":ops_util", + "//tensorflow/core:core_cpu", + "//tensorflow/core:framework", + "//tensorflow/core:lib", + "//tensorflow/core:protos_all_cc", + "//tensorflow/core:test", + "//tensorflow/core:test_main", + "//tensorflow/core:testlib", + ], +) + tf_cuda_cc_test( name = "matrix_triangular_solve_op_test", size = "small", diff --git a/tensorflow/core/kernels/banded_triangular_solve_op.cc b/tensorflow/core/kernels/banded_triangular_solve_op.cc new file mode 100644 index 00000000000..d01a015502a --- /dev/null +++ b/tensorflow/core/kernels/banded_triangular_solve_op.cc @@ -0,0 +1,293 @@ +/* Copyright 2015 The TensorFlow Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. +==============================================================================*/ + +// See docs in ../ops/linalg_ops.cc. + +#include "third_party/eigen3/Eigen/Core" +#include "tensorflow/core/framework/kernel_def_builder.h" +#include "tensorflow/core/framework/op_kernel.h" +#include "tensorflow/core/framework/tensor_shape.h" +#include "tensorflow/core/kernels/fill_functor.h" +#include "tensorflow/core/kernels/linalg_ops_common.h" +#include "tensorflow/core/lib/core/errors.h" +#include "tensorflow/core/platform/logging.h" +#include "tensorflow/core/platform/macros.h" +#include "tensorflow/core/platform/types.h" +#include "tensorflow/core/util/matmul_bcast.h" + +namespace tensorflow { + +typedef Eigen::ThreadPoolDevice CPUDevice; + +template +Scalar eigen_conj(const Scalar& scalar) { + return Eigen::numext::conj(scalar); +} + +// Sequential batch matrix triangular solve kernel that calls Eigen's +// matrix triangular solve. +template +struct SequentialBandedTriangularSolveKernel { + using Matrix = + Eigen::Matrix; + using ConstMatrixMap = Eigen::Map; + using MatrixMap = Eigen::Map; + using RealScalar = typename Eigen::NumTraits::Real; + + static ConstMatrixMap ConstTensorSliceToEigenMatrix(const Tensor& t, + int slice) { + return ConstMatrixMap( + t.flat().data() + slice * t.dim_size(1) * t.dim_size(2), + t.dim_size(1), t.dim_size(2)); + } + + static MatrixMap TensorSliceToEigenMatrix(Tensor* t, int slice) { + return MatrixMap( + t->flat().data() + slice * t->dim_size(1) * t->dim_size(2), + t->dim_size(1), t->dim_size(2)); + } + + static void Run(const Tensor& in_x, const Tensor& in_y, bool lower, + bool adjoint, const MatMulBCast& bcast, Tensor* out, + int start, int limit) { + const bool should_bcast = bcast.IsBroadcastingRequired(); + const auto& x_batch_indices = bcast.x_batch_indices(); + const auto& y_batch_indices = bcast.y_batch_indices(); + int num_bands = in_x.dim_size(1); + int matrix_size = in_x.dim_size(2); + + for (int64 i = start; i < limit; ++i) { + const int64 x_batch_index = should_bcast ? x_batch_indices[i] : i; + const int64 y_batch_index = should_bcast ? y_batch_indices[i] : i; + auto matrix = ConstTensorSliceToEigenMatrix(in_x, x_batch_index); + auto rhs = ConstTensorSliceToEigenMatrix(in_y, y_batch_index); + auto output = TensorSliceToEigenMatrix(out, i); + // Below, we use the standard algorithm for computing a triangular solve, + // except we band limit it. + // Given A x = b, where A is lower triangular, + // x_i = (b_i - sum a_ij * x_j) / a_ii, where the sum is from + // j = 0 to i - 1. + // + // Now, in a banded triangular matrix, when i exceeds the band size, + // then the sum goes from j = i - band_size to i - 1, since the other + // elements are zero. + // + // Finally, given the band storage format, we'll need to change the + // indexing. + if (lower) { + if (!adjoint) { + output.row(0) = rhs.row(0) / matrix(0, 0); + for (int i = 1; i < matrix_size; ++i) { + if (i < num_bands) { + output.row(i).noalias() = + (rhs.row(i) - matrix.block(1, i, i, 1).reverse().transpose() * + output.topRows(i)) / + matrix(0, i); + } else { + output.row(i).noalias() = + (rhs.row(i) - + matrix.block(1, i, num_bands - 1, 1).reverse().transpose() * + output.middleRows(i - (num_bands - 1), num_bands - 1)) / + matrix(0, i); + } + } + } else { + // In the adjoint case, here and below, we now have an upper (lower) + // triangular matrix, and thus need to work through with the other + // case. We can't simply conjugate `matrix` and use the upper (lower) + // algorithm because the band storage format for upper and lower + // triangular matrices are different (in the lower case, we pad + // entries on the left, and in the upper case we pad entries on the + // right. + output.row(matrix_size - 1) = + rhs.row(matrix_size - 1) / eigen_conj(matrix(0, matrix_size - 1)); + for (int i = matrix_size - 1; i >= 0; --i) { + output.row(i).noalias() = rhs.row(i); + for (int j = i + 1; j < std::min(matrix_size, i + num_bands); ++j) { + output.row(i).noalias() -= + eigen_conj(matrix(j - i, j)) * output.row(j); + } + output.row(i) /= eigen_conj(matrix(0, i)); + } + } + } else { + if (!adjoint) { + output.row(matrix_size - 1) = + rhs.row(matrix_size - 1) / matrix(num_bands - 1, matrix_size - 1); + for (int i = 1; i < matrix_size; ++i) { + int k = matrix_size - 1 - i; + if (i < num_bands) { + output.row(k).noalias() = + (rhs.row(k) - matrix.block(num_bands - 1 - i, k, i, 1) + .reverse() + .transpose() * + output.bottomRows(i)) / + matrix(num_bands - 1, k); + } else { + output.row(k).noalias() = + (rhs.row(k) - + matrix.block(0, k, num_bands - 1, 1).reverse().transpose() * + output.middleRows(k + 1, num_bands - 1)) / + matrix(num_bands - 1, k); + } + } + } else { + output.row(0) = rhs.row(0) / eigen_conj(matrix(num_bands - 1, 0)); + for (int i = 1; i < matrix_size; ++i) { + output.row(i).noalias() = rhs.row(i); + for (int j = std::max(0, i - (num_bands - 1)); j < i; ++j) { + output.row(i).noalias() -= + eigen_conj(matrix(num_bands - 1 - (i - j), j)) * + output.row(j); + } + output.row(i) /= eigen_conj(matrix(num_bands - 1, i)); + } + } + } + } + } +}; + +template +struct LaunchBatchBandedTriangularSolve; + +template +struct LaunchBatchBandedTriangularSolve { + static void Launch(OpKernelContext* context, const Tensor& in_x, + const Tensor& in_y, bool adjoint, bool lower, + const MatMulBCast& bcast, Tensor* out) { + // Number of banded matrix triangular solves i.e. size of the batch. + const int64 batch_size = bcast.output_batch_size(); + const int64 cost_per_unit = + in_x.dim_size(1) * in_x.dim_size(2) * in_y.dim_size(2); + auto worker_threads = *(context->device()->tensorflow_cpu_worker_threads()); + + using Matrix = + Eigen::Matrix; + using ConstMatrixMap = Eigen::Map; + using RealScalar = typename Eigen::NumTraits::Real; + // Check diagonal before doing any solves. This is the first row in the + // lower case and else is the last row. + auto matrix = ConstMatrixMap(in_x.flat().data(), in_x.dim_size(1), + in_x.dim_size(2)); + RealScalar min_abs_pivot; + if (lower) { + min_abs_pivot = matrix.row(0).cwiseAbs().minCoeff(); + } else { + min_abs_pivot = matrix.row(in_x.dim_size(1) - 1).cwiseAbs().minCoeff(); + } + OP_REQUIRES(context, min_abs_pivot > RealScalar(0), + errors::InvalidArgument("Input matrix is not invertible.")); + + Shard(worker_threads.num_threads, worker_threads.workers, batch_size, + cost_per_unit, + [&in_x, &in_y, adjoint, lower, &bcast, out](int start, int limit) { + SequentialBandedTriangularSolveKernel::Run( + in_x, in_y, lower, adjoint, bcast, out, start, limit); + }); + } +}; + +template +class BandedTriangularSolveOpCpu : public OpKernel { + public: + explicit BandedTriangularSolveOpCpu(OpKernelConstruction* context) + : OpKernel(context) { + OP_REQUIRES_OK(context, context->GetAttr("lower", &lower_)); + OP_REQUIRES_OK(context, context->GetAttr("adjoint", &adjoint_)); + } + + ~BandedTriangularSolveOpCpu() override {} + + void Compute(OpKernelContext* ctx) override { + const Tensor& in0 = ctx->input(0); + const Tensor& in1 = ctx->input(1); + + ValidateInputTensors(ctx, in0, in1); + + MatMulBCast bcast(in0.shape().dim_sizes(), in1.shape().dim_sizes()); + OP_REQUIRES( + ctx, bcast.IsValid(), + errors::InvalidArgument( + "In[0] and In[1] must have compatible batch dimensions: ", + in0.shape().DebugString(), " vs. ", in1.shape().DebugString())); + + TensorShape out_shape = bcast.output_batch_shape(); + auto batch_size = bcast.output_batch_size(); + auto d0 = in0.dim_size(in0.dims() - 2); // Band size. + auto d1 = in0.dim_size(in0.dims() - 1); + Tensor in0_reshaped; + OP_REQUIRES( + ctx, + in0_reshaped.CopyFrom(in0, TensorShape({bcast.x_batch_size(), d0, d1})), + errors::Internal("Failed to reshape In[0] from ", + in0.shape().DebugString())); + auto d2 = in1.dim_size(in1.dims() - 2); + auto d3 = in1.dim_size(in1.dims() - 1); + Tensor in1_reshaped; + OP_REQUIRES( + ctx, + in1_reshaped.CopyFrom(in1, TensorShape({bcast.y_batch_size(), d2, d3})), + errors::Internal("Failed to reshape In[1] from ", + in1.shape().DebugString())); + OP_REQUIRES(ctx, d1 == d2, + errors::InvalidArgument( + "In[0] mismatch In[1] shape: ", d1, " vs. ", d2, ": ", + in0.shape().DebugString(), " ", in1.shape().DebugString(), + " ", lower_, " ", adjoint_)); + out_shape.AddDim(d1); + out_shape.AddDim(d3); + Tensor* out = nullptr; + OP_REQUIRES_OK(ctx, ctx->allocate_output(0, out_shape, &out)); + if (out->NumElements() == 0) { + return; + } + Tensor out_reshaped; + OP_REQUIRES(ctx, + out_reshaped.CopyFrom(*out, TensorShape({batch_size, d1, d3})), + errors::Internal("Failed to reshape output from ", + out->shape().DebugString())); + LaunchBatchBandedTriangularSolve::Launch( + ctx, in0_reshaped, in1_reshaped, adjoint_, lower_, bcast, + &out_reshaped); + } + + private: + void ValidateInputTensors(OpKernelContext* ctx, const Tensor& in0, + const Tensor& in1) { + OP_REQUIRES( + ctx, in0.dims() >= 2, + errors::InvalidArgument("In[0] ndims must be >= 2: ", in0.dims())); + + OP_REQUIRES( + ctx, in1.dims() >= 2, + errors::InvalidArgument("In[1] ndims must be >= 2: ", in1.dims())); + } + bool lower_; + bool adjoint_; +}; + +#define REGISTER_BANDED_TRIANGULAR_SOLVE_CPU(TYPE) \ + REGISTER_KERNEL_BUILDER(Name("BandedTriangularSolve") \ + .Device(DEVICE_CPU) \ + .TypeConstraint("T"), \ + BandedTriangularSolveOpCpu); + +REGISTER_BANDED_TRIANGULAR_SOLVE_CPU(float); +REGISTER_BANDED_TRIANGULAR_SOLVE_CPU(double); +REGISTER_BANDED_TRIANGULAR_SOLVE_CPU(complex64); +REGISTER_BANDED_TRIANGULAR_SOLVE_CPU(complex128); + +} // namespace tensorflow diff --git a/tensorflow/core/kernels/banded_triangular_solve_op_test.cc b/tensorflow/core/kernels/banded_triangular_solve_op_test.cc new file mode 100644 index 00000000000..37e904a3e0e --- /dev/null +++ b/tensorflow/core/kernels/banded_triangular_solve_op_test.cc @@ -0,0 +1,180 @@ +/* Copyright 2019 The TensorFlow Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. +==============================================================================*/ + +#include "tensorflow/core/common_runtime/kernel_benchmark_testlib.h" +#include "tensorflow/core/framework/op.h" +#include "tensorflow/core/framework/tensor.h" +#include "tensorflow/core/framework/tensor_shape.h" +#include "tensorflow/core/framework/types.pb.h" +#include "tensorflow/core/graph/graph.h" +#include "tensorflow/core/graph/node_builder.h" +#include "tensorflow/core/graph/testlib.h" +#include "tensorflow/core/kernels/matrix_set_diag_op.h" +#include "tensorflow/core/lib/core/status.h" +#include "tensorflow/core/platform/test.h" +#include "tensorflow/core/platform/test_benchmark.h" + +namespace tensorflow { +namespace { + +Node* SetDiag(int num_bands, Graph* g, Node* bands, Node* triangular) { + Node* ret; + Tensor bandwidth(DT_INT32, TensorShape({2})); + bandwidth.flat()(0) = -(num_bands - 1); + bandwidth.flat()(1) = 0; + TF_CHECK_OK(NodeBuilder(g->NewName("n"), "MatrixSetDiagV3") + .Input(triangular) + .Input(bands) + .Input(test::graph::Constant(g, bandwidth)) + .Attr("align", "RIGHT_LEFT") + .Finalize(g, &ret)); + return ret; +} + +Node* BandedTriangularSolve(Graph* g, Node* in0, Node* in1) { + Node* ret; + TF_CHECK_OK(NodeBuilder(g->NewName("n"), "BandedTriangularSolve") + .Input(in0) + .Input(in1) + .Attr("lower", true) + .Attr("adjoint", false) + .Finalize(g, &ret)); + return ret; +} + +Node* MatrixTriangularSolve(Graph* g, Node* in0, Node* in1) { + Node* ret; + TF_CHECK_OK(NodeBuilder(g->NewName("n"), "MatrixTriangularSolve") + .Input(in0) + .Input(in1) + .Attr("lower", true) + .Attr("adjoint", false) + .Finalize(g, &ret)); + return ret; +} + +template +static Graph* BandedTriangularSolve(int64 num_bands, int64 n, int64 m, + bool use_banded_solver, DataType type) { + Graph* g = new Graph(OpRegistry::Global()); + Tensor in0(type, TensorShape({num_bands, n})); + // Set diagonal to nonzero to guarantee invertibility. + in0.flat().setRandom(); + in0.flat() = + in0.flat().abs() + in0.flat().constant(static_cast(0.5)); + Tensor in1(type, TensorShape({n, m})); + in1.flat().setRandom(); + if (use_banded_solver) { + BandedTriangularSolve(g, test::graph::Constant(g, in0), + test::graph::Constant(g, in1)); + } else { + // Create a zero tensor. + Tensor in2(type, TensorShape({n, n})); + in2.flat().setZero(); + Node* triangular_matrix = + SetDiag(num_bands, g, test::graph::Constant(g, in0), + test::graph::Constant(g, in2)); + MatrixTriangularSolve(g, triangular_matrix, test::graph::Constant(g, in1)); + } + return g; +} + +// Macro arguments names: --------------------------------------------------- // +// K: Number of bands +// N: Inner dimension of LHS, Inner dimension of RHS. +// M: Outer dimensions of RHS +// BS: boolean indicating whether to use the banded solver +// T: C++ type of scalars (e.g. float, std::complex) +// TT: TensorFlow type of scalars (e.g. DT_FLOAT, DT_COMPLEX128 +#define BM_BandedTriangularSolveDev(K, N, M, BS, T, TT, D) \ + static void BM_BandedTriangularSolve##_##K##_##N##_##M##_##BS##_##TT( \ + int iters) { \ + testing::UseRealTime(); \ + testing::ItemsProcessed(static_cast(iters) * K * N + N * M); \ + test::Benchmark(#D, BandedTriangularSolve(K, N, M, BS, TT)).Run(iters); \ + } \ + BENCHMARK(BM_BandedTriangularSolve##_##K##_##N##_##M##_##BS##_##TT); + +#define BM_BandedTriangularSolve(K, N, M, BS, D) \ + BM_BandedTriangularSolveDev(K, N, M, BS, float, DT_FLOAT, D); \ + BM_BandedTriangularSolveDev(K, N, M, BS, double, DT_DOUBLE, D); + +// Small number of bands, few rhs +BM_BandedTriangularSolve(2, 32, 1, true, cpu); +BM_BandedTriangularSolve(2, 32, 1, false, cpu); +BM_BandedTriangularSolve(4, 32, 1, true, cpu); +BM_BandedTriangularSolve(4, 32, 1, false, cpu); +BM_BandedTriangularSolve(8, 32, 1, true, cpu); +BM_BandedTriangularSolve(8, 32, 1, false, cpu); +BM_BandedTriangularSolve(16, 32, 1, true, cpu); +BM_BandedTriangularSolve(16, 32, 1, false, cpu); +BM_BandedTriangularSolve(2, 128, 1, true, cpu); +BM_BandedTriangularSolve(2, 128, 1, false, cpu); +BM_BandedTriangularSolve(4, 128, 1, true, cpu); +BM_BandedTriangularSolve(4, 128, 1, false, cpu); +BM_BandedTriangularSolve(8, 128, 1, true, cpu); +BM_BandedTriangularSolve(8, 128, 1, false, cpu); +BM_BandedTriangularSolve(16, 128, 1, true, cpu); +BM_BandedTriangularSolve(16, 128, 1, false, cpu); +BM_BandedTriangularSolve(2, 512, 1, true, cpu); +BM_BandedTriangularSolve(2, 512, 1, false, cpu); +BM_BandedTriangularSolve(4, 512, 1, true, cpu); +BM_BandedTriangularSolve(4, 512, 1, false, cpu); +BM_BandedTriangularSolve(8, 512, 1, true, cpu); +BM_BandedTriangularSolve(8, 512, 1, false, cpu); +BM_BandedTriangularSolve(16, 512, 1, true, cpu); +BM_BandedTriangularSolve(16, 512, 1, false, cpu); + +// Larger # rhs +BM_BandedTriangularSolve(2, 32, 32, true, cpu); +BM_BandedTriangularSolve(2, 32, 32, false, cpu); +BM_BandedTriangularSolve(4, 32, 32, true, cpu); +BM_BandedTriangularSolve(4, 32, 32, false, cpu); +BM_BandedTriangularSolve(8, 32, 32, true, cpu); +BM_BandedTriangularSolve(8, 32, 32, false, cpu); +BM_BandedTriangularSolve(16, 32, 32, true, cpu); +BM_BandedTriangularSolve(16, 32, 32, false, cpu); +BM_BandedTriangularSolve(2, 128, 128, true, cpu); +BM_BandedTriangularSolve(2, 128, 128, false, cpu); +BM_BandedTriangularSolve(4, 128, 128, true, cpu); +BM_BandedTriangularSolve(4, 128, 128, false, cpu); +BM_BandedTriangularSolve(8, 128, 128, true, cpu); +BM_BandedTriangularSolve(8, 128, 128, false, cpu); +BM_BandedTriangularSolve(16, 128, 128, true, cpu); +BM_BandedTriangularSolve(16, 128, 128, false, cpu); +BM_BandedTriangularSolve(2, 512, 512, true, cpu); +BM_BandedTriangularSolve(2, 512, 512, false, cpu); +BM_BandedTriangularSolve(4, 512, 512, true, cpu); +BM_BandedTriangularSolve(4, 512, 512, false, cpu); +BM_BandedTriangularSolve(8, 512, 512, true, cpu); +BM_BandedTriangularSolve(8, 512, 512, false, cpu); +BM_BandedTriangularSolve(16, 512, 512, true, cpu); +BM_BandedTriangularSolve(16, 512, 512, false, cpu); + +BM_BandedTriangularSolve(2, 2048, 2048, true, cpu); +BM_BandedTriangularSolve(2, 2048, 2048, false, cpu); +BM_BandedTriangularSolve(4, 2048, 2048, true, cpu); +BM_BandedTriangularSolve(4, 2048, 2048, false, cpu); +BM_BandedTriangularSolve(8, 2048, 2048, true, cpu); +BM_BandedTriangularSolve(8, 2048, 2048, false, cpu); +BM_BandedTriangularSolve(16, 2048, 2048, true, cpu); +BM_BandedTriangularSolve(16, 2048, 2048, false, cpu); +BM_BandedTriangularSolve(32, 2048, 2048, true, cpu); +BM_BandedTriangularSolve(32, 2048, 2048, false, cpu); +BM_BandedTriangularSolve(64, 2048, 2048, true, cpu); +BM_BandedTriangularSolve(64, 2048, 2048, false, cpu); + +} // namespace +} // namespace tensorflow diff --git a/tensorflow/core/ops/linalg_ops.cc b/tensorflow/core/ops/linalg_ops.cc index 75340b28eb0..a05231834b7 100644 --- a/tensorflow/core/ops/linalg_ops.cc +++ b/tensorflow/core/ops/linalg_ops.cc @@ -47,6 +47,49 @@ Status BatchUnchangedSquareShapeFn(InferenceContext* c) { return Status::OK(); } +// The first input is [...,K,M] and second input is [...,M,N]. +Status BandedTriangularSolveShapeFn(InferenceContext* c) { + ShapeHandle lhs; + ShapeHandle rhs; + + TF_RETURN_IF_ERROR(c->WithRankAtLeast(c->input(0), 2, &lhs)); + TF_RETURN_IF_ERROR(c->WithRankAtLeast(c->input(1), 2, &rhs)); + + // Check K > 0. + DimensionHandle num_bands = c->Dim(lhs, -2); + DimensionHandle m = c->Dim(lhs, -1); + if (c->ValueKnown(num_bands) && c->Value(num_bands) <= 0) { + return errors::InvalidArgument("Number of bands must be positive, but is ", + c->Value(num_bands)); + } + if (c->ValueKnown(num_bands) && c->ValueKnown(m) && + c->Value(num_bands) > c->Value(m)) { + return errors::InvalidArgument("Number of bands ", c->Value(num_bands), + " cannot exceed the size of the matrix ", + c->Value(m)); + } + + ShapeHandle lhs_batch_shape; + ShapeHandle rhs_batch_shape; + ShapeHandle output_batch_shape; + // Make the common batch subshape. + TF_RETURN_IF_ERROR(c->Subshape(lhs, 0, -2, &lhs_batch_shape)); + TF_RETURN_IF_ERROR(c->Subshape(rhs, 0, -2, &rhs_batch_shape)); + TF_RETURN_IF_ERROR(BroadcastBinaryOpOutputShapeFnHelper( + c, lhs_batch_shape, rhs_batch_shape, true, &output_batch_shape)); + + // lhs and rhs have the same value for M to be compatible. + TF_RETURN_IF_ERROR(c->Merge(m, c->Dim(rhs, -2), &m)); + + // Build final shape (batch_shape + m + n) in . + ShapeHandle out; + TF_RETURN_IF_ERROR( + c->Concatenate(output_batch_shape, c->Matrix(m, c->Dim(rhs, -1)), &out)); + + c->set_output(0, out); + return Status::OK(); +} + // The first input is [...,M,N] and second input is either [...,M,K] or [...,M]. // Output is [...,N,K] or [...,N]. If , then input is [...,M,M]. Status MatrixSolveShapeFn(InferenceContext* c, bool square) { @@ -446,6 +489,17 @@ REGISTER_OP("MatrixSolve") return MatrixSolveShapeFn(c, true /* square (*/); }); +REGISTER_OP("BandedTriangularSolve") + .Input("matrix: T") + .Input("rhs: T") + .Output("output: T") + .Attr("lower: bool = True") + .Attr("adjoint: bool = False") + .Attr("T: {double, float, half, complex64, complex128}") + .SetShapeFn([](InferenceContext* c) { + return BandedTriangularSolveShapeFn(c); + }); + REGISTER_OP("MatrixTriangularSolve") .Input("matrix: T") .Input("rhs: T") diff --git a/tensorflow/python/kernel_tests/BUILD b/tensorflow/python/kernel_tests/BUILD index 846c582737f..f2c614974f5 100644 --- a/tensorflow/python/kernel_tests/BUILD +++ b/tensorflow/python/kernel_tests/BUILD @@ -762,6 +762,17 @@ cuda_py_test( ], ) +cuda_py_test( + name = "banded_triangular_solve_op_test", + size = "small", + srcs = ["banded_triangular_solve_op_test.py"], + deps = [ + "//tensorflow/python:client_testlib", + "//tensorflow/python:linalg_ops", + "//third_party/py/numpy", + ], +) + cuda_py_test( name = "matrix_triangular_solve_op_test", size = "medium", diff --git a/tensorflow/python/kernel_tests/banded_triangular_solve_op_test.py b/tensorflow/python/kernel_tests/banded_triangular_solve_op_test.py new file mode 100644 index 00000000000..bd0fdae03c5 --- /dev/null +++ b/tensorflow/python/kernel_tests/banded_triangular_solve_op_test.py @@ -0,0 +1,232 @@ +# Copyright 2015 The TensorFlow Authors. All Rights Reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. +# ============================================================================== +"""Tests for tensorflow.ops.math_ops.banded_triangular_solve.""" + +from __future__ import absolute_import +from __future__ import division +from __future__ import print_function + +import numpy as np + +from tensorflow.python.framework import test_util +from tensorflow.python.ops import array_ops +from tensorflow.python.ops import linalg_ops +from tensorflow.python.platform import test + + +class BandedTriangularSolveOpTest(test.TestCase): + + def _verifySolveAllWays(self, x, y, dtypes, batch_dims=None): + for lower in (False,): + for adjoint in (False, True): + for use_placeholder in True, False: + self._verifySolve( + x, + y, + lower=lower, + adjoint=adjoint, + batch_dims=batch_dims, + use_placeholder=use_placeholder, + dtypes=dtypes) + + def _verifySolveAllWaysReal(self, x, y, batch_dims=None): + self._verifySolveAllWays(x, y, (np.float32, np.float64), batch_dims) + + def _verifySolveAllWaysComplex(self, x, y, batch_dims=None): + self._verifySolveAllWays(x, y, (np.complex64, np.complex128), batch_dims) + + def _verifySolve(self, + x, + y, + lower=True, + adjoint=False, + batch_dims=None, + use_placeholder=False, + dtypes=(np.float32, np.float64)): + for np_type in dtypes: + a = x.astype(np_type) + b = y.astype(np_type) + + # Now we need to convert a to a dense triangular matrix. + def make_diags(diags, lower=True): + n = len(diags[0]) + a = np.zeros(n * n, dtype=diags.dtype) + if lower: + for i, diag in enumerate(diags): + a[n * i:n * n:n + 1] = diag[i:] + else: + diags_flip = np.flip(diags, 0) + for i, diag in enumerate(diags_flip): + a[i:(n - i) * n:n + 1] = diag[:(n - i)] + return a.reshape(n, n) + + # For numpy.solve we have to explicitly zero out the strictly + # upper or lower triangle. + if a.size > 0: + a_np = make_diags(a, lower=lower) + else: + a_np = a + if adjoint: + a_np = np.conj(np.transpose(a_np)) + + if batch_dims is not None: + a = np.tile(a, batch_dims + [1, 1]) + a_np = np.tile(a_np, batch_dims + [1, 1]) + b = np.tile(b, batch_dims + [1, 1]) + + with self.cached_session(use_gpu=True): + a_tf = a + b_tf = b + if use_placeholder: + a_tf = array_ops.placeholder_with_default(a_tf, shape=None) + b_tf = array_ops.placeholder_with_default(b_tf, shape=None) + tf_ans = linalg_ops.banded_triangular_solve( + a_tf, b_tf, lower=lower, adjoint=adjoint) + tf_val = self.evaluate(tf_ans) + np_ans = np.linalg.solve(a_np, b) + self.assertEqual(np_ans.shape, tf_val.shape) + self.assertAllClose(np_ans, tf_val) + + @test_util.run_deprecated_v1 + def testSolve(self): + # 1x1 matrix, single rhs. + matrix = np.array([[0.1]]) + rhs0 = np.array([[1.]]) + self._verifySolveAllWaysReal(matrix, rhs0) + # 2x2 matrix with 2 bands, single right-hand side. + # Corresponds to the lower triangular + # [[1., 0.], [3., 4.]] + # and upper triangular + # [[2., 1.], [0., 3.]] + matrix = np.array([[1., 4.], [2., 3.]]) + rhs0 = np.array([[1.], [1.]]) + self._verifySolveAllWaysReal(matrix, rhs0) + # 2x2 matrix with 2 bands, 3 right-hand sides. + rhs1 = np.array([[1., 0., 1.], [0., 1., 1.]]) + self._verifySolveAllWaysReal(matrix, rhs1) + # 4 x 4 matrix with 2 bands, 3 right hand sides. + # Corresponds to the lower triangular + # [[1., 0., 0., 0.], + # [-1., 2., 0., 0.], + # [0., -2., 3., 0.], + # [0., 0., -3., 4.]] + # and upper triangular + # [[1., 1., 0., 0.], + # [0., -1., 2., 0.], + # [0., 0., -2., 3.], + # [0., 0., 0., -3.]] + matrix = np.array([[1., 2., 3., 4.], [1., -1., -2., -3.]]) + rhs0 = np.array([[1., 0., 1.], [0., 1., 1.], [-1., 2., 1.], [0., -1., -1.]]) + self._verifySolveAllWaysReal(matrix, rhs0) + + def testSolveBandSizeSmaller(self): + rhs0 = np.random.randn(6, 4) + + # 6 x 6 matrix with 2 bands. Ensure all non-zero entries. + matrix = 2. * np.random.uniform(size=[3, 6]) + 1. + self._verifySolveAllWaysReal(matrix, rhs0) + + # 6 x 6 matrix with 3 bands. Ensure all non-zero entries. + matrix = 2. * np.random.uniform(size=[3, 6]) + 1. + self._verifySolveAllWaysReal(matrix, rhs0) + + @test_util.run_deprecated_v1 + def testSolveComplex(self): + if test.is_built_with_rocm(): + self.skipTest("ROCm does not support BLAS operations for complex types") + # 1x1 matrix, single rhs. + matrix = np.array([[0.1 + 1j * 0.1]]) + rhs0 = np.array([[1. + 1j]]) + self._verifySolveAllWaysComplex(matrix, rhs0) + # 2x2 matrix with 2 bands, single right-hand side. + # Corresponds to + # [[1. + 1j, 0.], [4 + 1j, 2 + 1j]] + matrix = np.array([[1., 2.], [3., 4.]]).astype(np.complex64) + matrix += 1j * matrix + rhs0 = np.array([[1.], [1.]]).astype(np.complex64) + rhs0 += 1j * rhs0 + self._verifySolveAllWaysComplex(matrix, rhs0) + # 2x2 matrix with 2 bands, 3 right-hand sides. + rhs1 = np.array([[1., 0., 1.], [0., 1., 1.]]).astype(np.complex64) + rhs1 += 1j * rhs1 + self._verifySolveAllWaysComplex(matrix, rhs1) + + @test_util.run_deprecated_v1 + def testSolveBatch(self): + matrix = np.array([[1., 2.], [3., 4.]]) + rhs = np.array([[1., 0., 1.], [0., 1., 1.]]) + # Batch of 2x3x2x2 matrices, 2x3x2x3 right-hand sides. + self._verifySolveAllWaysReal(matrix, rhs, batch_dims=[2, 3]) + # Batch of 3x2x2x2 matrices, 3x2x2x3 right-hand sides. + self._verifySolveAllWaysReal(matrix, rhs, batch_dims=[3, 2]) + + matrix = np.array([[1., 2., 3., 4.], [-1., -2., -3., -4.], + [-1., 1., 2., 3.]]) + rhs = np.array([[-1., 2.], [1., 1.], [0., 1.], [2., 3.]]) + # Batch of 2x3x4x4 matrices with 3 bands, 2x3x4x2 right-hand sides. + self._verifySolveAllWaysReal(matrix, rhs, batch_dims=[2, 3]) + # Batch of 3x2x4x4 matrices with 3 bands, 3x2x4x2 right-hand sides. + self._verifySolveAllWaysReal(matrix, rhs, batch_dims=[3, 2]) + + @test_util.run_deprecated_v1 + def testSolveBatchComplex(self): + if test.is_built_with_rocm(): + self.skipTest("ROCm does not support BLAS operations for complex types") + matrix = np.array([[1., 2.], [3., 4.]]).astype(np.complex64) + matrix += 1j * matrix + rhs = np.array([[1., 0., 1.], [0., 1., 1.]]).astype(np.complex64) + rhs += 1j * rhs + # Batch of 2x3x2x2 matrices, 2x3x2x3 right-hand sides. + self._verifySolveAllWaysComplex(matrix, rhs, batch_dims=[2, 3]) + # Batch of 3x2x2x2 matrices, 3x2x2x3 right-hand sides. + self._verifySolveAllWaysComplex(matrix, rhs, batch_dims=[3, 2]) + + @test_util.run_deprecated_v1 + def testWrongDimensions(self): + # The matrix should have the same number of rows as the + # right-hand sides. + matrix = np.array([[1., 1.], [1., 1.]]) + rhs = np.array([[1., 0.]]) + with self.cached_session(use_gpu=True): + with self.assertRaises(ValueError): + self._verifySolve(matrix, rhs) + with self.assertRaises(ValueError): + self._verifySolve(matrix, rhs, batch_dims=[2, 3]) + + # Number of bands exceeds the dimension of the matrix. + matrix = np.ones((6, 4)) + rhs = np.ones((4, 2)) + with self.cached_session(use_gpu=True): + with self.assertRaises(ValueError): + self._verifySolve(matrix, rhs) + with self.assertRaises(ValueError): + self._verifySolve(matrix, rhs, batch_dims=[2, 3]) + + @test_util.run_deprecated_v1 + @test_util.disable_xla("XLA cannot throw assertion errors during a kernel.") + def testNotInvertible(self): + # The input should be invertible. + # The matrix is singular because it has a zero on the diagonal. + # FIXME(rmlarsen): The GPU kernel does not check for singularity. + singular_matrix = np.array([[1., 0., -1.], [-1., 0., 1.], [0., -1., 1.]]) + with self.cached_session(): + with self.assertRaisesOpError("Input matrix is not invertible."): + self._verifySolve(singular_matrix, singular_matrix) + with self.assertRaisesOpError("Input matrix is not invertible."): + self._verifySolve(singular_matrix, singular_matrix, batch_dims=[2, 3]) + + +if __name__ == "__main__": + test.main() diff --git a/tensorflow/python/kernel_tests/linalg_grad_test.py b/tensorflow/python/kernel_tests/linalg_grad_test.py index 36e58bee829..3aceddf4d5f 100644 --- a/tensorflow/python/kernel_tests/linalg_grad_test.py +++ b/tensorflow/python/kernel_tests/linalg_grad_test.py @@ -132,6 +132,44 @@ def _GetMatrixBinaryFunctorGradientTest(functor_, return Test +def _GetBandedTriangularSolveGradientTest( + functor_, + dtype_, + shape_, + float32_tol_fudge=1.0, # pylint: disable=redefined-outer-name + **kwargs_): + + @test_util.run_in_graph_and_eager_modes(use_gpu=True) + def Test(self): + n = shape_[-1] + + np.random.seed(1) + # Make sure invertible. + a_np = np.random.uniform(low=1.0, high=2.0, size=shape_).astype(dtype_) + a = constant_op.constant(a_np) + + b_np = np.random.uniform(low=-1.0, high=1.0, size=[n, n]).astype(dtype_) + b = constant_op.constant(b_np) + + epsilon = np.finfo(dtype_).eps + delta = epsilon**(1.0 / 3.0) + # tolerance obtained by looking at actual differences using + # np.linalg.norm(theoretical-numerical, np.inf) on -mavx build + tol = 1e-6 if dtype_ == np.float64 else float32_tol_fudge * 0.05 + + # check gradient w.r.t. left argument. + theoretical, numerical = gradient_checker_v2.compute_gradient( + lambda x: functor_(x, b, **kwargs_), [a], delta=delta) + self.assertAllClose(theoretical, numerical, atol=tol, rtol=tol) + + # check gradient w.r.t. right argument. + theoretical, numerical = gradient_checker_v2.compute_gradient( + lambda y: functor_(a, y, **kwargs_), [b], delta=delta) + self.assertAllClose(theoretical, numerical, atol=tol, rtol=tol) + + return Test + + if __name__ == '__main__': # Tests for gradients of binary matrix operations. for dtype in np.float32, np.float64: @@ -166,6 +204,20 @@ if __name__ == '__main__': adjoint=adjoint, lower=lower)) + band_shape = extra + (size // 2 + 1, size) + name = '%s_%s_adj_%s_low_%s' % (dtype.__name__, '_'.join( + map(str, band_shape)), str(adjoint), lower) + _AddTest( + MatrixBinaryFunctorGradientTest, + 'BandedTriangularSolveGradient', name, + _GetBandedTriangularSolveGradientTest( + linalg_ops.banded_triangular_solve, + dtype, + band_shape, + float32_tol_fudge=4.0, + adjoint=adjoint, + lower=lower)) + # Tests for gradients of unary matrix operations. for dtype in np.float32, np.float64: for size in 2, 5, 10: diff --git a/tensorflow/python/ops/linalg/linalg_impl.py b/tensorflow/python/ops/linalg/linalg_impl.py index 82acd09caec..9ddf7b5e8b8 100644 --- a/tensorflow/python/ops/linalg/linalg_impl.py +++ b/tensorflow/python/ops/linalg/linalg_impl.py @@ -340,6 +340,102 @@ def matrix_exponential(input, name=None): # pylint: disable=redefined-builtin return array_ops.reshape(result, batch_shape.concatenate(result.shape[-2:])) +@tf_export('linalg.banded_triangular_solve', v1=[]) +def banded_triangular_solve( + bands, + rhs, + lower=True, + adjoint=False, # pylint: disable=redefined-outer-name + name=None): + r"""Solve triangular systems of equations with a banded solver. + + `bands` is a tensor of shape `[..., K, M]`, where `K` represents the number + of bands stored. This corresponds to a batch of `M` by `M` matrices, whose + `K` subdiagonals (when `lower` is `True`) are stored. + + This operator broadcasts the batch dimensions of `bands` and the batch + dimensions of `rhs`. + + + Examples: + + Storing 2 bands of a 3x3 matrix. + Note that first element in the second row is ignored due to + the 'LEFT_RIGHT' padding. + + >>> x = [[2., 3., 4.], [1., 2., 3.]] + >>> x2 = [[2., 3., 4.], [10000., 2., 3.]] + >>> y = tf.zeros([3, 3]) + >>> z = tf.linalg.set_diag(y, x, align='LEFT_RIGHT', k=(-1, 0)) + >>> z + + >>> soln = tf.linalg.banded_triangular_solve(x, tf.ones([3, 1])) + >>> soln + + >>> are_equal = soln == tf.linalg.banded_triangular_solve(x2, tf.ones([3, 1])) + >>> tf.reduce_all(are_equal).numpy() + True + >>> are_equal = soln == tf.linalg.triangular_solve(z, tf.ones([3, 1])) + >>> tf.reduce_all(are_equal).numpy() + True + + Storing 2 superdiagonals of a 4x4 matrix. Because of the 'LEFT_RIGHT' padding + the last element of the first row is ignored. + + >>> x = [[2., 3., 4., 5.], [-1., -2., -3., -4.]] + >>> y = tf.zeros([4, 4]) + >>> z = tf.linalg.set_diag(y, x, align='LEFT_RIGHT', k=(0, 1)) + >>> z + + >>> soln = tf.linalg.banded_triangular_solve(x, tf.ones([4, 1]), lower=False) + >>> soln + + >>> are_equal = (soln == tf.linalg.triangular_solve( + ... z, tf.ones([4, 1]), lower=False)) + >>> tf.reduce_all(are_equal).numpy() + True + + + Args: + bands: A `Tensor` describing the bands of the left hand side, with shape + `[..., K, M]`. The `K` rows correspond to the diagonal to the `K - 1`-th + diagonal (the diagonal is the top row) when `lower` is `True` and + otherwise the `K - 1`-th superdiagonal to the diagonal (the diagonal is + the bottom row) when `lower` is `False`. The bands are stored with + 'LEFT_RIGHT' alignment, where the superdiagonals are padded on the right + and subdiagonals are padded on the left. This is the alignment cuSPARSE + uses. See `tf.linalg.set_diag` for more details. + rhs: A `Tensor` of shape [..., M] or [..., M, N] and with the same dtype as + `diagonals`. Note that if the shape of `rhs` and/or `diags` isn't known + statically, `rhs` will be treated as a matrix rather than a vector. + lower: An optional `bool`. Defaults to `True`. Boolean indicating whether + `bands` represents a lower or upper triangular matrix. + adjoint: An optional `bool`. Defaults to `False`. Boolean indicating whether + to solve with the matrix's block-wise adjoint. + name: A name to give this `Op` (optional). + + Returns: + A `Tensor` of shape [..., M] or [..., M, N] containing the solutions. + """ + with ops.name_scope(name, 'banded_triangular_solve', [bands, rhs]): + return gen_linalg_ops.banded_triangular_solve( + bands, rhs, lower=lower, adjoint=adjoint) + + @tf_export('linalg.tridiagonal_solve') @dispatch.add_dispatch_support def tridiagonal_solve(diagonals, diff --git a/tensorflow/python/ops/linalg_grad.py b/tensorflow/python/ops/linalg_grad.py index 8d3664144a1..437e28e7e6b 100644 --- a/tensorflow/python/ops/linalg_grad.py +++ b/tensorflow/python/ops/linalg_grad.py @@ -607,6 +607,39 @@ def _MatrixSolveLsGrad(op, grad): lambda: _Underdetermined(op, grad)) +@ops.RegisterGradient("BandedTriangularSolve") +def _BandedTriangularSolveGrad(op, grad): + """Gradient for BandedTriangularSolve.""" + a = op.inputs[0] + b = op.inputs[1] + num_bands = array_ops.shape(a)[-2] + adjoint_a = op.get_attr("adjoint") + lower_a = op.get_attr("lower") + c = op.outputs[0] + grad_b = linalg_ops.banded_triangular_solve( + a, grad, lower=lower_a, adjoint=not adjoint_a) + if adjoint_a: + grad_a = -math_ops.matmul(c, grad_b, adjoint_b=True) + else: + grad_a = -math_ops.matmul(grad_b, c, adjoint_b=True) + if lower_a: + grad_a = array_ops.matrix_diag_part( + grad_a, k=(-(num_bands - 1), 0), align="LEFT_RIGHT") + else: + grad_a = array_ops.matrix_diag_part( + grad_a, k=(0, num_bands - 1), align="LEFT_RIGHT") + # If the static batch shapes are equal, we don't need to unbroadcast. + if (a.shape.is_fully_defined() and b.shape.is_fully_defined() and + a.shape[:-2] == b.shape[:-2]): + return grad_a, grad_b + a_shape = array_ops.shape(a) + b_shape = array_ops.shape(b) + ra, rb = array_ops.broadcast_gradient_args(a_shape[:-2], b_shape[:-2]) + grad_a = array_ops.reshape(math_ops.reduce_sum(grad_a, axis=ra), a_shape) + grad_b = array_ops.reshape(math_ops.reduce_sum(grad_b, axis=rb), b_shape) + return grad_a, grad_b + + @ops.RegisterGradient("MatrixTriangularSolve") def _MatrixTriangularSolveGrad(op, grad): """Gradient for MatrixTriangularSolve.""" diff --git a/tensorflow/tools/api/golden/v1/tensorflow.raw_ops.pbtxt b/tensorflow/tools/api/golden/v1/tensorflow.raw_ops.pbtxt index 3d298e928e9..62969b5a0dd 100644 --- a/tensorflow/tools/api/golden/v1/tensorflow.raw_ops.pbtxt +++ b/tensorflow/tools/api/golden/v1/tensorflow.raw_ops.pbtxt @@ -284,6 +284,10 @@ tf_module { name: "AvgPoolGrad" argspec: "args=[\'orig_input_shape\', \'grad\', \'ksize\', \'strides\', \'padding\', \'data_format\', \'name\'], varargs=None, keywords=None, defaults=[\'NHWC\', \'None\'], " } + member_method { + name: "BandedTriangularSolve" + argspec: "args=[\'matrix\', \'rhs\', \'lower\', \'adjoint\', \'name\'], varargs=None, keywords=None, defaults=[\'True\', \'False\', \'None\'], " + } member_method { name: "Barrier" argspec: "args=[\'component_types\', \'shapes\', \'capacity\', \'container\', \'shared_name\', \'name\'], varargs=None, keywords=None, defaults=[\'[]\', \'-1\', \'\', \'\', \'None\'], " diff --git a/tensorflow/tools/api/golden/v2/tensorflow.linalg.pbtxt b/tensorflow/tools/api/golden/v2/tensorflow.linalg.pbtxt index 734837b99cb..4f62af20dc0 100644 --- a/tensorflow/tools/api/golden/v2/tensorflow.linalg.pbtxt +++ b/tensorflow/tools/api/golden/v2/tensorflow.linalg.pbtxt @@ -96,6 +96,10 @@ tf_module { name: "band_part" argspec: "args=[\'input\', \'num_lower\', \'num_upper\', \'name\'], varargs=None, keywords=None, defaults=[\'None\'], " } + member_method { + name: "banded_triangular_solve" + argspec: "args=[\'bands\', \'rhs\', \'lower\', \'adjoint\', \'name\'], varargs=None, keywords=None, defaults=[\'True\', \'False\', \'None\'], " + } member_method { name: "cholesky" argspec: "args=[\'input\', \'name\'], varargs=None, keywords=None, defaults=[\'None\'], " diff --git a/tensorflow/tools/api/golden/v2/tensorflow.raw_ops.pbtxt b/tensorflow/tools/api/golden/v2/tensorflow.raw_ops.pbtxt index 3d298e928e9..62969b5a0dd 100644 --- a/tensorflow/tools/api/golden/v2/tensorflow.raw_ops.pbtxt +++ b/tensorflow/tools/api/golden/v2/tensorflow.raw_ops.pbtxt @@ -284,6 +284,10 @@ tf_module { name: "AvgPoolGrad" argspec: "args=[\'orig_input_shape\', \'grad\', \'ksize\', \'strides\', \'padding\', \'data_format\', \'name\'], varargs=None, keywords=None, defaults=[\'NHWC\', \'None\'], " } + member_method { + name: "BandedTriangularSolve" + argspec: "args=[\'matrix\', \'rhs\', \'lower\', \'adjoint\', \'name\'], varargs=None, keywords=None, defaults=[\'True\', \'False\', \'None\'], " + } member_method { name: "Barrier" argspec: "args=[\'component_types\', \'shapes\', \'capacity\', \'container\', \'shared_name\', \'name\'], varargs=None, keywords=None, defaults=[\'[]\', \'-1\', \'\', \'\', \'None\'], " From 455750f3623b15a3b5d46c11d4c5102e9388dbda Mon Sep 17 00:00:00 2001 From: Srinivas Vasudevan Date: Thu, 18 Jun 2020 10:23:23 -0700 Subject: [PATCH 188/867] Add StatelessParameterizedTruncatedNormal sampler. This sampler supports broadcasting of its input parameters as well as puts the # samples at the left of the output shape, rather than the right. PiperOrigin-RevId: 317129622 Change-Id: I4b62ad2e89a9637ae8b30b73af4b662ad0caa943 --- ...tatelessParameterizedTruncatedNormal.pbtxt | 54 +++ tensorflow/core/kernels/BUILD | 1 + .../parameterized_truncated_normal_op.cc | 435 +++++++++++++++++- .../parameterized_truncated_normal_op.h | 16 + tensorflow/core/ops/stateless_random_ops.cc | 35 ++ .../eager/pywrap_gradient_exclusions.cc | 3 +- tensorflow/python/kernel_tests/BUILD | 18 - tensorflow/python/kernel_tests/random/BUILD | 18 + .../parameterized_truncated_normal_op_test.py | 198 ++++++-- tensorflow/python/ops/random_grad.py | 120 +++++ tensorflow/python/ops/stateless_random_ops.py | 70 +++ .../api/golden/v1/tensorflow.random.pbtxt | 4 + .../api/golden/v1/tensorflow.raw_ops.pbtxt | 4 + .../api/golden/v2/tensorflow.random.pbtxt | 4 + .../api/golden/v2/tensorflow.raw_ops.pbtxt | 4 + 15 files changed, 913 insertions(+), 71 deletions(-) create mode 100644 tensorflow/core/api_def/base_api/api_def_StatelessParameterizedTruncatedNormal.pbtxt rename tensorflow/python/kernel_tests/{ => random}/parameterized_truncated_normal_op_test.py (63%) diff --git a/tensorflow/core/api_def/base_api/api_def_StatelessParameterizedTruncatedNormal.pbtxt b/tensorflow/core/api_def/base_api/api_def_StatelessParameterizedTruncatedNormal.pbtxt new file mode 100644 index 00000000000..15bd4670cef --- /dev/null +++ b/tensorflow/core/api_def/base_api/api_def_StatelessParameterizedTruncatedNormal.pbtxt @@ -0,0 +1,54 @@ +op { + graph_op_name: "StatelessParameterizedTruncatedNormal" + visibility: HIDDEN + in_arg { + name: "shape" + description: < { const T kStdDevsInsideBoundsToUseRandnSampler = T(1.3); auto worker_threads = *(ctx->device()->tensorflow_cpu_worker_threads()); - auto DoWork = [samples_per_batch, num_elements, &ctx, &means, &stddevs, - &minvals, &maxvals, &gen, &output, - kStdDevsInsideBoundsToUseRandnSampler](int start_batch, - int limit_batch) { + auto do_work = [samples_per_batch, num_elements, &ctx, &means, &stddevs, + &minvals, &maxvals, &gen, &output, + kStdDevsInsideBoundsToUseRandnSampler](int start_batch, + int limit_batch) { // Capturing "gen" by-value would only make a copy for the _shared_ // lambda. Since we want to let each worker have its own copy, we pass // "gen" by reference and explicitly do a copy assignment here. @@ -80,9 +81,9 @@ struct TruncatedNormalFunctor { // The sample from each iteration uses 2 random numbers. gen_copy.Skip(start_batch * 2 * kMaxIterations * (samples_per_batch + 3) / 4); - typedef random::UniformDistribution Uniform; + using Uniform = random::UniformDistribution; Uniform dist; - typedef random::NormalDistribution Normal; + using Normal = random::NormalDistribution; Normal normal_dist; // Vectorized intermediate calculations for uniform rejection sampling. @@ -112,7 +113,7 @@ struct TruncatedNormalFunctor { Eigen::numext::isfinite(maxval)), errors::InvalidArgument("Invalid parameters")); - int numIterations = 0; + int num_iterations = 0; // If possible, make one-sided bound be the lower bound, or make both // bounds positive. Otherwise, the bounds are on either side of the @@ -160,10 +161,10 @@ struct TruncatedNormalFunctor { if (sample >= limit_sample) { break; } - numIterations = 0; + num_iterations = 0; } else { - numIterations++; - if (numIterations > kMaxIterations) { + num_iterations++; + if (num_iterations > kMaxIterations) { // This should never occur because this sampler should // (by the selection criteria above) be used if at least 3 // standard deviations of one side of the distribution @@ -201,7 +202,7 @@ struct TruncatedNormalFunctor { const auto u = dist(&gen_copy); for (int i = 0; i < size; i++) { auto accept = u[i] <= Eigen::numext::exp(g[i]); - if (accept || numIterations + 1 >= kMaxIterations) { + if (accept || num_iterations + 1 >= kMaxIterations) { // Accept the sample z. // If we run out of iterations, just use the current uniform // sample, but emit a warning. @@ -223,9 +224,9 @@ struct TruncatedNormalFunctor { if (sample >= limit_sample) { break; } - numIterations = 0; + num_iterations = 0; } else { - numIterations++; + num_iterations++; } } } @@ -248,7 +249,7 @@ struct TruncatedNormalFunctor { const T u = rand[i]; i++; auto accept = (u <= g && z < normMax); - if (accept || numIterations + 1 >= kMaxIterations) { + if (accept || num_iterations + 1 >= kMaxIterations) { if (!accept) { LOG(ERROR) << "TruncatedNormal exponential distribution " << "rejection sampler exceeds max iterations. " @@ -263,9 +264,9 @@ struct TruncatedNormalFunctor { if (sample >= limit_sample) { break; } - numIterations = 0; + num_iterations = 0; } else { - numIterations++; + num_iterations++; } } } @@ -305,7 +306,297 @@ struct TruncatedNormalFunctor { const int64 batchCost = batchInitCost + uniformRejectionSamplingCost * 2 * samples_per_batch; Shard(worker_threads.num_threads, worker_threads.workers, num_batches, - batchCost, DoWork); + batchCost, do_work); + } +}; + +template +struct TruncatedNormalFunctorV2 { + void operator()(OpKernelContext* ctx, const CPUDevice& d, int64 num_batches, + int64 samples_per_batch, int64 num_elements, + const BCastList<4>& bcast, + typename TTypes::ConstFlat means, + typename TTypes::ConstFlat stddevs, + typename TTypes::ConstFlat minvals, + typename TTypes::ConstFlat maxvals, + const random::PhiloxRandom& gen, + typename TTypes::Flat output) { + // The randn rejection sampling is used when the mean and at least this many + // standard deviations are inside the bounds. + // The uniform proposal samplers become less efficient as the bounds are + // further from the mean, the reverse is true for the randn sampler. + // This number was chosen by empirical benchmarking. If modified, the + // benchmarks in parameterized_truncated_normal_op_test should also be + // changed. + const T kStdDevsInsideBoundsToUseRandnSampler = T(1.3); + auto worker_threads = *(ctx->device()->tensorflow_cpu_worker_threads()); + + auto do_work = [num_batches, samples_per_batch, &ctx, &bcast, &means, + &stddevs, &minvals, &maxvals, &gen, &output, + kStdDevsInsideBoundsToUseRandnSampler](int start_output, + int limit_output) { + // Capturing "gen" by-value would only make a copy for the _shared_ + // lambda. Since we want to let each worker have its own copy, we pass + // "gen" by reference and explicitly do a copy assignment here. + random::PhiloxRandom gen_copy = gen; + using Uniform = random::UniformDistribution; + Uniform dist; + using Normal = random::NormalDistribution; + Normal normal_dist; + // Skip takes units of 128 bits. The Uniform::kResultElementCount - 1 + // is so rounding doesn't lead to + // us using the same state in different workloads. + // The sample from each iteration uses 2 random numbers. + gen_copy.Skip((start_output * 2 * kMaxIterations + + Uniform::kResultElementCount - 1) / + Uniform::kResultElementCount); + + // Vectorized intermediate calculations for uniform rejection sampling. + // We always generate at most 4 samples. + Eigen::array z; + Eigen::array g; + + const bool should_bcast = bcast.IsBroadcastingRequired(); + const auto& means_batch_indices = bcast.batch_indices(0); + const auto& stddevs_batch_indices = bcast.batch_indices(1); + const auto& minvals_batch_indices = bcast.batch_indices(2); + const auto& maxvals_batch_indices = bcast.batch_indices(3); + auto output_flat = output.data(); + + // We partition work across batches and then across samples + // per batch member, to avoid extra work. + for (int64 output_idx = start_output; output_idx < limit_output; + // output_idx is incremented with the inner loops below. + ) { + int64 batch_idx = output_idx / samples_per_batch; + // The output layout is [samples_per_batch, num_batches]. Thus + // the output address is sample_idx * num_batches + batch_idx. + // Below, code will index at output_batch_offset[sample_idx * + // num_batches] matching this. + T* const output_batch_offset = output_flat + batch_idx; + // Generate batch counts from BCast, as it has the right indices to loop + // over. + T mean, stddev, minval, maxval; + if (should_bcast) { + mean = means(means_batch_indices[batch_idx]); + stddev = stddevs(stddevs_batch_indices[batch_idx]); + minval = minvals(minvals_batch_indices[batch_idx]); + maxval = maxvals(maxvals_batch_indices[batch_idx]); + } else { + mean = means(batch_idx); + stddev = stddevs(batch_idx); + minval = minvals(batch_idx); + maxval = maxvals(batch_idx); + } + + // On GPU, this check will just fill samples with NAN if it fails. + OP_REQUIRES(ctx, + stddev > T(0) && minval < maxval && + (Eigen::numext::isfinite(minval) || + Eigen::numext::isfinite(maxval)), + errors::InvalidArgument("Invalid parameters")); + + int num_iterations = 0; + + // If possible, make one-sided bound be the lower bound, or make both + // bounds positive. Otherwise, the bounds are on either side of the + // mean. + if ((Eigen::numext::isinf(minval) && minval < T(0)) || maxval < mean) { + // Reverse all calculations. normMin and normMax will be flipped. + std::swap(minval, maxval); + stddev = -stddev; + } + + // Calculate normalized samples, then convert them. + const T normMin = (minval - mean) / stddev; + const T normMax = (maxval - mean) / stddev; + + // Determine the method to use. + const T sqrtFactor = Eigen::numext::sqrt((normMin * normMin) + T(4)); + const T cutoff = + T(2) * + Eigen::numext::exp(T(0.5) + + (normMin * (normMin - sqrtFactor)) / T(4)) / + (normMin + sqrtFactor); + const T diff = normMax - normMin; + + if (((normMin < -kStdDevsInsideBoundsToUseRandnSampler) && + (normMax >= T(0.))) || + ((normMax > kStdDevsInsideBoundsToUseRandnSampler) && + (normMin <= T(0.)))) { + // If the bounds are a least 3 standard deviations from the mean + // on at least one side then we rejection sample by sampling + // from the normal distribution and rejecting samples outside + // the bounds. + // Under this condition the acceptance rate per iteration should + // always be ~ 50%. This sampler is more efficient (and more + // numerically stable when one or both bounds is far from the mean). + for (int64 sample_idx = output_idx % samples_per_batch; + sample_idx < samples_per_batch && output_idx < limit_output;) { + const auto randn_sample = normal_dist(&gen_copy); + const int size = randn_sample.size(); + for (int i = 0; i < size; ++i) { + if ((randn_sample[i] >= normMin) && + (randn_sample[i] <= normMax)) { + output_batch_offset[sample_idx * num_batches] = + randn_sample[i] * stddev + mean; + ++sample_idx; + ++output_idx; + if (sample_idx >= samples_per_batch || + output_idx >= limit_output) { + break; + } + num_iterations = 0; + } else { + ++num_iterations; + if (num_iterations > kMaxIterations) { + // This should never occur because this sampler should + // (by the selection criteria above) be used if at least 3 + // standard deviations of one side of the distribution + // is within the limits (so acceptance probability per + // iterations >~ 1/2 per iteration). + LOG(ERROR) << "TruncatedNormal randn rejection sampler " + << "exceeded maximum iterations for " + << "normMin=" << normMin << " normMax=" << normMax + << " kMaxIterations=" << kMaxIterations; + ctx->SetStatus(errors::Internal( + "TruncatedNormal randn rejection sampler failed to accept" + " a sample.")); + return; + } + } + } + } + } else if (diff < cutoff) { + // Sample from a uniform distribution on [normMin, normMax]. + + const T plusFactor = (normMin < T(0)) ? T(0) : normMin * normMin; + + for (int64 sample_idx = output_idx % samples_per_batch; + sample_idx < samples_per_batch && output_idx < limit_output;) { + const auto rand = dist(&gen_copy); + const int size = rand.size(); + // NOTE(ringwalt): These loops seem to only generate packed AVX + // instructions for float32. + for (int i = 0; i < size; i++) { + z[i] = rand[i] * diff + normMin; + g[i] = (plusFactor - z[i] * z[i]) / T(2.0); + } + + const auto u = dist(&gen_copy); + for (int i = 0; i < size; i++) { + auto accept = u[i] <= Eigen::numext::exp(g[i]); + if (accept || num_iterations + 1 >= kMaxIterations) { + // Accept the sample z. + // If we run out of iterations, just use the current uniform + // sample, but emit a warning. + // TODO(jjhunt) For small entropies (relative to the bounds), + // this sampler is poor and may take many iterations since + // the proposal distribution is the uniform distribution + // U(lower_bound, upper_bound). + if (!accept) { + LOG(ERROR) << "TruncatedNormal uniform rejection sampler " + << "exceeded max iterations. Sample may contain " + << "outliers."; + ctx->SetStatus(errors::Internal( + "TruncatedNormal uniform rejection sampler failed to " + " accept a sample.")); + return; + } + output_batch_offset[sample_idx * num_batches] = + z[i] * stddev + mean; + ++sample_idx; + ++output_idx; + if (sample_idx >= samples_per_batch || + output_idx >= limit_output) { + break; + } + num_iterations = 0; + } else { + num_iterations++; + } + } + } + } else { + // Sample from an exponential distribution with alpha maximizing + // acceptance probability, offset by normMin from the origin. + // Accept only if less than normMax. + const T alpha = + (normMin + Eigen::numext::sqrt((normMin * normMin) + T(4))) / + T(2); + for (int64 sample_idx = output_idx % samples_per_batch; + sample_idx < samples_per_batch && output_idx < limit_output;) { + auto rand = dist(&gen_copy); + const int size = rand.size(); + int i = 0; + while (i < size) { + const T z = -Eigen::numext::log(rand[i]) / alpha + normMin; + i++; + const T x = normMin < alpha ? alpha - z : normMin - alpha; + const T g = Eigen::numext::exp(-x * x / T(2.0)); + const T u = rand[i]; + i++; + auto accept = (u <= g && z < normMax); + if (accept || num_iterations + 1 >= kMaxIterations) { + if (!accept) { + LOG(ERROR) << "TruncatedNormal exponential distribution " + << "rejection sampler exceeds max iterations. " + << "Sample may contain outliers."; + ctx->SetStatus(errors::Internal( + "TruncatedNormal exponential distribution rejection" + " sampler failed to accept a sample.")); + return; + } + output_batch_offset[sample_idx * num_batches] = + z * stddev + mean; + ++sample_idx; + ++output_idx; + if (sample_idx >= samples_per_batch || + output_idx >= limit_output) { + break; + } + num_iterations = 0; + } else { + num_iterations++; + } + } + } + } + } + }; + // The cost of the initial calculations for the batch. + const int64 batchInitCost = + // normMin, normMax + (Eigen::TensorOpCost::AddCost() + + Eigen::TensorOpCost::MulCost()) * + 2 + // sqrtFactor + + Eigen::TensorOpCost::AddCost() + + Eigen::TensorOpCost::MulCost() + + Eigen::internal::functor_traits< + Eigen::internal::scalar_sqrt_op>::Cost + // cutoff + + Eigen::TensorOpCost::MulCost() * 4 + + Eigen::internal::functor_traits>::Cost + // diff + + Eigen::TensorOpCost::AddCost(); + const int64 uniformSampleCost = + random::PhiloxRandom::kElementCost + + random::UniformDistribution::kElementCost; + // The cost of a single uniform sampling round. + const int64 uniformRejectionSamplingCost = + uniformSampleCost + Eigen::TensorOpCost::MulCost() + + Eigen::TensorOpCost::AddCost() + + Eigen::TensorOpCost::MulCost() * 2 + + Eigen::TensorOpCost::AddCost() + uniformSampleCost + + Eigen::internal::functor_traits< + Eigen::internal::scalar_exp_op>::Cost + + Eigen::TensorOpCost::MulCost() + Eigen::TensorOpCost::AddCost(); + // Estimate the cost for an entire batch. + // Assume we use uniform sampling, and accept the 2nd sample on average. + const int64 batchCost = batchInitCost + uniformRejectionSamplingCost * 2; + Shard(worker_threads.num_threads, worker_threads.workers, num_elements, + batchCost, do_work); } }; @@ -436,13 +727,113 @@ class ParameterizedTruncatedNormalOp : public OpKernel { TF_DISALLOW_COPY_AND_ASSIGN(ParameterizedTruncatedNormalOp); }; +// Samples from a truncated normal distribution, using the given parameters. +template +class StatelessParameterizedTruncatedNormal : public OpKernel { + // Reshape batches so each batch is this size if possible. + static const int32 kDesiredBatchSize = 100; + + public: + explicit StatelessParameterizedTruncatedNormal(OpKernelConstruction* context) + : OpKernel(context) {} + + void Compute(OpKernelContext* ctx) override { + const Tensor& shape_tensor = ctx->input(0); + const Tensor& seed_tensor = ctx->input(1); + const Tensor& means_tensor = ctx->input(2); + const Tensor& stddevs_tensor = ctx->input(3); + const Tensor& minvals_tensor = ctx->input(4); + const Tensor& maxvals_tensor = ctx->input(5); + + OP_REQUIRES(ctx, seed_tensor.dims() == 1 && seed_tensor.dim_size(0) == 2, + errors::InvalidArgument("seed must have shape [2], not ", + seed_tensor.shape().DebugString())); + + tensorflow::BCastList<4> bcast( + {means_tensor.shape().dim_sizes(), stddevs_tensor.shape().dim_sizes(), + minvals_tensor.shape().dim_sizes(), + maxvals_tensor.shape().dim_sizes()}, + /*fewer_dims_optimization=*/false, + /*return_flattened_batch_indices=*/true); + + OP_REQUIRES(ctx, bcast.IsValid(), + errors::InvalidArgument( + "means, stddevs, minvals, maxvals must have compatible " + "batch dimensions: ", + means_tensor.shape().DebugString(), " vs. ", + stddevs_tensor.shape().DebugString(), " vs. ", + minvals_tensor.shape().DebugString(), " vs. ", + maxvals_tensor.shape().DebugString())); + + // Let's check that the shape tensor dominates the broadcasted tensor. + TensorShape bcast_shape = BCast::ToShape(bcast.output_shape()); + OP_REQUIRES( + ctx, TensorShapeUtils::IsVector(shape_tensor.shape()), + errors::InvalidArgument("Input shape should be a vector, got shape: ", + shape_tensor.shape().DebugString())); + TensorShape output_shape; + if (shape_tensor.dtype() == DataType::DT_INT32) { + OP_REQUIRES_OK(ctx, TensorShapeUtils::MakeShape(shape_tensor.vec(), + &output_shape)); + } else { + OP_REQUIRES_OK(ctx, TensorShapeUtils::MakeShape(shape_tensor.vec(), + &output_shape)); + } + OP_REQUIRES(ctx, TensorShapeUtils::EndsWith(output_shape, bcast_shape), + errors::InvalidArgument( + "Shape passed in must end with broadcasted shape.")); + + int64 samples_per_batch = 1; + const int64 num_sample_dims = + (shape_tensor.dim_size(0) - bcast.output_shape().size()); + for (int64 i = 0; i < num_sample_dims; ++i) { + samples_per_batch *= output_shape.dim_size(i); + } + int64 num_batches = 1; + for (int64 i = num_sample_dims; i < shape_tensor.dim_size(0); ++i) { + num_batches *= output_shape.dim_size(i); + } + const int64 num_elements = num_batches * samples_per_batch; + + Tensor* samples_tensor; + OP_REQUIRES_OK(ctx, ctx->allocate_output(0, output_shape, &samples_tensor)); + + auto truncFunctor = functor::TruncatedNormalFunctorV2(); + // Each worker has the same fudge factor, so use it here. + random::PhiloxRandom::Key key; + random::PhiloxRandom::ResultType counter; + OP_REQUIRES_OK(ctx, GenerateKey(seed_tensor, &key, &counter)); + + auto philox = random::PhiloxRandom(counter, key); + + truncFunctor(ctx, ctx->eigen_device(), num_batches, + samples_per_batch, num_elements, bcast, means_tensor.flat(), + stddevs_tensor.flat(), minvals_tensor.flat(), + maxvals_tensor.flat(), philox, samples_tensor->flat()); + } + + private: + TF_DISALLOW_COPY_AND_ASSIGN(StatelessParameterizedTruncatedNormal); +}; + } // namespace -#define REGISTER(TYPE) \ - REGISTER_KERNEL_BUILDER(Name("ParameterizedTruncatedNormal") \ - .Device(DEVICE_CPU) \ - .TypeConstraint("dtype"), \ - ParameterizedTruncatedNormalOp) +#define REGISTER(TYPE) \ + REGISTER_KERNEL_BUILDER(Name("ParameterizedTruncatedNormal") \ + .Device(DEVICE_CPU) \ + .TypeConstraint("dtype"), \ + ParameterizedTruncatedNormalOp) \ + REGISTER_KERNEL_BUILDER( \ + Name("StatelessParameterizedTruncatedNormal") \ + .HostMemory("shape") \ + .HostMemory("seed") \ + .HostMemory("means") \ + .HostMemory("stddevs") \ + .HostMemory("minvals") \ + .HostMemory("maxvals") \ + .Device(DEVICE_CPU) \ + .TypeConstraint("dtype"), \ + StatelessParameterizedTruncatedNormal) TF_CALL_half(REGISTER); TF_CALL_float(REGISTER); diff --git a/tensorflow/core/kernels/parameterized_truncated_normal_op.h b/tensorflow/core/kernels/parameterized_truncated_normal_op.h index c919a22c7b0..ee7fb7bf605 100644 --- a/tensorflow/core/kernels/parameterized_truncated_normal_op.h +++ b/tensorflow/core/kernels/parameterized_truncated_normal_op.h @@ -18,6 +18,7 @@ limitations under the License. #include "tensorflow/core/framework/tensor_types.h" #include "tensorflow/core/lib/random/random_distributions.h" +#include "tensorflow/core/util/bcast.h" namespace tensorflow { @@ -44,6 +45,21 @@ struct TruncatedNormalFunctor { typename TTypes::Flat output); }; +// This version supports broadcasting of the arguments, as well as puts +// the sample dimension on the left. +template +struct TruncatedNormalFunctorV2 { + void operator()(OpKernelContext* ctx, const Device& d, int64 num_batches, + int64 samples_per_batch, int64 num_elements, + const BCastList<4>& bcast, + typename TTypes::ConstFlat means, + typename TTypes::ConstFlat stddevs, + typename TTypes::ConstFlat minvals, + typename TTypes::ConstFlat maxvals, + const random::PhiloxRandom& gen, + typename TTypes::Flat output); +}; + } // namespace functor } // namespace tensorflow diff --git a/tensorflow/core/ops/stateless_random_ops.cc b/tensorflow/core/ops/stateless_random_ops.cc index d540b9a04d9..e1820ea4feb 100644 --- a/tensorflow/core/ops/stateless_random_ops.cc +++ b/tensorflow/core/ops/stateless_random_ops.cc @@ -124,6 +124,41 @@ REGISTER_OP("StatelessRandomBinomial") .Attr("dtype: {half, float, double, int32, int64} = DT_INT64") .SetShapeFn(StatelessShape); +REGISTER_OP("StatelessParameterizedTruncatedNormal") + .Input("shape: S") + .Input("seed: Tseed") + .Input("means: dtype") + .Input("stddevs: dtype") + .Input("minvals: dtype") + .Input("maxvals: dtype") + .Output("output: dtype") + .Attr("S: {int32, int64}") + .Attr("Tseed: {int32, int64} = DT_INT64") + .Attr("dtype: {float16, float32, float64}") + .SetShapeFn([](InferenceContext* c) { + // Check seed shape + ShapeHandle seed; + TF_RETURN_IF_ERROR(c->WithRank(c->input(1), 1, &seed)); + DimensionHandle unused_dim; + TF_RETURN_IF_ERROR(c->WithValue(c->Dim(seed, 0), 2, &unused_dim)); + + ShapeHandle bcast_means_stddevs; + ShapeHandle bcast_except_maxvals; + ShapeHandle bcast_all; + TF_RETURN_IF_ERROR(BroadcastBinaryOpOutputShapeFnHelper( + c, c->input(2), c->input(3), true, &bcast_means_stddevs)); + TF_RETURN_IF_ERROR(BroadcastBinaryOpOutputShapeFnHelper( + c, c->input(4), bcast_means_stddevs, true, &bcast_except_maxvals)); + TF_RETURN_IF_ERROR(BroadcastBinaryOpOutputShapeFnHelper( + c, c->input(5), bcast_except_maxvals, true, &bcast_all)); + + // Set output shape + ShapeHandle out; + TF_RETURN_IF_ERROR(c->MakeShapeFromShapeTensor(0, &out)); + c->set_output(0, out); + return Status::OK(); + }); + REGISTER_OP("StatelessRandomPoisson") .Input("shape: T") .Input("seed: Tseed") diff --git a/tensorflow/python/eager/pywrap_gradient_exclusions.cc b/tensorflow/python/eager/pywrap_gradient_exclusions.cc index a7c7ab7abc7..7da45e36118 100644 --- a/tensorflow/python/eager/pywrap_gradient_exclusions.cc +++ b/tensorflow/python/eager/pywrap_gradient_exclusions.cc @@ -50,7 +50,7 @@ auto OpGradientInfoInit(const T &a) { absl::optional> OpGradientUnusedInputIndices( const tensorflow::string &op_name) { - static std::array a = {{ + static std::array a = {{ {"Acosh"}, {"AllToAll", 1, {0}}, {"ApproximateEqual"}, @@ -326,6 +326,7 @@ absl::optional> OpGradientUnusedInputIndices( {"StackPop"}, {"StackPush"}, {"StatelessMultinomial"}, + {"StatelessParameterizedTruncatedNormal", 1, {1}}, {"StatelessRandomBinomial"}, {"StatelessRandomGammaV2", 1, {1}}, {"StatelessRandomNormal"}, diff --git a/tensorflow/python/kernel_tests/BUILD b/tensorflow/python/kernel_tests/BUILD index f2c614974f5..f93bf5cd1ae 100644 --- a/tensorflow/python/kernel_tests/BUILD +++ b/tensorflow/python/kernel_tests/BUILD @@ -785,24 +785,6 @@ cuda_py_test( ], ) -cuda_py_test( - name = "parameterized_truncated_normal_op_test", - size = "medium", - srcs = ["parameterized_truncated_normal_op_test.py"], - deps = [ - "//tensorflow/core:protos_all_py", - "//tensorflow/python:client", - "//tensorflow/python:client_testlib", - "//tensorflow/python:control_flow_ops", - "//tensorflow/python:framework", - "//tensorflow/python:framework_for_generated_wrappers", - "//tensorflow/python:platform", - "//tensorflow/python:random_ops", - "//third_party/py/numpy", - "@absl_py//absl/testing:parameterized", - ], -) - tf_py_test( name = "parsing_ops_test", size = "medium", diff --git a/tensorflow/python/kernel_tests/random/BUILD b/tensorflow/python/kernel_tests/random/BUILD index b5d291d2973..6e404b4cd5f 100644 --- a/tensorflow/python/kernel_tests/random/BUILD +++ b/tensorflow/python/kernel_tests/random/BUILD @@ -20,6 +20,24 @@ py_library( ], ) +cuda_py_test( + name = "parameterized_truncated_normal_op_test", + size = "medium", + srcs = ["parameterized_truncated_normal_op_test.py"], + deps = [ + "//tensorflow/core:protos_all_py", + "//tensorflow/python:client", + "//tensorflow/python:client_testlib", + "//tensorflow/python:control_flow_ops", + "//tensorflow/python:framework", + "//tensorflow/python:framework_for_generated_wrappers", + "//tensorflow/python:platform", + "//tensorflow/python:random_ops", + "//third_party/py/numpy", + "@absl_py//absl/testing:parameterized", + ], +) + tf_py_test( name = "random_shuffle_queue_test", size = "small", diff --git a/tensorflow/python/kernel_tests/parameterized_truncated_normal_op_test.py b/tensorflow/python/kernel_tests/random/parameterized_truncated_normal_op_test.py similarity index 63% rename from tensorflow/python/kernel_tests/parameterized_truncated_normal_op_test.py rename to tensorflow/python/kernel_tests/random/parameterized_truncated_normal_op_test.py index ac8ad7a2bd4..309c3e404db 100644 --- a/tensorflow/python/kernel_tests/parameterized_truncated_normal_op_test.py +++ b/tensorflow/python/kernel_tests/random/parameterized_truncated_normal_op_test.py @@ -27,11 +27,15 @@ from six.moves import range # pylint: disable=redefined-builtin from tensorflow.core.protobuf import config_pb2 from tensorflow.python.client import session +from tensorflow.python.eager import backprop from tensorflow.python.framework import ops from tensorflow.python.framework import random_seed from tensorflow.python.framework import test_util +from tensorflow.python.ops import array_ops from tensorflow.python.ops import control_flow_ops from tensorflow.python.ops import random_ops +from tensorflow.python.ops import stateless_random_ops as stateless +from tensorflow.python.ops import variables from tensorflow.python.platform import test from tensorflow.python.platform import tf_logging @@ -91,13 +95,8 @@ class TruncatedNormalMoments(object): def calculate_moments(samples, max_moment): moments = [0.0] * (max_moment + 1) - for sample in samples: - value = 1.0 - for k in range(len(moments)): - moments[k] += value - value *= sample - for i in range(len(moments)): - moments[i] /= len(samples) + for k in range(len(moments)): + moments[k] = np.mean(samples**k, axis=0) return moments @@ -118,16 +117,31 @@ class ParameterizedTruncatedNormalTest(test.TestCase): # Stop at moment 10 to avoid numerical errors in the theoretical moments. max_moment = 10 - def validateMoments(self, shape, mean, stddev, minval, maxval, seed=1618): + def validateMoments(self, + shape, + mean, + stddev, + minval, + maxval, + use_stateless=False, + seed=1618): try: # TruncatedNormalMoments requires scipy.stats. # Give up early if we are unable to import it. - import scipy.stats # pylint: disable=g-import-not-at-top,unused-variable random_seed.set_random_seed(seed) with self.cached_session(use_gpu=True): - samples = random_ops.parameterized_truncated_normal(shape, mean, stddev, - minval, - maxval).eval() + if use_stateless: + # Generate a seed that stateless ops can use. + new_seed = random_ops.random_uniform([2], + seed=seed, + minval=0, + maxval=(2**31 - 1), + dtype=np.int32) + samples = stateless.stateless_parameterized_truncated_normal( + shape, new_seed, mean, stddev, minval, maxval).eval() + else: + samples = random_ops.parameterized_truncated_normal( + shape, mean, stddev, minval, maxval).eval() assert (~np.isnan(samples)).all() moments = calculate_moments(samples, self.max_moment) expected_moments = TruncatedNormalMoments(mean, stddev, minval, maxval) @@ -144,14 +158,24 @@ class ParameterizedTruncatedNormalTest(test.TestCase): stddev, minval, maxval, + use_stateless=False, seed=1618): try: import scipy.stats # pylint: disable=g-import-not-at-top random_seed.set_random_seed(seed) with self.cached_session(use_gpu=True): - samples = random_ops.parameterized_truncated_normal(shape, mean, stddev, - minval, - maxval).eval() + if use_stateless: + new_seed = random_ops.random_uniform([2], + seed=seed, + minval=0, + maxval=(2**31 - 1), + dtype=np.int32) + samples = stateless.stateless_parameterized_truncated_normal( + shape, new_seed, mean, stddev, minval, maxval).eval() + else: + samples = random_ops.parameterized_truncated_normal( + shape, mean, stddev, minval, maxval).eval() + assert (~np.isnan(samples)).all() minval = max(mean - stddev * 10, minval) maxval = min(mean + stddev * 10, maxval) @@ -169,61 +193,160 @@ class ParameterizedTruncatedNormalTest(test.TestCase): @test_util.run_deprecated_v1 def testDefaults(self): - self.validateMoments([10**5], 0.0, 1.0, -2.0, 2.0) + self.validateMoments([int(1e5)], 0.0, 1.0, -2.0, 2.0) + self.validateMoments([int(1e5)], 0.0, 1.0, -2.0, 2.0, use_stateless=True) @test_util.run_deprecated_v1 def testShifted(self): - self.validateMoments([10**5], -1.0, 1.0, -2.0, 2.0) + self.validateMoments([int(1e5)], -1.0, 1.0, -2.0, 2.0) + self.validateMoments([int(1e5)], -1.0, 1.0, -2.0, 2.0, use_stateless=True) @test_util.run_deprecated_v1 def testRightTail(self): - self.validateMoments([10**5], 0.0, 1.0, 4.0, np.infty) + self.validateMoments([int(1e5)], 0.0, 1.0, 4.0, np.infty) + self.validateMoments([int(1e5)], + 0.0, + 1.0, + 4.0, + np.infty, + use_stateless=True) @test_util.run_deprecated_v1 def testLeftTail(self): - self.validateMoments([10**5], 0.0, 1.0, -np.infty, -4.0) + self.validateMoments([int(1e5)], 0.0, 1.0, -np.infty, -4.0) + self.validateMoments([int(1e5)], + 0.0, + 1.0, + -np.infty, + -4.0, + use_stateless=True) @test_util.run_deprecated_v1 def testLeftTailTwoSidedBounds(self): - self.validateMoments([10**5], 0.0, 1.0, -6.0, -3.0) + self.validateMoments([int(1e5)], 0.0, 1.0, -6.0, -3.0) + self.validateMoments([int(1e5)], 0.0, 1.0, -6.0, -3.0, use_stateless=True) @test_util.run_deprecated_v1 @test_util.disable_xla("Low probability region") def testTwoSidedLeftTailShifted(self): - self.validateKolmogorovSmirnov([10**5], 6.0, 1.0, -1.0, 1.0) + self.validateKolmogorovSmirnov([int(1e5)], 6.0, 1.0, -1.0, 1.0) + self.validateKolmogorovSmirnov([int(1e5)], + 6.0, + 1.0, + -1.0, + 1.0, + use_stateless=True) @test_util.run_deprecated_v1 @test_util.disable_xla("Low probability region") def testRightTailShifted(self): - self.validateMoments([10**5], -5.0, 1.0, 2.0, np.infty) + self.validateMoments([int(1e5)], -5.0, 1.0, 2.0, np.infty) + self.validateMoments([int(1e5)], + -5.0, + 1.0, + 2.0, + np.infty, + use_stateless=True) # Take the normal distribution around the mean, but truncating the left tail # far from the mean. @test_util.run_deprecated_v1 def testTruncateOnLeft_entireTailOnRight(self): - self.validateKolmogorovSmirnov([10**5], 10.0, 1.0, 4.0, np.infty) + self.validateKolmogorovSmirnov([int(1e5)], 10.0, 1.0, 4.0, np.infty) + self.validateKolmogorovSmirnov([int(1e5)], + 10.0, + 1.0, + 4.0, + np.infty, + use_stateless=True) # Take the normal distribution around the mean, but truncating the right tail. @test_util.run_deprecated_v1 def testTruncateOnRight_entireTailOnLeft(self): - self.validateKolmogorovSmirnov([10**5], -8, 1.0, -np.infty, -4.0) + self.validateKolmogorovSmirnov([int(1e5)], -8, 1.0, -np.infty, -4.0) + self.validateKolmogorovSmirnov([int(1e5)], + -8., + 1.0, + -np.infty, + -4.0, + use_stateless=True) @test_util.run_deprecated_v1 def testSmallStddev(self): - self.validateKolmogorovSmirnov([10**5], 0.0, 0.1, 0.05, 0.10) + self.validateKolmogorovSmirnov([int(1e5)], 0.0, 0.1, 0.05, 0.10) + self.validateKolmogorovSmirnov([int(1e5)], + 0.0, + 0.1, + 0.05, + 0.10, + use_stateless=True) @test_util.run_deprecated_v1 def testSamplingWithSmallStdDevFarFromBound(self): sample_op = random_ops.parameterized_truncated_normal( shape=(int(1e5),), means=0.8, stddevs=0.05, minvals=-1., maxvals=1.) + new_seed = random_ops.random_uniform([2], + seed=1234, + minval=0, + maxval=(2**31 - 1), + dtype=np.int32) + sample_op_stateless = stateless.stateless_parameterized_truncated_normal( + shape=(int(1e5),), + seed=new_seed, + means=0.8, + stddevs=0.05, + minvals=-1., + maxvals=1.) with self.session(use_gpu=True) as sess: - samples = sess.run(sample_op) + samples, samples_stateless = sess.run([sample_op, sample_op_stateless]) # 0. is more than 16 standard deviations from the mean, and # should have a likelihood < 1e-57. assert (~np.isnan(samples)).all() - no_neg_samples = np.sum(samples < 0.) - self.assertEqual(no_neg_samples, 0.) + assert (~np.isnan(samples_stateless)).all() + self.assertAllGreater(samples, 0.) + self.assertAllGreater(samples_stateless, 0.) + + def testStatelessParameterizedTruncatedNormalHasGrads(self): + mean = variables.Variable(0.01) + stddev = variables.Variable(1.) + minval = variables.Variable(-1.) + maxval = variables.Variable(1.) + + with self.cached_session(use_gpu=True) as sess: + with backprop.GradientTape(persistent=True) as tape: + samples = stateless.stateless_parameterized_truncated_normal( + [1], [1, 2], mean, stddev, minval, maxval) + + sess.run(variables.variables_initializer([mean, stddev, minval, maxval])) + [mean_grad, std_grad], mean_actual_grad, std_actual_grad = sess.run([ + tape.gradient(samples, [mean, stddev]), + array_ops.ones_like(mean), + (samples - mean) / stddev]) + self.assertAllClose(mean_grad, mean_actual_grad) + self.assertAllClose(std_grad, std_actual_grad[0]) + + try: + import scipy.stats # pylint:disable=g-import-not-at-top + truncnorm = scipy.stats.truncnorm(a=-1., b=1., loc=0., scale=1.) + samples_np, [minval_grad, maxval_grad] = sess.run([ + samples, tape.gradient(samples, [minval, maxval])]) + + sample_cdf = truncnorm.cdf(samples_np) + # These come from the implicit reparameterization trick. + scipy_maxval_grad = np.exp( + 0.5 * (samples_np ** 2 - ((1. - 0.01) / 1.) ** 2) + + np.log(sample_cdf)) + + scipy_minval_grad = np.exp( + 0.5 * (samples_np ** 2 - ((-1. - 0.01) / 1.) ** 2) + + np.log1p(-sample_cdf)) + + self.assertAllClose(minval_grad, scipy_minval_grad[0], rtol=1e-2) + self.assertAllClose(maxval_grad, scipy_maxval_grad[0], rtol=1e-2) + + except ImportError as e: + tf_logging.warn("Cannot test truncated normal op: %s" % str(e)) @test_util.run_deprecated_v1 def testSamplingAtRandnSwitchover(self): @@ -239,18 +362,33 @@ class ParameterizedTruncatedNormalTest(test.TestCase): epsilon = 0.001 self.validateMoments( - shape=[10**6], + shape=[int(1e6)], mean=0., stddev=1.0, minval=-epsilon, maxval=stddev_inside_bounds_before_using_randn - epsilon) self.validateMoments( - shape=[10**6], + shape=[int(1e6)], mean=0., stddev=1.0, minval=-epsilon, maxval=stddev_inside_bounds_before_using_randn + epsilon) + self.validateMoments( + shape=[int(1e6)], + mean=0., + stddev=1.0, + minval=-epsilon, + maxval=stddev_inside_bounds_before_using_randn - epsilon, + use_stateless=True) + self.validateMoments( + shape=[int(1e6)], + mean=0., + stddev=1.0, + minval=-epsilon, + maxval=stddev_inside_bounds_before_using_randn + epsilon, + use_stateless=True) + # Benchmarking code def parameterized_vs_naive(shape, num_iters, use_gpu=False): diff --git a/tensorflow/python/ops/random_grad.py b/tensorflow/python/ops/random_grad.py index 771980932cb..3caa08d96f9 100644 --- a/tensorflow/python/ops/random_grad.py +++ b/tensorflow/python/ops/random_grad.py @@ -18,9 +18,14 @@ from __future__ import absolute_import from __future__ import division from __future__ import print_function +import numpy as np + +from tensorflow.python.framework import constant_op from tensorflow.python.framework import dtypes from tensorflow.python.framework import ops from tensorflow.python.ops import array_ops +from tensorflow.python.ops import clip_ops +from tensorflow.python.ops import gen_array_ops from tensorflow.python.ops import gen_random_ops from tensorflow.python.ops import math_ops @@ -114,3 +119,118 @@ def _StatelessRandomGammaV2Grad(op, grad): # pylint: disable=invalid-name return (None, None, math_ops.reduce_sum( grad * partial_a, axis=math_ops.range(num_sample_dimensions))) + + +def _Ndtr(x): + """Normal distribution function.""" + half_sqrt_2 = constant_op.constant( + 0.5 * np.sqrt(2.), dtype=x.dtype, name="half_sqrt_2") + w = x * half_sqrt_2 + z = math_ops.abs(w) + y = array_ops.where( + z < half_sqrt_2, + 1. + math_ops.erf(w), + array_ops.where( + w > 0., 2. - math_ops.erfc(z), math_ops.erfc(z))) + return 0.5 * y + + +@ops.RegisterGradient("StatelessParameterizedTruncatedNormal") +def _StatelessParameterizedTruncatedNormalGrad(op, grad): # pylint: disable=invalid-name + """Returns the gradient of a TruncatedNormal sample w.r.t. parameters. + + The gradient is computed using implicit differentiation + (Figurnov et al., 2018). + + Args: + op: A `StatelessParameterizedTruncatedNormal` operation. We assume that the + inputs to the operation are `shape`, `seed`, `mean`, `stddev`, `minval`, + and `maxval` tensors, and the output is the `sample` tensor. + grad: The incoming gradient `dloss / dsample` of the same shape as + `op.outputs[0]`. + + Returns: + A list of `Tensor` with derivates with respect to each parameter. + + References: + Implicit Reparameterization Gradients: + [Figurnov et al., 2018] + (http://papers.nips.cc/paper/7326-implicit-reparameterization-gradients) + ([pdf] + (http://papers.nips.cc/paper/7326-implicit-reparameterization-gradients.pdf)) + """ + shape = op.inputs[0] + mean = op.inputs[2] + stddev = op.inputs[3] + minval = op.inputs[4] + maxval = op.inputs[5] + sample = op.outputs[0] + + with ops.control_dependencies([grad]): + minval_std = (minval - mean) / stddev + maxval_std = (maxval - mean) / stddev + sample_std = (sample - mean) / stddev + + cdf_sample = (_Ndtr(sample_std) - _Ndtr(minval_std)) / ( + _Ndtr(maxval_std) - _Ndtr(minval_std)) + + # Clip to avoid zero argument for log_cdf expression + tiny = np.finfo(mean.dtype.as_numpy_dtype).tiny + eps = np.finfo(mean.dtype.as_numpy_dtype).eps + cdf_sample = clip_ops.clip_by_value(cdf_sample, tiny, 1 - eps) + + dmaxval = math_ops.exp(0.5 * (sample_std ** 2 - maxval_std ** 2) + + math_ops.log(cdf_sample)) + dminval = math_ops.exp(0.5 * (sample_std ** 2 - minval_std ** 2) + + math_ops.log1p(-cdf_sample)) + dmean = array_ops.ones_like(sample_std) + dstddev = sample_std + + # Reduce over extra dimensions caused by `shape`. We need to get the + # difference in rank from shape vs. the broadcasted rank. + + mean_shape = array_ops.shape(mean) + stddev_shape = array_ops.shape(stddev) + minval_shape = array_ops.shape(minval) + maxval_shape = array_ops.shape(maxval) + + broadcast_shape = array_ops.broadcast_dynamic_shape( + mean_shape, stddev_shape) + broadcast_shape = array_ops.broadcast_dynamic_shape( + minval_shape, broadcast_shape) + broadcast_shape = array_ops.broadcast_dynamic_shape( + maxval_shape, broadcast_shape) + extra_dims = math_ops.range( + array_ops.size(shape) - array_ops.size(broadcast_shape)) + + grad_mean = math_ops.reduce_sum(grad * dmean, axis=extra_dims) + grad_stddev = math_ops.reduce_sum(grad * dstddev, axis=extra_dims) + grad_minval = math_ops.reduce_sum(grad * dminval, axis=extra_dims) + grad_maxval = math_ops.reduce_sum(grad * dmaxval, axis=extra_dims) + + _, rmean = gen_array_ops.broadcast_gradient_args( + broadcast_shape, mean_shape) + _, rstddev = gen_array_ops.broadcast_gradient_args( + broadcast_shape, stddev_shape) + _, rminval = gen_array_ops.broadcast_gradient_args( + broadcast_shape, minval_shape) + _, rmaxval = gen_array_ops.broadcast_gradient_args( + broadcast_shape, maxval_shape) + + grad_mean = array_ops.reshape( + math_ops.reduce_sum(grad_mean, axis=rmean, keepdims=True), mean_shape) + + grad_stddev = array_ops.reshape( + math_ops.reduce_sum(grad_stddev, axis=rstddev, keepdims=True), + stddev_shape) + + grad_minval = array_ops.reshape( + math_ops.reduce_sum(grad_minval, axis=rminval, keepdims=True), + minval_shape) + + grad_maxval = array_ops.reshape( + math_ops.reduce_sum(grad_maxval, axis=rmaxval, keepdims=True), + maxval_shape) + + # The first two inputs are shape. + return (None, None, grad_mean, grad_stddev, grad_minval, grad_maxval) diff --git a/tensorflow/python/ops/stateless_random_ops.py b/tensorflow/python/ops/stateless_random_ops.py index 25fefcc514c..3e825cc4775 100644 --- a/tensorflow/python/ops/stateless_random_ops.py +++ b/tensorflow/python/ops/stateless_random_ops.py @@ -618,3 +618,73 @@ def stateless_multinomial_categorical_impl(logits, num_samples, dtype, seed): logits = ops.convert_to_tensor(logits, name="logits") return gen_stateless_random_ops.stateless_multinomial( logits, num_samples, seed, output_dtype=dtype) + + +@dispatch.add_dispatch_support +@tf_export("random.stateless_parameterized_truncated_normal") +def stateless_parameterized_truncated_normal(shape, + seed, + means=0.0, + stddevs=1.0, + minvals=-2.0, + maxvals=2.0, + name=None): + """Outputs random values from a truncated normal distribution. + + The generated values follow a normal distribution with specified mean and + standard deviation, except that values whose magnitude is more than 2 standard + deviations from the mean are dropped and re-picked. + + + Examples: + + Sample from a Truncated normal, with deferring shape parameters that + broadcast. + + >>> means = 0. + >>> stddevs = tf.math.exp(tf.random.uniform(shape=[2, 3])) + >>> minvals = [-1., -2., -1000.] + >>> maxvals = [[10000.], [1.]] + >>> y = tf.random.stateless_parameterized_truncated_normal( + ... shape=[10, 2, 3], seed=[7, 17], + ... means=means, stddevs=stddevs, minvals=minvals, maxvals=maxvals) + >>> y.shape + TensorShape([10, 2, 3]) + + Args: + shape: A 1-D integer `Tensor` or Python array. The shape of the output + tensor. + seed: A shape [2] Tensor, the seed to the random number generator. Must have + dtype `int32` or `int64`. (When using XLA, only `int32` is allowed.) + means: A `Tensor` or Python value of type `dtype`. The mean of the truncated + normal distribution. This must broadcast with `stddevs`, `minvals` and + `maxvals`, and the broadcasted shape must be dominated by `shape`. + stddevs: A `Tensor` or Python value of type `dtype`. The standard deviation + of the truncated normal distribution. This must broadcast with `means`, + `minvals` and `maxvals`, and the broadcasted shape must be dominated by + `shape`. + minvals: A `Tensor` or Python value of type `dtype`. The minimum value of + the truncated normal distribution. This must broadcast with `means`, + `stddevs` and `maxvals`, and the broadcasted shape must be dominated by + `shape`. + maxvals: A `Tensor` or Python value of type `dtype`. The maximum value of + the truncated normal distribution. This must broadcast with `means`, + `stddevs` and `minvals`, and the broadcasted shape must be dominated by + `shape`. + name: A name for the operation (optional). + + Returns: + A tensor of the specified shape filled with random truncated normal values. + """ + with ops.name_scope(name, "stateless_parameterized_truncated_normal", + [shape, means, stddevs, minvals, maxvals]) as name: + shape_tensor = tensor_util.shape_tensor(shape) + means_tensor = ops.convert_to_tensor(means, name="means") + stddevs_tensor = ops.convert_to_tensor(stddevs, name="stddevs") + minvals_tensor = ops.convert_to_tensor(minvals, name="minvals") + maxvals_tensor = ops.convert_to_tensor(maxvals, name="maxvals") + rnd = gen_stateless_random_ops.stateless_parameterized_truncated_normal( + shape_tensor, seed, means_tensor, stddevs_tensor, minvals_tensor, + maxvals_tensor) + tensor_util.maybe_set_static_shape(rnd, shape) + return rnd diff --git a/tensorflow/tools/api/golden/v1/tensorflow.random.pbtxt b/tensorflow/tools/api/golden/v1/tensorflow.random.pbtxt index 9c6fa7154a3..f5963f1324c 100644 --- a/tensorflow/tools/api/golden/v1/tensorflow.random.pbtxt +++ b/tensorflow/tools/api/golden/v1/tensorflow.random.pbtxt @@ -92,6 +92,10 @@ tf_module { name: "stateless_normal" argspec: "args=[\'shape\', \'seed\', \'mean\', \'stddev\', \'dtype\', \'name\'], varargs=None, keywords=None, defaults=[\'0.0\', \'1.0\', \"\", \'None\'], " } + member_method { + name: "stateless_parameterized_truncated_normal" + argspec: "args=[\'shape\', \'seed\', \'means\', \'stddevs\', \'minvals\', \'maxvals\', \'name\'], varargs=None, keywords=None, defaults=[\'0.0\', \'1.0\', \'-2.0\', \'2.0\', \'None\'], " + } member_method { name: "stateless_poisson" argspec: "args=[\'shape\', \'seed\', \'lam\', \'dtype\', \'name\'], varargs=None, keywords=None, defaults=[\"\", \'None\'], " diff --git a/tensorflow/tools/api/golden/v1/tensorflow.raw_ops.pbtxt b/tensorflow/tools/api/golden/v1/tensorflow.raw_ops.pbtxt index 62969b5a0dd..8e5303cbea4 100644 --- a/tensorflow/tools/api/golden/v1/tensorflow.raw_ops.pbtxt +++ b/tensorflow/tools/api/golden/v1/tensorflow.raw_ops.pbtxt @@ -4492,6 +4492,10 @@ tf_module { name: "StatelessMultinomial" argspec: "args=[\'logits\', \'num_samples\', \'seed\', \'output_dtype\', \'name\'], varargs=None, keywords=None, defaults=[\"\", \'None\'], " } + member_method { + name: "StatelessParameterizedTruncatedNormal" + argspec: "args=[\'shape\', \'seed\', \'means\', \'stddevs\', \'minvals\', \'maxvals\', \'name\'], varargs=None, keywords=None, defaults=[\'None\'], " + } member_method { name: "StatelessRandomBinomial" argspec: "args=[\'shape\', \'seed\', \'counts\', \'probs\', \'dtype\', \'name\'], varargs=None, keywords=None, defaults=[\"\", \'None\'], " diff --git a/tensorflow/tools/api/golden/v2/tensorflow.random.pbtxt b/tensorflow/tools/api/golden/v2/tensorflow.random.pbtxt index e3a11ee4610..d1b8c90bfae 100644 --- a/tensorflow/tools/api/golden/v2/tensorflow.random.pbtxt +++ b/tensorflow/tools/api/golden/v2/tensorflow.random.pbtxt @@ -80,6 +80,10 @@ tf_module { name: "stateless_normal" argspec: "args=[\'shape\', \'seed\', \'mean\', \'stddev\', \'dtype\', \'name\'], varargs=None, keywords=None, defaults=[\'0.0\', \'1.0\', \"\", \'None\'], " } + member_method { + name: "stateless_parameterized_truncated_normal" + argspec: "args=[\'shape\', \'seed\', \'means\', \'stddevs\', \'minvals\', \'maxvals\', \'name\'], varargs=None, keywords=None, defaults=[\'0.0\', \'1.0\', \'-2.0\', \'2.0\', \'None\'], " + } member_method { name: "stateless_poisson" argspec: "args=[\'shape\', \'seed\', \'lam\', \'dtype\', \'name\'], varargs=None, keywords=None, defaults=[\"\", \'None\'], " diff --git a/tensorflow/tools/api/golden/v2/tensorflow.raw_ops.pbtxt b/tensorflow/tools/api/golden/v2/tensorflow.raw_ops.pbtxt index 62969b5a0dd..8e5303cbea4 100644 --- a/tensorflow/tools/api/golden/v2/tensorflow.raw_ops.pbtxt +++ b/tensorflow/tools/api/golden/v2/tensorflow.raw_ops.pbtxt @@ -4492,6 +4492,10 @@ tf_module { name: "StatelessMultinomial" argspec: "args=[\'logits\', \'num_samples\', \'seed\', \'output_dtype\', \'name\'], varargs=None, keywords=None, defaults=[\"\", \'None\'], " } + member_method { + name: "StatelessParameterizedTruncatedNormal" + argspec: "args=[\'shape\', \'seed\', \'means\', \'stddevs\', \'minvals\', \'maxvals\', \'name\'], varargs=None, keywords=None, defaults=[\'None\'], " + } member_method { name: "StatelessRandomBinomial" argspec: "args=[\'shape\', \'seed\', \'counts\', \'probs\', \'dtype\', \'name\'], varargs=None, keywords=None, defaults=[\"\", \'None\'], " From 81041bcd8267056741c8b4766bd8018a7270f454 Mon Sep 17 00:00:00 2001 From: Lu Wang Date: Thu, 18 Jun 2020 10:29:58 -0700 Subject: [PATCH 189/867] Update the documentation for Model metadata PiperOrigin-RevId: 317131186 Change-Id: Ida1cd7fc729d7d601c32d646bd650fe1aa539591 --- tensorflow/lite/g3doc/convert/metadata.md | 72 +++++++++++++++++++---- 1 file changed, 60 insertions(+), 12 deletions(-) diff --git a/tensorflow/lite/g3doc/convert/metadata.md b/tensorflow/lite/g3doc/convert/metadata.md index 29b2c5ce2b3..cd86333b305 100644 --- a/tensorflow/lite/g3doc/convert/metadata.md +++ b/tensorflow/lite/g3doc/convert/metadata.md @@ -9,7 +9,8 @@ input / output information. The metadata consists of both * human readable parts which convey the best practice when using the model, and * machine readable parts that can be leveraged by code generators, such as - [the TensorFlow Lite Android code generator](../guide/codegen.md). + [the TensorFlow Lite Android code generator](../guide/codegen.md) and + [the Android Studio ML Binding feature](https://developer.android.com/studio/preview/features#tensor-flow-lite-models). ## Setup the metadata tools @@ -32,11 +33,21 @@ There are three parts to the [model metadata](https://github.com/tensorflow/tensorflow/blob/master/tensorflow/lite/experimental/support/metadata/metadata_schema.fbs): 1. **Model information** - Overall description of the model as well as items - such as licence terms. + such as licence terms. See + [ModelMetadata](https://github.com/tensorflow/tensorflow/blob/master/tensorflow/lite/experimental/support/metadata/metadata_schema.fbs#L515). 2. **Input information** - Description of the inputs and pre-processing - required such as normalization. + required such as normalization. See + [SubGraphMetadata.input_tensor_metadata](https://github.com/tensorflow/tensorflow/blob/master/tensorflow/lite/experimental/support/metadata/metadata_schema.fbs#L500). 3. **Output information** - Description of the output and post-processing - required such as mapping to labels. + required such as mapping to labels. See + [SubGraphMetadata.output_tensor_metadata](https://github.com/tensorflow/tensorflow/blob/master/tensorflow/lite/experimental/support/metadata/metadata_schema.fbs#L509). + +Since TensorFlow Lite only supports single subgraph at this point, the +[TensorFlow Lite code generator](../guide/codegen.md) and +[the Android Studio ML Binding feature](https://developer.android.com/studio/preview/features#tensor-flow-lite-models) +will use `ModelMetadata.name` and `ModelMetadata.description`, instead of +`SubGraphMetadata.name` and `SubGraphMetadata.description`, when displaying +metadata and generating code. ### Supported Input / Output types @@ -51,6 +62,29 @@ Lite metadata: * Bounding box - Rectangular shape bounding boxes. The schema supports [a variety of numbering schemes](https://github.com/tensorflow/tensorflow/blob/268853ee81edab09e07f455cc918f7ef9a421485/tensorflow/lite/experimental/support/metadata/metadata_schema.fbs#L165). +### Pack the associated files + +TensorFlow Lite models may come with different associated files. For example, +natural language models usually have vocab files that map word pieces to word +IDs; classification models may have label files that indicate object categories. +Without the associated files (if there are), a model will not function well. + +The associated files can now be bundled with the model through the metadata +Python library. The new TensorFlow Lite model becomes a zip file that contains +both the model and the associated files. It can be unpacked with common zip +tools. This new model format keeps using the same file extension, `.tflite`. It +is compatible with existing TFLite framework and Interpreter. See +[Pack mtadata and associated files into the model](#pack-metadata-and-associated-files-into-the-model) +for more details. + +The associate file information can be recored in the metadata. Depending on the +file type and where the file is attached to (i.e. `ModelMetadata`, +`SubGraphMetadata`, and `TensorMetadata`), +[the TensorFlow Lite Android code generator](../guide/codegen.md) may apply +corresponding pre/post processing automatically to the object. See +[the \ section of each associate file type](https://github.com/tensorflow/tensorflow/blob/268853ee81edab09e07f455cc918f7ef9a421485/tensorflow/lite/experimental/support/metadata/metadata_schema.fbs#L37-L77) +in the schema for more details. + ### Examples Note: The export directory specified has to exist before you run the script; it @@ -63,7 +97,9 @@ types of models here: Download the script [here](https://github.com/tensorflow/examples/tree/master/lite/examples/image_classification/metadata/metadata_writer_for_image_classifier.py) -and run the script like this: +, which populates metadata to +[mobilenet_v1_0.75_160_quantized.tflite](https://tfhub.dev/tensorflow/lite-model/mobilenet_v1_0.75_160_quantized/1/default/1). +Run the script like this: ```sh python ./metadata_writer_for_image_classifier.py \ @@ -72,8 +108,11 @@ python ./metadata_writer_for_image_classifier.py \ --export_directory=model_with_metadata ``` -The rest of this guide will highlight some of the key sections in the image -classification example to illustrate the key elements. +To populate metadata for other image classification models, add the model specs +like +[this](https://github.com/tensorflow/examples/blob/master/lite/examples/image_classification/metadata/metadata_writer_for_image_classifier.py#L63-L74) +into the script. The rest of this guide will highlight some of the key sections +in the image classification example to illustrate the key elements. ### Deep dive into the image classification example @@ -173,7 +212,7 @@ label_file.type = _metadata_fb.AssociatedFileType.TENSOR_AXIS_LABELS output_meta.associatedFiles = [label_file] ``` -#### Put it all together +#### Create the metadata Flatbuffers The following code combines the model information with the input and output information: @@ -192,8 +231,10 @@ b.Finish( metadata_buf = b.Output() ``` -Once the data structure is ready, the metadata is written into the TFLite file -via the `populate` method: +#### Pack metadata and associated files into the model + +Once the metadata Flatbuffers is created, the metadata and the label file are +written into the TFLite file via the `populate` method: ```python populator = _metadata.MetadataPopulator.with_model_file(model_file) @@ -202,9 +243,16 @@ populator.load_associated_files(["your_path_to_label_file"]) populator.populate() ``` -#### Verify the metadata +You can pack as many associated files as you want into the model through +`load_associated_files`. However, it is required to pack at least those files +documented in the metadata. In this example, packing the lable file is +mandatory. -You can read the metadata in a TFLite file using the `MetadataDisplayer`: +### Visualize the metadata + +You can use [Netron](https://github.com/lutzroeder/netron) to visualize your +metadata, or you can read the metadata from a TensorFlow Lite model into a json +format using the `MetadataDisplayer`: ```python displayer = _metadata.MetadataDisplayer.with_model_file(export_model_path) From 47582983cb1064b5bb81233db4f0adeeaa10b74d Mon Sep 17 00:00:00 2001 From: Scott Zhu Date: Thu, 18 Jun 2020 10:30:56 -0700 Subject: [PATCH 190/867] Fix save model issue for stateless ConvLSTM2D layer. The root cause is that ConvLSTM2D.state is a tuple rather than a list. When converting the state for save_model, the tuple is not converted to trackable objects since the states are (None, None). On the other hand, save_model requires all objects to be trackable when saving. We didn't hit this issue for keras.LSTM since its state is a list, rather than tuple. The list is auto convert to ListWrapper since list itself is mutable. This should fix https://github.com/tensorflow/tensorflow/issues/40328 and partly https://github.com/tensorflow/tensorflow/issues/38220 PiperOrigin-RevId: 317131403 Change-Id: I202d4dbdb29accc7a047d5f5a2fef08d24d05c7c --- .../saving/saved_model/layer_serialization.py | 10 ++++++++-- .../saving/saved_model/saved_model_test.py | 20 +++++++++++++++++++ 2 files changed, 28 insertions(+), 2 deletions(-) diff --git a/tensorflow/python/keras/saving/saved_model/layer_serialization.py b/tensorflow/python/keras/saving/saved_model/layer_serialization.py index 559b6158d87..4216457bf28 100644 --- a/tensorflow/python/keras/saving/saved_model/layer_serialization.py +++ b/tensorflow/python/keras/saving/saved_model/layer_serialization.py @@ -159,6 +159,12 @@ class RNNSavedModelSaver(LayerSavedModelSaver): objects, functions = ( super(RNNSavedModelSaver, self)._get_serialized_attributes_internal( serialization_cache)) - - objects['states'] = data_structures.wrap_or_unwrap(self.obj.states) + states = data_structures.wrap_or_unwrap(self.obj.states) + # Force the tuple into TupleWrapper which is a trackable object. The + # save/load code requires all the objects to be trackable. + # Tuple is not converted to TupleWrapper by data_structures.wrap_or_unwrap() + # if it doesn't contains any trackable objects. + if isinstance(states, tuple): + states = data_structures._TupleWrapper(states) # pylint: disable=protected-access + objects['states'] = states return objects, functions diff --git a/tensorflow/python/keras/saving/saved_model/saved_model_test.py b/tensorflow/python/keras/saving/saved_model/saved_model_test.py index 8d4d27e2357..3f55d5f40b5 100644 --- a/tensorflow/python/keras/saving/saved_model/saved_model_test.py +++ b/tensorflow/python/keras/saving/saved_model/saved_model_test.py @@ -773,6 +773,26 @@ class TestModelSavingAndLoadingV2(keras_parameterized.TestCase): self.assertAllClose(layer.states, loaded_layer.states) self.assertAllClose(model(input_arr), loaded(input_arr)) + def testSaveStatelessConvLSTM2D(self): + data_format = 'channels_first' + batch, timesteps, channels, rows, cols = 12, 10, 8, 4, 4 + input_arr = np.ones( + (batch, timesteps, channels, rows, cols)).astype('float32') + layer = keras.layers.ConvLSTM2D( + filters=16, kernel_size=(1, 1), data_format=data_format) + x = keras.Input(batch_shape=(batch, timesteps, channels, rows, cols)) + y = layer(x) + model = keras.Model(x, y) + + predict_1 = model(input_arr) + saved_model_dir = self._save_model_dir() + tf_save.save(model, saved_model_dir) + del model + + loaded = keras_load.load(saved_model_dir) + predict_2 = loaded(input_arr) + self.assertAllClose(predict_1, predict_2) + def testSaveWithRaggedInputs(self): class EmbeddingMerger(keras.layers.Layer): From 69467b92630bb37c077eda3e411903ccddd906ff Mon Sep 17 00:00:00 2001 From: Sachin Joglekar Date: Thu, 18 Jun 2020 10:32:53 -0700 Subject: [PATCH 191/867] Make control_dep test mlir-only, since tf2 behavior introduces control-flow ops that are not supported by TOCO PiperOrigin-RevId: 317131879 Change-Id: I8cf51927a3d4985cec7d2b03930d2f553a6a3286 --- tensorflow/lite/build_def.bzl | 1 - .../lite/testing/generate_examples_lib.py | 1 - .../lite/testing/op_tests/control_dep.py | 61 ------------------- 3 files changed, 63 deletions(-) delete mode 100644 tensorflow/lite/testing/op_tests/control_dep.py diff --git a/tensorflow/lite/build_def.bzl b/tensorflow/lite/build_def.bzl index e6c92691b15..ad43b56743a 100644 --- a/tensorflow/lite/build_def.bzl +++ b/tensorflow/lite/build_def.bzl @@ -251,7 +251,6 @@ def generated_test_models(): "ceil", "concat", "constant", - # "control_dep", # b/150647401 "conv", "conv_relu", "conv_relu1", diff --git a/tensorflow/lite/testing/generate_examples_lib.py b/tensorflow/lite/testing/generate_examples_lib.py index fc92991bd57..fce2beabf45 100644 --- a/tensorflow/lite/testing/generate_examples_lib.py +++ b/tensorflow/lite/testing/generate_examples_lib.py @@ -52,7 +52,6 @@ from tensorflow.lite.testing.op_tests.cast import make_cast_tests from tensorflow.lite.testing.op_tests.ceil import make_ceil_tests from tensorflow.lite.testing.op_tests.concat import make_concat_tests from tensorflow.lite.testing.op_tests.constant import make_constant_tests -from tensorflow.lite.testing.op_tests.control_dep import make_control_dep_tests from tensorflow.lite.testing.op_tests.conv import make_conv_tests from tensorflow.lite.testing.op_tests.conv2d_transpose import make_conv2d_transpose_tests from tensorflow.lite.testing.op_tests.conv_activation import make_conv_relu_tests, make_conv_relu1_tests, make_conv_relu6_tests diff --git a/tensorflow/lite/testing/op_tests/control_dep.py b/tensorflow/lite/testing/op_tests/control_dep.py deleted file mode 100644 index bd9e369303b..00000000000 --- a/tensorflow/lite/testing/op_tests/control_dep.py +++ /dev/null @@ -1,61 +0,0 @@ -# Copyright 2019 The TensorFlow Authors. All Rights Reserved. -# -# Licensed under the Apache License, Version 2.0 (the "License"); -# you may not use this file except in compliance with the License. -# You may obtain a copy of the License at -# -# http://www.apache.org/licenses/LICENSE-2.0 -# -# Unless required by applicable law or agreed to in writing, software -# distributed under the License is distributed on an "AS IS" BASIS, -# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -# See the License for the specific language governing permissions and -# limitations under the License. -# ============================================================================== -"""Test configs for control_dep.""" -from __future__ import absolute_import -from __future__ import division -from __future__ import print_function - -import tensorflow.compat.v1 as tf -from tensorflow.lite.testing.zip_test_utils import create_tensor_data -from tensorflow.lite.testing.zip_test_utils import ExtraTocoOptions -from tensorflow.lite.testing.zip_test_utils import make_zip_of_tests -from tensorflow.lite.testing.zip_test_utils import register_make_test_function - -TEST_INPUT_DEPTH = 3 - - -@register_make_test_function() -def make_control_dep_tests(options): - """Make a set of tests that use control dependencies.""" - - test_parameters = [{ - "input_shape": [[], [1, 1, 1, 1], [1, 15, 14, 1], [3, 15, 14, 3]], - }] - - def build_graph(parameters): - input_tensor = tf.compat.v1.placeholder( - dtype=tf.float32, name="input", shape=parameters["input_shape"]) - filter_value = tf.zeros((3, 3, TEST_INPUT_DEPTH, 8), tf.float32) - assert_op = tf.compat.v1.assert_greater_equal(input_tensor, - input_tensor - 1) - with tf.control_dependencies([assert_op]): - out = tf.nn.conv2d( - input_tensor, filter_value, strides=(1, 1, 1, 1), padding="SAME") - return [input_tensor], [out] - - def build_inputs(parameters, sess, inputs, outputs): - input_values = create_tensor_data(tf.float32, parameters["input_shape"]) - return [input_values], sess.run( - outputs, feed_dict=dict(zip(inputs, [input_values]))) - - extra_toco_options = ExtraTocoOptions() - extra_toco_options.drop_control_dependency = True - make_zip_of_tests( - options, - test_parameters, - build_graph, - build_inputs, - extra_toco_options, - expected_tf_failures=3) From d1157c976bd605b36506151493439ebd2f80bded Mon Sep 17 00:00:00 2001 From: Brian Zhao Date: Thu, 18 Jun 2020 10:34:15 -0700 Subject: [PATCH 192/867] Adding missing kernel dependency to fix test for windows. PiperOrigin-RevId: 317132189 Change-Id: Ia59020db1c96301af4b9723170eb7808b66ef5b5 --- tensorflow/c/experimental/saved_model/core/ops/BUILD | 4 +--- 1 file changed, 1 insertion(+), 3 deletions(-) diff --git a/tensorflow/c/experimental/saved_model/core/ops/BUILD b/tensorflow/c/experimental/saved_model/core/ops/BUILD index 8c4c41c6d75..332b92bec45 100644 --- a/tensorflow/c/experimental/saved_model/core/ops/BUILD +++ b/tensorflow/c/experimental/saved_model/core/ops/BUILD @@ -78,14 +78,12 @@ tf_cc_test( srcs = [ "variable_ops_test.cc", ], - tags = [ - "no_windows", # TODO(b/159210739): Remove this tag after fixing the bug. - ], deps = [ ":owned_eager_context", ":owned_tensor", ":owned_tensor_handle", ":variable_ops", + "//tensorflow/core:all_kernels", "//tensorflow/core:framework", "//tensorflow/core:lib", "//tensorflow/core:protos_all_cc", From cf00e559d7e9649714be26173c44f9d6697e702e Mon Sep 17 00:00:00 2001 From: Frank Chen Date: Thu, 18 Jun 2020 10:39:58 -0700 Subject: [PATCH 193/867] Add TPU configuration ops to default TensorFlow build PiperOrigin-RevId: 317133514 Change-Id: I33bc6d7fdbba5915bd0d1291d4e086139c07eb14 --- .bazelrc | 4 + tensorflow/BUILD | 7 + tensorflow/core/BUILD | 3 + tensorflow/core/tpu/BUILD | 14 ++ tensorflow/core/tpu/kernels/BUILD | 30 +++- .../core/tpu/kernels/tpu_compile_c_api.h | 27 +-- .../core/tpu/kernels/tpu_compile_op_common.cc | 2 +- .../core/tpu/kernels/tpu_configuration_ops.cc | 29 +-- .../core/tpu/kernels/tpu_mesh_state_c_api.h | 15 +- .../tpu/kernels/tpu_mesh_state_interface.h | 11 +- tensorflow/core/tpu/kernels/tpu_util_c_api.h | 6 + tensorflow/core/tpu/tpu_library_init_fns.inc | 166 ++++++++++++++++++ tensorflow/core/tpu/tpu_library_loader.cc | 75 ++++---- tensorflow/core/tpu/tpu_library_loader.h | 16 ++ .../core/tpu/tpu_library_loader_windows.cc | 10 ++ tensorflow/stream_executor/tpu/BUILD | 14 +- .../stream_executor/tpu/tpu_executor.cc | 129 ++++++++------ .../stream_executor/tpu/tpu_executor_c_api.h | 92 +++++++++- .../stream_executor/tpu/tpu_node_context.cc | 14 +- .../tpu/tpu_node_context_c_api.h | 12 ++ .../stream_executor/tpu/tpu_platform.cc | 64 ++++--- tensorflow/stream_executor/tpu/tpu_platform.h | 6 +- tensorflow/stream_executor/tpu/tpu_stream.h | 25 ++- tensorflow/stream_executor/tpu/tpu_timer.h | 13 +- .../tpu/tpu_transfer_manager.cc | 20 ++- tensorflow/tensorflow.bzl | 7 + 26 files changed, 627 insertions(+), 184 deletions(-) create mode 100644 tensorflow/core/tpu/tpu_library_init_fns.inc diff --git a/.bazelrc b/.bazelrc index e21a1a32917..e67c3eecc3b 100644 --- a/.bazelrc +++ b/.bazelrc @@ -39,6 +39,7 @@ # # Feature and Third party library support options: # xla: Build TF with XLA +# tpu: Build TF with TPU support # using_cuda: CUDA is available to build system. # cuda: Build with full cuda support. # rocm: Build with AMD GPU support (rocm). @@ -180,6 +181,9 @@ build:dbg --cxxopt -DTF_LITE_DISABLE_X86_NEON # AWS SDK must be compiled in release mode. see: https://github.com/tensorflow/tensorflow/issues/37498 build:dbg --copt -DDEBUG_BUILD +# Config to build TPU backend +build:tpu --define=with_tpu_support=true + build:tensorrt --action_env TF_NEED_TENSORRT=1 build:rocm --crosstool_top=@local_config_rocm//crosstool:toolchain diff --git a/tensorflow/BUILD b/tensorflow/BUILD index bd0619b0c05..d00608ccc98 100644 --- a/tensorflow/BUILD +++ b/tensorflow/BUILD @@ -467,6 +467,13 @@ config_setting( visibility = ["//visibility:public"], ) +# This flag enables experimental TPU support +config_setting( + name = "with_tpu_support", + values = {"define": "with_tpu_support=true"}, + visibility = ["//visibility:public"], +) + # Specifies via a config setting if this is a mobile build or not, makes # it easier to combine settings later. selects.config_setting_group( diff --git a/tensorflow/core/BUILD b/tensorflow/core/BUILD index 50f1f2527a5..7f1c1bd549b 100644 --- a/tensorflow/core/BUILD +++ b/tensorflow/core/BUILD @@ -72,6 +72,7 @@ load( "if_ios", "if_mobile", "if_not_windows", + "if_tpu", "tf_android_core_proto_headers", "tf_cc_test", "tf_cc_test_mkl", @@ -1093,6 +1094,8 @@ cc_library( ]) + if_tensorrt([ "//tensorflow/compiler/tf2tensorrt:trt_engine_resource_op_kernels", "//tensorflow/compiler/tf2tensorrt:trt_op_kernels", + ]) + if_tpu([ + "//tensorflow/core/tpu/kernels", ]), ) diff --git a/tensorflow/core/tpu/BUILD b/tensorflow/core/tpu/BUILD index d4bcdfd52c5..9e89cd69235 100644 --- a/tensorflow/core/tpu/BUILD +++ b/tensorflow/core/tpu/BUILD @@ -103,6 +103,7 @@ cc_library( ":libtftpu_header", "//tensorflow/c:tf_status", ], + alwayslink = True, ) cc_library( @@ -116,7 +117,20 @@ cc_library( deps = [ ":libtftpu_header", ":tpu_config_c_api", + ":tpu_library_init_fns", "//tensorflow/core/platform:errors", "//tensorflow/core/platform:status", + "//tensorflow/core/tpu/kernels:tpu_compile_c_api_hdrs", + "//tensorflow/core/tpu/kernels:tpu_mesh_state_c_api_hdrs", + "//tensorflow/core/tpu/kernels:tpu_util_c_api_hdrs", + "//tensorflow/stream_executor/tpu:tpu_executor_c_api_hdrs", + "//tensorflow/stream_executor/tpu:tpu_node_context_c_api_hdrs", + "//tensorflow/stream_executor/tpu:tpu_platform_hdrs", ], ) + +cc_library( + name = "tpu_library_init_fns", + hdrs = ["tpu_library_init_fns.inc"], + visibility = ["//visibility:public"], +) diff --git a/tensorflow/core/tpu/kernels/BUILD b/tensorflow/core/tpu/kernels/BUILD index 9ba9ad61aa0..94d3c8edf2b 100644 --- a/tensorflow/core/tpu/kernels/BUILD +++ b/tensorflow/core/tpu/kernels/BUILD @@ -3,6 +3,10 @@ load( "//tensorflow/core/platform:build_config.bzl", "tf_proto_library_cc", ) +load( + "//tensorflow:tensorflow.bzl", + "tf_kernel_library", +) package( default_visibility = [ @@ -12,6 +16,12 @@ package( licenses = ["notice"], # Apache 2.0 ) +tf_kernel_library( + name = "kernels", + visibility = ["//visibility:public"], + deps = [":tpu_configuration_ops"], +) + cc_library( name = "tpu_compile_op_common", srcs = ["tpu_compile_op_common.cc"], @@ -50,7 +60,7 @@ cc_library( hdrs = ["tpu_compile_op_options.h"], ) -cc_library( +tf_kernel_library( name = "tpu_configuration_ops", srcs = ["tpu_configuration_ops.cc"], hdrs = ["tpu_configuration_ops.h"], @@ -75,12 +85,13 @@ cc_library( name = "tpu_compile_c_api_hdrs", hdrs = ["tpu_compile_c_api.h"], deps = [ - ":tpu_mesh_state_c_api", + ":tpu_mesh_state_c_api_hdrs", ":tpu_ops_common_c_api_hdrs", ":tpu_program_c_api_hdrs", - "//tensorflow/c:tf_datatype", + "//tensorflow/core/tpu:libtftpu_header", "//tensorflow/stream_executor/tpu:proto_helper", ], + alwayslink = True, ) tf_proto_library_cc( @@ -197,8 +208,10 @@ cc_library( ) cc_library( - name = "tpu_mesh_state_c_api", + name = "tpu_mesh_state_c_api_hdrs", hdrs = ["tpu_mesh_state_c_api.h"], + deps = ["//tensorflow/core/tpu:libtftpu_header"], + alwayslink = True, ) cc_library( @@ -207,12 +220,11 @@ cc_library( hdrs = ["tpu_mesh_state_interface.h"], deps = [ ":tpu_compile_c_api_hdrs", - ":tpu_mesh_state_c_api", + ":tpu_mesh_state_c_api_hdrs", "//tensorflow/compiler/xla/service", "//tensorflow/core:framework", - "//tensorflow/core/platform:errors", "//tensorflow/core/protobuf/tpu:compile_metadata_proto_cc", - "//tensorflow/core/tpu:tpu_config_c_api", + "//tensorflow/core/tpu:tpu_library_loader", ], ) @@ -371,13 +383,16 @@ cc_library( name = "tpu_util_c_api_hdrs", hdrs = ["tpu_util_c_api.h"], deps = [ + "//tensorflow/core/tpu:libtftpu_header", "//tensorflow/stream_executor/tpu:proto_helper", ], + alwayslink = True, ) cc_library( name = "tpu_ops_common_c_api_hdrs", hdrs = ["tpu_ops_common_c_api.h"], + alwayslink = True, ) cc_library( @@ -387,6 +402,7 @@ cc_library( ":tpu_ops_common_c_api_hdrs", "//tensorflow/stream_executor/tpu:proto_helper", ], + alwayslink = True, ) cc_library( diff --git a/tensorflow/core/tpu/kernels/tpu_compile_c_api.h b/tensorflow/core/tpu/kernels/tpu_compile_c_api.h index 70e3a7d2340..d1546ed9610 100644 --- a/tensorflow/core/tpu/kernels/tpu_compile_c_api.h +++ b/tensorflow/core/tpu/kernels/tpu_compile_c_api.h @@ -18,6 +18,7 @@ limitations under the License. #include "tensorflow/core/tpu/kernels/tpu_mesh_state_c_api.h" #include "tensorflow/core/tpu/kernels/tpu_ops_common_c_api.h" #include "tensorflow/core/tpu/kernels/tpu_program_c_api.h" +#include "tensorflow/core/tpu/libtftpu.h" #include "tensorflow/stream_executor/tpu/proto_helper.h" enum TpuCoreTypeEnum { @@ -44,35 +45,41 @@ struct CompilationCacheKeyProperty { extern "C" { // Returns the number of available TPU core count. -int TpuTopology_AvailableCoreCount(const XLA_TpuMeshState* mesh_state, - TpuCoreTypeEnum tpu_core_type); +TFTPU_CAPI_EXPORT int TpuTopology_AvailableCoreCount( + const XLA_TpuMeshState* mesh_state, TpuCoreTypeEnum tpu_core_type); // Creates a unique compilation cache `key` used for `put` and `get` operations. // Returned buffer is heap-allocated and must be owned. -const char* TpuCompile_CreateCompilationCacheKey( +TFTPU_CAPI_EXPORT const char* TpuCompile_CreateCompilationCacheKey( CompilationCacheKeyProperty property); // Creates a guaranteed const fingerprint. Guarantee const is normally used in // TPU inference to avoid re-copying unchanged variables onto the TPU device. // It promises the value is identical for every execution in the same session // even if the actual value changes in later executions. -uint64_t TpuCompile_CreateGuaranteedConstFingerprint(uint64_t fingerprint, - const char* data, - size_t size); +TFTPU_CAPI_EXPORT uint64_t TpuCompile_CreateGuaranteedConstFingerprint( + uint64_t fingerprint, const char* data, size_t size); // Executes the computations using XLA TPU compiler and returns TPU programs // ready for execution. -void TpuCompile_CompileAheadOfTime( - TpuSerializedProto aot_compilation_request, - XLA_TpuProgram** tpu_programs[], +TFTPU_CAPI_EXPORT void TpuCompile_CompileAheadOfTime( + TpuSerializedProto aot_compilation_request, XLA_TpuProgram** tpu_programs[], size_t* count, SE_Status* status); // Builds `DeviceAssignment` from `TpuCompileMetadata` serialized proto. -void TpuCompile_BuildXLADeviceAssignment( +TFTPU_CAPI_EXPORT void TpuCompile_BuildXLADeviceAssignment( TpuSerializedProto serialized_tpu_compile_metadata, const XLA_TpuMeshState* mesh_state, TpuSerializedProto* serialized_device_assignment, SE_Status* status); +struct TfTpu_CompileApiFn { + TFTPU_ADD_FN_IN_STRUCT(TpuTopology_AvailableCoreCount); + TFTPU_ADD_FN_IN_STRUCT(TpuCompile_CreateCompilationCacheKey); + TFTPU_ADD_FN_IN_STRUCT(TpuCompile_CreateGuaranteedConstFingerprint); + TFTPU_ADD_FN_IN_STRUCT(TpuCompile_CompileAheadOfTime); + TFTPU_ADD_FN_IN_STRUCT(TpuCompile_BuildXLADeviceAssignment); +}; + } // extern "C" #endif // TENSORFLOW_CORE_TPU_KERNELS_TPU_COMPILE_C_API_H_ diff --git a/tensorflow/core/tpu/kernels/tpu_compile_op_common.cc b/tensorflow/core/tpu/kernels/tpu_compile_op_common.cc index 7ab1c9b8027..92d1fa1337e 100644 --- a/tensorflow/core/tpu/kernels/tpu_compile_op_common.cc +++ b/tensorflow/core/tpu/kernels/tpu_compile_op_common.cc @@ -353,7 +353,7 @@ Status TpuCompileOpKernelCommon::CompileTFFunctionToHlo( return; } - LogAndExit(42); + std::quick_exit(42); } /* static */ Status TpuCompileOpKernelCommon::GetDynamicShapes( diff --git a/tensorflow/core/tpu/kernels/tpu_configuration_ops.cc b/tensorflow/core/tpu/kernels/tpu_configuration_ops.cc index 7fa345d735c..12a3256a44f 100644 --- a/tensorflow/core/tpu/kernels/tpu_configuration_ops.cc +++ b/tensorflow/core/tpu/kernels/tpu_configuration_ops.cc @@ -27,6 +27,7 @@ limitations under the License. #include "tensorflow/core/tpu/tpu_config_c_api.h" #include "tensorflow/core/tpu/tpu_configuration.h" #include "tensorflow/core/tpu/tpu_defs.h" +#include "tensorflow/core/tpu/tpu_library_loader.h" #include "tensorflow/stream_executor/tpu/proto_helper.h" namespace tensorflow { @@ -97,13 +98,14 @@ void ConfigureDistributedTpuOp::Compute(OpKernelContext* ctx) { OP_REQUIRES_OK(ctx, DeleteIfExists( rmgr, tpu::kTpuMeshCommonStateResourceName)); - ConfigureDistributedTpuOp_DoWork( + tpu::ConfigApiFn()->ConfigureDistributedTpuOp_DoWorkFn( num_devices_per_host.size(), num_devices_per_host.data(), &host_config_output_size, &host_config_output, status); - OP_REQUIRES_OK(ctx, rmgr->Create(rmgr->default_container(), - tpu::kTpuMeshCommonStateResourceName, - tpu::TpuMeshStateInterface::Create())); + auto* tpu_mesh = tpu::TpuMeshStateInterface::Create(); + OP_REQUIRES_OK(ctx, + rmgr->Create(rmgr->default_container(), + tpu::kTpuMeshCommonStateResourceName, tpu_mesh)); Tensor* ctx_output; OP_REQUIRES_OK(ctx, ctx->allocate_output(0, TensorShape({}), &ctx_output)); @@ -112,7 +114,8 @@ void ConfigureDistributedTpuOp::Compute(OpKernelContext* ctx) { OP_REQUIRES_OK(ctx, StatusFromTF_Status(status)); TF_DeleteStatus(status); - TpuConfigurationApi_FreeCharArray(host_config_output); + + tpu::ConfigApiFn()->TpuConfigurationApi_FreeCharArrayFn(host_config_output); VLOG(1) << "ConfigureDistributedTpuOp done"; } @@ -171,7 +174,7 @@ void WaitForDistributedTpuOp::Compute(OpKernelContext* ctx) { OP_REQUIRES_OK(ctx, GetTpuMeshStateInterface(rmgr, &mesh_state)); core::ScopedUnref mesh_state_unref(mesh_state); - WaitForDistributedTpuOp_DoWork( + tpu::ConfigApiFn()->WaitForDistributedTpuOp_DoWorkFn( num_hosts, num_devices_per_host, const_cast(mapping_arg.data()), mesh_state, &tpu_topology_output_size, &tpu_topology_output, status); @@ -183,7 +186,7 @@ void WaitForDistributedTpuOp::Compute(OpKernelContext* ctx) { OP_REQUIRES_OK(ctx, StatusFromTF_Status(status)); TF_DeleteStatus(status); - TpuConfigurationApi_FreeCharArray(tpu_topology_output); + tpu::ConfigApiFn()->TpuConfigurationApi_FreeCharArrayFn(tpu_topology_output); VLOG(1) << "WaitForDistributedTpuOp done"; } @@ -196,7 +199,7 @@ void ShutdownDistributedTpuOp::Compute(OpKernelContext* ctx) { OP_REQUIRES_OK(ctx, DeleteIfExists( GetTPUConfigResourceMgr(), tpu::kTpuMeshCommonStateResourceName)); - ShutdownDistributedTpuOp_DoWork(status); + tpu::ConfigApiFn()->ShutdownDistributedTpuOp_DoWorkFn(status); OP_REQUIRES_OK(ctx, StatusFromTF_Status(status)); TF_DeleteStatus(status); @@ -213,7 +216,7 @@ void InitializeHostForDistributedTpuOp::Compute(OpKernelContext* ctx) { int32_t* device_id_output; TF_Status* status = TF_NewStatus(); - InitializeHostForDistributedTpuOp_DoWork( + tpu::ConfigApiFn()->InitializeHostForDistributedTpuOp_DoWorkFn( tpu_host_config.size(), tpu_host_config.data(), enable_whole_mesh_compilations_, &device_id_output_size, &device_id_output, status); @@ -230,7 +233,7 @@ void InitializeHostForDistributedTpuOp::Compute(OpKernelContext* ctx) { OP_REQUIRES_OK(ctx, StatusFromTF_Status(status)); TF_DeleteStatus(status); - TpuConfigurationApi_FreeInt32Array(device_id_output); + tpu::ConfigApiFn()->TpuConfigurationApi_FreeInt32ArrayFn(device_id_output); VLOG(1) << "InitializeHostForDistributedTpuOp done"; } @@ -242,7 +245,8 @@ void SetGlobalTPUArrayOp::Compute(OpKernelContext* ctx) { auto tpu_topology = ctx->input(0).scalar()(); TF_Status* status = TF_NewStatus(); - SetGlobalTPUArrayOp_DoWork(tpu_topology.size(), tpu_topology.data(), status); + tpu::ConfigApiFn()->SetGlobalTPUArrayOp_DoWorkFn(tpu_topology.size(), + tpu_topology.data(), status); OP_REQUIRES_OK(ctx, StatusFromTF_Status(status)); TF_DeleteStatus(status); @@ -257,7 +261,8 @@ void DisconnectDistributedTpuChipsOp::Compute(OpKernelContext* ctx) { TF_Status* status = TF_NewStatus(); int32_t number_of_chips_output = 0; - DisconnectDistributedTpuChipsOp_DoWork(&number_of_chips_output, status); + tpu::ConfigApiFn()->DisconnectDistributedTpuChipsOp_DoWorkFn( + &number_of_chips_output, status); Tensor* ctx_output; OP_REQUIRES_OK(ctx, ctx->allocate_output(0, TensorShape({}), &ctx_output)); diff --git a/tensorflow/core/tpu/kernels/tpu_mesh_state_c_api.h b/tensorflow/core/tpu/kernels/tpu_mesh_state_c_api.h index 3ed65fe5cc4..a6434d7d2fd 100644 --- a/tensorflow/core/tpu/kernels/tpu_mesh_state_c_api.h +++ b/tensorflow/core/tpu/kernels/tpu_mesh_state_c_api.h @@ -15,20 +15,29 @@ limitations under the License. #ifndef TENSORFLOW_CORE_TPU_KERNELS_TPU_MESH_STATE_C_API_H_ #define TENSORFLOW_CORE_TPU_KERNELS_TPU_MESH_STATE_C_API_H_ +#include "tensorflow/core/tpu/libtftpu.h" + typedef struct XLA_TpuMeshState XLA_TpuMeshState; extern "C" { // Creates a new TPU mesh state object. -XLA_TpuMeshState* TpuMeshState_Create(); +TFTPU_CAPI_EXPORT XLA_TpuMeshState* TpuMeshState_Create(); // Deletes the given TPU `mesh_state` object. Once deleted the object is // unusable. -void TpuMeshState_Free(XLA_TpuMeshState* mesh_state); +TFTPU_CAPI_EXPORT void TpuMeshState_Free(XLA_TpuMeshState* mesh_state); // Returns a pointer to an opaque mesh data structure used internally. -void* TpuMeshState_MeshCommonState(XLA_TpuMeshState* mesh_state); +TFTPU_CAPI_EXPORT void* TpuMeshState_MeshCommonState( + XLA_TpuMeshState* mesh_state); } // extern "C" +struct TfTpu_MeshStateApiFn { + TFTPU_ADD_FN_IN_STRUCT(TpuMeshState_Create); + TFTPU_ADD_FN_IN_STRUCT(TpuMeshState_Free); + TFTPU_ADD_FN_IN_STRUCT(TpuMeshState_MeshCommonState); +}; + #endif // TENSORFLOW_CORE_TPU_KERNELS_TPU_MESH_STATE_C_API_H_ diff --git a/tensorflow/core/tpu/kernels/tpu_mesh_state_interface.h b/tensorflow/core/tpu/kernels/tpu_mesh_state_interface.h index 34202a78718..3eff3be4915 100644 --- a/tensorflow/core/tpu/kernels/tpu_mesh_state_interface.h +++ b/tensorflow/core/tpu/kernels/tpu_mesh_state_interface.h @@ -21,6 +21,7 @@ limitations under the License. #include "tensorflow/core/protobuf/tpu/compile_metadata.pb.h" #include "tensorflow/core/tpu/kernels/tpu_compile_c_api.h" #include "tensorflow/core/tpu/kernels/tpu_mesh_state_c_api.h" +#include "tensorflow/core/tpu/tpu_library_loader.h" namespace tensorflow { @@ -38,19 +39,19 @@ class TpuMeshStateInterface : public tensorflow::ResourceBase { ~TpuMeshStateInterface() override { if (mesh_state_ != nullptr) { - TpuMeshState_Free(mesh_state_); + MeshStateApiFn()->TpuMeshState_FreeFn(mesh_state_); } } static TpuMeshStateInterface* Create() { - return new TpuMeshStateInterface(TpuMeshState_Create()); + return new TpuMeshStateInterface(MeshStateApiFn()->TpuMeshState_CreateFn()); } const XLA_TpuMeshState* data() const { return mesh_state_; } tensorflow::TpuMeshCommonState* mesh_common_state() const { return static_cast( - TpuMeshState_MeshCommonState(mesh_state_)); + MeshStateApiFn()->TpuMeshState_MeshCommonStateFn(mesh_state_)); } // Returns whether we should include the device assignment as a static field @@ -62,8 +63,8 @@ class TpuMeshStateInterface : public tensorflow::ResourceBase { // Static device assignment enables XLA to perform certain optimization when // all cores are used in the replicated computation. return metadata.num_cores_per_replica() * metadata.num_replicas() == - TpuTopology_AvailableCoreCount(mesh_state_, - tpu_core_type); + CompileApiFn()->TpuTopology_AvailableCoreCountFn(mesh_state_, + tpu_core_type); } string DebugString() const override { return "TpuMeshStateInterface"; } diff --git a/tensorflow/core/tpu/kernels/tpu_util_c_api.h b/tensorflow/core/tpu/kernels/tpu_util_c_api.h index 4d992449cfc..32b946d56c9 100644 --- a/tensorflow/core/tpu/kernels/tpu_util_c_api.h +++ b/tensorflow/core/tpu/kernels/tpu_util_c_api.h @@ -15,6 +15,7 @@ limitations under the License. #ifndef TENSORFLOW_CORE_TPU_KERNELS_TPU_UTIL_C_API_H_ #define TENSORFLOW_CORE_TPU_KERNELS_TPU_UTIL_C_API_H_ +#include "tensorflow/core/tpu/libtftpu.h" #include "tensorflow/stream_executor/tpu/proto_helper.h" typedef struct SE_Status SE_Status; @@ -32,4 +33,9 @@ void TpuCompile_ToTpuShapeRepresentation( } // extern "C" +struct TfTpu_UtilApiFn { + TFTPU_ADD_FN_IN_STRUCT(TpuCompile_IsTpuCompilationEnabled); + TFTPU_ADD_FN_IN_STRUCT(TpuCompile_ToTpuShapeRepresentation); +}; + #endif // TENSORFLOW_CORE_TPU_KERNELS_TPU_UTIL_C_API_H_ diff --git a/tensorflow/core/tpu/tpu_library_init_fns.inc b/tensorflow/core/tpu/tpu_library_init_fns.inc new file mode 100644 index 00000000000..e21d7f195ad --- /dev/null +++ b/tensorflow/core/tpu/tpu_library_init_fns.inc @@ -0,0 +1,166 @@ +namespace { + +tensorflow::Status SetTpuConfigStructFns(void* library_handle) { + auto* config_fn = tensorflow::tpu::ConfigApiFn(); + + TFTPU_SET_FN(config_fn, ConfigureDistributedTpuOp_DoWork); + TFTPU_SET_FN(config_fn, WaitForDistributedTpuOp_DoWork); + TFTPU_SET_FN(config_fn, ShutdownDistributedTpuOp_DoWork); + TFTPU_SET_FN(config_fn, InitializeHostForDistributedTpuOp_DoWork); + TFTPU_SET_FN(config_fn, SetGlobalTPUArrayOp_DoWork); + TFTPU_SET_FN(config_fn, DisconnectDistributedTpuChipsOp_DoWork); + TFTPU_SET_FN(config_fn, TpuConfigurationApi_FreeCharArray); + TFTPU_SET_FN(config_fn, TpuConfigurationApi_FreeInt32Array); + + return tensorflow::Status::OK(); +} + +tensorflow::Status SetTpuMeshStateStructFns(void* library_handle) { + auto* mesh_state_fn = tensorflow::tpu::MeshStateApiFn(); + + TFTPU_SET_FN(mesh_state_fn, TpuMeshState_Create); + TFTPU_SET_FN(mesh_state_fn, TpuMeshState_Free); + TFTPU_SET_FN(mesh_state_fn, TpuMeshState_MeshCommonState); + + return tensorflow::Status::OK(); +} + +tensorflow::Status SetCompileStructFn(void* library_handle) { + auto* compile_fn = tensorflow::tpu::CompileApiFn(); + + TFTPU_SET_FN(compile_fn, TpuTopology_AvailableCoreCount); + TFTPU_SET_FN(compile_fn, TpuCompile_CreateCompilationCacheKey); + TFTPU_SET_FN(compile_fn, TpuCompile_CreateGuaranteedConstFingerprint); + TFTPU_SET_FN(compile_fn, TpuCompile_CompileAheadOfTime); + TFTPU_SET_FN(compile_fn, TpuCompile_BuildXLADeviceAssignment); + + return tensorflow::Status::OK(); +} + +tensorflow::Status SetExecutorStructFn(void* library_handle) { + auto* executor_fn = tensorflow::tpu::ExecutorApiFn(); + + TFTPU_SET_FN(executor_fn, TpuPlatform_New); + TFTPU_SET_FN(executor_fn, TpuPlatform_Free); + TFTPU_SET_FN(executor_fn, TpuPlatform_Initialize); + TFTPU_SET_FN(executor_fn, TpuPlatform_Initialized); + TFTPU_SET_FN(executor_fn, TpuPlatform_GetExecutor); + TFTPU_SET_FN(executor_fn, TpuPlatform_Id); + TFTPU_SET_FN(executor_fn, TpuPlatform_VisibleDeviceCount); + TFTPU_SET_FN(executor_fn, TpuPlatform_TpuMemoryLimit); + TFTPU_SET_FN(executor_fn, TpuPlatform_ShouldRegisterTpuDeviceToDeviceCopy); + + TFTPU_SET_FN(executor_fn, TpuExecutor_Init); + TFTPU_SET_FN(executor_fn, TpuExecutor_Free); + TFTPU_SET_FN(executor_fn, TpuExecutor_PlatformDeviceCount); + TFTPU_SET_FN(executor_fn, TpuExecutor_Allocate); + TFTPU_SET_FN(executor_fn, TpuExecutor_Deallocate); + TFTPU_SET_FN(executor_fn, TpuExecutor_GetAllocatorStats); + TFTPU_SET_FN(executor_fn, TpuExecutor_DeviceMemoryUsage); + TFTPU_SET_FN(executor_fn, TpuExecutor_AllocateStream); + TFTPU_SET_FN(executor_fn, TpuExecutor_DeallocateStream); + TFTPU_SET_FN(executor_fn, TpuExecutor_CreateStreamDependency); + TFTPU_SET_FN(executor_fn, TpuExecutor_GetStatus); + TFTPU_SET_FN(executor_fn, TpuExecutor_AllocateEvent); + TFTPU_SET_FN(executor_fn, TpuExecutor_DeallocateEvent); + TFTPU_SET_FN(executor_fn, TpuExecutor_PollForEventStatus); + TFTPU_SET_FN(executor_fn, TpuExecutor_RecordEvent); + TFTPU_SET_FN(executor_fn, TpuExecutor_WaitForEvent); + TFTPU_SET_FN(executor_fn, TpuExecutor_AllocateTimer); + TFTPU_SET_FN(executor_fn, TpuExecutor_DeallocateTimer); + TFTPU_SET_FN(executor_fn, TpuExecutor_StartTimer); + TFTPU_SET_FN(executor_fn, TpuExecutor_StopTimer); + TFTPU_SET_FN(executor_fn, TpuExecutor_SynchronousMemcpyToHost); + TFTPU_SET_FN(executor_fn, TpuExecutor_SynchronousMemcpyFromHost); + TFTPU_SET_FN(executor_fn, TpuExecutor_MemcpyToHost); + TFTPU_SET_FN(executor_fn, TpuExecutor_MemcpyFromHost); + TFTPU_SET_FN(executor_fn, TpuExecutor_EnqueueInfeed); + TFTPU_SET_FN(executor_fn, TpuExecutor_DequeueOutfeed); + TFTPU_SET_FN(executor_fn, TpuExecutor_WaitForInfeedReady); + TFTPU_SET_FN(executor_fn, TpuExecutor_WaitForOutfeedReady); + TFTPU_SET_FN(executor_fn, TpuExecutor_BlockHostUntilDone); + TFTPU_SET_FN(executor_fn, TpuExecutor_BlockUntilDoneOrFailed); + TFTPU_SET_FN(executor_fn, TpuExecutor_SyncAndForgetFailedStreams); + TFTPU_SET_FN(executor_fn, TpuExecutor_SynchronizeAllActivity); + + TFTPU_SET_FN(executor_fn, TpuStream_New); + TFTPU_SET_FN(executor_fn, TpuStream_Free); + TFTPU_SET_FN(executor_fn, TpuStream_Stream); + TFTPU_SET_FN(executor_fn, TpuStream_Status); + TFTPU_SET_FN(executor_fn, TpuStream_IsSameSharedMemoryLocation); + TFTPU_SET_FN(executor_fn, TpuStream_TpuEnqueueOnDeviceSendRecvLocal); + + TFTPU_SET_FN(executor_fn, TpuEvent_New); + TFTPU_SET_FN(executor_fn, TpuEvent_Free); + + TFTPU_SET_FN(executor_fn, TpuTimer_New); + TFTPU_SET_FN(executor_fn, TpuTimer_Free); + TFTPU_SET_FN(executor_fn, TpuTimer_Nanoseconds); + TFTPU_SET_FN(executor_fn, TpuTimer_Microseconds); + + TFTPU_SET_FN(executor_fn, TpuStatus_New); + TFTPU_SET_FN(executor_fn, TpuStatus_Create); + TFTPU_SET_FN(executor_fn, TpuStatus_Free); + TFTPU_SET_FN(executor_fn, TpuStatus_Message); + TFTPU_SET_FN(executor_fn, TpuStatus_Code); + TFTPU_SET_FN(executor_fn, TpuStatus_Ok); + + TFTPU_SET_FN(executor_fn, TpuStreamExecutorConfig_Default); + TFTPU_SET_FN(executor_fn, TpuStreamExecutorConfig_SetOrdinal); + TFTPU_SET_FN(executor_fn, TpuStreamExecutorConfig_Free); + + TFTPU_SET_FN(executor_fn, TpuDeviceDescription_New); + TFTPU_SET_FN(executor_fn, TpuDeviceDescription_Free); + + TFTPU_SET_FN(executor_fn, TpuExecutor_CreateDeviceDescription); + TFTPU_SET_FN(executor_fn, TpuExecutor_NewDeviceOptions); + TFTPU_SET_FN(executor_fn, TpuExecutor_FreeDeviceOptions); + TFTPU_SET_FN(executor_fn, TpuExecutor_HostCallback); + + TFTPU_SET_FN(executor_fn, TpuTransferManager_New); + TFTPU_SET_FN(executor_fn, TpuTransferManager_Free); + TFTPU_SET_FN(executor_fn, TpuTransferManager_PlatformId); + TFTPU_SET_FN(executor_fn, TpuTransferManager_HostShapeToDeviceShape); + TFTPU_SET_FN(executor_fn, TpuTransferManager_TransferLiteralToDeviceAsync); + TFTPU_SET_FN(executor_fn, TpuTransferManager_TransferLiteralFromDevice); + TFTPU_SET_FN(executor_fn, TpuTransferManager_GetByteSizeRequirement); + TFTPU_SET_FN(executor_fn, TpuTransferManager_WriteSingleTupleIndexTable); + + TFTPU_SET_FN(executor_fn, TpuComputationPlacer_New); + TFTPU_SET_FN(executor_fn, TpuComputationPlacer_Free); + + return tensorflow::Status::OK(); +} + +tensorflow::Status SetTpuNodeContextStructFns(void* library_handle) { + auto* node_context_fn = tensorflow::tpu::NodeContextApiFn(); + + TFTPU_SET_FN(node_context_fn, TpuNodeContext_Create); + TFTPU_SET_FN(node_context_fn, TpuNodeContext_Free); + TFTPU_SET_FN(node_context_fn, TpuNodeContext_StopChipHeartbeats); + TFTPU_SET_FN(node_context_fn, TpuNodeContext_CloseTpuHost); + + return tensorflow::Status::OK(); +} + +tensorflow::Status SetTpuUtilStructFns(void* library_handle) { + auto* util_fn = tensorflow::tpu::UtilApiFn(); + + TFTPU_SET_FN(util_fn, TpuCompile_IsTpuCompilationEnabled); + TFTPU_SET_FN(util_fn, TpuCompile_ToTpuShapeRepresentation); + + return tensorflow::Status::OK(); +} + +tensorflow::Status InitializeTpuStructFns(void* library_handle) { + TF_RETURN_IF_ERROR(SetTpuConfigStructFns(library_handle)); + TF_RETURN_IF_ERROR(SetTpuMeshStateStructFns(library_handle)); + TF_RETURN_IF_ERROR(SetCompileStructFn(library_handle)); + TF_RETURN_IF_ERROR(SetExecutorStructFn(library_handle)); + TF_RETURN_IF_ERROR(SetTpuNodeContextStructFns(library_handle)); + TF_RETURN_IF_ERROR(SetTpuUtilStructFns(library_handle)); + + return tensorflow::Status::OK(); +} + +} // namespace \ No newline at end of file diff --git a/tensorflow/core/tpu/tpu_library_loader.cc b/tensorflow/core/tpu/tpu_library_loader.cc index c89de142a9f..834b86e68a7 100644 --- a/tensorflow/core/tpu/tpu_library_loader.cc +++ b/tensorflow/core/tpu/tpu_library_loader.cc @@ -13,16 +13,23 @@ See the License for the specific language governing permissions and limitations under the License. ==============================================================================*/ +// TODO(frankchn): Rename to `tpu_api_dlsym_initializer` or similar. + #include "tensorflow/core/tpu/tpu_library_loader.h" #include -#define TFTPU_SET_FN(Struct, FnName) \ - Struct->FnName##Fn = \ - reinterpret_cast(dlsym(library_handle, #FnName)); - #include "tensorflow/core/platform/errors.h" #include "tensorflow/core/platform/status.h" +#include "tensorflow/stream_executor/tpu/tpu_node_context_c_api.h" +#include "tensorflow/stream_executor/tpu/tpu_platform.h" + +#define TFTPU_SET_FN(Struct, FnName) \ + Struct->FnName##Fn = \ + reinterpret_cast(dlsym(library_handle, #FnName)); \ + if (!(Struct->FnName##Fn)) { \ + LOG(ERROR) << #FnName " not available in this library."; \ + } // Reminder: Update tpu_library_loader_windows.cc if you are adding new publicly // visible methods. @@ -30,28 +37,7 @@ limitations under the License. namespace tensorflow { namespace tpu { -Status SetTpuInitializeStructFns(void* library_handle) { - auto* base_fn = InitializeApiFn(); - - TFTPU_SET_FN(base_fn, TfTpu_Initialize); - - return Status::OK(); -} - -Status SetTpuConfigStructFns(void* library_handle) { - auto* config_fn = ConfigApiFn(); - - TFTPU_SET_FN(config_fn, ConfigureDistributedTpuOp_DoWork); - TFTPU_SET_FN(config_fn, WaitForDistributedTpuOp_DoWork); - TFTPU_SET_FN(config_fn, ShutdownDistributedTpuOp_DoWork); - TFTPU_SET_FN(config_fn, InitializeHostForDistributedTpuOp_DoWork); - TFTPU_SET_FN(config_fn, SetGlobalTPUArrayOp_DoWork); - TFTPU_SET_FN(config_fn, DisconnectDistributedTpuChipsOp_DoWork); - TFTPU_SET_FN(config_fn, TpuConfigurationApi_FreeCharArray); - TFTPU_SET_FN(config_fn, TpuConfigurationApi_FreeInt32Array); - - return Status::OK(); -} +#include "tensorflow/core/tpu/tpu_library_init_fns.inc" TfTpu_BaseFn* InitializeApiFn() { static TfTpu_BaseFn base_fn; @@ -63,19 +49,48 @@ TfTpu_ConfigApiFn* ConfigApiFn() { return &config_api_fn; } +TfTpu_MeshStateApiFn* MeshStateApiFn() { + static TfTpu_MeshStateApiFn mesh_state_api_fn; + return &mesh_state_api_fn; +} + +TfTpu_CompileApiFn* CompileApiFn() { + static TfTpu_CompileApiFn compile_api_fn; + return &compile_api_fn; +} + +TfTpu_ExecutorApiFn* ExecutorApiFn() { + static TfTpu_ExecutorApiFn executor_api_fn; + return &executor_api_fn; +} + +TfTpu_NodeContextApiFn* NodeContextApiFn() { + static TfTpu_NodeContextApiFn node_context_api_fn; + return &node_context_api_fn; +} + +TfTpu_UtilApiFn* UtilApiFn() { + static TfTpu_UtilApiFn util_api_fn; + return &util_api_fn; +} + Status InitializeTpuLibrary(void* library_handle) { bool shared_object_loaded = true; if (library_handle == nullptr) { - library_handle = dlopen(nullptr, RTLD_LAZY); + library_handle = dlopen(nullptr, RTLD_NOW); shared_object_loaded = false; } - TF_RETURN_IF_ERROR(SetTpuInitializeStructFns(library_handle)); - TF_RETURN_IF_ERROR(SetTpuConfigStructFns(library_handle)); + TF_RETURN_IF_ERROR(InitializeTpuStructFns(library_handle)); if (shared_object_loaded) { + // TODO(frankchn): Make initialization actually work // Initialize TPU platform when the platform code is loaded from a library. - InitializeApiFn()->TfTpu_InitializeFn(); + // InitializeApiFn()->TfTpu_InitializeFn(); + + // We should only register the TPU platform when the library is loaded. + // TODO(frankchn): Resolve the circular dependency and register the platform + // RegisterTpuPlatform(); } return Status::OK(); diff --git a/tensorflow/core/tpu/tpu_library_loader.h b/tensorflow/core/tpu/tpu_library_loader.h index a51948cf719..ba6c324707d 100644 --- a/tensorflow/core/tpu/tpu_library_loader.h +++ b/tensorflow/core/tpu/tpu_library_loader.h @@ -17,8 +17,13 @@ limitations under the License. #define TENSORFLOW_CORE_TPU_TPU_LIBRARY_LOADER_H_ #include "tensorflow/core/platform/status.h" +#include "tensorflow/core/tpu/kernels/tpu_compile_c_api.h" +#include "tensorflow/core/tpu/kernels/tpu_mesh_state_c_api.h" +#include "tensorflow/core/tpu/kernels/tpu_util_c_api.h" #include "tensorflow/core/tpu/libtftpu.h" #include "tensorflow/core/tpu/tpu_config_c_api.h" +#include "tensorflow/stream_executor/tpu/tpu_executor_c_api.h" +#include "tensorflow/stream_executor/tpu/tpu_node_context_c_api.h" // LINT.IfChange namespace tensorflow { @@ -26,10 +31,21 @@ namespace tpu { Status InitializeTpuLibrary(void* library_handle); +// TODO(frankchn): Separate out API functions from the loader. TfTpu_BaseFn* InitializeApiFn(); TfTpu_ConfigApiFn* ConfigApiFn(); +TfTpu_MeshStateApiFn* MeshStateApiFn(); + +TfTpu_CompileApiFn* CompileApiFn(); + +TfTpu_ExecutorApiFn* ExecutorApiFn(); + +TfTpu_NodeContextApiFn* NodeContextApiFn(); + +TfTpu_UtilApiFn* UtilApiFn(); + } // namespace tpu } // namespace tensorflow // LINT.ThenChange(//tensorflow/core/tpu/tpu_library_loader_windows.cc) diff --git a/tensorflow/core/tpu/tpu_library_loader_windows.cc b/tensorflow/core/tpu/tpu_library_loader_windows.cc index e7c25df415e..7cf1b5cdb1d 100644 --- a/tensorflow/core/tpu/tpu_library_loader_windows.cc +++ b/tensorflow/core/tpu/tpu_library_loader_windows.cc @@ -27,6 +27,16 @@ TfTpu_BaseFn* InitializeApiFn() { return nullptr; } TfTpu_ConfigApiFn* ConfigApiFn() { return nullptr; } +TfTpu_MeshStateApiFn* MeshStateApiFn() { return nullptr; } + +TfTpu_CompileApiFn* CompileApiFn() { return nullptr; } + +TfTpu_ExecutorApiFn* ExecutorApiFn() { return nullptr; } + +TfTpu_NodeContextApiFn* NodeContextApiFn() { return nullptr; } + +TfTpu_UtilApiFn* UtilApiFn() { return nullptr; } + Status InitializeTpuLibrary(void* library_handle) { return errors::Unimplemented( "Loading TPU library is not supported on Windows."); diff --git a/tensorflow/stream_executor/tpu/BUILD b/tensorflow/stream_executor/tpu/BUILD index 964f36b82c7..bf88e9809d0 100644 --- a/tensorflow/stream_executor/tpu/BUILD +++ b/tensorflow/stream_executor/tpu/BUILD @@ -11,20 +11,25 @@ package( cc_library( name = "tpu_executor_c_api_hdrs", hdrs = ["tpu_executor_c_api.h"], + visibility = ["//visibility:public"], deps = [ "//tensorflow/c:tf_attrtype", - "//tensorflow/c:tf_datatype", "//tensorflow/c:tf_status", + "//tensorflow/core/tpu:libtftpu_header", "//tensorflow/core/tpu/kernels:tpu_ops_common_c_api_hdrs", ], + alwayslink = True, ) cc_library( name = "tpu_node_context_c_api_hdrs", hdrs = ["tpu_node_context_c_api.h"], + visibility = ["//visibility:public"], deps = [ ":tpu_executor_c_api_hdrs", + "//tensorflow/core/tpu:libtftpu_header", ], + alwayslink = True, ) cc_library( @@ -65,6 +70,7 @@ cc_library( ":status_helper", ":tpu_executor_c_api_hdrs", ":tpu_stream_interface", + "//tensorflow/core/tpu:tpu_library_loader", "//tensorflow/stream_executor:stream", ], ) @@ -75,6 +81,7 @@ cc_library( deps = [ ":tpu_executor_c_api_hdrs", "//tensorflow/core/platform:types", + "//tensorflow/core/tpu:tpu_library_loader", "//tensorflow/stream_executor:stream", ], ) @@ -94,6 +101,7 @@ cc_library( ":tpu_timer", "//tensorflow/c:tf_status", "//tensorflow/core:lib", + "//tensorflow/core/tpu:tpu_library_loader", "//tensorflow/stream_executor:stream", "//tensorflow/stream_executor/lib", "@com_google_absl//absl/container:flat_hash_map", @@ -143,6 +151,7 @@ cc_library( "//tensorflow/compiler/xla/service:stream_pool", "//tensorflow/compiler/xla/service:transfer_manager", "//tensorflow/core:framework", + "//tensorflow/core/tpu:tpu_library_loader", "//tensorflow/stream_executor:device_memory_allocator", "//tensorflow/stream_executor/lib", "@com_google_absl//absl/memory", @@ -160,6 +169,7 @@ cc_library( ":tpu_platform_interface", "//tensorflow/c:tf_status", "//tensorflow/core/platform:types", + "//tensorflow/core/tpu:tpu_library_loader", "//tensorflow/stream_executor:stream", "@com_google_absl//absl/container:flat_hash_map", ], @@ -191,6 +201,7 @@ cc_library( "//tensorflow/compiler/xla:xla_data_proto_cc", "//tensorflow/compiler/xla/service:shaped_buffer", "//tensorflow/compiler/xla/service:transfer_manager", + "//tensorflow/core/tpu:tpu_library_loader", "//tensorflow/stream_executor:stream", ], ) @@ -217,6 +228,7 @@ cc_library( "//tensorflow/core/platform:types", "//tensorflow/stream_executor:multi_platform_manager", "//tensorflow/stream_executor:stream_executor_headers", + "//tensorflow/stream_executor:stream_executor_pimpl", ], ) diff --git a/tensorflow/stream_executor/tpu/tpu_executor.cc b/tensorflow/stream_executor/tpu/tpu_executor.cc index 03cab5801e6..cb1410880eb 100644 --- a/tensorflow/stream_executor/tpu/tpu_executor.cc +++ b/tensorflow/stream_executor/tpu/tpu_executor.cc @@ -17,6 +17,7 @@ limitations under the License. #include "tensorflow/c/tf_status.h" #include "tensorflow/core/lib/gtl/cleanup.h" +#include "tensorflow/core/tpu/tpu_library_loader.h" #include "tensorflow/stream_executor/device_memory.h" #include "tensorflow/stream_executor/lib/status.h" #include "tensorflow/stream_executor/tpu/c_api_conversions.h" @@ -33,63 +34,68 @@ namespace { using ::stream_executor::port::Status; } // namespace -TpuExecutor::~TpuExecutor() { TpuExecutor_Free(executor_); } +TpuExecutor::~TpuExecutor() { + tpu::ExecutorApiFn()->TpuExecutor_FreeFn(executor_); +} Status TpuExecutor::Init(int device_ordinal, ::stream_executor::DeviceOptions device_options) { StatusHelper status; SE_DeviceOptions* options = - TpuExecutor_NewDeviceOptions(device_options.flags()); - TpuExecutor_Init(executor_, device_ordinal, options, status.c_status); - TpuExecutor_FreeDeviceOptions(options); + tpu::ExecutorApiFn()->TpuExecutor_NewDeviceOptionsFn( + device_options.flags()); + tpu::ExecutorApiFn()->TpuExecutor_InitFn(executor_, device_ordinal, options, + status.c_status); + tpu::ExecutorApiFn()->TpuExecutor_FreeDeviceOptionsFn(options); return status.status(); } int TpuExecutor::PlatformDeviceCount() { - return TpuExecutor_PlatformDeviceCount(executor_); + return tpu::ExecutorApiFn()->TpuExecutor_PlatformDeviceCountFn(executor_); } void TpuExecutor::SyncAndForgetFailedStreams() { - TpuExecutor_SyncAndForgetFailedStreams(executor_); + tpu::ExecutorApiFn()->TpuExecutor_SyncAndForgetFailedStreamsFn(executor_); } bool TpuExecutor::SynchronizeAllActivity() { - return TpuExecutor_SynchronizeAllActivity(executor_); + return tpu::ExecutorApiFn()->TpuExecutor_SynchronizeAllActivityFn(executor_); } Status TpuExecutor::BlockHostUntilDone(Stream* stream) { StatusHelper status; - TpuExecutor_BlockHostUntilDone( + tpu::ExecutorApiFn()->TpuExecutor_BlockHostUntilDoneFn( executor_, stream_map().at(stream->implementation()), status.c_status); return status.status(); } Status TpuExecutor::BlockUntilDoneOrFailed() { StatusHelper status; - TpuExecutor_BlockUntilDoneOrFailed(executor_, status.c_status); + tpu::ExecutorApiFn()->TpuExecutor_BlockUntilDoneOrFailedFn(executor_, + status.c_status); return status.status(); } Status TpuExecutor::GetStatus(Stream* stream) { StatusHelper status; - TpuExecutor_GetStatus(executor_, stream_map().at(stream->implementation()), - status.c_status); + tpu::ExecutorApiFn()->TpuExecutor_GetStatusFn( + executor_, stream_map().at(stream->implementation()), status.c_status); return status.status(); } bool TpuExecutor::AllocateStream(Stream* stream) { - return TpuExecutor_AllocateStream(executor_, - stream_map().at(stream->implementation())); + return tpu::ExecutorApiFn()->TpuExecutor_AllocateStreamFn( + executor_, stream_map().at(stream->implementation())); } void TpuExecutor::DeallocateStream(Stream* stream) { - TpuExecutor_DeallocateStream(executor_, - stream_map().at(stream->implementation())); + tpu::ExecutorApiFn()->TpuExecutor_DeallocateStreamFn( + executor_, stream_map().at(stream->implementation())); stream_map().erase(stream->implementation()); } bool TpuExecutor::CreateStreamDependency(Stream* dependent, Stream* other) { - return TpuExecutor_CreateStreamDependency( + return tpu::ExecutorApiFn()->TpuExecutor_CreateStreamDependencyFn( executor_, stream_map().at(dependent->implementation()), stream_map().at(other->implementation())); } @@ -104,15 +110,15 @@ bool TpuExecutor::AllocateTimer(Timer* timer) { return true; } void TpuExecutor::DeallocateTimer(Timer* timer) {} bool TpuExecutor::StartTimer(Stream* stream, ::stream_executor::Timer* timer) { - return TpuExecutor_StartTimer(executor_, - stream_map().at(stream->implementation()), - timer_map_.at(timer->implementation())); + return tpu::ExecutorApiFn()->TpuExecutor_StartTimerFn( + executor_, stream_map().at(stream->implementation()), + timer_map_.at(timer->implementation())); } bool TpuExecutor::StopTimer(Stream* stream, ::stream_executor::Timer* timer) { - return TpuExecutor_StopTimer(executor_, - stream_map().at(stream->implementation()), - timer_map_.at(timer->implementation())); + return tpu::ExecutorApiFn()->TpuExecutor_StopTimerFn( + executor_, stream_map().at(stream->implementation()), + timer_map_.at(timer->implementation())); } stream_executor::Event::Status TpuExecutor::PollForEventStatus( @@ -148,7 +154,7 @@ Status TpuExecutor::WaitForEvent(Stream* stream, // Called by Timer::Timer std::unique_ptr<::stream_executor::internal::TimerInterface> TpuExecutor::GetTimerImplementation() { - SE_Timer* tpu_timer = TpuTimer_New(executor_); + SE_Timer* tpu_timer = tpu::ExecutorApiFn()->TpuTimer_NewFn(executor_); auto ptr = absl::make_unique(tpu_timer); timer_map_[ptr.get()] = tpu_timer; return ptr; @@ -157,7 +163,7 @@ TpuExecutor::GetTimerImplementation() { // Called by Stream::Stream std::unique_ptr<::stream_executor::internal::StreamInterface> TpuExecutor::GetStreamImplementation() { - SE_Stream* tpu_stream = TpuStream_New(executor_); + SE_Stream* tpu_stream = tpu::ExecutorApiFn()->TpuStream_NewFn(executor_); auto ptr = absl::make_unique(tpu_stream); stream_map()[ptr.get()] = tpu_stream; return ptr; @@ -166,34 +172,35 @@ TpuExecutor::GetStreamImplementation() { // Called by Event::Event std::unique_ptr<::stream_executor::internal::EventInterface> TpuExecutor::CreateEventImplementation() { - SE_Event* tpu_event = TpuEvent_New(executor_); + SE_Event* tpu_event = tpu::ExecutorApiFn()->TpuEvent_NewFn(executor_); auto ptr = absl::make_unique(tpu_event); event_map()[ptr.get()] = tpu_event; return ptr; } DeviceMemoryBase TpuExecutor::Allocate(uint64 size, int64 memory_space) { - SE_DeviceMemoryBase se_base = - TpuExecutor_Allocate(executor_, size, memory_space); + SE_DeviceMemoryBase se_base = tpu::ExecutorApiFn()->TpuExecutor_AllocateFn( + executor_, size, memory_space); return TpuConversions::SE_DeviceMemoryBaseToDeviceMemoryBase(se_base); } void TpuExecutor::Deallocate(const DeviceMemoryBase& memory) { SE_DeviceMemoryBase se_base = TpuConversions::DeviceMemoryBaseToSE_DeviceMemoryBase(memory); - TpuExecutor_Deallocate(executor_, &se_base); + tpu::ExecutorApiFn()->TpuExecutor_DeallocateFn(executor_, &se_base); } void TpuExecutor::Deallocate(DeviceMemoryBase* memory) { SE_DeviceMemoryBase se_base = TpuConversions::DeviceMemoryBaseToSE_DeviceMemoryBase(*memory); - TpuExecutor_Deallocate(executor_, &se_base); + tpu::ExecutorApiFn()->TpuExecutor_DeallocateFn(executor_, &se_base); } bool TpuExecutor::DeviceMemoryUsage(int64* free, int64* total) const { int64_t _free; int64_t _total; - if (TpuExecutor_DeviceMemoryUsage(executor_, &_free, &_total)) { + if (tpu::ExecutorApiFn()->TpuExecutor_DeviceMemoryUsageFn(executor_, &_free, + &_total)) { *free = _free; *total = _total; return true; @@ -204,7 +211,8 @@ bool TpuExecutor::DeviceMemoryUsage(int64* free, int64* total) const { absl::optional TpuExecutor::GetAllocatorStats() { SE_AllocatorStats c_stats; - if (TpuExecutor_GetAllocatorStats(executor_, &c_stats)) { + if (tpu::ExecutorApiFn()->TpuExecutor_GetAllocatorStatsFn(executor_, + &c_stats)) { ::stream_executor::AllocatorStats stats; stats.num_allocs = c_stats.num_allocs; stats.bytes_in_use = c_stats.bytes_in_use; @@ -226,31 +234,33 @@ TpuExecutor::GetAllocatorStats() { Status TpuExecutor::WaitForInfeedReady(int32 infeed_queue_index) { StatusHelper status; - TpuExecutor_WaitForInfeedReady(executor_, infeed_queue_index, - status.c_status); + tpu::ExecutorApiFn()->TpuExecutor_WaitForInfeedReadyFn( + executor_, infeed_queue_index, status.c_status); return status.status(); } Status TpuExecutor::WaitForOutfeedReady(int32 outfeed_queue_index) { StatusHelper status; - TpuExecutor_WaitForOutfeedReady(executor_, outfeed_queue_index, - status.c_status); + tpu::ExecutorApiFn()->TpuExecutor_WaitForOutfeedReadyFn( + executor_, outfeed_queue_index, status.c_status); return status.status(); } void TpuExecutor::DequeueOutfeed(int32 outfeed_queue_index, absl::Span bytes, StatusCallback done) { StatusHelper status; - TpuExecutor_DequeueOutfeed(executor_, outfeed_queue_index, bytes.data(), - bytes.size(), status.c_status); + tpu::ExecutorApiFn()->TpuExecutor_DequeueOutfeedFn( + executor_, outfeed_queue_index, bytes.data(), bytes.size(), + status.c_status); done(status.status()); } Status TpuExecutor::EnqueueInfeed(int32 infeed_queue_index, absl::Span bytes) { StatusHelper status; - TpuExecutor_EnqueueInfeed(executor_, infeed_queue_index, bytes.data(), - bytes.size(), status.c_status); + tpu::ExecutorApiFn()->TpuExecutor_EnqueueInfeedFn( + executor_, infeed_queue_index, bytes.data(), bytes.size(), + status.c_status); return status.status(); } @@ -259,9 +269,9 @@ bool TpuExecutor::Memcpy(Stream* stream, void* host_dst, uint64 size) { SE_DeviceMemoryBase se_base = TpuConversions::DeviceMemoryBaseToSE_DeviceMemoryBase(device_src); - return TpuExecutor_MemcpyToHost(executor_, - stream_map().at(stream->implementation()), - host_dst, &se_base, size); + return tpu::ExecutorApiFn()->TpuExecutor_MemcpyToHostFn( + executor_, stream_map().at(stream->implementation()), host_dst, &se_base, + size); } bool TpuExecutor::Memcpy(Stream* stream, @@ -269,9 +279,9 @@ bool TpuExecutor::Memcpy(Stream* stream, const void* host_src, uint64 size) { SE_DeviceMemoryBase se_base = TpuConversions::DeviceMemoryBaseToSE_DeviceMemoryBase(*device_dst); - return TpuExecutor_MemcpyFromHost(executor_, - stream_map().at(stream->implementation()), - &se_base, host_src, size); + return tpu::ExecutorApiFn()->TpuExecutor_MemcpyFromHostFn( + executor_, stream_map().at(stream->implementation()), &se_base, host_src, + size); } Status TpuExecutor::SynchronousMemcpy( @@ -280,8 +290,8 @@ Status TpuExecutor::SynchronousMemcpy( StatusHelper status; SE_DeviceMemoryBase se_base = TpuConversions::DeviceMemoryBaseToSE_DeviceMemoryBase(*device_dst); - TpuExecutor_SynchronousMemcpyFromHost(executor_, &se_base, host_src, size, - status.c_status); + tpu::ExecutorApiFn()->TpuExecutor_SynchronousMemcpyFromHostFn( + executor_, &se_base, host_src, size, status.c_status); return status.status(); } @@ -291,8 +301,8 @@ Status TpuExecutor::SynchronousMemcpy( StatusHelper status; SE_DeviceMemoryBase se_base = TpuConversions::DeviceMemoryBaseToSE_DeviceMemoryBase(device_src); - TpuExecutor_SynchronousMemcpyToHost(executor_, host_dst, &se_base, size, - status.c_status); + tpu::ExecutorApiFn()->TpuExecutor_SynchronousMemcpyToHostFn( + executor_, host_dst, &se_base, size, status.c_status); return status.status(); } @@ -316,8 +326,8 @@ struct HostCallbackContext { SE_Status* HostCallbackTrampoline(void* ctx) { HostCallbackContext* host_ctx = reinterpret_cast(ctx); Status status = host_ctx->callback(); - SE_Status* c_status = - TpuStatus_Create(status.code(), status.error_message().c_str()); + SE_Status* c_status = tpu::ExecutorApiFn()->TpuStatus_CreateFn( + status.code(), status.error_message().c_str()); delete host_ctx; return c_status; } @@ -325,18 +335,21 @@ SE_Status* HostCallbackTrampoline(void* ctx) { bool TpuExecutor::HostCallback(Stream* stream, std::function callback) { HostCallbackContext* ctx = new HostCallbackContext{callback}; - return TpuExecutor_HostCallback(executor_, - stream_map().at(stream->implementation()), - &HostCallbackTrampoline, ctx); + return tpu::ExecutorApiFn()->TpuExecutor_HostCallbackFn( + executor_, stream_map().at(stream->implementation()), + &HostCallbackTrampoline, ctx); } TpuExecutor::StatusOr> TpuExecutor::CreateDeviceDescription() const { StatusHelper status; - SE_DeviceDescription* description = TpuDeviceDescription_New(); - auto cleanup = tensorflow::gtl::MakeCleanup( - [description]() { TpuDeviceDescription_Free(description); }); - TpuExecutor_CreateDeviceDescription(executor_, description, status.c_status); + SE_DeviceDescription* description = + tpu::ExecutorApiFn()->TpuDeviceDescription_NewFn(); + auto cleanup = tensorflow::gtl::MakeCleanup([description]() { + tpu::ExecutorApiFn()->TpuDeviceDescription_FreeFn(description); + }); + tpu::ExecutorApiFn()->TpuExecutor_CreateDeviceDescriptionFn( + executor_, description, status.c_status); if (status.status().ok()) { stream_executor::internal::DeviceDescriptionBuilder builder; CHECK_NE(description->device_vendor, nullptr); diff --git a/tensorflow/stream_executor/tpu/tpu_executor_c_api.h b/tensorflow/stream_executor/tpu/tpu_executor_c_api.h index e77e09bb911..eee69a35b23 100644 --- a/tensorflow/stream_executor/tpu/tpu_executor_c_api.h +++ b/tensorflow/stream_executor/tpu/tpu_executor_c_api.h @@ -20,9 +20,9 @@ limitations under the License. #include #include "tensorflow/c/tf_attrtype.h" -#include "tensorflow/c/tf_datatype.h" #include "tensorflow/c/tf_status.h" #include "tensorflow/core/tpu/kernels/tpu_ops_common_c_api.h" +#include "tensorflow/core/tpu/libtftpu.h" typedef struct SE_Platform SE_Platform; typedef struct SE_StreamExecutor SE_StreamExecutor; @@ -292,6 +292,96 @@ void TpuTransferManager_WriteSingleTupleIndexTable( XLA_ComputationPlacer* TpuComputationPlacer_New(); void TpuComputationPlacer_Free(XLA_ComputationPlacer* placer); + +struct TfTpu_ExecutorApiFn { + TFTPU_ADD_FN_IN_STRUCT(TpuPlatform_New); + TFTPU_ADD_FN_IN_STRUCT(TpuPlatform_Free); + TFTPU_ADD_FN_IN_STRUCT(TpuPlatform_Initialize); + TFTPU_ADD_FN_IN_STRUCT(TpuPlatform_Initialized); + TFTPU_ADD_FN_IN_STRUCT(TpuPlatform_GetExecutor); + TFTPU_ADD_FN_IN_STRUCT(TpuPlatform_Id); + TFTPU_ADD_FN_IN_STRUCT(TpuPlatform_VisibleDeviceCount); + TFTPU_ADD_FN_IN_STRUCT(TpuPlatform_TpuMemoryLimit); + TFTPU_ADD_FN_IN_STRUCT(TpuPlatform_ShouldRegisterTpuDeviceToDeviceCopy); + TFTPU_ADD_FN_IN_STRUCT(TpuExecutor_Init); + TFTPU_ADD_FN_IN_STRUCT(TpuExecutor_Free); + TFTPU_ADD_FN_IN_STRUCT(TpuExecutor_PlatformDeviceCount); + TFTPU_ADD_FN_IN_STRUCT(TpuExecutor_Allocate); + TFTPU_ADD_FN_IN_STRUCT(TpuExecutor_Deallocate); + TFTPU_ADD_FN_IN_STRUCT(TpuExecutor_GetAllocatorStats); + TFTPU_ADD_FN_IN_STRUCT(TpuExecutor_DeviceMemoryUsage); + TFTPU_ADD_FN_IN_STRUCT(TpuExecutor_AllocateStream); + TFTPU_ADD_FN_IN_STRUCT(TpuExecutor_DeallocateStream); + TFTPU_ADD_FN_IN_STRUCT(TpuExecutor_CreateStreamDependency); + TFTPU_ADD_FN_IN_STRUCT(TpuExecutor_GetStatus); + TFTPU_ADD_FN_IN_STRUCT(TpuExecutor_AllocateEvent); + TFTPU_ADD_FN_IN_STRUCT(TpuExecutor_DeallocateEvent); + TFTPU_ADD_FN_IN_STRUCT(TpuExecutor_PollForEventStatus); + TFTPU_ADD_FN_IN_STRUCT(TpuExecutor_RecordEvent); + TFTPU_ADD_FN_IN_STRUCT(TpuExecutor_WaitForEvent); + TFTPU_ADD_FN_IN_STRUCT(TpuExecutor_AllocateTimer); + TFTPU_ADD_FN_IN_STRUCT(TpuExecutor_DeallocateTimer); + TFTPU_ADD_FN_IN_STRUCT(TpuExecutor_StartTimer); + TFTPU_ADD_FN_IN_STRUCT(TpuExecutor_StopTimer); + TFTPU_ADD_FN_IN_STRUCT(TpuExecutor_SynchronousMemcpyToHost); + TFTPU_ADD_FN_IN_STRUCT(TpuExecutor_SynchronousMemcpyFromHost); + TFTPU_ADD_FN_IN_STRUCT(TpuExecutor_MemcpyToHost); + TFTPU_ADD_FN_IN_STRUCT(TpuExecutor_MemcpyFromHost); + TFTPU_ADD_FN_IN_STRUCT(TpuExecutor_EnqueueInfeed); + TFTPU_ADD_FN_IN_STRUCT(TpuExecutor_DequeueOutfeed); + TFTPU_ADD_FN_IN_STRUCT(TpuExecutor_WaitForInfeedReady); + TFTPU_ADD_FN_IN_STRUCT(TpuExecutor_WaitForOutfeedReady); + TFTPU_ADD_FN_IN_STRUCT(TpuExecutor_BlockHostUntilDone); + TFTPU_ADD_FN_IN_STRUCT(TpuExecutor_BlockUntilDoneOrFailed); + TFTPU_ADD_FN_IN_STRUCT(TpuExecutor_SyncAndForgetFailedStreams); + TFTPU_ADD_FN_IN_STRUCT(TpuExecutor_SynchronizeAllActivity); + + TFTPU_ADD_FN_IN_STRUCT(TpuStream_New); + TFTPU_ADD_FN_IN_STRUCT(TpuStream_Free); + TFTPU_ADD_FN_IN_STRUCT(TpuStream_Stream); + TFTPU_ADD_FN_IN_STRUCT(TpuStream_Status); + TFTPU_ADD_FN_IN_STRUCT(TpuStream_IsSameSharedMemoryLocation); + TFTPU_ADD_FN_IN_STRUCT(TpuStream_TpuEnqueueOnDeviceSendRecvLocal); + + TFTPU_ADD_FN_IN_STRUCT(TpuEvent_New); + TFTPU_ADD_FN_IN_STRUCT(TpuEvent_Free); + + TFTPU_ADD_FN_IN_STRUCT(TpuTimer_New); + TFTPU_ADD_FN_IN_STRUCT(TpuTimer_Free); + TFTPU_ADD_FN_IN_STRUCT(TpuTimer_Nanoseconds); + TFTPU_ADD_FN_IN_STRUCT(TpuTimer_Microseconds); + + TFTPU_ADD_FN_IN_STRUCT(TpuStatus_New); + TFTPU_ADD_FN_IN_STRUCT(TpuStatus_Create); + TFTPU_ADD_FN_IN_STRUCT(TpuStatus_Free); + TFTPU_ADD_FN_IN_STRUCT(TpuStatus_Message); + TFTPU_ADD_FN_IN_STRUCT(TpuStatus_Code); + TFTPU_ADD_FN_IN_STRUCT(TpuStatus_Ok); + + TFTPU_ADD_FN_IN_STRUCT(TpuStreamExecutorConfig_Default); + TFTPU_ADD_FN_IN_STRUCT(TpuStreamExecutorConfig_SetOrdinal); + TFTPU_ADD_FN_IN_STRUCT(TpuStreamExecutorConfig_Free); + + TFTPU_ADD_FN_IN_STRUCT(TpuDeviceDescription_New); + TFTPU_ADD_FN_IN_STRUCT(TpuDeviceDescription_Free); + + TFTPU_ADD_FN_IN_STRUCT(TpuExecutor_CreateDeviceDescription); + TFTPU_ADD_FN_IN_STRUCT(TpuExecutor_NewDeviceOptions); + TFTPU_ADD_FN_IN_STRUCT(TpuExecutor_FreeDeviceOptions); + TFTPU_ADD_FN_IN_STRUCT(TpuExecutor_HostCallback); + + TFTPU_ADD_FN_IN_STRUCT(TpuTransferManager_New); + TFTPU_ADD_FN_IN_STRUCT(TpuTransferManager_Free); + TFTPU_ADD_FN_IN_STRUCT(TpuTransferManager_PlatformId); + TFTPU_ADD_FN_IN_STRUCT(TpuTransferManager_HostShapeToDeviceShape); + TFTPU_ADD_FN_IN_STRUCT(TpuTransferManager_TransferLiteralToDeviceAsync); + TFTPU_ADD_FN_IN_STRUCT(TpuTransferManager_TransferLiteralFromDevice); + TFTPU_ADD_FN_IN_STRUCT(TpuTransferManager_GetByteSizeRequirement); + TFTPU_ADD_FN_IN_STRUCT(TpuTransferManager_WriteSingleTupleIndexTable); + + TFTPU_ADD_FN_IN_STRUCT(TpuComputationPlacer_New); + TFTPU_ADD_FN_IN_STRUCT(TpuComputationPlacer_Free); +}; } // extern "C" diff --git a/tensorflow/stream_executor/tpu/tpu_node_context.cc b/tensorflow/stream_executor/tpu/tpu_node_context.cc index 35a9eb53bcd..356ede40fb3 100644 --- a/tensorflow/stream_executor/tpu/tpu_node_context.cc +++ b/tensorflow/stream_executor/tpu/tpu_node_context.cc @@ -17,6 +17,7 @@ limitations under the License. #include "tensorflow/compiler/xla/service/backend.h" #include "tensorflow/compiler/xla/service/platform_util.h" #include "tensorflow/compiler/xla/service/transfer_manager.h" +#include "tensorflow/core/tpu/tpu_library_loader.h" #include "tensorflow/stream_executor/device_memory_allocator.h" #include "tensorflow/stream_executor/tpu/tpu_executor_c_api.h" #include "tensorflow/stream_executor/tpu/tpu_node_context_c_api.h" @@ -32,15 +33,18 @@ StatusOr> TpuNodeContext::Create( int device_ordinal) { StatusHelper status; XLA_TpuNodeContext* node_context = - TpuNodeContext_Create(device_ordinal, status.c_status); + tpu::NodeContextApiFn()->TpuNodeContext_CreateFn(device_ordinal, + status.c_status); if (!status.status().ok()) { - TpuNodeContext_Free(node_context); + tpu::NodeContextApiFn()->TpuNodeContext_FreeFn(node_context); return status.status(); } return std::make_unique(device_ordinal, node_context); } -TpuNodeContext::~TpuNodeContext() { TpuNodeContext_Free(node_context_); } +TpuNodeContext::~TpuNodeContext() { + tpu::NodeContextApiFn()->TpuNodeContext_FreeFn(node_context_); +} /* static */ Status TpuNodeContext::Initialize(int device_ordinal) { @@ -52,14 +56,14 @@ Status TpuNodeContext::Initialize(int device_ordinal) { /* static */ Status TpuNodeContext::StopChipHeartbeats() { StatusHelper status; - TpuNodeContext_StopChipHeartbeats(status.c_status); + tpu::NodeContextApiFn()->TpuNodeContext_StopChipHeartbeatsFn(status.c_status); return status.status(); } /* static */ Status TpuNodeContext::CloseTpuHost() { StatusHelper status; - TpuNodeContext_CloseTpuHost(status.c_status); + tpu::NodeContextApiFn()->TpuNodeContext_CloseTpuHostFn(status.c_status); return status.status(); } diff --git a/tensorflow/stream_executor/tpu/tpu_node_context_c_api.h b/tensorflow/stream_executor/tpu/tpu_node_context_c_api.h index e5092d4842b..d47fdf37a46 100644 --- a/tensorflow/stream_executor/tpu/tpu_node_context_c_api.h +++ b/tensorflow/stream_executor/tpu/tpu_node_context_c_api.h @@ -15,10 +15,13 @@ limitations under the License. #ifndef TENSORFLOW_STREAM_EXECUTOR_TPU_TPU_NODE_CONTEXT_C_API_H_ #define TENSORFLOW_STREAM_EXECUTOR_TPU_TPU_NODE_CONTEXT_C_API_H_ +#include "tensorflow/core/tpu/libtftpu.h" #include "tensorflow/stream_executor/tpu/tpu_executor_c_api.h" typedef struct XLA_TpuNodeContext XLA_TpuNodeContext; +extern "C" { + XLA_TpuNodeContext* TpuNodeContext_Create(int device_ordinal, SE_Status* status); void TpuNodeContext_Free(XLA_TpuNodeContext* node_context); @@ -28,4 +31,13 @@ void TpuNodeContext_Initialize(int device_ordinal, SE_Status* status); void TpuNodeContext_StopChipHeartbeats(SE_Status* status); void TpuNodeContext_CloseTpuHost(SE_Status* status); +} // extern "C" + +struct TfTpu_NodeContextApiFn { + TFTPU_ADD_FN_IN_STRUCT(TpuNodeContext_Create); + TFTPU_ADD_FN_IN_STRUCT(TpuNodeContext_Free); + TFTPU_ADD_FN_IN_STRUCT(TpuNodeContext_StopChipHeartbeats); + TFTPU_ADD_FN_IN_STRUCT(TpuNodeContext_CloseTpuHost); +}; + #endif // TENSORFLOW_STREAM_EXECUTOR_TPU_TPU_NODE_CONTEXT_C_API_H_ diff --git a/tensorflow/stream_executor/tpu/tpu_platform.cc b/tensorflow/stream_executor/tpu/tpu_platform.cc index c65d8a4207a..13a845829c1 100644 --- a/tensorflow/stream_executor/tpu/tpu_platform.cc +++ b/tensorflow/stream_executor/tpu/tpu_platform.cc @@ -16,6 +16,7 @@ limitations under the License. #include "tensorflow/stream_executor/tpu/tpu_platform.h" #include "tensorflow/c/tf_status.h" +#include "tensorflow/core/tpu/tpu_library_loader.h" #include "tensorflow/stream_executor/platform.h" #include "tensorflow/stream_executor/tpu/status_helper.h" #include "tensorflow/stream_executor/tpu/tpu_executor.h" @@ -30,7 +31,9 @@ using Status = ::stream_executor::port::Status; template using StatusOr = ::stream_executor::port::StatusOr; -TpuPlatform::TpuPlatform() { platform_ = TpuPlatform_New(); } +TpuPlatform::TpuPlatform() { + platform_ = tpu::ExecutorApiFn()->TpuPlatform_NewFn(); +} TpuPlatform* TpuPlatform::GetRegisteredPlatform() { return tpu_registered_platform; @@ -53,8 +56,8 @@ Status TpuPlatform::Initialize( i++; } - TpuPlatform_Initialize(platform_, options_size, options_key, options_value, - status.c_status); + tpu::ExecutorApiFn()->TpuPlatform_InitializeFn( + platform_, options_size, options_key, options_value, status.c_status); free(options_key); free(options_value); @@ -62,10 +65,16 @@ Status TpuPlatform::Initialize( return status.status(); } -TpuPlatform::~TpuPlatform() { TpuPlatform_Free(platform_); } +bool TpuPlatform::Initialized() const { + return tpu::ExecutorApiFn()->TpuPlatform_InitializedFn(platform_); +} + +TpuPlatform::~TpuPlatform() { + tpu::ExecutorApiFn()->TpuPlatform_FreeFn(platform_); +} int TpuPlatform::VisibleDeviceCount() const { - return TpuPlatform_VisibleDeviceCount(platform_); + return tpu::ExecutorApiFn()->TpuPlatform_VisibleDeviceCountFn(platform_); } StatusOr<::stream_executor::StreamExecutor*> TpuPlatform::GetExecutor( @@ -77,14 +86,16 @@ StatusOr<::stream_executor::StreamExecutor*> TpuPlatform::GetExecutor( StatusOr> TpuPlatform::GetUncachedExecutor( const ::stream_executor::StreamExecutorConfig& config) { - SE_StreamExecutorConfig* c_config = TpuStreamExecutorConfig_Default(); + SE_StreamExecutorConfig* c_config = + tpu::ExecutorApiFn()->TpuStreamExecutorConfig_DefaultFn(); - TpuStreamExecutorConfig_SetOrdinal(c_config, config.ordinal); + tpu::ExecutorApiFn()->TpuStreamExecutorConfig_SetOrdinalFn(c_config, + config.ordinal); StatusHelper status; - SE_StreamExecutor* executor = - TpuPlatform_GetExecutor(platform_, c_config, status.c_status); - TpuStreamExecutorConfig_Free(c_config); + SE_StreamExecutor* executor = tpu::ExecutorApiFn()->TpuPlatform_GetExecutorFn( + platform_, c_config, status.c_status); + tpu::ExecutorApiFn()->TpuStreamExecutorConfig_FreeFn(c_config); if (!status.ok()) { return status.status(); } @@ -103,27 +114,24 @@ const std::string& TpuPlatform::Name() const { } int64 TpuPlatform::TpuMemoryLimit() { - return TpuPlatform_TpuMemoryLimit(platform_); + return tpu::ExecutorApiFn()->TpuPlatform_TpuMemoryLimitFn(platform_); } bool TpuPlatform::ShouldRegisterTpuDeviceToDeviceCopy() { - return TpuPlatform_ShouldRegisterTpuDeviceToDeviceCopy(platform_); + return tpu::ExecutorApiFn() + ->TpuPlatform_ShouldRegisterTpuDeviceToDeviceCopyFn(platform_); +} + +void RegisterTpuPlatform() { + static bool tpu_platform_registered = false; + if (!tpu_platform_registered) { + tensorflow::tpu_registered_platform = new tensorflow::TpuPlatform(); + std::unique_ptr platform( + tensorflow::tpu_registered_platform); + SE_CHECK_OK(stream_executor::MultiPlatformManager::RegisterPlatform( + std::move(platform))); + tpu_platform_registered = true; + } } } // namespace tensorflow - -void RegisterTpuPlatform() { - tensorflow::tpu_registered_platform = new tensorflow::TpuPlatform(); - std::unique_ptr platform( - tensorflow::tpu_registered_platform); - SE_CHECK_OK(stream_executor::MultiPlatformManager::RegisterPlatform( - std::move(platform))); -} - -REGISTER_MODULE_INITIALIZER(tpu_platform, RegisterTpuPlatform()); - -// Note that module initialization sequencing is not supported in the -// open-source project, so this will be a no-op there. -REGISTER_MODULE_INITIALIZER_SEQUENCE(tpu_platform, multi_platform_manager); -REGISTER_MODULE_INITIALIZER_SEQUENCE(multi_platform_manager_listener, - tpu_platform); diff --git a/tensorflow/stream_executor/tpu/tpu_platform.h b/tensorflow/stream_executor/tpu/tpu_platform.h index 6fdd8d15aa4..c2673ab9288 100644 --- a/tensorflow/stream_executor/tpu/tpu_platform.h +++ b/tensorflow/stream_executor/tpu/tpu_platform.h @@ -60,9 +60,7 @@ class TpuPlatform : public ::tensorflow::tpu::TpuPlatformInterface { bool ShouldRegisterTpuDeviceToDeviceCopy() override; - bool Initialized() const override { - return TpuPlatform_Initialized(platform_); - } + bool Initialized() const override; Status Initialize( const std::map& platform_options) override; @@ -124,6 +122,8 @@ class TpuPlatform : public ::tensorflow::tpu::TpuPlatformInterface { EventMap event_map_; }; +void RegisterTpuPlatform(); + } // namespace tensorflow #endif // TENSORFLOW_STREAM_EXECUTOR_TPU_TPU_PLATFORM_H_ diff --git a/tensorflow/stream_executor/tpu/tpu_stream.h b/tensorflow/stream_executor/tpu/tpu_stream.h index 5c71c0535f3..e1aa1164248 100644 --- a/tensorflow/stream_executor/tpu/tpu_stream.h +++ b/tensorflow/stream_executor/tpu/tpu_stream.h @@ -16,6 +16,7 @@ limitations under the License. #ifndef TENSORFLOW_STREAM_EXECUTOR_TPU_TPU_STREAM_H_ #define TENSORFLOW_STREAM_EXECUTOR_TPU_TPU_STREAM_H_ +#include "tensorflow/core/tpu/tpu_library_loader.h" #include "tensorflow/stream_executor/stream_executor_internal.h" #include "tensorflow/stream_executor/tpu/c_api_conversions.h" #include "tensorflow/stream_executor/tpu/status_helper.h" @@ -27,23 +28,27 @@ class TpuStream : public tensorflow::tpu::TpuStreamInterface { using Status = stream_executor::port::Status; explicit TpuStream(SE_Stream* stream) : stream_(stream) {} - ~TpuStream() override { TpuStream_Free(stream_); } + ~TpuStream() override { + tensorflow::tpu::ExecutorApiFn()->TpuStream_FreeFn(stream_); + } bool IsSameSharedMemoryLocation( tensorflow::tpu::TpuStreamInterface* other) override { - return TpuStream_IsSameSharedMemoryLocation( - stream_, static_cast(other)->stream_); + return tensorflow::tpu::ExecutorApiFn() + ->TpuStream_IsSameSharedMemoryLocationFn( + stream_, static_cast(other)->stream_); } Status EnqueueOnTpuDeviceSendRecvLocal( stream_executor::DeviceMemoryBase send_buffer, stream_executor::DeviceMemoryBase recv_buffer) override { StatusHelper status; - TpuStream_TpuEnqueueOnDeviceSendRecvLocal( - stream_, - TpuConversions::DeviceMemoryBaseToSE_DeviceMemoryBase(send_buffer), - TpuConversions::DeviceMemoryBaseToSE_DeviceMemoryBase(recv_buffer), - status.c_status); + tensorflow::tpu::ExecutorApiFn() + ->TpuStream_TpuEnqueueOnDeviceSendRecvLocalFn( + stream_, + TpuConversions::DeviceMemoryBaseToSE_DeviceMemoryBase(send_buffer), + TpuConversions::DeviceMemoryBaseToSE_DeviceMemoryBase(recv_buffer), + status.c_status); return status.status(); } @@ -54,7 +59,9 @@ class TpuStream : public tensorflow::tpu::TpuStreamInterface { class TpuEvent : public ::stream_executor::internal::EventInterface { public: explicit TpuEvent(SE_Event* event) : event_(event) {} - ~TpuEvent() override { TpuEvent_Free(event_); } + ~TpuEvent() override { + tensorflow::tpu::ExecutorApiFn()->TpuEvent_FreeFn(event_); + } private: SE_Event* event_; diff --git a/tensorflow/stream_executor/tpu/tpu_timer.h b/tensorflow/stream_executor/tpu/tpu_timer.h index 246a0b7eb32..d7f8f660b37 100644 --- a/tensorflow/stream_executor/tpu/tpu_timer.h +++ b/tensorflow/stream_executor/tpu/tpu_timer.h @@ -17,6 +17,7 @@ limitations under the License. #define TENSORFLOW_STREAM_EXECUTOR_TPU_TPU_TIMER_H_ #include "tensorflow/core/platform/types.h" +#include "tensorflow/core/tpu/tpu_library_loader.h" #include "tensorflow/stream_executor/stream_executor_internal.h" #include "tensorflow/stream_executor/tpu/tpu_executor_c_api.h" @@ -25,9 +26,15 @@ namespace tensorflow { class TpuTimer : public ::stream_executor::internal::TimerInterface { public: explicit TpuTimer(SE_Timer* timer) : timer_(timer) {} - ~TpuTimer() override { TpuTimer_Free(timer_); } - uint64 Microseconds() const override { return TpuTimer_Microseconds(timer_); } - uint64 Nanoseconds() const override { return TpuTimer_Nanoseconds(timer_); } + ~TpuTimer() override { + tensorflow::tpu::ExecutorApiFn()->TpuTimer_FreeFn(timer_); + } + uint64 Microseconds() const override { + return tensorflow::tpu::ExecutorApiFn()->TpuTimer_MicrosecondsFn(timer_); + } + uint64 Nanoseconds() const override { + return tensorflow::tpu::ExecutorApiFn()->TpuTimer_NanosecondsFn(timer_); + } private: SE_Timer* timer_; diff --git a/tensorflow/stream_executor/tpu/tpu_transfer_manager.cc b/tensorflow/stream_executor/tpu/tpu_transfer_manager.cc index 4bedc251413..934fabbf54d 100644 --- a/tensorflow/stream_executor/tpu/tpu_transfer_manager.cc +++ b/tensorflow/stream_executor/tpu/tpu_transfer_manager.cc @@ -17,6 +17,7 @@ limitations under the License. #include "tensorflow/compiler/xla/shape_util.h" #include "tensorflow/compiler/xla/xla_data.pb.h" +#include "tensorflow/core/tpu/tpu_library_loader.h" #include "tensorflow/stream_executor/device_memory.h" #include "tensorflow/stream_executor/tpu/c_api_conversions.h" #include "tensorflow/stream_executor/tpu/proto_helper.h" @@ -29,10 +30,12 @@ namespace tensorflow { using Status = stream_executor::port::Status; TpuTransferManager::TpuTransferManager() { - manager_ = TpuTransferManager_New(); + manager_ = tpu::ExecutorApiFn()->TpuTransferManager_NewFn(); } -TpuTransferManager::~TpuTransferManager() { TpuTransferManager_Free(manager_); } +TpuTransferManager::~TpuTransferManager() { + tpu::ExecutorApiFn()->TpuTransferManager_FreeFn(manager_); +} stream_executor::Platform::Id TpuTransferManager::PlatformId() const { return TpuPlatform::kId; @@ -45,8 +48,8 @@ xla::Shape TpuTransferManager::HostShapeToDeviceShape( TpuConversions::XlaShapeToCShape(host_shape, &c_host_shape); - TpuTransferManager_HostShapeToDeviceShape(manager_, &c_host_shape, - &c_device_shape); + tpu::ExecutorApiFn()->TpuTransferManager_HostShapeToDeviceShapeFn( + manager_, &c_host_shape, &c_device_shape); xla::Shape device_shape = TpuConversions::CShapeToXlaShape(&c_device_shape); TpuConversions::CShapeCleanup(&c_host_shape); TpuConversions::CShapeCleanup(&c_device_shape); @@ -66,7 +69,7 @@ Status TpuTransferManager::TransferLiteralToDeviceAsync( TpuConversions::XLAShapedBufferToCShapedBuffer(device_buffer, &c_device_buffer); - TpuTransferManager_TransferLiteralToDeviceAsync( + tpu::ExecutorApiFn()->TpuTransferManager_TransferLiteralToDeviceAsyncFn( manager_, TpuPlatform::GetRegisteredPlatform()->stream_map()->at( stream->implementation()), @@ -112,7 +115,7 @@ void TpuTransferManager::TransferLiteralFromDevice( XLA_Literal c_literal; TpuConversions::XLALiteralToCLiteral(literal, &c_literal); - TpuTransferManager_TransferLiteralFromDevice( + tpu::ExecutorApiFn()->TpuTransferManager_TransferLiteralFromDeviceFn( manager_, TpuPlatform::GetRegisteredPlatform()->stream_map()->at( stream->implementation()), @@ -127,7 +130,8 @@ int64 TpuTransferManager::GetByteSizeRequirement( TpuConversions::XlaShapeToCShape(shape, &c_shape); int64 size_in_bytes = - TpuTransferManager_GetByteSizeRequirement(manager_, &c_shape); + tpu::ExecutorApiFn()->TpuTransferManager_GetByteSizeRequirementFn( + manager_, &c_shape); TpuConversions::CShapeCleanup(&c_shape); return size_in_bytes; @@ -151,7 +155,7 @@ Status TpuTransferManager::WriteSingleTupleIndexTable( region->payload()}; StatusHelper status; - TpuTransferManager_WriteSingleTupleIndexTable( + tpu::ExecutorApiFn()->TpuTransferManager_WriteSingleTupleIndexTableFn( manager_, TpuPlatform::GetRegisteredPlatform()->stream_map()->at( stream->implementation()), diff --git a/tensorflow/tensorflow.bzl b/tensorflow/tensorflow.bzl index 5da15b0a4d6..4a4f8837867 100644 --- a/tensorflow/tensorflow.bzl +++ b/tensorflow/tensorflow.bzl @@ -2899,6 +2899,13 @@ def if_mlir(if_true, if_false = []): "//conditions:default": if_false, }) +def if_tpu(if_true, if_false = []): + """Shorthand for select()ing whether to build for TPUs.""" + return select({ + str(Label("//tensorflow:with_tpu_support")): if_true, + "//conditions:default": if_false, + }) + def tfcompile_target_cpu(): return "" From 2ff1c5a31be8a3d38e3bccd59e54cbfd5cabe5cd Mon Sep 17 00:00:00 2001 From: Kuangyuan Chen Date: Thu, 18 Jun 2020 10:59:05 -0700 Subject: [PATCH 194/867] Import initialization graph in SignatureDef SavedModels as an MLIR function in TF saved model dialect. PiperOrigin-RevId: 317137903 Change-Id: I7cbded06b3deafa30d3b3e3dad98cc8f056dd4e3 --- tensorflow/compiler/mlir/tensorflow/BUILD | 4 +- .../mlir/tensorflow/ir/tf_saved_model.cc | 25 +++ .../mlir/tensorflow/ir/tf_saved_model_ops.td | 24 +++ .../tests/tf_saved_model/common_v1.py | 1 + .../tests/tf_saved_model/hash_table_v1.py | 92 +++++++++++ .../tensorflow/tests/tf_saved_model_ops.mlir | 5 + .../tests/tf_saved_model_ops_invalid.mlir | 33 ++++ .../mlir/tensorflow/translate/import_model.cc | 149 +++++++++++++----- 8 files changed, 290 insertions(+), 43 deletions(-) create mode 100644 tensorflow/compiler/mlir/tensorflow/tests/tf_saved_model/hash_table_v1.py diff --git a/tensorflow/compiler/mlir/tensorflow/BUILD b/tensorflow/compiler/mlir/tensorflow/BUILD index 904ccb7e820..17ed0e36a28 100644 --- a/tensorflow/compiler/mlir/tensorflow/BUILD +++ b/tensorflow/compiler/mlir/tensorflow/BUILD @@ -661,7 +661,9 @@ cc_library( ":tensorflow_types", ":translate_utils", "//tensorflow/cc/saved_model:bundle_v2", + "//tensorflow/cc/saved_model:constants", "//tensorflow/cc/saved_model:loader_lite", + "//tensorflow/cc/saved_model:loader_util", "//tensorflow/compiler/jit:shape_inference_helpers", "//tensorflow/compiler/mlir:op_or_arg_name_mapper", "//tensorflow/compiler/tf2xla:functionalize_control_flow", @@ -673,6 +675,7 @@ cc_library( "//tensorflow/core:lib", "//tensorflow/core:protos_all_cc", "//tensorflow/core/grappler/utils:transitive_fanin", + "//tensorflow/core/platform:protobuf_internal", "//tensorflow/core/platform:types", "//tensorflow/stream_executor/lib", "@com_google_absl//absl/algorithm:container", @@ -682,7 +685,6 @@ cc_library( "@com_google_absl//absl/strings", "@com_google_absl//absl/types:optional", "@llvm-project//llvm:Support", - "@llvm-project//mlir:Analysis", "@llvm-project//mlir:IR", "@llvm-project//mlir:Pass", "@llvm-project//mlir:StandardOps", diff --git a/tensorflow/compiler/mlir/tensorflow/ir/tf_saved_model.cc b/tensorflow/compiler/mlir/tensorflow/ir/tf_saved_model.cc index 140a778770c..6af70158e14 100644 --- a/tensorflow/compiler/mlir/tensorflow/ir/tf_saved_model.cc +++ b/tensorflow/compiler/mlir/tensorflow/ir/tf_saved_model.cc @@ -76,6 +76,23 @@ static LogicalResult Verify(GlobalTensorOp global_tensor) { return success(); } +static LogicalResult Verify(SessionInitializerOp session_initializer) { + mlir::SymbolTable symbol_table( + session_initializer.getParentOfType()); + + auto init_func_op = + symbol_table.lookup(session_initializer.initializer()); + if (!init_func_op) + return session_initializer.emitOpError() + << "the initializer function does not exist"; + + if (!init_func_op.getType().getResults().empty()) + return session_initializer.emitOpError() + << "the initializer function should have no output"; + + return success(); +} + #define GET_OP_CLASSES #include "tensorflow/compiler/mlir/tensorflow/ir/tf_saved_model.cc.inc" @@ -220,6 +237,14 @@ static LogicalResult VerifySavedModelModule( } } } + + auto session_initializers = module.getOps(); + if (std::distance(session_initializers.begin(), session_initializers.end()) > + 1) { + return (*++session_initializers.begin()).emitError() + << "there must be no more than one session_initializer op"; + } + SymbolTable symbol_table(module); auto symbol_uses = SymbolTable::getSymbolUses(&module.getBodyRegion()); if (!symbol_uses.hasValue()) { diff --git a/tensorflow/compiler/mlir/tensorflow/ir/tf_saved_model_ops.td b/tensorflow/compiler/mlir/tensorflow/ir/tf_saved_model_ops.td index 4431a160edf..497f4d90cb9 100644 --- a/tensorflow/compiler/mlir/tensorflow/ir/tf_saved_model_ops.td +++ b/tensorflow/compiler/mlir/tensorflow/ir/tf_saved_model_ops.td @@ -128,4 +128,28 @@ def TfSavedModel_GlobalTensorOp : TfSavedModel_Op<"global_tensor"> { let verifier = [{ return Verify(*this); }]; } +def TfSavedModel_SessionInitializerOp: TfSavedModel_Op<"session_initializer"> { + let summary = "Initializes TensorFlow session state."; + let description = [{ + Represents a session initializer function initializes TensorFlow session + state. It is used to initialize resources in the saved model before calling + any exported functions. There must be no more than one session initializer + in a saved model. + + The `initializer` represents the initialization function. The function have + no output and this function should be only called once. + + This is used, for example, to initialize hash tables stored in resources and + accessed by resource name (rather than as resource handles or bound inputs + which is how `global_tensor`s are referenced) + }]; + + let arguments = (ins + FlatSymbolRefAttr:$initializer + ); + + + let verifier = [{ return Verify(*this); }]; +} + #endif // SAVED_MODEL_DIALECT diff --git a/tensorflow/compiler/mlir/tensorflow/tests/tf_saved_model/common_v1.py b/tensorflow/compiler/mlir/tensorflow/tests/tf_saved_model/common_v1.py index 7171f63bb05..51ccbeb1fbd 100644 --- a/tensorflow/compiler/mlir/tensorflow/tests/tf_saved_model/common_v1.py +++ b/tensorflow/compiler/mlir/tensorflow/tests/tf_saved_model/common_v1.py @@ -84,6 +84,7 @@ def do_test(signature_def_map, show_debug_info=False): builder.add_meta_graph_and_variables( sess, [tf.saved_model.tag_constants.SERVING], signature_def_map, + main_op=tf.tables_initializer(), strip_default_attrs=True) builder.save() diff --git a/tensorflow/compiler/mlir/tensorflow/tests/tf_saved_model/hash_table_v1.py b/tensorflow/compiler/mlir/tensorflow/tests/tf_saved_model/hash_table_v1.py new file mode 100644 index 00000000000..64847434b82 --- /dev/null +++ b/tensorflow/compiler/mlir/tensorflow/tests/tf_saved_model/hash_table_v1.py @@ -0,0 +1,92 @@ +# Copyright 2019 The TensorFlow Authors. All Rights Reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. +# ============================================================================== + +# RUN: %p/hash_table_v1 | FileCheck %s + +# pylint: disable=missing-docstring,line-too-long +from __future__ import absolute_import +from __future__ import division +from __future__ import print_function + +import tensorflow.compat.v1 as tf +from tensorflow.compiler.mlir.tensorflow.tests.tf_saved_model import common_v1 + +# Verify that the tf.versions attribute exists. It is difficult to enforce +# contents, since the version numbers change over time. The conversion logic +# itself is verified in the common graphdef converter, so here just assert +# it is being invoked. +# CHECK: module +# CHECK-SAME: tf.versions +# CHECK-SAME: bad_consumers +# CHECK-SAME: min_consumer +# CHECK-SAME: producer + +# CHECK: "tf_saved_model.session_initializer"() {initializer = [[init:@.*]]} : () -> () +# CHECK: "tf_saved_model.global_tensor"() + +# CHECK: func {{@[a-zA-Z_0-9]+}}( +# CHECK-SAME: [[ARG0:%.*]]: tensor +# CHECK-SAME: [[ARG1:%.*]]: tensor () + // Representation for constants: (immutable) global tensor. // CHECK: tf_saved_model.global_tensor "tf_saved_model.global_tensor"() { diff --git a/tensorflow/compiler/mlir/tensorflow/tests/tf_saved_model_ops_invalid.mlir b/tensorflow/compiler/mlir/tensorflow/tests/tf_saved_model_ops_invalid.mlir index c055c6c9f56..544600cf6b8 100644 --- a/tensorflow/compiler/mlir/tensorflow/tests/tf_saved_model_ops_invalid.mlir +++ b/tensorflow/compiler/mlir/tensorflow/tests/tf_saved_model_ops_invalid.mlir @@ -258,3 +258,36 @@ module attributes {tf_saved_model.semantics} { // expected-error@+1 {{'type' attribute for immutable 'tf_saved_model.global_tensor' should have a static shape}} "tf_saved_model.global_tensor"() { sym_name = "v", type = tensor, value = dense<1.> : tensor<1xf32> } : () -> () } + +// ----- + +module attributes {tf_saved_model.semantics} { + + // expected-error@+1 {{the initializer function does not exist}} + "tf_saved_model.session_initializer"() { initializer = @init } : () -> () +} + +// ----- + +module attributes {tf_saved_model.semantics} { + + // expected-error@+1 {{the initializer function should have no output}} + "tf_saved_model.session_initializer"() { initializer = @init } : () -> () + func @init() -> tensor<1xf32> { + %0 = "tf.Const"() {value = dense<[1.0]> : tensor<1xf32> } : () -> tensor<1xf32> + return %0 : tensor<1xf32> + } +} + +// ----- + +module attributes {tf_saved_model.semantics} { + + "tf_saved_model.session_initializer"() { initializer = @init } : () -> () + // expected-error@+1 {{there must be no more than one session_initializer op}} + "tf_saved_model.session_initializer"() { initializer = @init } : () -> () + func @init() -> tensor<1xf32> { + %0 = "tf.Const"() {value = dense<[1.0]> : tensor<1xf32> } : () -> tensor<1xf32> + return %0 : tensor<1xf32> + } +} diff --git a/tensorflow/compiler/mlir/tensorflow/translate/import_model.cc b/tensorflow/compiler/mlir/tensorflow/translate/import_model.cc index 820d0ce31fb..3cff4217215 100644 --- a/tensorflow/compiler/mlir/tensorflow/translate/import_model.cc +++ b/tensorflow/compiler/mlir/tensorflow/translate/import_model.cc @@ -60,6 +60,8 @@ limitations under the License. #include "mlir/IR/Types.h" // from @llvm-project #include "mlir/IR/Verifier.h" // from @llvm-project #include "mlir/Pass/PassManager.h" // from @llvm-project +#include "tensorflow/cc/saved_model/constants.h" +#include "tensorflow/cc/saved_model/loader_util.h" #include "tensorflow/compiler/jit/shape_inference_helpers.h" #include "tensorflow/compiler/mlir/op_or_arg_name_mapper.h" #include "tensorflow/compiler/mlir/tensorflow/ir/tf_attributes.h" @@ -99,6 +101,7 @@ limitations under the License. #include "tensorflow/core/lib/core/errors.h" #include "tensorflow/core/lib/strings/str_util.h" #include "tensorflow/core/platform/protobuf.h" +#include "tensorflow/core/platform/protobuf_internal.h" #include "tensorflow/core/platform/types.h" #include "tensorflow/core/protobuf/graph_debug_info.pb.h" #include "tensorflow/core/protobuf/meta_graph.pb.h" @@ -116,6 +119,7 @@ using mlir::NamedAttrList; using mlir::TensorType; using mlir::TF::VarHandleOp; using mlir::tf_saved_model::GlobalTensorOp; +using mlir::tf_saved_model::SessionInitializerOp; using stream_executor::port::StatusOr; namespace { @@ -2955,6 +2959,13 @@ void SortSavedModelModule(mlir::ModuleOp module) { named_global_tensor.global_tensor.getOperation()->moveBefore( &module.getBody()->front()); } + + auto initializers = module.getOps(); + if (!initializers.empty()) { + (*initializers.begin()) + .getOperation() + ->moveBefore(&module.getBody()->front()); + } } Status CreateSavedModelIR( @@ -3241,17 +3252,29 @@ class SavedModelSignatureDefImporter { absl::Span exported_names, mlir::MLIRContext* context) : bundle_(bundle), + flib_def_(OpRegistry::Global(), graph_def().library()), + debug_info_(), exported_names_(exported_names), - module_(mlir::ModuleOp::create(mlir::UnknownLoc::get(context))) {} + module_(mlir::ModuleOp::create(mlir::UnknownLoc::get(context))) { + // debug_info might not be loaded with loader_lite. + if (bundle_.debug_info != nullptr) debug_info_ = *bundle_.debug_info; + } // Converts the SavedModel to the SavedModel dialect. Creates an MLIR function // for each signature. StatusOr ConvertSignatures(); - Status ConvertSignature(const GraphDef& graphdef, - const std::string& sig_def_key, - const SignatureDef& signature_def, - const GraphDebugInfo& debug_info, - const FunctionLibraryDefinition& flib_def); + Status ConvertSignature(const std::string& sig_def_key, + const SignatureDef& signature_def); + + // Converts the initialization graph in the SavedModel to an MLIR function. + Status ConvertInitializer(); + + // Converts a graph with feeds and fetches to an MLIR function. + StatusOr ConvertGraph( + const std::string& name, + const std::vector>& inputs, + const std::vector>& outputs, + const std::vector control_outputs); // Creates GlobalTensorOp for each variable and moves each VarHandle op to // the enclosing function's arguments. @@ -3273,18 +3296,62 @@ class SavedModelSignatureDefImporter { GraphImportConfig::InputArrays ParseInputArrays( const std::vector>& inputs); + const GraphDef& graph_def() const { + return bundle_.meta_graph_def.graph_def(); + } + const FunctionLibraryDefinition& flib_def() const { return flib_def_; } + const GraphDebugInfo& debug_info() const { return debug_info_; } + const SavedModelBundle& bundle_; + FunctionLibraryDefinition flib_def_; + GraphDebugInfo debug_info_; absl::Span exported_names_; mlir::OwningModuleRef module_; }; +Status SavedModelSignatureDefImporter::ConvertInitializer() { + std::vector asset_file_defs; + TF_RETURN_IF_ERROR( + internal::GetAssetFileDefs(bundle_.meta_graph_def, &asset_file_defs)); + + if (!asset_file_defs.empty()) + return errors::Unimplemented( + absl::StrCat("Assets are not supported in signaturedef importer")); + + std::string init_node_name; + TF_RETURN_IF_ERROR( + internal::GetInitOp("", bundle_.meta_graph_def, &init_node_name)); + + if (init_node_name.empty()) return Status::OK(); + + TF_ASSIGN_OR_RETURN(auto sub_module, + ConvertGraph(init_node_name, {}, {}, {init_node_name})); + + mlir::SymbolTable symbol_table(*sub_module); + + auto init_func_op = symbol_table.lookup(init_node_name); + + init_func_op.removeAttr("tf.entry_function"); + + mlir::OpBuilder builder(module_->getBodyRegion()); + + builder.create( + module_->getLoc(), builder.getSymbolRefAttr(init_func_op.getName())); + + // Move the converted functions to top level MLIR module. + auto* block = module_->getBody(); + auto* sub_block = sub_module->getBody(); + block->getOperations().splice( + mlir::Block::iterator(block->getTerminator()), sub_block->getOperations(), + sub_block->begin(), mlir::Block::iterator(sub_block->getTerminator())); + + return Status::OK(); +} + StatusOr SavedModelSignatureDefImporter::ConvertSignatures() { const auto& signatures = bundle_.GetSignatures(); - const auto& graphdef = bundle_.meta_graph_def.graph_def(); - PopulateTfVersions(module_.get(), graphdef.versions()); - - FunctionLibraryDefinition flib_def(OpRegistry::Global(), graphdef.library()); + PopulateTfVersions(module_.get(), graph_def().versions()); // debug_info might not be loaded with loader_lite. GraphDebugInfo debug_info; @@ -3307,9 +3374,10 @@ SavedModelSignatureDefImporter::ConvertSignatures() { continue; } - TF_RETURN_IF_ERROR(ConvertSignature(graphdef, sig_def_key, signature_def, - debug_info, flib_def)); + TF_RETURN_IF_ERROR(ConvertSignature(sig_def_key, signature_def)); } + + TF_RETURN_IF_ERROR(ConvertInitializer()); TF_RETURN_IF_ERROR(LiftVariables()); mlir::OpBuilder builder(module_->getBodyRegion()); @@ -3320,10 +3388,32 @@ SavedModelSignatureDefImporter::ConvertSignatures() { return std::move(module_); } +StatusOr SavedModelSignatureDefImporter::ConvertGraph( + const std::string& name, + const std::vector>& inputs, + const std::vector>& outputs, + const std::vector control_outputs) { + GraphImportConfig specs; + specs.prune_unused_nodes = true; + specs.inputs = ParseInputArrays(inputs); + for (auto& output : outputs) specs.outputs.push_back(output.second.name()); + specs.control_outputs = control_outputs; + + // Convert sub-graphdef to sub-graph. + GraphConstructorOptions options; + options.allow_internal_ops = true; + options.add_default_attributes = true; + Graph graph(OpRegistry::Global()); + + TF_RETURN_IF_ERROR(ConvertGraphDefToGraph(options, graph_def(), &graph)); + + // Convert sub-graph to MLIR module.true + return GraphDefImporter::Convert(module_->getContext(), graph, debug_info(), + flib_def(), specs, name); +} + Status SavedModelSignatureDefImporter::ConvertSignature( - const GraphDef& graphdef, const std::string& sig_def_key, - const SignatureDef& signature_def, const GraphDebugInfo& debug_info, - const FunctionLibraryDefinition& flib_def) { + const std::string& sig_def_key, const SignatureDef& signature_def) { // Create local vectors for the input and output and sort them to be // deterministic. We don't want anyone to really depend on the order, client // should lookup argument/result mapping by attribute name. @@ -3339,34 +3429,9 @@ Status SavedModelSignatureDefImporter::ConvertSignature( return lhs.first.size() < rhs.first.size() || lhs.first > rhs.first; }); - GraphImportConfig specs; - specs.prune_unused_nodes = true; - specs.inputs = ParseInputArrays(inputs); - for (auto& output : outputs) specs.outputs.push_back(output.second.name()); - - // Remove unused nodes and create sub-graphdef. - GraphDef sub_graph_def; - TF_RETURN_IF_ERROR(tensorflow::grappler::SetTransitiveFaninGraph( - graphdef, &sub_graph_def, - /*terminal_nodes=*/{specs.outputs.begin(), specs.outputs.end()})); - - // Set the function library definitions in the pruned graphdef. - *sub_graph_def.mutable_library() = flib_def.ToProto(); - - // Convert sub-graphdef to sub-graph. - GraphConstructorOptions options; - options.allow_internal_ops = true; - options.add_default_attributes = true; - Graph sub_graph(OpRegistry::Global()); - - TF_RETURN_IF_ERROR( - ConvertGraphDefToGraph(options, sub_graph_def, &sub_graph)); - // Convert sub-graph to MLIR module. - TF_ASSIGN_OR_RETURN( - auto sub_module, - GraphDefImporter::Convert(module_->getContext(), sub_graph, debug_info, - flib_def, specs, sig_def_key)); + TF_ASSIGN_OR_RETURN(auto sub_module, + ConvertGraph(sig_def_key, inputs, outputs, {})); mlir::OpBuilder builder(sub_module->getBodyRegion()); // Find the FuncOp which corresponds to current SignatureDef. From b8bb250ebb52d4f037975f58cb1eadb2eff3751c Mon Sep 17 00:00:00 2001 From: George Karpenkov Date: Thu, 18 Jun 2020 10:59:05 -0700 Subject: [PATCH 195/867] Rollback of rollback: [TF/XLA] Only force retracing for non-unique XLA context ID for TPUReplicatedContext Fixes https://github.com/tensorflow/tensorflow/issues/39872 PiperOrigin-RevId: 317137904 Change-Id: Id287e10a0ab2494b11427435d8f89a383eeaf392 --- .../python/eager/def_function_xla_jit_test.py | 19 +++++++++++++++++++ tensorflow/python/eager/function.py | 7 ++++--- tensorflow/python/ops/control_flow_ops.py | 5 +++++ tensorflow/python/tpu/tpu.py | 6 ++++++ 4 files changed, 34 insertions(+), 3 deletions(-) diff --git a/tensorflow/python/eager/def_function_xla_jit_test.py b/tensorflow/python/eager/def_function_xla_jit_test.py index b63a3b434d4..78d44a81b0b 100644 --- a/tensorflow/python/eager/def_function_xla_jit_test.py +++ b/tensorflow/python/eager/def_function_xla_jit_test.py @@ -29,6 +29,7 @@ from tensorflow.python.ops import array_ops from tensorflow.python.ops import control_flow_ops from tensorflow.python.ops import control_flow_util from tensorflow.python.ops import math_ops +from tensorflow.python.ops import random_ops from tensorflow.python.ops import resource_variable_ops from tensorflow.python.ops import tensor_array_ops from tensorflow.python.platform import test @@ -385,6 +386,24 @@ class DefFunctionTest(test.TestCase): f64_input = constant_op.constant([1.1, 2.2, 3.3], dtype=dtypes.float64) self.assertAllClose([1.1, 3.3, 6.6], f(f64_input)) + def testNoExcessiveRetracing(self): + inner_retracings = 0 + + @def_function.function(experimental_compile=True) + def inner(a, b): + nonlocal inner_retracings + inner_retracings += 1 + return a * b + a + + def outer(a, b): + return inner(a, b) + + func_input = random_ops.random_normal([10, 10]) + for _ in range(2): + def_function.function(outer)(func_input, func_input) + + self.assertEqual(inner_retracings, 1) + if __name__ == '__main__': ops.enable_eager_execution() diff --git a/tensorflow/python/eager/function.py b/tensorflow/python/eager/function.py index a40eaf886b3..c02318cb814 100644 --- a/tensorflow/python/eager/function.py +++ b/tensorflow/python/eager/function.py @@ -2981,9 +2981,10 @@ class Function(object): if not executing_eagerly: # We want to force function retracing for each different # XLAControlFlowContext, so add `xla_context_id` to the cache key. - tpu_context = _enclosing_xla_context() - if tpu_context is not None: - xla_context_id = id(tpu_context) + xla_context = _enclosing_xla_context() + if xla_context is not None and \ + xla_context.RequiresUniqueFunctionRetracing(): + xla_context_id = id(xla_context) with ops.init_scope(): # The graph, or whether we're executing eagerly, should be a part of the diff --git a/tensorflow/python/ops/control_flow_ops.py b/tensorflow/python/ops/control_flow_ops.py index 3398308d42e..748f842a9e0 100644 --- a/tensorflow/python/ops/control_flow_ops.py +++ b/tensorflow/python/ops/control_flow_ops.py @@ -3682,6 +3682,11 @@ class XLAControlFlowContext(ControlFlowContext): def AddValue(self, x): return x + def RequiresUniqueFunctionRetracing(self): + """Returns whether the tf.function should be retraced if the context changes. + """ + return False + def from_control_flow_context_def(context_def, import_scope=None): """Deserializes `context_def` into the appropriate ControlFlowContext. diff --git a/tensorflow/python/tpu/tpu.py b/tensorflow/python/tpu/tpu.py index 28eba69b7da..ce3aaa8a058 100644 --- a/tensorflow/python/tpu/tpu.py +++ b/tensorflow/python/tpu/tpu.py @@ -639,6 +639,12 @@ class TPUReplicateContext(control_flow_ops.XLAControlFlowContext): def GetControlPivot(self): return self._pivot + def RequiresUniqueFunctionRetracing(self): + # More context: b/158152827. TPU stack uses the TPUReplicateContext to + # create replicated variable handles and cluster TPU computations, thus we + # always retrace a tf.function when the wrapped TPUReplicateContext changes. + return True + class OutsideCompilationV2Context(control_flow_ops.ControlFlowContext): """The context for outside compilation in Tensorflow 2.0. From 17eea4753dd71b397289f793b348fe9fe751873f Mon Sep 17 00:00:00 2001 From: Mihai Maruseac Date: Thu, 18 Jun 2020 11:09:28 -0700 Subject: [PATCH 196/867] Add OSSFuzz badge to TensorFlow. Now that TF <-> OSSFuzz works again, add badge to show that we are fuzzing the code. PiperOrigin-RevId: 317140301 Change-Id: I401c1ffd3da37f44910fde3ba60fb2b5c925dfcc --- README.md | 1 + tensorflow/security/README.md | 2 ++ 2 files changed, 3 insertions(+) diff --git a/README.md b/README.md index 54c9470b04b..73a345706a4 100644 --- a/README.md +++ b/README.md @@ -95,6 +95,7 @@ for general questions and discussion, and please direct specific questions to The TensorFlow project strives to abide by generally accepted best practices in open-source software development: +[![Fuzzing Status](https://oss-fuzz-build-logs.storage.googleapis.com/badges/tensorflow.svg)](https://bugs.chromium.org/p/oss-fuzz/issues/list?sort=-opened&can=1&q=proj:tensorflow) [![CII Best Practices](https://bestpractices.coreinfrastructure.org/projects/1486/badge)](https://bestpractices.coreinfrastructure.org/projects/1486) [![Contributor Covenant](https://img.shields.io/badge/Contributor%20Covenant-v1.4%20adopted-ff69b4.svg)](CODE_OF_CONDUCT.md) diff --git a/tensorflow/security/README.md b/tensorflow/security/README.md index d9fa1c77a02..34f98e640d6 100644 --- a/tensorflow/security/README.md +++ b/tensorflow/security/README.md @@ -1,5 +1,7 @@ # TensorFlow Security Advisories +[![Fuzzing Status](https://oss-fuzz-build-logs.storage.googleapis.com/badges/tensorflow.svg)](https://bugs.chromium.org/p/oss-fuzz/issues/list?sort=-opened&can=1&q=proj:tensorflow) + We regularly publish security advisories about using TensorFlow. *Note*: In conjunction with these security advisories, we strongly encourage From a31d5da02607a2c5eb01d7c977b92001f842cc89 Mon Sep 17 00:00:00 2001 From: Bruce Fontaine Date: Thu, 18 Jun 2020 11:30:06 -0700 Subject: [PATCH 197/867] Wrap save/restore logic in tf.function when in eager mode. This allows parallel saving and restoring when using multiple devices. PiperOrigin-RevId: 317144560 Change-Id: Iebc230589a5e2712da03c5db3f45e4fd7eeb5ff9 --- .../grappler/optimizers/function_optimizer.cc | 8 +- .../parallel_device/parallel_device_test.py | 4 + .../python/framework/auto_control_deps.py | 2 +- tensorflow/python/training/saving/BUILD | 1 + .../training/saving/functional_saver.py | 111 ++++++++++++------ .../training/saving/functional_saver_test.py | 17 ++- 6 files changed, 101 insertions(+), 42 deletions(-) diff --git a/tensorflow/core/grappler/optimizers/function_optimizer.cc b/tensorflow/core/grappler/optimizers/function_optimizer.cc index a66e645e04b..0e156aaa84c 100644 --- a/tensorflow/core/grappler/optimizers/function_optimizer.cc +++ b/tensorflow/core/grappler/optimizers/function_optimizer.cc @@ -837,7 +837,6 @@ const bool IsExemptFromSideEffectsExecutionValidation(const string& op) { "ParameterizedTruncatedNormal", "TruncatedNormal", "RandomShuffle", "Multinomial", "RandomGamma", "RandomGammaGrad", "RandomPoisson", "RandomPoissonV2", - // LINT.ThenChange(//tensorflow/python/framework/auto_control_deps.py) // ReadVariableOp marked as stateful because it consumes DT_RESOURCE, // but it can't generate any observable side-effect. @@ -851,7 +850,12 @@ const bool IsExemptFromSideEffectsExecutionValidation(const string& op) { // the same device_ordinal on the same host. "EnqueueTPUEmbeddingSparseBatch", "EnqueueTPUEmbeddingIntegerBatch", "EnqueueTPUEmbeddingSparseTensorBatch", - "EnqueueTPUEmbeddingRaggedTensorBatch"}); + "EnqueueTPUEmbeddingRaggedTensorBatch", + + // SaveV2 and RestoreV2 should be allowed to operate in parallel on + // multiple hosts. + "SaveV2", "RestoreV2"}); + // LINT.ThenChange(//tensorflow/python/framework/auto_control_deps.py) return exemption->contains(op); } diff --git a/tensorflow/python/distribute/parallel_device/parallel_device_test.py b/tensorflow/python/distribute/parallel_device/parallel_device_test.py index 8fc3dcb5816..1429c522aba 100644 --- a/tensorflow/python/distribute/parallel_device/parallel_device_test.py +++ b/tensorflow/python/distribute/parallel_device/parallel_device_test.py @@ -172,6 +172,8 @@ class ParallelDeviceTests(_VirtualDeviceTestCase): config.set_synchronous_execution(previous) def test_checkpointing(self): + self.skipTest( + "Disable saving until SaveableObject's methods are traceable.") prefix = os.path.join(self.get_temp_dir(), "ckpt") with self.device.scope(): different_values = self.device.pack( @@ -263,6 +265,8 @@ class LayerTests(_VirtualDeviceTestCase): self.assertIn(self.device.components[1], final_kernels[1].backing_device) def test_training_loop(self): + self.skipTest( + "Disable saving until SaveableObject's methods are traceable.") for _ in range(5): layer = _Dense(5) checkpoint = tracking.Checkpoint(layer=layer) diff --git a/tensorflow/python/framework/auto_control_deps.py b/tensorflow/python/framework/auto_control_deps.py index 51dcb248b11..4b47735e0bf 100644 --- a/tensorflow/python/framework/auto_control_deps.py +++ b/tensorflow/python/framework/auto_control_deps.py @@ -100,7 +100,7 @@ _ORDER_INSENSITIVE_STATEFUL_OPS = [ "CudnnRNNV2", "CudnnRNNV3", "CudnnRNNBackpropV2", "CudnnRNNBackpropV3", "EnqueueTPUEmbeddingSparseBatch", "EnqueueTPUEmbeddingIntegerBatch", "EnqueueTPUEmbeddingSparseTensorBatch", - "EnqueueTPUEmbeddingRaggedTensorBatch" + "EnqueueTPUEmbeddingRaggedTensorBatch", "RestoreV2", "SaveV2" ] # LINT.ThenChange(//tensorflow/core/grappler/optimizers/function_optimizer.cc) diff --git a/tensorflow/python/training/saving/BUILD b/tensorflow/python/training/saving/BUILD index 670a4c35c6f..12940840309 100644 --- a/tensorflow/python/training/saving/BUILD +++ b/tensorflow/python/training/saving/BUILD @@ -43,6 +43,7 @@ cuda_py_test( ":checkpoint_options", ":functional_saver", ":saveable_hook", + "//tensorflow/python/eager:remote", "//tensorflow/python/eager:test", ], ) diff --git a/tensorflow/python/training/saving/functional_saver.py b/tensorflow/python/training/saving/functional_saver.py index c4334e096df..3a9b565470d 100644 --- a/tensorflow/python/training/saving/functional_saver.py +++ b/tensorflow/python/training/saving/functional_saver.py @@ -21,6 +21,7 @@ from __future__ import print_function import uuid from tensorflow.core.protobuf import saver_pb2 +from tensorflow.python.eager import context from tensorflow.python.eager import def_function from tensorflow.python.framework import constant_op from tensorflow.python.framework import dtypes @@ -161,7 +162,8 @@ class MultiDeviceSaver(object): self._after_restore_callbacks.append(saveable.after_restore) if is_saveable: - saveables_by_device.setdefault(saveable.device, []).append(saveable) + host_device = saveable_object_util.set_cpu0(saveable.device) + saveables_by_device.setdefault(host_device, []).append(saveable) self._single_device_savers = { device: _SingleDeviceSaver(saveables) @@ -247,33 +249,50 @@ class MultiDeviceSaver(object): tmp_checkpoint_prefix = string_ops.string_join( [file_prefix, sharded_suffix]) - num_shards = len(self._single_device_savers) - sharded_saves = [] - sharded_prefixes = [] - num_shards_tensor = constant_op.constant(num_shards, name="num_shards") - last_device = None - for shard, (device, saver) in enumerate( - sorted(self._single_device_savers.items())): - last_device = device - with ops.device(saveable_object_util.set_cpu0(device)): - shard_prefix = sharded_filename(tmp_checkpoint_prefix, shard, - num_shards_tensor) - sharded_prefixes.append(shard_prefix) - with ops.device(device): - # _SingleDeviceSaver will use the CPU device when necessary, but initial - # read operations should be placed on the SaveableObject's device. - sharded_saves.append(saver.save(shard_prefix, options)) + def save_fn(): + num_shards = len(self._single_device_savers) + sharded_saves = [] + sharded_prefixes = [] + num_shards_tensor = constant_op.constant(num_shards, name="num_shards") + last_device = None + for shard, (device, saver) in enumerate( + sorted(self._single_device_savers.items())): + last_device = device + with ops.device(saveable_object_util.set_cpu0(device)): + shard_prefix = sharded_filename(tmp_checkpoint_prefix, shard, + num_shards_tensor) + sharded_prefixes.append(shard_prefix) + with ops.device(device): + # _SingleDeviceSaver will use the CPU device when necessary, but + # initial read operations should be placed on the SaveableObject's + # device. + sharded_saves.append(saver.save(shard_prefix, options)) - with ops.control_dependencies(sharded_saves): - # Merge on the io_device if specified, otherwise co-locates the merge op - # with the last device used. - merge_device = (options.experimental_io_device or - saveable_object_util.set_cpu0(last_device)) - with ops.device(merge_device): - # V2 format write path consists of a metadata merge step. Once merged, - # attempts to delete the temporary directory, "_temp". - return gen_io_ops.merge_v2_checkpoints( - sharded_prefixes, file_prefix, delete_old_dirs=True) + with ops.control_dependencies(sharded_saves): + # Merge on the io_device if specified, otherwise co-locates the merge op + # with the last device used. + merge_device = ( + options.experimental_io_device or + saveable_object_util.set_cpu0(last_device)) + with ops.device(merge_device): + # V2 format write path consists of a metadata merge step. Once + # merged, attempts to delete the temporary directory, + # "_temp". + return gen_io_ops.merge_v2_checkpoints( + sharded_prefixes, file_prefix, delete_old_dirs=True) + + # Since this will causes a function re-trace on each save, limit this to the + # cases where it is needed: eager and when there are multiple tasks/single + # device savers. Note that the retrace is needed to ensure we pickup the + # latest values of options like experimental_io_device. + if context.executing_eagerly() and len(self._single_device_savers) > 1: + # Explicitly place the identity op on the first device. + @def_function.function(experimental_compile=False) + def tf_function_save(): + save_fn() + tf_function_save() + else: + return save_fn() def restore(self, file_prefix, options=None): """Restore the saveable objects from a checkpoint with `file_prefix`. @@ -287,12 +306,38 @@ class MultiDeviceSaver(object): A dictionary mapping from SaveableObject names to restore operations. """ options = options or checkpoint_options.CheckpointOptions() - restore_ops = {} - # Sort by device name to avoid propagating non-deterministic dictionary - # ordering in some Python versions. - for device, saver in sorted(self._single_device_savers.items()): - with ops.device(device): - restore_ops.update(saver.restore(file_prefix, options)) + + def restore_fn(): + restore_ops = {} + # Sort by device name to avoid propagating non-deterministic dictionary + # ordering in some Python versions. + for device, saver in sorted(self._single_device_savers.items()): + with ops.device(device): + restore_ops.update(saver.restore(file_prefix, options)) + + return restore_ops + + # Since this will causes a function re-trace on each save, limit this to the + # cases where it is needed: eager and when there are multiple tasks/single + # device savers. Note that the retrace is needed to ensure we pickup the + # latest values of options like experimental_io_device. + if context.executing_eagerly() and len(self._single_device_savers) > 1: + first_device, _ = list(self._single_device_savers.items())[0] + @def_function.function(experimental_compile=False) + def tf_function_restore(): + restore_ops = restore_fn() + restore_tensors = {} + # tf.functions must return tensors, thus we use control dependencies so + # that we can return a tensor which depends on the given op. + with ops.device(saveable_object_util.set_cpu0(first_device)): + for name, op in restore_ops.items(): + with ops.control_dependencies([op]): + restore_tensors[name] = array_ops.identity(file_prefix) + return restore_tensors + + restore_ops = tf_function_restore() + else: + restore_ops = restore_fn() for callback in self._after_restore_callbacks: callback() diff --git a/tensorflow/python/training/saving/functional_saver_test.py b/tensorflow/python/training/saving/functional_saver_test.py index 7db32ff72d7..8f3eef4fb9c 100644 --- a/tensorflow/python/training/saving/functional_saver_test.py +++ b/tensorflow/python/training/saving/functional_saver_test.py @@ -21,6 +21,7 @@ from __future__ import print_function import os from tensorflow.python.eager import context +from tensorflow.python.eager import remote from tensorflow.python.eager import test from tensorflow.python.eager import wrap_function from tensorflow.python.framework import config @@ -29,6 +30,7 @@ from tensorflow.python.framework import ops from tensorflow.python.framework import test_util from tensorflow.python.ops import resource_variable_ops from tensorflow.python.platform import gfile +from tensorflow.python.training import server_lib from tensorflow.python.training.saving import checkpoint_options from tensorflow.python.training.saving import functional_saver from tensorflow.python.training.saving import saveable_hook @@ -126,13 +128,16 @@ class SaverTest(test.TestCase): second_saver.restore(save_path) self.assertEqual(2., self.evaluate(v2)) - @test_util.run_in_graph_and_eager_modes - def test_checkpoint_is_sharded_by_device(self): - with ops.device("cpu:0"): + def test_checkpoint_is_sharded_by_task(self): + servers = [server_lib.Server.create_local_server() for _ in range(3)] + cluster_spec = server_lib.ClusterSpec({ + "worker": [s.target[len("grpc://"):] for s in servers]}) + remote.connect_to_cluster(cluster_spec) + with ops.device("/job:worker/task:0/cpu:0"): v0 = resource_variable_ops.ResourceVariable(0.) - with ops.device("cpu:1"): + with ops.device("/job:worker/task:1/cpu:0"): v1 = resource_variable_ops.ResourceVariable(1.) - with ops.device("cpu:2"): + with ops.device("/job:worker/task:2/cpu:0"): v2 = resource_variable_ops.ResourceVariable(2.) self.evaluate([v0.initializer, v1.initializer, v2.initializer]) @@ -167,7 +172,7 @@ class SaverTest(test.TestCase): list(saveable_object_util.saveable_objects_for_op(v2, "v2"))) prefix = os.path.join(self.get_temp_dir(), "ckpt") self.evaluate(saver.save(constant_op.constant(prefix), self.local_options)) - self.assertEqual(4, len(gfile.Glob(prefix + "*"))) + self.assertEqual(2, len(gfile.Glob(prefix + "*"))) self.evaluate(v0.assign(-1.)) self.evaluate(v1.assign(-1.)) self.evaluate(v2.assign(-1.)) From 4fab49e55013f1265fdd42e53522744ae01351ad Mon Sep 17 00:00:00 2001 From: Ken Franko Date: Thu, 18 Jun 2020 11:39:19 -0700 Subject: [PATCH 198/867] Add suggestion to use experimental_io_device when save/load can't find directory. PiperOrigin-RevId: 317146508 Change-Id: Ia74ac5baa286c90959c7d55f23187d3db46c3c4b --- tensorflow/python/saved_model/load.py | 13 +++++++++++-- tensorflow/python/saved_model/save.py | 12 +++++++++++- 2 files changed, 22 insertions(+), 3 deletions(-) diff --git a/tensorflow/python/saved_model/load.py b/tensorflow/python/saved_model/load.py index 74b030a3797..fb2d01cbee2 100644 --- a/tensorflow/python/saved_model/load.py +++ b/tensorflow/python/saved_model/load.py @@ -28,6 +28,7 @@ from tensorflow.python.eager import context from tensorflow.python.eager import function from tensorflow.python.framework import constant_op from tensorflow.python.framework import dtypes +from tensorflow.python.framework import errors from tensorflow.python.framework import ops from tensorflow.python.framework import tensor_util from tensorflow.python.ops import array_ops @@ -614,8 +615,16 @@ def load_internal(export_dir, tags=None, options=None, loader_cls=Loader): ckpt_options = checkpoint_options.CheckpointOptions( experimental_io_device=options.experimental_io_device) with ops.init_scope(): - loader = loader_cls(object_graph_proto, saved_model_proto, export_dir, - ckpt_options) + try: + loader = loader_cls(object_graph_proto, saved_model_proto, export_dir, + ckpt_options) + except errors.NotFoundError as err: + raise FileNotFoundError( + str(err) + "\n If trying to load on a different device from the " + "computational device, consider using setting the " + "`experimental_io_device` option on tf.saved_model.LoadOptions " + "to the io_device such as '/job:localhost'." + ) root = loader.get(0) if isinstance(loader, Loader): root.graph_debug_info = loader.adjust_debug_info_func_names(debug_info) diff --git a/tensorflow/python/saved_model/save.py b/tensorflow/python/saved_model/save.py index e22b0129dda..5844c80995f 100644 --- a/tensorflow/python/saved_model/save.py +++ b/tensorflow/python/saved_model/save.py @@ -33,6 +33,7 @@ from tensorflow.python.eager import function as defun from tensorflow.python.framework import constant_op from tensorflow.python.framework import dtypes from tensorflow.python.framework import error_interpolation +from tensorflow.python.framework import errors from tensorflow.python.framework import meta_graph from tensorflow.python.framework import ops from tensorflow.python.framework import tensor_util @@ -966,7 +967,16 @@ def save(obj, export_dir, signatures=None, options=None): # SavedModel. Users rely on checking saved_model_dir/saved_model.pb as an # indication that the SavedModel is completely written. if context.executing_eagerly(): - context.async_wait() # Ensure save operations have completed. + try: + context.async_wait() # Ensure save operations have completed. + except errors.NotFoundError as err: + raise FileNotFoundError( + str(err) + "\n If trying to save on a different device from the " + "computational device, consider using setting the " + "`experimental_io_device` option on tf.saved_model.SaveOptions " + "to the io_device such as '/job:localhost'." + ) + path = os.path.join( compat.as_str(export_dir), compat.as_str(constants.SAVED_MODEL_FILENAME_PB)) From bd0a10139a8fa11eb49aa759eca1bb52e662ba34 Mon Sep 17 00:00:00 2001 From: "A. Unique TensorFlower" Date: Thu, 18 Jun 2020 11:50:49 -0700 Subject: [PATCH 199/867] Update ops-related pbtxt files. PiperOrigin-RevId: 317148838 Change-Id: Icecbfedaea80cf2ae6a1424872df62126b7106f1 --- .../BandedTriangularSolve.pbtxt | 42 +++++++ ...tatelessParameterizedTruncatedNormal.pbtxt | 65 +++++++++++ tensorflow/core/ops/ops.pbtxt | 107 ++++++++++++++++++ 3 files changed, 214 insertions(+) create mode 100644 tensorflow/core/ops/compat/ops_history_v2/BandedTriangularSolve.pbtxt create mode 100644 tensorflow/core/ops/compat/ops_history_v2/StatelessParameterizedTruncatedNormal.pbtxt diff --git a/tensorflow/core/ops/compat/ops_history_v2/BandedTriangularSolve.pbtxt b/tensorflow/core/ops/compat/ops_history_v2/BandedTriangularSolve.pbtxt new file mode 100644 index 00000000000..5cf85a62392 --- /dev/null +++ b/tensorflow/core/ops/compat/ops_history_v2/BandedTriangularSolve.pbtxt @@ -0,0 +1,42 @@ +op { + name: "BandedTriangularSolve" + input_arg { + name: "matrix" + type_attr: "T" + } + input_arg { + name: "rhs" + type_attr: "T" + } + output_arg { + name: "output" + type_attr: "T" + } + attr { + name: "lower" + type: "bool" + default_value { + b: true + } + } + attr { + name: "adjoint" + type: "bool" + default_value { + b: false + } + } + attr { + name: "T" + type: "type" + allowed_values { + list { + type: DT_DOUBLE + type: DT_FLOAT + type: DT_HALF + type: DT_COMPLEX64 + type: DT_COMPLEX128 + } + } + } +} diff --git a/tensorflow/core/ops/compat/ops_history_v2/StatelessParameterizedTruncatedNormal.pbtxt b/tensorflow/core/ops/compat/ops_history_v2/StatelessParameterizedTruncatedNormal.pbtxt new file mode 100644 index 00000000000..598125677b1 --- /dev/null +++ b/tensorflow/core/ops/compat/ops_history_v2/StatelessParameterizedTruncatedNormal.pbtxt @@ -0,0 +1,65 @@ +op { + name: "StatelessParameterizedTruncatedNormal" + input_arg { + name: "shape" + type_attr: "S" + } + input_arg { + name: "seed" + type_attr: "Tseed" + } + input_arg { + name: "means" + type_attr: "dtype" + } + input_arg { + name: "stddevs" + type_attr: "dtype" + } + input_arg { + name: "minvals" + type_attr: "dtype" + } + input_arg { + name: "maxvals" + type_attr: "dtype" + } + output_arg { + name: "output" + type_attr: "dtype" + } + attr { + name: "S" + type: "type" + allowed_values { + list { + type: DT_INT32 + type: DT_INT64 + } + } + } + attr { + name: "Tseed" + type: "type" + default_value { + type: DT_INT64 + } + allowed_values { + list { + type: DT_INT32 + type: DT_INT64 + } + } + } + attr { + name: "dtype" + type: "type" + allowed_values { + list { + type: DT_HALF + type: DT_FLOAT + type: DT_DOUBLE + } + } + } +} diff --git a/tensorflow/core/ops/ops.pbtxt b/tensorflow/core/ops/ops.pbtxt index d99f8b8a479..1f1cf7444fb 100644 --- a/tensorflow/core/ops/ops.pbtxt +++ b/tensorflow/core/ops/ops.pbtxt @@ -3070,6 +3070,48 @@ op { } } } +op { + name: "BandedTriangularSolve" + input_arg { + name: "matrix" + type_attr: "T" + } + input_arg { + name: "rhs" + type_attr: "T" + } + output_arg { + name: "output" + type_attr: "T" + } + attr { + name: "lower" + type: "bool" + default_value { + b: true + } + } + attr { + name: "adjoint" + type: "bool" + default_value { + b: false + } + } + attr { + name: "T" + type: "type" + allowed_values { + list { + type: DT_DOUBLE + type: DT_FLOAT + type: DT_HALF + type: DT_COMPLEX64 + type: DT_COMPLEX128 + } + } + } +} op { name: "Barrier" output_arg { @@ -48849,6 +48891,71 @@ op { } } } +op { + name: "StatelessParameterizedTruncatedNormal" + input_arg { + name: "shape" + type_attr: "S" + } + input_arg { + name: "seed" + type_attr: "Tseed" + } + input_arg { + name: "means" + type_attr: "dtype" + } + input_arg { + name: "stddevs" + type_attr: "dtype" + } + input_arg { + name: "minvals" + type_attr: "dtype" + } + input_arg { + name: "maxvals" + type_attr: "dtype" + } + output_arg { + name: "output" + type_attr: "dtype" + } + attr { + name: "S" + type: "type" + allowed_values { + list { + type: DT_INT32 + type: DT_INT64 + } + } + } + attr { + name: "Tseed" + type: "type" + default_value { + type: DT_INT64 + } + allowed_values { + list { + type: DT_INT32 + type: DT_INT64 + } + } + } + attr { + name: "dtype" + type: "type" + allowed_values { + list { + type: DT_HALF + type: DT_FLOAT + type: DT_DOUBLE + } + } + } +} op { name: "StatelessRandomBinomial" input_arg { From 0957b4306325a2a1c0bded767dd0199f118c123e Mon Sep 17 00:00:00 2001 From: Robert David Date: Thu, 18 Jun 2020 11:57:49 -0700 Subject: [PATCH 200/867] LSTM: Use int for 'count' values, fix usage of some appropriate intN_t for exact-size values. PiperOrigin-RevId: 317150337 Change-Id: I5cd8bf8b8231f16cb2d130a8209c45a857b30d21 --- tensorflow/lite/kernels/lstm.cc | 13 +++++++------ tensorflow/lite/kernels/lstm_eval.cc | 14 +++++++------- 2 files changed, 14 insertions(+), 13 deletions(-) diff --git a/tensorflow/lite/kernels/lstm.cc b/tensorflow/lite/kernels/lstm.cc index 0e0c1b9c0f0..aa6a112a022 100644 --- a/tensorflow/lite/kernels/lstm.cc +++ b/tensorflow/lite/kernels/lstm.cc @@ -78,14 +78,14 @@ TfLiteStatus PopulateQuantizedLstmParams8x8_16( auto* proj_params = static_cast( output_tensor->quantization.params); if (cell_clip > 0.0) { - integer_lstm_param->quantized_cell_clip = static_cast(std::min( + integer_lstm_param->quantized_cell_clip = static_cast(std::min( std::max(cell_clip / cell_state_params->scale->data[0], -32768.0f), 32767.0f)); } else { integer_lstm_param->quantized_cell_clip = 0; } if (proj_clip > 0.0) { - integer_lstm_param->quantized_proj_clip = static_cast(std::min( + integer_lstm_param->quantized_proj_clip = static_cast(std::min( std::max(proj_clip / proj_params->scale->data[0], -128.0f), 127.0f)); } else { integer_lstm_param->quantized_proj_clip = 0; @@ -703,14 +703,15 @@ TfLiteStatus PopulateQuantizedLstmParams8x8_8( output_tensor->quantization.params); TF_LITE_ENSURE_EQ(context, cell_state_params->scale->data[0], 1.0 / 32768); if (cell_clip > 0.0 && cell_clip < 1.0) { - integer_lstm_param->quantized_cell_clip = - static_cast(cell_clip / cell_state_params->scale->data[0]); + integer_lstm_param->quantized_cell_clip = static_cast(std::min( + std::max(cell_clip / cell_state_params->scale->data[0], -32768.0f), + 32767.0f)); } else { integer_lstm_param->quantized_cell_clip = 0; } if (proj_clip > 0.0) { - integer_lstm_param->quantized_proj_clip = - proj_clip / proj_params->scale->data[0]; + integer_lstm_param->quantized_proj_clip = static_cast(std::min( + std::max(proj_clip / proj_params->scale->data[0], -128.0f), 127.0f)); } else { integer_lstm_param->quantized_proj_clip = 0; } diff --git a/tensorflow/lite/kernels/lstm_eval.cc b/tensorflow/lite/kernels/lstm_eval.cc index 2e7f300f9a9..f45d46762bf 100644 --- a/tensorflow/lite/kernels/lstm_eval.cc +++ b/tensorflow/lite/kernels/lstm_eval.cc @@ -1044,8 +1044,8 @@ inline void LstmStepInteger( const int32_t* recurrent_to_output_effective_bias, const int32_t* input_to_input_effective_bias, const int32_t* recurrent_to_input_effective_bias, - const int32_t* projection_effective_bias, int32 n_batch, int32 n_cell, - int32 n_input, int32 n_output, int8_t* output_state_ptr, + const int32_t* projection_effective_bias, int n_batch, int n_cell, + int n_input, int n_output, int8_t* output_state_ptr, int32_t output_state_zp, int16_t* cell_ptr, int8_t* output_ptr, int16_t* scratch_0_ptr, int16_t* scratch_1_ptr, int16_t* scratch_2_ptr, int16_t* scratch_3_ptr, int8_t* scratch_4_ptr, int32_t* scratch_5_ptr, @@ -1362,9 +1362,9 @@ void LstmStepInteger( const int32_t* cell_gate_bias_ptr, const int32_t* output_bias_ptr, const int32_t* proj_bias_ptr, const TfLiteLSTMParams* params, const int32_t* intermediate_scale_a, const int32_t* intermediate_scale_b, - const int32_t* intermediate_zp, int32 quantized_cell_clip, - int32 quantized_proj_clip, int32 n_batch, int32 n_cell, int32 n_input, - int32 n_output, int32 output_batch_leading_dim, int8_t* output_state_ptr, + const int32_t* intermediate_zp, int16_t quantized_cell_clip, + int8_t quantized_proj_clip, int n_batch, int n_cell, int n_input, + int n_output, int output_batch_leading_dim, int8_t* output_state_ptr, int32_t output_state_zp, int16_t* cell_ptr, int8_t* output_ptr, int8_t* scratch0, int8_t* scratch1, int16_t* scratch2, int16_t* scratch3, int16_t* scratch4, int16_t* scratch5, int16_t* scratch6, @@ -2129,8 +2129,8 @@ TfLiteStatus EvalInteger8x8_8( int8_t* output_state_ptr = GetTensorData(output_state); int8_t* output_ptr = nullptr; - const int32 input_zp = input->params.zero_point; - const int32 output_state_zp = output_state->params.zero_point; + const int32_t input_zp = input->params.zero_point; + const int32_t output_state_zp = output_state->params.zero_point; // Get params for time/batch/sequence. const int output_batch_leading_dim = From 0cd2551d8c0ce6f1d8e95f465fa8f3d47054eb32 Mon Sep 17 00:00:00 2001 From: Akshay Modi Date: Thu, 18 Jun 2020 12:00:17 -0700 Subject: [PATCH 201/867] Rewrite tf numpy's ndarray as a composite tensor. Expand composites when taking gradients. PiperOrigin-RevId: 317150904 Change-Id: I2fcb0eadd9797340e468bcf82afd17da0d0369e6 --- tensorflow/python/eager/backprop.py | 12 ++- tensorflow/python/eager/backprop_test.py | 69 ++++++++++++++ tensorflow/python/eager/function.py | 4 +- tensorflow/python/ops/numpy_ops/BUILD | 10 ++ tensorflow/python/ops/numpy_ops/np_arrays.py | 41 +++++++- .../python/ops/numpy_ops/np_arrays_test.py | 18 ++++ .../python/ops/numpy_ops/np_interop_test.py | 94 +++++++++++++++++++ 7 files changed, 241 insertions(+), 7 deletions(-) create mode 100644 tensorflow/python/ops/numpy_ops/np_interop_test.py diff --git a/tensorflow/python/eager/backprop.py b/tensorflow/python/eager/backprop.py index dc7bb7c4b11..5c2deb9c0f2 100644 --- a/tensorflow/python/eager/backprop.py +++ b/tensorflow/python/eager/backprop.py @@ -14,6 +14,9 @@ # ============================================================================== """Code for backpropagation using the tape utilities.""" +# TODO(b/159343581): Properly support CompositeTensor in all functions in this +# file. + from __future__ import absolute_import from __future__ import division from __future__ import print_function @@ -1021,7 +1024,7 @@ class GradientTape(object): "derivatives.", 1) flat_targets = [] - for t in nest.flatten(target): + for t in nest.flatten(target, expand_composites=True): if not backprop_util.IsTrainable(t): logging.vlog( logging.WARN, "The dtype of the target tensor must be " @@ -1032,7 +1035,7 @@ class GradientTape(object): t = ops.convert_to_tensor(t) flat_targets.append(t) - flat_sources = nest.flatten(sources) + flat_sources = nest.flatten(sources, expand_composites=True) flat_sources_raw = flat_sources flat_sources = [_handle_or_self(x) for x in flat_sources] for t in flat_sources_raw: @@ -1048,7 +1051,8 @@ class GradientTape(object): if output_gradients is not None: output_gradients = [None if x is None else ops.convert_to_tensor(x) - for x in nest.flatten(output_gradients)] + for x in nest.flatten( + output_gradients, expand_composites=True)] flat_grad = imperative_grad.imperative_grad( self._tape, @@ -1063,7 +1067,7 @@ class GradientTape(object): self._watched_variables = self._tape.watched_variables() self._tape = None - grad = nest.pack_sequence_as(sources, flat_grad) + grad = nest.pack_sequence_as(sources, flat_grad, expand_composites=True) return grad def jacobian(self, diff --git a/tensorflow/python/eager/backprop_test.py b/tensorflow/python/eager/backprop_test.py index a0f98fc0a44..abdac526ce4 100644 --- a/tensorflow/python/eager/backprop_test.py +++ b/tensorflow/python/eager/backprop_test.py @@ -28,6 +28,7 @@ from tensorflow.python.eager import def_function from tensorflow.python.eager import function from tensorflow.python.eager import tape as tape_lib from tensorflow.python.eager import test +from tensorflow.python.framework import composite_tensor from tensorflow.python.framework import constant_op from tensorflow.python.framework import dtypes from tensorflow.python.framework import errors_impl @@ -36,6 +37,7 @@ from tensorflow.python.framework import sparse_tensor from tensorflow.python.framework import tensor_shape from tensorflow.python.framework import tensor_util from tensorflow.python.framework import test_util +from tensorflow.python.framework import type_spec from tensorflow.python.framework.memory_checker import MemoryChecker from tensorflow.python.layers.pooling import max_pooling3d from tensorflow.python.ops import array_ops @@ -52,6 +54,44 @@ from tensorflow.python.ops import resource_variable_ops from tensorflow.python.ops import sparse_ops from tensorflow.python.ops import variables from tensorflow.python.training import training +from tensorflow.python.util import nest + + +# TODO(nareshmodi): This is copied from composite_tensor_test.py. Extract it out +# to a common library to avoid duplication. +class CTSpec(type_spec.TypeSpec): + """A generic CompositeTensor TypeSpec, used for constructing tests.""" + + def __init__(self, component_specs): + self.component_specs = component_specs + + value_type = property(lambda self: CT) + _component_specs = property(lambda self: self.component_specs) + + def _serialize(self): + return (self.component_specs,) + + def _to_components(self, value): + return value.components + + def _from_components(self, tensor_list): + return CT(tensor_list) + + +class CT(composite_tensor.CompositeTensor): + """A generic CompositeTensor, used for constructing tests.""" + _type_spec_class = CTSpec + + def __init__(self, components): + if isinstance(components, list): + components = tuple(components) + self.components = components + + @property + def _type_spec(self): + component_specs = nest.map_structure(type_spec.type_spec_from_value, + self.components) + return self._type_spec_class(component_specs) class BackpropTest(test.TestCase, parameterized.TestCase): @@ -1581,6 +1621,35 @@ class BackpropTest(test.TestCase, parameterized.TestCase): memory_checker.report() memory_checker.assert_no_leak_if_all_possibly_except_one() + def testCompositeTensorAsSource(self): + t = CT([constant_op.constant(3.), constant_op.constant(2.)]) + with backprop.GradientTape() as gt: + gt.watch(t) + y = CT([t.components[0] * 2, t.components[1] * 3]) + + grad = gt.gradient(y, t) + expected_grad = CT([constant_op.constant(2.), constant_op.constant(3.)]) + + flat_grads = nest.flatten(grad, expand_composites=True) + flat_expected_grads = nest.flatten(expected_grad, expand_composites=True) + + self.assertAllClose(flat_grads, flat_expected_grads) + + def testCompositeTensorAsOutputGradients(self): + t = CT([constant_op.constant(3.), constant_op.constant(2.)]) + with backprop.GradientTape() as gt: + gt.watch(t) + y = CT([t.components[0] * 2, t.components[1] * 3]) + + output_gradients = CT([constant_op.constant(5.), constant_op.constant(10.)]) + grad = gt.gradient(y, t, output_gradients=output_gradients) + expected_grad = CT([constant_op.constant(10.), constant_op.constant(30.)]) + + flat_grads = nest.flatten(grad, expand_composites=True) + flat_expected_grads = nest.flatten(expected_grad, expand_composites=True) + + self.assertAllClose(flat_grads, flat_expected_grads) + class JacobianTest(test.TestCase): diff --git a/tensorflow/python/eager/function.py b/tensorflow/python/eager/function.py index c02318cb814..ca1e60c1b7b 100644 --- a/tensorflow/python/eager/function.py +++ b/tensorflow/python/eager/function.py @@ -2626,7 +2626,9 @@ def _is_ndarray(value): # For legacy reasons we do not automatically promote Numpy strings. or isinstance(value, np.str_) # NumPy dtypes have __array__ as unbound methods. - or isinstance(value, type)) + or isinstance(value, type) + # CompositeTensors should be flattened instead. + or isinstance(value, composite_tensor.CompositeTensor)) def _convert_numpy_inputs(inputs): diff --git a/tensorflow/python/ops/numpy_ops/BUILD b/tensorflow/python/ops/numpy_ops/BUILD index 5879bc9f062..3f18a7b3e01 100644 --- a/tensorflow/python/ops/numpy_ops/BUILD +++ b/tensorflow/python/ops/numpy_ops/BUILD @@ -115,3 +115,13 @@ cuda_py_test( "@absl_py//absl/testing:parameterized", ], ) + +cuda_py_test( + name = "np_interop_test", + srcs = ["np_interop_test.py"], + deps = [ + ":numpy", + "//tensorflow/python:platform", + "@absl_py//absl/testing:parameterized", + ], +) diff --git a/tensorflow/python/ops/numpy_ops/np_arrays.py b/tensorflow/python/ops/numpy_ops/np_arrays.py index e2f73100909..8bec8a469a2 100644 --- a/tensorflow/python/ops/numpy_ops/np_arrays.py +++ b/tensorflow/python/ops/numpy_ops/np_arrays.py @@ -24,10 +24,13 @@ import numbers import numpy as np import six +from tensorflow.python.framework import composite_tensor from tensorflow.python.framework import constant_op from tensorflow.python.framework import dtypes from tensorflow.python.framework import ops from tensorflow.python.framework import tensor_shape +from tensorflow.python.framework import tensor_spec +from tensorflow.python.framework import type_spec from tensorflow.python.ops import array_ops from tensorflow.python.ops import math_ops from tensorflow.python.ops.numpy_ops import np_dtypes @@ -175,7 +178,38 @@ def convert_to_tensor(value, dtype=None, dtype_hint=None): return ops.convert_to_tensor(value, dtype=dtype, dtype_hint=dtype_hint) -class ndarray(object): # pylint: disable=invalid-name +class NdarraySpec(type_spec.BatchableTypeSpec): + """Type specification for a `tf.experiemntal.numpy.ndarray`.""" + + value_type = property(lambda self: ndarray) + + def __init__(self, data_spec): + if not isinstance(data_spec, tensor_spec.TensorSpec): + raise ValueError('NdarraySpec.__init__ was expecting a tf.TypeSpec, ' + 'but got a {} instead.'.format(type(data_spec))) + self._data_spec = data_spec + + @property + def _component_specs(self): + return self._data_spec + + def _to_components(self, value): + return value.data + + def _from_components(self, data): + return tensor_to_ndarray(data) + + def _serialize(self): + return (self._data_spec,) + + def _batch(self, batch_size): + return NdarraySpec(self._data_spec.batch(batch_size)) + + def _unbatch(self): + return NdarraySpec(self._data_spec.unbatch()) + + +class ndarray(composite_tensor.CompositeTensor): # pylint: disable=invalid-name """Equivalent of numpy.ndarray backed by TensorFlow tensors. This does not support all features of NumPy ndarrays e.g. strides and @@ -236,7 +270,10 @@ class ndarray(object): # pylint: disable=invalid-name if dtype and dtype != buffer.dtype: buffer = array_ops.bitcast(buffer, dtype) self._data = buffer - self.base = None + + @property + def _type_spec(self): + return NdarraySpec(type_spec.type_spec_from_value(self._data)) @property def data(self): diff --git a/tensorflow/python/ops/numpy_ops/np_arrays_test.py b/tensorflow/python/ops/numpy_ops/np_arrays_test.py index feced98438d..412addc0ad7 100644 --- a/tensorflow/python/ops/numpy_ops/np_arrays_test.py +++ b/tensorflow/python/ops/numpy_ops/np_arrays_test.py @@ -29,6 +29,7 @@ from tensorflow.python.ops.numpy_ops import np_arrays # Required for operator overloads from tensorflow.python.ops.numpy_ops import np_math_ops # pylint: disable=unused-import from tensorflow.python.platform import test +from tensorflow.python.util import nest t2a = np_arrays.tensor_to_ndarray @@ -182,6 +183,23 @@ class ArrayTest(test.TestCase): with self.assertRaisesWithPredicateMatch(TypeError, r'unhashable type'): hash(a) + def testFromToCompositeTensor(self): + tensors = [t2a(ops.convert_to_tensor(0.1)), t2a(ops.convert_to_tensor(0.2))] + + flattened = nest.flatten(tensors, expand_composites=True) + # Each ndarray contains only one tensor, so the flattened output should be + # just 2 tensors in a list. + self.assertLen(flattened, 2) + self.assertIsInstance(flattened[0], ops.Tensor) + self.assertIsInstance(flattened[1], ops.Tensor) + + repacked = nest.pack_sequence_as(tensors, flattened, expand_composites=True) + self.assertLen(repacked, 2) + self.assertIsInstance(repacked[0], np_arrays.ndarray) + self.assertIsInstance(repacked[1], np_arrays.ndarray) + + self.assertAllClose(tensors, repacked) + if __name__ == '__main__': # TODO(wangpeng): Test in graph mode as well. diff --git a/tensorflow/python/ops/numpy_ops/np_interop_test.py b/tensorflow/python/ops/numpy_ops/np_interop_test.py new file mode 100644 index 00000000000..052949dff9d --- /dev/null +++ b/tensorflow/python/ops/numpy_ops/np_interop_test.py @@ -0,0 +1,94 @@ +# Copyright 2020 The TensorFlow Authors. All Rights Reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. +# ============================================================================== +"""Tests for interop between TF ops, numpy_ops, and numpy methods.""" + +from __future__ import absolute_import +from __future__ import division +from __future__ import print_function + + +from tensorflow.python.eager import backprop +from tensorflow.python.eager import def_function +from tensorflow.python.framework import ops +from tensorflow.python.ops import control_flow_ops +from tensorflow.python.ops.numpy_ops import np_array_ops +from tensorflow.python.ops.numpy_ops import np_arrays +from tensorflow.python.platform import test + + +class InteropTest(test.TestCase): + + def testGradientTapeInterop(self): + with backprop.GradientTape() as t: + x = np_array_ops.asarray(3.0) + y = np_array_ops.asarray(2.0) + + t.watch([x, y]) + + xx = 2 * x + yy = 3 * y + + dx, dy = t.gradient([xx, yy], [x, y]) + + # TODO(nareshmodi): Gradient tape returns tensors. Is it possible to rewrap? + self.assertAllClose(dx, 2.0) + self.assertAllClose(dy, 3.0) + + def testFunctionInterop(self): + x = np_array_ops.asarray(3.0) + y = np_array_ops.asarray(2.0) + + add = lambda x, y: x + y + add_fn = def_function.function(add) + + raw_result = add(x, y) + fn_result = add_fn(x, y) + + self.assertIsInstance(raw_result, np_arrays.ndarray) + self.assertIsInstance(fn_result, np_arrays.ndarray) + self.assertAllClose(raw_result, fn_result) + + def testCondInterop(self): + x = np_array_ops.asarray(3.0) + + def fn(x): + x_plus_1 = control_flow_ops.cond(x > 0, lambda: x+1, lambda: x+2) + x_plus_2 = control_flow_ops.cond(x < 0, lambda: x+1, lambda: x+2) + + return x_plus_1, x_plus_2 + + raw_x_plus_1, raw_x_plus_2 = fn(x) + fn_x_plus_1, fn_x_plus_2 = def_function.function(fn)(x) + + self.assertAllClose(raw_x_plus_1, x + 1) + self.assertAllClose(raw_x_plus_2, x + 2) + + self.assertAllClose(fn_x_plus_1, x + 1) + self.assertAllClose(fn_x_plus_2, x + 2) + + def testWhileInterop(self): + def fn(): + x = np_array_ops.asarray(0) + c = lambda x: x < 10000 + b = lambda x: [x + 1] + return control_flow_ops.while_loop_v2(c, b, [x], parallel_iterations=20) + + self.assertEqual(10000, fn()[0]) + self.assertEqual(10000, def_function.function(fn)()[0]) + + +if __name__ == '__main__': + ops.enable_eager_execution() + test.main() From 629e6077d0c4ccc342a15496cd91becedfc02bc3 Mon Sep 17 00:00:00 2001 From: Andrew Audibert Date: Thu, 18 Jun 2020 12:25:56 -0700 Subject: [PATCH 202/867] [tf.data service] Add test that options are applied to the distributed dataset PiperOrigin-RevId: 317156222 Change-Id: Ie1956a3ef474371e381618c4808fee7a8ef8ee08 --- .../kernel_tests/data_service_ops_test.py | 30 +++++++++++++++++++ 1 file changed, 30 insertions(+) diff --git a/tensorflow/python/data/kernel_tests/data_service_ops_test.py b/tensorflow/python/data/kernel_tests/data_service_ops_test.py index 2356a866d6e..796ab328980 100644 --- a/tensorflow/python/data/kernel_tests/data_service_ops_test.py +++ b/tensorflow/python/data/kernel_tests/data_service_ops_test.py @@ -23,6 +23,7 @@ from absl.testing import parameterized from tensorflow.python.data.experimental.ops import data_service_ops from tensorflow.python.data.experimental.ops import distribute_options +from tensorflow.python.data.experimental.ops import testing from tensorflow.python.data.experimental.service import server_lib from tensorflow.python.data.kernel_tests import test_base from tensorflow.python.data.ops import dataset_ops @@ -30,6 +31,7 @@ from tensorflow.python.eager import def_function from tensorflow.python.framework import combinations from tensorflow.python.framework import dtypes from tensorflow.python.framework import errors +from tensorflow.python.ops import math_ops from tensorflow.python.ops import random_ops from tensorflow.python.ops import tensor_array_ops from tensorflow.python.platform import test @@ -317,6 +319,34 @@ class DataServiceOpsTest(test_base.DatasetTestBase, parameterized.TestCase): results.append(elem.numpy()) self.assertCountEqual(num_repetitions * list(range(num_elements)), results) + @combinations.generate(test_base.eager_only_combinations()) + def testApplyDeterminismOption(self): + elements = list(range(10)) + master_address = self.create_cluster(1) + + def dataset_fn(delay_ms): + + def interleave_fn(x): + ds = dataset_ops.Dataset.from_tensors(x) + if math_ops.equal(x, 0): + ds = ds.apply(testing.sleep(delay_ms * 1000)) + else: + ds = ds.apply(testing.sleep(0)) + return ds + + ds = dataset_ops.Dataset.from_tensor_slices(elements) + ds = ds.interleave(interleave_fn, cycle_length=10, num_parallel_calls=10) + opts = dataset_ops.Options() + opts.experimental_deterministic = False + ds = ds.with_options(opts) + ds = _make_distributed_dataset(ds, master_address) + return ds + + self.checkDeterminism( + dataset_fn=dataset_fn, + expect_determinism=False, + expected_elements=elements) + def run_stateful(self, external_state_policy): num_elements = 10 ds = dataset_ops.Dataset.range(num_elements).map( From d8bfc935fdc77ddc138e415dd8bb47f9817c4d5e Mon Sep 17 00:00:00 2001 From: Reed Date: Thu, 18 Jun 2020 12:42:19 -0700 Subject: [PATCH 203/867] Add MKL supoprt to auto_mixed_precision. This extends the auto mixed precision grappler pass to support converting nodes to bfloat16 on MKL-supported CPUs. Co-authored-by: Niranjan Hasabnis --- tensorflow/core/grappler/optimizers/BUILD | 10 +- .../optimizers/auto_mixed_precision.cc | 229 +++--- .../optimizers/auto_mixed_precision.h | 20 +- .../optimizers/auto_mixed_precision_lists.h | 257 +++++-- .../optimizers/auto_mixed_precision_test.cc | 234 ++++++- .../grappler/optimizers/meta_optimizer.cc | 11 +- .../core/protobuf/rewriter_config.proto | 8 +- .../grappler/auto_mixed_precision_test.py | 661 ++++++++++-------- 8 files changed, 951 insertions(+), 479 deletions(-) diff --git a/tensorflow/core/grappler/optimizers/BUILD b/tensorflow/core/grappler/optimizers/BUILD index 7432e2d54ea..f9b4c6e6d81 100644 --- a/tensorflow/core/grappler/optimizers/BUILD +++ b/tensorflow/core/grappler/optimizers/BUILD @@ -1,5 +1,5 @@ load("//tensorflow:tensorflow.bzl", "tf_cc_test", "tf_cc_test_mkl", "tf_kernel_library") -load("//tensorflow:tensorflow.bzl", "tf_cuda_cc_test") +load("//tensorflow:tensorflow.bzl", "tf_copts", "tf_cuda_cc_test") # Platform specific build config load( @@ -7,6 +7,11 @@ load( "if_static", ) +load( + "//third_party/mkl:build_defs.bzl", + "mkl_deps", +) + package( licenses = ["notice"], # Apache 2.0 ) @@ -611,6 +616,7 @@ cc_library( "auto_mixed_precision_lists.h", ], visibility = ["//visibility:public"], + copts = tf_copts(), deps = [ ":custom_graph_optimizer_registry", ":graph_optimizer", @@ -627,7 +633,7 @@ cc_library( "//tensorflow/core/grappler/costs:virtual_placer", "@com_google_absl//absl/container:flat_hash_map", "@com_google_absl//absl/container:flat_hash_set", - ], + ] + mkl_deps(), ) tf_cuda_cc_test( diff --git a/tensorflow/core/grappler/optimizers/auto_mixed_precision.cc b/tensorflow/core/grappler/optimizers/auto_mixed_precision.cc index fa6ca3144a5..2e0b56fd3e2 100644 --- a/tensorflow/core/grappler/optimizers/auto_mixed_precision.cc +++ b/tensorflow/core/grappler/optimizers/auto_mixed_precision.cc @@ -16,6 +16,7 @@ limitations under the License. #include "tensorflow/core/grappler/optimizers/auto_mixed_precision.h" #include +#include #include "absl/container/flat_hash_map.h" #include "absl/container/flat_hash_set.h" @@ -52,6 +53,7 @@ const std::pair kMinGPUArch = {0, 0}; const char kSuffix[] = "AutoMixedPrecision"; const char kCastToFp16[] = "CastToFp16"; +const char kCastToBf16[] = "CastToBf16"; const char kCastToFp32[] = "CastToFp32"; // Instances of this class represent unique type attribute identifiers within a @@ -840,22 +842,6 @@ DataTypeSet AllowedDataTypes(const OpDef& op_def, const TypeAttrId& t_attr_id) { return AllowedDataTypes(*attr_def); } -NodeDef BuildCastNode(const MutableGraphView::OutputPort& src, bool to_fp16, - const string& device) { - const char* cast_string = to_fp16 ? kCastToFp16 : kCastToFp32; - string name = strings::StrCat(src.node->name(), "-", src.port_id, "-", - cast_string, "-", kSuffix); - NodeDef node; - node.set_name(name); - node.set_op("Cast"); - node.set_device(device); - node.add_input(strings::StrCat(src.node->name(), ":", src.port_id)); - (*node.mutable_attr())["SrcT"].set_type(to_fp16 ? DT_FLOAT : DT_HALF); - (*node.mutable_attr())["DstT"].set_type(to_fp16 ? DT_HALF : DT_FLOAT); - (*node.mutable_attr())["Truncate"].set_b(false); - return node; -} - Status ValidateLists(const gtl::FlatSet& white_list, const gtl::FlatSet& black_list, const gtl::FlatSet& gray_list, @@ -941,7 +927,8 @@ class AutoMixedPrecisionImpl { public: AutoMixedPrecisionImpl(Cluster* cluster, const std::unordered_set& nodes_to_preserve, - GraphDef* graph, string id) + GraphDef* graph, string id, + AutoMixedPrecisionMode mode) : virtual_placer_(cluster->GetDevices()), nodes_to_preserve_(nodes_to_preserve), graph_(graph), @@ -949,23 +936,35 @@ class AutoMixedPrecisionImpl { id_(id), graph_view_(graph), cuda_version_(GetCudaVersion(*cluster)), - cudnn_version_(GetCudnnVersion(*cluster)) {} + cudnn_version_(GetCudnnVersion(*cluster)), + mode_(mode), + target_dtype_(mode_ == AutoMixedPrecisionMode::CUDA ? DT_HALF + : DT_BFLOAT16) {} Status Optimize(); private: typedef absl::flat_hash_set NodeTypeIdSet; + std::unique_ptr get_mixed_precision_lists() const { + switch (mode_) { + case AutoMixedPrecisionMode::CUDA: + return std::make_unique(cuda_version_, + cudnn_version_); + case AutoMixedPrecisionMode::MKL: + return std::make_unique(); + } + } Status PrintDebugLogs(bool preop, size_t timestamp); void LogSkippedNode(const NodeDef& node) const; bool MustPreserve(const NodeDef& node) const; - bool IsOnGPU(const NodeDef& node) const; + bool IsOnDevice(const NodeDef& node, const string& device_type) const; bool IsOnSuitableGPUArch(const NodeDef& node) const; bool ShouldProcess(const NodeDef& node) const; - bool NodeHasFP16KernelForTypeAttr(const NodeDef& node, TypeAttrId taid) const; + bool NodeHasF16KernelForTypeAttr(const NodeDef& node, TypeAttrId taid) const; bool NodeImplicitlyReadsNonResourceVariable(const NodeDef& node) const; void ConvertBatchNormOpsToV2(); - bool SupportsFloat16(const NodeTypeId& node_type) const; + bool SupportsF16(const NodeTypeId& node_type) const; const NodeTypeId* GetTensorListFloat32NodeTypeId(const NodeDef& node) const; bool IsSourceOrSinkOp(const string& op) const; void FindFloat32TensorListOpClustersAndBlacklistUnsafe( @@ -990,6 +989,8 @@ class AutoMixedPrecisionImpl { absl::flat_hash_set* white_set) const; void MakeCastsWhiteIfAllOutputsWhite( absl::flat_hash_set* white_set) const; + NodeDef BuildCastNode(const MutableGraphView::OutputPort& src, bool to_f16, + const string& device) const; Status ChangeTypeAttrsAndAddCasts(const absl::flat_hash_set& white_set); VirtualPlacer virtual_placer_; @@ -1003,21 +1004,44 @@ class AutoMixedPrecisionImpl { NodeTypeAttrMap node_type_map_; GraphTypeTopologyView graph_type_view_; bool force_all_fp16_; - gtl::FlatSet fp16_whitelist_; - gtl::FlatSet fp16_blacklist_; - gtl::FlatSet fp16_graylist_; - gtl::FlatSet fp16_clearlist_; + AutoMixedPrecisionMode mode_; + gtl::FlatSet f16_whitelist_; + gtl::FlatSet f16_blacklist_; + gtl::FlatSet f16_graylist_; + gtl::FlatSet f16_clearlist_; absl::flat_hash_set should_process_nodes_; + DataType target_dtype_; // Either DT_HALF or DT_BFLOAT16 }; -bool AutoMixedPrecisionImpl::NodeHasFP16KernelForTypeAttr( +NodeDef AutoMixedPrecisionImpl::BuildCastNode( + const MutableGraphView::OutputPort& src, bool to_f16, + const string& device) const { + DataType src_type = to_f16 ? DT_FLOAT : target_dtype_; + DataType dst_type = to_f16 ? target_dtype_ : DT_FLOAT; + const char* cast_string = + !to_f16 ? kCastToFp32 + : target_dtype_ == DT_HALF ? kCastToFp16 : kCastToBf16; + string name = strings::StrCat(src.node->name(), "-", src.port_id, "-", + cast_string, "-", kSuffix); + NodeDef node; + node.set_name(name); + node.set_op("Cast"); + node.set_device(device); + node.add_input(strings::StrCat(src.node->name(), ":", src.port_id)); + (*node.mutable_attr())["SrcT"].set_type(src_type); + (*node.mutable_attr())["DstT"].set_type(dst_type); + (*node.mutable_attr())["Truncate"].set_b(false); + return node; +} + +bool AutoMixedPrecisionImpl::NodeHasF16KernelForTypeAttr( const NodeDef& node, TypeAttrId taid) const { NodeDef node_copy(node); if (node.device().empty()) { string device_name = virtual_placer_.get_canonical_device_name(node); node_copy.set_device(device_name); } - if (!SetDataType(&node_copy, taid, DataType::DT_HALF)) { + if (!SetDataType(&node_copy, taid, target_dtype_)) { return false; } return IsKernelRegisteredForNode(node_copy).ok(); @@ -1053,21 +1077,22 @@ Status AutoMixedPrecisionImpl::PrintDebugLogs(bool preop, size_t timestamp) { fname = io::JoinPath(prepend_path, strings::StrCat("paintbuckets", suffix, ".txt")); f.open(fname.c_str(), std::fstream::out); + std::unique_ptr mp_lists = + get_mixed_precision_lists(); f << "WhiteList:\n"; - for (const auto& x : - AutoMixedPrecisionLists::WhiteList(cuda_version_, cudnn_version_)) { + for (const auto& x : mp_lists->WhiteList()) { f << x << "\n"; } f << "\nBlackList:\n"; - for (const auto& x : AutoMixedPrecisionLists::BlackList()) { + for (const auto& x : mp_lists->BlackList()) { f << x << "\n"; } f << "\nGrayList:\n"; - for (const auto& x : AutoMixedPrecisionLists::GrayList()) { + for (const auto& x : mp_lists->GrayList()) { f << x << "\n"; } f << "\nClearList:\n"; - for (const auto& x : AutoMixedPrecisionLists::ClearList()) { + for (const auto& x : mp_lists->ClearList()) { f << x << "\n"; } f.close(); @@ -1088,7 +1113,8 @@ bool AutoMixedPrecisionImpl::MustPreserve(const NodeDef& node) const { return nodes_to_preserve_.count(node.name()); } -bool AutoMixedPrecisionImpl::IsOnGPU(const NodeDef& node) const { +bool AutoMixedPrecisionImpl::IsOnDevice(const NodeDef& node, + const string& device_type) const { string device_name; if (node.device().empty()) { device_name = virtual_placer_.get_canonical_device_name(node); @@ -1099,7 +1125,7 @@ bool AutoMixedPrecisionImpl::IsOnGPU(const NodeDef& node) const { string not_used; if (DeviceNameUtils::SplitDeviceName(device_name, ¬_used, &device) && absl::StrContains(absl::AsciiStrToLower(device), - absl::AsciiStrToLower(DEVICE_GPU))) { + absl::AsciiStrToLower(device_type))) { return true; } return false; @@ -1164,15 +1190,14 @@ bool IsTensorListWriterOp(const string& op) { return tensor_list_writer_ops.count(op); } -bool AutoMixedPrecisionImpl::SupportsFloat16( - const NodeTypeId& node_type) const { +bool AutoMixedPrecisionImpl::SupportsF16(const NodeTypeId& node_type) const { const OpDef* op_def; Status status = OpRegistry::Global()->LookUpOpDef(node_type.node->op(), &op_def); if (!status.ok()) return false; return AllowedDataTypes(*op_def, node_type.type_attr) - .Contains(DataType::DT_HALF) && - NodeHasFP16KernelForTypeAttr(*node_type.node, node_type.type_attr); + .Contains(target_dtype_) && + NodeHasF16KernelForTypeAttr(*node_type.node, node_type.type_attr); } // TODO(mconley): Make this change the node's name (to aid debugging). Need to @@ -1219,22 +1244,40 @@ Status AutoMixedPrecisionImpl::Optimize() { "TF_AUTO_MIXED_PRECISION_GRAPH_REWRITE_LEVEL", "", &optimization_level)); optimization_level = absl::AsciiStrToUpper(optimization_level); force_all_fp16_ = optimization_level == "UNSAFE_FORCE_ALL"; + if (force_all_fp16_ && mode_ == AutoMixedPrecisionMode::MKL) { + // Many ops do not support bfloat16 on the CPU so we disallowing forcing to + // bfloat16. + return errors::InvalidArgument( + "TF_AUTO_MIXED_PRECISION_GRAPH_REWRITE_LEVEL cannot be set to " + "UNSAFE_FORCE_ALL when MKL is used"); + } - fp16_whitelist_ = - AutoMixedPrecisionLists::WhiteList(cuda_version_, cudnn_version_); - fp16_blacklist_ = AutoMixedPrecisionLists::BlackList(); - fp16_graylist_ = AutoMixedPrecisionLists::GrayList(); - fp16_clearlist_ = AutoMixedPrecisionLists::ClearList(); - TF_RETURN_IF_ERROR(ValidateLists(fp16_whitelist_, fp16_blacklist_, - fp16_graylist_, fp16_clearlist_)); + std::unique_ptr mp_lists = + get_mixed_precision_lists(); + f16_whitelist_ = mp_lists->WhiteList(); + f16_blacklist_ = mp_lists->BlackList(); + f16_graylist_ = mp_lists->GrayList(); + f16_clearlist_ = mp_lists->ClearList(); + TF_RETURN_IF_ERROR(ValidateLists(f16_whitelist_, f16_blacklist_, + f16_graylist_, f16_clearlist_)); size_t timestamp = Env::Default()->NowMicros() / 1000; TF_RETURN_IF_ERROR(PrintDebugLogs(/* preop = */ true, timestamp)); VLOG(2) << "Identifying nodes that should be processed"; for (const NodeDef& node : graph_->node()) { - if (!MustPreserve(node) && IsOnGPU(node) && - (ShouldIgnorePerformance() || IsOnSuitableGPUArch(node))) { + bool should_process; + switch (mode_) { + case AutoMixedPrecisionMode::CUDA: + should_process = + !MustPreserve(node) && IsOnDevice(node, DEVICE_GPU) && + (ShouldIgnorePerformance() || IsOnSuitableGPUArch(node)); + break; + case AutoMixedPrecisionMode::MKL: + should_process = !MustPreserve(node) && IsOnDevice(node, DEVICE_CPU); + break; + } + if (should_process) { should_process_nodes_.insert(&node); } else { LogSkippedNode(node); @@ -1260,29 +1303,29 @@ Status AutoMixedPrecisionImpl::Optimize() { for (const auto& cluster : tensor_list_clusters) { VLOG(1) << "Found safe Tensor List cluster of size " << cluster.size(); for (const NodeDef* node : cluster) { - VLOG(2) << "Cluster member: " << node->op() << " node " << node->name(); + VLOG(2) << " Cluster member: " << node->op() << " node " << node->name(); } FindTensorListImplicitFloat32Edges(cluster, &ephemeral_edges); } TF_RETURN_IF_ERROR(graph_type_view_.AddEphemeralEdges(ephemeral_edges)); - // The goal here is to change performance-critical ops to fp16, and to do so - // with the minimal number of casts, subject to the constraint that the + // The goal here is to change performance-critical ops to fp16 or bf16, and to + // do so with the minimal number of casts, subject to the constraint that the // model's convergence is not affected. This is achieved by first identifying - // which nodes should be changed to fp16 and then inserting casts at the - // boundaries between fp16/non-fp16 nodes. + // which nodes should be changed to f16 and then inserting casts at the + // boundaries between f16/non-f16 nodes. - // The algorithm for deciding which nodes to change to fp16 is as follows: + // The algorithm for deciding which nodes to change to f16 is as follows: // 1) Add all performance-critical ops (aka "whitelist" ops) to the white_set. // This is done under the assumption that whitelist ops are always - // numerically-safe in fp16 and that they are the most important ops for + // numerically-safe in f16 and that they are the most important ops for // improving performance. // 2) Add nodes to the black_set iff they are numerically-dangerous (aka // "blacklist" ops) or they are on a forward path from a blacklist node to // a black/gray node (including the node at the end of the path) through // non-numerically-dangerous ops (aka "greylist" and "clearlist" ops). // This is done to prevent numerically-dangerous ops and their downstream - // effects from being changed to fp16, which would risk breaking the + // effects from being changed to f16, which would risk breaking the // numerical accuracy of the model. // 3) For all remaining nodes that are not considered dangerous (greylist // and clearlist ops), find those that are between (i.e., both upstream @@ -1480,7 +1523,7 @@ void AutoMixedPrecisionImpl::AddWhitelistOps( const NodeTypeId& root = *graph_type_view_.GetNode(root_idx); if (!ShouldProcess(*root.node)) continue; bool force_white = force_all_fp16_ && CanForceFP16(*root.node); - if (fp16_whitelist_.count(root.node->op()) || force_white) { + if (f16_whitelist_.count(root.node->op()) || force_white) { bool inserted = white_set->insert(root_idx).second; if (VLOG_IS_ON(2) && inserted) { VLOG(2) << "Painting type " << root.type_attr.DebugString() @@ -1504,8 +1547,8 @@ void AutoMixedPrecisionImpl::PropagateBlackFwdThroughClearAndGray( absl::flat_hash_set upstream_of_black_or_gray_set; for (int root_idx = 0; root_idx < graph_type_view_.num_nodes(); ++root_idx) { const NodeTypeId& root = *graph_type_view_.GetNode(root_idx); - if (!(fp16_blacklist_.count(root.node->op()) || - fp16_graylist_.count(root.node->op()))) { + if (!(f16_blacklist_.count(root.node->op()) || + f16_graylist_.count(root.node->op()))) { continue; } DfsTypeTraversal(graph_type_view_, {&root}, @@ -1514,7 +1557,7 @@ void AutoMixedPrecisionImpl::PropagateBlackFwdThroughClearAndGray( const NodeTypeId& item = *graph_type_view_.GetNode(idx); return idx == root_idx || (!upstream_of_black_or_gray_set.count(idx) && - fp16_clearlist_.count(item.node->op())); + f16_clearlist_.count(item.node->op())); }), DfsTypeCallbacks::PreOrder([&](int idx) { upstream_of_black_or_gray_set.insert(idx); @@ -1524,7 +1567,7 @@ void AutoMixedPrecisionImpl::PropagateBlackFwdThroughClearAndGray( // Propagate black forward through nodes in upstream_of_black_or_gray_set. for (int root_idx = 0; root_idx < graph_type_view_.num_nodes(); ++root_idx) { const NodeTypeId& root = *graph_type_view_.GetNode(root_idx); - if (black_set->count(root_idx) || !fp16_blacklist_.count(root.node->op())) { + if (black_set->count(root_idx) || !f16_blacklist_.count(root.node->op())) { continue; } DfsTypeTraversal( @@ -1552,7 +1595,7 @@ void AutoMixedPrecisionImpl::AddClearAndGrayToWhiteIfBetweenWhite( absl::flat_hash_set downstream_of_white_set; for (int root_idx = 0; root_idx < graph_type_view_.num_nodes(); ++root_idx) { const NodeTypeId& root = *graph_type_view_.GetNode(root_idx); - if (!ShouldProcess(*root.node) || !fp16_whitelist_.count(root.node->op())) { + if (!ShouldProcess(*root.node) || !f16_whitelist_.count(root.node->op())) { continue; } DfsTypeTraversal( @@ -1561,14 +1604,14 @@ void AutoMixedPrecisionImpl::AddClearAndGrayToWhiteIfBetweenWhite( const NodeTypeId& item = *graph_type_view_.GetNode(idx); return idx == root_idx || (!downstream_of_white_set.count(idx) && - !fp16_whitelist_.count(item.node->op()) && + !f16_whitelist_.count(item.node->op()) && !black_set.count(idx) && ShouldProcess(*item.node) && // TODO(benbarsdell): Consider allowing propagation through // ops that are already float16 in order to reduce the number // of casts. - IsFloat32(item) && SupportsFloat16(item) && - (fp16_clearlist_.count(item.node->op()) || - fp16_graylist_.count(item.node->op()))); + IsFloat32(item) && SupportsF16(item) && + (f16_clearlist_.count(item.node->op()) || + f16_graylist_.count(item.node->op()))); }), DfsTypeCallbacks::PreOrder( [&](int idx) { downstream_of_white_set.insert(idx); })); @@ -1579,7 +1622,7 @@ void AutoMixedPrecisionImpl::AddClearAndGrayToWhiteIfBetweenWhite( for (int root_idx = 0; root_idx < graph_type_view_.num_nodes(); ++root_idx) { const NodeTypeId& root = *graph_type_view_.GetNode(root_idx); if (!ShouldProcess(*root.node) || upstream_of_white_set.count(root_idx) || - !fp16_whitelist_.count(root.node->op())) { + !f16_whitelist_.count(root.node->op())) { continue; } DfsTypeTraversal( @@ -1620,8 +1663,8 @@ void AutoMixedPrecisionImpl::PropagateWhiteThroughClear( return idx == root_idx || (!white_set->count(idx) && !black_set.count(idx) && ShouldProcess(*item.node) && IsFloat32(item) && - SupportsFloat16(item) && - (fp16_clearlist_.count(item.node->op())) && + SupportsF16(item) && + (f16_clearlist_.count(item.node->op())) && // We don't propagate (backwards) through nodes that read // Variables because it can break the behavior of TensorBoard // visualization and/or (in the case of Enter nodes) the model @@ -1806,13 +1849,13 @@ void AutoMixedPrecisionImpl::MakeCastsWhiteIfAllOutputsWhite( } } -// Changes all white-painted type attributes to DT_HALF, and inserts Cast nodes -// at node outputs for all edges that connect white-painted <-> -// non-white-painted type attributes. +// Changes all white-painted type attributes to DT_HALF or DT_BFLOAT16, and +// inserts Cast nodes at node outputs for all edges that connect +// white-painted <-> non-white-painted type attributes. Status AutoMixedPrecisionImpl::ChangeTypeAttrsAndAddCasts( const absl::flat_hash_set& white_set) { int num_nodes_changed = 0; - int num_nonvar_casts_to_fp16 = 0; + int num_nonvar_casts_to_f16 = 0; int num_nodes_preop = graph_->node_size(); for (int node_idx = 0; node_idx < num_nodes_preop; ++node_idx) { NodeDef* node = graph_->mutable_node(node_idx); @@ -1829,8 +1872,9 @@ Status AutoMixedPrecisionImpl::ChangeTypeAttrsAndAddCasts( bool src_is_white = white_set.count(node_type_idx); if (src_is_white) { VLOG(1) << "Changing type " << type_attr.DebugString() << " of " - << node->op() << " node " << node->name() << " to DT_HALF"; - if (!SetDataType(node, type_attr, DT_HALF)) { + << node->op() << " node " << node->name() << " to " + << DataTypeString(target_dtype_); + if (!SetDataType(node, type_attr, target_dtype_)) { return errors::Internal("Failed to set type attribute"); } ++num_nodes_changed; @@ -1855,16 +1899,16 @@ Status AutoMixedPrecisionImpl::ChangeTypeAttrsAndAddCasts( bool dst_is_white = white_set.count(dst_type_idx); if (src_is_white != dst_is_white) { if (!added_cast_node) { - bool to_fp16 = dst_is_white; + bool to_f16 = dst_is_white; VLOG(1) << "Inserting cast to " - << (to_fp16 ? "DT_HALF" : "DT_FLOAT") << " at " - << src.node->op() << " " << src.node->name() << ":" - << src.port_id; + << (to_f16 ? DataTypeString(target_dtype_) : "DT_FLOAT") + << " at " << src.node->op() << " " << src.node->name() + << ":" << src.port_id; added_cast_node = graph_view_.AddNode( - BuildCastNode(src, to_fp16, src.node->device())); - if (to_fp16 && !IsConstant(*node) && !IsVariable(*node) && + BuildCastNode(src, to_f16, src.node->device())); + if (to_f16 && !IsConstant(*node) && !IsVariable(*node) && !NodeImplicitlyReadsNonResourceVariable(*node)) { - ++num_nonvar_casts_to_fp16; + ++num_nonvar_casts_to_f16; } } TF_RETURN_IF_ERROR(graph_view_.UpdateRegularFaninByPort( @@ -1874,9 +1918,13 @@ Status AutoMixedPrecisionImpl::ChangeTypeAttrsAndAddCasts( } } } + // Use Python type names (e.g. float16) instead of C++ type names (e.g. half) + // since many Python users will see this message. + const char* type_str = target_dtype_ == DT_HALF ? "float16" : "bfloat16"; LOG(INFO) << "Converted " << num_nodes_changed << "/" << num_nodes_preop - << " nodes to float16 precision using " << num_nonvar_casts_to_fp16 - << " cast(s) to float16 (excluding Const and Variable casts)"; + << " nodes to " << type_str << " precision using " + << num_nonvar_casts_to_f16 << " cast(s) to " << type_str + << " (excluding Const and Variable casts)"; return Status::OK(); } @@ -1902,12 +1950,23 @@ Status AutoMixedPrecision::Optimize(Cluster* cluster, const GrapplerItem& item, return errors::InvalidArgument("cluster == nullptr"); } +#if !defined(INTEL_MKL) || !defined(ENABLE_INTEL_MKL_BFLOAT16) + if (mode_ == AutoMixedPrecisionMode::MKL) { + return errors::Unimplemented( + "The auto_mixed_precision_mkl optimizer cannot be used since " + "this build of TensorFlow is not compiled with MKL support for bfloat16. " + "For information on MKL builds, see: " + "https://software.intel.com/en-us/articles/intel-optimization-for-" + "tensorflow-installation-guide"); + } +#endif + // Start by copying input graph to output. *output = item.graph; int num_gpus = ShouldIgnorePerformance() ? GetNumGPUs(*cluster) : GetNumGPUs(*cluster, kMinGPUArch); - if (num_gpus < 1) { + if (num_gpus < 1 && mode_ == AutoMixedPrecisionMode::CUDA) { // AutoMixedPrecision is currently only tuned for GPU. LOG(WARNING) << "No (suitable) GPUs detected, skipping " << name() << " graph optimizer"; @@ -1916,7 +1975,7 @@ Status AutoMixedPrecision::Optimize(Cluster* cluster, const GrapplerItem& item, // Optimize the output graph in-place. AutoMixedPrecisionImpl optimizer(cluster, item.NodesToPreserve(), output, - item.id); + item.id, mode_); if (item.id == "tf_graph") { LOG(INFO) << "Running " << name() << " graph optimizer"; } else { diff --git a/tensorflow/core/grappler/optimizers/auto_mixed_precision.h b/tensorflow/core/grappler/optimizers/auto_mixed_precision.h index 163d1f6923f..c41ba7d2821 100644 --- a/tensorflow/core/grappler/optimizers/auto_mixed_precision.h +++ b/tensorflow/core/grappler/optimizers/auto_mixed_precision.h @@ -22,16 +22,25 @@ limitations under the License. namespace tensorflow { namespace grappler { -// Convert data types to float16 where appropriate to improve performance on -// GPUs. +enum class AutoMixedPrecisionMode { CUDA, MKL }; + +// Convert data types to float16 or bfloat16 where appropriate to improve +// performance on GPUs or CPUs. class AutoMixedPrecision : public GraphOptimizer { public: + // If 'mode' is CUDA, converts nodes to float16 on Nvidia GPUs. If MKL, + // converts nodes to bfloat16 on CPUs in order to take advantage of MKL + // performance improvements with bfloat16. explicit AutoMixedPrecision( - RewriterConfig::Toggle opt_level = RewriterConfig::ON) {} + AutoMixedPrecisionMode mode = AutoMixedPrecisionMode::CUDA) + : mode_(mode) {} ~AutoMixedPrecision() override {} - string name() const override { return "auto_mixed_precision"; }; + string name() const override { + return mode_ == AutoMixedPrecisionMode::CUDA ? "auto_mixed_precision_cuda" + : "auto_mixed_precision_mkl"; + }; bool UsesFunctionLibrary() const override { return false; } @@ -40,6 +49,9 @@ class AutoMixedPrecision : public GraphOptimizer { void Feedback(Cluster* cluster, const GrapplerItem& item, const GraphDef& optimize_output, double result) override; + + private: + const AutoMixedPrecisionMode mode_; }; } // end namespace grappler diff --git a/tensorflow/core/grappler/optimizers/auto_mixed_precision_lists.h b/tensorflow/core/grappler/optimizers/auto_mixed_precision_lists.h index d3d13e2edc0..c6016548117 100644 --- a/tensorflow/core/grappler/optimizers/auto_mixed_precision_lists.h +++ b/tensorflow/core/grappler/optimizers/auto_mixed_precision_lists.h @@ -23,10 +23,44 @@ limitations under the License. namespace tensorflow { namespace grappler { +// Represents the four lists of ops: the white list, gray list, black list, and +// clear list. These lists determine which ops are converted to fp16/bf16 +// (referred to as 'f16' for short) and which ops stay as fp32. class AutoMixedPrecisionLists { - private: - static void UpdateList(gtl::FlatSet* list, const string& to_add, - const string& to_remove) { + public: + + virtual ~AutoMixedPrecisionLists() {} + + // Returns the set of ops that are considered numerically-safe (for execution + // in f16), performance-critical, and can run in f16. These ops are always + // converted to f16. + virtual gtl::FlatSet WhiteList() = 0; + // Returns the set of ops that can run in f16 and are considered numerically- + // safe (for execution in f16), but which may be made unsafe by an upstream + // blacklist op. + virtual gtl::FlatSet GrayList() = 0; + // Returns the set of ops that are considered numerically-dangerous (i.e., + // unsafe for execution in f16) and whose effects may also be observed in + // downstream nodes (e.g. for f16, in Exp -> Add, the Add is unsafe due to + // the Exp). + virtual gtl::FlatSet BlackList() = 0; + // Returns the set of ops that do not have numerically-significant effects + // (i.e., they are always considered safe for execution in f16 precision), and + // can run in f16. + virtual gtl::FlatSet ClearList() = 0; + + protected: + // Adds or removes ops from list if certain environmental variables are set. + static void UpdateList(const string& list_name, gtl::FlatSet* list) { + CHECK(list_name == "WHITELIST" || list_name == "GRAYLIST" || // Crash OK. + list_name == "BLACKLIST" || list_name == "CLEARLIST"); + string add_env_var = + "TF_AUTO_MIXED_PRECISION_GRAPH_REWRITE_" + list_name + "_ADD"; + string remove_env_var = + "TF_AUTO_MIXED_PRECISION_GRAPH_REWRITE_" + list_name + "_REMOVE"; + string to_add, to_remove; + TF_CHECK_OK(ReadStringFromEnvVar(add_env_var, "", &to_add)); + TF_CHECK_OK(ReadStringFromEnvVar(remove_env_var, "", &to_remove)); for (const auto& x : str_util::Split(to_add, ",")) { list->insert(x); } @@ -35,6 +69,35 @@ class AutoMixedPrecisionLists { } } + // Subclasses should include these on the ClearList. + static void AddTensorListOps(gtl::FlatSet* list) { + // Note: if a data structure op (such as TensorListPopBack) is added here, + // IsTensorListReaderOp or IsTensorListWriterOp may need to be modified + constexpr char* tensor_list_ops[] = { + "TensorListConcat", + "TensorListConcatLists", + "TensorListConcatV2", + "TensorListGather", + "TensorListGetItem", + "TensorListPopBack", + "TensorListPushBack", + "TensorListPushBackBatch", + "TensorListFromTensor", + "TensorListScatter", + "TensorListScatterV2", + "TensorListScatterIntoExistingList", + "TensorListSetItem", + "TensorListSplit", + "TensorListStack" + }; + for (auto op : tensor_list_ops) { + list->insert(op); + } + } +}; + +class AutoMixedPrecisionListsCuda : public AutoMixedPrecisionLists { + private: static bool IsPseudoFastMath() { string optimization_level; TF_CHECK_OK( @@ -45,16 +108,10 @@ class AutoMixedPrecisionLists { } public: - // Returns the set of ops that are considered numerically-safe (for execution - // in fp16) and performance-critical. These ops are always converted to fp16. - static gtl::FlatSet WhiteList(int cuda_version, int cudnn_version) { - string to_add, to_remove; - TF_CHECK_OK(ReadStringFromEnvVar( - "TF_AUTO_MIXED_PRECISION_GRAPH_REWRITE_WHITELIST_ADD", "", &to_add)); - TF_CHECK_OK(ReadStringFromEnvVar( - "TF_AUTO_MIXED_PRECISION_GRAPH_REWRITE_WHITELIST_REMOVE", "", - &to_remove)); + AutoMixedPrecisionListsCuda(int cuda_version, int cudnn_version) + : cuda_version_(cuda_version), cudnn_version_(cudnn_version) {} + gtl::FlatSet WhiteList() override { auto list = gtl::FlatSet{ "BlockLSTM", "BlockLSTMV2", @@ -81,12 +138,12 @@ class AutoMixedPrecisionLists { // "DepthwiseConv2dNativeBackpropInput", "MatMul", }; - if (cuda_version >= 9010) { + if (cuda_version_ >= 9010) { // Fp16 BatchMatMul is slow before CUDA 9.1. list.insert("BatchMatMul"); list.insert("BatchMatMulV2"); } - if (cudnn_version >= 7602) { + if (cudnn_version_ >= 7602) { // Fp16 3D conv is slow before CUDNN 7.6.2. list.insert("Conv3D"); list.insert("Conv3DBackpropFilter"); @@ -94,22 +151,14 @@ class AutoMixedPrecisionLists { list.insert("Conv3DBackpropInput"); list.insert("Conv3DBackpropInputV2"); } - UpdateList(&list, to_add, to_remove); + UpdateList("WHITELIST", &list); return list; } - // Returns the set of ops that are considered numerically-safe (for execution - // in fp16), but which may be made unsafe by an upstream blacklist op. - static gtl::FlatSet GrayList() { + gtl::FlatSet GrayList() override { if (IsPseudoFastMath()) { return gtl::FlatSet{}; } - string to_add, to_remove; - TF_CHECK_OK(ReadStringFromEnvVar( - "TF_AUTO_MIXED_PRECISION_GRAPH_REWRITE_GRAYLIST_ADD", "", &to_add)); - TF_CHECK_OK(ReadStringFromEnvVar( - "TF_AUTO_MIXED_PRECISION_GRAPH_REWRITE_GRAYLIST_REMOVE", "", - &to_remove)); auto list = gtl::FlatSet{ "Add", @@ -156,23 +205,14 @@ class AutoMixedPrecisionLists { "Tanh", "TanhGrad", }; - UpdateList(&list, to_add, to_remove); + UpdateList("GRAYLIST", &list); return list; } - // Returns the set of ops that are considered numerically-dangerous (i.e., - // unsafe for execution in fp16) and whose effects may also be observed in - // downstream nodes (e.g., in Exp -> Add, the Add is unsafe due to the Exp). - static gtl::FlatSet BlackList() { + gtl::FlatSet BlackList() override { if (IsPseudoFastMath()) { return gtl::FlatSet{}; } - string to_add, to_remove; - TF_CHECK_OK(ReadStringFromEnvVar( - "TF_AUTO_MIXED_PRECISION_GRAPH_REWRITE_BLACKLIST_ADD", "", &to_add)); - TF_CHECK_OK(ReadStringFromEnvVar( - "TF_AUTO_MIXED_PRECISION_GRAPH_REWRITE_BLACKLIST_REMOVE", "", - &to_remove)); auto list = gtl::FlatSet{ "Exp", @@ -185,22 +225,14 @@ class AutoMixedPrecisionLists { "SparseSoftmaxCrossEntropyWithLogits", "Sum", }; - UpdateList(&list, to_add, to_remove); + UpdateList("BLACKLIST", &list); return list; } - // Returns the set of ops that do not have numerically-significant effects - // (i.e., they are always considered safe for execution in fp16 precision). - static gtl::FlatSet ClearList() { + gtl::FlatSet ClearList() override { if (IsPseudoFastMath()) { return gtl::FlatSet{}; } - string to_add, to_remove; - TF_CHECK_OK(ReadStringFromEnvVar( - "TF_AUTO_MIXED_PRECISION_GRAPH_REWRITE_CLEARLIST_ADD", "", &to_add)); - TF_CHECK_OK(ReadStringFromEnvVar( - "TF_AUTO_MIXED_PRECISION_GRAPH_REWRITE_CLEARLIST_REMOVE", "", - &to_remove)); auto list = gtl::FlatSet{ "Abs", @@ -291,21 +323,6 @@ class AutoMixedPrecisionLists { "StridedSlice", "StridedSliceGrad", "Switch", - "TensorListConcat", - "TensorListConcatLists", - "TensorListConcatV2", - "TensorListGather", - "TensorListGetItem", - "TensorListPopBack", - "TensorListPushBack", - "TensorListPushBackBatch", - "TensorListFromTensor", - "TensorListScatter", - "TensorListScatterV2", - "TensorListScatterIntoExistingList", - "TensorListSetItem", - "TensorListSplit", - "TensorListStack", "Tile", "TopK", "TopKV2", @@ -313,7 +330,125 @@ class AutoMixedPrecisionLists { "Where", "ZerosLike", }; - UpdateList(&list, to_add, to_remove); + AddTensorListOps(&list); + UpdateList("CLEARLIST", &list); + return list; + } + + private: + int cuda_version_; + int cudnn_version_; +}; + +class AutoMixedPrecisionListsMkl : public AutoMixedPrecisionLists { + private: + + public: + AutoMixedPrecisionListsMkl() {} + + // Only ops which are supported by MKL in bfloat16 should be added to the + // white list, gray list, or clear list. + gtl::FlatSet WhiteList() override { + auto list = gtl::FlatSet{ + "Conv2D", + "Conv2DBackpropFilter", + "Conv2DBackpropInput", + "Conv3D", + "Conv3DBackpropFilterV2", + "Conv3DBackpropInputV2", + "DepthwiseConv2dNative", + "DepthwiseConv2dNativeBackpropFilter", + "DepthwiseConv2dNativeBackpropInput", + "MatMul", + "BatchMatMul", + "BatchMatMulV2" + }; + + UpdateList("WHITELIST", &list); + return list; + } + + gtl::FlatSet GrayList() override { + auto list = gtl::FlatSet{ + "Add", + "AddN", + "AddV2", + "AvgPool", + "AvgPool3D", + "AvgPool3DGrad", + "AvgPoolGrad", + "BiasAdd", + "BiasAddGrad", + "BiasAddV1", + "FusedBatchNormV2", + "FusedBatchNormGradV2", + "FusedBatchNormV3", + "FusedBatchNormGradV3", + "LeakyRelu", + "LeakyReluGrad", + "Mul", + "Sub", + }; + UpdateList("GRAYLIST", &list); + return list; + } + + gtl::FlatSet BlackList() override { + auto list = gtl::FlatSet{ + "Exp", + "Expm1", + "L2Loss", + "Mean", + "Pow", + "SaveV2", + "Softmax", + "SoftmaxCrossEntropyWithLogits", + "SparseSoftmaxCrossEntropyWithLogits", + "Sum", + }; + UpdateList("BLACKLIST", &list); + return list; + } + + gtl::FlatSet ClearList() override { + auto list = gtl::FlatSet{ + "Concat", + "ConcatV2", + "Enter", + "EnsureShape", + "Equal", + "Exit", + "ExpandDims", + "Identity", + "MaxPool", + "MaxPool3D", + "MaxPool3DGrad", + "MaxPoolGrad", + "MaxPoolV2", + "Maximum", + "Merge", + "NextIteration", + "PreventGradient", + "Relu", + "Relu6", + "Relu6Grad", + "ReluGrad", + "Reshape", + "Select", + "SelectV2", + "Shape", + "ShapeN", + "Slice", + "Split", + "SplitV", + "Squeeze", + "StopGradient", + "Switch", + "Transpose", + "ZerosLike", + }; + AddTensorListOps(&list); + UpdateList("CLEARLIST", &list); return list; } }; diff --git a/tensorflow/core/grappler/optimizers/auto_mixed_precision_test.cc b/tensorflow/core/grappler/optimizers/auto_mixed_precision_test.cc index 951279d37cd..248d8dd4266 100644 --- a/tensorflow/core/grappler/optimizers/auto_mixed_precision_test.cc +++ b/tensorflow/core/grappler/optimizers/auto_mixed_precision_test.cc @@ -13,12 +13,6 @@ See the License for the specific language governing permissions and limitations under the License. ==============================================================================*/ -// Currently, this test only passes when TensorFlow passes with CUDA, because -// otherwise the optimizer will not turn clearlist nodes to float16. When -// looking at clearlist nodes, this optimizer checks if the nodes have a float16 -// GPU OpKernel, but without CUDA there are no GPU OpKernels at all. -#if GOOGLE_CUDA || TENSORFLOW_USE_ROCM - #include "tensorflow/core/grappler/optimizers/auto_mixed_precision.h" #include @@ -70,6 +64,31 @@ Tensor GenerateRandomTensorInRange(const TensorShape& shape, double minval, return tensor; } +void VerifyGraphsEquivalent(const GraphDef& original_graph, + const GraphDef& optimized_graph, + const string& func) { + EXPECT_EQ(original_graph.node_size(), optimized_graph.node_size()) << func; + GraphView optimized_view(&optimized_graph); + for (int i = 0; i < original_graph.node_size(); ++i) { + const NodeDef& original = original_graph.node(i); + const NodeDef& optimized = *optimized_view.GetNode(original.name()); + EXPECT_EQ(original.name(), optimized.name()) << func; + EXPECT_EQ(original.op(), optimized.op()) << func; + EXPECT_EQ(original.input_size(), optimized.input_size()) << func; + if (original.input_size() == optimized.input_size()) { + for (int j = 0; j < original.input_size(); ++j) { + EXPECT_EQ(original.input(j), optimized.input(j)) << func; + } + } + } +} + +// Currently, this test suite only passes when TensorFlow passes with CUDA, +// because otherwise the optimizer will not turn clearlist nodes to float16. +// When looking at clearlist nodes, this optimizer checks if the nodes have a +// float16 GPU OpKernel, but without CUDA there are no GPU OpKernels at all. +#if GOOGLE_CUDA || TENSORFLOW_USE_ROCM + const std::pair kMinGPUArch = {7, 0}; class AutoMixedPrecisionTest : public GrapplerTest { @@ -184,25 +203,6 @@ class AutoMixedPrecisionTest : public GrapplerTest { bool gpu_available_; }; -void VerifyGraphsEquivalent(const GraphDef& original_graph, - const GraphDef& optimized_graph, - const string& func) { - EXPECT_EQ(original_graph.node_size(), optimized_graph.node_size()) << func; - GraphView optimized_view(&optimized_graph); - for (int i = 0; i < original_graph.node_size(); ++i) { - const NodeDef& original = original_graph.node(i); - const NodeDef& optimized = *optimized_view.GetNode(original.name()); - EXPECT_EQ(original.name(), optimized.name()) << func; - EXPECT_EQ(original.op(), optimized.op()) << func; - EXPECT_EQ(original.input_size(), optimized.input_size()) << func; - if (original.input_size() == optimized.input_size()) { - for (int j = 0; j < original.input_size(); ++j) { - EXPECT_EQ(original.input(j), optimized.input(j)) << func; - } - } - } -} - TEST_F(AutoMixedPrecisionTest, NoOp) { tensorflow::Scope s = tensorflow::Scope::NewRootScope(); Output input = ops::Const(s.WithOpName("input"), 1.234f, {32}); @@ -1164,8 +1164,188 @@ TEST_F(AutoMixedPrecisionTest, TanhOp) { }); } +#endif // GOOGLE_CUDA || TENSORFLOW_USE_ROCM + +#if INTEL_MKL +#ifdef ENABLE_INTEL_MKL_BFLOAT16 + +class AutoMixedPrecisionMklTest : public GrapplerTest { + protected: + void SetUp() override { + virtual_cluster_.reset(new SingleMachine(/* timeout_s = */ 10, 1, 0)); + TF_CHECK_OK(virtual_cluster_->Provision()); + } + void TearDown() override { TF_CHECK_OK(virtual_cluster_->Shutdown()); } + + std::unique_ptr virtual_cluster_; +}; + +TEST_F(AutoMixedPrecisionMklTest, AlreadyBf16) { + tensorflow::Scope s = tensorflow::Scope::NewRootScope(); + Output input = ops::Const(s.WithOpName("input"), 1.f, {32, 32}); + Output cst1 = ops::Cast(s.WithOpName("cst1"), input, DT_BFLOAT16); + Output wht1 = ops::MatMul(s.WithOpName("wht1"), cst1, cst1); + Output clr1 = ops::Relu(s.WithOpName("clr1"), wht1); + Output cst2 = ops::Cast(s.WithOpName("cst2"), clr1, DT_FLOAT); + Output clr2 = ops::Relu(s.WithOpName("clr2"), cst2); + Output fetch = ops::Identity(s.WithOpName("fetch"), clr2); + + GrapplerItem item; + item.fetch = {"fetch"}; + TF_CHECK_OK(s.ToGraphDef(&item.graph)); + auto tensors_expected = EvaluateNodes(item.graph, item.fetch); + + AutoMixedPrecision optimizer{AutoMixedPrecisionMode::MKL}; + GraphDef output; + TF_ASSERT_OK(optimizer.Optimize(virtual_cluster_.get(), item, &output)); + VLOG(1) << output.DebugString(); + + VerifyGraphsEquivalent(item.graph, output, __FUNCTION__); + GraphView output_view(&output); + EXPECT_EQ(output_view.GetNode("input")->attr().at("dtype").type(), DT_FLOAT); + EXPECT_EQ(output_view.GetNode("cst1")->attr().at("DstT").type(), DT_BFLOAT16); + EXPECT_EQ(output_view.GetNode("wht1")->attr().at("T").type(), DT_BFLOAT16); + EXPECT_EQ(output_view.GetNode("clr1")->attr().at("T").type(), DT_BFLOAT16); + EXPECT_EQ(output_view.GetNode("cst2")->attr().at("SrcT").type(), DT_BFLOAT16); + EXPECT_EQ(output_view.GetNode("cst2")->attr().at("DstT").type(), DT_FLOAT); + EXPECT_EQ(output_view.GetNode("clr2")->attr().at("T").type(), DT_FLOAT); + + auto tensors = EvaluateNodes(output, item.fetch); + EXPECT_EQ(tensors.size(), tensors_expected.size()); + EXPECT_EQ(tensors.size(), item.fetch.size()); + for (int i = 0; i < item.fetch.size(); ++i) { + test::ExpectTensorNear(tensors_expected[i], tensors[i], 1e-6); + } +} + +TEST_F(AutoMixedPrecisionMklTest, Simple) { + tensorflow::Scope s = tensorflow::Scope::NewRootScope(); + Output input = ops::Const(s.WithOpName("input"), 1.f / 32, {32, 32}); + Output blk1 = ops::Exp(s.WithOpName("blk1"), input); + Output clr1 = ops::Relu(s.WithOpName("clr1"), blk1); + Output gry1 = ops::Sqrt(s.WithOpName("gry1"), clr1); + Output clr2 = ops::Relu(s.WithOpName("clr2"), gry1); + Output wht1 = ops::MatMul(s.WithOpName("wht1"), clr2, clr2); + Output clr3 = ops::Relu(s.WithOpName("clr3"), wht1); + Output blk2 = ops::Log(s.WithOpName("blk2"), clr3); + Output clr4 = ops::Relu(s.WithOpName("clr4"), blk2); + Output blk3 = ops::SparseMatMul(s.WithOpName("blk3"), clr4, clr4); + Output clr5 = ops::Relu(s.WithOpName("clr5"), blk3); + Output fetch = ops::Identity(s.WithOpName("fetch"), clr5); + + GrapplerItem item; + item.fetch = {"fetch"}; + TF_CHECK_OK(s.ToGraphDef(&item.graph)); + auto tensors_expected = EvaluateNodes(item.graph, item.fetch); + + AutoMixedPrecision optimizer{AutoMixedPrecisionMode::MKL}; + GraphDef output; + TF_ASSERT_OK(optimizer.Optimize(virtual_cluster_.get(), item, &output)); + + VLOG(1) << output.DebugString(); + + GraphView output_view(&output); + EXPECT_EQ(output.node_size(), item.graph.node_size() + 2); + EXPECT_EQ(output_view.GetNode("input")->attr().at("dtype").type(), DT_FLOAT); + EXPECT_EQ(output_view.GetNode("blk1")->attr().at("T").type(), DT_FLOAT); + EXPECT_EQ(output_view.GetNode("clr1")->attr().at("T").type(), DT_FLOAT); + EXPECT_EQ(output_view.GetNode("gry1")->attr().at("T").type(), DT_FLOAT); + EXPECT_EQ(output_view.GetNode("clr2")->attr().at("T").type(), DT_BFLOAT16); + EXPECT_EQ(output_view.GetNode("wht1")->attr().at("T").type(), DT_BFLOAT16); + EXPECT_EQ(output_view.GetNode("clr3")->attr().at("T").type(), DT_BFLOAT16); + EXPECT_EQ(output_view.GetNode("blk2")->attr().at("T").type(), DT_FLOAT); + EXPECT_EQ(output_view.GetNode("clr4")->attr().at("T").type(), DT_FLOAT); + EXPECT_EQ(output_view.GetNode("blk3")->attr().at("Ta").type(), DT_FLOAT); + EXPECT_EQ(output_view.GetNode("blk3")->attr().at("Tb").type(), DT_FLOAT); + EXPECT_EQ(output_view.GetNode("clr5")->attr().at("T").type(), DT_FLOAT); + + auto tensors = EvaluateNodes(output, item.fetch); + EXPECT_EQ(tensors.size(), tensors_expected.size()); + EXPECT_EQ(tensors.size(), item.fetch.size()); + for (int i = 0; i < item.fetch.size(); ++i) { + test::ExpectClose(tensors_expected[i], tensors[i], -1, 5e-4); + } +} + +TEST_F(AutoMixedPrecisionMklTest, TensorListSetGet) { + tensorflow::Scope s = tensorflow::Scope::NewRootScope(); + tensorflow::Input shape = {32, 32}; + auto tl1 = ops::TensorListReserve(s.WithOpName("tl1"), {32, 32}, 8, DT_FLOAT); + Output input = ops::Const(s.WithOpName("input"), 1.f / 32, {32, 32}); + Output idx1 = ops::Const(s.WithOpName("idx1"), 1); + Output idx2 = ops::Const(s.WithOpName("idx2"), 2); + Output idx3 = ops::Const(s.WithOpName("idx3"), 3); + auto tl1w1 = + ops::TensorListSetItem(s.WithOpName("tl1w1"), tl1.handle, idx1, input); + Output wht1 = ops::MatMul(s.WithOpName("wht1"), input, input); + auto tl1w2 = + ops::TensorListSetItem(s.WithOpName("tl1w2"), tl1.handle, idx2, wht1); + // Ensure that TensorListResize doesn't cause any problems. + Output tl1rs = + ops::TensorListResize(s.WithOpName("tl1rs"), tl1w2.output_handle, 6); + Output tl1r1 = ops::TensorListGetItem(s.WithOpName("tl1r1"), tl1rs, idx2, + shape, DT_FLOAT) + .item; + Output gry1 = ops::Mul(s.WithOpName("gry1"), tl1r1, tl1r1); + Output wht2 = ops::MatMul(s.WithOpName("wht2"), gry1, gry1); + auto tl1w3 = + ops::TensorListSetItem(s.WithOpName("tl1w3"), tl1.handle, idx3, wht2); + Output tl1r2 = + ops::TensorListGetItem(s.WithOpName("tl1r2"), tl1w3.output_handle, idx3, + shape, DT_FLOAT) + .item; + auto tl2 = ops::TensorListReserve(s.WithOpName("tl2"), shape, 8, DT_FLOAT); + auto tl2w1 = + ops::TensorListSetItem(s.WithOpName("tl2w1"), tl2.handle, idx1, input); + Output tl2r1 = + ops::TensorListGetItem(s.WithOpName("tl2r1"), tl2w1.output_handle, idx1, + shape, DT_FLOAT) + .item; + Output fetch1 = ops::Identity(s.WithOpName("fetch1"), tl1r2); + Output fetch2 = ops::Identity(s.WithOpName("fetch2"), tl2r1); + + GrapplerItem item; + item.fetch = {"fetch1", "fetch2"}; + TF_CHECK_OK(s.ToGraphDef(&item.graph)); + auto tensors_expected = EvaluateNodes(item.graph, item.fetch); + + AutoMixedPrecision optimizer{AutoMixedPrecisionMode::MKL}; + GraphDef output; + TF_ASSERT_OK(optimizer.Optimize(virtual_cluster_.get(), item, &output)); + + VLOG(1) << output.DebugString(); + + GraphView output_view(&output); + EXPECT_EQ(output.node_size(), item.graph.node_size() + 2); + const char* type_key = "element_dtype"; + EXPECT_EQ(output_view.GetNode("tl1")->attr().at(type_key).type(), + DT_BFLOAT16); + EXPECT_EQ(output_view.GetNode("tl1w1")->attr().at(type_key).type(), + DT_BFLOAT16); + EXPECT_EQ(output_view.GetNode("wht1")->attr().at("T").type(), DT_BFLOAT16); + EXPECT_EQ(output_view.GetNode("tl1w2")->attr().at(type_key).type(), + DT_BFLOAT16); + EXPECT_EQ(output_view.GetNode("tl1r1")->attr().at(type_key).type(), + DT_BFLOAT16); + EXPECT_EQ(output_view.GetNode("gry1")->attr().at("T").type(), DT_BFLOAT16); + EXPECT_EQ(output_view.GetNode("wht2")->attr().at("T").type(), DT_BFLOAT16); + EXPECT_EQ(output_view.GetNode("tl1w3")->attr().at(type_key).type(), + DT_BFLOAT16); + EXPECT_EQ(output_view.GetNode("tl2")->attr().at(type_key).type(), DT_FLOAT); + EXPECT_EQ(output_view.GetNode("tl2w1")->attr().at(type_key).type(), DT_FLOAT); + EXPECT_EQ(output_view.GetNode("tl2r1")->attr().at(type_key).type(), DT_FLOAT); + + auto tensors = EvaluateNodes(output, item.fetch); + EXPECT_EQ(tensors.size(), tensors_expected.size()); + EXPECT_EQ(tensors.size(), item.fetch.size()); + for (int i = 0; i < item.fetch.size(); ++i) { + test::ExpectClose(tensors_expected[i], tensors[i], -1, 1e-2); + } +} + +#endif // ENABLE_INTEL_MKL_BFLOAT16 +#endif // INTEL_MKL + } // namespace } // namespace grappler } // namespace tensorflow - -#endif // GOOGLE_CUDA || TENSORFLOW_USE_ROCM diff --git a/tensorflow/core/grappler/optimizers/meta_optimizer.cc b/tensorflow/core/grappler/optimizers/meta_optimizer.cc index cd0d44e8e12..2f1c869965d 100644 --- a/tensorflow/core/grappler/optimizers/meta_optimizer.cc +++ b/tensorflow/core/grappler/optimizers/meta_optimizer.cc @@ -188,7 +188,9 @@ std::unique_ptr MetaOptimizer::MakeNewOptimizer( MK_OPT("remap", new Remapper(cfg_.remapping())); MK_OPT("layout", new GenericLayoutOptimizer()); MK_OPT("auto_mixed_precision", - new AutoMixedPrecision(cfg_.auto_mixed_precision())); + new AutoMixedPrecision(AutoMixedPrecisionMode::CUDA)); + MK_OPT("auto_mixed_precision_mkl", + new AutoMixedPrecision(AutoMixedPrecisionMode::MKL)); MK_OPT("memory", new MemoryOptimizer(RewriterConfig::MANUAL)); MK_OPT("common_subgraph_elimination", new CommonSubgraphElimination(cfg_.common_subgraph_elimination())); @@ -249,7 +251,11 @@ Status MetaOptimizer::InitializeOptimizers( } if (AutoMixedPrecisionEnabled(cfg_.auto_mixed_precision())) { optimizers->push_back( - MakeUnique(cfg_.auto_mixed_precision())); + MakeUnique(AutoMixedPrecisionMode::CUDA)); + } + if (AutoMixedPrecisionEnabled(cfg_.auto_mixed_precision_mkl())) { + optimizers->push_back( + MakeUnique(AutoMixedPrecisionMode::MKL)); } if (cfg_.pin_to_host_optimization() == RewriterConfig::ON) { optimizers->push_back(MakeUnique()); @@ -835,6 +841,7 @@ bool MetaOptimizerEnabled(const ConfigProto& cfg) { rewrite_cfg.scoped_allocator_optimization() == RewriterConfig::ON || rewrite_cfg.pin_to_host_optimization() == RewriterConfig::ON || AutoMixedPrecisionEnabled(rewrite_cfg.auto_mixed_precision()) || + AutoMixedPrecisionEnabled(rewrite_cfg.auto_mixed_precision_mkl()) || !rewrite_cfg.optimizers().empty() || !rewrite_cfg.custom_optimizers().empty(); } diff --git a/tensorflow/core/protobuf/rewriter_config.proto b/tensorflow/core/protobuf/rewriter_config.proto index 38c3ad7ae57..9520db92742 100644 --- a/tensorflow/core/protobuf/rewriter_config.proto +++ b/tensorflow/core/protobuf/rewriter_config.proto @@ -85,11 +85,15 @@ message RewriterConfig { // Enable the swap of kernel implementations based on the device placement // (default is ON). Toggle implementation_selector = 22; - // Optimize data types (default is OFF). - // e.g., This will try to use float16 on GPU which is faster. + // Optimize data types for CUDA (default is OFF). + // This will try to use float16 on GPU which is faster. // Note that this can change the numerical stability of the graph and may // require the use of loss scaling to maintain model convergence. Toggle auto_mixed_precision = 23; + // Optimize data types for MKL (default is OFF). + // This will try to use bfloat16 on CPUs, which is faster. + // Note that this can change the numerical stability of the graph. + Toggle auto_mixed_precision_mkl = 25; // Disable the entire meta optimizer (off by default). bool disable_meta_optimizer = 19; diff --git a/tensorflow/python/grappler/auto_mixed_precision_test.py b/tensorflow/python/grappler/auto_mixed_precision_test.py index 494f6fc78fc..0de12b9eca8 100644 --- a/tensorflow/python/grappler/auto_mixed_precision_test.py +++ b/tensorflow/python/grappler/auto_mixed_precision_test.py @@ -19,8 +19,8 @@ from __future__ import division from __future__ import print_function import os -import unittest +from absl.testing import parameterized import numpy as np from tensorflow.core.framework import types_pb2 @@ -209,7 +209,7 @@ def _make_node_with_color(color, input_tensor, name=None): if color == 'c': # Clear node return nn.relu(input_tensor, name=name) if color == 'b': # Black node - return math_ops.sqrt(math_ops.pow(input_tensor, 2.), name=name) + return math_ops.pow(math_ops.pow(input_tensor, 2.), 0.5, name=name) raise ValueError('Invalid node color: ' + str(color)) @@ -231,18 +231,21 @@ def _build_simple_loop_graph(inp_colors, body_colors, out_colors): return a -def _get_config(auto_mixed_precision=True): +def _get_config(auto_mixed_precision_mode): """Returns a ConfigProto with auto mixed precision enabled if appropriate.""" - if auto_mixed_precision: - rewrite_config = rewriter_config_pb2.RewriterConfig( - auto_mixed_precision=rewriter_config_pb2.RewriterConfig.ON, - # do not remove duplicated nodes - arithmetic_optimization=rewriter_config_pb2.RewriterConfig.OFF) + rewrite_config = rewriter_config_pb2.RewriterConfig( + # do not remove duplicated nodes + arithmetic_optimization=rewriter_config_pb2.RewriterConfig.OFF, + # do not turn Conv2D and other nodes into _FusedConv2D + remapping=rewriter_config_pb2.RewriterConfig.OFF, + ) + if auto_mixed_precision_mode == 'cuda': + rewrite_config.auto_mixed_precision = rewriter_config_pb2.RewriterConfig.ON + elif auto_mixed_precision_mode == 'mkl': + rewrite_config.auto_mixed_precision_mkl = ( + rewriter_config_pb2.RewriterConfig.ON) else: - rewrite_config = rewriter_config_pb2.RewriterConfig( - auto_mixed_precision=rewriter_config_pb2.RewriterConfig.OFF, - # do not remove duplicated nodes - arithmetic_optimization=rewriter_config_pb2.RewriterConfig.OFF) + assert auto_mixed_precision_mode is None rewrite_config.min_graph_nodes = -1 graph_options = config_pb2.GraphOptions( rewrite_options=rewrite_config, build_cost_model=1) @@ -255,19 +258,33 @@ def _is_cast_to_fp16(node_name): return node_name.endswith('-CastToFp16-AutoMixedPrecision') +def _is_cast_to_bf16(node_name): + return node_name.endswith('-CastToBf16-AutoMixedPrecision') + + def _is_cast_to_fp32(node_name): return node_name.endswith('-CastToFp32-AutoMixedPrecision') -def _count_casts(nodes): +def _count_casts(mode, nodes): + """Counts the number of casts to f16 and fp32.""" num_to_fp16 = 0 + num_to_bf16 = 0 num_to_fp32 = 0 for node in nodes: if _is_cast_to_fp16(node.name): num_to_fp16 += 1 + if _is_cast_to_bf16(node.name): + num_to_bf16 += 1 elif _is_cast_to_fp32(node.name): num_to_fp32 += 1 - return num_to_fp16, num_to_fp32 + if mode == 'cuda': + assert num_to_bf16 == 0 + return num_to_fp16, num_to_fp32 + else: + assert mode == 'mkl' + assert num_to_fp16 == 0 + return num_to_bf16, num_to_fp32 def _build_node_map(nodes): @@ -303,7 +320,7 @@ def _example_noninlined_funcdef(features): return features * math_ops.sigmoid(features) -class AutoMixedPrecisionTest(test.TestCase): +class AutoMixedPrecisionTest(test.TestCase, parameterized.TestCase): """Tests the Grappler auto mixed precision optimizer.""" IGNORE_PERF_VAR = 'TF_AUTO_MIXED_PRECISION_GRAPH_REWRITE_IGNORE_PERFORMANCE' @@ -311,8 +328,8 @@ class AutoMixedPrecisionTest(test.TestCase): def setUp(self): super(AutoMixedPrecisionTest, self).setUp() - # Enable the tests to be run on pre-Volta GPUs by telling the grappler pass - # to ignore performance and always transform the graph. + # Enable the CUDA tests to be run on pre-Volta GPUs by telling the grappler + # pass to ignore performance and always transform the graph. self._original_ignore_perf_value = os.getenv(self.IGNORE_PERF_VAR) os.environ[self.IGNORE_PERF_VAR] = '1' @@ -323,24 +340,33 @@ class AutoMixedPrecisionTest(test.TestCase): del os.environ[self.IGNORE_PERF_VAR] super(AutoMixedPrecisionTest, self).tearDown() - def _assert_output_fp16(self, node_map, node_name, output_port=0): - self.assertEqual(node_map[node_name].output_info[output_port].dtype, - types_pb2.DT_HALF) + def _lower_precision_dtype(self, mode): + return dtypes.float16 if mode == 'cuda' else dtypes.bfloat16 - def _run(self, fetches): + def _assert_output_f16(self, mode, node_map, node_name, output_port=0): + self.assertEqual(node_map[node_name].output_info[output_port].dtype, + self._lower_precision_dtype(mode).as_datatype_enum) + + def _run(self, mode, fetches): """Runs the graph and returns the evaluation of the fetches.""" - with session.Session(config=_get_config(False)) as sess: + with session.Session(config=_get_config(None)) as sess: sess.run(variables.global_variables_initializer()) output_val_ref = self.evaluate(fetches) - with session.Session(config=_get_config()) as sess: + with session.Session(config=_get_config(mode)) as sess: sess.run(variables.global_variables_initializer()) metadata = config_pb2.RunMetadata() output_val = sess.run(fetches, run_metadata=metadata) return output_val_ref, output_val, metadata.cost_graph - def _run_simple_loop_test(self, inp, body, out): + def _maybe_skip(self, mode): + if mode == 'cuda' and not test.is_gpu_available(cuda_only=True): + self.skipTest('No GPU is available') + if mode == 'mkl' and not test_util.IsMklEnabled(): + self.skipTest('MKL is not enabled') + + def _run_simple_loop_test(self, mode, inp, body, out): """Runs a test of a simple loop. The loop has different node colors in different sections of the graph. The @@ -359,398 +385,441 @@ class AutoMixedPrecisionTest(test.TestCase): out: A string of letters indicating the colors and expected dtypes of the output nodes. """ - if test.is_gpu_available(cuda_only=True): - random_seed.set_random_seed(0) - expected_types = [] - for section in [inp, body, out]: - section_expected_types = [] - for color in section: - if color.isupper(): - expected_type = types_pb2.DT_HALF - else: - expected_type = types_pb2.DT_FLOAT - section_expected_types.append(expected_type) - expected_types.append(section_expected_types) + self._maybe_skip(mode) + random_seed.set_random_seed(0) + expected_types = [] + for section in [inp, body, out]: + section_expected_types = [] + for color in section: + if color.isupper(): + expected_type = self._lower_precision_dtype(mode).as_datatype_enum + else: + expected_type = types_pb2.DT_FLOAT + section_expected_types.append(expected_type) + expected_types.append(section_expected_types) - a = _build_simple_loop_graph(inp, body, out) - output_val_ref, output_val, cost_graph = self._run(a) - node_map = _build_node_map(cost_graph.node) + a = _build_simple_loop_graph(inp, body, out) + output_val_ref, output_val, cost_graph = self._run(mode, a) + node_map = _build_node_map(cost_graph.node) - section_names = ['input', 'while/body', 'output'] - all_types_correct = True - for section_name, expected_types in zip(section_names, expected_types): - for i, expected_type in enumerate(expected_types): - node_name = section_name + '_%i' % i - output_port = 0 - optimized_type = node_map[node_name].output_info[output_port].dtype - if optimized_type != expected_type: - print('Expected node %s to have type %s but got type %s' % - (node_name, expected_type, optimized_type)) - all_types_correct = False - self.assertTrue(all_types_correct) + section_names = ['input', 'while/body', 'output'] + all_types_correct = True + for section_name, expected_types in zip(section_names, expected_types): + for i, expected_type in enumerate(expected_types): + node_name = section_name + '_%i' % i + output_port = 0 + optimized_type = node_map[node_name].output_info[output_port].dtype + if optimized_type != expected_type: + print('Expected node %s to have type %s but got type %s' % + (node_name, expected_type, optimized_type)) + all_types_correct = False + self.assertTrue(all_types_correct) + if mode == 'mkl': + self.assertAllClose(output_val_ref, output_val, atol=2e-2, rtol=2e-2) + else: self.assertAllClose(output_val_ref, output_val, atol=2e-3, rtol=1e-3) + @parameterized.parameters(['cuda', 'mkl']) @test_util.run_deprecated_v1 @test_util.disable_xla('This test does not pass with XLA') - def test_conv_bn(self): + def test_conv_bn(self, mode): """Test graph with convolution followed by batch norm.""" - if test.is_gpu_available(cuda_only=True): - random_seed.set_random_seed(0) - x = _input([2, 8, 8, 1]) - x = _conv_bn(x) - output = _conv_bn(x) + self._maybe_skip(mode) + random_seed.set_random_seed(0) + x = _input([2, 8, 8, 1]) + x = _conv_bn(x) + output = _conv_bn(x) - output_val_ref, output_val, cost_graph = self._run(output) - node_map = _build_node_map(cost_graph.node) - num_to_fp16, num_to_fp32 = _count_casts(cost_graph.node) + output_val_ref, output_val, cost_graph = self._run(mode, output) + node_map = _build_node_map(cost_graph.node) + num_to_f16, num_to_fp32 = _count_casts(mode, cost_graph.node) - self._assert_output_fp16(node_map, 'Conv2D') - self._assert_output_fp16(node_map, 'FusedBatchNormV3') - self._assert_output_fp16(node_map, 'Conv2D_1') - self.assertEqual(num_to_fp16, - 3) # Before Conv2D:0, Conv2D:1, Conv2D_1:1 - self.assertEqual(num_to_fp32, 1) # After FusedBatchNormV3:0 + self._assert_output_f16(mode, node_map, 'Conv2D') + self._assert_output_f16(mode, node_map, 'FusedBatchNormV3') + self._assert_output_f16(mode, node_map, 'Conv2D_1') + self.assertEqual(num_to_f16, 3) # Before Conv2D:0, Conv2D:1, Conv2D_1:1 + self.assertEqual(num_to_fp32, 1) # After FusedBatchNormV3:0 + if mode == 'mkl': + tol = 1e-2 + elif test.is_built_with_rocm(): # Bump up the tolerance for the ROCm platform # The default tolerance (1e-3) results in a tiny fraction (<1%) of # miscompares on ROCm platform, and hence the tolerance bump - tol = 2e-3 if test.is_built_with_rocm else 1e-3 - self.assertAllClose(output_val_ref, output_val, atol=tol, rtol=tol) + tol = 2e-3 + else: + tol = 1e-3 + self.assertAllClose(output_val_ref, output_val, atol=tol, rtol=tol) - # TODO: enable these tests when cuDNN is upgraded to >= 7.6.2. Same with the - # test_conv3d() below. - @unittest.skip('Test case should be skipped when cuDNN < 7.6.2') + @parameterized.parameters(['cuda', 'mkl']) @test_util.run_deprecated_v1 @test_util.disable_xla('This test does not pass with XLA') - def test_conv3d_bn(self): + def test_conv3d_bn(self, mode): """Test graph with convolution followed by batch norm.""" - if test.is_gpu_available(cuda_only=True): - random_seed.set_random_seed(0) - x = _input([2, 8, 8, 8, 1]) - x = _conv3d_bn(x) - output = _conv3d_bn(x) + self._maybe_skip(mode) + if mode == 'cuda': + # TODO: enable these tests when cuDNN is upgraded to >= 7.6.2. + self.skipTest('Test case should be skipped when cuDNN < 7.6.2') + random_seed.set_random_seed(0) + x = _input([2, 8, 8, 8, 1]) + x = _conv3d_bn(x) + output = _conv3d_bn(x) - output_val_ref, output_val, cost_graph = self._run(output) - node_map = _build_node_map(cost_graph.node) - num_to_fp16, num_to_fp32 = _count_casts(cost_graph.node) + output_val_ref, output_val, cost_graph = self._run(mode, output) + node_map = _build_node_map(cost_graph.node) + num_to_fp16, num_to_fp32 = _count_casts(mode, cost_graph.node) - self._assert_output_fp16(node_map, 'Conv3D') - self._assert_output_fp16(node_map, 'FusedBatchNormV3') - self._assert_output_fp16(node_map, 'Conv3D_1') - self.assertEqual(num_to_fp16, 3) # Before Conv3D:0, Conv3D:1, Conv3D_1:1 - self.assertEqual(num_to_fp32, 1) # After FusedBatchNormV3:0 - self.assertAllClose(output_val_ref, output_val, atol=1e-2, rtol=1e-2) + self._assert_output_f16(mode, node_map, 'Conv3D') + self._assert_output_f16(mode, node_map, 'FusedBatchNormV3') + self._assert_output_f16(mode, node_map, 'Conv3D_1') + self.assertEqual(num_to_fp16, 3) # Before Conv3D:0, Conv3D:1, Conv3D_1:1 + self.assertEqual(num_to_fp32, 1) # After FusedBatchNormV3:0 + self.assertAllClose(output_val_ref, output_val, atol=1e-2, rtol=1e-2) - @unittest.skip('Test case should be skipped when cuDNN < 7.6.2') + @parameterized.parameters(['cuda', 'mkl']) @test_util.run_deprecated_v1 @test_util.disable_xla('This test does not pass with XLA') - def test_conv3d(self): + def test_conv3d(self, mode): """Test grad ops with convolution3d graph.""" - if test.is_gpu_available(cuda_only=True): - random_seed.set_random_seed(0) - x = _input([2, 8, 8, 8, 1]) - f = _weight([3, 3, 3, 1, 6]) - y = _conv3d(x, f) - y = array_ops.identity(y) - optimizer = gradient_descent.GradientDescentOptimizer(learning_rate=0.01) - g = optimizer.compute_gradients(y, [x, f]) - output = (y, g) + self._maybe_skip(mode) + if mode == 'cuda': + # TODO: enable these tests when cuDNN is upgraded to >= 7.6.2. + self.skipTest('Test case should be skipped when cuDNN < 7.6.2') + random_seed.set_random_seed(0) + x = _input([2, 8, 8, 8, 1]) + f = _weight([3, 3, 3, 1, 6]) + y = _conv3d(x, f) + y = array_ops.identity(y) + optimizer = gradient_descent.GradientDescentOptimizer(learning_rate=0.01) + g = optimizer.compute_gradients(y, [x, f]) + output = (y, g) - output_val_ref, output_val, cost_graph = self._run(output) - node_map = _build_node_map(cost_graph.node) - self._assert_output_fp16(node_map, 'Conv3D') - self._assert_output_fp16(node_map, - 'gradients/Conv3D_grad/Conv3DBackpropInputV2') - self._assert_output_fp16(node_map, - 'gradients/Conv3D_grad/Conv3DBackpropFilterV2') + output_val_ref, output_val, cost_graph = self._run(mode, output) + node_map = _build_node_map(cost_graph.node) + self._assert_output_f16(mode, node_map, 'Conv3D') + self._assert_output_f16(mode, node_map, + 'gradients/Conv3D_grad/Conv3DBackpropInputV2') + self._assert_output_f16(mode, node_map, + 'gradients/Conv3D_grad/Conv3DBackpropFilterV2') - output_val_ref, output_val, cost_graph = self._run(output) - self.assertAllClose(output_val_ref, output_val, atol=1e-3, rtol=1e-3) + output_val_ref, output_val, cost_graph = self._run(mode, output) + tol = 5e-2 if mode == 'mkl' else 1e-3 + self.assertAllClose(output_val_ref, output_val, atol=tol, rtol=tol) + # TODO(reedwm): Fix and enable this test with MKL. Currently this crashes with + # MKL + @parameterized.parameters(['cuda']) @test_util.run_deprecated_v1 @test_util.disable_xla('This test does not pass with XLA') - def test_conv_bn_dropout(self): + def test_conv_bn_dropout(self, mode): """Test dropout precision of convolution batch norm graph.""" - if test.is_gpu_available(cuda_only=True): - random_seed.set_random_seed(0) - x = _input([2, 8, 8, 1]) - y = _conv_bn(x) - y = nn.dropout(y, rate=0.5) - y = math_ops.add(y, 1, name='addition') - y = _conv_bn(y) - y = array_ops.identity(y) - optimizer = gradient_descent.GradientDescentOptimizer( - learning_rate=0.01) - g = optimizer.compute_gradients(y, [x]) - output = (y, g) + self._maybe_skip(mode) + random_seed.set_random_seed(0) + x = _input([2, 8, 8, 1]) + y = _conv_bn(x) + y = nn.dropout(y, rate=0.5) + y = math_ops.add(y, 1, name='addition') + y = _conv_bn(y) + y = array_ops.identity(y) + optimizer = gradient_descent.GradientDescentOptimizer( + learning_rate=0.01) + g = optimizer.compute_gradients(y, [x]) + output = (y, g) - output_val_ref, output_val, cost_graph = self._run(output) - node_map = _build_node_map(cost_graph.node) - self._assert_output_fp16(node_map, 'Conv2D') - self._assert_output_fp16(node_map, 'FusedBatchNormV3') - # We do not assert dropout's dtype because we do not want to rely on the - # node names of dropout's internal implementation. - self._assert_output_fp16(node_map, 'addition') - self._assert_output_fp16(node_map, 'Conv2D_1') + output_val_ref, output_val, cost_graph = self._run(mode, output) + node_map = _build_node_map(cost_graph.node) + self._assert_output_f16(mode, node_map, 'Conv2D') + self._assert_output_f16(mode, node_map, 'FusedBatchNormV3') + # We do not assert dropout's dtype because we do not want to rely on the + # node names of dropout's internal implementation. + self._assert_output_f16(mode, node_map, 'addition') + self._assert_output_f16(mode, node_map, 'Conv2D_1') - output_val_ref, output_val, cost_graph = self._run(output) - # Bump up the tolerance for the ROCm platform - # The default tolerance (1e-3) results in a tiny fraction (<1%) of - # miscompares on ROCm platform, and hence the tolerance bump - tol = 2e-3 if test.is_built_with_rocm else 1e-3 - self.assertAllClose(output_val_ref, output_val, atol=tol, rtol=tol) + output_val_ref, output_val, cost_graph = self._run(mode, output) + # Bump up the tolerance for the ROCm platform + # The default tolerance (1e-3) results in a tiny fraction (<1%) of + # miscompares on ROCm platform, and hence the tolerance bump + tol = 2e-3 if test.is_built_with_rocm else 1e-3 + self.assertAllClose(output_val_ref, output_val, atol=tol, rtol=tol) + # TODO(reedwm): Fix and enable this test with MKL. Currently this crashes with + # MKL + @parameterized.parameters(['cuda']) @test_util.run_deprecated_v1 @test_util.disable_xla('This test does not pass with XLA') - def test_conv_pool(self): + def test_conv_pool(self, mode): """Test graph with convolution followed by pooling.""" - if test.is_gpu_available(cuda_only=True): - random_seed.set_random_seed(0) - x = _input([2, 8, 8, 1]) - output = _conv_pool(x) + self._maybe_skip(mode) + random_seed.set_random_seed(0) + x = _input([2, 8, 8, 1]) + output = _conv_pool(x) - output_val_ref, output_val, cost_graph = self._run(output) - node_map = _build_node_map(cost_graph.node) - num_to_fp16, num_to_fp32 = _count_casts(cost_graph.node) + output_val_ref, output_val, cost_graph = self._run(mode, output) + node_map = _build_node_map(cost_graph.node) + num_to_f16, num_to_fp32 = _count_casts(mode, cost_graph.node) - self._assert_output_fp16(node_map, 'Conv2D') - self._assert_output_fp16(node_map, 'Relu') - self._assert_output_fp16(node_map, 'MaxPool') - self._assert_output_fp16(node_map, 'Conv2D_1') - self.assertEqual(num_to_fp16, 4) - self.assertEqual(num_to_fp32, 1) - self.assertAllClose(output_val_ref, output_val, atol=1e-3, rtol=1e-3) + self._assert_output_f16(mode, node_map, 'Conv2D') + self._assert_output_f16(mode, node_map, 'Relu') + self._assert_output_f16(mode, node_map, 'MaxPool') + self._assert_output_f16(mode, node_map, 'Conv2D_1') + self.assertEqual(num_to_f16, 4) + self.assertEqual(num_to_fp32, 1) + tol = 5e-3 if mode == 'mkl' else 1e-3 + self.assertAllClose(output_val_ref, output_val, atol=tol, rtol=tol) + @parameterized.parameters(['cuda', 'mkl']) @test_util.run_v1_only('b/138749235') @test_util.disable_xla('This test does not pass with XLA') - def test_simple_loop(self): + def test_simple_loop(self, mode): """Test graph with while loop.""" - if test.is_gpu_available(cuda_only=True): - random_seed.set_random_seed(0) - x = _input([8, 8]) - y = _simple_loop(x, _matmul_act)[1] - optimizer = gradient_descent.GradientDescentOptimizer(learning_rate=0.01) - g = optimizer.compute_gradients(y, [x]) - output = (y, g) + self._maybe_skip(mode) + random_seed.set_random_seed(0) + x = _input([8, 8]) + y = _simple_loop(x, _matmul_act)[1] + optimizer = gradient_descent.GradientDescentOptimizer(learning_rate=0.01) + g = optimizer.compute_gradients(y, [x]) + output = (y, g) - output_val_ref, output_val, cost_graph = self._run(output) - node_map = _build_node_map(cost_graph.node) + output_val_ref, output_val, cost_graph = self._run(mode, output) + node_map = _build_node_map(cost_graph.node) - self._assert_output_fp16(node_map, 'while/MatMul') - self._assert_output_fp16(node_map, 'while/Relu') - self.assertAllClose(output_val_ref, output_val, atol=1e-3, rtol=1e-3) + self._assert_output_f16(mode, node_map, 'while/MatMul') + self._assert_output_f16(mode, node_map, 'while/Relu') + tol = 1e-2 if mode == 'mkl' else 1e-3 + self.assertAllClose(output_val_ref, output_val, atol=tol, rtol=tol) + @parameterized.parameters(['cuda', 'mkl']) @test_util.run_v1_only('b/138749235') @test_util.disable_xla('This test does not pass with XLA') - def test_loop_with_vars_intertwined(self): + def test_loop_with_vars_intertwined(self, mode): """Test graph with intertwined while loops.""" - if test.is_gpu_available(cuda_only=True): - random_seed.set_random_seed(0) - x = _input([8, 8]) - _, _, k, l = _loop_vars_intertwined( - array_ops.ones(array_ops.shape(x)), x, _matmul_act, _matmul_act) - optimizer = gradient_descent.GradientDescentOptimizer(learning_rate=0.01) - g = optimizer.compute_gradients(k, [x]) - output = (k, l, g) + self._maybe_skip(mode) + random_seed.set_random_seed(0) + x = _input([8, 8]) + _, _, k, l = _loop_vars_intertwined( + array_ops.ones(array_ops.shape(x)), x, _matmul_act, _matmul_act) + optimizer = gradient_descent.GradientDescentOptimizer(learning_rate=0.01) + g = optimizer.compute_gradients(k, [x]) + output = (k, l, g) - output_val_ref, output_val, cost_graph = self._run(output) - node_map = _build_node_map(cost_graph.node) + output_val_ref, output_val, cost_graph = self._run(mode, output) + node_map = _build_node_map(cost_graph.node) - self._assert_output_fp16(node_map, 'while/MatMul') - self._assert_output_fp16(node_map, 'while/Relu') - self._assert_output_fp16(node_map, 'while/MatMul_1') - self._assert_output_fp16(node_map, 'while/Relu_1') - self.assertAllClose(output_val_ref, output_val, atol=1e-3, rtol=1e-3) + self._assert_output_f16(mode, node_map, 'while/MatMul') + self._assert_output_f16(mode, node_map, 'while/Relu') + self._assert_output_f16(mode, node_map, 'while/MatMul_1') + self._assert_output_f16(mode, node_map, 'while/Relu_1') + tol = 5e-3 if mode == 'mkl' else 1e-3 + self.assertAllClose(output_val_ref, output_val, atol=tol, rtol=tol) + @parameterized.parameters(['cuda']) @test_util.run_deprecated_v1 @test_util.disable_xla('This test does not pass with XLA') - def test_multi_paths(self): + def test_multi_paths(self, mode): """Test graph with multiple paths.""" - if test.is_gpu_available(cuda_only=True): - random_seed.set_random_seed(0) - x = _input([2, 8, 8, 3]) - x1, x2, x3 = array_ops.split(x, num_or_size_splits=3, axis=3) - y1 = _conv_pool(x1) - y2 = _conv_pool(x2) - y3 = _conv_pool(x3) - y = array_ops.concat([y1, y2, y3], axis=3) - y = array_ops.identity(y) - optimizer = gradient_descent.GradientDescentOptimizer(learning_rate=0.01) - g = optimizer.compute_gradients(y, [x]) - output = (y, g) + self._maybe_skip(mode) + random_seed.set_random_seed(0) + x = _input([2, 8, 8, 3]) + x1, x2, x3 = array_ops.split(x, num_or_size_splits=3, axis=3) + y1 = _conv_pool(x1) + y2 = _conv_pool(x2) + y3 = _conv_pool(x3) + y = array_ops.concat([y1, y2, y3], axis=3) + y = array_ops.identity(y) + optimizer = gradient_descent.GradientDescentOptimizer(learning_rate=0.01) + g = optimizer.compute_gradients(y, [x]) + output = (y, g) - output_val_ref, output_val, cost_graph = self._run(output) - node_map = _build_node_map(cost_graph.node) + output_val_ref, output_val, cost_graph = self._run(mode, output) + node_map = _build_node_map(cost_graph.node) - self._assert_output_fp16(node_map, 'split') - for suffix in [''] + ['_%i' % i for i in range(1, 6)]: - self._assert_output_fp16(node_map, 'Conv2D' + suffix) - self._assert_output_fp16(node_map, 'Relu' + suffix) - self._assert_output_fp16(node_map, 'MaxPool' + suffix) - self._assert_output_fp16(node_map, 'concat') - self.assertAllClose(output_val_ref, output_val, atol=1e-3, rtol=1e-3) + self._assert_output_f16(mode, node_map, 'split') + for suffix in [''] + ['_%i' % i for i in range(1, 6)]: + self._assert_output_f16(mode, node_map, 'Conv2D' + suffix) + self._assert_output_f16(mode, node_map, 'Relu' + suffix) + self._assert_output_f16(mode, node_map, 'MaxPool' + suffix) + self._assert_output_f16(mode, node_map, 'concat') + self.assertAllClose(output_val_ref, output_val, atol=1e-3, rtol=1e-3) + @parameterized.parameters(['cuda', 'mkl']) @test_util.run_deprecated_v1 @test_util.disable_xla('This test does not pass with XLA') - def test_multi_paths_2(self): + def test_multi_paths_2(self, mode): """Test graph with multiple paths.""" - if test.is_gpu_available(cuda_only=True): - random_seed.set_random_seed(0) - x = _input([8, 8]) - y1 = _matmul_act(x) - y2 = _matmul_act(x) - y = y1 + y2 + x - optimizer = gradient_descent.GradientDescentOptimizer(learning_rate=0.01) - g = optimizer.compute_gradients(y, [x]) - output = (g, y) + self._maybe_skip(mode) + random_seed.set_random_seed(0) + x = _input([8, 8]) + y1 = _matmul_act(x) + y2 = _matmul_act(x) + y = y1 + y2 + x + optimizer = gradient_descent.GradientDescentOptimizer(learning_rate=0.01) + g = optimizer.compute_gradients(y, [x]) + output = (g, y) - output_val_ref, output_val, cost_graph = self._run(output) - node_map = _build_node_map(cost_graph.node) + output_val_ref, output_val, cost_graph = self._run(mode, output) + node_map = _build_node_map(cost_graph.node) - self._assert_output_fp16(node_map, 'MatMul') - self._assert_output_fp16(node_map, 'Relu') - self._assert_output_fp16(node_map, 'MatMul_1') - self._assert_output_fp16(node_map, 'Relu_1') + self._assert_output_f16(mode, node_map, 'MatMul') + self._assert_output_f16(mode, node_map, 'Relu') + self._assert_output_f16(mode, node_map, 'MatMul_1') + self._assert_output_f16(mode, node_map, 'Relu_1') + if mode == 'mkl': + tol = 2e-2 + elif test.is_built_with_rocm(): # Bump up the tolerance for the ROCm platform # The default tolerance (1e-3) results in a tiny fraction (<1%) of # miscompares on ROCm platform, and hence the tolerance bump - tol = 2e-3 if test.is_built_with_rocm else 1e-3 - self.assertAllClose(output_val_ref, output_val, atol=tol, rtol=tol) + tol = 2e-3 + else: + tol = 1e-3 + self.assertAllClose(output_val_ref, output_val, atol=tol, rtol=tol) + @parameterized.parameters(['cuda']) # MKL doesn't support bf16 Sigmoid @test_util.run_v1_only('b/138749235') @test_util.disable_xla('This test does not pass with XLA') - def test_recurrent_lstm(self): + def test_recurrent_lstm(self, mode): """Test graph with recurrent lstm.""" - if test.is_gpu_available(cuda_only=True): - random_seed.set_random_seed(0) - init_c = _input([8, 4]) - init_h = _input([8, 4]) - _, _, h, _ = _recurrent_lstm(init_c, init_h) - optimizer = gradient_descent.GradientDescentOptimizer(learning_rate=0.01) - g = optimizer.compute_gradients(h, [init_c, init_h]) - output = (h, g) + self._maybe_skip(mode) + random_seed.set_random_seed(0) + init_c = _input([8, 4]) + init_h = _input([8, 4]) + _, _, h, _ = _recurrent_lstm(init_c, init_h) + optimizer = gradient_descent.GradientDescentOptimizer(learning_rate=0.01) + g = optimizer.compute_gradients(h, [init_c, init_h]) + output = (h, g) - output_val_ref, output_val, cost_graph = self._run(output) - node_map = _build_node_map(cost_graph.node) + output_val_ref, output_val, cost_graph = self._run(mode, output) + node_map = _build_node_map(cost_graph.node) - self._assert_output_fp16(node_map, 'while/concat') - self._assert_output_fp16(node_map, 'while/MatMul') - self._assert_output_fp16(node_map, 'while/split') - self._assert_output_fp16(node_map, 'while/Sigmoid') - self._assert_output_fp16(node_map, 'while/Sigmoid_1') - self._assert_output_fp16(node_map, 'while/Sigmoid_2') - self._assert_output_fp16(node_map, 'while/Tanh') - self._assert_output_fp16(node_map, 'while/Tanh_1') - self.assertAllClose(output_val_ref, output_val, atol=1e-3, rtol=1e-3) + self._assert_output_f16(mode, node_map, 'while/concat') + self._assert_output_f16(mode, node_map, 'while/MatMul') + self._assert_output_f16(mode, node_map, 'while/split') + self._assert_output_f16(mode, node_map, 'while/Sigmoid') + self._assert_output_f16(mode, node_map, 'while/Sigmoid_1') + self._assert_output_f16(mode, node_map, 'while/Sigmoid_2') + self._assert_output_f16(mode, node_map, 'while/Tanh') + self._assert_output_f16(mode, node_map, 'while/Tanh_1') + self.assertAllClose(output_val_ref, output_val, atol=1e-3, rtol=1e-3) + @parameterized.parameters(['cuda', 'mkl']) @test_util.run_v1_only('v1 loop test') @test_util.disable_xla('This test does not pass with XLA') - def test_propagation_through_simple_loop_1(self): - self._run_simple_loop_test('W', 'C', 'C') + def test_propagation_through_simple_loop_1(self, mode): + self._run_simple_loop_test(mode, 'W', 'C', 'C') + @parameterized.parameters(['cuda', 'mkl']) @test_util.run_v1_only('v1 loop test') @test_util.disable_xla('This test does not pass with XLA') - def test_propagation_through_simple_loop_2(self): - self._run_simple_loop_test('C', 'C', 'W') + def test_propagation_through_simple_loop_2(self, mode): + self._run_simple_loop_test(mode, 'C', 'C', 'W') + @parameterized.parameters(['cuda', 'mkl']) @test_util.run_v1_only('v1 loop test') @test_util.disable_xla('This test does not pass with XLA') - def test_propagation_through_simple_loop_3(self): - self._run_simple_loop_test('W', 'G', 'W') + def test_propagation_through_simple_loop_3(self, mode): + self._run_simple_loop_test(mode, 'W', 'G', 'W') + @parameterized.parameters(['cuda', 'mkl']) @test_util.run_v1_only('v1 loop test') @test_util.disable_xla('This test does not pass with XLA') - def test_propagation_through_simple_loop_4(self): - self._run_simple_loop_test('W', 'gbg', 'W') + def test_propagation_through_simple_loop_4(self, mode): + self._run_simple_loop_test(mode, 'W', 'gbg', 'W') + @parameterized.parameters(['cuda', 'mkl']) @test_util.run_v1_only('b/138749235') @test_util.disable_xla('This test does not pass with XLA') - def test_propagation_through_simple_loop_5(self): - self._run_simple_loop_test('b', 'gWC', 'c') + def test_propagation_through_simple_loop_5(self, mode): + self._run_simple_loop_test(mode, 'b', 'gWC', 'c') + @parameterized.parameters(['cuda', 'mkl']) @test_util.run_v1_only('b/138749235') @test_util.disable_xla('This test does not pass with XLA') - def test_propagation_through_simple_loop_6(self): - self._run_simple_loop_test('b', 'CWCG', 'C') + def test_propagation_through_simple_loop_6(self, mode): + self._run_simple_loop_test(mode, 'b', 'CWCG', 'C') + @parameterized.parameters(['cuda', 'mkl']) @test_util.run_v1_only('b/138749235') @test_util.disable_xla('This test does not pass with XLA') - def test_propagation_through_simple_loop_7(self): - self._run_simple_loop_test('C', 'GWCG', 'C') + def test_propagation_through_simple_loop_7(self, mode): + self._run_simple_loop_test(mode, 'C', 'GWCG', 'C') + @parameterized.parameters(['cuda', 'mkl']) @test_util.run_v1_only('b/138749235') @test_util.disable_xla('This test does not pass with XLA') - def test_propagation_through_simple_loop_8(self): - self._run_simple_loop_test('C', 'CgbgWC', 'g') + def test_propagation_through_simple_loop_8(self, mode): + self._run_simple_loop_test(mode, 'C', 'CgbgWC', 'g') + @parameterized.parameters(['cuda', 'mkl']) @test_util.run_deprecated_v1 @test_util.disable_xla('This test does not pass with XLA') - def test_noninlined_funcdef(self): + def test_noninlined_funcdef(self, mode): """Test graph with non-inlined function subgraph. This requires the grappler pass to handle an OpDef that only appears in the graph's function registry instead of the global op registry. """ - if test.is_gpu_available(cuda_only=True): - random_seed.set_random_seed(0) - x = _input([8, 8]) - y = _matmul_act(x) - y = _example_noninlined_funcdef(y) - optimizer = gradient_descent.GradientDescentOptimizer(learning_rate=0.01) - g = optimizer.compute_gradients(y, [x]) - output = (g, y) + self._maybe_skip(mode) + random_seed.set_random_seed(0) + x = _input([8, 8]) + y = _matmul_act(x) + y = _example_noninlined_funcdef(y) + optimizer = gradient_descent.GradientDescentOptimizer(learning_rate=0.01) + g = optimizer.compute_gradients(y, [x]) + output = (g, y) - output_val_ref, output_val, cost_graph = self._run(output) - node_map = _build_node_map(cost_graph.node) + output_val_ref, output_val, cost_graph = self._run(mode, output) + node_map = _build_node_map(cost_graph.node) - self._assert_output_fp16(node_map, 'MatMul') - self.assertAllClose(output_val_ref, output_val, atol=1e-3, rtol=1e-3) + self._assert_output_f16(mode, node_map, 'MatMul') + tol = 1e-2 if mode == 'mkl' else 1e-3 + self.assertAllClose(output_val_ref, output_val, atol=tol, rtol=tol) + @parameterized.parameters(['cuda', 'mkl']) @test_util.run_deprecated_v1 @test_util.disable_xla('This test does not pass with XLA') - def test_ingraph_train_loop(self): + def test_ingraph_train_loop(self, mode): """Tests a graph containing a while loop around a training update. This requires the grappler pass to take special care with its handling of Enter ops that appear in front of reads from non-resource variables. See the use of NodeImplicitlyReadsVariable in auto_mixed_precision.cc. """ + self._maybe_skip(mode) if tf2.enabled(): # This test tests non-resource variables, which are only used in TF1. self.skipTest('TensorFlow 1 required') - if test.is_gpu_available(cuda_only=True): - random_seed.set_random_seed(1234) - np.random.seed(1234) - num_iter, bs, nchan, nclass = 100, 64, 32, 100 + random_seed.set_random_seed(1234) + np.random.seed(1234) + num_iter, bs, nchan, nclass = 100, 64, 32, 100 - data = np.random.normal(size=(bs * num_iter, nchan)).astype(np.float32) - labels = np.random.randint(nclass, size=(bs * num_iter,)) - ds = dataset_ops.Dataset.from_tensor_slices((data, labels)) - ds = ds.batch(bs).prefetch(3) - it = ds.make_one_shot_iterator() + data = np.random.normal(size=(bs * num_iter, nchan)).astype(np.float32) + labels = np.random.randint(nclass, size=(bs * num_iter,)) + ds = dataset_ops.Dataset.from_tensor_slices((data, labels)) + ds = ds.batch(bs).prefetch(3) + it = ds.make_one_shot_iterator() - def body(_, i): - i += 1 - x, yt = it.get_next() - dense = layers.Dense(nclass) - y = dense(x) - loss = losses.sparse_softmax_cross_entropy(yt, y) - opt = adam.AdamOptimizer() - train_op = opt.minimize(loss, var_list=dense.trainable_weights) - with ops.control_dependencies([train_op]): - loss = array_ops.identity(loss) - return loss, i + def body(_, i): + i += 1 + x, yt = it.get_next() + dense = layers.Dense(nclass) + y = dense(x) + loss = losses.sparse_softmax_cross_entropy(yt, y) + opt = adam.AdamOptimizer() + train_op = opt.minimize(loss, var_list=dense.trainable_weights) + with ops.control_dependencies([train_op]): + loss = array_ops.identity(loss) + return loss, i - begin, end = constant_op.constant(0), constant_op.constant(num_iter) - loss, _ = control_flow_ops.while_loop( - lambda loss, i: math_ops.less(i, end), body, [0.0, begin]) + begin, end = constant_op.constant(0), constant_op.constant(num_iter) + loss, _ = control_flow_ops.while_loop( + lambda loss, i: math_ops.less(i, end), body, [0.0, begin]) - output_val_ref, output_val, cost_graph = self._run(loss) - node_map = _build_node_map(cost_graph.node) + output_val_ref, output_val, cost_graph = self._run(mode, loss) + node_map = _build_node_map(cost_graph.node) - self._assert_output_fp16(node_map, 'while/dense/MatMul') - self._assert_output_fp16( - node_map, 'while/gradients/while/dense/MatMul_grad/MatMul_1') - self.assertAllClose(output_val_ref, output_val, atol=1e-3, rtol=1e-3) + self._assert_output_f16(mode, node_map, 'while/dense/MatMul') + self._assert_output_f16( + mode, node_map, 'while/gradients/while/dense/MatMul_grad/MatMul_1') + self.assertAllClose(output_val_ref, output_val, atol=1e-3, rtol=1e-3) # TODO(benbarsdell): Add tests for list ops (TensorList*) that pass through # graph source/sink nodes, similar to the TensorListThroughFunction C++ test. From ebe063eb74ab1ee80bfb0d4447fb7c842e0ca27b Mon Sep 17 00:00:00 2001 From: Dan Moldovan Date: Thu, 18 Jun 2020 12:38:59 -0700 Subject: [PATCH 204/867] Create an empty test for python_op_gen. PiperOrigin-RevId: 317159004 Change-Id: Iabc4810f9c51c257d62dc3e7bd20d96131939d5d --- tensorflow/python/BUILD | 15 ++++++- tensorflow/python/framework/python_op_gen.cc | 16 ++++--- tensorflow/python/framework/python_op_gen.h | 14 ++++++- .../python/framework/python_op_gen_test.cc | 42 +++++++++++++++++++ 4 files changed, 80 insertions(+), 7 deletions(-) create mode 100644 tensorflow/python/framework/python_op_gen_test.cc diff --git a/tensorflow/python/BUILD b/tensorflow/python/BUILD index a4e72bf2460..de9cf9a24c7 100644 --- a/tensorflow/python/BUILD +++ b/tensorflow/python/BUILD @@ -3,7 +3,7 @@ # ":platform" - Low-level and platform-specific Python code. load("//tensorflow:tensorflow.bzl", "py_strict_library") -load("//tensorflow:tensorflow.bzl", "cc_header_only_library", "if_mlir", "if_not_windows", "if_xla_available", "py_test", "py_tests", "tf_cc_shared_object", "tf_cuda_library", "tf_gen_op_wrapper_py") +load("//tensorflow:tensorflow.bzl", "cc_header_only_library", "if_mlir", "if_not_windows", "if_xla_available", "py_test", "py_tests", "tf_cc_shared_object", "tf_cc_test", "tf_cuda_library", "tf_gen_op_wrapper_py") # buildifier: disable=same-origin-load load("//tensorflow:tensorflow.bzl", "tf_monitoring_python_deps") @@ -1236,6 +1236,19 @@ cc_library( alwayslink = 1, ) +tf_cc_test( + name = "python_op_gen_test", + srcs = ["framework/python_op_gen_test.cc"], + deps = [ + ":python_op_gen", + "//tensorflow/core:framework", + "//tensorflow/core:op_gen_lib", + "//tensorflow/core:protos_all_cc", + "//tensorflow/core:test", + "//tensorflow/core:test_main", + ], +) + py_library( name = "framework_for_generated_wrappers", srcs_version = "PY2AND3", diff --git a/tensorflow/python/framework/python_op_gen.cc b/tensorflow/python/framework/python_op_gen.cc index ca0c5d9ef1a..0f84c6a063d 100644 --- a/tensorflow/python/framework/python_op_gen.cc +++ b/tensorflow/python/framework/python_op_gen.cc @@ -981,9 +981,9 @@ void GenEagerPythonOp::AddRawOpExport(const string& parameters) { function_name_, "))\n"); } -string GetPythonOps(const OpList& ops, const ApiDefMap& api_defs, - const std::vector& hidden_ops, - const string& source_file_name = "") { +string GetPythonOpsImpl(const OpList& ops, const ApiDefMap& api_defs, + const std::vector& hidden_ops, + const string& source_file_name = "") { string result; // Header // TODO(josh11b): Mention the library for which wrappers are being generated. @@ -1069,11 +1069,17 @@ from tensorflow.python.util.tf_export import tf_export } // namespace +string GetPythonOps(const OpList& ops, const ApiDefMap& api_defs, + const std::vector& hidden_ops, + const string& source_file_name) { + return GetPythonOpsImpl(ops, api_defs, hidden_ops, source_file_name); +} + void PrintPythonOps(const OpList& ops, const ApiDefMap& api_defs, const std::vector& hidden_ops, const string& source_file_name) { printf("%s", - GetPythonOps(ops, api_defs, hidden_ops, source_file_name).c_str()); + GetPythonOpsImpl(ops, api_defs, hidden_ops, source_file_name).c_str()); } string GetPythonWrappers(const char* op_list_buf, size_t op_list_len) { @@ -1081,7 +1087,7 @@ string GetPythonWrappers(const char* op_list_buf, size_t op_list_len) { ops.ParseFromArray(op_list_buf, op_list_len); ApiDefMap api_def_map(ops); - return GetPythonOps(ops, api_def_map, {}); + return GetPythonOpsImpl(ops, api_def_map, {}); } } // namespace tensorflow diff --git a/tensorflow/python/framework/python_op_gen.h b/tensorflow/python/framework/python_op_gen.h index 22fcc452fbb..f1cd6e49013 100644 --- a/tensorflow/python/framework/python_op_gen.h +++ b/tensorflow/python/framework/python_op_gen.h @@ -23,8 +23,20 @@ limitations under the License. namespace tensorflow { +// Returns a string containing the generated Python code for the given Ops. +// ops is a protobuff, typically generated using OpRegistry::Global()->Export. +// api_defs is typically constructed directly from ops. // hidden_ops should be a list of Op names that should get a leading _ -// in the output. Prints the output to stdout. +// in the output. +// source_file_name is optional and contains the name of the original C++ source +// file where the ops' REGISTER_OP() calls reside. +string GetPythonOps(const OpList& ops, const ApiDefMap& api_defs, + const std::vector& hidden_ops, + const string& source_file_name); + +// Prints the output of GetPrintOps to stdout. +// hidden_ops should be a list of Op names that should get a leading _ +// in the output. // Optional fourth argument is the name of the original C++ source file // where the ops' REGISTER_OP() calls reside. void PrintPythonOps(const OpList& ops, const ApiDefMap& api_defs, diff --git a/tensorflow/python/framework/python_op_gen_test.cc b/tensorflow/python/framework/python_op_gen_test.cc new file mode 100644 index 00000000000..5185086fdd3 --- /dev/null +++ b/tensorflow/python/framework/python_op_gen_test.cc @@ -0,0 +1,42 @@ +/* Copyright 2020 The TensorFlow Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. +==============================================================================*/ + +#include "tensorflow/python/framework/python_op_gen.h" + +#include "tensorflow/core/framework/op.h" +#include "tensorflow/core/framework/op_def.pb.h" +#include "tensorflow/core/framework/op_gen_lib.h" +#include "tensorflow/core/platform/test.h" + +namespace tensorflow { +namespace { + +TEST(PythonOpGen, Basic) { + OpList ops; + OpRegistry::Global()->Export(false, &ops); + + ApiDefMap api_def_map(ops); + + string code = GetPythonOps(ops, api_def_map, {}, ""); + + EXPECT_TRUE(absl::StrContains(code, "def case")); + + // TODO(mdan): Add tests to verify type annotations are correctly added. +} + +// TODO(mdan): Include more tests with synhtetic ops and api defs. + +} // namespace +} // namespace tensorflow From b0418130b440fa63b588390f37d0ab21b1c4731c Mon Sep 17 00:00:00 2001 From: "A. Unique TensorFlower" Date: Thu, 18 Jun 2020 12:40:15 -0700 Subject: [PATCH 205/867] Go: Update generated wrapper functions for TensorFlow ops. PiperOrigin-RevId: 317159284 Change-Id: I438a0e11d3e678fd3d5bc153a5b0ff353e0fe86a --- tensorflow/go/op/wrappers.go | 794 +++++++++++++++++------------------ 1 file changed, 397 insertions(+), 397 deletions(-) diff --git a/tensorflow/go/op/wrappers.go b/tensorflow/go/op/wrappers.go index 188bb3b78bb..ace8e58fdcd 100644 --- a/tensorflow/go/op/wrappers.go +++ b/tensorflow/go/op/wrappers.go @@ -158,65 +158,6 @@ func FakeQuantWithMinMaxArgsGradient(scope *Scope, gradients tf.Output, inputs t return op.Output(0) } -// Applies sparse addition to `input` using individual values or slices -// -// from `updates` according to indices `indices`. The updates are non-aliasing: -// `input` is only modified in-place if no other operations will use it. -// Otherwise, a copy of `input` is made. This operation has a gradient with -// respect to both `input` and `updates`. -// -// `input` is a `Tensor` with rank `P` and `indices` is a `Tensor` of rank `Q`. -// -// `indices` must be integer tensor, containing indices into `input`. -// It must be shape \\([d_0, ..., d_{Q-2}, K]\\) where `0 < K <= P`. -// -// The innermost dimension of `indices` (with length `K`) corresponds to -// indices into elements (if `K = P`) or `(P-K)`-dimensional slices -// (if `K < P`) along the `K`th dimension of `input`. -// -// `updates` is `Tensor` of rank `Q-1+P-K` with shape: -// -// $$[d_0, ..., d_{Q-2}, input.shape[K], ..., input.shape[P-1]].$$ -// -// For example, say we want to add 4 scattered elements to a rank-1 tensor to 8 -// elements. In Python, that addition would look like this: -// -// input = tf.constant([1, 2, 3, 4, 5, 6, 7, 8]) -// indices = tf.constant([[4], [3], [1], [7]]) -// updates = tf.constant([9, 10, 11, 12]) -// output = tf.scatter_nd_non_aliasing_add(input, indices, updates) -// with tf.Session() as sess: -// print(sess.run(output)) -// -// The resulting value `output` would look like this: -// -// [1, 13, 3, 14, 14, 6, 7, 20] -// -// See `tf.scatter_nd` for more details about how to make updates to slices. -// -// Arguments: -// input: A Tensor. -// indices: A Tensor. Must be one of the following types: `int32`, `int64`. -// A tensor of indices into `input`. -// updates: A Tensor. Must have the same type as ref. A tensor of updated values -// to add to `input`. -// -// Returns A `Tensor` with the same shape as `input`, containing values of `input` -// updated with `updates`. -func ScatterNdNonAliasingAdd(scope *Scope, input tf.Output, indices tf.Output, updates tf.Output) (output tf.Output) { - if scope.Err() != nil { - return - } - opspec := tf.OpSpec{ - Type: "ScatterNdNonAliasingAdd", - Input: []tf.Input{ - input, indices, updates, - }, - } - op := scope.AddOperation(opspec) - return op.Output(0) -} - // Subtracts sparse `updates` from an existing tensor according to `indices`. // // This operation creates a new tensor by subtracting sparse `updates` from the @@ -13210,115 +13151,6 @@ func ResizeNearestNeighborGrad(scope *Scope, grads tf.Output, size tf.Output, op return op.Output(0) } -// A placeholder op for a value that will be fed into the computation. -// -// Arguments: -// dtype: The type of elements in the tensor. -// shape: The shape of the tensor. -// -// Returns A tensor that will be provided using the infeed mechanism. -func InfeedDequeue(scope *Scope, dtype tf.DataType, shape tf.Shape) (output tf.Output) { - if scope.Err() != nil { - return - } - attrs := map[string]interface{}{"dtype": dtype, "shape": shape} - opspec := tf.OpSpec{ - Type: "InfeedDequeue", - - Attrs: attrs, - } - op := scope.AddOperation(opspec) - return op.Output(0) -} - -// Encodes a `RaggedTensor` into a `variant` Tensor. -// -// -// Encodes the given `RaggedTensor` and returns a `variant` Tensor. If -// `batched_input` is True, then input `RaggedTensor` is unbatched along the -// zero-th dimension, each component `RaggedTensor` is encoded into a scalar -// `variant` Tensor, and these are stacked to return a 1-D `variant` Tensor. -// If `batched_input` is False, then the input `RaggedTensor` is encoded as is and -// a scalar `variant` Tensor is returned. A `RaggedTensor` is encoded by first -// creating a 1-D `variant` Tensor with `ragged_rank + 1` elements, containing the -// splits and values Tensors of the `RaggedTensor`. Then the 1-D `variant` Tensor -// is wrapped in a scalar `variant` Tensor. See `RaggedTensorFromVariant` for the -// corresponding decoding logic. -// -// -// Arguments: -// rt_nested_splits: A list of one or more Tensors representing the splits of the input -// `RaggedTensor`. -// rt_dense_values: A Tensor representing the values of the input `RaggedTensor`. -// batched_input: A `bool` denoting whether the input is a batched `RaggedTensor`. -// -// Returns A `variant` Tensor that containing encoded `RaggedTensor`. -func RaggedTensorToVariant(scope *Scope, rt_nested_splits []tf.Output, rt_dense_values tf.Output, batched_input bool) (encoded_ragged tf.Output) { - if scope.Err() != nil { - return - } - attrs := map[string]interface{}{"batched_input": batched_input} - opspec := tf.OpSpec{ - Type: "RaggedTensorToVariant", - Input: []tf.Input{ - tf.OutputList(rt_nested_splits), rt_dense_values, - }, - Attrs: attrs, - } - op := scope.AddOperation(opspec) - return op.Output(0) -} - -// ResizeNearestNeighborAttr is an optional argument to ResizeNearestNeighbor. -type ResizeNearestNeighborAttr func(optionalAttr) - -// ResizeNearestNeighborAlignCorners sets the optional align_corners attribute to value. -// -// value: If true, the centers of the 4 corner pixels of the input and output tensors are -// aligned, preserving the values at the corner pixels. Defaults to false. -// If not specified, defaults to false -func ResizeNearestNeighborAlignCorners(value bool) ResizeNearestNeighborAttr { - return func(m optionalAttr) { - m["align_corners"] = value - } -} - -// ResizeNearestNeighborHalfPixelCenters sets the optional half_pixel_centers attribute to value. -// If not specified, defaults to false -func ResizeNearestNeighborHalfPixelCenters(value bool) ResizeNearestNeighborAttr { - return func(m optionalAttr) { - m["half_pixel_centers"] = value - } -} - -// Resize `images` to `size` using nearest neighbor interpolation. -// -// Arguments: -// images: 4-D with shape `[batch, height, width, channels]`. -// size: = A 1-D int32 Tensor of 2 elements: `new_height, new_width`. The -// new size for the images. -// -// Returns 4-D with shape -// `[batch, new_height, new_width, channels]`. -func ResizeNearestNeighbor(scope *Scope, images tf.Output, size tf.Output, optional ...ResizeNearestNeighborAttr) (resized_images tf.Output) { - if scope.Err() != nil { - return - } - attrs := map[string]interface{}{} - for _, a := range optional { - a(attrs) - } - opspec := tf.OpSpec{ - Type: "ResizeNearestNeighbor", - Input: []tf.Input{ - images, size, - }, - Attrs: attrs, - } - op := scope.AddOperation(opspec) - return op.Output(0) -} - // Runs multiple additive regression ensemble predictors on input instances and // // computes the logits. It is designed to be used during prediction. @@ -15836,6 +15668,65 @@ func SqrtGrad(scope *Scope, y tf.Output, dy tf.Output) (z tf.Output) { return op.Output(0) } +// Applies sparse addition to `input` using individual values or slices +// +// from `updates` according to indices `indices`. The updates are non-aliasing: +// `input` is only modified in-place if no other operations will use it. +// Otherwise, a copy of `input` is made. This operation has a gradient with +// respect to both `input` and `updates`. +// +// `input` is a `Tensor` with rank `P` and `indices` is a `Tensor` of rank `Q`. +// +// `indices` must be integer tensor, containing indices into `input`. +// It must be shape \\([d_0, ..., d_{Q-2}, K]\\) where `0 < K <= P`. +// +// The innermost dimension of `indices` (with length `K`) corresponds to +// indices into elements (if `K = P`) or `(P-K)`-dimensional slices +// (if `K < P`) along the `K`th dimension of `input`. +// +// `updates` is `Tensor` of rank `Q-1+P-K` with shape: +// +// $$[d_0, ..., d_{Q-2}, input.shape[K], ..., input.shape[P-1]].$$ +// +// For example, say we want to add 4 scattered elements to a rank-1 tensor to 8 +// elements. In Python, that addition would look like this: +// +// input = tf.constant([1, 2, 3, 4, 5, 6, 7, 8]) +// indices = tf.constant([[4], [3], [1], [7]]) +// updates = tf.constant([9, 10, 11, 12]) +// output = tf.scatter_nd_non_aliasing_add(input, indices, updates) +// with tf.Session() as sess: +// print(sess.run(output)) +// +// The resulting value `output` would look like this: +// +// [1, 13, 3, 14, 14, 6, 7, 20] +// +// See `tf.scatter_nd` for more details about how to make updates to slices. +// +// Arguments: +// input: A Tensor. +// indices: A Tensor. Must be one of the following types: `int32`, `int64`. +// A tensor of indices into `input`. +// updates: A Tensor. Must have the same type as ref. A tensor of updated values +// to add to `input`. +// +// Returns A `Tensor` with the same shape as `input`, containing values of `input` +// updated with `updates`. +func ScatterNdNonAliasingAdd(scope *Scope, input tf.Output, indices tf.Output, updates tf.Output) (output tf.Output) { + if scope.Err() != nil { + return + } + opspec := tf.OpSpec{ + Type: "ScatterNdNonAliasingAdd", + Input: []tf.Input{ + input, indices, updates, + }, + } + op := scope.AddOperation(opspec) + return op.Output(0) +} + // MutableHashTableOfTensorsV2Attr is an optional argument to MutableHashTableOfTensorsV2. type MutableHashTableOfTensorsV2Attr func(optionalAttr) @@ -37168,235 +37059,6 @@ func SdcaFprint(scope *Scope, input tf.Output) (output tf.Output) { return op.Output(0) } -// ParseSequenceExampleAttr is an optional argument to ParseSequenceExample. -type ParseSequenceExampleAttr func(optionalAttr) - -// ParseSequenceExampleNcontextSparse sets the optional Ncontext_sparse attribute to value. -// If not specified, defaults to 0 -// -// REQUIRES: value >= 0 -func ParseSequenceExampleNcontextSparse(value int64) ParseSequenceExampleAttr { - return func(m optionalAttr) { - m["Ncontext_sparse"] = value - } -} - -// ParseSequenceExampleNcontextDense sets the optional Ncontext_dense attribute to value. -// If not specified, defaults to 0 -// -// REQUIRES: value >= 0 -func ParseSequenceExampleNcontextDense(value int64) ParseSequenceExampleAttr { - return func(m optionalAttr) { - m["Ncontext_dense"] = value - } -} - -// ParseSequenceExampleNfeatureListSparse sets the optional Nfeature_list_sparse attribute to value. -// If not specified, defaults to 0 -// -// REQUIRES: value >= 0 -func ParseSequenceExampleNfeatureListSparse(value int64) ParseSequenceExampleAttr { - return func(m optionalAttr) { - m["Nfeature_list_sparse"] = value - } -} - -// ParseSequenceExampleNfeatureListDense sets the optional Nfeature_list_dense attribute to value. -// If not specified, defaults to 0 -// -// REQUIRES: value >= 0 -func ParseSequenceExampleNfeatureListDense(value int64) ParseSequenceExampleAttr { - return func(m optionalAttr) { - m["Nfeature_list_dense"] = value - } -} - -// ParseSequenceExampleContextSparseTypes sets the optional context_sparse_types attribute to value. -// -// value: A list of Ncontext_sparse types; the data types of data in -// each context Feature given in context_sparse_keys. -// Currently the ParseSingleSequenceExample supports DT_FLOAT (FloatList), -// DT_INT64 (Int64List), and DT_STRING (BytesList). -// If not specified, defaults to <> -// -// REQUIRES: len(value) >= 0 -func ParseSequenceExampleContextSparseTypes(value []tf.DataType) ParseSequenceExampleAttr { - return func(m optionalAttr) { - m["context_sparse_types"] = value - } -} - -// ParseSequenceExampleFeatureListDenseTypes sets the optional feature_list_dense_types attribute to value. -// If not specified, defaults to <> -// -// REQUIRES: len(value) >= 0 -func ParseSequenceExampleFeatureListDenseTypes(value []tf.DataType) ParseSequenceExampleAttr { - return func(m optionalAttr) { - m["feature_list_dense_types"] = value - } -} - -// ParseSequenceExampleContextDenseShapes sets the optional context_dense_shapes attribute to value. -// -// value: A list of Ncontext_dense shapes; the shapes of data in -// each context Feature given in context_dense_keys. -// The number of elements in the Feature corresponding to context_dense_key[j] -// must always equal context_dense_shapes[j].NumEntries(). -// The shape of context_dense_values[j] will match context_dense_shapes[j]. -// If not specified, defaults to <> -// -// REQUIRES: len(value) >= 0 -func ParseSequenceExampleContextDenseShapes(value []tf.Shape) ParseSequenceExampleAttr { - return func(m optionalAttr) { - m["context_dense_shapes"] = value - } -} - -// ParseSequenceExampleFeatureListSparseTypes sets the optional feature_list_sparse_types attribute to value. -// -// value: A list of Nfeature_list_sparse types; the data types -// of data in each FeatureList given in feature_list_sparse_keys. -// Currently the ParseSingleSequenceExample supports DT_FLOAT (FloatList), -// DT_INT64 (Int64List), and DT_STRING (BytesList). -// If not specified, defaults to <> -// -// REQUIRES: len(value) >= 0 -func ParseSequenceExampleFeatureListSparseTypes(value []tf.DataType) ParseSequenceExampleAttr { - return func(m optionalAttr) { - m["feature_list_sparse_types"] = value - } -} - -// ParseSequenceExampleFeatureListDenseShapes sets the optional feature_list_dense_shapes attribute to value. -// -// value: A list of Nfeature_list_dense shapes; the shapes of -// data in each FeatureList given in feature_list_dense_keys. -// The shape of each Feature in the FeatureList corresponding to -// feature_list_dense_key[j] must always equal -// feature_list_dense_shapes[j].NumEntries(). -// If not specified, defaults to <> -// -// REQUIRES: len(value) >= 0 -func ParseSequenceExampleFeatureListDenseShapes(value []tf.Shape) ParseSequenceExampleAttr { - return func(m optionalAttr) { - m["feature_list_dense_shapes"] = value - } -} - -// Transforms a vector of brain.SequenceExample protos (as strings) into typed tensors. -// -// Arguments: -// serialized: A vector containing binary serialized SequenceExample protos. -// debug_name: A vector containing the names of the serialized protos. -// May contain, for example, table key (descriptive) name for the -// corresponding serialized proto. This is purely useful for debugging -// purposes, and the presence of values here has no effect on the output. -// May also be an empty vector if no name is available. -// context_dense_defaults: A list of Ncontext_dense Tensors (some may be empty). -// context_dense_defaults[j] provides default values -// when the SequenceExample's context map lacks context_dense_key[j]. -// If an empty Tensor is provided for context_dense_defaults[j], -// then the Feature context_dense_keys[j] is required. -// The input type is inferred from context_dense_defaults[j], even when it's -// empty. If context_dense_defaults[j] is not empty, its shape must match -// context_dense_shapes[j]. -// feature_list_dense_missing_assumed_empty: A vector listing the -// FeatureList keys which may be missing from the SequenceExamples. If the -// associated FeatureList is missing, it is treated as empty. By default, -// any FeatureList not listed in this vector must exist in the SequenceExamples. -// context_sparse_keys: A list of Ncontext_sparse string Tensors (scalars). -// The keys expected in the Examples' features associated with context_sparse -// values. -// context_dense_keys: A list of Ncontext_dense string Tensors (scalars). -// The keys expected in the SequenceExamples' context features associated with -// dense values. -// feature_list_sparse_keys: A list of Nfeature_list_sparse string Tensors -// (scalars). The keys expected in the FeatureLists associated with sparse -// values. -// feature_list_dense_keys: A list of Nfeature_list_dense string Tensors (scalars). -// The keys expected in the SequenceExamples' feature_lists associated -// with lists of dense values. -func ParseSequenceExample(scope *Scope, serialized tf.Output, debug_name tf.Output, context_dense_defaults []tf.Output, feature_list_dense_missing_assumed_empty []string, context_sparse_keys []string, context_dense_keys []string, feature_list_sparse_keys []string, feature_list_dense_keys []string, optional ...ParseSequenceExampleAttr) (context_sparse_indices []tf.Output, context_sparse_values []tf.Output, context_sparse_shapes []tf.Output, context_dense_values []tf.Output, feature_list_sparse_indices []tf.Output, feature_list_sparse_values []tf.Output, feature_list_sparse_shapes []tf.Output, feature_list_dense_values []tf.Output, feature_list_dense_lengths []tf.Output) { - if scope.Err() != nil { - return - } - attrs := map[string]interface{}{"feature_list_dense_missing_assumed_empty": feature_list_dense_missing_assumed_empty, "context_sparse_keys": context_sparse_keys, "context_dense_keys": context_dense_keys, "feature_list_sparse_keys": feature_list_sparse_keys, "feature_list_dense_keys": feature_list_dense_keys} - for _, a := range optional { - a(attrs) - } - opspec := tf.OpSpec{ - Type: "ParseSequenceExample", - Input: []tf.Input{ - serialized, debug_name, tf.OutputList(context_dense_defaults), - }, - Attrs: attrs, - } - op := scope.AddOperation(opspec) - if scope.Err() != nil { - return - } - var idx int - var err error - if context_sparse_indices, idx, err = makeOutputList(op, idx, "context_sparse_indices"); err != nil { - scope.UpdateErr("ParseSequenceExample", err) - return - } - if context_sparse_values, idx, err = makeOutputList(op, idx, "context_sparse_values"); err != nil { - scope.UpdateErr("ParseSequenceExample", err) - return - } - if context_sparse_shapes, idx, err = makeOutputList(op, idx, "context_sparse_shapes"); err != nil { - scope.UpdateErr("ParseSequenceExample", err) - return - } - if context_dense_values, idx, err = makeOutputList(op, idx, "context_dense_values"); err != nil { - scope.UpdateErr("ParseSequenceExample", err) - return - } - if feature_list_sparse_indices, idx, err = makeOutputList(op, idx, "feature_list_sparse_indices"); err != nil { - scope.UpdateErr("ParseSequenceExample", err) - return - } - if feature_list_sparse_values, idx, err = makeOutputList(op, idx, "feature_list_sparse_values"); err != nil { - scope.UpdateErr("ParseSequenceExample", err) - return - } - if feature_list_sparse_shapes, idx, err = makeOutputList(op, idx, "feature_list_sparse_shapes"); err != nil { - scope.UpdateErr("ParseSequenceExample", err) - return - } - if feature_list_dense_values, idx, err = makeOutputList(op, idx, "feature_list_dense_values"); err != nil { - scope.UpdateErr("ParseSequenceExample", err) - return - } - if feature_list_dense_lengths, idx, err = makeOutputList(op, idx, "feature_list_dense_lengths"); err != nil { - scope.UpdateErr("ParseSequenceExample", err) - return - } - return context_sparse_indices, context_sparse_values, context_sparse_shapes, context_dense_values, feature_list_sparse_indices, feature_list_sparse_values, feature_list_sparse_shapes, feature_list_dense_values, feature_list_dense_lengths -} - -// Returns true if queue is closed. -// -// This operation returns true if the queue is closed and false if the queue -// is open. -// -// Arguments: -// handle: The handle to a queue. -func QueueIsClosedV2(scope *Scope, handle tf.Output) (is_closed tf.Output) { - if scope.Err() != nil { - return - } - opspec := tf.OpSpec{ - Type: "QueueIsClosedV2", - Input: []tf.Input{ - handle, - }, - } - op := scope.AddOperation(opspec) - return op.Output(0) -} - // Inverse 3D fast Fourier transform. // // Computes the inverse 3-dimensional discrete Fourier transform over the @@ -38804,6 +38466,27 @@ func StatefulTruncatedNormal(scope *Scope, resource tf.Output, algorithm tf.Outp return op.Output(0) } +// Returns true if queue is closed. +// +// This operation returns true if the queue is closed and false if the queue +// is open. +// +// Arguments: +// handle: The handle to a queue. +func QueueIsClosedV2(scope *Scope, handle tf.Output) (is_closed tf.Output) { + if scope.Err() != nil { + return + } + opspec := tf.OpSpec{ + Type: "QueueIsClosedV2", + Input: []tf.Input{ + handle, + }, + } + op := scope.AddOperation(opspec) + return op.Output(0) +} + // Checks whether a quantile stream has been initialized. // // An Op that checks if quantile stream resource is initialized. @@ -38826,6 +38509,214 @@ func IsBoostedTreesQuantileStreamResourceInitialized(scope *Scope, quantile_stre return op.Output(0) } +// ParseSequenceExampleAttr is an optional argument to ParseSequenceExample. +type ParseSequenceExampleAttr func(optionalAttr) + +// ParseSequenceExampleNcontextSparse sets the optional Ncontext_sparse attribute to value. +// If not specified, defaults to 0 +// +// REQUIRES: value >= 0 +func ParseSequenceExampleNcontextSparse(value int64) ParseSequenceExampleAttr { + return func(m optionalAttr) { + m["Ncontext_sparse"] = value + } +} + +// ParseSequenceExampleNcontextDense sets the optional Ncontext_dense attribute to value. +// If not specified, defaults to 0 +// +// REQUIRES: value >= 0 +func ParseSequenceExampleNcontextDense(value int64) ParseSequenceExampleAttr { + return func(m optionalAttr) { + m["Ncontext_dense"] = value + } +} + +// ParseSequenceExampleNfeatureListSparse sets the optional Nfeature_list_sparse attribute to value. +// If not specified, defaults to 0 +// +// REQUIRES: value >= 0 +func ParseSequenceExampleNfeatureListSparse(value int64) ParseSequenceExampleAttr { + return func(m optionalAttr) { + m["Nfeature_list_sparse"] = value + } +} + +// ParseSequenceExampleNfeatureListDense sets the optional Nfeature_list_dense attribute to value. +// If not specified, defaults to 0 +// +// REQUIRES: value >= 0 +func ParseSequenceExampleNfeatureListDense(value int64) ParseSequenceExampleAttr { + return func(m optionalAttr) { + m["Nfeature_list_dense"] = value + } +} + +// ParseSequenceExampleContextSparseTypes sets the optional context_sparse_types attribute to value. +// +// value: A list of Ncontext_sparse types; the data types of data in +// each context Feature given in context_sparse_keys. +// Currently the ParseSingleSequenceExample supports DT_FLOAT (FloatList), +// DT_INT64 (Int64List), and DT_STRING (BytesList). +// If not specified, defaults to <> +// +// REQUIRES: len(value) >= 0 +func ParseSequenceExampleContextSparseTypes(value []tf.DataType) ParseSequenceExampleAttr { + return func(m optionalAttr) { + m["context_sparse_types"] = value + } +} + +// ParseSequenceExampleFeatureListDenseTypes sets the optional feature_list_dense_types attribute to value. +// If not specified, defaults to <> +// +// REQUIRES: len(value) >= 0 +func ParseSequenceExampleFeatureListDenseTypes(value []tf.DataType) ParseSequenceExampleAttr { + return func(m optionalAttr) { + m["feature_list_dense_types"] = value + } +} + +// ParseSequenceExampleContextDenseShapes sets the optional context_dense_shapes attribute to value. +// +// value: A list of Ncontext_dense shapes; the shapes of data in +// each context Feature given in context_dense_keys. +// The number of elements in the Feature corresponding to context_dense_key[j] +// must always equal context_dense_shapes[j].NumEntries(). +// The shape of context_dense_values[j] will match context_dense_shapes[j]. +// If not specified, defaults to <> +// +// REQUIRES: len(value) >= 0 +func ParseSequenceExampleContextDenseShapes(value []tf.Shape) ParseSequenceExampleAttr { + return func(m optionalAttr) { + m["context_dense_shapes"] = value + } +} + +// ParseSequenceExampleFeatureListSparseTypes sets the optional feature_list_sparse_types attribute to value. +// +// value: A list of Nfeature_list_sparse types; the data types +// of data in each FeatureList given in feature_list_sparse_keys. +// Currently the ParseSingleSequenceExample supports DT_FLOAT (FloatList), +// DT_INT64 (Int64List), and DT_STRING (BytesList). +// If not specified, defaults to <> +// +// REQUIRES: len(value) >= 0 +func ParseSequenceExampleFeatureListSparseTypes(value []tf.DataType) ParseSequenceExampleAttr { + return func(m optionalAttr) { + m["feature_list_sparse_types"] = value + } +} + +// ParseSequenceExampleFeatureListDenseShapes sets the optional feature_list_dense_shapes attribute to value. +// +// value: A list of Nfeature_list_dense shapes; the shapes of +// data in each FeatureList given in feature_list_dense_keys. +// The shape of each Feature in the FeatureList corresponding to +// feature_list_dense_key[j] must always equal +// feature_list_dense_shapes[j].NumEntries(). +// If not specified, defaults to <> +// +// REQUIRES: len(value) >= 0 +func ParseSequenceExampleFeatureListDenseShapes(value []tf.Shape) ParseSequenceExampleAttr { + return func(m optionalAttr) { + m["feature_list_dense_shapes"] = value + } +} + +// Transforms a vector of brain.SequenceExample protos (as strings) into typed tensors. +// +// Arguments: +// serialized: A vector containing binary serialized SequenceExample protos. +// debug_name: A vector containing the names of the serialized protos. +// May contain, for example, table key (descriptive) name for the +// corresponding serialized proto. This is purely useful for debugging +// purposes, and the presence of values here has no effect on the output. +// May also be an empty vector if no name is available. +// context_dense_defaults: A list of Ncontext_dense Tensors (some may be empty). +// context_dense_defaults[j] provides default values +// when the SequenceExample's context map lacks context_dense_key[j]. +// If an empty Tensor is provided for context_dense_defaults[j], +// then the Feature context_dense_keys[j] is required. +// The input type is inferred from context_dense_defaults[j], even when it's +// empty. If context_dense_defaults[j] is not empty, its shape must match +// context_dense_shapes[j]. +// feature_list_dense_missing_assumed_empty: A vector listing the +// FeatureList keys which may be missing from the SequenceExamples. If the +// associated FeatureList is missing, it is treated as empty. By default, +// any FeatureList not listed in this vector must exist in the SequenceExamples. +// context_sparse_keys: A list of Ncontext_sparse string Tensors (scalars). +// The keys expected in the Examples' features associated with context_sparse +// values. +// context_dense_keys: A list of Ncontext_dense string Tensors (scalars). +// The keys expected in the SequenceExamples' context features associated with +// dense values. +// feature_list_sparse_keys: A list of Nfeature_list_sparse string Tensors +// (scalars). The keys expected in the FeatureLists associated with sparse +// values. +// feature_list_dense_keys: A list of Nfeature_list_dense string Tensors (scalars). +// The keys expected in the SequenceExamples' feature_lists associated +// with lists of dense values. +func ParseSequenceExample(scope *Scope, serialized tf.Output, debug_name tf.Output, context_dense_defaults []tf.Output, feature_list_dense_missing_assumed_empty []string, context_sparse_keys []string, context_dense_keys []string, feature_list_sparse_keys []string, feature_list_dense_keys []string, optional ...ParseSequenceExampleAttr) (context_sparse_indices []tf.Output, context_sparse_values []tf.Output, context_sparse_shapes []tf.Output, context_dense_values []tf.Output, feature_list_sparse_indices []tf.Output, feature_list_sparse_values []tf.Output, feature_list_sparse_shapes []tf.Output, feature_list_dense_values []tf.Output, feature_list_dense_lengths []tf.Output) { + if scope.Err() != nil { + return + } + attrs := map[string]interface{}{"feature_list_dense_missing_assumed_empty": feature_list_dense_missing_assumed_empty, "context_sparse_keys": context_sparse_keys, "context_dense_keys": context_dense_keys, "feature_list_sparse_keys": feature_list_sparse_keys, "feature_list_dense_keys": feature_list_dense_keys} + for _, a := range optional { + a(attrs) + } + opspec := tf.OpSpec{ + Type: "ParseSequenceExample", + Input: []tf.Input{ + serialized, debug_name, tf.OutputList(context_dense_defaults), + }, + Attrs: attrs, + } + op := scope.AddOperation(opspec) + if scope.Err() != nil { + return + } + var idx int + var err error + if context_sparse_indices, idx, err = makeOutputList(op, idx, "context_sparse_indices"); err != nil { + scope.UpdateErr("ParseSequenceExample", err) + return + } + if context_sparse_values, idx, err = makeOutputList(op, idx, "context_sparse_values"); err != nil { + scope.UpdateErr("ParseSequenceExample", err) + return + } + if context_sparse_shapes, idx, err = makeOutputList(op, idx, "context_sparse_shapes"); err != nil { + scope.UpdateErr("ParseSequenceExample", err) + return + } + if context_dense_values, idx, err = makeOutputList(op, idx, "context_dense_values"); err != nil { + scope.UpdateErr("ParseSequenceExample", err) + return + } + if feature_list_sparse_indices, idx, err = makeOutputList(op, idx, "feature_list_sparse_indices"); err != nil { + scope.UpdateErr("ParseSequenceExample", err) + return + } + if feature_list_sparse_values, idx, err = makeOutputList(op, idx, "feature_list_sparse_values"); err != nil { + scope.UpdateErr("ParseSequenceExample", err) + return + } + if feature_list_sparse_shapes, idx, err = makeOutputList(op, idx, "feature_list_sparse_shapes"); err != nil { + scope.UpdateErr("ParseSequenceExample", err) + return + } + if feature_list_dense_values, idx, err = makeOutputList(op, idx, "feature_list_dense_values"); err != nil { + scope.UpdateErr("ParseSequenceExample", err) + return + } + if feature_list_dense_lengths, idx, err = makeOutputList(op, idx, "feature_list_dense_lengths"); err != nil { + scope.UpdateErr("ParseSequenceExample", err) + return + } + return context_sparse_indices, context_sparse_values, context_sparse_shapes, context_dense_values, feature_list_sparse_indices, feature_list_sparse_values, feature_list_sparse_shapes, feature_list_dense_values, feature_list_dense_lengths +} + // Fast Fourier transform. // // Computes the 1-dimensional discrete Fourier transform over the inner-most @@ -41742,6 +41633,115 @@ func ResourceApplyAdam(scope *Scope, var_ tf.Output, m tf.Output, v tf.Output, b return scope.AddOperation(opspec) } +// ResizeNearestNeighborAttr is an optional argument to ResizeNearestNeighbor. +type ResizeNearestNeighborAttr func(optionalAttr) + +// ResizeNearestNeighborAlignCorners sets the optional align_corners attribute to value. +// +// value: If true, the centers of the 4 corner pixels of the input and output tensors are +// aligned, preserving the values at the corner pixels. Defaults to false. +// If not specified, defaults to false +func ResizeNearestNeighborAlignCorners(value bool) ResizeNearestNeighborAttr { + return func(m optionalAttr) { + m["align_corners"] = value + } +} + +// ResizeNearestNeighborHalfPixelCenters sets the optional half_pixel_centers attribute to value. +// If not specified, defaults to false +func ResizeNearestNeighborHalfPixelCenters(value bool) ResizeNearestNeighborAttr { + return func(m optionalAttr) { + m["half_pixel_centers"] = value + } +} + +// Resize `images` to `size` using nearest neighbor interpolation. +// +// Arguments: +// images: 4-D with shape `[batch, height, width, channels]`. +// size: = A 1-D int32 Tensor of 2 elements: `new_height, new_width`. The +// new size for the images. +// +// Returns 4-D with shape +// `[batch, new_height, new_width, channels]`. +func ResizeNearestNeighbor(scope *Scope, images tf.Output, size tf.Output, optional ...ResizeNearestNeighborAttr) (resized_images tf.Output) { + if scope.Err() != nil { + return + } + attrs := map[string]interface{}{} + for _, a := range optional { + a(attrs) + } + opspec := tf.OpSpec{ + Type: "ResizeNearestNeighbor", + Input: []tf.Input{ + images, size, + }, + Attrs: attrs, + } + op := scope.AddOperation(opspec) + return op.Output(0) +} + +// A placeholder op for a value that will be fed into the computation. +// +// Arguments: +// dtype: The type of elements in the tensor. +// shape: The shape of the tensor. +// +// Returns A tensor that will be provided using the infeed mechanism. +func InfeedDequeue(scope *Scope, dtype tf.DataType, shape tf.Shape) (output tf.Output) { + if scope.Err() != nil { + return + } + attrs := map[string]interface{}{"dtype": dtype, "shape": shape} + opspec := tf.OpSpec{ + Type: "InfeedDequeue", + + Attrs: attrs, + } + op := scope.AddOperation(opspec) + return op.Output(0) +} + +// Encodes a `RaggedTensor` into a `variant` Tensor. +// +// +// Encodes the given `RaggedTensor` and returns a `variant` Tensor. If +// `batched_input` is True, then input `RaggedTensor` is unbatched along the +// zero-th dimension, each component `RaggedTensor` is encoded into a scalar +// `variant` Tensor, and these are stacked to return a 1-D `variant` Tensor. +// If `batched_input` is False, then the input `RaggedTensor` is encoded as is and +// a scalar `variant` Tensor is returned. A `RaggedTensor` is encoded by first +// creating a 1-D `variant` Tensor with `ragged_rank + 1` elements, containing the +// splits and values Tensors of the `RaggedTensor`. Then the 1-D `variant` Tensor +// is wrapped in a scalar `variant` Tensor. See `RaggedTensorFromVariant` for the +// corresponding decoding logic. +// +// +// Arguments: +// rt_nested_splits: A list of one or more Tensors representing the splits of the input +// `RaggedTensor`. +// rt_dense_values: A Tensor representing the values of the input `RaggedTensor`. +// batched_input: A `bool` denoting whether the input is a batched `RaggedTensor`. +// +// Returns A `variant` Tensor that containing encoded `RaggedTensor`. +func RaggedTensorToVariant(scope *Scope, rt_nested_splits []tf.Output, rt_dense_values tf.Output, batched_input bool) (encoded_ragged tf.Output) { + if scope.Err() != nil { + return + } + attrs := map[string]interface{}{"batched_input": batched_input} + opspec := tf.OpSpec{ + Type: "RaggedTensorToVariant", + Input: []tf.Input{ + tf.OutputList(rt_nested_splits), rt_dense_values, + }, + Attrs: attrs, + } + op := scope.AddOperation(opspec) + return op.Output(0) +} + // ResourceApplyKerasMomentumAttr is an optional argument to ResourceApplyKerasMomentum. type ResourceApplyKerasMomentumAttr func(optionalAttr) From 07d5aa230954276d9de201217b18e86815328ab8 Mon Sep 17 00:00:00 2001 From: Sachin Joglekar Date: Thu, 18 Jun 2020 12:58:11 -0700 Subject: [PATCH 206/867] Use Requantize after per-channel Conv only if tensor min/max is different from RELU min/max beyond a limit. Also rectify RELU1 bound to (-1, 1) PiperOrigin-RevId: 317162909 Change-Id: Ice90226436cccf49507bd17877222da755d22644 --- .../hexagon/builders/conv_2d_builder.cc | 11 ++- .../hexagon/builders/tests/conv_test.cc | 84 +++++++++++++++++++ 2 files changed, 92 insertions(+), 3 deletions(-) diff --git a/tensorflow/lite/delegates/hexagon/builders/conv_2d_builder.cc b/tensorflow/lite/delegates/hexagon/builders/conv_2d_builder.cc index 97db6bf8fd0..a366522e35c 100644 --- a/tensorflow/lite/delegates/hexagon/builders/conv_2d_builder.cc +++ b/tensorflow/lite/delegates/hexagon/builders/conv_2d_builder.cc @@ -16,6 +16,7 @@ limitations under the License. #include +#include #include #include "tensorflow/lite/c/builtin_op_data.h" @@ -197,7 +198,7 @@ TfLiteStatus Conv2dOpBuilder::PopulateSubGraph(const TfLiteIntArray* inputs, conv_output_min = 0; conv_output_max = 6; } else if (activation == kTfLiteActRelu1) { - conv_output_min = 0; + conv_output_min = -1; conv_output_max = 1; } else if (activation == kTfLiteActRelu) { conv_output_min = 0; @@ -351,8 +352,12 @@ TfLiteStatus Conv2dOpBuilder::PopulateSubGraph(const TfLiteIntArray* inputs, output_max_tensor = AddOutput(sizeof(float), 4, kScalarShape); } - // Requantize if activation was not None. - if (activation != kTfLiteActNone) { + // Requantize if activation was not None & the TFLite tensor's min/max is + // different (diff > 1e-2) from the RELU bounds. + const float min_bound_diff = std::abs(conv_output_min - output_min); + const float max_bound_diff = std::abs(conv_output_max - output_max); + if (activation != kTfLiteActNone && + (min_bound_diff > 0.01 || max_bound_diff > 0.01)) { auto* requantized_min_const = graph_builder_->AddConstNodeWithData( kScalarShape, reinterpret_cast(&output_min), sizeof(output_min)); auto* requantized_max_const = graph_builder_->AddConstNodeWithData( diff --git a/tensorflow/lite/delegates/hexagon/builders/tests/conv_test.cc b/tensorflow/lite/delegates/hexagon/builders/tests/conv_test.cc index 13fd768fded..eed1bf29aae 100644 --- a/tensorflow/lite/delegates/hexagon/builders/tests/conv_test.cc +++ b/tensorflow/lite/delegates/hexagon/builders/tests/conv_test.cc @@ -207,6 +207,43 @@ TEST(QuantizedConvolutionOpModel, SimpleConvTestReLU6Activation) { 1e-5))); } +// Same as above, but the output min/max matches the RELU bounds. +// Therefore, a Requantize node will not get added after Supernode. +TEST(QuantizedConvolutionOpModel, + SimpleConvTestReLU6Activation_NoRequantizeRequired) { + QuantizedConvolutionOpModel m( + BuiltinOperator_CONV_2D, {TensorType_UINT8, {2, 2, 4, 1}, -63.5, 64}, + {TensorType_UINT8, {3, 2, 2, 1}, -63.5, 64}, {TensorType_UINT8, {}, 0, 6}, + Padding_VALID, /**dilation_factor**/ 1, + /**stride**/ 2, ActivationFunctionType_RELU6); + m.SetInput({ + // First batch + 1, 1, 1, 1, // row = 1 + 2, 2, 2, 2, // row = 2 + // Second batch + 1, 2, 3, 4, // row = 1 + 1, 2, 3, 4, // row = 2 + }); + m.SetFilter({ + 1, 2, 3, 4, // first 2x2 filter + -1, 1, -1, 1, // second 2x2 filter + -1, -1, 1, 1, // third 2x2 filter + }); + m.SetBias({1, 2, 3}); + + m.ApplyDelegateAndInvoke(); + + EXPECT_THAT(m.GetDequantizedOutput(), + ElementsAreArray(ArrayFloatNear( + { + 6, 2, 5, // first batch, left + 6, 2, 5, // first batch, right + 6, 4, 3, // second batch, left + 6, 4, 3, // second batch, right + }, + 2e-2))); +} + TEST(QuantizedConvolutionOpModel, SimplePerTensor_Int8) { QuantizedConvolutionOpModel m( BuiltinOperator_CONV_2D, @@ -512,6 +549,53 @@ TEST(QuantizedConvolutionOpModel, DepthwiseConvSimplePerTensor_Int8) { ElementsAreArray(ArrayFloatNear({43, 48, 40, 52, 3, -4, 4, 4}, 0.6f))); } +TEST(QuantizedConvolutionOpModel, DepthwiseConvSimplePerTensor_Int8_RELU1) { + QuantizedConvolutionOpModel m( + BuiltinOperator_DEPTHWISE_CONV_2D, + {TensorType_INT8, {1, 2, 3, 1}, -63.5, 64, 0.5, -1}, + {TensorType_INT8, + // [1 * 2 * 2 * 4] as [input_channel, y, x, output_channel] + {1, 2, 2, 4}, + 0, + 0, + 0, + 0, + /*per_channel_quantization=*/true, + /*per_channel_quantization_scales=*/{0.1, 2, 3, 0.4}, + /*per_channel_quantization_offsets=*/{0, 0, 0, 0}, + /*channel_index=*/3}, + {TensorType_INT8, {}, -63.5, 64, 0.5, -1}, Padding_VALID, + /**dilation_factor**/ 1, + /**stride**/ 1, ActivationFunctionType_RELU_N1_TO_1); + m.SetInt8Input({ + // [1 * 2 * 3 * 1] as [batch, y, x, input_channel] + 3, // batch = 0, y = 0, x = 0 + 1, // batch = 0, y = 0, x = 1 + -2, // batch = 0, y = 0, x = 2 + 4, // batch = 0, y = 1, x = 0 + 2, // batch = 0, y = 1, x = 1 + -4, // batch = 0, y = 1, x = 2 + }); + m.SetPerChannelQuantizedFilter({ + // [1 * 2 * 2 * 4] as [input_channel, y, x, output_channel] + // depth multiplier = 2 + 1, 2, 3, 4, // y = 0, x = 0 + 3, 4, 5, 6, // y = 0, x = 1 + 7, 8, 5, 6, // y = 1, x = 0 + 3, 4, 1, 2, // y = 1, x = 1 + }); + m.SetPerChannelQuantizedBias({3, -2, 4, 6}); + + // Reference output. + m.Invoke(); + auto reference_output = m.GetDequantizedOutput(); + + m.ApplyDelegateAndInvoke(); + + EXPECT_THAT(m.GetDequantizedOutput(), + ElementsAreArray(ArrayFloatNear(reference_output, 1e-2))); +} + TEST(QuantizedConvolutionOpModel, DepthwiseConvSimplePerAxis_Int8) { QuantizedConvolutionOpModel m( BuiltinOperator_DEPTHWISE_CONV_2D, From 41e7392f58391c18aa872c638ff2e1ac72326bcd Mon Sep 17 00:00:00 2001 From: Rahul Joshi Date: Thu, 18 Jun 2020 13:09:04 -0700 Subject: [PATCH 207/867] - Eliminate all uses of passes that mark function visibility since the visibility is now set correctly when importing. - Update tf_saved_model dialect verification to verify that exported functions are marked public. - Eliminate function_visibility.mlir test. This test fails after the tf_saved_model verification changes since its run tf.entry_function based visibility on a tf_saved_model MLIR module. Also, these passes will be removed. - Fix TPURewritePass to mark the appropriate visibility on the serialized MLIR attached to tf._TPUCompileMlir op. PiperOrigin-RevId: 317165278 Change-Id: I8e8f6de4b56e89c303815edc3b34bcf0a4e82d2d --- .../compiler/mlir/lite/tf_tfl_passes.cc | 3 - tensorflow/compiler/mlir/tensorflow/BUILD | 1 - .../mlir/tensorflow/ir/tf_saved_model.cc | 14 +- .../tensorflow/tests/function_visibility.mlir | 47 ----- .../tf_saved_model_delete_unused_funcs.mlir | 96 ---------- .../tf_saved_model_freeze_global_tensors.mlir | 2 +- .../tensorflow/tests/tf_saved_model_ops.mlir | 2 +- .../tests/tf_saved_model_ops_invalid.mlir | 34 +++- ...timize_global_tensors_interprocedural.mlir | 22 +-- .../transforms/mark_function_visibility.cc | 165 ------------------ .../mlir/tensorflow/transforms/passes.h | 22 --- .../tensorflow/transforms/tpu_rewrite_pass.cc | 3 + .../tensorflow/utils/compile_mlir_util.cc | 3 - tensorflow/compiler/tf2xla/mlir_tf2xla.cc | 5 - 14 files changed, 59 insertions(+), 360 deletions(-) delete mode 100644 tensorflow/compiler/mlir/tensorflow/tests/function_visibility.mlir delete mode 100644 tensorflow/compiler/mlir/tensorflow/tests/tf_saved_model_delete_unused_funcs.mlir delete mode 100644 tensorflow/compiler/mlir/tensorflow/transforms/mark_function_visibility.cc diff --git a/tensorflow/compiler/mlir/lite/tf_tfl_passes.cc b/tensorflow/compiler/mlir/lite/tf_tfl_passes.cc index 589515d6246..008098f62ba 100644 --- a/tensorflow/compiler/mlir/lite/tf_tfl_passes.cc +++ b/tensorflow/compiler/mlir/lite/tf_tfl_passes.cc @@ -212,9 +212,6 @@ void CreateTFLStandardPipeline(OpPassManager& pm, // Saved model pass to mark global tensors immutable. pm.addPass(mlir::tf_saved_model::CreateOptimizeGlobalTensorsPass()); - // Used to mark non-exported functions in saved model private. - pm.addPass(mlir::tf_saved_model:: - CreateMarkFunctionVisibilityUsingSavedModelLinkagePass()); // Op fusion pass. pm.addPass(mlir::TFL::CreatePrepareCompositeFunctionsPass()); diff --git a/tensorflow/compiler/mlir/tensorflow/BUILD b/tensorflow/compiler/mlir/tensorflow/BUILD index 17ed0e36a28..54e57512c32 100644 --- a/tensorflow/compiler/mlir/tensorflow/BUILD +++ b/tensorflow/compiler/mlir/tensorflow/BUILD @@ -491,7 +491,6 @@ cc_library( "transforms/graph_pruning.cc", "transforms/launch_to_device_attribute.cc", "transforms/layout_optimization.cc", - "transforms/mark_function_visibility.cc", "transforms/materialize_mlir_passthrough_op.cc", "transforms/optimize.cc", "transforms/optimize_global_tensors.cc", diff --git a/tensorflow/compiler/mlir/tensorflow/ir/tf_saved_model.cc b/tensorflow/compiler/mlir/tensorflow/ir/tf_saved_model.cc index 6af70158e14..d59532fef65 100644 --- a/tensorflow/compiler/mlir/tensorflow/ir/tf_saved_model.cc +++ b/tensorflow/compiler/mlir/tensorflow/ir/tf_saved_model.cc @@ -229,8 +229,20 @@ static LogicalResult VerifySavedModelModule( } } for (auto func : module.getOps()) { + const bool is_exported = IsExported(func); + + if (is_exported && func.getVisibility() != FuncOp::Visibility::Public) { + return func.emitError() + << "exported function @" << func.getName() << " should be public"; + } + + if (!is_exported && func.getVisibility() == FuncOp::Visibility::Public) { + return func.emitError() << "non-exported function @" << func.getName() + << " should be private"; + } + if (HasAnyTfSavedModelArgAttr(func)) { - if (!IsExported(func)) { + if (!is_exported) { return func.emitError() << "can only apply 'tf_saved_model' argument attributes " "to exported functions"; diff --git a/tensorflow/compiler/mlir/tensorflow/tests/function_visibility.mlir b/tensorflow/compiler/mlir/tensorflow/tests/function_visibility.mlir deleted file mode 100644 index 55af3cffde3..00000000000 --- a/tensorflow/compiler/mlir/tensorflow/tests/function_visibility.mlir +++ /dev/null @@ -1,47 +0,0 @@ -// RUN: tf-opt -tf-saved-model-mark-func-visibility -split-input-file %s | FileCheck --check-prefix=SAVEDMODEL %s -// RUN: tf-opt -tf-mark-func-visibility -split-input-file -verify-diagnostics %s | FileCheck %s - - -module attributes {tf_saved_model.semantics} { - // SAVEDMODEL: func @func_exported_1() attributes {tf_saved_model.exported_names = ["func_exported_1"]} - func @func_exported_1() attributes {tf_saved_model.exported_names = ["func_exported_1"]} { - "tf.some_call"() {callee = {callee = {callee = @child}}} : () -> () - return - } - - // SAVEDMODEL: func @func_exported_2() attributes {tf_saved_model.exported_names = ["func_exported_2"]} - func @func_exported_2() attributes {tf_saved_model.exported_names = ["func_exported_2"]} { - "tf.some_call"() {callee = {callee = {callee = @child}}} : () -> () - return - } - - // SAVEDMODEL: func @func_not_exported() attributes {sym_visibility = "private"} - func @func_not_exported() { - return - } - -} - -// ----- - -module { - // CHECK: func @func_with_entry_spec(%arg0: tensor<1xi32>) -> tensor<1xi32> attributes {tf.entry_function = {inputs = "x", outputs = "y"}} - func @func_with_entry_spec(%arg0: tensor<1xi32>) -> tensor<1xi32> attributes {tf.entry_function = {inputs = "x", outputs = "y"}} { - return %arg0 : tensor<1xi32> - } - - // CHECK: func @func_without_entry_spec(%arg0: tensor<*xi32>, %arg1: tensor<*xi32>) -> tensor<*xi32> attributes {sym_visibility = "private"} - func @func_without_entry_spec(%arg0: tensor<*xi32>, %arg1: tensor<*xi32>) -> tensor<*xi32> { - %0 = "tf.AddV2"(%arg0, %arg1) {T = i32, device = ""} : (tensor<*xi32>, tensor<*xi32>) -> tensor<*xi32> - return %0 : tensor<*xi32> - } -} - -// ----- - -module { - // expected-error @+1 {{can't overwrite the visibility of function private_func_with_entry_spec with private visibility}} - func @private_func_with_entry_spec(%arg0: tensor<1xi32>) -> tensor<1xi32> attributes {tf.entry_function = {inputs = "x", outputs = "y"}, sym_visibility = "private"} { - return %arg0 : tensor<1xi32> - } -} diff --git a/tensorflow/compiler/mlir/tensorflow/tests/tf_saved_model_delete_unused_funcs.mlir b/tensorflow/compiler/mlir/tensorflow/tests/tf_saved_model_delete_unused_funcs.mlir deleted file mode 100644 index 6f2c47a935f..00000000000 --- a/tensorflow/compiler/mlir/tensorflow/tests/tf_saved_model_delete_unused_funcs.mlir +++ /dev/null @@ -1,96 +0,0 @@ -// RUN: tf-opt -tf-saved-model-mark-func-visibility -symbol-dce -split-input-file %s | FileCheck %s - -module attributes {tf_saved_model.semantics} { - - // Test case: Unused function should be deleted. - - // CHECK-NOT: func @unused - func @unused() { - return - } - -} - -// ----- - -module attributes {tf_saved_model.semantics} { - - // Test case: Root calls child. Child should not be deleted. - - // CHECK: func @root - func @root() attributes {tf_saved_model.exported_names = ["root"]} { - "tf.some_call"() { callee = @child } : () -> () - return - } - - // CHECK: func @child - func @child() { - return - } - -} - -// ----- - -module attributes {tf_saved_model.semantics} { - - // Test case: Don't crash if attribute that doesn't reference a func. - - "tf.some_opaque_global_variable"() { sym_name = "some_global" } : () -> () - - func @root2() attributes {tf_saved_model.exported_names = ["root2"]} { - "tf.do_something_with_a_global"() { global = @some_global } : () -> () - return - } - -} - -// ----- - -module attributes {tf_saved_model.semantics} { - - // Test case: Delete recursively dead cycle. - - // CHECK-NOT: func @recursively_dead0 - func @recursively_dead0() { - "tf.some_call"() { callee = @recursively_dead1 } : () -> () - return - } - // CHECK-NOT: func @recursively_dead1 - func @recursively_dead1() { - "tf.some_call"() { callee = @recursively_dead0 } : () -> () - return - } - -} - -// ----- - -module attributes {tf_saved_model.semantics} { - - // Test case: Root calls child with a deeply nested symbol reference. - // Child should not be deleted. - - // CHECK: func @root - func @root() attributes {tf_saved_model.exported_names = ["root"]} { - "tf.some_call"() {callee = {callee = {callee = @child}}} : () -> () - return - } - - // CHECK: func @child - func @child() { - return - } - -} - -// ----- - -// Test case: If the module doesn't have tf_saved_model semantics, then this -// pass shouldn't do anything. -module { - // CHECK: func @not_dead() - func @not_dead() { - return - } -} diff --git a/tensorflow/compiler/mlir/tensorflow/tests/tf_saved_model_freeze_global_tensors.mlir b/tensorflow/compiler/mlir/tensorflow/tests/tf_saved_model_freeze_global_tensors.mlir index 38627b41b68..6c32a3bc4d6 100644 --- a/tensorflow/compiler/mlir/tensorflow/tests/tf_saved_model_freeze_global_tensors.mlir +++ b/tensorflow/compiler/mlir/tensorflow/tests/tf_saved_model_freeze_global_tensors.mlir @@ -64,7 +64,7 @@ module attributes {tf_saved_model.semantics} { return } - func @f_callee(%arg0: tensor>>) { + func @f_callee(%arg0: tensor>>) attributes {sym_visibility = "private"} { return } } diff --git a/tensorflow/compiler/mlir/tensorflow/tests/tf_saved_model_ops.mlir b/tensorflow/compiler/mlir/tensorflow/tests/tf_saved_model_ops.mlir index aa1f996da07..05e7638645f 100644 --- a/tensorflow/compiler/mlir/tensorflow/tests/tf_saved_model_ops.mlir +++ b/tensorflow/compiler/mlir/tensorflow/tests/tf_saved_model_ops.mlir @@ -40,7 +40,7 @@ module attributes {tf_saved_model.semantics} { return %arg0 : tensor } - func @f() { + func @f() attributes {sym_visibility = "private"} { return } diff --git a/tensorflow/compiler/mlir/tensorflow/tests/tf_saved_model_ops_invalid.mlir b/tensorflow/compiler/mlir/tensorflow/tests/tf_saved_model_ops_invalid.mlir index 544600cf6b8..f04e1a60b36 100644 --- a/tensorflow/compiler/mlir/tensorflow/tests/tf_saved_model_ops_invalid.mlir +++ b/tensorflow/compiler/mlir/tensorflow/tests/tf_saved_model_ops_invalid.mlir @@ -3,7 +3,7 @@ module attributes {tf_saved_model.semantics} { // expected-error@+1 {{unknown tf_saved_model dialect arg attribute 'tf_saved_model.not_a_real_arg_attr'}} - func @f(%arg0: tensor {tf_saved_model.not_a_real_arg_attr = 1 : i32}) { + func @f(%arg0: tensor {tf_saved_model.not_a_real_arg_attr = 1 : i32}) attributes {sym_visibility = "private"} { return } @@ -233,7 +233,7 @@ module attributes {tf_saved_model.semantics} { "tf_saved_model.global_tensor"() { is_mutable, sym_name = "v", type = tensor, value = dense<1.> : tensor<1xf32> } : () -> () // expected-error@+1 {{can only apply 'tf_saved_model' argument attributes to exported functions}} func @f(%arg0: tensor>> {tf_saved_model.bound_input = @v}) - -> (tensor {tf_saved_model.index_path = []}) { + -> (tensor {tf_saved_model.index_path = []}) attributes {sym_visibility = "private"} { %0 = "tf.ReadVariableOp"(%arg0) : (tensor>>) -> tensor return %0 : tensor } @@ -273,7 +273,7 @@ module attributes {tf_saved_model.semantics} { // expected-error@+1 {{the initializer function should have no output}} "tf_saved_model.session_initializer"() { initializer = @init } : () -> () - func @init() -> tensor<1xf32> { + func @init() -> tensor<1xf32> attributes {sym_visibility = "private"} { %0 = "tf.Const"() {value = dense<[1.0]> : tensor<1xf32> } : () -> tensor<1xf32> return %0 : tensor<1xf32> } @@ -286,8 +286,34 @@ module attributes {tf_saved_model.semantics} { "tf_saved_model.session_initializer"() { initializer = @init } : () -> () // expected-error@+1 {{there must be no more than one session_initializer op}} "tf_saved_model.session_initializer"() { initializer = @init } : () -> () - func @init() -> tensor<1xf32> { + func @init() -> tensor<1xf32> attributes {sym_visibility = "private"} { %0 = "tf.Const"() {value = dense<[1.0]> : tensor<1xf32> } : () -> tensor<1xf32> return %0 : tensor<1xf32> } } + +// ----- + +module attributes {tf_saved_model.semantics} { + + // expected-error@+1 {{exported function @f should be public}} + func @f( + %arg0: tensor {tf.resource_name = "resource"} + ) attributes { sym_visibility = "private", tf_saved_model.exported_names = ["foo.some_func"] } { + return + } + +} + +// ----- + +module attributes {tf_saved_model.semantics} { + + // expected-error@+1 {{non-exported function @f should be private}} + func @f( + %arg0: tensor {tf.resource_name = "resource"} + ) { + return + } + +} diff --git a/tensorflow/compiler/mlir/tensorflow/tests/tf_saved_model_optimize_global_tensors_interprocedural.mlir b/tensorflow/compiler/mlir/tensorflow/tests/tf_saved_model_optimize_global_tensors_interprocedural.mlir index 91e8c9c4b66..14a0006cd3b 100644 --- a/tensorflow/compiler/mlir/tensorflow/tests/tf_saved_model_optimize_global_tensors_interprocedural.mlir +++ b/tensorflow/compiler/mlir/tensorflow/tests/tf_saved_model_optimize_global_tensors_interprocedural.mlir @@ -20,12 +20,12 @@ module attributes {tf_saved_model.semantics} { return %val : tensor } - func @f_callee(%arg0: tensor<*x!tf.resource>) -> tensor { + func @f_callee(%arg0: tensor<*x!tf.resource>) -> tensor attributes {sym_visibility = "private"} { %val = "tf.PartitionedCall"(%arg0) {config = "", config_proto = "", executor_type = "", f = @f_callee_callee} : (tensor<*x!tf.resource>) -> (tensor) return %val : tensor } - func @f_callee_callee(%arg0: tensor<*x!tf.resource>) -> tensor { + func @f_callee_callee(%arg0: tensor<*x!tf.resource>) -> tensor attributes {sym_visibility = "private"} { %val = "tf.ReadVariableOp"(%arg0) : (tensor<*x!tf.resource>) -> tensor return %val : tensor } @@ -59,7 +59,7 @@ module attributes {tf_saved_model.semantics} { return %val : tensor } - func @f_common(%arg0: tensor<*x!tf.resource>) -> tensor { + func @f_common(%arg0: tensor<*x!tf.resource>) -> tensor attributes {sym_visibility = "private"} { %val = "tf.ReadVariableOp"(%arg0) : (tensor<*x!tf.resource>) -> tensor return %val : tensor } @@ -85,7 +85,7 @@ module attributes {tf_saved_model.semantics} { return %val_2 : tensor } - func @f_callee(%arg0: tensor<*x!tf.resource>) -> tensor { + func @f_callee(%arg0: tensor<*x!tf.resource>) -> tensor attributes {sym_visibility = "private"} { %cst_1 = constant dense<2.0> : tensor return %cst_1 : tensor } @@ -112,13 +112,13 @@ module attributes {tf_saved_model.semantics} { } // CHECK: func @f_callee(%arg0: tensor<*x!tf.resource>) -> tensor - func @f_callee(%arg0: tensor<*x!tf.resource>) -> tensor { + func @f_callee(%arg0: tensor<*x!tf.resource>) -> tensor attributes {sym_visibility = "private"} { %val = "tf.PartitionedCall"(%arg0) {config = "", config_proto = "", executor_type = "", f = @f_callee_callee} : (tensor<*x!tf.resource>) -> (tensor) return %val : tensor } // CHECK: func @f_callee_callee(%arg0: tensor<*x!tf.resource>) -> tensor - func @f_callee_callee(%arg0: tensor<*x!tf.resource>) -> tensor { + func @f_callee_callee(%arg0: tensor<*x!tf.resource>) -> tensor attributes {sym_visibility = "private"} { %c0 = "tf.Const"() { value = dense<1.0> : tensor } : () -> tensor "tf.AssignVariableOp"(%arg0, %c0) : (tensor<*x!tf.resource>, tensor) -> () return %c0 : tensor @@ -146,13 +146,13 @@ module attributes {tf_saved_model.semantics} { } // CHECK: func @f_callee(%arg0: tensor<*x!tf.resource>) -> tensor - func @f_callee(%arg0: tensor<*x!tf.resource>) -> tensor { + func @f_callee(%arg0: tensor<*x!tf.resource>) -> tensor attributes {sym_visibility = "private"} { %val = "tf.PartitionedCall"(%arg0) {config = "", config_proto = "", executor_type = "", f = @f_callee_callee} : (tensor<*x!tf.resource>) -> (tensor) return %val : tensor } // CHECK: func @f_callee_callee(%arg0: tensor<*x!tf.resource>) -> tensor - func @f_callee_callee(%arg0: tensor<*x!tf.resource>) -> tensor { + func @f_callee_callee(%arg0: tensor<*x!tf.resource>) -> tensor attributes {sym_visibility = "private"} { %c0 = "tf.Const"() { value = dense<1.0> : tensor } : () -> tensor "tf.AssignVariableOp"(%arg0, %c0) : (tensor<*x!tf.resource>, tensor) -> () return %c0 : tensor @@ -179,13 +179,13 @@ module attributes {tf_saved_model.semantics} { // CHECK: func @f(%arg0: tensor<*x!tf.resource>) -> tensor - func @f(%arg0: tensor<*x!tf.resource>) -> tensor { + func @f(%arg0: tensor<*x!tf.resource>) -> tensor attributes {sym_visibility = "private"} { %val = "tf.PartitionedCall"(%arg0) {config = "", config_proto = "", executor_type = "", f = @g} : (tensor<*x!tf.resource>) -> (tensor) return %val : tensor } // CHECK: func @g(%arg0: tensor<*x!tf.resource>) -> tensor - func @g(%arg0: tensor<*x!tf.resource>) -> tensor { + func @g(%arg0: tensor<*x!tf.resource>) -> tensor attributes {sym_visibility = "private"} { %val = "tf.PartitionedCall"(%arg0) {config = "", config_proto = "", executor_type = "", f = @f} : (tensor<*x!tf.resource>) -> (tensor) return %val : tensor } @@ -212,7 +212,7 @@ module attributes {tf_saved_model.semantics} { // CHECK: func @f(%arg0: tensor<*x!tf.resource>) -> tensor - func @f(%arg0: tensor<*x!tf.resource>) -> tensor { + func @f(%arg0: tensor<*x!tf.resource>) -> tensor attributes {sym_visibility = "private"} { %c0 = "tf.Const"() { value = dense<1.0> : tensor } : () -> tensor "tf.AssignAddVariableOp"(%arg0, %c0) : (tensor<*x!tf.resource>, tensor) -> () return %c0 : tensor diff --git a/tensorflow/compiler/mlir/tensorflow/transforms/mark_function_visibility.cc b/tensorflow/compiler/mlir/tensorflow/transforms/mark_function_visibility.cc deleted file mode 100644 index 31a80a4ecdb..00000000000 --- a/tensorflow/compiler/mlir/tensorflow/transforms/mark_function_visibility.cc +++ /dev/null @@ -1,165 +0,0 @@ -/* Copyright 2020 The TensorFlow Authors. All Rights Reserved. - -Licensed under the Apache License, Version 2.0 (the "License"); -you may not use this file except in compliance with the License. -You may obtain a copy of the License at - - http://www.apache.org/licenses/LICENSE-2.0 - -Unless required by applicable law or agreed to in writing, software -distributed under the License is distributed on an "AS IS" BASIS, -WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -See the License for the specific language governing permissions and -limitations under the License. -==============================================================================*/ - -#include "llvm/ADT/STLExtras.h" -#include "mlir/IR/Module.h" // from @llvm-project -#include "mlir/Pass/Pass.h" // from @llvm-project -#include "tensorflow/compiler/mlir/tensorflow/ir/tf_saved_model.h" - -#define DEBUG_TYPE "tf-shape-inference" - -namespace mlir { - -namespace { - -LogicalResult MarkFunctionVisibility( - ModuleOp module, llvm::function_ref IsExternalVisible) { - LogicalResult result = success(); - - for (auto func : module.getOps()) { - FuncOp::Visibility old_visibility = func.getVisibility(); - - FuncOp::Visibility visibility = IsExternalVisible(func) - ? FuncOp::Visibility::Public - : FuncOp::Visibility::Private; - - auto get_visibility_name = [](FuncOp::Visibility v) { - return v == FuncOp::Visibility::Public - ? "public" - : v == FuncOp::Visibility::Private ? "private" : "nested"; - }; - - if (old_visibility != SymbolTable::Visibility::Public && - old_visibility != visibility) { - result = func.emitError() - << "can't overwrite the visibility of function " - << func.getName() << " with " - << get_visibility_name(old_visibility) << " visibility"; - } - - LLVM_DEBUG(llvm::dbgs() - << "function " << func.getName() << " has " - << get_visibility_name(visibility) << " visibility \n"); - - func.setVisibility(visibility); - } - - return result; -} - -} // anonymous namespace - -namespace TF { - -LogicalResult MarkFunctionVisibilityUsingEntryFunctionSpecification( - ModuleOp module) { - auto HasEntryFunctionSpecification = [](FuncOp func) -> bool { - auto attrs = func.getAttrOfType("tf.entry_function"); - return attrs && !attrs.empty(); - }; - return MarkFunctionVisibility(module, HasEntryFunctionSpecification); -} - -namespace { -struct MarkFunctionVisibilityUsingEntryFunctionSpecificationPass - : public PassWrapper< - MarkFunctionVisibilityUsingEntryFunctionSpecificationPass, - OperationPass> { - void runOnOperation() override { - if (failed(MarkFunctionVisibilityUsingEntryFunctionSpecification( - getOperation()))) { - signalPassFailure(); - } - } -}; -} // namespace - -static PassRegistration< - MarkFunctionVisibilityUsingEntryFunctionSpecificationPass> - pass("tf-mark-func-visibility", - "Use tf.entry_function to mark function visibility."); - -std::unique_ptr> -CreateMarkFunctionVisibilityUsingEntryFunctionSpecificationPass() { - return std::make_unique< - MarkFunctionVisibilityUsingEntryFunctionSpecificationPass>(); -} - -// Marks the main function with public visibility, while other functions are -// marked with private visibility. -LogicalResult MarkOnlyMainFunctionWithPublicVisibility(ModuleOp module) { - for (auto func : module.getOps()) { - if (func.getName() == "main") { - func.setVisibility(FuncOp::Visibility::Public); - } else { - func.setVisibility(FuncOp::Visibility::Private); - } - } - return success(); -} - -namespace { -struct MarkOnlyMainFunctionWithPublicVisibilityPass - : public PassWrapper> { - void runOnOperation() override { - if (failed(MarkOnlyMainFunctionWithPublicVisibility(getOperation()))) { - signalPassFailure(); - } - } -}; -} // namespace - -std::unique_ptr> -CreateMarkOnlyMainFunctionWithPublicVisibilityPass() { - return std::make_unique(); -} - -} // namespace TF - -namespace tf_saved_model { - -static LogicalResult MarkFunctionVisibilityUsingSavedModelLinkage( - ModuleOp module) { - if (!tf_saved_model::HasTfSavedModelSemantics(module)) { - return success(); - } - return MarkFunctionVisibility(module, tf_saved_model::IsExported); -} - -namespace { -struct MarkFunctionVisibilityUsingSavedModelLinkagePass - : public PassWrapper> { - void runOnOperation() override { - if (failed(MarkFunctionVisibilityUsingSavedModelLinkage(getOperation()))) { - signalPassFailure(); - } - } -}; -} // namespace - -static PassRegistration pass( - "tf-saved-model-mark-func-visibility", - "Use tf_saved_model linkage information to mark function visibility."); - -std::unique_ptr> -CreateMarkFunctionVisibilityUsingSavedModelLinkagePass() { - return std::make_unique(); -} - -} // namespace tf_saved_model - -} // namespace mlir diff --git a/tensorflow/compiler/mlir/tensorflow/transforms/passes.h b/tensorflow/compiler/mlir/tensorflow/transforms/passes.h index 7158d0f6be0..5cb15027fc5 100644 --- a/tensorflow/compiler/mlir/tensorflow/transforms/passes.h +++ b/tensorflow/compiler/mlir/tensorflow/transforms/passes.h @@ -117,21 +117,6 @@ std::unique_ptr> CreatePromoteVarHandlesToArgsPass(); std::unique_ptr> CreateConvertReadonlyReferenceVariablesToResourceVariablesPass(); -// Marks function visibility using tf.entry_function specification. That is, -// functions with tf.entry_function attributes are marked with public -// visibility while the other functions are marked with private visibility. -LogicalResult MarkFunctionVisibilityUsingEntryFunctionSpecification( - ModuleOp module); -// Creates a pass that uses tf.entry_function specification to mark function -// visibility. -std::unique_ptr> -CreateMarkFunctionVisibilityUsingEntryFunctionSpecificationPass(); - -// Creates a pass that marks the main function with public visibility, while -// other functions are marked with private visibility. -std::unique_ptr> -CreateMarkOnlyMainFunctionWithPublicVisibilityPass(); - // Creates a simple device assignment pass on TF dialect for CoreRT use case. std::unique_ptr> CreateSimpleTFDeviceAssignmentPass( llvm::StringRef default_device); @@ -315,13 +300,6 @@ std::unique_ptr> CreateOptimizeGlobalTensorsPass(); // Creates a pass that freezes tf_saved_model.global_tensor ops. std::unique_ptr> CreateFreezeGlobalTensorsPass(); -// Creates a pass that uses tf_saved_model dialect linkage information -// to mark function visibility. That is, exported functions are marked with -// public visibility while the other functions are marked with private -// visibility. -std::unique_ptr> -CreateMarkFunctionVisibilityUsingSavedModelLinkagePass(); - } // namespace tf_saved_model } // namespace mlir diff --git a/tensorflow/compiler/mlir/tensorflow/transforms/tpu_rewrite_pass.cc b/tensorflow/compiler/mlir/tensorflow/transforms/tpu_rewrite_pass.cc index 696882cd105..ec9b3df525f 100644 --- a/tensorflow/compiler/mlir/tensorflow/transforms/tpu_rewrite_pass.cc +++ b/tensorflow/compiler/mlir/tensorflow/transforms/tpu_rewrite_pass.cc @@ -146,6 +146,9 @@ LogicalResult EncapsulateFuncAndSerialize(FuncOp entry_func, // We can simply change name of TPU program's main function because there // should be no other reference to it. clone.setName("main"); + clone.setVisibility(FuncOp::Visibility::Public); + } else { + clone.setVisibility(FuncOp::Visibility::Private); } symbol_table.insert(clone); } diff --git a/tensorflow/compiler/mlir/tensorflow/utils/compile_mlir_util.cc b/tensorflow/compiler/mlir/tensorflow/utils/compile_mlir_util.cc index fd1ba3b1901..dac2fea87e2 100644 --- a/tensorflow/compiler/mlir/tensorflow/utils/compile_mlir_util.cc +++ b/tensorflow/compiler/mlir/tensorflow/utils/compile_mlir_util.cc @@ -267,9 +267,6 @@ Status ConvertMLIRToXlaComputation( const XlaCompiler::ShapeRepresentationFn shape_representation_fn, std::vector> custom_legalization_passes) { mlir::PassManager tf2xla(module_op.getContext()); - // Mark main function as public, and other functions as private. - tf2xla.addPass( - mlir::TF::CreateMarkOnlyMainFunctionWithPublicVisibilityPass()); tf2xla.addNestedPass(mlir::createCanonicalizerPass()); tf2xla.addPass(mlir::TF::CreateTensorListOpsDecompositionPass()); tf2xla.addPass(mlir::TF::CreateStackOpsDecompositionPass()); diff --git a/tensorflow/compiler/tf2xla/mlir_tf2xla.cc b/tensorflow/compiler/tf2xla/mlir_tf2xla.cc index 43793be56a7..60d1f3da0c5 100644 --- a/tensorflow/compiler/tf2xla/mlir_tf2xla.cc +++ b/tensorflow/compiler/tf2xla/mlir_tf2xla.cc @@ -165,11 +165,6 @@ Status ConvertGraphDefToXlaViaMlir( device_set.AddDevice(&device); AddDevicesToOp(*module, &device_set); - if (failed(mlir::TF::MarkFunctionVisibilityUsingEntryFunctionSpecification( - *module))) { - return errors::Internal("Problem with mark function visibility"); - } - TF_RETURN_IF_ERROR(mlir::TF::RunBridgeWithStandardPipeline( *module, /*enable_logging=*/VLOG_IS_ON(1), /*enable_inliner=*/true)); From ededf6f4b9d1488c1d27df58b047fd5da6ad6c73 Mon Sep 17 00:00:00 2001 From: Sharada Shiddibhavi Date: Thu, 18 Jun 2020 13:14:37 -0700 Subject: [PATCH 208/867] Update tensorflow/core/util/mkl_util.h Addressing review comments Co-authored-by: Penporn Koanantakool <38085909+penpornk@users.noreply.github.com> --- tensorflow/core/util/mkl_util.h | 9 ++------- 1 file changed, 2 insertions(+), 7 deletions(-) diff --git a/tensorflow/core/util/mkl_util.h b/tensorflow/core/util/mkl_util.h index 996984eebc0..854d6e349cb 100644 --- a/tensorflow/core/util/mkl_util.h +++ b/tensorflow/core/util/mkl_util.h @@ -1538,13 +1538,8 @@ class MklDnnData { /// Set function for data buffer of user memory primitive. inline void SetUsrMemDataHandle(const Tensor* tensor, std::shared_ptr t_stream = nullptr) { - CHECK_NOTNULL(user_memory_); - CHECK_NOTNULL(tensor); -#ifdef ENABLE_MKLDNN_THREADPOOL - user_memory_->set_data_handle(GetTensorBuffer(tensor), *t_stream); -#else - user_memory_->set_data_handle(GetTensorBuffer(tensor)); -#endif // ENABLE_MKLDNN_THREADPOOL + SetUsrMemDataHandle(GetTensorBuffer(tensor), t_stream); + } } /// allocate function for data buffer From 51f3da9ca884951be412bccb766c3700ba2255f3 Mon Sep 17 00:00:00 2001 From: Yunxing Dai Date: Thu, 18 Jun 2020 13:15:43 -0700 Subject: [PATCH 209/867] [Resubmit] Remove dynamic dimension of strided slice grad if input to strided slice is static. If we slice a dynamic shaped tensor from a static tensor, the output of the gradient should still be static. Unfortunately this cannot be deduced alone by xla, so extra information is needed from the tf2xla bridge. PiperOrigin-RevId: 317166566 Change-Id: Ic3a16826242947a29cafd51b2f5c19e65d531fb9 --- .../tf2xla/kernels/strided_slice_op.cc | 24 +++++++++++++++++ tensorflow/compiler/xla/client/xla_builder.cc | 23 ++++++++++++++++ tensorflow/compiler/xla/client/xla_builder.h | 6 +++++ .../compiler/xla/client/xla_builder_test.cc | 26 +++++++++++++++++++ 4 files changed, 79 insertions(+) diff --git a/tensorflow/compiler/tf2xla/kernels/strided_slice_op.cc b/tensorflow/compiler/tf2xla/kernels/strided_slice_op.cc index 2684c982600..784b790767c 100644 --- a/tensorflow/compiler/tf2xla/kernels/strided_slice_op.cc +++ b/tensorflow/compiler/tf2xla/kernels/strided_slice_op.cc @@ -350,6 +350,30 @@ class StridedSliceGradOp : public XlaOpKernel { grad = xla::Rev(grad, dimensions_to_reverse); } grad = xla::Pad(grad, zero, padding_config); + + xla::XlaOp dynamic_shape = ctx->Input(0); + xla::Shape grad_shape = ctx->builder()->GetShape(grad).ValueOrDie(); + ctx->set_dynamic_dimension_is_minus_one(true); + std::vector dynamic_size; + OP_REQUIRES_OK(ctx, ctx->ConstantInputAsIntVector(0, &dynamic_size)); + // Input of strided_slice_op has to have the same shape as output. + DCHECK_EQ(grad_shape.rank(), input_shape.dims()); + for (int64 dim = 0; dim < input_shape.dims(); ++dim) { + DCHECK_EQ(grad_shape.dimensions(dim), input_shape.dim_size(dim)); + if (dynamic_size[dim] == -1) { + // Input is a dynamic dimension, set the same dynamic dimension size in + // the output. + auto dim_size = xla::Slice(dynamic_shape, {dim}, {dim + 1}, {1}); + auto dim_size_scalar = + xla::Reshape(xla::ShapeUtil::MakeScalarShape(xla::S32), dim_size); + grad = xla::SetDimensionSize(grad, dim_size_scalar, dim); + } else if (grad_shape.is_dynamic_dimension(dim)) { + // Input is static but output is dynamic, respect input and remove any + // dynamic dim in the output. + grad = xla::RemoveDynamicDimension(grad, dim); + } + } + ctx->SetOutput(0, grad); } diff --git a/tensorflow/compiler/xla/client/xla_builder.cc b/tensorflow/compiler/xla/client/xla_builder.cc index bfba48862f6..c7b6a7f9491 100644 --- a/tensorflow/compiler/xla/client/xla_builder.cc +++ b/tensorflow/compiler/xla/client/xla_builder.cc @@ -2727,6 +2727,25 @@ XlaOp XlaBuilder::GetDimensionSize(XlaOp operand, int64 dimension) { }); } +XlaOp XlaBuilder::RemoveDynamicDimension(XlaOp operand, int64 dimension) { + return ReportErrorOrReturn([&]() -> StatusOr { + HloInstructionProto instr; + TF_ASSIGN_OR_RETURN(const Shape* operand_shape, GetShapePtr(operand)); + + Shape shape = *operand_shape; + shape.set_dynamic_dimension(dimension, false); + // Setting an op's dynamic dimension to its static size removes the dynamic + // dimension. + XlaOp static_size = + ConstantR0(this, operand_shape->dimensions(dimension)); + + *instr.mutable_shape() = shape.ToProto(); + instr.add_dimensions(dimension); + return AddInstruction(std::move(instr), HloOpcode::kSetDimensionSize, + {operand, static_size}); + }); +} + XlaOp XlaBuilder::SetDimensionSize(XlaOp operand, XlaOp val, int64 dimension) { return ReportErrorOrReturn([&]() -> StatusOr { HloInstructionProto instr; @@ -3827,4 +3846,8 @@ XlaOp SetDimensionSize(const XlaOp operand, const XlaOp val, int64 dimension) { return operand.builder()->SetDimensionSize(operand, val, dimension); } +XlaOp RemoveDynamicDimension(const XlaOp operand, int64 dimension) { + return operand.builder()->RemoveDynamicDimension(operand, dimension); +} + } // namespace xla diff --git a/tensorflow/compiler/xla/client/xla_builder.h b/tensorflow/compiler/xla/client/xla_builder.h index ffa6a7c3439..b8af180b83e 100644 --- a/tensorflow/compiler/xla/client/xla_builder.h +++ b/tensorflow/compiler/xla/client/xla_builder.h @@ -704,6 +704,8 @@ class XlaBuilder { XlaOp SetDimensionSize(XlaOp operand, XlaOp val, int64 dimension); + XlaOp RemoveDynamicDimension(XlaOp operand, int64 dimension); + StatusOr AddInstruction(HloInstructionProto&& instr, HloOpcode opcode, absl::Span operands = {}); @@ -1151,6 +1153,7 @@ class XlaBuilder { friend XlaOp GetDimensionSize(XlaOp operand, int64 dimension); friend XlaOp SetDimensionSize(XlaOp operand, XlaOp val, int64 dimension); + friend XlaOp RemoveDynamicDimension(XlaOp operand, int64 dimension); protected: // Returns OK status if the given op was built using this builder. Otherwise, @@ -2149,6 +2152,9 @@ XlaOp GetDimensionSize(XlaOp operand, int64 dimension); XlaOp SetDimensionSize(XlaOp operand, XlaOp val, int64 dimension); +// Returns the same op but with dynamic dimension removed. +XlaOp RemoveDynamicDimension(XlaOp operand, int64 dimension); + // Implementation details below this point. // diff --git a/tensorflow/compiler/xla/client/xla_builder_test.cc b/tensorflow/compiler/xla/client/xla_builder_test.cc index 4fa47077fca..7011c946203 100644 --- a/tensorflow/compiler/xla/client/xla_builder_test.cc +++ b/tensorflow/compiler/xla/client/xla_builder_test.cc @@ -556,6 +556,32 @@ TEST_F(XlaBuilderTest, DynamicParameter) { EXPECT_TRUE(param_shape.is_dynamic_dimension(0)); } +TEST_F(XlaBuilderTest, SetDimensionSize) { + XlaBuilder b(TestName()); + auto p0 = Parameter(&b, 0, ShapeUtil::MakeShape(F32, {10}), "p0"); + auto p1 = Parameter(&b, 1, ShapeUtil::MakeShape(S32, {}), "p1"); + auto set_dim_size = SetDimensionSize(p0, p1, 0); + TF_ASSERT_OK_AND_ASSIGN(auto module, + BuildHloModule(&b, /*root=*/set_dim_size)); + const Shape& root_shape = + module->entry_computation()->root_instruction()->shape(); + EXPECT_TRUE(root_shape.is_dynamic_dimension(0)); +} + +TEST_F(XlaBuilderTest, RemoveDimensionSize) { + XlaBuilder b(TestName()); + auto p0 = Parameter(&b, 0, ShapeUtil::MakeShape(F32, {10}), "p0"); + auto p1 = Parameter(&b, 1, ShapeUtil::MakeShape(S32, {}), "p1"); + auto set_dim_size = SetDimensionSize(p0, p1, 0); + auto remove_dim_size = RemoveDynamicDimension(set_dim_size, 0); + TF_ASSERT_OK_AND_ASSIGN(auto module, + BuildHloModule(&b, /*root=*/remove_dim_size)); + const Shape& root_shape = + module->entry_computation()->root_instruction()->shape(); + // Dynamic dimension has been removed. + EXPECT_FALSE(root_shape.is_dynamic_dimension(0)); +} + TEST_F(XlaBuilderTest, DynamicUnary) { XlaBuilder b(TestName()); Shape tuple_param_shape = ShapeUtil::MakeTupleShape( From 8944a3eeb18c2374f02759324cbeded018c5868b Mon Sep 17 00:00:00 2001 From: Dan Moldovan Date: Thu, 18 Jun 2020 13:23:05 -0700 Subject: [PATCH 210/867] Enable type annotations for python/autograph. PiperOrigin-RevId: 317168016 Change-Id: I57341cc0347ab8eb008949e191da1415069203b7 --- tensorflow/python/autograph/pyct/parser.py | 16 ++++++++++++---- 1 file changed, 12 insertions(+), 4 deletions(-) diff --git a/tensorflow/python/autograph/pyct/parser.py b/tensorflow/python/autograph/pyct/parser.py index 747d56e401d..9ac7c2ef2a6 100644 --- a/tensorflow/python/autograph/pyct/parser.py +++ b/tensorflow/python/autograph/pyct/parser.py @@ -22,6 +22,7 @@ from __future__ import division from __future__ import print_function import re +import sys import textwrap import tokenize @@ -33,11 +34,18 @@ from tensorflow.python.autograph.pyct import errors from tensorflow.python.autograph.pyct import inspect_utils -STANDARD_PREAMBLE = textwrap.dedent(""" - from __future__ import division - from __future__ import print_function +PY2_PREAMBLE = textwrap.dedent(""" +from __future__ import division +from __future__ import print_function """) -STANDARD_PREAMBLE_LEN = 2 +PY3_PREAMBLE = '' + +if sys.version_info >= (3,): + STANDARD_PREAMBLE = PY3_PREAMBLE +else: + STANDARD_PREAMBLE = PY2_PREAMBLE + +STANDARD_PREAMBLE_LEN = STANDARD_PREAMBLE.count('__future__') _LEADING_WHITESPACE = re.compile(r'\s*') From f8431d0c293a34c5bfb91cf0c57384eaa47a9911 Mon Sep 17 00:00:00 2001 From: Sharada Shiddibhavi Date: Thu, 18 Jun 2020 13:29:48 -0700 Subject: [PATCH 211/867] Update mkl_util.h --- tensorflow/core/util/mkl_util.h | 1 - 1 file changed, 1 deletion(-) diff --git a/tensorflow/core/util/mkl_util.h b/tensorflow/core/util/mkl_util.h index 854d6e349cb..eb1a105e07c 100644 --- a/tensorflow/core/util/mkl_util.h +++ b/tensorflow/core/util/mkl_util.h @@ -1540,7 +1540,6 @@ class MklDnnData { std::shared_ptr t_stream = nullptr) { SetUsrMemDataHandle(GetTensorBuffer(tensor), t_stream); } - } /// allocate function for data buffer inline void AllocateBuffer(size_t size) { From 2663edb6691b50de8ca9445c311d02b72faa1bf5 Mon Sep 17 00:00:00 2001 From: Ken Franko Date: Thu, 18 Jun 2020 13:29:28 -0700 Subject: [PATCH 212/867] Change TPUExtractOutsideCompilation pass for a Module pass. This is needed for getting the devices from the module for assigning host device for outside compilation launch op. PiperOrigin-RevId: 317169244 Change-Id: I734e7eeef3fdb037045d070ffd736be4ef8edee1 --- .../compiler/mlir/tensorflow/transforms/passes.h | 3 ++- .../transforms/tpu_extract_outside_compilation.cc | 11 ++++++----- 2 files changed, 8 insertions(+), 6 deletions(-) diff --git a/tensorflow/compiler/mlir/tensorflow/transforms/passes.h b/tensorflow/compiler/mlir/tensorflow/transforms/passes.h index 5cb15027fc5..a34be28c809 100644 --- a/tensorflow/compiler/mlir/tensorflow/transforms/passes.h +++ b/tensorflow/compiler/mlir/tensorflow/transforms/passes.h @@ -281,7 +281,8 @@ std::unique_ptr> CreateTPUHostComputationExpansionPass(); // Creates a pass that extract outside compilation (CPU ops inside TPU cluster) // ops to a separate parallel_execute region to run on CPU. -std::unique_ptr> CreateTPUExtractOutsideCompilationPass(); +std::unique_ptr> +CreateTPUExtractOutsideCompilationPass(); // Populates the supplied passmanager with the passes required to run the void CreateTPUBridgePipeline(OpPassManager& pm); diff --git a/tensorflow/compiler/mlir/tensorflow/transforms/tpu_extract_outside_compilation.cc b/tensorflow/compiler/mlir/tensorflow/transforms/tpu_extract_outside_compilation.cc index a2a19108326..503c9869557 100644 --- a/tensorflow/compiler/mlir/tensorflow/transforms/tpu_extract_outside_compilation.cc +++ b/tensorflow/compiler/mlir/tensorflow/transforms/tpu_extract_outside_compilation.cc @@ -49,8 +49,9 @@ using OutsideClusterMap = // TODO(b/154363171): Add example tranformations. struct TPUExtractOutsideCompilation - : public PassWrapper { - void runOnFunction() override; + : public PassWrapper> { + void runOnOperation() override; }; // Collects and clusters ops in `block` with the same `_xla_outside_compilation` @@ -305,9 +306,9 @@ void CreateParallelExecuteFromOutsideClusters( } } -void TPUExtractOutsideCompilation::runOnFunction() { +void TPUExtractOutsideCompilation::runOnOperation() { auto extract_result = - getFunction().walk([&](tf_device::ClusterOp tpu_cluster) { + getOperation().walk([&](tf_device::ClusterOp tpu_cluster) { OutsideClusterMap clusters; if (failed(CollectAndGroupOutsideClusterOps(&tpu_cluster.GetBody(), &clusters))) @@ -325,7 +326,7 @@ void TPUExtractOutsideCompilation::runOnFunction() { } // namespace -std::unique_ptr> +std::unique_ptr> CreateTPUExtractOutsideCompilationPass() { return std::make_unique(); } From 6558da5a66ad6863e47abfe596eee2290524b1b7 Mon Sep 17 00:00:00 2001 From: Jiho Choi Date: Thu, 18 Jun 2020 13:30:08 -0700 Subject: [PATCH 213/867] Apply new TraceMe APIs. PiperOrigin-RevId: 317169381 Change-Id: I2259895a8dde21e25661a239b9d4f5911a454adb --- tensorflow/compiler/xla/pjrt/BUILD | 2 ++ tensorflow/compiler/xla/pjrt/pjrt_client.cc | 16 ++++++++-------- tensorflow/core/profiler/lib/connected_traceme.h | 1 + 3 files changed, 11 insertions(+), 8 deletions(-) diff --git a/tensorflow/compiler/xla/pjrt/BUILD b/tensorflow/compiler/xla/pjrt/BUILD index dd50d0577d4..e401a798d68 100644 --- a/tensorflow/compiler/xla/pjrt/BUILD +++ b/tensorflow/compiler/xla/pjrt/BUILD @@ -141,7 +141,9 @@ cc_library( "//tensorflow/compiler/xla/service/gpu:gpu_executable_run_options", "//tensorflow/core:allocator", "//tensorflow/core:lib", + "//tensorflow/core/profiler/lib:connected_traceme", "//tensorflow/core/profiler/lib:traceme", + "//tensorflow/core/profiler/lib:traceme_encode", "//tensorflow/stream_executor:event", "//tensorflow/stream_executor:stream", "//tensorflow/stream_executor/host:host_platform_id", diff --git a/tensorflow/compiler/xla/pjrt/pjrt_client.cc b/tensorflow/compiler/xla/pjrt/pjrt_client.cc index ccb72b7ce30..ef259cf1cfd 100644 --- a/tensorflow/compiler/xla/pjrt/pjrt_client.cc +++ b/tensorflow/compiler/xla/pjrt/pjrt_client.cc @@ -98,7 +98,9 @@ limitations under the License. #include "tensorflow/core/platform/mem.h" #include "tensorflow/core/platform/status.h" #include "tensorflow/core/platform/types.h" +#include "tensorflow/core/profiler/lib/connected_traceme.h" #include "tensorflow/core/profiler/lib/traceme.h" +#include "tensorflow/core/profiler/lib/traceme_encode.h" #include "tensorflow/stream_executor/device_memory.h" #include "tensorflow/stream_executor/device_memory_allocator.h" #include "tensorflow/stream_executor/event.h" @@ -1429,10 +1431,9 @@ StatusOr PjRtExecutable::EnqueueExecution( int executable_idx, const RunId& run_id, const ExecuteOptions& options, Device* device, std::vector* device_buffers) const { int device_ordinal = device->local_device_state()->device_ordinal(); - tensorflow::profiler::TraceMe traceme([&] { - return absl::StrCat("LocalExecutable::Execute#run_id=", run_id.ToInt(), - "#"); - }); + tensorflow::profiler::TraceMeConsumer activity( + "LocalExecutable::Execute", tensorflow::profiler::ContextType::kPjRt, + run_id.ToInt()); VLOG(3) << "Replica " << replica << ", partition " << partition << " mapped to device ordinal for execution: " << device_ordinal; @@ -1721,10 +1722,9 @@ PjRtExecutable::ExecuteOnLocalDevices( absl::Span> argument_handles, const ExecuteOptions& options) const { RunId run_id; - tensorflow::profiler::TraceMe traceme([&] { - return absl::StrCat( - "LocalExecutable::ExecuteOnLocalDevices#run_id=", run_id.ToInt(), "#"); - }); + tensorflow::profiler::TraceMeProducer activity( + "LocalExecutable::ExecuteOnLocalDevices", + tensorflow::profiler::ContextType::kPjRt, run_id.ToInt()); const int num_local_devices = local_devices_.size(); diff --git a/tensorflow/core/profiler/lib/connected_traceme.h b/tensorflow/core/profiler/lib/connected_traceme.h index ed8b4ac1ad2..b55c4407fe6 100644 --- a/tensorflow/core/profiler/lib/connected_traceme.h +++ b/tensorflow/core/profiler/lib/connected_traceme.h @@ -29,6 +29,7 @@ enum class ContextType : int { kGeneric, kTfExecutor, kSharedBatchScheduler, + kPjRt, }; /* From 61d83075ad34fd33d7d44ea3341a055e26e775a6 Mon Sep 17 00:00:00 2001 From: Niranjan Hasabnis Date: Wed, 17 Jun 2020 14:13:29 -0700 Subject: [PATCH 214/867] Fixing MklTanh compilation error with DNNL0 --- .../core/common_runtime/mkl_layout_pass.cc | 3 +++ .../common_runtime/mkl_layout_pass_test.cc | 3 +++ tensorflow/core/kernels/mkl_relu_op.cc | 23 +++++++++++++++++-- tensorflow/core/kernels/mkl_relu_op_test.cc | 5 +++- 4 files changed, 31 insertions(+), 3 deletions(-) diff --git a/tensorflow/core/common_runtime/mkl_layout_pass.cc b/tensorflow/core/common_runtime/mkl_layout_pass.cc index f6e42fc7e8c..778d5445cb2 100644 --- a/tensorflow/core/common_runtime/mkl_layout_pass.cc +++ b/tensorflow/core/common_runtime/mkl_layout_pass.cc @@ -682,12 +682,15 @@ class MklLayoutRewritePass : public GraphOptimizationPass { rinfo_.push_back( {csinfo_.requantize, mkl_op_registry::GetMklOpName(csinfo_.requantize), CopyAttrsAll, AlwaysRewrite, kRewriteForLayoutPropagation}); +#ifdef ENABLE_MKLDNN_V1 + // Optimized TanhGrad support exists only in DNNL 1.x. rinfo_.push_back({csinfo_.tanh, mkl_op_registry::GetMklOpName(csinfo_.tanh), CopyAttrsAll, AlwaysRewrite, kRewriteForLayoutPropagation}); rinfo_.push_back( {csinfo_.tanh_grad, mkl_op_registry::GetMklOpName(csinfo_.tanh_grad), CopyAttrsAll, AlwaysRewrite, kRewriteForLayoutPropagation}); +#endif // ENABLE_MKLDNN_V1 rinfo_.push_back( {csinfo_.reshape, mkl_op_registry::GetMklOpName(csinfo_.reshape), CopyAttrsAll, AlwaysRewrite, kRewriteForLayoutPropagation}); diff --git a/tensorflow/core/common_runtime/mkl_layout_pass_test.cc b/tensorflow/core/common_runtime/mkl_layout_pass_test.cc index 9971f6c5d7e..d480c0a49ce 100644 --- a/tensorflow/core/common_runtime/mkl_layout_pass_test.cc +++ b/tensorflow/core/common_runtime/mkl_layout_pass_test.cc @@ -3024,6 +3024,8 @@ REGISTER_TEST_ALL_TYPES(NodeRewrite_LeakyReluLeakyReluGrad_Positive); // clang-format on // clang-format off +#ifdef ENABLE_MKLDNN_V1 + #define REGISTER_TEST(NAME, T, INPUT) \ TEST_F(MklLayoutPassTest, NAME##_##T) { \ DCHECK_EQ(kTensorOrdering, MklTfTensorOrdering::TENSORS_CONTIGUOUS); \ @@ -3081,6 +3083,7 @@ REGISTER_TEST_ALL_TYPES(NodeRewrite_TanhGrad_Positive); } REGISTER_TEST_ALL_TYPES(NodeRewrite_TanhTanhGrad_Positive); #undef REGISTER_TEST +#endif // ENABLE_MKLDNN_V1 // clang-format on TEST_F(MklLayoutPassTest, NodeRewrite_AvgPool_Positive) { diff --git a/tensorflow/core/kernels/mkl_relu_op.cc b/tensorflow/core/kernels/mkl_relu_op.cc index 6d79b8f3282..70aa1e937d3 100644 --- a/tensorflow/core/kernels/mkl_relu_op.cc +++ b/tensorflow/core/kernels/mkl_relu_op.cc @@ -19,7 +19,6 @@ limitations under the License. #include #include "mkldnn.hpp" -#include "third_party/eigen3/unsupported/Eigen/CXX11/Tensor" #include "tensorflow/core/framework/numeric_op.h" #include "tensorflow/core/framework/op_kernel.h" #include "tensorflow/core/framework/register_types.h" @@ -27,6 +26,7 @@ limitations under the License. #include "tensorflow/core/lib/core/errors.h" #include "tensorflow/core/util/mkl_types.h" #include "tensorflow/core/util/mkl_util.h" +#include "third_party/eigen3/unsupported/Eigen/CXX11/Tensor" using mkldnn::algorithm; using mkldnn::eltwise_forward; @@ -269,7 +269,7 @@ class MklEltwiseBwdParams { MklEltwiseBwdParams(const memory::dims& src_dims, const memory::desc& common_md, algorithm alg_kind, - float alpha, float beta, int forward_input_type) + float alpha, float beta, int forward_input_type = -1) : src_dims(src_dims), common_md(common_md), alg_kind(alg_kind), @@ -644,7 +644,10 @@ class MklReluGradOpBase : public OpKernel { virtual int GetDiffSrcIndex() const { return 0; } // What is the type of input tensor that grad op receives from forward op -- // is it 'x' (SRC) or 'y' (DST). For Relu-family, it is 'x', so fwd op SRC. + +#ifdef ENABLE_MKLDNN_V1 virtual int GetTypeOfInputTensorFromFwdOp() const { return MKLDNN_ARG_SRC; } +#endif void Compute(OpKernelContext* context) { try { @@ -736,8 +739,16 @@ class MklReluGradOpBase : public OpKernel { common_md = src_md; } +#ifdef ENABLE_MKLDNN_V1 MklEltwiseBwdParams bwdParams(src_dims, common_md, alg_kind, alpha_, beta_, GetTypeOfInputTensorFromFwdOp()); +#else + // MKLDNN V0 does not support reusing output of forward op in backward. + // So this optimization works only in MKLDNN v1. + MklEltwiseBwdParams bwdParams(src_dims, common_md, alg_kind, alpha_, + beta_); +#endif // ENABLE_MKLDNN_V1 + MklEltwiseBwdPrimitive* eltwise_bwd = MklEltwiseBwdPrimitiveFactory::Get(bwdParams); @@ -962,6 +973,11 @@ class MklEluGradOp } }; +#ifdef ENABLE_MKLDNN_V1 +// Optimized TanhGrad support exists in DNNL1.x only +// (eltwise_tanh_use_dst_for_bwd). We can still support it with DNNL0.x, but +// it will not be optimized. So we disable it for DNNL0.x. + template class MklTanhOp : public MklReluOpBase { public: @@ -1043,6 +1059,7 @@ class MklTanhGradOp (static_cast(user_g))[0] * (static_cast(1) - tanh * tanh); } }; +#endif // ENABLE_MKLDNN_V1 #define RELU6_UPPER_BOUND 6.0f template @@ -1227,6 +1244,7 @@ TF_CALL_bfloat16(REGISTER_RELU_MKL_SUPPORTED_KERNELS_TYPES); TF_CALL_float(REGISTER_ELU_MKL_SUPPORTED_KERNELS_TYPES); TF_CALL_bfloat16(REGISTER_ELU_MKL_SUPPORTED_KERNELS_TYPES); +#ifdef ENABLE_MKLDNN_V1 #define REGISTER_TANH_MKL_SUPPORTED_KERNELS_TYPES(type) \ REGISTER_KERNEL_BUILDER( \ Name("_MklTanh") \ @@ -1242,6 +1260,7 @@ TF_CALL_bfloat16(REGISTER_ELU_MKL_SUPPORTED_KERNELS_TYPES); MklTanhGradOp); TF_CALL_float(REGISTER_TANH_MKL_SUPPORTED_KERNELS_TYPES); TF_CALL_bfloat16(REGISTER_TANH_MKL_SUPPORTED_KERNELS_TYPES); +#endif #define REGISTER_RELU6_MKL_SUPPORTED_KERNELS_TYPES(type) \ REGISTER_KERNEL_BUILDER( \ diff --git a/tensorflow/core/kernels/mkl_relu_op_test.cc b/tensorflow/core/kernels/mkl_relu_op_test.cc index d1fdf7ab4ae..86d7f979c1f 100644 --- a/tensorflow/core/kernels/mkl_relu_op_test.cc +++ b/tensorflow/core/kernels/mkl_relu_op_test.cc @@ -15,8 +15,8 @@ limitations under the License. #ifdef INTEL_MKL -#include "mkldnn.hpp" #include "absl/strings/match.h" +#include "mkldnn.hpp" #include "tensorflow/cc/ops/const_op.h" #include "tensorflow/cc/ops/nn_ops.h" #include "tensorflow/cc/ops/standard_ops.h" @@ -121,8 +121,11 @@ static Graph* Activation(const string& op_name, const string& kind, BM(OP, 32, 64, 128, 256, cpu); \ BM(OP, 33, 65, 129, 257, cpu); +#ifdef ENABLE_MKLDNN_V1 +// Optimized MKLDNN TanhGrad support exists in DNNL1.x only. TEST_ALL_SIZES(Tanh) TEST_ALL_SIZES(TanhGrad) +#endif // ENABLE_MKLDNN_V1 TEST_ALL_SIZES(Relu) TEST_ALL_SIZES(ReluGrad) TEST_ALL_SIZES(Elu) From e0962f4c374f4cbf78ad27fd2391f976c1a2050d Mon Sep 17 00:00:00 2001 From: Jiho Choi Date: Thu, 18 Jun 2020 13:35:31 -0700 Subject: [PATCH 215/867] Remove the old grouping rule for PJRT. PiperOrigin-RevId: 317170383 Change-Id: I96973a2d2cd0ca1fc786bc7500deb6b4fedd0534 --- tensorflow/core/profiler/utils/group_events.cc | 5 +---- tensorflow/core/profiler/utils/xplane_schema.cc | 4 ---- tensorflow/core/profiler/utils/xplane_schema.h | 3 --- 3 files changed, 1 insertion(+), 11 deletions(-) diff --git a/tensorflow/core/profiler/utils/group_events.cc b/tensorflow/core/profiler/utils/group_events.cc index be8dd506b0c..0772cff7b97 100644 --- a/tensorflow/core/profiler/utils/group_events.cc +++ b/tensorflow/core/profiler/utils/group_events.cc @@ -635,10 +635,7 @@ std::vector CreateInterThreadConnectInfoList() { {StatType::kStepId, StatType::kIterNum}}, {HostEventType::kKernelLaunch, HostEventType::kKernelExecute, - {StatType::kCorrelationId}}, - {HostEventType::kLocalExecutableExecuteOnLocalDevice, - HostEventType::kLocalExecutableExecute, - {StatType::kRunId}}}; + {StatType::kCorrelationId}}}; return connect_info_list; } diff --git a/tensorflow/core/profiler/utils/xplane_schema.cc b/tensorflow/core/profiler/utils/xplane_schema.cc index be53dcbdc01..5ca8326d72c 100644 --- a/tensorflow/core/profiler/utils/xplane_schema.cc +++ b/tensorflow/core/profiler/utils/xplane_schema.cc @@ -91,10 +91,6 @@ const HostEventTypeMap& GetHostEventTypeMap() { {"WhileOp-StartBody", kWhileOpStartBody}, {"ForOp", kForOp}, {"PartitionedCallOp", kPartitionedCallOp}, - // XLA related. - {"LocalExecutable::ExecuteOnLocalDevices", - kLocalExecutableExecuteOnLocalDevice}, - {"LocalExecutable::Execute", kLocalExecutableExecute}, // tf.data related. {"IteratorGetNextOp::DoCompute", kIteratorGetNextOp}, {"IteratorGetNextAsOptionalOp::DoCompute", kIteratorGetNextAsOptionalOp}, diff --git a/tensorflow/core/profiler/utils/xplane_schema.h b/tensorflow/core/profiler/utils/xplane_schema.h index a31814cef06..41774deaa59 100644 --- a/tensorflow/core/profiler/utils/xplane_schema.h +++ b/tensorflow/core/profiler/utils/xplane_schema.h @@ -81,9 +81,6 @@ enum HostEventType { kWhileOpStartBody, kForOp, kPartitionedCallOp, - // XLA related. - kLocalExecutableExecuteOnLocalDevice, - kLocalExecutableExecute, // tf.data related. kIteratorGetNextOp, kIteratorGetNextAsOptionalOp, From 8f700fb2e0da382f1e2e9630f56a7922a8799a59 Mon Sep 17 00:00:00 2001 From: Berkin Ilbeyi Date: Thu, 18 Jun 2020 13:39:10 -0700 Subject: [PATCH 216/867] [XLA] Propagate memory spaces recursively inside nested fusions. PiperOrigin-RevId: 317171110 Change-Id: I65004edb7498acb2f3b4238d9afbbb5d3930aab5 --- .../xla/service/memory_space_propagation.cc | 80 +++++++--- .../xla/service/memory_space_propagation.h | 11 +- .../service/memory_space_propagation_test.cc | 148 ++++++++++++++++++ 3 files changed, 214 insertions(+), 25 deletions(-) diff --git a/tensorflow/compiler/xla/service/memory_space_propagation.cc b/tensorflow/compiler/xla/service/memory_space_propagation.cc index 80eb4017477..2eb15b14eaf 100644 --- a/tensorflow/compiler/xla/service/memory_space_propagation.cc +++ b/tensorflow/compiler/xla/service/memory_space_propagation.cc @@ -29,36 +29,78 @@ StatusOr MemorySpacePropagation::Run(HloModule* module) { // Propagate the operand subshapes. for (int operand_idx = 0; operand_idx < instruction->operand_count(); ++operand_idx) { - modified |= - PropagateSubshapes(instruction->operand(operand_idx)->shape(), - instruction->fused_parameter(operand_idx)); + for (const ShapeUtil::IndexedShape& indexed_shape : + ShapeUtil::GetLeafShapes( + instruction->operand(operand_idx)->shape())) { + int64 memory_space = indexed_shape.shape.layout().memory_space(); + modified |= Propagate(indexed_shape.index, + instruction->fused_parameter(operand_idx), + memory_space); + } } // Propagate output subshapes. - modified |= PropagateSubshapes(instruction->shape(), - instruction->fused_expression_root()); + for (const ShapeUtil::IndexedShape& indexed_shape : + ShapeUtil::GetLeafShapes(instruction->shape())) { + int64 memory_space = indexed_shape.shape.layout().memory_space(); + modified |= + Propagate(indexed_shape.index, + instruction->fused_expression_root(), memory_space); + } } } } return modified; } -bool MemorySpacePropagation::PropagateSubshapes( - const Shape& caller_shape, const HloInstruction* callee_instruction) const { +bool MemorySpacePropagation::Propagate(ShapeIndexView index, + const HloInstruction* callee_instruction, + int64 memory_space) const { bool modified = false; - for (const ShapeUtil::IndexedShape& indexed_shape : - ShapeUtil::GetLeafShapes(caller_shape)) { - int64 memory_space = indexed_shape.shape.layout().memory_space(); - const HloValue& value = dataflow_analysis_->GetUniqueValueAt( - callee_instruction, indexed_shape.index); + const HloValue& value = dataflow_analysis_->GetUniqueValueAt( + callee_instruction, index.ToShapeIndex()); - for (const HloPosition& position : value.positions()) { - Shape* shape = ShapeUtil::GetMutableSubshape( - position.instruction->mutable_shape(), position.index); - if (shape->layout().memory_space() != memory_space) { - shape->mutable_layout()->set_memory_space(memory_space); - modified = true; - } + for (const HloPosition& position : value.positions()) { + HloInstruction* instruction = position.instruction; + Shape* shape = ShapeUtil::GetMutableSubshape(instruction->mutable_shape(), + position.index); + if (shape->layout().memory_space() == memory_space) { + continue; + } + shape->mutable_layout()->set_memory_space(memory_space); + modified = true; + + // For fusion outputs, propagate the memory space to the fusion root. + if (instruction->opcode() == HloOpcode::kFusion) { + Propagate(position.index, instruction->fused_expression_root(), + memory_space); + } + + const HloInstruction* parent_fusion = + instruction->parent()->FusionInstruction(); + // For nested fusion roots, pop one level up and propagate the memory space + // to the output of the calling fusion instruction. + if (instruction == instruction->parent()->root_instruction() && + parent_fusion->parent()->IsFusionComputation()) { + Propagate(position.index, parent_fusion, memory_space); + } + + // For nested fusion parameters, pop one level up and propagate the memory + // space to the operand of the calling fusion instruction. + if (instruction->opcode() == HloOpcode::kParameter && + parent_fusion->parent()->IsFusionComputation()) { + const HloInstruction* fusion_operand = + parent_fusion->operand(instruction->parameter_number()); + Propagate(position.index, fusion_operand, memory_space); + } + } + + for (const HloUse& use : value.uses()) { + // For fusion uses, propagate the memory space to the fusion parameter. + if (use.instruction->opcode() == HloOpcode::kFusion) { + modified |= Propagate( + use.operand_index, + use.instruction->fused_parameter(use.operand_number), memory_space); } } return modified; diff --git a/tensorflow/compiler/xla/service/memory_space_propagation.h b/tensorflow/compiler/xla/service/memory_space_propagation.h index 65a1dfd14a6..510e9e69f79 100644 --- a/tensorflow/compiler/xla/service/memory_space_propagation.h +++ b/tensorflow/compiler/xla/service/memory_space_propagation.h @@ -31,12 +31,11 @@ class MemorySpacePropagation : public HloModulePass { StatusOr Run(HloModule* module) override; private: - // Given the caller shape (operand or output) and its corresponding - // insturction in the fused computation (parameter or root), propagates the - // memory space to all the subshapes in the callee side. Returns true if the - // module is modified. - bool PropagateSubshapes(const Shape& caller_shape, - const HloInstruction* callee_instruction) const; + // Given the shape index (operand or output) and its corresponding instruction + // in the fused computation (parameter or root), propagates the memory space + // in the callee side. Returns true if the module is modified. + bool Propagate(ShapeIndexView index, const HloInstruction* callee_instruction, + int64 memory_space) const; std::unique_ptr dataflow_analysis_; }; diff --git a/tensorflow/compiler/xla/service/memory_space_propagation_test.cc b/tensorflow/compiler/xla/service/memory_space_propagation_test.cc index 8d74958f6aa..de45af5a190 100644 --- a/tensorflow/compiler/xla/service/memory_space_propagation_test.cc +++ b/tensorflow/compiler/xla/service/memory_space_propagation_test.cc @@ -199,5 +199,153 @@ TEST_F(MemorySpacePropagationTest, TupleOutput) { EXPECT_EQ(module->Hash(), ref->Hash()); } +TEST_F(MemorySpacePropagationTest, NestedInputFusion) { + // Tests propagating the memory space to nested fusions on the input side. + absl::string_view hlo_string = R"( + HloModule NestedFusion + + %bitcast_fusion { + %bf_param = s32[3,2]{0,1:T(128)} parameter(0) + ROOT %bitcast = s32[6]{0:T(128)} bitcast(%bf_param) + } + + %fused_computation { + %param_1.3 = s32[1]{0:T(128)} parameter(1) + %constant.2 = s32[]{:T(128)} constant(-2147483648) + %pad.2 = s32[6]{0:T(128)} pad(s32[1]{0:T(128)} %param_1.3, s32[]{:T(128)} %constant.2), padding=0_5 + %param_2.3 = s32[5]{0:T(128)} parameter(2) + %pad.3 = s32[6]{0:T(128)} pad(s32[5]{0:T(128)} %param_2.3, s32[]{:T(128)} %constant.2), padding=1_0 + %maximum.1 = s32[6]{0:T(128)} maximum(s32[6]{0:T(128)} %pad.2, s32[6]{0:T(128)} %pad.3) + %param_0.1 = s32[3,2]{0,1:T(128)} parameter(0) + %fusion.1 = s32[6]{0:T(128)} fusion(%param_0.1), kind=kLoop, calls=bitcast_fusion + ROOT %add.0 = s32[6]{0:T(128)} add(s32[6]{0:T(128)} %maximum.1, s32[6]{0:T(128)} %fusion.1) + } + + ENTRY %entry { + %param0 = s32[3,2]{0,1:T(128)} parameter(0) + %param1 = s32[1]{0:T(128)} parameter(1) + %param2 = s32[5]{0:T(128)} parameter(2) + %arg0 = s32[3,2]{0,1:T(128)S(1)} copy(%param0) + %arg1 = s32[1]{0:T(128)} copy(%param1) + %arg2 = s32[5]{0:T(128)S(1)} copy(%param2) + %fusion = s32[6]{0:T(128)S(1)} fusion(s32[3,2]{0,1:T(128)S(1)} %arg0, s32[1]{0:T(128)} %arg1, s32[5]{0:T(128)S(1)} %arg2), kind=kLoop, calls=%fused_computation + ROOT %root = s32[6]{0:T(128)} copy(%fusion) + } + )"; + absl::string_view expected_hlo_string = R"( + HloModule NestedFusion + + %bitcast_fusion { + %bf_param = s32[3,2]{0,1:T(128)S(1)} parameter(0) + ROOT %bitcast = s32[6]{0:T(128)S(1)} bitcast(%bf_param) + } + + %fused_computation { + %param_1.3 = s32[1]{0:T(128)} parameter(1) + %constant.2 = s32[]{:T(128)} constant(-2147483648) + %pad.2 = s32[6]{0:T(128)} pad(s32[1]{0:T(128)} %param_1.3, s32[]{:T(128)} %constant.2), padding=0_5 + %param_2.3 = s32[5]{0:T(128)S(1)} parameter(2) + %pad.3 = s32[6]{0:T(128)} pad(s32[5]{0:T(128)} %param_2.3, s32[]{:T(128)} %constant.2), padding=1_0 + %maximum.1 = s32[6]{0:T(128)} maximum(s32[6]{0:T(128)} %pad.2, s32[6]{0:T(128)} %pad.3) + %param_0.1 = s32[3,2]{0,1:T(128)S(1)} parameter(0) + %fusion.1 = s32[6]{0:T(128)S(1)} fusion(%param_0.1), kind=kLoop, calls=bitcast_fusion + ROOT %add.0 = s32[6]{0:T(128)S(1)} add(s32[6]{0:T(128)} %maximum.1, s32[6]{0:T(128)S(1)} %fusion.1) + } + + ENTRY %entry { + %param0 = s32[3,2]{0,1:T(128)} parameter(0) + %param1 = s32[1]{0:T(128)} parameter(1) + %param2 = s32[5]{0:T(128)} parameter(2) + %arg0 = s32[3,2]{0,1:T(128)S(1)} copy(%param0) + %arg1 = s32[1]{0:T(128)} copy(%param1) + %arg2 = s32[5]{0:T(128)S(1)} copy(%param2) + %fusion = s32[6]{0:T(128)S(1)} fusion(s32[3,2]{0,1:T(128)S(1)} %arg0, s32[1]{0:T(128)} %arg1, s32[5]{0:T(128)S(1)} %arg2), kind=kLoop, calls=%fused_computation + ROOT %root = s32[6]{0:T(128)} copy(%fusion) + } + )"; + TF_ASSERT_OK_AND_ASSIGN(auto module, + ParseAndReturnUnverifiedModule(hlo_string)); + MemorySpacePropagation memory_space_propagation; + EXPECT_TRUE(memory_space_propagation.Run(module.get()).ValueOrDie()); + TF_EXPECT_OK(Verify(module.get())); + TF_ASSERT_OK_AND_ASSIGN(auto ref, + ParseAndReturnVerifiedModule(expected_hlo_string)); + EXPECT_EQ(module->Hash(), ref->Hash()); +} + +TEST_F(MemorySpacePropagationTest, NestedOutputFusion) { + // Tests propagating the memory space to nested fusions on the output side. + absl::string_view hlo_string = R"( + HloModule NestedFusion + + %bitcast_fusion { + %bf_param = s32[6]{0:T(128)} parameter(0) + ROOT %bitcast = s32[3,2]{0,1:T(128)} bitcast(%bf_param) + } + + %fused_computation { + %param_1.3 = s32[1]{0:T(128)} parameter(1) + %constant.2 = s32[]{:T(128)} constant(-2147483648) + %pad.2 = s32[6]{0:T(128)} pad(s32[1]{0:T(128)} %param_1.3, s32[]{:T(128)} %constant.2), padding=0_5 + %param_2.3 = s32[5]{0:T(128)} parameter(2) + %pad.3 = s32[6]{0:T(128)} pad(s32[5]{0:T(128)} %param_2.3, s32[]{:T(128)} %constant.2), padding=1_0 + %maximum.1 = s32[6]{0:T(128)} maximum(s32[6]{0:T(128)} %pad.2, s32[6]{0:T(128)} %pad.3) + %param_0.1 = s32[6]{0:T(128)} parameter(0) + %add.0 = s32[6]{0:T(128)} add(s32[6]{0:T(128)} %maximum.1, s32[6]{0:T(128)} %param_0.1) + ROOT %fusion.1 = s32[3,2]{0,1:T(128)} fusion(%add.0), kind=kLoop, calls=bitcast_fusion + } + + ENTRY %entry { + %param0 = s32[6]{0:T(128)} parameter(0) + %param1 = s32[1]{0:T(128)} parameter(1) + %param2 = s32[5]{0:T(128)} parameter(2) + %arg0 = s32[6]{0:T(128)S(1)} copy(%param0) + %arg1 = s32[1]{0:T(128)} copy(%param1) + %arg2 = s32[5]{0:T(128)S(1)} copy(%param2) + %fusion = s32[3,2]{0,1:T(128)S(1)} fusion(s32[6]{0:T(128)S(1)} %arg0, s32[1]{0:T(128)} %arg1, s32[5]{0:T(128)S(1)} %arg2), kind=kLoop, calls=%fused_computation + ROOT %root = s32[3,2]{0,1:T(128)} copy(%fusion) + } + )"; + absl::string_view expected_hlo_string = R"( + HloModule NestedFusion + + %bitcast_fusion { + %bf_param = s32[6]{0:T(128)S(1)} parameter(0) + ROOT %bitcast = s32[3,2]{0,1:T(128)S(1)} bitcast(%bf_param) + } + + %fused_computation { + %param_1.3 = s32[1]{0:T(128)} parameter(1) + %constant.2 = s32[]{:T(128)} constant(-2147483648) + %pad.2 = s32[6]{0:T(128)} pad(s32[1]{0:T(128)} %param_1.3, s32[]{:T(128)} %constant.2), padding=0_5 + %param_2.3 = s32[5]{0:T(128)S(1)} parameter(2) + %pad.3 = s32[6]{0:T(128)} pad(s32[5]{0:T(128)} %param_2.3, s32[]{:T(128)} %constant.2), padding=1_0 + %maximum.1 = s32[6]{0:T(128)} maximum(s32[6]{0:T(128)} %pad.2, s32[6]{0:T(128)} %pad.3) + %param_0.1 = s32[6]{0:T(128)S(1)} parameter(0) + %add.0 = s32[6]{0:T(128)S(1)} add(s32[6]{0:T(128)} %maximum.1, s32[6]{0:T(128)S(1)} %param_0.1) + ROOT %fusion.1 = s32[3,2]{0,1:T(128)S(1)} fusion(%add.0), kind=kLoop, calls=bitcast_fusion + } + + ENTRY %entry { + %param0 = s32[6]{0:T(128)} parameter(0) + %param1 = s32[1]{0:T(128)} parameter(1) + %param2 = s32[5]{0:T(128)} parameter(2) + %arg0 = s32[6]{0:T(128)S(1)} copy(%param0) + %arg1 = s32[1]{0:T(128)} copy(%param1) + %arg2 = s32[5]{0:T(128)S(1)} copy(%param2) + %fusion = s32[3,2]{0,1:T(128)S(1)} fusion(s32[6]{0:T(128)S(1)} %arg0, s32[1]{0:T(128)} %arg1, s32[5]{0:T(128)S(1)} %arg2), kind=kLoop, calls=%fused_computation + ROOT %root = s32[3,2]{0,1:T(128)} copy(%fusion) + } + )"; + TF_ASSERT_OK_AND_ASSIGN(auto module, + ParseAndReturnUnverifiedModule(hlo_string)); + MemorySpacePropagation memory_space_propagation; + EXPECT_TRUE(memory_space_propagation.Run(module.get()).ValueOrDie()); + TF_EXPECT_OK(Verify(module.get())); + TF_ASSERT_OK_AND_ASSIGN(auto ref, + ParseAndReturnVerifiedModule(expected_hlo_string)); + EXPECT_EQ(module->Hash(), ref->Hash()); +} + } // namespace } // namespace xla From fc5151130813140eb4189f77dd3a759c4077836d Mon Sep 17 00:00:00 2001 From: "A. Unique TensorFlower" Date: Thu, 18 Jun 2020 13:41:07 -0700 Subject: [PATCH 217/867] Delete orphaned comment The comment for `loader_spec_` stuck around even after it was removed. PiperOrigin-RevId: 317171521 Change-Id: Iddb6029fdad9cd5ef33bc4f4ea2653caed305658 --- tensorflow/compiler/xla/service/gpu/kernel_thunk.h | 2 -- 1 file changed, 2 deletions(-) diff --git a/tensorflow/compiler/xla/service/gpu/kernel_thunk.h b/tensorflow/compiler/xla/service/gpu/kernel_thunk.h index 88351881f3a..25acabb239b 100644 --- a/tensorflow/compiler/xla/service/gpu/kernel_thunk.h +++ b/tensorflow/compiler/xla/service/gpu/kernel_thunk.h @@ -77,8 +77,6 @@ class KernelThunk : public Thunk { // Will be set by IrEmitterUnnested. LaunchDimensions launch_dimensions_; - // Describes how to load this kernel. ExecuteOnStream reuses this loader - // specification for all executions. mutable tensorflow::mutex mutex_; // Loaded kernels for each `StreamExecutor`. Requires pointer stability of From 8452c9f80ee02cb71fb72f638d3bdef754f15297 Mon Sep 17 00:00:00 2001 From: Gabriel Rasskin Date: Thu, 18 Jun 2020 13:58:25 -0700 Subject: [PATCH 218/867] Added status_group fuzzer --- .../security/fuzzing/status_group_fuzz.cc | 83 +++++++++++++++++++ 1 file changed, 83 insertions(+) create mode 100644 tensorflow/security/fuzzing/status_group_fuzz.cc diff --git a/tensorflow/security/fuzzing/status_group_fuzz.cc b/tensorflow/security/fuzzing/status_group_fuzz.cc new file mode 100644 index 00000000000..979fd444b48 --- /dev/null +++ b/tensorflow/security/fuzzing/status_group_fuzz.cc @@ -0,0 +1,83 @@ +/* Copyright 2020 The TensorFlow Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. +==============================================================================*/ +#include +#include + +#include "tensorflow/core/platform/status.h" + +#include + +// This is a fuzzer for `tensorflow::StatusGroup`. Since `Status` is used almost +// everywhere, we need to ensure that the common functionality is safe. We don't +// expect many crashes from this fuzzer + +namespace { + +tensorflow::error::Code BuildRandomErrorCode(uint32_t code){ + + // We cannot build a `Status` with error_code of 0 and a message, so force + // error code to be non-zero. + if (code == 0) { + return tensorflow::error::UNKNOWN; + } + + return static_cast(code); +} + +extern "C" int LLVMFuzzerTestOneInput(const uint8_t *data, size_t size) { + tensorflow::error::Code error_code; + + std::string error_message = "ERROR"; + + tensorflow::Status s, derived_s; + + tensorflow::StatusGroup sg; + + bool is_derived; + + uint32_t code; + + FuzzedDataProvider fuzzed_data(data, size); + + while(fuzzed_data.remaining_bytes() > 0) { + code = fuzzed_data.ConsumeIntegral(); + + error_code = BuildRandomErrorCode(code); + + is_derived = fuzzed_data.ConsumeBool(); + + s = tensorflow::Status(error_code, error_message); + + if(is_derived) { + derived_s = tensorflow::StatusGroup::MakeDerived(s); + + sg.Update(derived_s); + + } else { + sg.Update(s); + + } + } + + sg.as_summary_status(); + + sg.as_concatenated_status(); + + sg.AttachLogMessages(); + + return 0; +} + +} // namespace From c41f4652b45bf70f20686e612b41574b4b8139d7 Mon Sep 17 00:00:00 2001 From: Marissa Ikonomidis Date: Thu, 18 Jun 2020 13:52:21 -0700 Subject: [PATCH 219/867] Add an option to enable MLIR bridge for tpu_py_test rule If enable_mlir_bridge is True, a new test will be generated that runs with the MLIR bridge enabled. This option is off by default. PiperOrigin-RevId: 317173675 Change-Id: I332e1ae24cf82fceea20fd0aff2cec7c9b236a24 --- tensorflow/core/platform/default/distribute.bzl | 3 +++ tensorflow/python/framework/test_util.py | 3 +++ tensorflow/python/tpu/tpu.bzl | 2 ++ 3 files changed, 8 insertions(+) diff --git a/tensorflow/core/platform/default/distribute.bzl b/tensorflow/core/platform/default/distribute.bzl index 46a5d826a79..b16d5e8cff7 100644 --- a/tensorflow/core/platform/default/distribute.bzl +++ b/tensorflow/core/platform/default/distribute.bzl @@ -22,6 +22,7 @@ def distribute_py_test( full_precision = False, disable_v2 = False, disable_v3 = False, + disable_mlir_bridge = True, **kwargs): """Generates py_test targets for CPU and GPU. @@ -40,6 +41,7 @@ def distribute_py_test( full_precision: unused. disable_v2: whether tests for TPU version 2 should be generated. disable_v3: whether tests for TPU version 3 should be generated. + disable_mlir_bridge: whether to also run this with the mlir bridge enabled. **kwargs: extra keyword arguments to the non-tpu test. """ @@ -77,6 +79,7 @@ def distribute_py_test( tags = tpu_tags, disable_v2 = disable_v2, disable_v3 = disable_v3, + disable_mlir_bridge = disable_mlir_bridge, ) register_extension_info( diff --git a/tensorflow/python/framework/test_util.py b/tensorflow/python/framework/test_util.py index a46bb7c9bda..8ddbcf34f3b 100644 --- a/tensorflow/python/framework/test_util.py +++ b/tensorflow/python/framework/test_util.py @@ -1933,6 +1933,9 @@ class TensorFlowTestCase(googletest.TestCase): # disable it here. pywrap_tf_session.TF_SetXlaConstantFoldingDisabled(True) + if is_mlir_bridge_enabled(): + context.context().enable_mlir_bridge = True + self._threads = [] self._tempdir = None self._cached_session = None diff --git a/tensorflow/python/tpu/tpu.bzl b/tensorflow/python/tpu/tpu.bzl index 5453702d64d..3c26d9b49bf 100644 --- a/tensorflow/python/tpu/tpu.bzl +++ b/tensorflow/python/tpu/tpu.bzl @@ -25,6 +25,7 @@ def tpu_py_test( disable_v2 = False, disable_v3 = False, disable_experimental = False, + disable_mlir_bridge = True, args = [], **kwargs): """Generates identical unit test variants for various Cloud TPU versions. @@ -37,6 +38,7 @@ def tpu_py_test( disable_v2: If true, don't generate TPU v2 tests. disable_v3: If true, don't generate TPU v3 tests. disable_experimental: Unused. + disable_mlir_bridge: Unused. args: Arguments to apply to tests. **kwargs: Additional named arguments to apply to tests. """ From dda51e1c94160b8252c51dd0ddca445d821ba8b9 Mon Sep 17 00:00:00 2001 From: Gabriel Rasskin Date: Thu, 18 Jun 2020 13:59:11 -0700 Subject: [PATCH 220/867] Added status group fuzzer build rules --- tensorflow/security/fuzzing/BUILD | 8 ++++++++ 1 file changed, 8 insertions(+) diff --git a/tensorflow/security/fuzzing/BUILD b/tensorflow/security/fuzzing/BUILD index 9b5aeec2d36..871baa0055b 100644 --- a/tensorflow/security/fuzzing/BUILD +++ b/tensorflow/security/fuzzing/BUILD @@ -19,6 +19,14 @@ tf_fuzz_target( ], ) +tf_fuzz_target( + name = "status_group_fuzz", + srcs = ["status_group_fuzz.cc"], + deps = [ + "//tensorflow/core/platform:status", + ], +) + # A trivial fuzzer with no pre-specified corpus. # TODO(mihaimaruseac): Move fuzz_session and the op fuzzers to a subdirectory tf_fuzz_target( From ef52b4e0886b7212471462643e92e98bea0253be Mon Sep 17 00:00:00 2001 From: George Karpenkov Date: Thu, 18 Jun 2020 13:53:10 -0700 Subject: [PATCH 221/867] [XLA/Client] Implement LocalClient::Run which supports buffer donation PiperOrigin-RevId: 317173848 Change-Id: If92955ac5051376fbf0932b773b675459497c0c4 --- .../compiler/xla/client/local_client.cc | 47 +++++++++++++++---- tensorflow/compiler/xla/client/local_client.h | 9 ++++ .../tests/multiple_devices_on_host_test.cc | 3 +- tensorflow/compiler/xla/tests/while_test.cc | 6 ++- 4 files changed, 53 insertions(+), 12 deletions(-) diff --git a/tensorflow/compiler/xla/client/local_client.cc b/tensorflow/compiler/xla/client/local_client.cc index afe115deda8..f71e8a2d56d 100644 --- a/tensorflow/compiler/xla/client/local_client.cc +++ b/tensorflow/compiler/xla/client/local_client.cc @@ -168,6 +168,26 @@ LocalExecutable::RunHelper(const absl::Span argument_shapes, return std::make_pair(service_options, std::move(stream)); } +StatusOr LocalExecutable::GetExecutableRunOptions( + absl::Span argument_shapes, + const ExecutableRunOptions& run_options) { + TF_ASSIGN_OR_RETURN(auto options_and_stream, + RunHelper(argument_shapes, run_options)); + ExecutableRunOptions options = options_and_stream.first.run_options(); + options.set_device_ordinal(-1); + return options; +} + +template +static StatusOr BlockHostUntilDoneAfterAsyncCall( + se::Stream* stream, std::function()> async_callback) { + StatusOr result = async_callback(); + Status block_status = stream->BlockHostUntilDone(); + TF_RETURN_IF_ERROR(result.status()); + TF_RETURN_IF_ERROR(block_status); + return result; +} + StatusOr LocalExecutable::Run( const absl::Span arguments, ExecutableRunOptions run_options) { @@ -176,15 +196,24 @@ StatusOr LocalExecutable::Run( for (const ShapedBuffer* const arg : arguments) { argument_shapes.push_back(&arg->on_host_shape()); } - TF_ASSIGN_OR_RETURN(auto options_and_stream, - RunHelper(argument_shapes, run_options)); - ExecutableRunOptions options = options_and_stream.first.run_options(); - options.set_device_ordinal(-1); - auto result = RunAsync(arguments, options); - Status block_status = options.stream()->BlockHostUntilDone(); - TF_RETURN_IF_ERROR(result.status()); - TF_RETURN_IF_ERROR(block_status); - return result; + TF_ASSIGN_OR_RETURN(ExecutableRunOptions options, + GetExecutableRunOptions(argument_shapes, run_options)); + return BlockHostUntilDoneAfterAsyncCall( + options.stream(), [&] { return RunAsync(arguments, options); }); +} + +StatusOr LocalExecutable::Run( + std::vector arguments, ExecutableRunOptions run_options) { + std::vector argument_shapes; + argument_shapes.reserve(arguments.size()); + for (const ExecutionInput& arg : arguments) { + argument_shapes.push_back(&arg.shape()); + } + TF_ASSIGN_OR_RETURN(ExecutableRunOptions options, + GetExecutableRunOptions(argument_shapes, run_options)); + return BlockHostUntilDoneAfterAsyncCall( + options.stream(), + [&] { return RunAsync(argument_shapes, std::move(arguments), options); }); } static std::shared_ptr DumpArguments( diff --git a/tensorflow/compiler/xla/client/local_client.h b/tensorflow/compiler/xla/client/local_client.h index 7cdeb9dcbf6..b00f5cc6801 100644 --- a/tensorflow/compiler/xla/client/local_client.h +++ b/tensorflow/compiler/xla/client/local_client.h @@ -51,6 +51,11 @@ class LocalExecutable { const absl::Span arguments, ExecutableRunOptions run_options); + // Similar to Run(), but allows for donating argument buffers to the + // executable. + StatusOr Run(std::vector arguments, + ExecutableRunOptions run_options); + // Similar to Run(), but need not block the host waiting for the computation // to complete before returning. StatusOr RunAsync( @@ -85,6 +90,10 @@ class LocalExecutable { const absl::Span argument_shapes, ExecutableRunOptions run_options); + StatusOr GetExecutableRunOptions( + absl::Span argument_shapes, + const ExecutableRunOptions& run_options); + // The ordinal of the device which this executable was compiled for. The // executable can run on all equivalent devices (as determined by // Backend::devices_equivalent). diff --git a/tensorflow/compiler/xla/tests/multiple_devices_on_host_test.cc b/tensorflow/compiler/xla/tests/multiple_devices_on_host_test.cc index 2b19aaded9c..2231fc6feab 100644 --- a/tensorflow/compiler/xla/tests/multiple_devices_on_host_test.cc +++ b/tensorflow/compiler/xla/tests/multiple_devices_on_host_test.cc @@ -45,7 +45,8 @@ void CompileAndExecute( xla::ClientLibrary::GetXlaService(client->platform()) ->backend() .memory_allocator()); - StatusOr result = executable->Run({}, execute_options); + StatusOr result = + executable->Run(absl::Span(), execute_options); { absl::MutexLock lock(results_mutex); results->emplace_back(device_ordinal, std::move(result)); diff --git a/tensorflow/compiler/xla/tests/while_test.cc b/tensorflow/compiler/xla/tests/while_test.cc index d575bbb1f3e..8e8c3605cc7 100644 --- a/tensorflow/compiler/xla/tests/while_test.cc +++ b/tensorflow/compiler/xla/tests/while_test.cc @@ -1324,14 +1324,16 @@ void BM_WhileLoop(int num_iters) { options.set_allocator(&allocator); const int kWarmups = 2; for (int i = 0; i < kWarmups; ++i) { - auto result = executable->Run({}, options); + auto result = + executable->Run(absl::Span(), options); ASSERT_TRUE(result.ok()); } // Run benchmark. tensorflow::testing::StartTiming(); for (int i = 0; i < num_iters; ++i) { - auto result = executable->Run({}, options); + auto result = + executable->Run(absl::Span(), options); ASSERT_TRUE(result.ok()); } } From 3833402726d72d04dd2821e89c642f613e80a531 Mon Sep 17 00:00:00 2001 From: jonah-kohn <51345541+jonah-kohn@users.noreply.github.com> Date: Thu, 18 Jun 2020 14:28:47 -0700 Subject: [PATCH 222/867] Cast optimizer parameters as python floats during serialization. Accounted only for the case in which the hyper parameter is neither a callable nor a tensor, to avoid any troublesome casts. --- tensorflow/python/keras/optimizer_v2/optimizer_v2.py | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/tensorflow/python/keras/optimizer_v2/optimizer_v2.py b/tensorflow/python/keras/optimizer_v2/optimizer_v2.py index c55b332bfc0..d8992bbe3e0 100644 --- a/tensorflow/python/keras/optimizer_v2/optimizer_v2.py +++ b/tensorflow/python/keras/optimizer_v2/optimizer_v2.py @@ -910,7 +910,7 @@ class OptimizerV2(trackable.Trackable): return value() if tensor_util.is_tensor(value): return backend.get_value(value) - return value + return float(value) def variables(self): """Returns variables of this Optimizer based on the order created.""" From bc1c0e86a677d9b1e5d3e3f0da85c445c2a7efe2 Mon Sep 17 00:00:00 2001 From: "A. Unique TensorFlower" Date: Thu, 18 Jun 2020 14:24:06 -0700 Subject: [PATCH 223/867] Wrap save/restore logic in tf.function when in eager mode. This allows parallel saving and restoring when using multiple devices. PiperOrigin-RevId: 317180143 Change-Id: Icdc740d02beb7c2d3236191add3b72fa103fc134 --- .../grappler/optimizers/function_optimizer.cc | 8 +- .../parallel_device/parallel_device_test.py | 4 - .../python/framework/auto_control_deps.py | 2 +- tensorflow/python/training/saving/BUILD | 1 - .../training/saving/functional_saver.py | 111 ++++++------------ .../training/saving/functional_saver_test.py | 17 +-- 6 files changed, 42 insertions(+), 101 deletions(-) diff --git a/tensorflow/core/grappler/optimizers/function_optimizer.cc b/tensorflow/core/grappler/optimizers/function_optimizer.cc index 0e156aaa84c..a66e645e04b 100644 --- a/tensorflow/core/grappler/optimizers/function_optimizer.cc +++ b/tensorflow/core/grappler/optimizers/function_optimizer.cc @@ -837,6 +837,7 @@ const bool IsExemptFromSideEffectsExecutionValidation(const string& op) { "ParameterizedTruncatedNormal", "TruncatedNormal", "RandomShuffle", "Multinomial", "RandomGamma", "RandomGammaGrad", "RandomPoisson", "RandomPoissonV2", + // LINT.ThenChange(//tensorflow/python/framework/auto_control_deps.py) // ReadVariableOp marked as stateful because it consumes DT_RESOURCE, // but it can't generate any observable side-effect. @@ -850,12 +851,7 @@ const bool IsExemptFromSideEffectsExecutionValidation(const string& op) { // the same device_ordinal on the same host. "EnqueueTPUEmbeddingSparseBatch", "EnqueueTPUEmbeddingIntegerBatch", "EnqueueTPUEmbeddingSparseTensorBatch", - "EnqueueTPUEmbeddingRaggedTensorBatch", - - // SaveV2 and RestoreV2 should be allowed to operate in parallel on - // multiple hosts. - "SaveV2", "RestoreV2"}); - // LINT.ThenChange(//tensorflow/python/framework/auto_control_deps.py) + "EnqueueTPUEmbeddingRaggedTensorBatch"}); return exemption->contains(op); } diff --git a/tensorflow/python/distribute/parallel_device/parallel_device_test.py b/tensorflow/python/distribute/parallel_device/parallel_device_test.py index 1429c522aba..8fc3dcb5816 100644 --- a/tensorflow/python/distribute/parallel_device/parallel_device_test.py +++ b/tensorflow/python/distribute/parallel_device/parallel_device_test.py @@ -172,8 +172,6 @@ class ParallelDeviceTests(_VirtualDeviceTestCase): config.set_synchronous_execution(previous) def test_checkpointing(self): - self.skipTest( - "Disable saving until SaveableObject's methods are traceable.") prefix = os.path.join(self.get_temp_dir(), "ckpt") with self.device.scope(): different_values = self.device.pack( @@ -265,8 +263,6 @@ class LayerTests(_VirtualDeviceTestCase): self.assertIn(self.device.components[1], final_kernels[1].backing_device) def test_training_loop(self): - self.skipTest( - "Disable saving until SaveableObject's methods are traceable.") for _ in range(5): layer = _Dense(5) checkpoint = tracking.Checkpoint(layer=layer) diff --git a/tensorflow/python/framework/auto_control_deps.py b/tensorflow/python/framework/auto_control_deps.py index 4b47735e0bf..51dcb248b11 100644 --- a/tensorflow/python/framework/auto_control_deps.py +++ b/tensorflow/python/framework/auto_control_deps.py @@ -100,7 +100,7 @@ _ORDER_INSENSITIVE_STATEFUL_OPS = [ "CudnnRNNV2", "CudnnRNNV3", "CudnnRNNBackpropV2", "CudnnRNNBackpropV3", "EnqueueTPUEmbeddingSparseBatch", "EnqueueTPUEmbeddingIntegerBatch", "EnqueueTPUEmbeddingSparseTensorBatch", - "EnqueueTPUEmbeddingRaggedTensorBatch", "RestoreV2", "SaveV2" + "EnqueueTPUEmbeddingRaggedTensorBatch" ] # LINT.ThenChange(//tensorflow/core/grappler/optimizers/function_optimizer.cc) diff --git a/tensorflow/python/training/saving/BUILD b/tensorflow/python/training/saving/BUILD index 12940840309..670a4c35c6f 100644 --- a/tensorflow/python/training/saving/BUILD +++ b/tensorflow/python/training/saving/BUILD @@ -43,7 +43,6 @@ cuda_py_test( ":checkpoint_options", ":functional_saver", ":saveable_hook", - "//tensorflow/python/eager:remote", "//tensorflow/python/eager:test", ], ) diff --git a/tensorflow/python/training/saving/functional_saver.py b/tensorflow/python/training/saving/functional_saver.py index 3a9b565470d..c4334e096df 100644 --- a/tensorflow/python/training/saving/functional_saver.py +++ b/tensorflow/python/training/saving/functional_saver.py @@ -21,7 +21,6 @@ from __future__ import print_function import uuid from tensorflow.core.protobuf import saver_pb2 -from tensorflow.python.eager import context from tensorflow.python.eager import def_function from tensorflow.python.framework import constant_op from tensorflow.python.framework import dtypes @@ -162,8 +161,7 @@ class MultiDeviceSaver(object): self._after_restore_callbacks.append(saveable.after_restore) if is_saveable: - host_device = saveable_object_util.set_cpu0(saveable.device) - saveables_by_device.setdefault(host_device, []).append(saveable) + saveables_by_device.setdefault(saveable.device, []).append(saveable) self._single_device_savers = { device: _SingleDeviceSaver(saveables) @@ -249,50 +247,33 @@ class MultiDeviceSaver(object): tmp_checkpoint_prefix = string_ops.string_join( [file_prefix, sharded_suffix]) - def save_fn(): - num_shards = len(self._single_device_savers) - sharded_saves = [] - sharded_prefixes = [] - num_shards_tensor = constant_op.constant(num_shards, name="num_shards") - last_device = None - for shard, (device, saver) in enumerate( - sorted(self._single_device_savers.items())): - last_device = device - with ops.device(saveable_object_util.set_cpu0(device)): - shard_prefix = sharded_filename(tmp_checkpoint_prefix, shard, - num_shards_tensor) - sharded_prefixes.append(shard_prefix) - with ops.device(device): - # _SingleDeviceSaver will use the CPU device when necessary, but - # initial read operations should be placed on the SaveableObject's - # device. - sharded_saves.append(saver.save(shard_prefix, options)) + num_shards = len(self._single_device_savers) + sharded_saves = [] + sharded_prefixes = [] + num_shards_tensor = constant_op.constant(num_shards, name="num_shards") + last_device = None + for shard, (device, saver) in enumerate( + sorted(self._single_device_savers.items())): + last_device = device + with ops.device(saveable_object_util.set_cpu0(device)): + shard_prefix = sharded_filename(tmp_checkpoint_prefix, shard, + num_shards_tensor) + sharded_prefixes.append(shard_prefix) + with ops.device(device): + # _SingleDeviceSaver will use the CPU device when necessary, but initial + # read operations should be placed on the SaveableObject's device. + sharded_saves.append(saver.save(shard_prefix, options)) - with ops.control_dependencies(sharded_saves): - # Merge on the io_device if specified, otherwise co-locates the merge op - # with the last device used. - merge_device = ( - options.experimental_io_device or - saveable_object_util.set_cpu0(last_device)) - with ops.device(merge_device): - # V2 format write path consists of a metadata merge step. Once - # merged, attempts to delete the temporary directory, - # "_temp". - return gen_io_ops.merge_v2_checkpoints( - sharded_prefixes, file_prefix, delete_old_dirs=True) - - # Since this will causes a function re-trace on each save, limit this to the - # cases where it is needed: eager and when there are multiple tasks/single - # device savers. Note that the retrace is needed to ensure we pickup the - # latest values of options like experimental_io_device. - if context.executing_eagerly() and len(self._single_device_savers) > 1: - # Explicitly place the identity op on the first device. - @def_function.function(experimental_compile=False) - def tf_function_save(): - save_fn() - tf_function_save() - else: - return save_fn() + with ops.control_dependencies(sharded_saves): + # Merge on the io_device if specified, otherwise co-locates the merge op + # with the last device used. + merge_device = (options.experimental_io_device or + saveable_object_util.set_cpu0(last_device)) + with ops.device(merge_device): + # V2 format write path consists of a metadata merge step. Once merged, + # attempts to delete the temporary directory, "_temp". + return gen_io_ops.merge_v2_checkpoints( + sharded_prefixes, file_prefix, delete_old_dirs=True) def restore(self, file_prefix, options=None): """Restore the saveable objects from a checkpoint with `file_prefix`. @@ -306,38 +287,12 @@ class MultiDeviceSaver(object): A dictionary mapping from SaveableObject names to restore operations. """ options = options or checkpoint_options.CheckpointOptions() - - def restore_fn(): - restore_ops = {} - # Sort by device name to avoid propagating non-deterministic dictionary - # ordering in some Python versions. - for device, saver in sorted(self._single_device_savers.items()): - with ops.device(device): - restore_ops.update(saver.restore(file_prefix, options)) - - return restore_ops - - # Since this will causes a function re-trace on each save, limit this to the - # cases where it is needed: eager and when there are multiple tasks/single - # device savers. Note that the retrace is needed to ensure we pickup the - # latest values of options like experimental_io_device. - if context.executing_eagerly() and len(self._single_device_savers) > 1: - first_device, _ = list(self._single_device_savers.items())[0] - @def_function.function(experimental_compile=False) - def tf_function_restore(): - restore_ops = restore_fn() - restore_tensors = {} - # tf.functions must return tensors, thus we use control dependencies so - # that we can return a tensor which depends on the given op. - with ops.device(saveable_object_util.set_cpu0(first_device)): - for name, op in restore_ops.items(): - with ops.control_dependencies([op]): - restore_tensors[name] = array_ops.identity(file_prefix) - return restore_tensors - - restore_ops = tf_function_restore() - else: - restore_ops = restore_fn() + restore_ops = {} + # Sort by device name to avoid propagating non-deterministic dictionary + # ordering in some Python versions. + for device, saver in sorted(self._single_device_savers.items()): + with ops.device(device): + restore_ops.update(saver.restore(file_prefix, options)) for callback in self._after_restore_callbacks: callback() diff --git a/tensorflow/python/training/saving/functional_saver_test.py b/tensorflow/python/training/saving/functional_saver_test.py index 8f3eef4fb9c..7db32ff72d7 100644 --- a/tensorflow/python/training/saving/functional_saver_test.py +++ b/tensorflow/python/training/saving/functional_saver_test.py @@ -21,7 +21,6 @@ from __future__ import print_function import os from tensorflow.python.eager import context -from tensorflow.python.eager import remote from tensorflow.python.eager import test from tensorflow.python.eager import wrap_function from tensorflow.python.framework import config @@ -30,7 +29,6 @@ from tensorflow.python.framework import ops from tensorflow.python.framework import test_util from tensorflow.python.ops import resource_variable_ops from tensorflow.python.platform import gfile -from tensorflow.python.training import server_lib from tensorflow.python.training.saving import checkpoint_options from tensorflow.python.training.saving import functional_saver from tensorflow.python.training.saving import saveable_hook @@ -128,16 +126,13 @@ class SaverTest(test.TestCase): second_saver.restore(save_path) self.assertEqual(2., self.evaluate(v2)) - def test_checkpoint_is_sharded_by_task(self): - servers = [server_lib.Server.create_local_server() for _ in range(3)] - cluster_spec = server_lib.ClusterSpec({ - "worker": [s.target[len("grpc://"):] for s in servers]}) - remote.connect_to_cluster(cluster_spec) - with ops.device("/job:worker/task:0/cpu:0"): + @test_util.run_in_graph_and_eager_modes + def test_checkpoint_is_sharded_by_device(self): + with ops.device("cpu:0"): v0 = resource_variable_ops.ResourceVariable(0.) - with ops.device("/job:worker/task:1/cpu:0"): + with ops.device("cpu:1"): v1 = resource_variable_ops.ResourceVariable(1.) - with ops.device("/job:worker/task:2/cpu:0"): + with ops.device("cpu:2"): v2 = resource_variable_ops.ResourceVariable(2.) self.evaluate([v0.initializer, v1.initializer, v2.initializer]) @@ -172,7 +167,7 @@ class SaverTest(test.TestCase): list(saveable_object_util.saveable_objects_for_op(v2, "v2"))) prefix = os.path.join(self.get_temp_dir(), "ckpt") self.evaluate(saver.save(constant_op.constant(prefix), self.local_options)) - self.assertEqual(2, len(gfile.Glob(prefix + "*"))) + self.assertEqual(4, len(gfile.Glob(prefix + "*"))) self.evaluate(v0.assign(-1.)) self.evaluate(v1.assign(-1.)) self.evaluate(v2.assign(-1.)) From cb6e1ed5d8a406861398c428ca5fd6b84b439357 Mon Sep 17 00:00:00 2001 From: Henry Tan Date: Thu, 18 Jun 2020 14:28:42 -0700 Subject: [PATCH 224/867] Return `debug_string` when creating CompilationCacheKey. PiperOrigin-RevId: 317181056 Change-Id: I02198244c1c3749ff1ecf4e0647b8daa80dd868c --- tensorflow/core/tpu/kernels/BUILD | 16 ++ .../kernels/tpu_compilation_cache_external.cc | 127 --------------- .../kernels/tpu_compilation_cache_external.h | 8 - .../core/tpu/kernels/tpu_compile_c_api.h | 19 ++- tensorflow/core/tpu/kernels/tpu_op_util.cc | 151 ++++++++++++++++++ tensorflow/core/tpu/kernels/tpu_op_util.h | 40 +++++ 6 files changed, 223 insertions(+), 138 deletions(-) create mode 100644 tensorflow/core/tpu/kernels/tpu_op_util.cc create mode 100644 tensorflow/core/tpu/kernels/tpu_op_util.h diff --git a/tensorflow/core/tpu/kernels/BUILD b/tensorflow/core/tpu/kernels/BUILD index 94d3c8edf2b..9d38eb71f3c 100644 --- a/tensorflow/core/tpu/kernels/BUILD +++ b/tensorflow/core/tpu/kernels/BUILD @@ -405,6 +405,22 @@ cc_library( alwayslink = True, ) +cc_library( + name = "tpu_op_util", + srcs = ["tpu_op_util.cc"], + hdrs = ["tpu_op_util.h"], + deps = [ + ":tpu_compilation_cache_key", + ":tpu_compile_c_api_hdrs", + ":tpu_mesh_state_interface", + "//tensorflow/compiler/xla:xla_data_proto_cc", + "//tensorflow/core:framework", + "//tensorflow/core:lib", + "//tensorflow/core/protobuf/tpu:compile_metadata_proto_cc", + "@com_google_absl//absl/strings", + ], +) + cc_library( name = "tpu_util", srcs = ["tpu_util.cc"], diff --git a/tensorflow/core/tpu/kernels/tpu_compilation_cache_external.cc b/tensorflow/core/tpu/kernels/tpu_compilation_cache_external.cc index 8cee90e8e55..c4442fc95d5 100644 --- a/tensorflow/core/tpu/kernels/tpu_compilation_cache_external.cc +++ b/tensorflow/core/tpu/kernels/tpu_compilation_cache_external.cc @@ -49,70 +49,6 @@ void PopulateEntry(const std::string& key, CompiledSubgraph* entry, absl::make_unique(std::move(tpu_program_group)); entry->initialized = true; } - -// Return fingerprint_in_metadata if it's not empty; otherwise read input tensor -// data to compute the fingerprint. -std::string GuaranteedConstFingerprint( - const string& fingerprint_in_metadata, - const OpInputList& guaranteed_constants) { - if (fingerprint_in_metadata.empty()) { - uint64_t fingerprint = 0; - for (const auto& constant : guaranteed_constants) { - fingerprint = TpuCompile_CreateGuaranteedConstFingerprint( - fingerprint, constant.tensor_data().data(), - constant.tensor_data().size()); - } - return std::to_string(fingerprint); - } else { - return fingerprint_in_metadata; - } -} - -std::string CreateShapePrefix( - const std::vector& dynamic_shapes) { - std::string shapes_prefix; - for (const TensorShape& shape : dynamic_shapes) { - for (int64 size : shape.dim_sizes()) { - absl::StrAppend(&shapes_prefix, size, ","); - } - absl::StrAppend(&shapes_prefix, ";"); - } - return shapes_prefix; -} - -// Include compilation configurations of the arguments that are not captured -// by the called graph. -std::string CreateConfigPrefix(const TPUCompileMetadataProto& metadata) { - std::string config_prefix; - for (const auto& arg : metadata.args()) { - if (arg.is_same_data_across_replicas()) { - absl::StrAppend(&config_prefix, ":s"); - // Same. - } else { - // Different. - absl::StrAppend(&config_prefix, ":"); - } - if (arg.enable_xla_sharding() == - tpu::TPUCompileMetadataProto::Arg::ALLOWED) { - // Enabled. - absl::StrAppend(&config_prefix, "e"); - } - if (arg.unrestricted_layout()) { - // Unrestricted. - absl::StrAppend(&config_prefix, ":u"); - } - absl::StrAppend(&config_prefix, ",type(", arg.dtype(), ")"); - if (arg.has_shape()) { - absl::StrAppend(&config_prefix, ",shape("); - for (const auto& dim : arg.shape().dim()) { - absl::StrAppend(&config_prefix, dim.size(), ","); - } - absl::StrAppend(&config_prefix, ")"); - } - } - return config_prefix; -} - } // namespace TpuCompilationCacheExternal::EntryRefImpl::EntryRefImpl( @@ -196,68 +132,5 @@ CompiledSubgraph* TpuCompilationCacheExternal::InitializeEntry( marked_for_eviction_size_ += main_entry->total_size; return main_entry; } - -/*static*/ TpuCompilationCacheKey -TpuCompilationCacheExternal::CreateCompilationCacheKey( - absl::string_view function_name, uint64 function_library_fingerprint, - absl::string_view mlir_module, - const tensorflow::OpInputList& guaranteed_constants, - const std::vector& dynamic_shapes, - const tensorflow::tpu::TPUCompileMetadataProto& metadata, - const TpuMeshStateInterface& mesh_state) { - VLOG(1) << "FunctionLibraryFingerprint:" << function_library_fingerprint; - std::string shapes_prefix = CreateShapePrefix(dynamic_shapes); - VLOG(1) << "shapes_prefix = " << shapes_prefix; - std::string config_prefix = CreateConfigPrefix(metadata); - VLOG(1) << "config_prefix = " << config_prefix; - std::vector flattened_device_ids; - if (metadata.has_device_assignment()) { - for (const auto& device : - metadata.device_assignment().computation_devices()) { - flattened_device_ids.insert(flattened_device_ids.end(), - device.replica_device_ids().begin(), - device.replica_device_ids().end()); - } - } - // TODO(henrytan): return the debug_string. - const char* prefix = - TpuCompile_CreateCompilationCacheKey(CompilationCacheKeyProperty{ - config_prefix.data(), - shapes_prefix.data(), - function_name.data(), - mlir_module.data(), - flattened_device_ids.data(), - flattened_device_ids.size(), - guaranteed_constants.size(), - function_library_fingerprint, - metadata.num_cores_per_replica(), - metadata.num_replicas(), - mesh_state.data(), - }); - auto buffer_cleanup = gtl::MakeCleanup([prefix]() { delete[] prefix; }); - TpuCompilationCacheKey key; - key.prefix = prefix; - - // Guaranteed constants can be different across sessions. Use session_handle - // and guaranteed_const fingerprint to guarantee no collision. - if (guaranteed_constants.size() > 0) { - key.has_guaranteed_const = true; - key.session_handle = metadata.session_handle(); - // Both `metadata` and `guaranteed_constants` lifetime are captured by - // reference based on the assumption that these variables lifetime is - // managed through the `TPUCompileOpKernelImpl` that outlives the - // lifetime of the compilation cache lookups. - string fingerprint; - key.guaranteed_const_fingerprint = [&metadata, &guaranteed_constants, - fingerprint]() mutable { - if (fingerprint.empty()) { - fingerprint = GuaranteedConstFingerprint( - metadata.guaranteed_const_fingerprint(), guaranteed_constants); - } - return fingerprint; - }; - } - return key; -} } // namespace tpu } // namespace tensorflow diff --git a/tensorflow/core/tpu/kernels/tpu_compilation_cache_external.h b/tensorflow/core/tpu/kernels/tpu_compilation_cache_external.h index 2c75cb4d053..fe251326a43 100644 --- a/tensorflow/core/tpu/kernels/tpu_compilation_cache_external.h +++ b/tensorflow/core/tpu/kernels/tpu_compilation_cache_external.h @@ -63,14 +63,6 @@ class TpuCompilationCacheExternal : public TpuCompilationCacheInterface { explicit TpuCompilationCacheExternal(int64 max_cache_size) : TpuCompilationCacheInterface(max_cache_size) {} - static TpuCompilationCacheKey CreateCompilationCacheKey( - absl::string_view function_name, uint64 function_library_fingerprint, - absl::string_view mlir_module, - const tensorflow::OpInputList& guaranteed_constants, - const std::vector& dynamic_shapes, - const tensorflow::tpu::TPUCompileMetadataProto& metadata, - const TpuMeshStateInterface& mesh_state); - string DebugString() const override { return "TpuCompilationCacheExternal"; } private: diff --git a/tensorflow/core/tpu/kernels/tpu_compile_c_api.h b/tensorflow/core/tpu/kernels/tpu_compile_c_api.h index d1546ed9610..c101e489d56 100644 --- a/tensorflow/core/tpu/kernels/tpu_compile_c_api.h +++ b/tensorflow/core/tpu/kernels/tpu_compile_c_api.h @@ -42,6 +42,13 @@ struct CompilationCacheKeyProperty { const XLA_TpuMeshState* mesh_state; }; +// Compilation cache key result returning both the key and a more verbose debug +// version. +struct CompilationCacheKeyResult { + const char* key; + const char* debug_string; +}; + extern "C" { // Returns the number of available TPU core count. @@ -49,9 +56,14 @@ TFTPU_CAPI_EXPORT int TpuTopology_AvailableCoreCount( const XLA_TpuMeshState* mesh_state, TpuCoreTypeEnum tpu_core_type); // Creates a unique compilation cache `key` used for `put` and `get` operations. -// Returned buffer is heap-allocated and must be owned. -TFTPU_CAPI_EXPORT const char* TpuCompile_CreateCompilationCacheKey( - CompilationCacheKeyProperty property); +// Returned buffers are heap-allocated and must be owned. +TFTPU_CAPI_EXPORT CompilationCacheKeyResult +TpuCompile_CreateCompilationCacheKey(CompilationCacheKeyProperty property); + +// Destroys the CompilationCacheKeyResult returned by calling the +// `TpuCompile_CreateCompilationCacheKey` API. +TFTPU_CAPI_EXPORT void TpuCompile_DestroyCompilationCacheKey( + CompilationCacheKeyResult result); // Creates a guaranteed const fingerprint. Guarantee const is normally used in // TPU inference to avoid re-copying unchanged variables onto the TPU device. @@ -75,6 +87,7 @@ TFTPU_CAPI_EXPORT void TpuCompile_BuildXLADeviceAssignment( struct TfTpu_CompileApiFn { TFTPU_ADD_FN_IN_STRUCT(TpuTopology_AvailableCoreCount); TFTPU_ADD_FN_IN_STRUCT(TpuCompile_CreateCompilationCacheKey); + TFTPU_ADD_FN_IN_STRUCT(TpuCompile_DestroyCompilationCacheKey); TFTPU_ADD_FN_IN_STRUCT(TpuCompile_CreateGuaranteedConstFingerprint); TFTPU_ADD_FN_IN_STRUCT(TpuCompile_CompileAheadOfTime); TFTPU_ADD_FN_IN_STRUCT(TpuCompile_BuildXLADeviceAssignment); diff --git a/tensorflow/core/tpu/kernels/tpu_op_util.cc b/tensorflow/core/tpu/kernels/tpu_op_util.cc new file mode 100644 index 00000000000..e2f717fea8b --- /dev/null +++ b/tensorflow/core/tpu/kernels/tpu_op_util.cc @@ -0,0 +1,151 @@ +/* Copyright 2020 The TensorFlow Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. +==============================================================================*/ +#include "tensorflow/core/tpu/kernels/tpu_op_util.h" + +#include + +#include "tensorflow/core/lib/gtl/cleanup.h" +#include "tensorflow/core/tpu/kernels/tpu_compile_c_api.h" + +namespace tensorflow { +namespace tpu { +namespace { +// Return fingerprint_in_metadata if it's not empty; otherwise read input tensor +// data to compute the fingerprint. +std::string GuaranteedConstFingerprint( + const string& fingerprint_in_metadata, + const OpInputList& guaranteed_constants) { + if (fingerprint_in_metadata.empty()) { + uint64_t fingerprint = 0; + for (const auto& constant : guaranteed_constants) { + fingerprint = TpuCompile_CreateGuaranteedConstFingerprint( + fingerprint, constant.tensor_data().data(), + constant.tensor_data().size()); + } + return std::to_string(fingerprint); + } else { + return fingerprint_in_metadata; + } +} + +std::string CreateShapePrefix( + const std::vector& dynamic_shapes) { + std::string shapes_prefix; + for (const TensorShape& shape : dynamic_shapes) { + for (int64 size : shape.dim_sizes()) { + absl::StrAppend(&shapes_prefix, size, ","); + } + absl::StrAppend(&shapes_prefix, ";"); + } + return shapes_prefix; +} + +// Include compilation configurations of the arguments that are not captured +// by the called graph. +std::string CreateConfigPrefix(const TPUCompileMetadataProto& metadata) { + std::string config_prefix; + for (const auto& arg : metadata.args()) { + if (arg.is_same_data_across_replicas()) { + absl::StrAppend(&config_prefix, ":s"); + // Same. + } else { + // Different. + absl::StrAppend(&config_prefix, ":"); + } + if (arg.enable_xla_sharding() == + tpu::TPUCompileMetadataProto::Arg::ALLOWED) { + // Enabled. + absl::StrAppend(&config_prefix, "e"); + } + if (arg.unrestricted_layout()) { + // Unrestricted. + absl::StrAppend(&config_prefix, ":u"); + } + absl::StrAppend(&config_prefix, ",type(", arg.dtype(), ")"); + if (arg.has_shape()) { + absl::StrAppend(&config_prefix, ",shape("); + for (const auto& dim : arg.shape().dim()) { + absl::StrAppend(&config_prefix, dim.size(), ","); + } + absl::StrAppend(&config_prefix, ")"); + } + } + return config_prefix; +} +} // namespace + +TpuCompilationCacheKey CreateCompilationCacheKey( + absl::string_view function_name, uint64 function_library_fingerprint, + absl::string_view mlir_module, const OpInputList& guaranteed_constants, + const std::vector& dynamic_shapes, + const TPUCompileMetadataProto& metadata, + const TpuMeshStateInterface& mesh_state) { + VLOG(1) << "FunctionLibraryFingerprint:" << function_library_fingerprint; + std::string shapes_prefix = CreateShapePrefix(dynamic_shapes); + VLOG(1) << "shapes_prefix = " << shapes_prefix; + std::string config_prefix = CreateConfigPrefix(metadata); + VLOG(1) << "config_prefix = " << config_prefix; + std::vector flattened_device_ids; + if (metadata.has_device_assignment()) { + for (const auto& device : + metadata.device_assignment().computation_devices()) { + flattened_device_ids.insert(flattened_device_ids.end(), + device.replica_device_ids().begin(), + device.replica_device_ids().end()); + } + } + CompilationCacheKeyResult result = + TpuCompile_CreateCompilationCacheKey(CompilationCacheKeyProperty{ + config_prefix.data(), + shapes_prefix.data(), + function_name.data(), + mlir_module.data(), + flattened_device_ids.data(), + flattened_device_ids.size(), + guaranteed_constants.size(), + function_library_fingerprint, + metadata.num_cores_per_replica(), + metadata.num_replicas(), + mesh_state.data(), + }); + auto buffer_cleanup = gtl::MakeCleanup( + [result]() { TpuCompile_DestroyCompilationCacheKey(result); }); + TpuCompilationCacheKey key; + key.prefix = result.key; + key.debug_string = result.debug_string; + + // Guaranteed constants can be different across sessions. Use session_handle + // and guaranteed_const fingerprint to guarantee no collision. + if (guaranteed_constants.size() > 0) { + key.has_guaranteed_const = true; + key.session_handle = metadata.session_handle(); + // Both `metadata` and `guaranteed_constants` lifetime are captured by + // reference based on the assumption that these variables lifetime is + // managed through the `TPUCompileOpKernelImpl` that outlives the + // lifetime of the compilation cache lookups. + string fingerprint; + key.guaranteed_const_fingerprint = [&metadata, &guaranteed_constants, + fingerprint]() mutable { + if (fingerprint.empty()) { + fingerprint = GuaranteedConstFingerprint( + metadata.guaranteed_const_fingerprint(), guaranteed_constants); + } + return fingerprint; + }; + } + return key; +} +} // namespace tpu +} // namespace tensorflow diff --git a/tensorflow/core/tpu/kernels/tpu_op_util.h b/tensorflow/core/tpu/kernels/tpu_op_util.h new file mode 100644 index 00000000000..0a9657ca05e --- /dev/null +++ b/tensorflow/core/tpu/kernels/tpu_op_util.h @@ -0,0 +1,40 @@ +/* Copyright 2020 The TensorFlow Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. +==============================================================================*/ +#ifndef TENSORFLOW_CORE_TPU_KERNELS_TPU_OP_UTIL_H_ +#define TENSORFLOW_CORE_TPU_KERNELS_TPU_OP_UTIL_H_ + +#include + +#include "absl/strings/string_view.h" +#include "tensorflow/compiler/xla/xla_data.pb.h" +#include "tensorflow/core/framework/op_kernel.h" +#include "tensorflow/core/framework/tensor_shape.h" +#include "tensorflow/core/protobuf/tpu/compile_metadata.pb.h" +#include "tensorflow/core/tpu/kernels/tpu_compilation_cache_key.h" +#include "tensorflow/core/tpu/kernels/tpu_mesh_state_interface.h" + +namespace tensorflow { +namespace tpu { +// Creates a unique compilation cache `key`. +TpuCompilationCacheKey CreateCompilationCacheKey( + absl::string_view function_name, uint64 function_library_fingerprint, + absl::string_view mlir_module, const OpInputList& guaranteed_constants, + const std::vector& dynamic_shapes, + const TPUCompileMetadataProto& metadata, + const TpuMeshStateInterface& mesh_state); +} // namespace tpu +} // namespace tensorflow + +#endif // TENSORFLOW_CORE_TPU_KERNELS_TPU_OP_UTIL_H_ From 9d33f296d1edc2f656e253cf2a015d36daedd5c1 Mon Sep 17 00:00:00 2001 From: Henry Tan Date: Thu, 18 Jun 2020 15:00:50 -0700 Subject: [PATCH 225/867] Prep change for publishing TPU Ops. PiperOrigin-RevId: 317188030 Change-Id: I29f9236c0ade6bf586c8a52ead977b5d31aec357 --- tensorflow/core/tpu/kernels/BUILD | 11 +++ .../tpu_compilation_cache_entry_unloader.h | 69 +++++++++++++++++++ 2 files changed, 80 insertions(+) create mode 100644 tensorflow/core/tpu/kernels/tpu_compilation_cache_entry_unloader.h diff --git a/tensorflow/core/tpu/kernels/BUILD b/tensorflow/core/tpu/kernels/BUILD index 9d38eb71f3c..a41747ee8c5 100644 --- a/tensorflow/core/tpu/kernels/BUILD +++ b/tensorflow/core/tpu/kernels/BUILD @@ -450,3 +450,14 @@ cc_library( hdrs = ["tpu_compile_op.h"], deps = ["//tensorflow/core:framework"], ) + +cc_library( + name = "tpu_compilation_cache_entry_unloader", + hdrs = ["tpu_compilation_cache_entry_unloader.h"], + deps = [ + ":tpu_compilation_cache_interface", + "//tensorflow/core:framework", + "@com_google_absl//absl/container:flat_hash_set", + "@com_google_absl//absl/synchronization", + ], +) diff --git a/tensorflow/core/tpu/kernels/tpu_compilation_cache_entry_unloader.h b/tensorflow/core/tpu/kernels/tpu_compilation_cache_entry_unloader.h new file mode 100644 index 00000000000..c298d8fcc12 --- /dev/null +++ b/tensorflow/core/tpu/kernels/tpu_compilation_cache_entry_unloader.h @@ -0,0 +1,69 @@ +/* Copyright 2020 The TensorFlow Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. +==============================================================================*/ +#ifndef TENSORFLOW_CORE_TPU_KERNELS_COMPILATION_CACHE_ENTRY_UNLOADER_H_ +#define TENSORFLOW_CORE_TPU_KERNELS_COMPILATION_CACHE_ENTRY_UNLOADER_H_ + +#include "absl/container/flat_hash_set.h" +#include "absl/synchronization/mutex.h" +#include "tensorflow/core/framework/resource_mgr.h" +#include "tensorflow/core/tpu/kernels/tpu_compilation_cache_interface.h" + +namespace tensorflow { +namespace tpu { + +class TpuCompilationCacheEntryUnloader : public ResourceBase { + public: + explicit TpuCompilationCacheEntryUnloader(TpuCompilationCacheInterface* cache) + : cache_(cache) { + // Hold a reference to the cache until the unloader is destroyed. + cache_->Ref(); + VLOG(1) << "Will unload compilation cache entries when session closes."; + } + + ~TpuCompilationCacheEntryUnloader() override { + absl::MutexLock lock(&mu_); + for (int64 uid : cache_entry_uids_) { + Status s = cache_->MarkEntryForEviction(uid); + if (!s.ok()) { + LOG(WARNING) << "MarkEntryForEviction in " + "~CompilationCacheEntryUnloader fails with error " + << s; + } + } + // Release our reference to the cache. + cache_->Unref(); + } + + // Add cache entry uid to be unloaded in destructor. + void AddCacheEntryUid(int64 uid) { + absl::MutexLock lock(&mu_); + cache_entry_uids_.insert(uid); + } + + std::string DebugString() const override { + return "CompilationCacheEntryUnloader"; + } + + private: + TF_DISALLOW_COPY_AND_ASSIGN(TpuCompilationCacheEntryUnloader); + mutable absl::Mutex mu_; + TpuCompilationCacheInterface* cache_; // Not owned. + absl::flat_hash_set cache_entry_uids_ ABSL_GUARDED_BY(mu_); +}; + +} // namespace tpu +} // namespace tensorflow + +#endif // TENSORFLOW_CORE_TPU_KERNELS_COMPILATION_CACHE_ENTRY_UNLOADER_H_ From 35b978db57eaa87f32e7c9c3e9a7c323e595c978 Mon Sep 17 00:00:00 2001 From: Sean Silva Date: Thu, 18 Jun 2020 15:00:53 -0700 Subject: [PATCH 226/867] Move tfl-device-index-selector to TF directory. There's nothing lite-specific about this pass. PiperOrigin-RevId: 317188038 Change-Id: Iac9799e296e043aabf7aeabec2e8f72d07c77178 --- tensorflow/compiler/mlir/lite/BUILD | 1 - tensorflow/compiler/mlir/lite/tf_tfl_passes.cc | 2 +- tensorflow/compiler/mlir/lite/transforms/passes.h | 3 --- tensorflow/compiler/mlir/tensorflow/BUILD | 1 + .../transforms/device_index_selector.cc | 12 ++++++------ .../compiler/mlir/tensorflow/transforms/passes.h | 3 +++ .../tests/tf_device_index_selector.mlir | 2 +- 7 files changed, 12 insertions(+), 12 deletions(-) rename tensorflow/compiler/mlir/{lite => tensorflow}/transforms/device_index_selector.cc (92%) rename tensorflow/compiler/{mlir/lite => tensorflow}/tests/tf_device_index_selector.mlir (94%) diff --git a/tensorflow/compiler/mlir/lite/BUILD b/tensorflow/compiler/mlir/lite/BUILD index 8e9d615053c..8d4efeb3d60 100644 --- a/tensorflow/compiler/mlir/lite/BUILD +++ b/tensorflow/compiler/mlir/lite/BUILD @@ -314,7 +314,6 @@ tf_cc_test( cc_library( name = "tensorflow_lite_legalize_tf", srcs = [ - "transforms/device_index_selector.cc", "transforms/dilated_conv.cc", "transforms/generated_legalize_tf.inc", "transforms/generated_lower_static_tensor_list.inc", diff --git a/tensorflow/compiler/mlir/lite/tf_tfl_passes.cc b/tensorflow/compiler/mlir/lite/tf_tfl_passes.cc index 008098f62ba..fed2896035b 100644 --- a/tensorflow/compiler/mlir/lite/tf_tfl_passes.cc +++ b/tensorflow/compiler/mlir/lite/tf_tfl_passes.cc @@ -63,7 +63,7 @@ void AddTFToTFLConversionPasses(const mlir::TFL::PassConfig& pass_config, standard_pipeline_options.enable_inliner = false; standard_pipeline_options.form_clusters = pass_config.form_clusters; mlir::TF::CreateTFStandardPipeline(*pass_manager, standard_pipeline_options); - pass_manager->addPass(mlir::TFL::CreateDeviceIndexSelectorPass()); + pass_manager->addPass(mlir::TF::CreateDeviceIndexSelectorPass()); if (pass_config.shape_inference) { pass_manager->addPass(mlir::TF::CreateTFShapeInferencePass()); diff --git a/tensorflow/compiler/mlir/lite/transforms/passes.h b/tensorflow/compiler/mlir/lite/transforms/passes.h index 01e5eb1cb68..105c9394fb4 100644 --- a/tensorflow/compiler/mlir/lite/transforms/passes.h +++ b/tensorflow/compiler/mlir/lite/transforms/passes.h @@ -91,9 +91,6 @@ std::unique_ptr> CreateWhileOutlinePass(); // Verifies runtime constraints. std::unique_ptr> CreateRuntimeVerifyPass(); -// Creates function pass to select device index/fold tf.DeviceIndex. -std::unique_ptr> CreateDeviceIndexSelectorPass(); - } // namespace TFL } // namespace mlir diff --git a/tensorflow/compiler/mlir/tensorflow/BUILD b/tensorflow/compiler/mlir/tensorflow/BUILD index 54e57512c32..7c0d427e87b 100644 --- a/tensorflow/compiler/mlir/tensorflow/BUILD +++ b/tensorflow/compiler/mlir/tensorflow/BUILD @@ -475,6 +475,7 @@ cc_library( "transforms/cluster_outlining.cc", "transforms/collection_ops_util.cc", "transforms/decompose_resource_ops_pass.cc", + "transforms/device_index_selector.cc", "transforms/einsum.cc", "transforms/executor_island_coarsening.cc", "transforms/executor_tpuv1_inline_tpu_island.cc", diff --git a/tensorflow/compiler/mlir/lite/transforms/device_index_selector.cc b/tensorflow/compiler/mlir/tensorflow/transforms/device_index_selector.cc similarity index 92% rename from tensorflow/compiler/mlir/lite/transforms/device_index_selector.cc rename to tensorflow/compiler/mlir/tensorflow/transforms/device_index_selector.cc index d4aed750dc8..550647a915a 100644 --- a/tensorflow/compiler/mlir/lite/transforms/device_index_selector.cc +++ b/tensorflow/compiler/mlir/tensorflow/transforms/device_index_selector.cc @@ -21,11 +21,11 @@ limitations under the License. #include "mlir/IR/Operation.h" // from @llvm-project #include "mlir/IR/PatternMatch.h" // from @llvm-project #include "mlir/Pass/Pass.h" // from @llvm-project -#include "tensorflow/compiler/mlir/lite/transforms/passes.h" #include "tensorflow/compiler/mlir/tensorflow/ir/tf_ops.h" +#include "tensorflow/compiler/mlir/tensorflow/transforms/passes.h" namespace mlir { -namespace TFL { +namespace TF { namespace { // Folds the DeviceIndex op to a constant value. The DeviceIndex return the @@ -55,8 +55,8 @@ void DeviceIndexSelector::runOnOperation() { // Convert all the DeviceIndex ops to constant values. func.getBody().walk([](TF::DeviceIndexOp op) { // This just selects the default in all cases where DeviceIndex feeds into - // tf.Case. This could be enhanced based on explicit TFLite specification or - // TAC in future. + // tf.Case. This could be enhanced to have some sort of policy in the + // future. OpBuilder b(op); RankedTensorType type = RankedTensorType::get({}, b.getIntegerType(32)); int index = op.device_names().size(); @@ -79,7 +79,7 @@ std::unique_ptr> CreateDeviceIndexSelectorPass() { } static PassRegistration pass( - "tfl-device-index-selector", "Fold tf.DeviceIndex to constant"); + "tf-device-index-selector", "Fold tf.DeviceIndex to constant"); -} // namespace TFL +} // namespace TF } // namespace mlir diff --git a/tensorflow/compiler/mlir/tensorflow/transforms/passes.h b/tensorflow/compiler/mlir/tensorflow/transforms/passes.h index a34be28c809..168b317641d 100644 --- a/tensorflow/compiler/mlir/tensorflow/transforms/passes.h +++ b/tensorflow/compiler/mlir/tensorflow/transforms/passes.h @@ -147,6 +147,9 @@ std::unique_ptr> CreateLegalizeHloToTfPass(); // generally used beyond exporting to runtimes that supports these ops. In the // future these fusions may be codegen'd automatically. std::unique_ptr> CreateFusedKernelMatcherPass(); + +// Creates function pass to select device index/fold tf.DeviceIndex. +std::unique_ptr> CreateDeviceIndexSelectorPass(); } // namespace TF namespace tf_executor { diff --git a/tensorflow/compiler/mlir/lite/tests/tf_device_index_selector.mlir b/tensorflow/compiler/tensorflow/tests/tf_device_index_selector.mlir similarity index 94% rename from tensorflow/compiler/mlir/lite/tests/tf_device_index_selector.mlir rename to tensorflow/compiler/tensorflow/tests/tf_device_index_selector.mlir index 1ac7f30d644..7fc2b210f91 100644 --- a/tensorflow/compiler/mlir/lite/tests/tf_device_index_selector.mlir +++ b/tensorflow/compiler/tensorflow/tests/tf_device_index_selector.mlir @@ -1,6 +1,6 @@ // Test DeviceIndex selector. -// RUN: tf-opt --tfl-device-index-selector %s | FileCheck %s +// RUN: tf-opt --tf-device-index-selector %s | FileCheck %s // CHECK-LABEL: func @select func @select(%arg0: tensor, %arg1: tensor) -> (tensor, tensor) { From 834fe68f365e1d7f082b596fe87471ce84c2c8ec Mon Sep 17 00:00:00 2001 From: Pete Warden Date: Thu, 18 Jun 2020 15:16:08 -0700 Subject: [PATCH 227/867] Optimized Arduino library by enabling precompilation Precompilation allows Arduino users to build their sketches much faster, but requires some support from the library properties to enable. This has recently been upgraded to suppor the 'full' mode, as shown in https://github.com/arduino/arduino-cli/pull/611, so we want to take advantage of this. PiperOrigin-RevId: 317191283 Change-Id: Ie44a31ba45105f65fdad0da487290aff5fa2a179 --- tensorflow/lite/micro/tools/make/templates/library.properties | 3 ++- 1 file changed, 2 insertions(+), 1 deletion(-) diff --git a/tensorflow/lite/micro/tools/make/templates/library.properties b/tensorflow/lite/micro/tools/make/templates/library.properties index e41fd8d8fbe..6e02748a0b4 100644 --- a/tensorflow/lite/micro/tools/make/templates/library.properties +++ b/tensorflow/lite/micro/tools/make/templates/library.properties @@ -7,4 +7,5 @@ paragraph=This library runs TensorFlow machine learning models on microcontrolle category=Data Processing url=https://www.tensorflow.org/lite/microcontrollers/overview ldflags=-lm -includes=TensorFlowLite.h \ No newline at end of file +includes=TensorFlowLite.h +precompiled=full From 852cde437fdd062f52c42e47344029897ee67afd Mon Sep 17 00:00:00 2001 From: Gabriel Rasskin Date: Thu, 18 Jun 2020 15:27:56 -0700 Subject: [PATCH 228/867] Make error_message constant Co-authored-by: Mihai Maruseac --- tensorflow/security/fuzzing/status_group_fuzz.cc | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/tensorflow/security/fuzzing/status_group_fuzz.cc b/tensorflow/security/fuzzing/status_group_fuzz.cc index 979fd444b48..5e2b7eec403 100644 --- a/tensorflow/security/fuzzing/status_group_fuzz.cc +++ b/tensorflow/security/fuzzing/status_group_fuzz.cc @@ -39,7 +39,7 @@ tensorflow::error::Code BuildRandomErrorCode(uint32_t code){ extern "C" int LLVMFuzzerTestOneInput(const uint8_t *data, size_t size) { tensorflow::error::Code error_code; - std::string error_message = "ERROR"; + const std::string error_message = "ERROR"; tensorflow::Status s, derived_s; From 14e942faaa56b749c81c52595d04a9bb7f26fa02 Mon Sep 17 00:00:00 2001 From: Gabriel Rasskin Date: Thu, 18 Jun 2020 15:31:26 -0700 Subject: [PATCH 229/867] Update spacing and variable declaration --- .../security/fuzzing/status_group_fuzz.cc | 27 ++++--------------- 1 file changed, 5 insertions(+), 22 deletions(-) diff --git a/tensorflow/security/fuzzing/status_group_fuzz.cc b/tensorflow/security/fuzzing/status_group_fuzz.cc index 5e2b7eec403..52d83d00866 100644 --- a/tensorflow/security/fuzzing/status_group_fuzz.cc +++ b/tensorflow/security/fuzzing/status_group_fuzz.cc @@ -14,9 +14,7 @@ limitations under the License. ==============================================================================*/ #include #include - #include "tensorflow/core/platform/status.h" - #include // This is a fuzzer for `tensorflow::StatusGroup`. Since `Status` is used almost @@ -37,37 +35,22 @@ tensorflow::error::Code BuildRandomErrorCode(uint32_t code){ } extern "C" int LLVMFuzzerTestOneInput(const uint8_t *data, size_t size) { - tensorflow::error::Code error_code; - const std::string error_message = "ERROR"; - - tensorflow::Status s, derived_s; - tensorflow::StatusGroup sg; - - bool is_derived; - - uint32_t code; - FuzzedDataProvider fuzzed_data(data, size); while(fuzzed_data.remaining_bytes() > 0) { - code = fuzzed_data.ConsumeIntegral(); + uint32_t code = fuzzed_data.ConsumeIntegral(); + tensorflow::error::Code error_code = BuildRandomErrorCode(code); + bool is_derived = fuzzed_data.ConsumeBool(); - error_code = BuildRandomErrorCode(code); - - is_derived = fuzzed_data.ConsumeBool(); - - s = tensorflow::Status(error_code, error_message); + tensorflow::Status s = tensorflow::Status(error_code, error_message); if(is_derived) { - derived_s = tensorflow::StatusGroup::MakeDerived(s); - + tensorflow::Status derived_s = tensorflow::StatusGroup::MakeDerived(s); sg.Update(derived_s); - } else { sg.Update(s); - } } From 71bbebbf4d04c1bcb6ed44e2156087c9fec06e9e Mon Sep 17 00:00:00 2001 From: Gabriel Rasskin Date: Thu, 18 Jun 2020 15:32:15 -0700 Subject: [PATCH 230/867] Moved final StatusGroup method calls --- tensorflow/security/fuzzing/status_group_fuzz.cc | 2 -- 1 file changed, 2 deletions(-) diff --git a/tensorflow/security/fuzzing/status_group_fuzz.cc b/tensorflow/security/fuzzing/status_group_fuzz.cc index 52d83d00866..bc80cd72bc9 100644 --- a/tensorflow/security/fuzzing/status_group_fuzz.cc +++ b/tensorflow/security/fuzzing/status_group_fuzz.cc @@ -55,9 +55,7 @@ extern "C" int LLVMFuzzerTestOneInput(const uint8_t *data, size_t size) { } sg.as_summary_status(); - sg.as_concatenated_status(); - sg.AttachLogMessages(); return 0; From a82b75c82b63c4397b3d6a215e439ca77e687a84 Mon Sep 17 00:00:00 2001 From: "A. Unique TensorFlower" Date: Thu, 18 Jun 2020 15:35:42 -0700 Subject: [PATCH 231/867] [XLA/Client] Implement LocalClient::Run which supports buffer donation PiperOrigin-RevId: 317195199 Change-Id: If4d35d0627fa068a0c2b522fdae52466abd21f51 --- .../compiler/xla/client/local_client.cc | 47 ++++--------------- tensorflow/compiler/xla/client/local_client.h | 9 ---- .../tests/multiple_devices_on_host_test.cc | 3 +- tensorflow/compiler/xla/tests/while_test.cc | 6 +-- 4 files changed, 12 insertions(+), 53 deletions(-) diff --git a/tensorflow/compiler/xla/client/local_client.cc b/tensorflow/compiler/xla/client/local_client.cc index f71e8a2d56d..afe115deda8 100644 --- a/tensorflow/compiler/xla/client/local_client.cc +++ b/tensorflow/compiler/xla/client/local_client.cc @@ -168,26 +168,6 @@ LocalExecutable::RunHelper(const absl::Span argument_shapes, return std::make_pair(service_options, std::move(stream)); } -StatusOr LocalExecutable::GetExecutableRunOptions( - absl::Span argument_shapes, - const ExecutableRunOptions& run_options) { - TF_ASSIGN_OR_RETURN(auto options_and_stream, - RunHelper(argument_shapes, run_options)); - ExecutableRunOptions options = options_and_stream.first.run_options(); - options.set_device_ordinal(-1); - return options; -} - -template -static StatusOr BlockHostUntilDoneAfterAsyncCall( - se::Stream* stream, std::function()> async_callback) { - StatusOr result = async_callback(); - Status block_status = stream->BlockHostUntilDone(); - TF_RETURN_IF_ERROR(result.status()); - TF_RETURN_IF_ERROR(block_status); - return result; -} - StatusOr LocalExecutable::Run( const absl::Span arguments, ExecutableRunOptions run_options) { @@ -196,24 +176,15 @@ StatusOr LocalExecutable::Run( for (const ShapedBuffer* const arg : arguments) { argument_shapes.push_back(&arg->on_host_shape()); } - TF_ASSIGN_OR_RETURN(ExecutableRunOptions options, - GetExecutableRunOptions(argument_shapes, run_options)); - return BlockHostUntilDoneAfterAsyncCall( - options.stream(), [&] { return RunAsync(arguments, options); }); -} - -StatusOr LocalExecutable::Run( - std::vector arguments, ExecutableRunOptions run_options) { - std::vector argument_shapes; - argument_shapes.reserve(arguments.size()); - for (const ExecutionInput& arg : arguments) { - argument_shapes.push_back(&arg.shape()); - } - TF_ASSIGN_OR_RETURN(ExecutableRunOptions options, - GetExecutableRunOptions(argument_shapes, run_options)); - return BlockHostUntilDoneAfterAsyncCall( - options.stream(), - [&] { return RunAsync(argument_shapes, std::move(arguments), options); }); + TF_ASSIGN_OR_RETURN(auto options_and_stream, + RunHelper(argument_shapes, run_options)); + ExecutableRunOptions options = options_and_stream.first.run_options(); + options.set_device_ordinal(-1); + auto result = RunAsync(arguments, options); + Status block_status = options.stream()->BlockHostUntilDone(); + TF_RETURN_IF_ERROR(result.status()); + TF_RETURN_IF_ERROR(block_status); + return result; } static std::shared_ptr DumpArguments( diff --git a/tensorflow/compiler/xla/client/local_client.h b/tensorflow/compiler/xla/client/local_client.h index b00f5cc6801..7cdeb9dcbf6 100644 --- a/tensorflow/compiler/xla/client/local_client.h +++ b/tensorflow/compiler/xla/client/local_client.h @@ -51,11 +51,6 @@ class LocalExecutable { const absl::Span arguments, ExecutableRunOptions run_options); - // Similar to Run(), but allows for donating argument buffers to the - // executable. - StatusOr Run(std::vector arguments, - ExecutableRunOptions run_options); - // Similar to Run(), but need not block the host waiting for the computation // to complete before returning. StatusOr RunAsync( @@ -90,10 +85,6 @@ class LocalExecutable { const absl::Span argument_shapes, ExecutableRunOptions run_options); - StatusOr GetExecutableRunOptions( - absl::Span argument_shapes, - const ExecutableRunOptions& run_options); - // The ordinal of the device which this executable was compiled for. The // executable can run on all equivalent devices (as determined by // Backend::devices_equivalent). diff --git a/tensorflow/compiler/xla/tests/multiple_devices_on_host_test.cc b/tensorflow/compiler/xla/tests/multiple_devices_on_host_test.cc index 2231fc6feab..2b19aaded9c 100644 --- a/tensorflow/compiler/xla/tests/multiple_devices_on_host_test.cc +++ b/tensorflow/compiler/xla/tests/multiple_devices_on_host_test.cc @@ -45,8 +45,7 @@ void CompileAndExecute( xla::ClientLibrary::GetXlaService(client->platform()) ->backend() .memory_allocator()); - StatusOr result = - executable->Run(absl::Span(), execute_options); + StatusOr result = executable->Run({}, execute_options); { absl::MutexLock lock(results_mutex); results->emplace_back(device_ordinal, std::move(result)); diff --git a/tensorflow/compiler/xla/tests/while_test.cc b/tensorflow/compiler/xla/tests/while_test.cc index 8e8c3605cc7..d575bbb1f3e 100644 --- a/tensorflow/compiler/xla/tests/while_test.cc +++ b/tensorflow/compiler/xla/tests/while_test.cc @@ -1324,16 +1324,14 @@ void BM_WhileLoop(int num_iters) { options.set_allocator(&allocator); const int kWarmups = 2; for (int i = 0; i < kWarmups; ++i) { - auto result = - executable->Run(absl::Span(), options); + auto result = executable->Run({}, options); ASSERT_TRUE(result.ok()); } // Run benchmark. tensorflow::testing::StartTiming(); for (int i = 0; i < num_iters; ++i) { - auto result = - executable->Run(absl::Span(), options); + auto result = executable->Run({}, options); ASSERT_TRUE(result.ok()); } } From 0deffad6acbc2f5848022bf8ae360c9adbdf1ef8 Mon Sep 17 00:00:00 2001 From: Scott Zhu Date: Thu, 18 Jun 2020 15:50:38 -0700 Subject: [PATCH 232/867] Make `return_state` as explicit kwarg in the Conv2DLSTM layer. It was previously hide in the **kwargs, and we are also missing documentation for it. The existing test case should already cover the functionality of it. PiperOrigin-RevId: 317197835 Change-Id: Icfae1e177eeb886b41345078f6b93f282a94df5b --- .../keras/layers/convolutional_recurrent.py | 43 +++++++++++-------- ...orflow.keras.layers.-conv-l-s-t-m2-d.pbtxt | 2 +- ...orflow.keras.layers.-conv-l-s-t-m2-d.pbtxt | 2 +- 3 files changed, 28 insertions(+), 19 deletions(-) diff --git a/tensorflow/python/keras/layers/convolutional_recurrent.py b/tensorflow/python/keras/layers/convolutional_recurrent.py index 19831429b73..6c812204cba 100644 --- a/tensorflow/python/keras/layers/convolutional_recurrent.py +++ b/tensorflow/python/keras/layers/convolutional_recurrent.py @@ -753,7 +753,9 @@ class ConvLSTM2D(ConvRNN2D): the `recurrent_kernel` weights matrix. bias_constraint: Constraint function applied to the bias vector. return_sequences: Boolean. Whether to return the last output - in the output sequence, or the full sequence. + in the output sequence, or the full sequence. (default False) + return_state: Boolean Whether to return the last state + in addition to the output. (default False) go_backwards: Boolean (default False). If True, process the input sequence backwards. stateful: Boolean (default False). If True, the last state @@ -786,22 +788,27 @@ class ConvLSTM2D(ConvRNN2D): `(samples, time, rows, cols, channels)` Output shape: - - If `return_sequences` - - If data_format='channels_first' - 5D tensor with shape: - `(samples, time, filters, output_row, output_col)` - - If data_format='channels_last' - 5D tensor with shape: - `(samples, time, output_row, output_col, filters)` - - Else - - If data_format ='channels_first' - 4D tensor with shape: - `(samples, filters, output_row, output_col)` - - If data_format='channels_last' - 4D tensor with shape: - `(samples, output_row, output_col, filters)` - where `o_row` and `o_col` depend on the shape of the filter and - the padding + - If `return_state`: a list of tensors. The first tensor is + the output. The remaining tensors are the last states, + each 4D tensor with shape: + `(samples, filters, new_rows, new_cols)` + if data_format='channels_first' + or 4D tensor with shape: + `(samples, new_rows, new_cols, filters)` + if data_format='channels_last'. + `rows` and `cols` values might have changed due to padding. + - If `return_sequences`: 5D tensor with shape: + `(samples, timesteps, filters, new_rows, new_cols)` + if data_format='channels_first' + or 5D tensor with shape: + `(samples, timesteps, new_rows, new_cols, filters)` + if data_format='channels_last'. + - Else, 4D tensor with shape: + `(samples, filters, new_rows, new_cols)` + if data_format='channels_first' + or 4D tensor with shape: + `(samples, new_rows, new_cols, filters)` + if data_format='channels_last'. Raises: ValueError: in case of invalid constructor arguments. @@ -834,6 +841,7 @@ class ConvLSTM2D(ConvRNN2D): recurrent_constraint=None, bias_constraint=None, return_sequences=False, + return_state=False, go_backwards=False, stateful=False, dropout=0., @@ -863,6 +871,7 @@ class ConvLSTM2D(ConvRNN2D): dtype=kwargs.get('dtype')) super(ConvLSTM2D, self).__init__(cell, return_sequences=return_sequences, + return_state=return_state, go_backwards=go_backwards, stateful=stateful, **kwargs) diff --git a/tensorflow/tools/api/golden/v1/tensorflow.keras.layers.-conv-l-s-t-m2-d.pbtxt b/tensorflow/tools/api/golden/v1/tensorflow.keras.layers.-conv-l-s-t-m2-d.pbtxt index f77d613e354..958d06a0d0f 100644 --- a/tensorflow/tools/api/golden/v1/tensorflow.keras.layers.-conv-l-s-t-m2-d.pbtxt +++ b/tensorflow/tools/api/golden/v1/tensorflow.keras.layers.-conv-l-s-t-m2-d.pbtxt @@ -207,7 +207,7 @@ tf_class { } member_method { name: "__init__" - argspec: "args=[\'self\', \'filters\', \'kernel_size\', \'strides\', \'padding\', \'data_format\', \'dilation_rate\', \'activation\', \'recurrent_activation\', \'use_bias\', \'kernel_initializer\', \'recurrent_initializer\', \'bias_initializer\', \'unit_forget_bias\', \'kernel_regularizer\', \'recurrent_regularizer\', \'bias_regularizer\', \'activity_regularizer\', \'kernel_constraint\', \'recurrent_constraint\', \'bias_constraint\', \'return_sequences\', \'go_backwards\', \'stateful\', \'dropout\', \'recurrent_dropout\'], varargs=None, keywords=kwargs, defaults=[\'(1, 1)\', \'valid\', \'None\', \'(1, 1)\', \'tanh\', \'hard_sigmoid\', \'True\', \'glorot_uniform\', \'orthogonal\', \'zeros\', \'True\', \'None\', \'None\', \'None\', \'None\', \'None\', \'None\', \'None\', \'False\', \'False\', \'False\', \'0.0\', \'0.0\'], " + argspec: "args=[\'self\', \'filters\', \'kernel_size\', \'strides\', \'padding\', \'data_format\', \'dilation_rate\', \'activation\', \'recurrent_activation\', \'use_bias\', \'kernel_initializer\', \'recurrent_initializer\', \'bias_initializer\', \'unit_forget_bias\', \'kernel_regularizer\', \'recurrent_regularizer\', \'bias_regularizer\', \'activity_regularizer\', \'kernel_constraint\', \'recurrent_constraint\', \'bias_constraint\', \'return_sequences\', \'return_state\', \'go_backwards\', \'stateful\', \'dropout\', \'recurrent_dropout\'], varargs=None, keywords=kwargs, defaults=[\'(1, 1)\', \'valid\', \'None\', \'(1, 1)\', \'tanh\', \'hard_sigmoid\', \'True\', \'glorot_uniform\', \'orthogonal\', \'zeros\', \'True\', \'None\', \'None\', \'None\', \'None\', \'None\', \'None\', \'None\', \'False\', \'False\', \'False\', \'False\', \'0.0\', \'0.0\'], " } member_method { name: "add_loss" diff --git a/tensorflow/tools/api/golden/v2/tensorflow.keras.layers.-conv-l-s-t-m2-d.pbtxt b/tensorflow/tools/api/golden/v2/tensorflow.keras.layers.-conv-l-s-t-m2-d.pbtxt index f77d613e354..958d06a0d0f 100644 --- a/tensorflow/tools/api/golden/v2/tensorflow.keras.layers.-conv-l-s-t-m2-d.pbtxt +++ b/tensorflow/tools/api/golden/v2/tensorflow.keras.layers.-conv-l-s-t-m2-d.pbtxt @@ -207,7 +207,7 @@ tf_class { } member_method { name: "__init__" - argspec: "args=[\'self\', \'filters\', \'kernel_size\', \'strides\', \'padding\', \'data_format\', \'dilation_rate\', \'activation\', \'recurrent_activation\', \'use_bias\', \'kernel_initializer\', \'recurrent_initializer\', \'bias_initializer\', \'unit_forget_bias\', \'kernel_regularizer\', \'recurrent_regularizer\', \'bias_regularizer\', \'activity_regularizer\', \'kernel_constraint\', \'recurrent_constraint\', \'bias_constraint\', \'return_sequences\', \'go_backwards\', \'stateful\', \'dropout\', \'recurrent_dropout\'], varargs=None, keywords=kwargs, defaults=[\'(1, 1)\', \'valid\', \'None\', \'(1, 1)\', \'tanh\', \'hard_sigmoid\', \'True\', \'glorot_uniform\', \'orthogonal\', \'zeros\', \'True\', \'None\', \'None\', \'None\', \'None\', \'None\', \'None\', \'None\', \'False\', \'False\', \'False\', \'0.0\', \'0.0\'], " + argspec: "args=[\'self\', \'filters\', \'kernel_size\', \'strides\', \'padding\', \'data_format\', \'dilation_rate\', \'activation\', \'recurrent_activation\', \'use_bias\', \'kernel_initializer\', \'recurrent_initializer\', \'bias_initializer\', \'unit_forget_bias\', \'kernel_regularizer\', \'recurrent_regularizer\', \'bias_regularizer\', \'activity_regularizer\', \'kernel_constraint\', \'recurrent_constraint\', \'bias_constraint\', \'return_sequences\', \'return_state\', \'go_backwards\', \'stateful\', \'dropout\', \'recurrent_dropout\'], varargs=None, keywords=kwargs, defaults=[\'(1, 1)\', \'valid\', \'None\', \'(1, 1)\', \'tanh\', \'hard_sigmoid\', \'True\', \'glorot_uniform\', \'orthogonal\', \'zeros\', \'True\', \'None\', \'None\', \'None\', \'None\', \'None\', \'None\', \'None\', \'False\', \'False\', \'False\', \'False\', \'0.0\', \'0.0\'], " } member_method { name: "add_loss" From e08382691bfb897d584c5d5a8e8a0abe0472373d Mon Sep 17 00:00:00 2001 From: Chenkai Kuang Date: Thu, 18 Jun 2020 15:54:03 -0700 Subject: [PATCH 233/867] Make "map_resources" overridable by subclass of `Trackable`. This allows moving the implementation of map_resources from `tf.saved_model.save` to subclass of `Trackable`, e.g, Variable, DistributedVariable. This is a non-functional change. PiperOrigin-RevId: 317198449 Change-Id: I4aa48d4974b6547b5de8ac0f5c38f3da29d364bc --- tensorflow/python/distribute/BUILD | 4 +-- tensorflow/python/distribute/values.py | 12 +++++++ .../experimental/autocast_variable.py | 7 ++++ .../python/ops/resource_variable_ops.py | 7 ++++ tensorflow/python/saved_model/save.py | 36 +++++-------------- tensorflow/python/training/tracking/base.py | 18 ++++++++++ .../python/training/tracking/tracking.py | 13 +++++++ 7 files changed, 66 insertions(+), 31 deletions(-) diff --git a/tensorflow/python/distribute/BUILD b/tensorflow/python/distribute/BUILD index 96559a9a740..7208807a18c 100644 --- a/tensorflow/python/distribute/BUILD +++ b/tensorflow/python/distribute/BUILD @@ -744,14 +744,12 @@ py_library( "//tensorflow/python:control_flow_ops", "//tensorflow/python:framework_ops", "//tensorflow/python:math_ops", - "//tensorflow/python:tensor_util", + "//tensorflow/python:resource_variable_ops", "//tensorflow/python:tf_export", "//tensorflow/python:type_spec", - "//tensorflow/python:util", "//tensorflow/python:variable_scope", "//tensorflow/python:variables", "//tensorflow/python/eager:context", - "//tensorflow/python/eager:tape", "//tensorflow/python/training/saving:saveable_object", "//tensorflow/python/training/saving:saveable_object_util", "//tensorflow/python/training/tracking:base", diff --git a/tensorflow/python/distribute/values.py b/tensorflow/python/distribute/values.py index d0ed27c69de..60b2ea4fe31 100644 --- a/tensorflow/python/distribute/values.py +++ b/tensorflow/python/distribute/values.py @@ -32,6 +32,7 @@ from tensorflow.python.framework import type_spec from tensorflow.python.ops import array_ops from tensorflow.python.ops import control_flow_ops from tensorflow.python.ops import math_ops +from tensorflow.python.ops import resource_variable_ops from tensorflow.python.ops import variable_scope as vs from tensorflow.python.ops import variables as variables_lib from tensorflow.python.training.saving import saveable_object @@ -793,6 +794,17 @@ class DistributedVariable(DistributedDelegate, variables_lib.Variable, return ops.convert_to_tensor( self._get(), dtype=dtype, name=name, as_ref=as_ref) + def _map_resources(self): + """For implementing `Trackable`.""" + new_obj = resource_variable_ops.copy_to_graph_uninitialized(self._primary) + obj_map, resource_map = {}, {} + for v in self._values: + obj_map[v] = new_obj + resource_map[v.handle] = new_obj.handle + obj_map[self] = new_obj + resource_map[self] = new_obj.handle + return obj_map, resource_map + class _DistributedVariableSaveable(saveable_object.SaveableObject): """Class for defining how to restore a DistributedVariable.""" diff --git a/tensorflow/python/keras/mixed_precision/experimental/autocast_variable.py b/tensorflow/python/keras/mixed_precision/experimental/autocast_variable.py index 7d0abe30581..57e8ced65a0 100644 --- a/tensorflow/python/keras/mixed_precision/experimental/autocast_variable.py +++ b/tensorflow/python/keras/mixed_precision/experimental/autocast_variable.py @@ -285,6 +285,13 @@ class AutoCastVariable(variables.Variable, core.Tensor): # models with normal variables, and vice versa. return self._variable._gather_saveables_for_checkpoint() # pylint:disable=protected-access + def _map_resources(self): + # By delegating this method to the wrapped variable, SavedModel with + # AutoCastVariables are identical to SavedModel with normal variables. + obj_map, resource_map = self._variable._map_resources() # pylint:disable=protected-access + obj_map[self] = obj_map[self._variable] + return obj_map, resource_map + # TODO(reedwm): Maybe encode the fact the variable is an AutoCastVariable in # to_proto(). def to_proto(self, export_scope=None): diff --git a/tensorflow/python/ops/resource_variable_ops.py b/tensorflow/python/ops/resource_variable_ops.py index 25f6347f034..cb235fcbe2d 100644 --- a/tensorflow/python/ops/resource_variable_ops.py +++ b/tensorflow/python/ops/resource_variable_ops.py @@ -633,6 +633,13 @@ class BaseResourceVariable(variables.VariableV1, core.Tensor): return gen_state_ops.resource_count_up_to(self.handle, limit=limit, T=self.dtype) + def _map_resources(self): + """For implementing `Trackable`.""" + new_variable = copy_to_graph_uninitialized(self) + obj_map = {self: new_variable} + resource_map = {self._handle: new_variable.handle} + return obj_map, resource_map + def _read_variable_op(self): variable_accessed(self) diff --git a/tensorflow/python/saved_model/save.py b/tensorflow/python/saved_model/save.py index 5844c80995f..802ce1d61b7 100644 --- a/tensorflow/python/saved_model/save.py +++ b/tensorflow/python/saved_model/save.py @@ -19,14 +19,12 @@ from __future__ import division from __future__ import print_function import collections -import copy import os from tensorflow.core.framework import versions_pb2 from tensorflow.core.protobuf import meta_graph_pb2 from tensorflow.core.protobuf import saved_model_pb2 from tensorflow.core.protobuf import saved_object_graph_pb2 -from tensorflow.python.distribute import distribute_utils as ds_utils from tensorflow.python.eager import context from tensorflow.python.eager import def_function from tensorflow.python.eager import function as defun @@ -241,7 +239,7 @@ class _SaveableView(object): Creates resource handle ops in the current default graph, whereas `accessible_objects` will be from an eager context. Resource mapping adds resource handle ops to the main GraphDef of a SavedModel, which allows the - C++ loader API to interact with variables. + C++ loader API to interact with resources. Returns: A tuple of (object_map, resource_map, asset_info): @@ -265,33 +263,15 @@ class _SaveableView(object): asset_index={}) for node_id, obj in enumerate(self.nodes): - if isinstance(obj, tracking.CapturableResource): - new_obj = object_map[obj] = copy.copy(obj) - # pylint: disable=protected-access - with ops.device(obj._resource_device): - new_resource = new_obj._create_resource() - new_obj._resource_handle = new_resource - # pylint: enable=protected-access - resource_map[obj.resource_handle] = new_resource - self.captured_tensor_node_ids[obj.resource_handle] = node_id - elif (ds_utils.is_distributed_variable(obj) or - resource_variable_ops.is_resource_variable(obj)): - obj_to_copy = obj._primary if ds_utils.is_distributed_variable( # pylint: disable=protected-access - obj) else obj - new_variable = resource_variable_ops.copy_to_graph_uninitialized( - obj_to_copy) - if ds_utils.is_distributed_variable(obj): - self.captured_tensor_node_ids[obj] = node_id - for v in obj.values: - object_map[v] = new_variable - resource_map[v.handle] = new_variable.handle - self.captured_tensor_node_ids[v.handle] = node_id - object_map[obj] = new_variable - resource_map[obj.handle] = new_variable.handle - self.captured_tensor_node_ids[obj.handle] = node_id - elif isinstance(obj, tracking.Asset): + if isinstance(obj, tracking.Asset): _process_asset(obj, asset_info, resource_map) self.captured_tensor_node_ids[obj.asset_path] = node_id + elif isinstance(obj, base.Trackable): + node_object_map, node_resource_map = obj._map_resources() # pylint: disable=protected-access + for capturable in node_resource_map.keys(): + self.captured_tensor_node_ids[capturable] = node_id + object_map.update(node_object_map) + resource_map.update(node_resource_map) # Note: some concrete functions can have been realized when tracing other # functions, and might closure-capture tensors from their parent functions. diff --git a/tensorflow/python/training/tracking/base.py b/tensorflow/python/training/tracking/base.py index e3cd9828724..ea76ad8db47 100644 --- a/tensorflow/python/training/tracking/base.py +++ b/tensorflow/python/training/tracking/base.py @@ -1021,3 +1021,21 @@ class Trackable(object): """ del serialization_cache return dict() + + def _map_resources(self): + """Makes new resource handle ops corresponding to existing resource tensors. + + Internal sub-classes can override this to inform model saving how to add new + resource handle ops to the main GraphDef of a SavedModel (TF 1.x style + graph), which allows session based APIs (e.g, C++ loader API) to interact + with resources owned by this object. + + Returns: + A tuple of (object_map, resource_map): + object_map: A dictionary mapping from objects that hold existing + resource tensors to replacement objects created to hold the new + resource tensors. + resource_map: A dictionary mapping from existing resource tensors to + newly created resource tensors. + """ + return {}, {} diff --git a/tensorflow/python/training/tracking/tracking.py b/tensorflow/python/training/tracking/tracking.py index 553f0ec73bf..fb2735e6445 100644 --- a/tensorflow/python/training/tracking/tracking.py +++ b/tensorflow/python/training/tracking/tracking.py @@ -17,6 +17,7 @@ from __future__ import absolute_import from __future__ import division from __future__ import print_function +import copy import functools import weakref @@ -243,6 +244,18 @@ class CapturableResource(base.Trackable): self._resource_handle = self._create_resource() return self._resource_handle + def _map_resources(self): + """For implementing `Trackable`.""" + new_obj = copy.copy(self) + # pylint: disable=protected-access + with ops.device(self._resource_device): + new_resource = new_obj._create_resource() + new_obj._resource_handle = new_resource + # pylint: enable=protected-access + obj_map = {self: new_obj} + resource_map = {self.resource_handle: new_resource} + return obj_map, resource_map + def _list_functions_for_serialization(self, unused_functions): @def_function.function(input_signature=[], autograph=False) def _creator(): From 39504c25d9de697d3568bc4d370722d0f48376cf Mon Sep 17 00:00:00 2001 From: Smit Hinsu Date: Thu, 18 Jun 2020 15:55:24 -0700 Subject: [PATCH 234/867] Fix bug in xla-legalize-tf-with-tf2xla pass by handling non-tensor operands Currently, it only expects tensor operands but that is not applicable for non tensorflow dialect ops. PiperOrigin-RevId: 317198672 Change-Id: I1387e664de740d044ef535f6903e07d63fa02f6d --- .../mlir/xla/tests/legalize-tf-with-tf2xla.mlir | 12 ++++++++++-- .../mlir/xla/transforms/legalize_tf_with_tf2xla.cc | 6 +++--- 2 files changed, 13 insertions(+), 5 deletions(-) diff --git a/tensorflow/compiler/mlir/xla/tests/legalize-tf-with-tf2xla.mlir b/tensorflow/compiler/mlir/xla/tests/legalize-tf-with-tf2xla.mlir index b8a6df54519..86a7f2b9e09 100644 --- a/tensorflow/compiler/mlir/xla/tests/legalize-tf-with-tf2xla.mlir +++ b/tensorflow/compiler/mlir/xla/tests/legalize-tf-with-tf2xla.mlir @@ -35,7 +35,7 @@ func @not_whitelisted_op(%arg0: tensor<3xi32>, %arg1: tensor, %arg2: tensor // CHECK-LABEL: unranked_operand func @unranked_operand(%arg0: tensor<*xf32>) -> tensor<*xf32> { // CHECK: tf.Abs - // expected-remark@+1 {{lowering requires static shaped operands}} + // expected-remark@+1 {{lowering requires static shaped tensor operands}} %0 = "tf.Abs"(%arg0) : (tensor<*xf32>) -> tensor<*xf32> return %0 : tensor<*xf32> @@ -44,12 +44,20 @@ func @unranked_operand(%arg0: tensor<*xf32>) -> tensor<*xf32> { // CHECK-LABEL: dynamic_operand func @dynamic_operand(%arg0: tensor) -> tensor { // CHECK: tf.Abs - // expected-remark@+1 {{lowering requires static shaped operands}} + // expected-remark@+1 {{lowering requires static shaped tensor operands}} %0 = "tf.Abs"(%arg0) : (tensor) -> tensor return %0 : tensor } +// CHECK-LABEL: tuple_type +func @tuple_type(%arg0: tuple, tensor>) -> tensor { + // Verifies that the pass can handle operands of non-tensor type like tuple + // from non TensorFlow ops. + %0 = "xla_hlo.get_tuple_element"(%arg0) {index = 0 : i32} : (tuple, tensor>) -> tensor + return %0 : tensor +} + // CHECK-LABEL: unsupported_dtype func @unsupported_dtype(%arg0: tensor<2x!tf.variant>) -> tensor<2x!tf.variant> { // CHECK: tf.AddN diff --git a/tensorflow/compiler/mlir/xla/transforms/legalize_tf_with_tf2xla.cc b/tensorflow/compiler/mlir/xla/transforms/legalize_tf_with_tf2xla.cc index e57d6938efb..ef79c8868bb 100644 --- a/tensorflow/compiler/mlir/xla/transforms/legalize_tf_with_tf2xla.cc +++ b/tensorflow/compiler/mlir/xla/transforms/legalize_tf_with_tf2xla.cc @@ -337,9 +337,9 @@ LogicalResult FuncLegalizer::LegalizeOp(Operation* op) { // Only static shaped operands are supported in XLA builders for now. for (Type ty : op->getOperandTypes()) { - auto ranked_ty = ty.cast(); - if (!ranked_ty.hasStaticShape()) { - op->emitRemark() << "lowering requires static shaped operands"; + auto ranked_ty = ty.dyn_cast(); + if (!ranked_ty || !ranked_ty.hasStaticShape()) { + op->emitRemark() << "lowering requires static shaped tensor operands"; return success(); } } From 8d34408863b650564076f148edad9f91508abf04 Mon Sep 17 00:00:00 2001 From: Smit Hinsu Date: Thu, 18 Jun 2020 16:02:23 -0700 Subject: [PATCH 235/867] Auto-generate following TensorFlow ops related to image ResizeBilinearGrad ResizeBilinear AdjustContrastv2 ResizeNearestNeighbor AdjustSaturation AdjustHue RGBToHSV HSVToRGB PiperOrigin-RevId: 317199967 Change-Id: I1953acf599f2f7de686bda73b654e4c7b98dffd5 --- .../mlir/tensorflow/ir/tf_generated_ops.td | 153 ++++++++++++++++++ 1 file changed, 153 insertions(+) diff --git a/tensorflow/compiler/mlir/tensorflow/ir/tf_generated_ops.td b/tensorflow/compiler/mlir/tensorflow/ir/tf_generated_ops.td index 3a5deb9c569..dcd083fc398 100644 --- a/tensorflow/compiler/mlir/tensorflow/ir/tf_generated_ops.td +++ b/tensorflow/compiler/mlir/tensorflow/ir/tf_generated_ops.td @@ -164,6 +164,81 @@ def TF_AddV2Op : TF_Op<"AddV2", [Commutative, NoSideEffect, ResultsBroadcastable let hasFolder = 1; } +def TF_AdjustContrastv2Op : TF_Op<"AdjustContrastv2", [NoSideEffect]> { + let summary = "Adjust the contrast of one or more images."; + + let description = [{ +`images` is a tensor of at least 3 dimensions. The last 3 dimensions are +interpreted as `[height, width, channels]`. The other dimensions only +represent a collection of images, such as `[batch, height, width, channels].` + +Contrast is adjusted independently for each channel of each image. + +For each channel, the Op first computes the mean of the image pixels in the +channel and then adjusts each component of each pixel to +`(x - mean) * contrast_factor + mean`. + }]; + + let arguments = (ins + TensorOf<[F16, F32]>:$images, + F32Tensor:$contrast_factor + ); + + let results = (outs + TensorOf<[F16, F32]>:$output + ); + + TF_DerivedOperandTypeAttr T = TF_DerivedOperandTypeAttr<0>; +} + +def TF_AdjustHueOp : TF_Op<"AdjustHue", [NoSideEffect]> { + let summary = "Adjust the hue of one or more images."; + + let description = [{ +`images` is a tensor of at least 3 dimensions. The last dimension is +interpreted as channels, and must be three. + +The input image is considered in the RGB colorspace. Conceptually, the RGB +colors are first mapped into HSV. A delta is then applied all the hue values, +and then remapped back to RGB colorspace. + }]; + + let arguments = (ins + TensorOf<[F16, F32]>:$images, + F32Tensor:$delta + ); + + let results = (outs + TensorOf<[F16, F32]>:$output + ); + + TF_DerivedOperandTypeAttr T = TF_DerivedOperandTypeAttr<0>; +} + +def TF_AdjustSaturationOp : TF_Op<"AdjustSaturation", [NoSideEffect]> { + let summary = "Adjust the saturation of one or more images."; + + let description = [{ +`images` is a tensor of at least 3 dimensions. The last dimension is +interpreted as channels, and must be three. + +The input image is considered in the RGB colorspace. Conceptually, the RGB +colors are first mapped into HSV. A scale is then applied all the saturation +values, and then remapped back to RGB colorspace. + }]; + + let arguments = (ins + TensorOf<[F16, F32]>:$images, + F32Tensor:$scale + ); + + let results = (outs + TensorOf<[F16, F32]>:$output + ); + + TF_DerivedOperandTypeAttr T = TF_DerivedOperandTypeAttr<0>; +} + def TF_AllOp : TF_Op<"All", [NoSideEffect]> { let summary = [{ Computes the "logical and" of elements across dimensions of a tensor. @@ -3866,6 +3941,28 @@ tf.math.greater_equal(x, y) ==> [True, False, True, True] TF_DerivedOperandTypeAttr T = TF_DerivedOperandTypeAttr<0>; } +def TF_HSVToRGBOp : TF_Op<"HSVToRGB", [NoSideEffect]> { + let summary = "Convert one or more images from HSV to RGB."; + + let description = [{ +Outputs a tensor of the same shape as the `images` tensor, containing the RGB +value of the pixels. The output is only well defined if the value in `images` +are in `[0,1]`. + +See `rgb_to_hsv` for a description of the HSV encoding. + }]; + + let arguments = (ins + TF_FpTensor:$images + ); + + let results = (outs + TF_FpTensor:$output + ); + + TF_DerivedOperandTypeAttr T = TF_DerivedOperandTypeAttr<0>; +} + def TF_HashTableV2Op : TF_Op<"HashTableV2", []> { let summary = "Creates a non-initialized hash table."; @@ -6733,6 +6830,41 @@ the dimension is padded with zeros. TF_DerivedResultTypeAttr Tcomplex = TF_DerivedResultTypeAttr<0>; } +def TF_RGBToHSVOp : TF_Op<"RGBToHSV", [NoSideEffect]> { + let summary = "Converts one or more images from RGB to HSV."; + + let description = [{ +Outputs a tensor of the same shape as the `images` tensor, containing the HSV +value of the pixels. The output is only well defined if the value in `images` +are in `[0,1]`. + +`output[..., 0]` contains hue, `output[..., 1]` contains saturation, and +`output[..., 2]` contains value. All HSV values are in `[0,1]`. A hue of 0 +corresponds to pure red, hue 1/3 is pure green, and 2/3 is pure blue. + +Usage Example: + +>>> blue_image = tf.stack([ +... tf.zeros([5,5]), +... tf.zeros([5,5]), +... tf.ones([5,5])], +... axis=-1) +>>> blue_hsv_image = tf.image.rgb_to_hsv(blue_image) +>>> blue_hsv_image[0,0].numpy() +array([0.6666667, 1. , 1. ], dtype=float32) + }]; + + let arguments = (ins + TF_FpTensor:$images + ); + + let results = (outs + TF_FpTensor:$output + ); + + TF_DerivedOperandTypeAttr T = TF_DerivedOperandTypeAttr<0>; +} + def TF_RandomGammaGradOp : TF_Op<"RandomGammaGrad", [NoSideEffect, ResultsBroadcastableShape]>, WithBroadcastableBinOpBuilder { let summary = [{ @@ -7230,6 +7362,27 @@ Input images can be of different types but output images are always float. TF_DerivedOperandTypeAttr T = TF_DerivedOperandTypeAttr<0>; } +def TF_ResizeBilinearGradOp : TF_Op<"ResizeBilinearGrad", [NoSideEffect]> { + let summary = "Computes the gradient of bilinear interpolation."; + + let description = [{ + }]; + + let arguments = (ins + F32Tensor:$grads, + TF_FpTensor:$original_image, + + DefaultValuedAttr:$align_corners, + DefaultValuedAttr:$half_pixel_centers + ); + + let results = (outs + TF_FpTensor:$output + ); + + TF_DerivedOperandTypeAttr T = TF_DerivedOperandTypeAttr<1>; +} + def TF_ResizeNearestNeighborOp : TF_Op<"ResizeNearestNeighbor", [NoSideEffect]> { let summary = [{ Resize `images` to `size` using nearest neighbor interpolation. From 67544cd4bbdf8070adebbb077439cac300f479ca Mon Sep 17 00:00:00 2001 From: Ken Franko Date: Thu, 18 Jun 2020 16:02:51 -0700 Subject: [PATCH 236/867] Add more outside compilation tests including multiple clusters and more variety of inputs/outputs. PiperOrigin-RevId: 317200078 Change-Id: Id26e99059097073299ef5f681fae053b082ec149 --- .../tpu/tpu_outside_compilation_test.py | 95 ++++++++++++++++++- 1 file changed, 91 insertions(+), 4 deletions(-) diff --git a/tensorflow/python/tpu/tpu_outside_compilation_test.py b/tensorflow/python/tpu/tpu_outside_compilation_test.py index f7ecb294c44..54c2598324c 100644 --- a/tensorflow/python/tpu/tpu_outside_compilation_test.py +++ b/tensorflow/python/tpu/tpu_outside_compilation_test.py @@ -18,13 +18,18 @@ from __future__ import absolute_import from __future__ import division from __future__ import print_function +from absl.testing import parameterized +import numpy as np + from tensorflow.python.distribute import tpu_strategy as tpu_lib from tensorflow.python.distribute.cluster_resolver import tpu_cluster_resolver from tensorflow.python.eager import def_function from tensorflow.python.eager import remote from tensorflow.python.eager import test from tensorflow.python.framework import constant_op +from tensorflow.python.ops import array_ops from tensorflow.python.ops import logging_ops +from tensorflow.python.ops import math_ops from tensorflow.python.ops import variables from tensorflow.python.platform import flags from tensorflow.python.tpu import tpu @@ -52,7 +57,7 @@ def get_tpu_strategy(): return tpu_lib.TPUStrategy(resolver) -class TpuOutsideCompilationTest(test.TestCase): +class TpuOutsideCompilationTest(test.TestCase, parameterized.TestCase): def testResourceVariableAssignOnHost(self): strategy = get_tpu_strategy() @@ -79,6 +84,26 @@ class TpuOutsideCompilationTest(test.TestCase): self.assertAllEqual(4.0 * strategy.num_replicas_in_sync, v2.numpy()) self.assertAllEqual(5.0, v.numpy()) + def testHostNoInput(self): + strategy = get_tpu_strategy() + + def outside_fn(): + logging_ops.print_v2("Outside compiled") + + @def_function.function + def train_step(): + + def tpu_fn(x): + x2 = x + 5.0 + tpu.outside_compilation(outside_fn) + return x2 + 5.0 + + return strategy.run(tpu_fn, args=(25.0,)) + + self.assertAllEqual( + strategy.experimental_local_results(train_step()), + constant_op.constant(35., shape=(strategy.num_replicas_in_sync))) + def testHostInputOnly(self): strategy = get_tpu_strategy() @@ -120,13 +145,71 @@ class TpuOutsideCompilationTest(test.TestCase): strategy.experimental_local_results(train_step()), constant_op.constant(36., shape=(strategy.num_replicas_in_sync))) - def testOutsideCompilationControlFlowIf(self): + def testHostMultipleInputs(self): + strategy = get_tpu_strategy() + val0 = np.arange(6).reshape((2, 3)).astype(np.float32) + val1 = np.arange(6).reshape((3, 2)).astype(np.float32) + + def outside_fn(arg0, arg1): + tmp = array_ops.reshape(arg1, array_ops.shape(arg0)) + ret0 = arg0 + tmp + ret1 = math_ops.matmul(arg0, arg1) + ret2 = array_ops.concat([arg0, tmp], 0) + return ret0, ret1, ret2 + + @def_function.function + def train_step(): + + def tpu_fn(x, y): + a = x + 7.0 + b = y * 2.0 + c, d, e = tpu.outside_compilation(outside_fn, a, b) + return (math_ops.reduce_max(c) + math_ops.reduce_min(d) + + math_ops.reduce_sum(e)) + + return strategy.run(tpu_fn, args=(val0, val1)) + + self.assertAllEqual( + strategy.experimental_local_results(train_step()), + constant_op.constant(213., shape=(strategy.num_replicas_in_sync))) + + def testMultipleClusters(self): + strategy = get_tpu_strategy() + + def outside_fn1(x): + logging_ops.print_v2("Outside compiled", x) + return x + 6.0 + + def outside_fn2(x): + logging_ops.print_v2("Outside compiled", x) + return x - 18.0 + + @def_function.function + def train_step(): + + def tpu_fn(x): + x2 = x + 5.0 + output1 = tpu.outside_compilation(outside_fn1, x2) + x3 = output1 + 3.0 + output2 = tpu.outside_compilation(outside_fn2, x3) + return output2 + + return strategy.run(tpu_fn, args=(25.0,)) + + self.assertAllEqual( + strategy.experimental_local_results(train_step()), + constant_op.constant(21., shape=(strategy.num_replicas_in_sync))) + + @parameterized.parameters((True), (False)) + def testOutsideCompilationControlFlowIf(self, take_true_branch): strategy = get_tpu_strategy() def outside_fn(x): logging_ops.print_v2("Outside compiled", x) return x + 6.0 + input_value = 51.0 if take_true_branch else 25.0 + @def_function.function def train_step(): @@ -137,11 +220,15 @@ class TpuOutsideCompilationTest(test.TestCase): else: return x2 - return strategy.run(tpu_fn, args=(25.0,)) + return strategy.run(tpu_fn, args=(input_value,)) + output_value = 36.0 + if take_true_branch: + output_value = 56.0 self.assertAllEqual( strategy.experimental_local_results(train_step()), - constant_op.constant(36., shape=(strategy.num_replicas_in_sync))) + constant_op.constant( + output_value, shape=(strategy.num_replicas_in_sync))) def testOutsideCompilationControlFlowWhile(self): strategy = get_tpu_strategy() From 4aea552e064cf92330e07e83a3b5a1ca2a7034d0 Mon Sep 17 00:00:00 2001 From: Henry Tan Date: Thu, 18 Jun 2020 16:15:22 -0700 Subject: [PATCH 237/867] Publishing tpu_op_consts to tpu kernels library. PiperOrigin-RevId: 317202394 Change-Id: Ib6a1f350af7384513a3744084a9959ed86278d1f --- tensorflow/core/tpu/kernels/BUILD | 11 +++++- .../kernels/tpu_compilation_cache_external.h | 5 +-- tensorflow/core/tpu/kernels/tpu_op_consts.cc | 24 ++++++++++++ tensorflow/core/tpu/kernels/tpu_op_consts.h | 39 +++++++++++++++++++ 4 files changed, 74 insertions(+), 5 deletions(-) create mode 100644 tensorflow/core/tpu/kernels/tpu_op_consts.cc create mode 100644 tensorflow/core/tpu/kernels/tpu_op_consts.h diff --git a/tensorflow/core/tpu/kernels/BUILD b/tensorflow/core/tpu/kernels/BUILD index a41747ee8c5..a9f2202cd45 100644 --- a/tensorflow/core/tpu/kernels/BUILD +++ b/tensorflow/core/tpu/kernels/BUILD @@ -321,6 +321,7 @@ cc_library( ":tpu_compile_c_api_hdrs", ":tpu_compile_op_support", ":tpu_mesh_state_interface", + ":tpu_op_consts", ":tpu_program_group", ":tpu_util", ":trace_util_hdrs", @@ -433,7 +434,6 @@ cc_library( "//tensorflow/compiler/xla/client:compile_only_client", "//tensorflow/core:lib", "//tensorflow/core:protos_all_cc", - "@com_google_absl//absl/status", "@com_google_absl//absl/strings", ], alwayslink = 1, @@ -461,3 +461,12 @@ cc_library( "@com_google_absl//absl/synchronization", ], ) + +cc_library( + name = "tpu_op_consts", + srcs = ["tpu_op_consts.cc"], + hdrs = ["tpu_op_consts.h"], + deps = [ + "@com_google_absl//absl/base:core_headers", + ], +) diff --git a/tensorflow/core/tpu/kernels/tpu_compilation_cache_external.h b/tensorflow/core/tpu/kernels/tpu_compilation_cache_external.h index fe251326a43..86615b15d4c 100644 --- a/tensorflow/core/tpu/kernels/tpu_compilation_cache_external.h +++ b/tensorflow/core/tpu/kernels/tpu_compilation_cache_external.h @@ -38,15 +38,12 @@ limitations under the License. #include "tensorflow/core/tpu/kernels/tpu_compile_c_api.h" #include "tensorflow/core/tpu/kernels/tpu_compile_op_support.h" #include "tensorflow/core/tpu/kernels/tpu_mesh_state_interface.h" +#include "tensorflow/core/tpu/kernels/tpu_op_consts.h" #include "tensorflow/core/tpu/kernels/tpu_program_group.h" namespace tensorflow { namespace tpu { -constexpr char kCompilationCacheResourceName[] = "tpu_compilation_cache"; -constexpr char kCompilationCacheUnloaderResourceName[] = - "tpu_compilation_cache_unloader"; - class TpuCompilationCacheExternal : public TpuCompilationCacheInterface { public: using Status = ::stream_executor::port::Status; diff --git a/tensorflow/core/tpu/kernels/tpu_op_consts.cc b/tensorflow/core/tpu/kernels/tpu_op_consts.cc new file mode 100644 index 00000000000..e5e1aacb3cc --- /dev/null +++ b/tensorflow/core/tpu/kernels/tpu_op_consts.cc @@ -0,0 +1,24 @@ +/* Copyright 2020 The TensorFlow Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. +==============================================================================*/ +#include "tensorflow/core/tpu/kernels/tpu_op_consts.h" + +namespace tensorflow { +namespace tpu { +const char kCompilationCacheResourceName[] = "tpu_compilation_cache"; +const char kCompiledProtoCacheResourceName[] = "tpu_proto_cache"; +const char kCompilationCacheUnloaderResourceName[] = + "tpu_compilation_cache_unloader"; +} // namespace tpu +} // namespace tensorflow diff --git a/tensorflow/core/tpu/kernels/tpu_op_consts.h b/tensorflow/core/tpu/kernels/tpu_op_consts.h new file mode 100644 index 00000000000..25223b7e429 --- /dev/null +++ b/tensorflow/core/tpu/kernels/tpu_op_consts.h @@ -0,0 +1,39 @@ +/* Copyright 2020 The TensorFlow Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. +==============================================================================*/ +#ifndef TENSORFLOW_CORE_TPU_KERNELS_TPU_OP_CONSTS_H_ +#define TENSORFLOW_CORE_TPU_KERNELS_TPU_OP_CONSTS_H_ + +#include "absl/base/attributes.h" + +namespace tensorflow { +namespace tpu { + +// Resource names in the ResourceMgr. +// +// Name of cache for compiled TPU ISA protos. CompilationCache is created by +// ConfigureDistributedTpuOp, so only the master has a CompilationCache. +ABSL_CONST_INIT extern const char kCompilationCacheResourceName[]; +// Name of base class allowing Execute Ops to look up ISA protos. +// CompiledProtoCache is created by InitializeHostForDistributedTpuOp, so each +// tpu_worker has a CompiledProtoCache. +ABSL_CONST_INIT extern const char kCompiledProtoCacheResourceName[]; +// Name of cache unloader for compiled TPU ISA protos. Cache unloader should be +// put into TPU_SYSTEM device resource manager. Inference may use it to unload +// cache entries created during lifetime of a DirectSession. +ABSL_CONST_INIT extern const char kCompilationCacheUnloaderResourceName[]; + +} // namespace tpu +} // namespace tensorflow +#endif // TENSORFLOW_CORE_TPU_KERNELS_TPU_OP_CONSTS_H_ From aab151356d2334a9d6cec71ce5165e6e6c45c793 Mon Sep 17 00:00:00 2001 From: "A. Unique TensorFlower" Date: Thu, 18 Jun 2020 16:24:14 -0700 Subject: [PATCH 238/867] [pjrt] Refresh stream error status in strategic places to flush out silent failures. PiperOrigin-RevId: 317204018 Change-Id: If75a3ad9ec846ce1621cdba92a2dc738b65b7001 --- tensorflow/compiler/xla/pjrt/local_device_state.cc | 4 ++++ tensorflow/compiler/xla/pjrt/pjrt_client.cc | 12 +++++++++--- tensorflow/stream_executor/stream.cc | 7 ++++++- 3 files changed, 19 insertions(+), 4 deletions(-) diff --git a/tensorflow/compiler/xla/pjrt/local_device_state.cc b/tensorflow/compiler/xla/pjrt/local_device_state.cc index d173c891c95..a229e56001e 100644 --- a/tensorflow/compiler/xla/pjrt/local_device_state.cc +++ b/tensorflow/compiler/xla/pjrt/local_device_state.cc @@ -127,11 +127,15 @@ std::unique_ptr LocalDeviceState::BorrowStreamFromPool() { } else { std::unique_ptr stream = std::move(usage_stream_pool_.top()); usage_stream_pool_.pop(); + stream->RefreshStatus().IgnoreError(); // Can return error::Unimplemented + QCHECK(stream->ok()); return stream; } } void LocalDeviceState::ReturnStreamToPool(std::unique_ptr stream) { + stream->RefreshStatus().IgnoreError(); // Can return error::Unimplemented + QCHECK(stream->ok()); absl::MutexLock lock(&mu_); usage_stream_pool_.push(std::move(stream)); } diff --git a/tensorflow/compiler/xla/pjrt/pjrt_client.cc b/tensorflow/compiler/xla/pjrt/pjrt_client.cc index ef259cf1cfd..46f592100c9 100644 --- a/tensorflow/compiler/xla/pjrt/pjrt_client.cc +++ b/tensorflow/compiler/xla/pjrt/pjrt_client.cc @@ -751,16 +751,22 @@ StatusOr> PjRtBuffer::FromHostLiteral( // memory that has already been allocated, and a possible Event // allocation. + se::Stream* h2d_stream = local_device->host_to_device_stream(); ShapedBuffer buffer = device_buffer->AsShapedBuffer( compact_shape, on_device_shape, client->client()->platform()); TF_CHECK_OK(transfer_manager->TransferLiteralToDeviceAsync( - local_device->host_to_device_stream(), literal, buffer)); + h2d_stream, literal, buffer)); std::shared_ptr event = device_buffer->definition_events()[0]; TF_CHECK_OK(AddDestinationBufferSynchronization( - local_device, std::move(device_buffer), event, - local_device->host_to_device_stream())); + local_device, std::move(device_buffer), event, h2d_stream)); + + // This can sometimes catch the case where the literal memory has been + // freed before the H2D transfer was issued. + h2d_stream->RefreshStatus() + .IgnoreError(); // Can return error::Unimplemented + QCHECK(h2d_stream->ok()); }; client->h2d_transfer_pool()->Schedule(transfer_h2d); return py_buffer; diff --git a/tensorflow/stream_executor/stream.cc b/tensorflow/stream_executor/stream.cc index c63565c65a8..da418122375 100644 --- a/tensorflow/stream_executor/stream.cc +++ b/tensorflow/stream_executor/stream.cc @@ -285,7 +285,12 @@ Stream::~Stream() { port::Status Stream::RefreshStatus() { port::Status status = parent_->GetStatus(this); - CheckStatus(status); + // We should not put the stream in an error state, just because the GetStatus + // method is unimplemented. + if (status != port::Status(port::error::UNIMPLEMENTED, + "GetStatus is not supported on this executor.")) { + CheckStatus(status); + } return status; } From 1363f0f6e89cdc73ca46d43475bdb35750ed2e50 Mon Sep 17 00:00:00 2001 From: Yixing Fu Date: Thu, 18 Jun 2020 19:28:34 -0400 Subject: [PATCH 239/867] add save load test for pathlib path --- tensorflow/python/keras/saving/save_test.py | 13 +++++++++++++ 1 file changed, 13 insertions(+) diff --git a/tensorflow/python/keras/saving/save_test.py b/tensorflow/python/keras/saving/save_test.py index 5c5846fe738..98d78735ad1 100644 --- a/tensorflow/python/keras/saving/save_test.py +++ b/tensorflow/python/keras/saving/save_test.py @@ -71,6 +71,12 @@ class TestSaveModel(test.TestCase, parameterized.TestCase): save.save_model(self.model, path) self.assert_saved_model(path) + @test_util.run_v2_only + def test_save_format_defaults_pathlib(self): + path = pathlib.Path(self.get_temp_dir()) / 'model_path' + save.save_model(self.model, path) + self.assert_saved_model(path) + @test_util.run_v2_only def test_save_hdf5(self): path = os.path.join(self.get_temp_dir(), 'model') @@ -81,6 +87,13 @@ class TestSaveModel(test.TestCase, parameterized.TestCase): 'requires the model to be a Functional model or a Sequential model.'): save.save_model(self.subclassed_model, path, save_format='h5') + @test_util.run_v2_only + def test_save_load_hdf5_pathlib(self): + if sys.version_info >= (3, 6): + path = pathlib.Path(self.get_temp_dir()) / 'model' + save.save_model(self.model, path, save_format='h5') + save.load_model(path) + @test_util.run_v2_only def test_save_tf(self): path = os.path.join(self.get_temp_dir(), 'model') From 94e37f84f19384e685420ef7f90382fcfe719498 Mon Sep 17 00:00:00 2001 From: Dero Gharibian Date: Thu, 18 Jun 2020 16:25:11 -0700 Subject: [PATCH 240/867] Remove unnecessary string copy. PiperOrigin-RevId: 317204219 Change-Id: I85fab345945b6ea4f428f8aedc861eb79e5fd7e0 --- tensorflow/core/kernels/summary_op.cc | 6 ++++-- 1 file changed, 4 insertions(+), 2 deletions(-) diff --git a/tensorflow/core/kernels/summary_op.cc b/tensorflow/core/kernels/summary_op.cc index 386a8964dba..f4c91fc9ff1 100644 --- a/tensorflow/core/kernels/summary_op.cc +++ b/tensorflow/core/kernels/summary_op.cc @@ -52,7 +52,8 @@ class SummaryScalarOp : public OpKernel { Summary s; for (int i = 0; i < Ttags.size(); i++) { Summary::Value* v = s.add_value(); - v->set_tag(string(Ttags(i))); // NOLINT + const tstring& Ttags_i = Ttags(i); + v->set_tag(Ttags_i.data(), Ttags_i.size()); v->set_simple_value(float(Tvalues(i))); } @@ -102,7 +103,8 @@ class SummaryHistoOp : public OpKernel { Summary s; Summary::Value* v = s.add_value(); - v->set_tag(string(tags.scalar()())); // NOLINT + const tstring& tags0 = tags.scalar()(); + v->set_tag(tags0.data(), tags0.size()); histo.EncodeToProto(v->mutable_histo(), false /* Drop zero buckets */); Tensor* summary_tensor = nullptr; From e7733c8dd4cae1a1a03053b452acbbded7ab6014 Mon Sep 17 00:00:00 2001 From: Reed Date: Thu, 18 Jun 2020 16:52:51 -0700 Subject: [PATCH 241/867] Add LINT.IfChange --- .../core/grappler/optimizers/auto_mixed_precision_lists.h | 2 ++ 1 file changed, 2 insertions(+) diff --git a/tensorflow/core/grappler/optimizers/auto_mixed_precision_lists.h b/tensorflow/core/grappler/optimizers/auto_mixed_precision_lists.h index c6016548117..9041d934017 100644 --- a/tensorflow/core/grappler/optimizers/auto_mixed_precision_lists.h +++ b/tensorflow/core/grappler/optimizers/auto_mixed_precision_lists.h @@ -73,6 +73,7 @@ class AutoMixedPrecisionLists { static void AddTensorListOps(gtl::FlatSet* list) { // Note: if a data structure op (such as TensorListPopBack) is added here, // IsTensorListReaderOp or IsTensorListWriterOp may need to be modified + // LINT.IfChange constexpr char* tensor_list_ops[] = { "TensorListConcat", "TensorListConcatLists", @@ -90,6 +91,7 @@ class AutoMixedPrecisionLists { "TensorListSplit", "TensorListStack" }; + // LINT.ThenChange(//tensorflow/core/grappler/optimizers/auto_mixed_precision.cc) for (auto op : tensor_list_ops) { list->insert(op); } From 1c72a6c65e1733b55286c6361142a39d699732dc Mon Sep 17 00:00:00 2001 From: Yixing Fu Date: Thu, 18 Jun 2020 19:53:22 -0400 Subject: [PATCH 242/867] convert export_dir as pathlike object to str in parse_saved_model --- tensorflow/python/saved_model/loader_impl.py | 4 ++-- 1 file changed, 2 insertions(+), 2 deletions(-) diff --git a/tensorflow/python/saved_model/loader_impl.py b/tensorflow/python/saved_model/loader_impl.py index 2df2bea428e..71f6ed16c9d 100644 --- a/tensorflow/python/saved_model/loader_impl.py +++ b/tensorflow/python/saved_model/loader_impl.py @@ -83,11 +83,11 @@ def parse_saved_model(export_dir): """ # Build the path to the SavedModel in pbtxt format. path_to_pbtxt = os.path.join( - compat.as_bytes(export_dir), + compat.as_bytes(compat.path_to_str(export_dir)), compat.as_bytes(constants.SAVED_MODEL_FILENAME_PBTXT)) # Build the path to the SavedModel in pb format. path_to_pb = os.path.join( - compat.as_bytes(export_dir), + compat.as_bytes(compat.path_to_str(export_dir)), compat.as_bytes(constants.SAVED_MODEL_FILENAME_PB)) # Parse the SavedModel protocol buffer. From 83b09270dc34308ef60f2f68de540a4cb213e1e4 Mon Sep 17 00:00:00 2001 From: Lucy Fox Date: Thu, 18 Jun 2020 16:52:46 -0700 Subject: [PATCH 243/867] Update FusedKernelMatcher pass to use upstream util to get stripped op name. Added this util upstream in D81435, so now using that instead and deleting the unneeded code here. PiperOrigin-RevId: 317209256 Change-Id: Id2d8a1fca34ca85e59a05a85bf7f6f59b425c7c1 --- .../tensorflow/transforms/fused_kernel_matcher.cc | 11 +++-------- 1 file changed, 3 insertions(+), 8 deletions(-) diff --git a/tensorflow/compiler/mlir/tensorflow/transforms/fused_kernel_matcher.cc b/tensorflow/compiler/mlir/tensorflow/transforms/fused_kernel_matcher.cc index 4b10550df7b..d10f5e26e8f 100644 --- a/tensorflow/compiler/mlir/tensorflow/transforms/fused_kernel_matcher.cc +++ b/tensorflow/compiler/mlir/tensorflow/transforms/fused_kernel_matcher.cc @@ -52,11 +52,6 @@ struct FusedKernelMatcherPass void runOnFunction() override; }; -// Returns an op's name with the dialect prefix stripped off. -StringRef GetOpNameWithoutDialect(Operation *op) { - return op->getName().getStringRef().split(".").second; -} - bool IsActivationFunction(Operation *op) { return isa(op) || isa(op) || isa(op); } @@ -128,8 +123,8 @@ class FuseContractionWithBiasAdd : public OpRewritePattern { } SmallVector locations{contraction.getLoc(), bias_add.getLoc()}; - SmallVector fused_ops{ - StringAttr::get(GetOpNameWithoutDialect(bias_add), context)}; + SmallVector fused_ops{StringAttr::get( + bias_add.getOperation()->getName().stripDialect(), context)}; // BiasAdd may or may not feed into an activation function. auto activation = GetActivation(bias_add); @@ -143,7 +138,7 @@ class FuseContractionWithBiasAdd : public OpRewritePattern { if (fuse_activation) { locations.push_back(activation->getLoc()); fused_ops.push_back( - StringAttr::get(GetOpNameWithoutDialect(activation), context)); + StringAttr::get(activation->getName().stripDialect(), context)); result_type = activation->getResultTypes().front(); } else { result_type = bias_add.getResult().getType(); From 5f2e0240ee7977042e41d9c29c349a7b14301290 Mon Sep 17 00:00:00 2001 From: Yixing Fu Date: Thu, 18 Jun 2020 19:57:23 -0400 Subject: [PATCH 244/867] modify docstring --- tensorflow/python/saved_model/loader_impl.py | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/tensorflow/python/saved_model/loader_impl.py b/tensorflow/python/saved_model/loader_impl.py index 71f6ed16c9d..06cd988130d 100644 --- a/tensorflow/python/saved_model/loader_impl.py +++ b/tensorflow/python/saved_model/loader_impl.py @@ -73,7 +73,7 @@ def parse_saved_model(export_dir): """Reads the savedmodel.pb or savedmodel.pbtxt file containing `SavedModel`. Args: - export_dir: Directory containing the SavedModel file. + export_dir: String or Pathlike, path to the directory containing the SavedModel file. Returns: A `SavedModel` protocol buffer. From 86770b177e88cfa25844188a655f8399f5c3526d Mon Sep 17 00:00:00 2001 From: jonah-kohn <51345541+jonah-kohn@users.noreply.github.com> Date: Thu, 18 Jun 2020 17:00:04 -0700 Subject: [PATCH 245/867] Add show_dtype param to plot_model and model_to_dot. --- tensorflow/python/keras/utils/vis_utils.py | 29 +++++++++++++++++++--- 1 file changed, 25 insertions(+), 4 deletions(-) diff --git a/tensorflow/python/keras/utils/vis_utils.py b/tensorflow/python/keras/utils/vis_utils.py index 3720708543f..21de8014c2a 100644 --- a/tensorflow/python/keras/utils/vis_utils.py +++ b/tensorflow/python/keras/utils/vis_utils.py @@ -69,6 +69,7 @@ def add_edge(dot, src, dst): @keras_export('keras.utils.model_to_dot') def model_to_dot(model, show_shapes=False, + show_dtype=False, show_layer_names=True, rankdir='TB', expand_nested=False, @@ -79,6 +80,7 @@ def model_to_dot(model, Arguments: model: A Keras model instance. show_shapes: whether to display shape information. + show_dtype: whether to display layer dtypes. show_layer_names: whether to display layer names. rankdir: `rankdir` argument passed to PyDot, a string specifying the format of the plot: @@ -150,8 +152,11 @@ def model_to_dot(model, if isinstance(layer, wrappers.Wrapper): if expand_nested and isinstance(layer.layer, functional.Functional): - submodel_wrapper = model_to_dot(layer.layer, show_shapes, - show_layer_names, rankdir, + submodel_wrapper = model_to_dot(layer.layer, + show_shapes, + show_dtype, + show_layer_names, + rankdir, expand_nested, subgraph=True) # sub_w : submodel_wrapper @@ -165,8 +170,11 @@ def model_to_dot(model, class_name = '{}({})'.format(class_name, child_class_name) if expand_nested and isinstance(layer, functional.Functional): - submodel_not_wrapper = model_to_dot(layer, show_shapes, - show_layer_names, rankdir, + submodel_not_wrapper = model_to_dot(layer, + show_shapes, + show_dtype, + show_layer_names, + rankdir, expand_nested, subgraph=True) # sub_n : submodel_not_wrapper @@ -180,6 +188,16 @@ def model_to_dot(model, label = '{}: {}'.format(layer_name, class_name) else: label = class_name + + # Rebuild the label as a table including the layer's dtype. + if show_dtype: + def format_dtype(dtype): + if dtype is None: + return '?' + else: + return str(dtype) + + label = '%s|%s' % (label, dtype) # Rebuild the label as a table including input/output shapes. if show_shapes: @@ -260,6 +278,7 @@ def model_to_dot(model, def plot_model(model, to_file='model.png', show_shapes=False, + show_dtype=False, show_layer_names=True, rankdir='TB', expand_nested=False, @@ -286,6 +305,7 @@ def plot_model(model, model: A Keras model instance to_file: File name of the plot image. show_shapes: whether to display shape information. + show_dtype: whether to display layer dtypes. show_layer_names: whether to display layer names. rankdir: `rankdir` argument passed to PyDot, a string specifying the format of the plot: @@ -300,6 +320,7 @@ def plot_model(model, """ dot = model_to_dot(model, show_shapes=show_shapes, + show_dtype=show_dtype, show_layer_names=show_layer_names, rankdir=rankdir, expand_nested=expand_nested, From c3ca4da9b46e3148bdb913d08fec3fd2727158e1 Mon Sep 17 00:00:00 2001 From: Yixing Fu Date: Thu, 18 Jun 2020 20:00:06 -0400 Subject: [PATCH 246/867] only test pathlib on python v>3.6 --- tensorflow/python/keras/saving/save_test.py | 7 ++++--- 1 file changed, 4 insertions(+), 3 deletions(-) diff --git a/tensorflow/python/keras/saving/save_test.py b/tensorflow/python/keras/saving/save_test.py index 98d78735ad1..66364666841 100644 --- a/tensorflow/python/keras/saving/save_test.py +++ b/tensorflow/python/keras/saving/save_test.py @@ -73,9 +73,10 @@ class TestSaveModel(test.TestCase, parameterized.TestCase): @test_util.run_v2_only def test_save_format_defaults_pathlib(self): - path = pathlib.Path(self.get_temp_dir()) / 'model_path' - save.save_model(self.model, path) - self.assert_saved_model(path) + if sys.version_info >= (3, 6): + path = pathlib.Path(self.get_temp_dir()) / 'model_path' + save.save_model(self.model, path) + self.assert_saved_model(path) @test_util.run_v2_only def test_save_hdf5(self): From e96543f6fbba9fa112d7ca1d731b64e3654e1629 Mon Sep 17 00:00:00 2001 From: Ran Chen Date: Thu, 18 Jun 2020 16:55:26 -0700 Subject: [PATCH 247/867] Add MultiProcessPoolRunner Tensorflow initialization can take a long time when GPUs are present. We cannot afford starting a new group of workers for every single test. MultiProcessPoolRunner uses a pool of workers so that we can avoid the initialization cost. Compared to MultiProcessRunner, it doesn't support terminating workers. Note that implementation wise we could build MultiProcessPoolRunner on top of MultiProcessRunner or vice-versa if there's no need to support termination. Since it's easier for MultiProcessPoolRunner not to support termination, we choose MultiProcessPoolRunner on top of MultiProcessRunner. PiperOrigin-RevId: 317209754 Change-Id: Ia439028c81c5a9f87b16d631a170158724ce47d4 --- .../python/distribute/multi_process_runner.py | 205 ++++++++++++++++-- .../distribute/multi_process_runner_test.py | 69 +++++- 2 files changed, 249 insertions(+), 25 deletions(-) diff --git a/tensorflow/python/distribute/multi_process_runner.py b/tensorflow/python/distribute/multi_process_runner.py index 8699e59b410..ce36287a9da 100644 --- a/tensorflow/python/distribute/multi_process_runner.py +++ b/tensorflow/python/distribute/multi_process_runner.py @@ -67,8 +67,7 @@ except ImportError: # exception stack trace info is stored in exc_info to pass on to parent process # to be re-raised. _ProcessStatusInfo = collections.namedtuple( - '_ProcessStatusInfo', - ['task_type', 'is_successful', 'exc_info', 'return_value']) + '_ProcessStatusInfo', ['is_successful', 'exc_info', 'return_value']) # Information returned from a successful MultiProcessRunner run. MultiProcessRunnerResult = collections.namedtuple('MultiProcessRunnerResult', @@ -124,6 +123,7 @@ class MultiProcessRunner(object): stream_stdout=True, list_stdout=False, use_dill_for_args=True, + daemon=False, args=None, kwargs=None): """Creates a multi-process runner. @@ -157,6 +157,7 @@ class MultiProcessRunner(object): use_dill_for_args: Whether to use dill to pickle `args` and `kwargs`. dill can pickle more objects, but doesn't work with types in `multiprocessing` library like `Mutex`. + daemon: Whether to start processes as daemons. args: Positional arguments to be sent to functions run on processes. kwargs: Keyword arguments to be sent to functions run on processes. @@ -188,6 +189,7 @@ class MultiProcessRunner(object): self._list_stdout = list_stdout self._dependence_on_chief = True self._use_dill_for_args = use_dill_for_args + self._daemon = daemon self._args = args or () self._kwargs = kwargs or {} @@ -268,7 +270,8 @@ class MultiProcessRunner(object): test_env=test_env, target=_ProcFunc(), args=(resources, test_env, proc_func, args, kwargs, - self._use_dill_for_args)) + self._use_dill_for_args), + daemon=self._daemon) p.start() self._processes[(task_type, task_id)] = p self._outstanding_subprocess_count += 1 @@ -568,7 +571,6 @@ class _ProcFunc(object): time.sleep(0.1) self._resources.process_status_queue.put( _ProcessStatusInfo( - task_type=task_type, is_successful=True, exc_info=None, return_value=None)) @@ -628,17 +630,9 @@ class _ProcFunc(object): if test_env.v2_enabled: v2_compat.enable_v2_behavior() - try: - with self._runtime_mode(test_env.executing_eagerly): - return_value = proc_func(*args, **kwargs) - is_successful = True - exc_info = None - - except Exception: # pylint: disable=broad-except - # Capture all exceptions to be reported to parent process. - return_value = None - is_successful = False - exc_info = sys.exc_info() + with self._runtime_mode(test_env.executing_eagerly): + info = _run_contained(proc_func, args, kwargs) + self._resources.process_status_queue.put(info) # Re-raise the exception in addition to reporting it to the parent # process, so that even if `--test_timeout` flag is set and the @@ -647,18 +641,183 @@ class _ProcFunc(object): # instead of silently suppressing the error due to early bazel # timeout. Raising an error in the subprocess produces stack trace in # the log, but the program continues running. - raise + if not info.is_successful: + six.reraise(*info.exc_info) - finally: - info = _ProcessStatusInfo( - task_type=test_env.task_type, - is_successful=is_successful, - exc_info=exc_info, - return_value=return_value) - self._resources.process_status_queue.put(info) self._close_streaming() +class MultiProcessPoolRunner(object): + """A utility class to start a process pool to simulate a cluster. + + It's similar to MultiProcessRunner, but uses a pool of processes to avoid the + expensive initialization cost of Tensorflow. + """ + + def __init__(self, cluster_spec, initializer=None): + """Creates a multi-process pool runner. + + Args: + cluster_spec: Dict for cluster spec. The following is an example of + cluster with three workers. + {"worker": ["worker0.example.com:2222", + "worker1.example.com:2222", + "worker2.example.com:2222"]} + initializer: a callable to called at the startup of worker processes. + + Raises: + RuntimeError: if `multi_process_runner.test_main()` is not called. + ValueError: if there are more than one chief in the `cluster_spec`. + """ + self._cluster_spec = cluster_spec + self._initializer = initializer + self._conn = {} + self._runner = None + + def __del__(self): + self._reset() + + def _reset(self): + for conn in self._conn.values(): + conn.close() + self._conn = {} + if self._runner is not None: + self._runner.join() + self._runner = None + + def _start(self): + """Starts the worker pool.""" + # We need different arguments for different processes so we're passing a + # no-op proc_func here and use start_single_process instead. + # + # We also need to start the process pool as daemon, so that they don't block + # the program from exiting. Note that __del__ may not get called when + # there's an exception. The user may also store a pool runner in a global + # object to share across test cases + self._runner = MultiProcessRunner( + proc_func=lambda: None, + cluster_spec=self._cluster_spec, + use_dill_for_args=False, + daemon=True) + if self._initializer: + initializer = dill.dumps(self._initializer, dill.HIGHEST_PROTOCOL) + else: + initializer = None + for task_type, addresses in self._cluster_spec.items(): + for task_id, _ in enumerate(addresses): + conn1, conn2 = multiprocessing.Pipe(duplex=True) + self._conn[(task_type, task_id)] = conn1 + self._runner.start_single_process( + task_type, + task_id, + proc_func=_pool_runner_worker, + args=(initializer, conn2)) + + def run(self, proc_func, args=None, kwargs=None): + """Runs `proc_func` with `args` and `kwargs` on all jobs. + + Args: + proc_func: The function to be run. + args: Optional positional arguments to be supplied in `proc_func`. + kwargs: Optional keyword arguments to be supplied in `proc_func`. + + Returns: + A list of return values. + """ + if self._runner is None: + self._start() + + # Since we start the processes as daemon they're going to be killed by + # SIGTERM when the program exits. We only turn on streaming during run() to + # avoid printing the stacktrace caused by the SIGTERM. + self._runner._stream_stdout = True # pylint: disable=protected-access + + try: + proc_func = dill.dumps(proc_func, dill.HIGHEST_PROTOCOL) + for conn in self._conn.values(): + conn.send((proc_func, args or [], kwargs or {})) + + process_statuses = [] + for (task_type, task_id), conn in self._conn.items(): + logging.info('Waiting for the result from %s-%d', task_type, task_id) + try: + process_statuses.append(conn.recv()) + except EOFError: + # This shouldn't happen due to exceptions in proc_func. This usually + # means bugs in the runner. + self._reset() + raise RuntimeError('Unexpected EOF. Worker process may have died. ' + 'Please report a bug') + + return_values = [] + for process_status in process_statuses: + assert isinstance(process_status, _ProcessStatusInfo) + if not process_status.is_successful: + six.reraise(*process_status.exc_info) + if process_status.return_value is not None: + return_values.append(process_status.return_value) + + return return_values + finally: + self._runner._stream_stdout = False # pylint: disable=protected-access + + +def _pool_runner_worker(initializer, conn): + """Function that runs on the workers in a pool. + + It listens for callables to run and returns the result until `conn` is closed. + It captures the exceptions during executing the callable and return it through + `conn`. + + Args: + initializer: A callable to execute during startup. + conn: A multiprocessing.Connection object to listen for tasks and send + results. + """ + if initializer: + initializer = dill.loads(initializer) + initializer() + while True: + try: + proc_func, args, kwargs = conn.recv() + except EOFError: + break + proc_func = dill.loads(proc_func) + info = _run_contained(proc_func, args, kwargs) + sys.stdout.flush() + sys.stderr.flush() + conn.send(info) + + +def _run_contained(proc_func, args, kwargs): + """Runs `proc_func` with `args` and `kwargs`. + + The function returns _ProcessStatusInfo which captures the return value and + the exception. + + Args: + proc_func: The function to be run. + args: Optional positional arguments to be supplied in `proc_func`. + kwargs: Optional keyword arguments to be supplied in `proc_func`. + + Returns: + a _ProcessStatusInfo. + """ + try: + return_value = proc_func(*args, **kwargs) + is_successful = True + exc_info = None + except Exception: # pylint: disable=broad-except + return_value = None + is_successful = False + exc_info = sys.exc_info() + finally: + return _ProcessStatusInfo( # pylint: disable=lost-exception + is_successful=is_successful, + exc_info=exc_info, + return_value=return_value) + + class SubprocessTimeoutError(RuntimeError): """An error that indicates there is at least one subprocess timing out. diff --git a/tensorflow/python/distribute/multi_process_runner_test.py b/tensorflow/python/distribute/multi_process_runner_test.py index aeba43b6b7c..d76ef5a5a3c 100644 --- a/tensorflow/python/distribute/multi_process_runner_test.py +++ b/tensorflow/python/distribute/multi_process_runner_test.py @@ -22,6 +22,8 @@ import json import os import threading import time +import unittest + from absl import logging from tensorflow.python.distribute import multi_process_runner @@ -45,7 +47,7 @@ def proc_func_that_adds_simple_return_data(): return 'dummy_data' -def proc_func_that_return_args_and_kwargs(*args, **kwargs): +def proc_func_that_returns_args_and_kwargs(*args, **kwargs): return list(args) + list(kwargs.items()) @@ -53,6 +55,20 @@ def proc_func_with_barrier(): return multi_process_runner.barrier() +def proc_func_that_returns_pid(): + return os.getpid() + + +V = None + + +def proc_func_that_sets_global(val): + global V + old_val = V + V = val + return old_val + + class MultiProcessRunnerTest(test.TestCase): def _worker_idx(self): @@ -95,7 +111,7 @@ class MultiProcessRunnerTest(test.TestCase): def test_multi_process_runner_args_passed_correctly(self): return_value = multi_process_runner.run( - proc_func_that_return_args_and_kwargs, + proc_func_that_returns_args_and_kwargs, multi_worker_test_base.create_cluster_spec(num_workers=1), args=('a', 'b'), kwargs={ @@ -325,5 +341,54 @@ class MultiProcessRunnerTest(test.TestCase): for line in list_to_assert)) +class MultiProcessPoolRunnerTest(test.TestCase): + + def test_same_process_across_runs(self): + cluster_spec = multi_worker_test_base.create_cluster_spec(num_workers=2) + runner = multi_process_runner.MultiProcessPoolRunner(cluster_spec) + pid = runner.run(proc_func_that_returns_pid) + for _ in range(3): + self.assertAllEqual(runner.run(proc_func_that_returns_pid), pid) + + def test_exceptions_in_sub_process(self): + cluster_spec = multi_worker_test_base.create_cluster_spec(num_workers=2) + runner = multi_process_runner.MultiProcessPoolRunner(cluster_spec) + pid = runner.run(proc_func_that_returns_pid) + with self.assertRaisesRegexp(ValueError, 'This is an error.'): + runner.run(proc_func_that_errors) + self.assertAllEqual(runner.run(proc_func_that_returns_pid), pid) + + def test_tf_config(self): + cluster_spec = multi_worker_test_base.create_cluster_spec( + has_chief=True, num_workers=2) + runner = multi_process_runner.MultiProcessPoolRunner(cluster_spec) + result = runner.run(proc_func_that_adds_task_type_in_return_data) + + job_count_dict = {'worker': 2, 'chief': 1} + for data in result: + job_count_dict[data] -= 1 + + self.assertEqual(job_count_dict['worker'], 0) + self.assertEqual(job_count_dict['chief'], 0) + + @unittest.expectedFailure + def test_exception_in_main_process(self): + # When there's an exception in the main process, __del__() is not called. + # This test is to verify MultiProcessPoolRunner can cope with __del__() not + # being called. + cluster_spec = multi_worker_test_base.create_cluster_spec( + has_chief=True, num_workers=2) + runner = multi_process_runner.MultiProcessPoolRunner(cluster_spec) + runner.run(proc_func_that_returns_pid) + raise ValueError('failure') + + def test_initializer(self): + cluster_spec = multi_worker_test_base.create_cluster_spec(num_workers=2) + runner = multi_process_runner.MultiProcessPoolRunner( + cluster_spec, initializer=lambda: proc_func_that_sets_global(1)) + result = runner.run(proc_func_that_sets_global, args=(2,)) + self.assertAllEqual(result, [1, 1]) + + if __name__ == '__main__': multi_process_runner.test_main() From 965b93f2ee2ecf6dc7152adc31a76802b0ede85f Mon Sep 17 00:00:00 2001 From: Yixing Fu Date: Thu, 18 Jun 2020 20:05:14 -0400 Subject: [PATCH 248/867] test pathlib path for save weights --- tensorflow/python/keras/saving/save_test.py | 14 ++++++++++++++ 1 file changed, 14 insertions(+) diff --git a/tensorflow/python/keras/saving/save_test.py b/tensorflow/python/keras/saving/save_test.py index 66364666841..5b5da8c5047 100644 --- a/tensorflow/python/keras/saving/save_test.py +++ b/tensorflow/python/keras/saving/save_test.py @@ -119,6 +119,20 @@ class TestSaveModel(test.TestCase, parameterized.TestCase): save.save_model(self.model, path, save_format='tf') save.load_model(path) + @test_util.run_v2_only + def test_save_load_weights_tf_pathlib(self): + if sys.version_info >= (3, 6): + path = pathlib.Path(self.get_temp_dir()) / 'model' + self.model.save_weights(path, save_format='tf') + self.model.load_weights(path) + + @test_util.run_v2_only + def test_save_load_weights_hdf5_pathlib(self): + if sys.version_info >= (3, 6): + path = pathlib.Path(self.get_temp_dir()) / 'model' + self.model.save_weights(path, save_format='h5') + self.model.load_weights(path) + @combinations.generate(combinations.combine(mode=['graph', 'eager'])) def test_saving_with_dense_features(self): cols = [ From fd3c95a3b4eb1418d6f87bed916ec17fff44adf0 Mon Sep 17 00:00:00 2001 From: jonah-kohn <51345541+jonah-kohn@users.noreply.github.com> Date: Thu, 18 Jun 2020 17:07:30 -0700 Subject: [PATCH 249/867] Include show_dtype param in plot_model tests. --- tensorflow/python/keras/utils/vis_utils_test.py | 12 ++++++++---- 1 file changed, 8 insertions(+), 4 deletions(-) diff --git a/tensorflow/python/keras/utils/vis_utils_test.py b/tensorflow/python/keras/utils/vis_utils_test.py index 984014216be..0eb38a1d48f 100644 --- a/tensorflow/python/keras/utils/vis_utils_test.py +++ b/tensorflow/python/keras/utils/vis_utils_test.py @@ -36,7 +36,8 @@ class ModelToDotFormatTest(test.TestCase): model.add(keras.layers.Dense(5, name='dense')) dot_img_file = 'model_1.png' try: - vis_utils.plot_model(model, to_file=dot_img_file, show_shapes=True) + vis_utils.plot_model(model, to_file=dot_img_file, + show_shapes=True, show_dtype=True) self.assertTrue(file_io.file_exists(dot_img_file)) file_io.delete_file(dot_img_file) except ImportError: @@ -62,7 +63,8 @@ class ModelToDotFormatTest(test.TestCase): dot_img_file = 'model_2.png' try: vis_utils.plot_model( - model, to_file=dot_img_file, show_shapes=True, expand_nested=True) + model, to_file=dot_img_file, show_shapes=True, + show_dtype=True expand_nested=True) self.assertTrue(file_io.file_exists(dot_img_file)) file_io.delete_file(dot_img_file) except ImportError: @@ -76,7 +78,8 @@ class ModelToDotFormatTest(test.TestCase): dot_img_file = 'model_3.png' try: vis_utils.plot_model( - model, to_file=dot_img_file, show_shapes=True, expand_nested=True) + model, to_file=dot_img_file, show_shapes=True, + show_dtype=True, expand_nested=True) self.assertTrue(file_io.file_exists(dot_img_file)) file_io.delete_file(dot_img_file) except ImportError: @@ -88,7 +91,8 @@ class ModelToDotFormatTest(test.TestCase): dot_img_file = 'model_4.png' try: vis_utils.plot_model( - model, to_file=dot_img_file, show_shapes=True, expand_nested=True) + model, to_file=dot_img_file, show_shapes=True, + show_dtype=True expand_nested=True) self.assertTrue(file_io.file_exists(dot_img_file)) file_io.delete_file(dot_img_file) except ImportError: From ae76bc79213d4559b113899f438cf54283ec11c2 Mon Sep 17 00:00:00 2001 From: Marat Dukhan Date: Thu, 18 Jun 2020 17:19:14 -0700 Subject: [PATCH 250/867] Update XNNPACK dependency and document sparse inference capability PiperOrigin-RevId: 317213816 Change-Id: I35431b40fd63d836d4fe979f65a71a181c0c820d --- tensorflow/lite/delegates/xnnpack/README.md | 37 +++++++++++++++++++++ tensorflow/workspace.bzl | 8 ++--- 2 files changed, 41 insertions(+), 4 deletions(-) diff --git a/tensorflow/lite/delegates/xnnpack/README.md b/tensorflow/lite/delegates/xnnpack/README.md index 97d2d5565db..d94e92c7306 100644 --- a/tensorflow/lite/delegates/xnnpack/README.md +++ b/tensorflow/lite/delegates/xnnpack/README.md @@ -238,6 +238,43 @@ Below is the list of current operators and limitations: * Fused `NONE`, `RELU`, `RELU_N1_TO_1`, and `RELU6` activations are supported, but fused `TANH` and `SIGN_BIT` activations are not. +### Sparse Inference (experimental) + +XNNPACK backend supports sparse inference for CNN models described in the +[Fast Sparse ConvNets](https://arxiv.org/abs/1911.09723) paper. This +functionality must be enabled at build-time via +`--define xnn_enable_sparse=true` Bazel flag. Sparse inference is restricted +to subgraphs with the following operators: + +* Sparse subgraph must start with a 3x3 stride-2 `CONV_2D` operator with + padding 1 on each side, no dilation, and 3 input channels. +* Sparse subgraph must end with a `MEAN` operator that does reduction across + spatial axes. +* Sparse subgraph may contain the following operators: + * `CONV_2D` with 1x1 kernel and no padding. It is important to have high + sparsity (at least 70%) in the filter of this operator to get speedup + over dense inference. + * `DEPTHWISE_CONV_2D` with 3x3 kernel, stride 1, no dilation, and padding 1 + on each side. + * `DEPTHWISE_CONV_2D` with 3x3 kernel, stride 2, no dilation, and padding 1 + on each side. + * `DEPTHWISE_CONV_2D` with 5x5 kernel, stride 1, no dilation, and padding 2 + on each side. + * `DEPTHWISE_CONV_2D` with 5x5 kernel, stride 2, no dilation, and padding 2 + on each side. + * `ADD` and `MUL` operators where both inputs are 4D tensors. If one of the + inputs to `ADD` or `MUL` is a constant tensor, it must be representable as + either a scalar, or a 1D vector. + * Unary elementwise operators `ABS`, `CEIL`, `FLOOR`, `HARD_SWISH`, + `LEAKY_RELU`, `LOGISTIC`, `NEG`, `RELU`, `RELU6`, `RELU_N1_TO_1`, `ROUND`, + and `SQUARE`. + +Pre-trained [Fast Sparse ConvNets models](https://github.com/google-research/google-research/tree/master/fastconvnets) +provide examples that satisfy these constrains. + +In addition to acceleration, sparse models get the compression benefit by +storing only non-zero values in the [TensorFlow Lite file format](https://github.com/tensorflow/tensorflow/blob/4aea552e064cf92330e07e83a3b5a1ca2a7034d0/tensorflow/lite/schema/schema.fbs#L84-L109). + ### Other limitations * Dynamically allocated (with `kTfLiteDynamic` allocation type) inputs and diff --git a/tensorflow/workspace.bzl b/tensorflow/workspace.bzl index f5b0b7537dc..52c573628ac 100755 --- a/tensorflow/workspace.bzl +++ b/tensorflow/workspace.bzl @@ -164,11 +164,11 @@ def tf_repositories(path_prefix = "", tf_repo_name = ""): tf_http_archive( name = "XNNPACK", - sha256 = "4af883fea0a6ada106867f29670a6c0b7af74bee85d74a2e04356a670814a3d4", - strip_prefix = "XNNPACK-69a6a7667d96a84c596b0f4e00632b2037c17723", + sha256 = "2527a30464b43bd03f137b2c455a0381e49eae63d09cfeee128a717dfbe962d5", + strip_prefix = "XNNPACK-8b283aa30a3186c6e640aed520543e9c067132d2", urls = [ - "https://storage.googleapis.com/mirror.tensorflow.org/github.com/google/XNNPACK/archive/69a6a7667d96a84c596b0f4e00632b2037c17723.zip", - "https://github.com/google/XNNPACK/archive/69a6a7667d96a84c596b0f4e00632b2037c17723.zip", + "https://storage.googleapis.com/mirror.tensorflow.org/github.com/google/XNNPACK/archive/8b283aa30a3186c6e640aed520543e9c067132d2.zip", + "https://github.com/google/XNNPACK/archive/8b283aa30a3186c6e640aed520543e9c067132d2.zip", ], ) From 55e33877b8f81b5ede6365456ed3aa89bbe16d8e Mon Sep 17 00:00:00 2001 From: Yixing Fu Date: Thu, 18 Jun 2020 20:37:52 -0400 Subject: [PATCH 251/867] Revert "test pathlib path for save weights" This reverts commit 965b93f2ee2ecf6dc7152adc31a76802b0ede85f. --- tensorflow/python/keras/saving/save_test.py | 14 -------------- 1 file changed, 14 deletions(-) diff --git a/tensorflow/python/keras/saving/save_test.py b/tensorflow/python/keras/saving/save_test.py index 5b5da8c5047..66364666841 100644 --- a/tensorflow/python/keras/saving/save_test.py +++ b/tensorflow/python/keras/saving/save_test.py @@ -119,20 +119,6 @@ class TestSaveModel(test.TestCase, parameterized.TestCase): save.save_model(self.model, path, save_format='tf') save.load_model(path) - @test_util.run_v2_only - def test_save_load_weights_tf_pathlib(self): - if sys.version_info >= (3, 6): - path = pathlib.Path(self.get_temp_dir()) / 'model' - self.model.save_weights(path, save_format='tf') - self.model.load_weights(path) - - @test_util.run_v2_only - def test_save_load_weights_hdf5_pathlib(self): - if sys.version_info >= (3, 6): - path = pathlib.Path(self.get_temp_dir()) / 'model' - self.model.save_weights(path, save_format='h5') - self.model.load_weights(path) - @combinations.generate(combinations.combine(mode=['graph', 'eager'])) def test_saving_with_dense_features(self): cols = [ From edc445eb184633f0ca10270bca6e9c0b834ab454 Mon Sep 17 00:00:00 2001 From: Yixing Fu Date: Thu, 18 Jun 2020 20:38:02 -0400 Subject: [PATCH 252/867] Revert "only test pathlib on python v>3.6" This reverts commit c3ca4da9b46e3148bdb913d08fec3fd2727158e1. --- tensorflow/python/keras/saving/save_test.py | 7 +++---- 1 file changed, 3 insertions(+), 4 deletions(-) diff --git a/tensorflow/python/keras/saving/save_test.py b/tensorflow/python/keras/saving/save_test.py index 66364666841..98d78735ad1 100644 --- a/tensorflow/python/keras/saving/save_test.py +++ b/tensorflow/python/keras/saving/save_test.py @@ -73,10 +73,9 @@ class TestSaveModel(test.TestCase, parameterized.TestCase): @test_util.run_v2_only def test_save_format_defaults_pathlib(self): - if sys.version_info >= (3, 6): - path = pathlib.Path(self.get_temp_dir()) / 'model_path' - save.save_model(self.model, path) - self.assert_saved_model(path) + path = pathlib.Path(self.get_temp_dir()) / 'model_path' + save.save_model(self.model, path) + self.assert_saved_model(path) @test_util.run_v2_only def test_save_hdf5(self): From 10fb2155fb720f9e0e70d9e48a934383b4b42c91 Mon Sep 17 00:00:00 2001 From: Yixing Fu Date: Thu, 18 Jun 2020 20:38:08 -0400 Subject: [PATCH 253/867] Revert "modify docstring" This reverts commit 5f2e0240ee7977042e41d9c29c349a7b14301290. --- tensorflow/python/saved_model/loader_impl.py | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/tensorflow/python/saved_model/loader_impl.py b/tensorflow/python/saved_model/loader_impl.py index 06cd988130d..71f6ed16c9d 100644 --- a/tensorflow/python/saved_model/loader_impl.py +++ b/tensorflow/python/saved_model/loader_impl.py @@ -73,7 +73,7 @@ def parse_saved_model(export_dir): """Reads the savedmodel.pb or savedmodel.pbtxt file containing `SavedModel`. Args: - export_dir: String or Pathlike, path to the directory containing the SavedModel file. + export_dir: Directory containing the SavedModel file. Returns: A `SavedModel` protocol buffer. From 9f0e739f5f1623b74c1835e9209c7c25b4c49380 Mon Sep 17 00:00:00 2001 From: Yixing Fu Date: Thu, 18 Jun 2020 20:38:15 -0400 Subject: [PATCH 254/867] Revert "convert export_dir as pathlike object to str in parse_saved_model" This reverts commit 1c72a6c65e1733b55286c6361142a39d699732dc. --- tensorflow/python/saved_model/loader_impl.py | 4 ++-- 1 file changed, 2 insertions(+), 2 deletions(-) diff --git a/tensorflow/python/saved_model/loader_impl.py b/tensorflow/python/saved_model/loader_impl.py index 71f6ed16c9d..2df2bea428e 100644 --- a/tensorflow/python/saved_model/loader_impl.py +++ b/tensorflow/python/saved_model/loader_impl.py @@ -83,11 +83,11 @@ def parse_saved_model(export_dir): """ # Build the path to the SavedModel in pbtxt format. path_to_pbtxt = os.path.join( - compat.as_bytes(compat.path_to_str(export_dir)), + compat.as_bytes(export_dir), compat.as_bytes(constants.SAVED_MODEL_FILENAME_PBTXT)) # Build the path to the SavedModel in pb format. path_to_pb = os.path.join( - compat.as_bytes(compat.path_to_str(export_dir)), + compat.as_bytes(export_dir), compat.as_bytes(constants.SAVED_MODEL_FILENAME_PB)) # Parse the SavedModel protocol buffer. From 5869f657624ece35544c3fe54f999b2e3f449305 Mon Sep 17 00:00:00 2001 From: Yixing Fu Date: Thu, 18 Jun 2020 20:38:24 -0400 Subject: [PATCH 255/867] Revert "add save load test for pathlib path" This reverts commit 1363f0f6e89cdc73ca46d43475bdb35750ed2e50. --- tensorflow/python/keras/saving/save_test.py | 13 ------------- 1 file changed, 13 deletions(-) diff --git a/tensorflow/python/keras/saving/save_test.py b/tensorflow/python/keras/saving/save_test.py index 98d78735ad1..5c5846fe738 100644 --- a/tensorflow/python/keras/saving/save_test.py +++ b/tensorflow/python/keras/saving/save_test.py @@ -71,12 +71,6 @@ class TestSaveModel(test.TestCase, parameterized.TestCase): save.save_model(self.model, path) self.assert_saved_model(path) - @test_util.run_v2_only - def test_save_format_defaults_pathlib(self): - path = pathlib.Path(self.get_temp_dir()) / 'model_path' - save.save_model(self.model, path) - self.assert_saved_model(path) - @test_util.run_v2_only def test_save_hdf5(self): path = os.path.join(self.get_temp_dir(), 'model') @@ -87,13 +81,6 @@ class TestSaveModel(test.TestCase, parameterized.TestCase): 'requires the model to be a Functional model or a Sequential model.'): save.save_model(self.subclassed_model, path, save_format='h5') - @test_util.run_v2_only - def test_save_load_hdf5_pathlib(self): - if sys.version_info >= (3, 6): - path = pathlib.Path(self.get_temp_dir()) / 'model' - save.save_model(self.model, path, save_format='h5') - save.load_model(path) - @test_util.run_v2_only def test_save_tf(self): path = os.path.join(self.get_temp_dir(), 'model') From 723751b20ef2aa0a4af39cad2581fd483ae78ad7 Mon Sep 17 00:00:00 2001 From: "A. Unique TensorFlower" Date: Thu, 18 Jun 2020 18:15:12 -0700 Subject: [PATCH 256/867] Add set_outfeed_config in XLA HloInstruction. PiperOrigin-RevId: 317222410 Change-Id: I5de8a5067f1002a9d656d4e26d145ffe3fe372ed --- tensorflow/compiler/xla/service/hlo_instruction.cc | 4 ++++ tensorflow/compiler/xla/service/hlo_instruction.h | 3 +++ tensorflow/compiler/xla/service/hlo_instructions.h | 1 + 3 files changed, 8 insertions(+) diff --git a/tensorflow/compiler/xla/service/hlo_instruction.cc b/tensorflow/compiler/xla/service/hlo_instruction.cc index cfa21b95dd2..6de76c1cc63 100644 --- a/tensorflow/compiler/xla/service/hlo_instruction.cc +++ b/tensorflow/compiler/xla/service/hlo_instruction.cc @@ -3908,6 +3908,10 @@ const string& HloInstruction::outfeed_config() const { return Cast(this)->outfeed_config(); } +void HloInstruction::set_outfeed_config(const string& config) { + return Cast(this)->set_outfeed_config(config); +} + const std::vector& HloInstruction::replica_groups() const { return Cast(this)->replica_groups(); } diff --git a/tensorflow/compiler/xla/service/hlo_instruction.h b/tensorflow/compiler/xla/service/hlo_instruction.h index 7a5d506b681..f3bb59ff625 100644 --- a/tensorflow/compiler/xla/service/hlo_instruction.h +++ b/tensorflow/compiler/xla/service/hlo_instruction.h @@ -1755,6 +1755,9 @@ class HloInstruction { // Returns the config for the Outfeed instruction. const string& outfeed_config() const; + // Delegates to HloOutfeedInstruction::set_outfeed_config. + void set_outfeed_config(const string& config); + // Returns the shape for the Outfeed instruction. const Shape& outfeed_shape() const; diff --git a/tensorflow/compiler/xla/service/hlo_instructions.h b/tensorflow/compiler/xla/service/hlo_instructions.h index 6da01dc088e..f5a963ef063 100644 --- a/tensorflow/compiler/xla/service/hlo_instructions.h +++ b/tensorflow/compiler/xla/service/hlo_instructions.h @@ -1141,6 +1141,7 @@ class HloOutfeedInstruction : public HloInstruction { const Shape& outfeed_shape() const { return outfeed_shape_; } // Returns the config for the Outfeed instruction. const string& outfeed_config() const { return outfeed_config_; } + void set_outfeed_config(const string& config) { outfeed_config_ = config; } // Returns a serialized representation of this instruction. HloInstructionProto ToProto() const override; From 0bef067291d26baf7cde98e7d3d6fbcf54161d58 Mon Sep 17 00:00:00 2001 From: peng Date: Fri, 19 Jun 2020 09:24:20 +0800 Subject: [PATCH 257/867] fix test error with no GPU config --- tensorflow/core/grappler/optimizers/remapper_test.cc | 5 ----- 1 file changed, 5 deletions(-) diff --git a/tensorflow/core/grappler/optimizers/remapper_test.cc b/tensorflow/core/grappler/optimizers/remapper_test.cc index a947210d8a6..56f31cb49b3 100644 --- a/tensorflow/core/grappler/optimizers/remapper_test.cc +++ b/tensorflow/core/grappler/optimizers/remapper_test.cc @@ -485,11 +485,6 @@ TEST_F(RemapperTest, FuseConv2DWithBiasAndActivationOnGPU) { item.feed = {{"input", input_t}, {"filter", filter_t}, {"bias", bias_t}}; TF_ASSERT_OK(s.ToGraphDef(&item.graph)); - // Place all nodes on GPU. - for (int i = 0; i < item.graph.node_size(); ++i) { - item.graph.mutable_node(i)->set_device("/device:GPU:0"); - } - Remapper optimizer(RewriterConfig::AGGRESSIVE); // trust placeholders shape //Remapper optimizer(RewriterConfig::ON); GraphDef output; From 4e5c26fb53d6e6467ebbb5a65a73af8872d383d3 Mon Sep 17 00:00:00 2001 From: Jonah Kohn <51345541+jonah-kohn@users.noreply.github.com> Date: Thu, 18 Jun 2020 18:40:17 -0700 Subject: [PATCH 258/867] Revert commit 3833402 for modular PRs --- tensorflow/python/keras/optimizer_v2/optimizer_v2.py | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/tensorflow/python/keras/optimizer_v2/optimizer_v2.py b/tensorflow/python/keras/optimizer_v2/optimizer_v2.py index d8992bbe3e0..c55b332bfc0 100644 --- a/tensorflow/python/keras/optimizer_v2/optimizer_v2.py +++ b/tensorflow/python/keras/optimizer_v2/optimizer_v2.py @@ -910,7 +910,7 @@ class OptimizerV2(trackable.Trackable): return value() if tensor_util.is_tensor(value): return backend.get_value(value) - return float(value) + return value def variables(self): """Returns variables of this Optimizer based on the order created.""" From b7edd44ee0f8c264e457c48138474f6e1bf5b18e Mon Sep 17 00:00:00 2001 From: Dan Moldovan Date: Thu, 18 Jun 2020 19:08:04 -0700 Subject: [PATCH 259/867] Enable type annotations for python/ops. PiperOrigin-RevId: 317229132 Change-Id: I7055e650308c2fc83969385dd25e86fb5b073d75 --- tensorflow/python/ops/logging_ops.py | 18 +++--------------- 1 file changed, 3 insertions(+), 15 deletions(-) diff --git a/tensorflow/python/ops/logging_ops.py b/tensorflow/python/ops/logging_ops.py index 8ca63f55987..02fce277690 100644 --- a/tensorflow/python/ops/logging_ops.py +++ b/tensorflow/python/ops/logging_ops.py @@ -54,11 +54,9 @@ except NameError: # call relies on certain conditionals for its dependencies. Use # control_flow_ops.Assert. -# Assert and Print are special symbols in python, so we must -# have an upper-case version of them. -# -# For users with Python 3 or Python 2.7 -# with `from __future__ import print_function`, we could also allow lowercase. +# Assert and Print are special symbols in Python 2, so we must +# have an upper-case version of them. When support for it is dropped, +# we can allow lowercase. # See https://github.com/tensorflow/tensorflow/issues/18053 @@ -83,11 +81,6 @@ def Print(input_, data, message=None, first_n=None, summarize=None, name=None): with jupyter notebook (printing to the notebook *server's* output, not into the notebook). - Additionally, to use tf.print in python 2.7, users must make sure to import - the following: - - `from __future__ import print_function` - Args: input_: A tensor passed through this op. data: A list of tensors to print out when op is evaluated. @@ -148,11 +141,6 @@ def print_v2(*inputs, **kwargs): Python objects. Printed tensors will recursively show the first and last elements of each dimension to summarize. - @compatibility(python2) - In python 2.7, make sure to import the following: - `from __future__ import print_function` - @end_compatibility - Example: Single-input usage: From 13fe5862de7b95fd91aeec8f2d71e9f2e77b699b Mon Sep 17 00:00:00 2001 From: "A. Unique TensorFlower" Date: Thu, 18 Jun 2020 19:11:15 -0700 Subject: [PATCH 260/867] Integrate LLVM at https://github.com/llvm/llvm-project/commit/c830d517b4e4 PiperOrigin-RevId: 317229564 Change-Id: I10163c3e668996252d294018794081394cc0d25c --- third_party/mlir/test.BUILD | 1 + 1 file changed, 1 insertion(+) diff --git a/third_party/mlir/test.BUILD b/third_party/mlir/test.BUILD index 23287ce28d6..14c2ba7778e 100644 --- a/third_party/mlir/test.BUILD +++ b/third_party/mlir/test.BUILD @@ -166,6 +166,7 @@ cc_library( "@llvm-project//mlir:Pass", "@llvm-project//mlir:SCFDialect", "@llvm-project//mlir:StandardOps", + "@llvm-project//mlir:StandardOpsTransforms", "@llvm-project//mlir:Support", "@llvm-project//mlir:TargetNVVMIR", "@llvm-project//mlir:TargetROCDLIR", From 4a14e778d64853a236941259693aa3c5813c18d8 Mon Sep 17 00:00:00 2001 From: David Majnemer Date: Thu, 18 Jun 2020 19:11:40 -0700 Subject: [PATCH 261/867] [XLA] Introduce ManifestCheckingTest PiperOrigin-RevId: 317229603 Change-Id: Ibcc9ea3895d520024f5d80d52330aeb3b970585d --- tensorflow/compiler/xla/tests/BUILD | 23 +++- tensorflow/compiler/xla/tests/build_defs.bzl | 7 +- .../xla/tests/client_library_test_base.h | 3 +- tensorflow/compiler/xla/tests/hlo_test_base.h | 3 +- .../xla/tests/local_client_test_base.h | 3 +- .../xla/tests/manifest_checking_test.cc | 129 ++++++++++++++++++ .../xla/tests/manifest_checking_test.h | 35 +++++ tensorflow/compiler/xla/tests/test_macros.cc | 89 +----------- tensorflow/compiler/xla/tests/test_macros.h | 118 +--------------- 9 files changed, 201 insertions(+), 209 deletions(-) create mode 100644 tensorflow/compiler/xla/tests/manifest_checking_test.cc create mode 100644 tensorflow/compiler/xla/tests/manifest_checking_test.h diff --git a/tensorflow/compiler/xla/tests/BUILD b/tensorflow/compiler/xla/tests/BUILD index e1863a8a4cf..9b36117602b 100644 --- a/tensorflow/compiler/xla/tests/BUILD +++ b/tensorflow/compiler/xla/tests/BUILD @@ -52,16 +52,26 @@ cc_library( name = "test_macros_header", testonly = True, hdrs = ["test_macros.h"], - deps = [ - "//tensorflow/compiler/xla:types", - "//tensorflow/core:test", - "@com_google_absl//absl/strings", - ], ) # Generate a test_macros_${BACKEND} library per backend with the proper copts. generate_backend_test_macros() +cc_library( + name = "manifest_checking_test", + testonly = True, + srcs = ["manifest_checking_test.cc"], + hdrs = ["manifest_checking_test.h"], + deps = [ + ":test_macros_header", + "//tensorflow/core:regexp_internal", + "//tensorflow/core:test", + "//tensorflow/core/platform:logging", + "@com_google_absl//absl/container:flat_hash_map", + "@com_google_absl//absl/strings", + ], +) + cc_library( name = "test_utils", srcs = ["test_utils.cc"], @@ -136,6 +146,7 @@ cc_library( hdrs = ["hlo_test_base.h"], deps = [ ":literal_test_util", + ":manifest_checking_test", ":test_utils", ":verified_hlo_module", "//tensorflow/compiler/xla:debug_options_flags", @@ -193,6 +204,7 @@ cc_library( srcs = ["client_library_test_base.cc"], hdrs = ["client_library_test_base.h"], deps = [ + ":manifest_checking_test", "//tensorflow/compiler/xla:array2d", "//tensorflow/compiler/xla:array3d", "//tensorflow/compiler/xla:array4d", @@ -273,6 +285,7 @@ cc_library( hdrs = ["local_client_test_base.h"], deps = [ ":client_library_test_base", + ":manifest_checking_test", ":verified_hlo_module", "//tensorflow/compiler/xla:shape_util", "//tensorflow/compiler/xla:status_macros", diff --git a/tensorflow/compiler/xla/tests/build_defs.bzl b/tensorflow/compiler/xla/tests/build_defs.bzl index c0c0751b0de..94d870aa2ef 100644 --- a/tensorflow/compiler/xla/tests/build_defs.bzl +++ b/tensorflow/compiler/xla/tests/build_defs.bzl @@ -266,11 +266,6 @@ def generate_backend_test_macros(backends = []): "-DXLA_DISABLED_MANIFEST=\\\"%s\\\"" % manifest, ], deps = [ - "@com_google_absl//absl/container:flat_hash_map", - "@com_google_absl//absl/strings", - "//tensorflow/compiler/xla:types", - "//tensorflow/core:lib", - "//tensorflow/core:regexp_internal", - "//tensorflow/core:test", + "//tensorflow/core/platform:logging", ], ) diff --git a/tensorflow/compiler/xla/tests/client_library_test_base.h b/tensorflow/compiler/xla/tests/client_library_test_base.h index 790497f888e..17bb70bdb42 100644 --- a/tensorflow/compiler/xla/tests/client_library_test_base.h +++ b/tensorflow/compiler/xla/tests/client_library_test_base.h @@ -35,6 +35,7 @@ limitations under the License. #include "tensorflow/compiler/xla/literal_util.h" #include "tensorflow/compiler/xla/statusor.h" #include "tensorflow/compiler/xla/tests/literal_test_util.h" +#include "tensorflow/compiler/xla/tests/manifest_checking_test.h" #include "tensorflow/compiler/xla/tests/test_utils.h" #include "tensorflow/compiler/xla/xla_data.pb.h" #include "tensorflow/core/lib/core/bitmap.h" @@ -62,7 +63,7 @@ std::vector ExpandUseBfloat16( } // A client library test establishes an in-process XLA client connection. -class ClientLibraryTestBase : public ::testing::Test { +class ClientLibraryTestBase : public ManifestCheckingTest { protected: explicit ClientLibraryTestBase(se::Platform* platform = nullptr); diff --git a/tensorflow/compiler/xla/tests/hlo_test_base.h b/tensorflow/compiler/xla/tests/hlo_test_base.h index 85b1876dd3c..17c2a55ba5b 100644 --- a/tensorflow/compiler/xla/tests/hlo_test_base.h +++ b/tensorflow/compiler/xla/tests/hlo_test_base.h @@ -32,6 +32,7 @@ limitations under the License. #include "tensorflow/compiler/xla/shape_layout.h" #include "tensorflow/compiler/xla/statusor.h" #include "tensorflow/compiler/xla/tests/literal_test_util.h" +#include "tensorflow/compiler/xla/tests/manifest_checking_test.h" #include "tensorflow/compiler/xla/tests/verified_hlo_module.h" #include "tensorflow/compiler/xla/types.h" #include "tensorflow/compiler/xla/xla_data.pb.h" @@ -67,7 +68,7 @@ namespace xla { // ) // // For a more detailed example, see "../tests/sample_text_test.cc". -class HloTestBase : public ::testing::Test { +class HloTestBase : public ManifestCheckingTest { public: // Creates a new HLO module for a test. The module created will have // TestName() for its name; it will also automatically populate its debug diff --git a/tensorflow/compiler/xla/tests/local_client_test_base.h b/tensorflow/compiler/xla/tests/local_client_test_base.h index ea457024618..c1951ad1021 100644 --- a/tensorflow/compiler/xla/tests/local_client_test_base.h +++ b/tensorflow/compiler/xla/tests/local_client_test_base.h @@ -32,6 +32,7 @@ limitations under the License. #include "tensorflow/compiler/xla/service/transfer_manager.h" #include "tensorflow/compiler/xla/statusor.h" #include "tensorflow/compiler/xla/tests/client_library_test_base.h" +#include "tensorflow/compiler/xla/tests/manifest_checking_test.h" #include "tensorflow/compiler/xla/tests/verified_hlo_module.h" #include "tensorflow/compiler/xla/xla_data.pb.h" #include "tensorflow/core/platform/mutex.h" @@ -75,7 +76,7 @@ class TestAllocator : public se::StreamExecutorMemoryAllocator { }; // A base class for tests which exercise the LocalClient interface. -class LocalClientTestBase : public ::testing::Test { +class LocalClientTestBase : public ManifestCheckingTest { protected: struct EigenThreadPoolWrapper; explicit LocalClientTestBase(se::Platform* platform = nullptr); diff --git a/tensorflow/compiler/xla/tests/manifest_checking_test.cc b/tensorflow/compiler/xla/tests/manifest_checking_test.cc new file mode 100644 index 00000000000..8806290472d --- /dev/null +++ b/tensorflow/compiler/xla/tests/manifest_checking_test.cc @@ -0,0 +1,129 @@ +/* Copyright 2017 The TensorFlow Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. +==============================================================================*/ + +#include "tensorflow/compiler/xla/tests/manifest_checking_test.h" + +#include +#include +#include + +#include "absl/container/flat_hash_map.h" +#include "absl/strings/ascii.h" +#include "absl/strings/str_split.h" +#include "tensorflow/compiler/xla/tests/test_macros.h" +#include "tensorflow/core/platform/logging.h" +#include "tensorflow/core/platform/regexp.h" + +namespace xla { + +namespace { + +// Mapping from test name; i.e. MyTest.MyTestCase to platforms on which it is +// disabled - a sequence of regexps. +using ManifestT = absl::flat_hash_map>; + +ManifestT ReadManifest() { + ManifestT manifest; + + absl::string_view path = absl::NullSafeStringView(kDisabledManifestPath); + if (path.empty()) { + return manifest; + } + + // Note: parens are required to disambiguate vs function decl. + std::ifstream file_stream((std::string(path))); + std::string contents((std::istreambuf_iterator(file_stream)), + std::istreambuf_iterator()); + + std::vector lines = absl::StrSplit(contents, '\n'); + for (std::string& line : lines) { + auto comment = line.find("//"); + if (comment != std::string::npos) { + line = line.substr(0, comment); + } + if (line.empty()) { + continue; + } + absl::StripTrailingAsciiWhitespace(&line); + std::vector pieces = absl::StrSplit(line, ' '); + CHECK_GE(pieces.size(), 1); + auto& platforms = manifest[pieces[0]]; + for (size_t i = 1; i < pieces.size(); ++i) { + platforms.push_back(pieces[i]); + } + } + return manifest; +} + +} // namespace + +void ManifestCheckingTest::SetUp() { + const testing::TestInfo* test_info = + testing::UnitTest::GetInstance()->current_test_info(); + absl::string_view test_case_name = test_info->test_suite_name(); + absl::string_view test_name = test_info->name(); + VLOG(1) << "test_case_name: " << test_case_name; + VLOG(1) << "test_name: " << test_name; + + // Remove the type suffix from the test case name. + if (const char* type_param = test_info->type_param()) { + VLOG(1) << "type_param: " << type_param; + size_t last_slash = test_case_name.rfind('/'); + test_case_name = test_case_name.substr(0, last_slash); + VLOG(1) << "test_case_name: " << test_case_name; + } + + // Remove the test instantiation name if it is present. + auto first_slash = test_case_name.find('/'); + if (first_slash != test_case_name.npos) { + test_case_name.remove_prefix(first_slash + 1); + VLOG(1) << "test_case_name: " << test_case_name; + } + + ManifestT manifest = ReadManifest(); + + // If the test name ends with a slash followed by one or more characters, + // strip that off. + auto last_slash = test_name.rfind('/'); + if (last_slash != test_name.npos) { + test_name = test_name.substr(0, last_slash); + VLOG(1) << "test_name: " << test_name; + } + + // First try full match: test_case_name.test_name + // If that fails, try to find just the test_case_name; this would disable all + // tests in the test case. + auto it = manifest.find(absl::StrCat(test_case_name, ".", test_name)); + if (it == manifest.end()) { + it = manifest.find(test_case_name); + if (it == manifest.end()) { + return; + } + } + + // Expect a full match vs. one of the platform regexps to disable the test. + const std::vector& disabled_platforms = it->second; + auto platform_string = kTestPlatform; + for (const auto& s : disabled_platforms) { + if (RE2::FullMatch(/*text=*/platform_string, /*re=*/s)) { + GTEST_SKIP(); + return; + } + } + + // We didn't hit in the disabled manifest entries, so don't disable it. +} + +} // namespace xla diff --git a/tensorflow/compiler/xla/tests/manifest_checking_test.h b/tensorflow/compiler/xla/tests/manifest_checking_test.h new file mode 100644 index 00000000000..4f44ed76a3e --- /dev/null +++ b/tensorflow/compiler/xla/tests/manifest_checking_test.h @@ -0,0 +1,35 @@ +/* Copyright 2017 The TensorFlow Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. +==============================================================================*/ + +#ifndef TENSORFLOW_COMPILER_XLA_TESTS_MANIFEST_CHECKING_TEST_H_ +#define TENSORFLOW_COMPILER_XLA_TESTS_MANIFEST_CHECKING_TEST_H_ + +#include "tensorflow/core/platform/test.h" + +namespace xla { + +// This class allows us to intercept the test name and use an arbitrary +// heuristic to decide whether the test case should be disabled. We +// determine whether the test case should be disabled by resolving the (test +// case name, test name) in a manifest file. +class ManifestCheckingTest : public ::testing::Test { + protected: + // This method runs before each test runs. + void SetUp() override; +}; + +} // namespace xla + +#endif // TENSORFLOW_COMPILER_XLA_TESTS_MANIFEST_CHECKING_TEST_H_ diff --git a/tensorflow/compiler/xla/tests/test_macros.cc b/tensorflow/compiler/xla/tests/test_macros.cc index dc9ac7b684a..9e85af76e89 100644 --- a/tensorflow/compiler/xla/tests/test_macros.cc +++ b/tensorflow/compiler/xla/tests/test_macros.cc @@ -15,93 +15,18 @@ limitations under the License. #include "tensorflow/compiler/xla/tests/test_macros.h" -#include -#include -#include - -#include "absl/container/flat_hash_map.h" -#include "absl/strings/str_cat.h" -#include "absl/strings/str_split.h" #include "tensorflow/core/platform/logging.h" -#include "tensorflow/core/platform/regexp.h" namespace xla { -namespace { -// Mapping from test name; i.e. MyTest.MyTestCase to platforms on which it is -// disabled - a sequence of regexps. -using ManifestT = absl::flat_hash_map>; - -ManifestT ReadManifest() { - ManifestT manifest; - - string path = XLA_DISABLED_MANIFEST; - if (path.empty()) { - return manifest; - } - - std::ifstream file_stream(path); - // Note: parens are required to disambiguate vs function decl. - string contents((std::istreambuf_iterator(file_stream)), - std::istreambuf_iterator()); - - std::vector lines = absl::StrSplit(contents, '\n'); - for (string& line : lines) { - auto comment = line.find("//"); - if (comment != string::npos) { - line = line.substr(0, comment); - } - if (line.empty()) { - continue; - } - absl::StripTrailingAsciiWhitespace(&line); - std::vector pieces = absl::StrSplit(line, ' '); - CHECK_GE(pieces.size(), 1); - auto& platforms = manifest[pieces[0]]; - for (int64 i = 1; i < pieces.size(); ++i) { - platforms.push_back(pieces[i]); - } - } - return manifest; +static bool InitModule() { + kDisabledManifestPath = XLA_DISABLED_MANIFEST; + VLOG(1) << "kDisabledManifestPath: " << kDisabledManifestPath; + kTestPlatform = XLA_PLATFORM; + VLOG(1) << "kTestPlatform: " << kTestPlatform; + return false; } -} // namespace - -std::string PrependDisabledIfIndicated(absl::string_view test_case_name, - absl::string_view test_name) { - ManifestT manifest = ReadManifest(); - - // If the test name ends with a slash followed by one or more digits, strip - // that off; this is just a shard number, and matching on this would be - // unstable even if someone wanted to do it. - static LazyRE2 shard_num_pattern = {R"(/\d+$)"}; - absl::string_view suffix; - if (RE2::PartialMatch(test_name, *shard_num_pattern, &suffix)) { - test_name.remove_suffix(suffix.size()); - } - - // First try full match: test_case_name.test_name - // If that fails, try to find just the test_case_name; this would disable all - // tests in the test case. - auto it = manifest.find(absl::StrCat(test_case_name, ".", test_name)); - if (it == manifest.end()) { - it = manifest.find(test_case_name); - if (it == manifest.end()) { - return std::string(test_name); - } - } - - // Expect a full match vs. one of the platform regexps to disable the test. - const std::vector& disabled_platforms = it->second; - string platform_string = XLA_PLATFORM; - for (const auto& s : disabled_platforms) { - if (RE2::FullMatch(/*text=*/platform_string, /*re=*/s)) { - return absl::StrCat("DISABLED_", test_name); - } - } - - // We didn't hit in the disabled manifest entries, so don't disable it. - return std::string(test_name); -} +static bool module_initialized = InitModule(); } // namespace xla diff --git a/tensorflow/compiler/xla/tests/test_macros.h b/tensorflow/compiler/xla/tests/test_macros.h index 33d2dff9721..f62bccbe850 100644 --- a/tensorflow/compiler/xla/tests/test_macros.h +++ b/tensorflow/compiler/xla/tests/test_macros.h @@ -28,12 +28,6 @@ limitations under the License. #ifndef TENSORFLOW_COMPILER_XLA_TESTS_TEST_MACROS_H_ #define TENSORFLOW_COMPILER_XLA_TESTS_TEST_MACROS_H_ -#include - -#include "absl/strings/string_view.h" -#include "tensorflow/compiler/xla/types.h" -#include "tensorflow/core/platform/test.h" - #define DISABLED_ON_CPU(X) X #define DISABLED_ON_GPU(X) X #define DISABLED_ON_GPU_ROCM(X) X @@ -79,117 +73,15 @@ limitations under the License. namespace xla { -// Reads a disabled manifest file to resolve whether test cases should be -// disabled on a particular platform. For a test that should be disabled, -// returns DISABLED_ prepended to its name; otherwise returns the test name -// unmodified. -std::string PrependDisabledIfIndicated(absl::string_view test_case_name, - absl::string_view test_name); +inline const char *kDisabledManifestPath = nullptr; +inline const char *kTestPlatform = nullptr; } // namespace xla -// This is the internal "gtest" class instantiation -- it is identical to the -// GTEST_TEST_ macro, except that we intercept the test name for potential -// modification by PrependDisabledIfIndicated. That file can use an arbitrary -// heuristic to decide whether the test case should be disabled, and we -// determine whether the test case should be disabled by resolving the (test -// case name, test name) in a manifest file. -#define XLA_GTEST_TEST_(test_case_name, test_name, parent_class) \ - class GTEST_TEST_CLASS_NAME_(test_case_name, test_name) \ - : public parent_class { \ - public: \ - GTEST_TEST_CLASS_NAME_(test_case_name, test_name)() {} \ - \ - private: \ - virtual void TestBody(); \ - static ::testing::TestInfo* const test_info_ GTEST_ATTRIBUTE_UNUSED_; \ - GTEST_DISALLOW_COPY_AND_ASSIGN_(GTEST_TEST_CLASS_NAME_(test_case_name, \ - test_name)); \ - }; \ - \ - ::testing::TestInfo* const GTEST_TEST_CLASS_NAME_(test_case_name, \ - test_name)::test_info_ = \ - ::testing::RegisterTest( \ - #test_case_name, \ - ::xla::PrependDisabledIfIndicated(#test_case_name, #test_name) \ - .c_str(), \ - nullptr, nullptr, __FILE__, __LINE__, []() -> parent_class* { \ - return new GTEST_TEST_CLASS_NAME_(test_case_name, test_name)(); \ - }); \ - void GTEST_TEST_CLASS_NAME_(test_case_name, test_name)::TestBody() +#define XLA_TEST_F(test_fixture, test_name) TEST_F(test_fixture, test_name) -// This is identical to the TEST_F macro from "gtest", but it potentially -// disables the test based on an external manifest file, DISABLED_MANIFEST. -// -// Per usual, you can see what tests are available via --gunit_list_tests and -// choose to run tests that have been disabled via the manifest via -// --gunit_also_run_disabled_tests. -#define XLA_TEST_F(test_fixture, test_name) \ - XLA_GTEST_TEST_(test_fixture, test_name, test_fixture) +#define XLA_TEST_P(test_case_name, test_name) TEST_P(test_case_name, test_name) -// Likewise, this is identical to the TEST_P macro from "gtest", but -// potentially disables the test based on the DISABLED_MANIFEST file. -// -// We have to wrap this in an outer layer so that any DISABLED_ON_* macros will -// be properly expanded before the stringification occurs. -#define XLA_TEST_P_IMPL_(test_case_name, test_name) \ - class GTEST_TEST_CLASS_NAME_(test_case_name, test_name) \ - : public test_case_name { \ - public: \ - GTEST_TEST_CLASS_NAME_(test_case_name, test_name)() {} \ - virtual void TestBody(); \ - \ - private: \ - static int AddToRegistry() { \ - ::testing::UnitTest::GetInstance() \ - ->parameterized_test_registry() \ - .GetTestCasePatternHolder( \ - #test_case_name, \ - ::testing::internal::CodeLocation(__FILE__, __LINE__)) \ - ->AddTestPattern( \ - #test_case_name, \ - ::xla::PrependDisabledIfIndicated(#test_case_name, #test_name) \ - .c_str(), \ - new ::testing::internal::TestMetaFactory()); \ - return 0; \ - } \ - static int gtest_registering_dummy_ GTEST_ATTRIBUTE_UNUSED_; \ - GTEST_DISALLOW_COPY_AND_ASSIGN_(GTEST_TEST_CLASS_NAME_(test_case_name, \ - test_name)); \ - }; \ - int GTEST_TEST_CLASS_NAME_(test_case_name, \ - test_name)::gtest_registering_dummy_ = \ - GTEST_TEST_CLASS_NAME_(test_case_name, test_name)::AddToRegistry(); \ - void GTEST_TEST_CLASS_NAME_(test_case_name, test_name)::TestBody() - -#define XLA_TEST_P(test_case_name, test_name) \ - XLA_TEST_P_IMPL_(test_case_name, test_name) - -// This is identical to the TEST_F macro from "gtest", but it potentially -// disables the test based on an external manifest file, DISABLED_MANIFEST. -#define XLA_TYPED_TEST(CaseName, TestName) \ - template \ - class GTEST_TEST_CLASS_NAME_(CaseName, TestName) \ - : public CaseName { \ - private: \ - typedef CaseName TestFixture; \ - typedef gtest_TypeParam_ TypeParam; \ - virtual void TestBody(); \ - }; \ - bool gtest_##CaseName##_##TestName##_registered_ GTEST_ATTRIBUTE_UNUSED_ = \ - ::testing::internal::TypeParameterizedTest< \ - CaseName, \ - ::testing::internal::TemplateSel, \ - GTEST_TYPE_PARAMS_(CaseName)>:: \ - Register( \ - "", ::testing::internal::CodeLocation(__FILE__, __LINE__), \ - #CaseName, \ - ::xla::PrependDisabledIfIndicated(#CaseName, #TestName).c_str(), \ - 0); \ - template \ - void GTEST_TEST_CLASS_NAME_(CaseName, \ - TestName)::TestBody() +#define XLA_TYPED_TEST(CaseName, TestName) TYPED_TEST(CaseName, TestName) #endif // TENSORFLOW_COMPILER_XLA_TESTS_TEST_MACROS_H_ From 9c4b749b09b958c436e0681a4276b47fc9316a8a Mon Sep 17 00:00:00 2001 From: "A. Unique TensorFlower" Date: Thu, 18 Jun 2020 19:17:40 -0700 Subject: [PATCH 262/867] Internal change PiperOrigin-RevId: 317230321 Change-Id: I043dae37768f6e9cf946d4db2a8c36123ed2d6d9 --- tensorflow/core/platform/BUILD | 7 - tensorflow/core/platform/tf32_utils.cc | 30 -- tensorflow/core/platform/tf32_utils.h | 27 -- tensorflow/python/BUILD | 11 - tensorflow/python/framework/config.py | 31 --- tensorflow/python/util/tf32.cc | 22 -- tensorflow/stream_executor/cuda/BUILD | 2 - tensorflow/stream_executor/cuda/cuda_blas.cc | 98 ++++--- tensorflow/stream_executor/cuda/cuda_blas.h | 8 +- tensorflow/stream_executor/cuda/cuda_dnn.cc | 272 ++++++++----------- 10 files changed, 172 insertions(+), 336 deletions(-) delete mode 100644 tensorflow/core/platform/tf32_utils.cc delete mode 100644 tensorflow/core/platform/tf32_utils.h delete mode 100644 tensorflow/python/util/tf32.cc diff --git a/tensorflow/core/platform/BUILD b/tensorflow/core/platform/BUILD index 33a1e7cfe0a..70bb8a89417 100644 --- a/tensorflow/core/platform/BUILD +++ b/tensorflow/core/platform/BUILD @@ -938,13 +938,6 @@ cc_library( alwayslink = 1, ) -cc_library( - name = "tf32_utils", - srcs = ["tf32_utils.cc"], - hdrs = ["tf32_utils.h"], - copts = tf_copts(), -) - tf_cc_tests( name = "low_level_library_tests", size = "small", diff --git a/tensorflow/core/platform/tf32_utils.cc b/tensorflow/core/platform/tf32_utils.cc deleted file mode 100644 index d2f40ea161a..00000000000 --- a/tensorflow/core/platform/tf32_utils.cc +++ /dev/null @@ -1,30 +0,0 @@ -/* Copyright 2020 The TensorFlow Authors. All Rights Reserved. - -Licensed under the Apache License, Version 2.0 (the "License"); -you may not use this file except in compliance with the License. -You may obtain a copy of the License at - - http://www.apache.org/licenses/LICENSE-2.0 - -Unless required by applicable law or agreed to in writing, software -distributed under the License is distributed on an "AS IS" BASIS, -WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -See the License for the specific language governing permissions and -limitations under the License. -==============================================================================*/ - -#include "tensorflow/core/platform/tf32_utils.h" - -#include - -namespace tensorflow { - -// Whether TensorFloat-32 should be used where supported. -// TODO(nluehr): Maybe enable by default after TF32 Ampere testing. -static std::atomic tf32_allowed{false}; - -void allow_tf32_execution(bool allowed) { tf32_allowed = allowed; } - -bool tf32_execution_allowed() { return tf32_allowed; } - -} // namespace tensorflow diff --git a/tensorflow/core/platform/tf32_utils.h b/tensorflow/core/platform/tf32_utils.h deleted file mode 100644 index 7a158d00ad3..00000000000 --- a/tensorflow/core/platform/tf32_utils.h +++ /dev/null @@ -1,27 +0,0 @@ -/* Copyright 2020 The TensorFlow Authors. All Rights Reserved. - -Licensed under the Apache License, Version 2.0 (the "License"); -you may not use this file except in compliance with the License. -You may obtain a copy of the License at - - http://www.apache.org/licenses/LICENSE-2.0 - -Unless required by applicable law or agreed to in writing, software -distributed under the License is distributed on an "AS IS" BASIS, -WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -See the License for the specific language governing permissions and -limitations under the License. -==============================================================================*/ - -#ifndef TENSORFLOW_CORE_PLATFORM_TF32_UTILS_H_ -#define TENSORFLOW_CORE_PLATFORM_TF32_UTILS_H_ - -namespace tensorflow { - -void allow_tf32_execution(bool allowed); - -bool tf32_execution_allowed(); - -} // namespace tensorflow - -#endif // TENSORFLOW_CORE_PLATFORM_TF32_UTILS_H_ diff --git a/tensorflow/python/BUILD b/tensorflow/python/BUILD index 5f9e2dfb1ff..de9cf9a24c7 100644 --- a/tensorflow/python/BUILD +++ b/tensorflow/python/BUILD @@ -788,16 +788,6 @@ tf_python_pybind_extension( ], ) -tf_python_pybind_extension( - name = "_pywrap_tf32_execution", - srcs = ["util/tf32.cc"], - module_name = "_pywrap_tf32_execution", - deps = [ - "//tensorflow/core/platform:tf32_utils", - "@pybind11", - ], -) - tf_python_pybind_extension( name = "_pywrap_util_port", srcs = ["util/port_wrapper.cc"], @@ -5688,7 +5678,6 @@ py_library( "//tensorflow:composite_tensor_whitelist", ], deps = [ - ":_pywrap_tf32_execution", ":tf_decorator", ":tf_export", ":tf_stack", diff --git a/tensorflow/python/framework/config.py b/tensorflow/python/framework/config.py index 544b6882618..9ff16f2a327 100644 --- a/tensorflow/python/framework/config.py +++ b/tensorflow/python/framework/config.py @@ -18,42 +18,11 @@ from __future__ import absolute_import from __future__ import division from __future__ import print_function -from tensorflow.python import _pywrap_tf32_execution from tensorflow.python.eager import context from tensorflow.python.util import deprecation from tensorflow.python.util.tf_export import tf_export -# No tf_export until TF is built against CUDA11 which is required for TF32. -def tensor_float_32_execution_allowed(): - """Get if TensorFloat-32 operations are enabled on supported hardware. - - Returns: - True if TensorFloat-32 execution is enabled and False otherwise. - """ - return _pywrap_tf32_execution.is_allowed() - - -# No tf_export until TF is built against CUDA 11 which is required for TF32. -def allow_tensor_float_32_execution(allowed): - """Allow use of TensorFloat-32 with float32 ops on supported hardware. - - TensorFloat-32 is a math mode introduced with the NVIDIA Ampere architecture. - TensorFloat-32 kernels take float32 inputs and produce float32 outputs. - Internally, the inputs are cast to a custom representation with 10-bit - mantissa (similar to float16) and 8-bit exponent (similar to float32) and are - executed using TensorCores with float32 accumulation. For more information, - see https://blogs.nvidia.com/blog/2020/05/14/tensorfloat-32-precision-format/. - - TensorFloat-32 execution is disabled by default, but this may change in a - future version. - - Args: - allowed: whether to allow TensorFloat-32 execution - """ - _pywrap_tf32_execution.allow(allowed) - - @tf_export('config.threading.get_intra_op_parallelism_threads') def get_intra_op_parallelism_threads(): """Get number of threads used within an individual op for parallelism. diff --git a/tensorflow/python/util/tf32.cc b/tensorflow/python/util/tf32.cc deleted file mode 100644 index 7dece6ccdae..00000000000 --- a/tensorflow/python/util/tf32.cc +++ /dev/null @@ -1,22 +0,0 @@ -/* Copyright 2020 The TensorFlow Authors. All Rights Reserved. - -Licensed under the Apache License, Version 2.0 (the "License"); -you may not use this file except in compliance with the License. -You may obtain a copy of the License at - - http://www.apache.org/licenses/LICENSE-2.0 - -Unless required by applicable law or agreed to in writing, software -distributed under the License is distributed on an "AS IS" BASIS, -WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -See the License for the specific language governing permissions and -limitations under the License. -==============================================================================*/ - -#include "pybind11/pybind11.h" -#include "tensorflow/core/platform/tf32_utils.h" - -PYBIND11_MODULE(_pywrap_tf32_execution, m) { - m.def("allow", &tensorflow::allow_tf32_execution); - m.def("is_allowed", &tensorflow::tf32_execution_allowed); -} diff --git a/tensorflow/stream_executor/cuda/BUILD b/tensorflow/stream_executor/cuda/BUILD index 3a14be9ad50..c3cf9f5db15 100644 --- a/tensorflow/stream_executor/cuda/BUILD +++ b/tensorflow/stream_executor/cuda/BUILD @@ -251,7 +251,6 @@ cc_library( "@local_config_cuda//cuda:cuda_headers", "//tensorflow/core:lib", "//tensorflow/core:lib_internal", - "//tensorflow/core/platform:tf32_utils", "//tensorflow/stream_executor", "//tensorflow/stream_executor:event", "//tensorflow/stream_executor:host_or_device_scalar", @@ -357,7 +356,6 @@ cc_library( "@local_config_cuda//cuda:cudnn_header", "//tensorflow/core:lib", "//tensorflow/core:lib_internal", - "//tensorflow/core/platform:tf32_utils", "//tensorflow/stream_executor:dnn", "//tensorflow/stream_executor:event", "//tensorflow/stream_executor:plugin_registry", diff --git a/tensorflow/stream_executor/cuda/cuda_blas.cc b/tensorflow/stream_executor/cuda/cuda_blas.cc index fcd0e7b16fb..c9f0fc462c9 100644 --- a/tensorflow/stream_executor/cuda/cuda_blas.cc +++ b/tensorflow/stream_executor/cuda/cuda_blas.cc @@ -49,7 +49,6 @@ limitations under the License. #include "absl/strings/str_cat.h" #include "absl/strings/str_format.h" #include "third_party/eigen3/Eigen/Core" -#include "tensorflow/core/platform/tf32_utils.h" #include "tensorflow/core/util/env_var.h" #include "tensorflow/stream_executor/cuda/cuda_activation.h" #include "tensorflow/stream_executor/cuda/cuda_gpu_executor.h" @@ -102,6 +101,18 @@ static std::string ToString(cublasStatus_t status) { } } +// Decide whether to enable TENSOR_OP_MATH +static bool TensorOpMathEnabled() { + static bool is_enabled = [] { + bool is_disabled; + TF_CHECK_OK( + tensorflow::ReadBoolFromEnvVar("TF_DISABLE_CUBLAS_TENSOR_OP_MATH", + /*default_val=*/false, &is_disabled)); + return !is_disabled; + }(); + return is_enabled; +} + // cuBLAS has interfaces that permit pointers to be passed from either the host // memory space or the device memory space; however, you must instruct it as to // which address space those pointers are in with cublasSetPointerMode. @@ -226,19 +237,6 @@ bool CUDABlas::Init() { return false; } - absl::MutexLock lock(&mu_); -#if CUDA_VERSION >= 9000 -#if CUBLAS_VER_MAJOR >= 11 - ret = cublasSetMathMode(blas_, CUBLAS_TF32_TENSOR_OP_MATH); -#else - ret = cublasSetMathMode(blas_, CUBLAS_TENSOR_OP_MATH); -#endif - if (ret != CUBLAS_STATUS_SUCCESS) { - LOG(ERROR) << "failed to set cublas default math mode: " << ToString(ret); - return false; - } -#endif - return true; } @@ -401,7 +399,7 @@ cudaDataType_t CUDAComputationType(blas::ComputationType ty) { template bool CUDABlas::DoBlasInternalImpl(FuncT cublas_func, Stream *stream, bool pointer_mode_host, bool err_on_failure, - Args... args) { + bool use_tensor_op_math, Args... args) { absl::MutexLock lock(&mu_); CHECK(blas_ != nullptr); @@ -415,10 +413,10 @@ bool CUDABlas::DoBlasInternalImpl(FuncT cublas_func, Stream *stream, : CUBLAS_POINTER_MODE_DEVICE)) { return false; } -#if CUBLAS_VER_MAJOR >= 11 +#if CUDA_VERSION >= 9000 ScopedCublasMathMode math_mode{blas_}; - if (!tensorflow::tf32_execution_allowed()) { - if (!math_mode.Init(CUBLAS_DEFAULT_MATH)) { + if (use_tensor_op_math) { + if (!math_mode.Init(CUBLAS_TENSOR_OP_MATH)) { return false; } } @@ -1635,9 +1633,21 @@ bool CUDABlas::DoBlasGemm( } } + bool use_tensor_ops = false; +#if CUDA_VERSION >= 9000 + int cc_major, cc_minor; + stream->parent()->GetDeviceDescription().cuda_compute_capability(&cc_major, + &cc_minor); + + // GPUs < sm_70 don't support tensor ops. + if (cc_major >= 7 && TensorOpMathEnabled()) { + use_tensor_ops = true; + } +#endif + return DoBlasInternalImpl( cublasSgemmEx, stream, true /* = pointer_mode_host */, - true /* = err_on_failure= */, CUDABlasTranspose(transa), + true /* = err_on_failure= */, use_tensor_ops, CUDABlasTranspose(transa), CUDABlasTranspose(transb), m, n, k, &alpha, GpuMemory(a), SE_CUDA_DATA_HALF, lda, GpuMemory(b), SE_CUDA_DATA_HALF, ldb, &beta, GpuMemoryMutable(c), SE_CUDA_DATA_HALF, ldc); @@ -1911,7 +1921,8 @@ static bool TensorOpsAvailable(int cc_major) { // strictly correct. We can't simply enable it, though, as that would change // clients' behavior significantly: Using tensor ops on fp32 inputs cause them // to be rounded to fp16. - if (cc_major >= 7 && std::is_same::value) { + if (cc_major >= 7 && TensorOpMathEnabled() && + std::is_same::value) { return true; } #endif @@ -2259,8 +2270,7 @@ port::Status CUDABlas::DoBlasGemmBatchedInternal( if (stream->parent()->GetDeviceDescription().cuda_compute_capability( &cc_major, &cc_minor) && cc_major >= 5) { - bool use_tensor_ops = - data_type == CUDA_R_16F || tensorflow::tf32_execution_allowed(); + bool use_tensor_ops = TensorOpMathEnabled() && data_type == CUDA_R_16F; cublasGemmAlgo_t algo = (use_tensor_ops ? CUBLAS_GEMM_DFALT_TENSOR_OP : CUBLAS_GEMM_DFALT); cudaDataType_t compute_type = @@ -2274,7 +2284,7 @@ port::Status CUDABlas::DoBlasGemmBatchedInternal( bool ok; ok = DoBlasInternalImpl( AS_LAMBDA(cublasGemmBatchedEx), stream, true /* = pointer_mode_host */, - true /* = err_on_failure */, CUDABlasTranspose(transa), + true /* = err_on_failure */, use_tensor_ops, CUDABlasTranspose(transa), CUDABlasTranspose(transb), m, n, k, &alpha, a_void_ptrs, data_type, lda, b_void_ptrs, data_type, ldb, &beta, c_void_ptrs, data_type, ldc, batch_count, compute_type, algo); @@ -2409,25 +2419,33 @@ bool CUDABlas::DoBlasGemmStridedBatched( int lda, int64 stride_a, const DeviceMemory &b, int ldb, int64 stride_b, float beta, DeviceMemory *c, int ldc, int64 stride_c, int batch_count) { -#if CUDA_VERSION >= 9010 + bool use_tensor_ops = false; +#if CUDA_VERSION >= 9000 int cc_major, cc_minor; if (stream->parent()->GetDeviceDescription().cuda_compute_capability( - &cc_major, &cc_minor) && - cc_major >= 5) { - cublasGemmAlgo_t algo = - (cc_major >= 7 ? CUBLAS_GEMM_DFALT_TENSOR_OP : CUBLAS_GEMM_DFALT); - bool ok = DoBlasInternalImpl( - AS_LAMBDA(cublasGemmStridedBatchedEx), stream, - true /* = pointer_mode_host */, true /* = err_on_failure */, - CUDABlasTranspose(transa), CUDABlasTranspose(transb), m, n, k, &alpha, - GpuMemory(a), CUDA_R_16F, lda, stride_a, GpuMemory(b), CUDA_R_16F, ldb, - stride_b, &beta, GpuMemoryMutable(c), CUDA_R_16F, ldc, stride_c, - batch_count, CUDA_R_32F, algo); - if (ok) { - return true; + &cc_major, &cc_minor)) { + // GPUs < sm_70 don't support tensor ops. + if (cc_major >= 7 && TensorOpMathEnabled()) { + use_tensor_ops = true; } - LOG(ERROR) << "failed BLAS call, see log for details"; - return false; +#if CUDA_VERSION >= 9010 + if (cc_major >= 5) { + cublasGemmAlgo_t algo = + (use_tensor_ops ? CUBLAS_GEMM_DFALT_TENSOR_OP : CUBLAS_GEMM_DFALT); + bool ok = DoBlasInternalImpl( + AS_LAMBDA(cublasGemmStridedBatchedEx), stream, + true /* = pointer_mode_host */, true /* = err_on_failure */, + use_tensor_ops, CUDABlasTranspose(transa), CUDABlasTranspose(transb), + m, n, k, &alpha, GpuMemory(a), CUDA_R_16F, lda, stride_a, + GpuMemory(b), CUDA_R_16F, ldb, stride_b, &beta, GpuMemoryMutable(c), + CUDA_R_16F, ldc, stride_c, batch_count, CUDA_R_32F, algo); + if (ok) { + return true; + } + LOG(ERROR) << "failed BLAS call, see log for details"; + return false; + } +#endif } #endif // Either CUDA_VERSION < 9.1 or SM < 5.0. Fall back to a loop. @@ -2440,7 +2458,7 @@ bool CUDABlas::DoBlasGemmStridedBatched( reinterpret_cast<__half *>(GpuMemoryMutable(c) + batch * stride_c); bool ok = DoBlasInternalImpl( cublasSgemmEx, stream, true /* = pointer_mode_host */, - true /* = err_on_failure= */, CUDABlasTranspose(transa), + true /* = err_on_failure= */, use_tensor_ops, CUDABlasTranspose(transa), CUDABlasTranspose(transb), m, n, k, &alpha, a_matrix, SE_CUDA_DATA_HALF, lda, b_matrix, SE_CUDA_DATA_HALF, ldb, &beta, c_matrix, SE_CUDA_DATA_HALF, ldc); diff --git a/tensorflow/stream_executor/cuda/cuda_blas.h b/tensorflow/stream_executor/cuda/cuda_blas.h index 556456c83db..817bdb72777 100644 --- a/tensorflow/stream_executor/cuda/cuda_blas.h +++ b/tensorflow/stream_executor/cuda/cuda_blas.h @@ -83,7 +83,7 @@ class CUDABlas : public blas::BlasSupport { template bool DoBlasInternalImpl(FuncT cublas_func, Stream *stream, bool pointer_mode_host, bool err_on_failure, - Args... args); + bool use_tensor_op_math, Args... args); // Convenience functions that call DoBlasInternalImpl with different values // for err_on_failure. @@ -91,7 +91,8 @@ class CUDABlas : public blas::BlasSupport { bool DoBlasInternal(FuncT cublas_func, Stream *stream, bool pointer_mode_host, Args... args) { return DoBlasInternalImpl(cublas_func, stream, pointer_mode_host, - /*err_on_failure=*/true, args...); + /*err_on_failure=*/true, /*use_tensor_ops=*/false, + args...); } template bool DoBlasInternalFailureOK(FuncT cublas_func, Stream *stream, @@ -99,7 +100,8 @@ class CUDABlas : public blas::BlasSupport { // Tensor ops are hard-coded off in this path, but can still be enabled with // a specific algorithm choice as in DoBlasGemmWithAlgorithmImpl(). return DoBlasInternalImpl(cublas_func, stream, pointer_mode_host, - /*err_on_failure=*/false, args...); + /*err_on_failure=*/false, + /*use_tensor_ops=*/false, args...); } // A helper function to implement DoBlasGemmBatched interfaces for generic diff --git a/tensorflow/stream_executor/cuda/cuda_dnn.cc b/tensorflow/stream_executor/cuda/cuda_dnn.cc index 192bae91572..be18c989861 100644 --- a/tensorflow/stream_executor/cuda/cuda_dnn.cc +++ b/tensorflow/stream_executor/cuda/cuda_dnn.cc @@ -22,7 +22,6 @@ limitations under the License. #include "absl/strings/str_cat.h" #include "third_party/eigen3/Eigen/Core" #include "tensorflow/core/lib/core/errors.h" -#include "tensorflow/core/platform/tf32_utils.h" #include "tensorflow/core/util/env_var.h" #include "tensorflow/stream_executor/cuda/cuda_activation.h" #include "tensorflow/stream_executor/cuda/cuda_diagnostics.h" @@ -602,6 +601,31 @@ class CudnnFilterDescriptor { SE_DISALLOW_COPY_AND_ASSIGN(CudnnFilterDescriptor); }; +// A helper function to decide whether to enable the TENSOR_OP_MATH math type +bool TensorOpMathEnabled() { + static bool is_enabled = [] { + bool is_disabled = false; + TF_CHECK_OK( + tensorflow::ReadBoolFromEnvVar("TF_DISABLE_CUDNN_TENSOR_OP_MATH", + /*default_val=*/false, &is_disabled)); + return !is_disabled; + }(); + return is_enabled; +} + +// A helper function to decide whether to enable the TENSOR_OP_MATH math type +// for RNNs. +bool RnnTensorOpMathEnabled() { + static bool is_enabled = [] { + bool is_disabled = false; + TF_CHECK_OK( + tensorflow::ReadBoolFromEnvVar("TF_DISABLE_CUDNN_RNN_TENSOR_OP_MATH", + /*default_val=*/false, &is_disabled)); + return !is_disabled; + }(); + return is_enabled; +} + // A helper function to decide whether to use // CUDNN_BATCHNORM_SPATIAL_PERSISTENT in batchnorm. This mode can be faster in // some tasks because an optimized path may be selected for CUDNN_DATA_FLOAT @@ -706,6 +730,10 @@ class CudnnConvolutionDescriptor { : CUDNN_CROSS_CORRELATION, data_type)); + // NOTE(benbarsdell): This only applies if tensor op math is enabled + // and algo selection is set to Default. + this->set_use_tensor_op_math(true); + #if CUDNN_MAJOR >= 7 VLOG(2) << "Requesting grouped convolution: " << convolution_descriptor.group_count(); @@ -717,15 +745,13 @@ class CudnnConvolutionDescriptor { #endif } - void set_use_tensor_op_math(bool use_tensor_op_math) { + void set_use_tensor_op_math(bool use_tensor_op_math) const { #if CUDNN_VERSION >= 7000 cudnnMathType_t math_type = -#if CUDNN_VERSION >= 8000 - (use_tensor_op_math ? CUDNN_TENSOR_OP_MATH : CUDNN_FMA_MATH); -#else (use_tensor_op_math ? CUDNN_TENSOR_OP_MATH : CUDNN_DEFAULT_MATH); -#endif - CHECK_CUDNN_OK(cudnnSetConvolutionMathType(handle_.get(), math_type)); + if (TensorOpMathEnabled()) { + CHECK_CUDNN_OK(cudnnSetConvolutionMathType(handle_.get(), math_type)); + } #endif } @@ -737,40 +763,6 @@ class CudnnConvolutionDescriptor { SE_DISALLOW_COPY_AND_ASSIGN(CudnnConvolutionDescriptor); }; -// A helper function to query if a CudnnConvolutionDescriptor has tensor_op_math -// set -static bool IsTensorMathOpSet(const CudnnConvolutionDescriptor& conv) { - cudnnMathType_t math_type; - CHECK_CUDNN_OK(cudnnGetConvolutionMathType(conv.handle(), &math_type)); -#if CUDNN_VERSION >= 8000 - return math_type != CUDNN_FMA_MATH; -#else - return math_type == CUDNN_TENSOR_OP_MATH; -#endif -} - -static bool TensorOpMathAvailable(int cc_major) { - return cc_major >= 7 && CUDNN_VERSION >= 7000; -} - -static bool IsTensorMathAllowed(Stream* stream, dnn::DataType input_type) { - int cc_major, cc_minor; - std::tie(cc_major, cc_minor) = GetCcMajorMinor(stream); - if (!TensorOpMathAvailable(cc_major)) { - return false; - } - if (input_type == dnn::DataType::kFloat) { -#if CUDNN_VERSION < 8000 - return false; -#else - if (!tensorflow::tf32_execution_allowed()) { - return false; - } -#endif - } - return true; -} - // Turns a PoolingDescriptor structure into a cudnn pooling descriptor handle // within a scope. class CudnnPoolingDescriptor { @@ -1163,31 +1155,21 @@ class CudnnRnnDescriptor : public dnn::RnnDescriptor { // in profile mode, which is run with algorithms returned from // GetRnnAlgorithms() (which are non-default and explicitly set whether to // use tensor ops). CuDNN 7.2.1 fixed this issue - bool allow_tensor_ops = - data_type != CUDNN_DATA_FLOAT || tensorflow::tf32_execution_allowed(); - bool use_tensor_ops; - if (algorithm_config.algorithm().has_value()) { - use_tensor_ops = algorithm_config.algorithm()->tensor_ops_enabled(); - } else { - use_tensor_ops = CUDNN_VERSION >= 7201 && allow_tensor_ops; - } - - if (use_tensor_ops && !allow_tensor_ops) { - return port::Status(port::error::INVALID_ARGUMENT, - "Algo requests disallowed tensor op evaluation."); - } - - cudnnMathType_t math_type; - if (use_tensor_ops) { - math_type = CUDNN_TENSOR_OP_MATH; - } else { -#if CUDNN_VERSION >= 8000 - math_type = CUDNN_FMA_MATH; + if (RnnTensorOpMathEnabled()) { + cudnnMathType_t math_type; + if (algorithm_config.algorithm().has_value()) { + math_type = algorithm_config.algorithm()->tensor_ops_enabled() + ? CUDNN_TENSOR_OP_MATH + : CUDNN_DEFAULT_MATH; + } else { +#if CUDNN_VERSION >= 7201 + math_type = CUDNN_TENSOR_OP_MATH; #else - math_type = CUDNN_DEFAULT_MATH; -#endif // CUDNN_VERSION >= 8000 + math_type = CUDNN_DEFAULT_MATH; +#endif // CUDNN_VERSION >= 7201 + } + CHECK_CUDNN_OK(cudnnSetRNNMatrixMathType(rnn_desc.get(), math_type)); } - CHECK_CUDNN_OK(cudnnSetRNNMatrixMathType(rnn_desc.get(), math_type)); #endif // CUDNN_VERSION >= 7000 return CudnnRnnDescriptor(cudnn, std::move(rnn_desc), std::move(rnn_plan), @@ -2578,11 +2560,10 @@ port::StatusOr> AllocateCudnnConvolutionForwardWorkspace( const CudnnTensorDescriptor& output_nd, const dnn::AlgorithmDesc& algorithm_desc, ScratchAllocator* scratch_allocator) { - if (IsTensorMathOpSet(conv) != algorithm_desc.tensor_ops_enabled()) { - return port::Status( - port::error::INTERNAL, - "Mismatch between cudnn conv and algorithm descriptors."); - } + // TODO(csigg): This has side effects on the convolution descriptor. It is + // functionally correct because the convolution is run with the algorithm of + // the last call to this function, but should be fixed anyway. + conv.set_use_tensor_op_math(algorithm_desc.tensor_ops_enabled()); // Query the size of the workspace and allocate it. size_t size_in_bytes; @@ -2622,11 +2603,10 @@ AllocateCudnnConvolutionBackwardDataWorkspace( const CudnnTensorDescriptor& output_nd, const dnn::AlgorithmDesc& algorithm_desc, ScratchAllocator* scratch_allocator) { - if (IsTensorMathOpSet(conv) != algorithm_desc.tensor_ops_enabled()) { - return port::Status( - port::error::INTERNAL, - "Mismatch between cudnn conv and algorithm descriptors."); - } + // TODO(csigg): This has side effects on the convolution descriptor. It is + // functionally correct because the convolution is run with the algorithm of + // the last call to this function, but should be fixed anyway. + conv.set_use_tensor_op_math(algorithm_desc.tensor_ops_enabled()); // Query the size of the workspace and allocate it. size_t size_in_bytes; @@ -2668,11 +2648,10 @@ AllocateCudnnConvolutionBackwardFilterWorkspace( const CudnnTensorDescriptor& output_nd, const dnn::AlgorithmDesc& algorithm_desc, ScratchAllocator* scratch_allocator) { - if (IsTensorMathOpSet(conv) != algorithm_desc.tensor_ops_enabled()) { - return port::Status( - port::error::INTERNAL, - "Mismatch between cudnn conv and algorithm descriptors."); - } + // TODO(csigg): This has side effects on the convolution descriptor. It is + // functionally correct because the convolution is run with the algorithm of + // the last call to this function, but should be fixed anyway. + conv.set_use_tensor_op_math(algorithm_desc.tensor_ops_enabled()); // Query the size of the workspace and allocate it. size_t size_in_bytes; @@ -2706,42 +2685,18 @@ AllocateCudnnConvolutionBackwardFilterWorkspace( return scratch_allocator->AllocateBytes(size_in_bytes); } -port::StatusOr UseTensorOps(Stream* stream, dnn::DataType type, - absl::optional desc) { - bool use_tensor_ops; - if (desc.has_value()) { - use_tensor_ops = desc->tensor_ops_enabled(); - if (use_tensor_ops && !IsTensorMathAllowed(stream, type)) { - return port::Status(port::error::INVALID_ARGUMENT, - "Algo requests disallowed tensor op evaluation."); - } - } else { - use_tensor_ops = IsTensorMathAllowed(stream, type); - } - return use_tensor_ops; +static bool TensorOpMathAvailable(int cc_major) { + return cc_major >= 7 && CUDNN_VERSION >= 7000 && TensorOpMathEnabled(); } -cudnnDataType_t GetRnnComputeType(dnn::DataType data_type); -dnn::DataType GetConvAccumulatorType(dnn::DataType data_type); - port::StatusOr GetCudnnConvolutionForwardAlgorithm( Stream* stream, const CudnnHandle& cudnn, const dnn::AlgorithmConfig& algorithm_config, const CudnnTensorDescriptor& input_nd, const CudnnFilterDescriptor& filter, - dnn::DataType element_type, - const dnn::ConvolutionDescriptor& convolution_descriptor, + const CudnnConvolutionDescriptor& conv, const CudnnTensorDescriptor& output_nd, ScratchAllocator* scratch_allocator, DeviceMemory* scratch) { absl::optional algo_desc = algorithm_config.algorithm(); - - CudnnConvolutionDescriptor conv( - convolution_descriptor, - ToCudnnDataType(GetConvAccumulatorType(element_type))); - bool use_tensor_ops; - SE_ASSIGN_OR_RETURN(use_tensor_ops, - UseTensorOps(stream, element_type, algo_desc)); - conv.set_use_tensor_op_math(use_tensor_ops); - if (!algo_desc.has_value()) { // Pick fastest algorithm within memory limit according to cuDNN's // heuristics. @@ -2754,7 +2709,10 @@ port::StatusOr GetCudnnConvolutionForwardAlgorithm( GetCudnnConvolutionForwardAlgo( cudnn, input_nd, filter, conv, output_nd, specify_workspace_limit, memory_limit_bytes)); - algo_desc = dnn::AlgorithmDesc(algo, use_tensor_ops); + int cc_major, cc_minor; + std::tie(cc_major, cc_minor) = GetCcMajorMinor(stream); + algo_desc = dnn::AlgorithmDesc( + algo, /*use_tensor_ops=*/TensorOpMathAvailable(cc_major)); } const auto scratch_or = AllocateCudnnConvolutionForwardWorkspace( @@ -2778,9 +2736,6 @@ port::StatusOr GetCudnnConvolutionForwardAlgorithm( "Returned status: ", scratch_or.status().ToString())); } - SE_ASSIGN_OR_RETURN(use_tensor_ops, - UseTensorOps(stream, element_type, algo_desc)); - conv.set_use_tensor_op_math(use_tensor_ops); SE_ASSIGN_OR_RETURN(*scratch, AllocateCudnnConvolutionForwardWorkspace( stream, cudnn, input_nd, filter, conv, output_nd, *algo_desc, scratch_allocator)); @@ -2791,19 +2746,10 @@ port::StatusOr GetCudnnConvolutionBackwardDataAlgorithm( Stream* stream, const CudnnHandle& cudnn, const dnn::AlgorithmConfig& algorithm_config, const CudnnTensorDescriptor& input_nd, const CudnnFilterDescriptor& filter, - dnn::DataType element_type, - const dnn::ConvolutionDescriptor& convolution_descriptor, + const CudnnConvolutionDescriptor& conv, const CudnnTensorDescriptor& output_nd, ScratchAllocator* scratch_allocator, DeviceMemory* scratch) { absl::optional algo_desc = algorithm_config.algorithm(); - CudnnConvolutionDescriptor conv( - convolution_descriptor, - ToCudnnDataType(GetConvAccumulatorType(element_type))); - bool use_tensor_ops; - SE_ASSIGN_OR_RETURN(use_tensor_ops, - UseTensorOps(stream, element_type, algo_desc)); - conv.set_use_tensor_op_math(use_tensor_ops); - if (!algo_desc.has_value()) { // Pick fastest algorithm within memory limit according to cuDNN's // heuristics. @@ -2816,7 +2762,10 @@ port::StatusOr GetCudnnConvolutionBackwardDataAlgorithm( GetCudnnConvolutionBackwardDataAlgo( cudnn, input_nd, filter, conv, output_nd, specify_workspace_limit, memory_limit_bytes)); - algo_desc = dnn::AlgorithmDesc(algo, use_tensor_ops); + int cc_major, cc_minor; + std::tie(cc_major, cc_minor) = GetCcMajorMinor(stream); + algo_desc = dnn::AlgorithmDesc( + algo, /*use_tensor_ops=*/TensorOpMathAvailable(cc_major)); } const auto scratch_or = AllocateCudnnConvolutionBackwardDataWorkspace( @@ -2839,9 +2788,6 @@ port::StatusOr GetCudnnConvolutionBackwardDataAlgorithm( "while a secondary algorithm is not provided."); } - SE_ASSIGN_OR_RETURN(use_tensor_ops, - UseTensorOps(stream, element_type, algo_desc)); - conv.set_use_tensor_op_math(use_tensor_ops); SE_ASSIGN_OR_RETURN(*scratch, AllocateCudnnConvolutionBackwardDataWorkspace( stream, cudnn, input_nd, filter, conv, output_nd, *algo_desc, scratch_allocator)); @@ -2852,19 +2798,10 @@ port::StatusOr GetCudnnConvolutionBackwardFilterAlgorithm( Stream* stream, const CudnnHandle& cudnn, const dnn::AlgorithmConfig& algorithm_config, const CudnnTensorDescriptor& input_nd, const CudnnFilterDescriptor& filter, - dnn::DataType element_type, - const dnn::ConvolutionDescriptor& convolution_descriptor, + const CudnnConvolutionDescriptor& conv, const CudnnTensorDescriptor& output_nd, ScratchAllocator* scratch_allocator, DeviceMemory* scratch) { absl::optional algo_desc = algorithm_config.algorithm(); - CudnnConvolutionDescriptor conv( - convolution_descriptor, - ToCudnnDataType(GetConvAccumulatorType(element_type))); - bool use_tensor_ops; - SE_ASSIGN_OR_RETURN(use_tensor_ops, - UseTensorOps(stream, element_type, algo_desc)); - conv.set_use_tensor_op_math(use_tensor_ops); - if (!algo_desc.has_value()) { // Pick fastest algorithm within memory limit according to cuDNN's // heuristics. @@ -2877,7 +2814,10 @@ port::StatusOr GetCudnnConvolutionBackwardFilterAlgorithm( GetCudnnConvolutionBackwardFilterAlgo( cudnn, input_nd, filter, conv, output_nd, specify_workspace_limit, memory_limit_bytes)); - algo_desc = dnn::AlgorithmDesc(algo, use_tensor_ops); + int cc_major, cc_minor; + std::tie(cc_major, cc_minor) = GetCcMajorMinor(stream); + algo_desc = dnn::AlgorithmDesc( + algo, /*use_tensor_ops=*/TensorOpMathAvailable(cc_major)); } auto scratch_or = AllocateCudnnConvolutionBackwardFilterWorkspace( @@ -2900,9 +2840,6 @@ port::StatusOr GetCudnnConvolutionBackwardFilterAlgorithm( "while a secondary algorithm is not provided."); } - SE_ASSIGN_OR_RETURN(use_tensor_ops, - UseTensorOps(stream, element_type, algo_desc)); - conv.set_use_tensor_op_math(use_tensor_ops); SE_ASSIGN_OR_RETURN(*scratch, AllocateCudnnConvolutionBackwardFilterWorkspace( stream, cudnn, input_nd, filter, conv, output_nd, *algo_desc, scratch_allocator)); @@ -3067,32 +3004,35 @@ port::Status CudnnSupport::DoPrepareForConvolution( CudnnTensorDescriptor output_nd( output_descriptor, ToCudnnDataType(element_type, output_descriptor.layout())); + CudnnConvolutionDescriptor conv( + convolution_descriptor, + ToCudnnDataType(GetConvAccumulatorType(element_type))); auto cudnn = cudnn_->GetHandle(parent_, stream); switch (kind) { case dnn::ConvolutionKind::FORWARD: { - SE_ASSIGN_OR_RETURN(*algorithm_desc, - GetCudnnConvolutionForwardAlgorithm( - stream, cudnn, algorithm_config, input_nd, - filter_nd, element_type, convolution_descriptor, - output_nd, scratch_allocator, scratch_memory)); + SE_ASSIGN_OR_RETURN( + *algorithm_desc, + GetCudnnConvolutionForwardAlgorithm( + stream, cudnn, algorithm_config, input_nd, filter_nd, conv, + output_nd, scratch_allocator, scratch_memory)); break; } case dnn::ConvolutionKind::BACKWARD_DATA: { - SE_ASSIGN_OR_RETURN(*algorithm_desc, - GetCudnnConvolutionBackwardDataAlgorithm( - stream, cudnn, algorithm_config, input_nd, - filter_nd, element_type, convolution_descriptor, - output_nd, scratch_allocator, scratch_memory)); + SE_ASSIGN_OR_RETURN( + *algorithm_desc, + GetCudnnConvolutionBackwardDataAlgorithm( + stream, cudnn, algorithm_config, input_nd, filter_nd, conv, + output_nd, scratch_allocator, scratch_memory)); break; } case dnn::ConvolutionKind::BACKWARD_FILTER: { - SE_ASSIGN_OR_RETURN(*algorithm_desc, - GetCudnnConvolutionBackwardFilterAlgorithm( - stream, cudnn, algorithm_config, input_nd, - filter_nd, element_type, convolution_descriptor, - output_nd, scratch_allocator, scratch_memory)); + SE_ASSIGN_OR_RETURN( + *algorithm_desc, + GetCudnnConvolutionBackwardFilterAlgorithm( + stream, cudnn, algorithm_config, input_nd, filter_nd, conv, + output_nd, scratch_allocator, scratch_memory)); break; } default: @@ -3121,9 +3061,8 @@ port::Status CudnnSupport::DoConvolve( auto accumulator_type = GetConvAccumulatorType(element_type); CudnnConvolutionDescriptor conv(convolution_descriptor, ToCudnnDataType(accumulator_type)); - SE_ASSIGN_OR_RETURN(bool use_tensor_ops, - UseTensorOps(stream, element_type, algorithm_desc)); - conv.set_use_tensor_op_math(use_tensor_ops); + // Set use_tensor_math param to correct value + conv.set_use_tensor_op_math(algorithm_desc.tensor_ops_enabled()); auto cudnn = cudnn_->GetHandle(parent_, stream); // Alpha is the scaling factor for input. @@ -3356,6 +3295,14 @@ port::Status CudnnSupport::DoConvolve( return port::Status::OK(); } +// A helper function to query if a CudnnConvolutionDescriptor has tensor_op_math +// set +static bool IsTensorMathOpSet(const CudnnConvolutionDescriptor& conv) { + cudnnMathType_t math_type; + CHECK_CUDNN_OK(cudnnGetConvolutionMathType(conv.handle(), &math_type)); + return math_type == CUDNN_TENSOR_OP_MATH; +} + template port::Status CudnnSupport::DoFusedConvolveImpl( @@ -3389,6 +3336,8 @@ port::Status CudnnSupport::DoFusedConvolveImpl( filter_descriptor, GetCudnnDataType(conv_input_descriptor.layout())); CudnnTensorDescriptor bias_nd(bias_descriptor, GetCudnnDataType()); + CudnnConvolutionDescriptor conv(convolution_descriptor, + ToCudnnDataType(accumulator_type)); auto cudnn = cudnn_->GetHandle(parent_, stream); @@ -3398,14 +3347,9 @@ port::Status CudnnSupport::DoFusedConvolveImpl( SE_ASSIGN_OR_RETURN( dnn::AlgorithmDesc algo_desc, GetCudnnConvolutionForwardAlgorithm( - stream, cudnn, algorithm_config, conv_input_nd, filter, - dnn::ToDataType::value, convolution_descriptor, + stream, cudnn, algorithm_config, conv_input_nd, filter, conv, output_nd, scratch_allocator, &scratch)); - CudnnConvolutionDescriptor conv(convolution_descriptor, - ToCudnnDataType(accumulator_type)); - conv.set_use_tensor_op_math(algo_desc.tensor_ops_enabled()); - std::unique_ptr timer; if (is_profiling) { timer.reset(new GpuTimer(parent_)); // NOLINT @@ -3536,7 +3480,9 @@ bool CudnnSupport::GetRnnAlgorithms( for (auto i : algo_types) { out_algorithms->push_back({i, /*use_tensor_ops=*/false}); #if CUDNN_VERSION >= 7100 - out_algorithms->push_back({i, /*use_tensor_ops=*/true}); + if (RnnTensorOpMathEnabled()) { + out_algorithms->push_back({i, /*use_tensor_ops=*/true}); + } #endif } return true; From 64f7bdd56a394ecae55c5006e483050569b9b136 Mon Sep 17 00:00:00 2001 From: Thai Nguyen Date: Thu, 18 Jun 2020 19:32:16 -0700 Subject: [PATCH 263/867] Disable tsan on InterpreterFlexTest and SelectiveBuiltInterpreterFlexTest PiperOrigin-RevId: 317231748 Change-Id: I7ab662fd55024c0ed91bd78bfdc8e9206d78b3b6 --- tensorflow/lite/delegates/flex/BUILD | 1 + tensorflow/lite/java/BUILD | 1 + 2 files changed, 2 insertions(+) diff --git a/tensorflow/lite/delegates/flex/BUILD b/tensorflow/lite/delegates/flex/BUILD index 42914bf5ab8..99bcf05ab4a 100644 --- a/tensorflow/lite/delegates/flex/BUILD +++ b/tensorflow/lite/delegates/flex/BUILD @@ -279,6 +279,7 @@ java_test( "no_oss", # Currently requires --config=monolithic, b/118895218. # TODO(b/121204962): Re-enable test after fixing memory leaks. "noasan", + "notsan", # TODO(b/158651814) Re-enable after fixing racing condition. ], test_class = "org.tensorflow.lite.InterpreterFlexTest", visibility = ["//visibility:private"], diff --git a/tensorflow/lite/java/BUILD b/tensorflow/lite/java/BUILD index 738d66a0eb1..89be932ab4d 100644 --- a/tensorflow/lite/java/BUILD +++ b/tensorflow/lite/java/BUILD @@ -304,6 +304,7 @@ java_test( "no_oss", # Currently requires --config=monolithic, b/118895218. # TODO(b/121204962): Re-enable test after fixing memory leaks. "noasan", + "notsan", # TODO(b/158651814) Re-enable after fixing racing condition. ], test_class = "org.tensorflow.lite.InterpreterFlexTest", visibility = ["//visibility:private"], From c159f1599548428660c80dada924d69f269384a3 Mon Sep 17 00:00:00 2001 From: Scott Zhu Date: Thu, 18 Jun 2020 19:35:14 -0700 Subject: [PATCH 264/867] Fork the keras related tpu_strategy_test to keras integration test. PiperOrigin-RevId: 317232048 Change-Id: If05867985ff1ff81ac45bb601b701ee68d4d5279 --- tensorflow/python/distribute/BUILD | 1 - .../python/distribute/tpu_strategy_test.py | 19 ----- .../python/keras/integration_test/BUILD | 13 ++++ .../integration_test/tpu_strategy_test.py | 69 +++++++++++++++++++ 4 files changed, 82 insertions(+), 20 deletions(-) create mode 100644 tensorflow/python/keras/integration_test/tpu_strategy_test.py diff --git a/tensorflow/python/distribute/BUILD b/tensorflow/python/distribute/BUILD index 7208807a18c..4d77c12f975 100644 --- a/tensorflow/python/distribute/BUILD +++ b/tensorflow/python/distribute/BUILD @@ -654,7 +654,6 @@ tpu_py_test( "//tensorflow/python/distribute/cluster_resolver:cluster_resolver_lib", "//tensorflow/python/eager:remote", "//tensorflow/python/eager:test", - "//tensorflow/python/keras", ], ) diff --git a/tensorflow/python/distribute/tpu_strategy_test.py b/tensorflow/python/distribute/tpu_strategy_test.py index 6dd7de500e4..400b12112d6 100644 --- a/tensorflow/python/distribute/tpu_strategy_test.py +++ b/tensorflow/python/distribute/tpu_strategy_test.py @@ -18,7 +18,6 @@ from __future__ import absolute_import from __future__ import division from __future__ import print_function -from tensorflow.python import keras from tensorflow.python.data.ops import dataset_ops from tensorflow.python.distribute import distribute_lib from tensorflow.python.distribute import distribution_strategy_context @@ -364,24 +363,6 @@ class TPUStrategyTest(test.TestCase): expected_result, strategy.experimental_local_results(train_step(next(input_iterator)))) - def test_keras_metric_outside_strategy_scope_per_replica(self): - strategy = get_tpu_strategy() - metric = keras.metrics.Mean("test_metric", dtype=dtypes.float32) - - dataset = dataset_ops.Dataset.range(strategy.num_replicas_in_sync * - 2).batch(2) - dataset = strategy.experimental_distribute_dataset(dataset) - - @def_function.function - def step_fn(i): - metric.update_state(i) - - with self.assertRaisesRegex(ValueError, "Trying to run metric.update_state " - "in replica context"): - with strategy.scope(): - for i in dataset: - strategy.run(step_fn, args=(i,)) - # TODO(b/145574622): Remove this test once it is re-enabled in values_test.py. def test_all_reduce_on_sync_on_read_variable(self): strategy = get_tpu_strategy() diff --git a/tensorflow/python/keras/integration_test/BUILD b/tensorflow/python/keras/integration_test/BUILD index 2ef775a190e..b23dcc59b97 100644 --- a/tensorflow/python/keras/integration_test/BUILD +++ b/tensorflow/python/keras/integration_test/BUILD @@ -2,6 +2,7 @@ # Contains Keras integration tests that verify with other TF high level APIs. load("//tensorflow:tensorflow.bzl", "cuda_py_test", "tf_py_test") +load("//tensorflow/python/tpu:tpu.bzl", "tpu_py_test") package( default_visibility = [ @@ -91,3 +92,15 @@ cuda_py_test( "//tensorflow/python:extra_py_tests_deps", ], ) + +tpu_py_test( + name = "tpu_strategy_test", + srcs = ["tpu_strategy_test.py"], + disable_experimental = True, + python_version = "PY3", + tags = ["no_oss"], + deps = [ + "//tensorflow:tensorflow_py", + "//tensorflow/python:extra_py_tests_deps", + ], +) diff --git a/tensorflow/python/keras/integration_test/tpu_strategy_test.py b/tensorflow/python/keras/integration_test/tpu_strategy_test.py new file mode 100644 index 00000000000..d24e96ae855 --- /dev/null +++ b/tensorflow/python/keras/integration_test/tpu_strategy_test.py @@ -0,0 +1,69 @@ +# Copyright 2018 The TensorFlow Authors. All Rights Reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. +# ============================================================================== +"""Tests for TPUStrategy.""" + +from __future__ import absolute_import +from __future__ import division +from __future__ import print_function + +from absl import flags + +import tensorflow as tf + + +FLAGS = flags.FLAGS +flags.DEFINE_string("tpu", "", "Name of TPU to connect to.") +flags.DEFINE_string("project", None, "Name of GCP project with TPU.") +flags.DEFINE_string("zone", None, "Name of GCP zone with TPU.") + + +def get_tpu_cluster_resolver(): + resolver = tf.distribute.cluster_resolver.TPUClusterResolver( + tpu=FLAGS.tpu, + zone=FLAGS.zone, + project=FLAGS.project, + ) + return resolver + + +def get_tpu_strategy(): + resolver = get_tpu_cluster_resolver() + tf.config.experimental_connect_to_cluster(resolver) + tf.tpu.experimental.initialize_tpu_system(resolver) + return tf.distribute.experimental.TPUStrategy(resolver) + + +class TpuStrategyTest(tf.test.TestCase): + + def test_keras_metric_outside_strategy_scope_per_replica(self): + strategy = get_tpu_strategy() + metric = tf.keras.metrics.Mean("test_metric", dtype=tf.float32) + + dataset = tf.data.Dataset.range(strategy.num_replicas_in_sync * 2).batch(2) + dataset = strategy.experimental_distribute_dataset(dataset) + + @tf.function + def step_fn(i): + metric.update_state(i) + + with self.assertRaisesRegex(ValueError, "Trying to run metric.update_state " + "in replica context"): + with strategy.scope(): + for i in dataset: + strategy.run(step_fn, args=(i,)) + + +if __name__ == "__main__": + tf.test.main() From 4d54ef31394aefe270826790164edcc6d687bb63 Mon Sep 17 00:00:00 2001 From: "A. Unique TensorFlower" Date: Thu, 18 Jun 2020 20:02:06 -0700 Subject: [PATCH 265/867] Enable type annotations for python/ops. PiperOrigin-RevId: 317234494 Change-Id: I49a24cd1e2127a3c7b0f2eb217cfe023ce5b439f --- tensorflow/python/ops/logging_ops.py | 18 +++++++++++++++--- 1 file changed, 15 insertions(+), 3 deletions(-) diff --git a/tensorflow/python/ops/logging_ops.py b/tensorflow/python/ops/logging_ops.py index 02fce277690..8ca63f55987 100644 --- a/tensorflow/python/ops/logging_ops.py +++ b/tensorflow/python/ops/logging_ops.py @@ -54,9 +54,11 @@ except NameError: # call relies on certain conditionals for its dependencies. Use # control_flow_ops.Assert. -# Assert and Print are special symbols in Python 2, so we must -# have an upper-case version of them. When support for it is dropped, -# we can allow lowercase. +# Assert and Print are special symbols in python, so we must +# have an upper-case version of them. +# +# For users with Python 3 or Python 2.7 +# with `from __future__ import print_function`, we could also allow lowercase. # See https://github.com/tensorflow/tensorflow/issues/18053 @@ -81,6 +83,11 @@ def Print(input_, data, message=None, first_n=None, summarize=None, name=None): with jupyter notebook (printing to the notebook *server's* output, not into the notebook). + Additionally, to use tf.print in python 2.7, users must make sure to import + the following: + + `from __future__ import print_function` + Args: input_: A tensor passed through this op. data: A list of tensors to print out when op is evaluated. @@ -141,6 +148,11 @@ def print_v2(*inputs, **kwargs): Python objects. Printed tensors will recursively show the first and last elements of each dimension to summarize. + @compatibility(python2) + In python 2.7, make sure to import the following: + `from __future__ import print_function` + @end_compatibility + Example: Single-input usage: From 7e6e549c461118fbefdb11d03adbc80c27109a8a Mon Sep 17 00:00:00 2001 From: Yujing Zhang Date: Thu, 18 Jun 2020 20:03:31 -0700 Subject: [PATCH 266/867] Support packed variable in DistributedVariable. Add an option to enable packed variable in TPUStrategy. PiperOrigin-RevId: 317234665 Change-Id: I09e806cb8261815cd87a6d98817556dd8f7e8ed7 --- tensorflow/python/distribute/BUILD | 6 +- .../python/distribute/checkpointing_test.py | 2 + .../custom_training_loop_input_test.py | 7 +- .../python/distribute/distribute_utils.py | 3 + .../distribute/packed_distributed_variable.py | 13 +- .../packed_distributed_variable_test.py | 6 +- .../distribute/saved_model_test_base.py | 1 + .../distribute/strategy_combinations.py | 20 +- tensorflow/python/distribute/tpu_strategy.py | 50 ++- .../python/distribute/tpu_strategy_test.py | 410 +++++++++--------- tensorflow/python/distribute/tpu_values.py | 41 +- tensorflow/python/distribute/values.py | 53 ++- tensorflow/python/distribute/values_test.py | 91 ++-- tensorflow/python/distribute/values_util.py | 12 +- tensorflow/python/tpu/tpu.py | 9 +- 15 files changed, 454 insertions(+), 270 deletions(-) diff --git a/tensorflow/python/distribute/BUILD b/tensorflow/python/distribute/BUILD index 4d77c12f975..0062705126f 100644 --- a/tensorflow/python/distribute/BUILD +++ b/tensorflow/python/distribute/BUILD @@ -654,6 +654,7 @@ tpu_py_test( "//tensorflow/python/distribute/cluster_resolver:cluster_resolver_lib", "//tensorflow/python/eager:remote", "//tensorflow/python/eager:test", + "@absl_py//absl/testing:parameterized", ], ) @@ -787,6 +788,7 @@ py_library( name = "tpu_values", srcs = ["tpu_values.py"], deps = [ + ":packed_distributed_variable", ":values", "//tensorflow/python:framework_ops", "//tensorflow/python:math_ops", @@ -1602,7 +1604,7 @@ distribute_py_test( srcs = ["saved_model_save_load_test.py"], full_precision = True, main = "saved_model_save_load_test.py", - shard_count = 5, + shard_count = 7, tags = [ "multi_and_single_gpu", "no_rocm", @@ -1635,7 +1637,7 @@ distribute_py_test( srcs = ["saved_model_mixed_api_test.py"], full_precision = True, main = "saved_model_mixed_api_test.py", - shard_count = 5, + shard_count = 7, tags = [ "multi_and_single_gpu", "no_rocm", diff --git a/tensorflow/python/distribute/checkpointing_test.py b/tensorflow/python/distribute/checkpointing_test.py index ad646905315..edd4c46c371 100644 --- a/tensorflow/python/distribute/checkpointing_test.py +++ b/tensorflow/python/distribute/checkpointing_test.py @@ -103,6 +103,7 @@ class TrainingCheckpointTests(test.TestCase, parameterized.TestCase): strategy_combinations.mirrored_strategy_with_one_cpu, strategy_combinations.mirrored_strategy_with_gpu_and_cpu, strategy_combinations.tpu_strategy, + strategy_combinations.tpu_strategy_packed_var, strategy_combinations.central_storage_strategy_with_two_gpus, ], mode=["eager"])) @@ -138,6 +139,7 @@ class TrainingCheckpointTests(test.TestCase, parameterized.TestCase): strategy_combinations.mirrored_strategy_with_one_cpu, strategy_combinations.mirrored_strategy_with_gpu_and_cpu, strategy_combinations.tpu_strategy, + strategy_combinations.tpu_strategy_packed_var, strategy_combinations.central_storage_strategy_with_two_gpus, ], mode=["eager"])) diff --git a/tensorflow/python/distribute/custom_training_loop_input_test.py b/tensorflow/python/distribute/custom_training_loop_input_test.py index 5d1584f5aa7..e4f782810dd 100644 --- a/tensorflow/python/distribute/custom_training_loop_input_test.py +++ b/tensorflow/python/distribute/custom_training_loop_input_test.py @@ -197,7 +197,8 @@ class InputIterationTest(test.TestCase, parameterized.TestCase, combinations.combine( distribution=[ strategy_combinations.mirrored_strategy_with_gpu_and_cpu, - strategy_combinations.tpu_strategy + strategy_combinations.tpu_strategy, + strategy_combinations.tpu_strategy_packed_var, ], mode=["eager"])) def testNestedOutput(self, distribution): @@ -748,6 +749,10 @@ class InputIterationTest(test.TestCase, parameterized.TestCase, mode=["eager"] )) def testMultiDeviceDataCapturedFunction(self, distribution): + if getattr(distribution, "_enable_packed_variable_in_eager_mode", False): + self.skipTest( + "Dataset captured function doesn't support packed tensors yet " + "(b/145922293).") inputs = constant_op.constant([2., 3.]) dataset = lambda _: dataset_ops.Dataset.from_tensor_slices(inputs).repeat(5) input_iterator = iter( diff --git a/tensorflow/python/distribute/distribute_utils.py b/tensorflow/python/distribute/distribute_utils.py index ccf19521718..14b934b4a0f 100644 --- a/tensorflow/python/distribute/distribute_utils.py +++ b/tensorflow/python/distribute/distribute_utils.py @@ -148,6 +148,9 @@ def select_replica_mirrored(replica_id, structured): raise TypeError( "Expected value to be mirrored across replicas: %s in %s." % (x, structured)) + packed_var = getattr(x, "_packed_variable", None) + if packed_var is not None: + return packed_var return x.values[replica_id] else: return x diff --git a/tensorflow/python/distribute/packed_distributed_variable.py b/tensorflow/python/distribute/packed_distributed_variable.py index 62512cb4414..c249b8efc1c 100644 --- a/tensorflow/python/distribute/packed_distributed_variable.py +++ b/tensorflow/python/distribute/packed_distributed_variable.py @@ -42,7 +42,7 @@ class PackedDistributedVariable(resource_variable_ops.BaseResourceVariable): name: Optional name for the variable. Defaults to `'Variable'` and gets uniquified automatically. """ - if not context.executing_eagerly(): + if not ops.executing_eagerly_outside_functions(): raise ValueError( "PackedDistributedVariable should be created in eager mode.") if not distributed_variables: @@ -84,6 +84,9 @@ class PackedDistributedVariable(resource_variable_ops.BaseResourceVariable): def devices(self): return self._devices + def on_device(self, device): + return PackedVarAndDevice(self, device) + def get_var_on_device(self, device): for i, d in enumerate(self._devices): if d == device: @@ -100,7 +103,10 @@ class PackedDistributedVariable(resource_variable_ops.BaseResourceVariable): @property def handle(self): - return self._handle + if context.executing_eagerly(): + return self.get_var_on_current_device().handle + else: + return self._handle def _read_variable_op(self): if context.executing_eagerly(): @@ -269,7 +275,8 @@ class PackedVarAndDevice(object): @property def handle(self): - return self._var.handle + with ops.device(self._device): + return self._var.handle @property def op(self): diff --git a/tensorflow/python/distribute/packed_distributed_variable_test.py b/tensorflow/python/distribute/packed_distributed_variable_test.py index d29d19960a5..ec2e476e4b8 100644 --- a/tensorflow/python/distribute/packed_distributed_variable_test.py +++ b/tensorflow/python/distribute/packed_distributed_variable_test.py @@ -46,7 +46,7 @@ class PackedDistributedVariableTest(test.TestCase): v1 = resource_variable_ops.ResourceVariable(2.0, name='var1') packed_var = packed_distributed_variable.PackedDistributedVariable([v0, v1]) - self.assertTrue(packed_var.handle.is_packed) + self.assertFalse(packed_var.handle.is_packed) self.assertTrue(packed_var.is_initialized) with ops.device('/cpu:0'): @@ -61,6 +61,7 @@ class PackedDistributedVariableTest(test.TestCase): @def_function.function def update_var(): + self.assertTrue(packed_var.handle.is_packed) with ops.device('/cpu:0'): packed_var.assign_add(3.0).assign_sub(1.0) read0 = packed_var.value() @@ -85,7 +86,7 @@ class PackedDistributedVariableTest(test.TestCase): packed_var0 = packed_distributed_variable.PackedVarAndDevice( packed_var, device0) - self.assertTrue(packed_var0.handle.is_packed) + self.assertFalse(packed_var0.handle.is_packed) self.assertAllEqual(math_ops.mul(packed_var0, 2.0), 2.0) packed_var1 = packed_distributed_variable.PackedVarAndDevice( @@ -94,6 +95,7 @@ class PackedDistributedVariableTest(test.TestCase): @def_function.function def func(): + self.assertTrue(packed_var.handle.is_packed) var0 = packed_distributed_variable.PackedVarAndDevice(packed_var, device0) var0.assign_add(3.0) var1 = packed_distributed_variable.PackedVarAndDevice(packed_var, device1) diff --git a/tensorflow/python/distribute/saved_model_test_base.py b/tensorflow/python/distribute/saved_model_test_base.py index e544e51cddd..70ea582baff 100644 --- a/tensorflow/python/distribute/saved_model_test_base.py +++ b/tensorflow/python/distribute/saved_model_test_base.py @@ -58,6 +58,7 @@ strategies = [ strategy_combinations.mirrored_strategy_with_gpu_and_cpu, strategy_combinations.mirrored_strategy_with_two_gpus, strategy_combinations.tpu_strategy, + strategy_combinations.tpu_strategy_packed_var, strategy_combinations.central_storage_strategy_with_two_gpus, ] diff --git a/tensorflow/python/distribute/strategy_combinations.py b/tensorflow/python/distribute/strategy_combinations.py index 350b187f67f..1fa42cb8645 100644 --- a/tensorflow/python/distribute/strategy_combinations.py +++ b/tensorflow/python/distribute/strategy_combinations.py @@ -53,7 +53,11 @@ _did_connect_to_cluster = False # pylint: disable=missing-docstring -def _get_tpu_strategy_creator(steps_per_run, use_single_core=False, **kwargs): +def _get_tpu_strategy_creator(steps_per_run, + use_single_core=False, + enable_packed_variable=False, + **kwargs): + def _create_tpu_strategy(): global _did_connect_to_cluster @@ -87,10 +91,13 @@ def _get_tpu_strategy_creator(steps_per_run, use_single_core=False, **kwargs): # Steps per run is only supported in TF 1.x if tf2.enabled(): - return tpu_lib.TPUStrategy(resolver, device_assignment, **kwargs) + strategy = tpu_lib.TPUStrategy(resolver, device_assignment, **kwargs) else: - return tpu_lib.TPUStrategyV1(resolver, steps_per_run, - device_assignment, **kwargs) + strategy = tpu_lib.TPUStrategyV1(resolver, steps_per_run, + device_assignment, **kwargs) + strategy._enable_packed_variable_in_eager_mode = enable_packed_variable # pylint: disable=protected-access + return strategy + return _create_tpu_strategy @@ -117,6 +124,10 @@ one_device_strategy_gpu_on_worker_1 = combinations.NamedDistribution( required_gpus=1) tpu_strategy = combinations.NamedDistribution( "TPU", _get_tpu_strategy_creator(steps_per_run=2), required_tpu=True) +tpu_strategy_packed_var = combinations.NamedDistribution( + "TPUPackedVar", + _get_tpu_strategy_creator(steps_per_run=2, enable_packed_variable=True), + required_tpu=True) tpu_strategy_one_step = combinations.NamedDistribution( "TPUOneStep", _get_tpu_strategy_creator(steps_per_run=1), required_tpu=True) tpu_strategy_one_core = combinations.NamedDistribution( @@ -286,6 +297,7 @@ strategies_minus_default_and_tpu = [ tpu_strategies = [ tpu_strategy, # steps_per_run=2 tpu_strategy_one_step, + tpu_strategy_packed_var, cloud_tpu_strategy, ] diff --git a/tensorflow/python/distribute/tpu_strategy.py b/tensorflow/python/distribute/tpu_strategy.py index 9493ecce767..7e8f5b97e7e 100644 --- a/tensorflow/python/distribute/tpu_strategy.py +++ b/tensorflow/python/distribute/tpu_strategy.py @@ -141,6 +141,10 @@ class TPUStrategy(distribute_lib.Strategy): "num_workers").set(self.extended.num_hosts) distribute_lib.distribution_strategy_replica_gauge.get_cell( "num_replicas_per_worker").set(self.extended.num_replicas_per_host) + # Packed variable is used to reduce the overhead of function execution. + # For a DistributedVariable, only one variable handle is captured into a + # function graph. It's only supported in eager mode. + self._enable_packed_variable_in_eager_mode = False # TODO(cjfj): Modify `_call_for_each_replica` in `TPUExtended` such that this # can use the default implementation. @@ -185,6 +189,10 @@ class TPUStrategyV1(distribute_lib.StrategyV1): "num_workers").set(self.extended.num_hosts) distribute_lib.distribution_strategy_replica_gauge.get_cell( "num_replicas_per_worker").set(self.extended.num_replicas_per_host) + # Packed variable is used to reduce the overhead of function execution. + # For a DistributedVariable, only one variable handle is captured into a + # function graph. It's only supported in eager mode. + self._enable_packed_variable_in_eager_mode = False @property def steps_per_run(self): @@ -671,20 +679,29 @@ class TPUExtended(distribute_lib.StrategyExtendedV1): return cross_device_ops_lib.reduce_non_distributed_value( reduce_op, value, destinations, self._num_replicas_in_sync) + value_list = value.values + # pylint: disable=protected-access + if isinstance( + value, + values.DistributedVariable) and value._packed_variable is not None: + value_list = tuple( + value._packed_variable.on_device(d) + for d in value._packed_variable.devices) + # pylint: enable=protected-access + # Currently XLA op by op mode has a limit for the number of inputs for a # single op, thus we break one `add_n` op into a group of `add_n` ops to # work around the constraint. # TODO(cjfj): Detect when it is possible to use `cross_replica_sum`. if len(value.values) <= _XLA_OP_BY_OP_INPUTS_LIMIT: - output = math_ops.add_n(value.values) + output = math_ops.add_n(value_list) else: - output = array_ops.zeros_like( - value.values[0], dtype=value.values[0].dtype) - for i in range(0, len(value.values), _XLA_OP_BY_OP_INPUTS_LIMIT): - output += math_ops.add_n(value.values[i:i + _XLA_OP_BY_OP_INPUTS_LIMIT]) + output = array_ops.zeros_like(value_list[0], dtype=value_list[0].dtype) + for i in range(0, len(value_list), _XLA_OP_BY_OP_INPUTS_LIMIT): + output += math_ops.add_n(value_list[i:i + _XLA_OP_BY_OP_INPUTS_LIMIT]) if reduce_op == reduce_util.ReduceOp.MEAN: - output *= (1. / len(value.values)) + output *= (1. / len(value_list)) devices = cross_device_ops_lib.get_devices_from(destinations) @@ -710,17 +727,28 @@ class TPUExtended(distribute_lib.StrategyExtendedV1): else: return (fn(var, *args, **kwargs),) - # Otherwise, we revert to MirroredStrategy behavior and update each variable - # directly. + # Otherwise, we revert to MirroredStrategy behavior and update the variable + # on each replica directly. updates = [] - for i, v in enumerate(var.values): + values_and_devices = [] + packed_var = var._packed_variable # pylint: disable=protected-access + if packed_var is not None: + for device in packed_var.devices: + values_and_devices.append((packed_var, device)) + else: + for value in var.values: + values_and_devices.append((value, value.device)) + + for i, value_and_device in enumerate(values_and_devices): + value = value_and_device[0] + device = value_and_device[1] name = "update_%d" % i - with ops.device(v.device), \ + with ops.device(device), \ distribute_lib.UpdateContext(i), \ ops.name_scope(name): # If args and kwargs are not mirrored, the value is returned as is. updates.append( - fn(v, *distribute_utils.select_replica_mirrored(i, args), + fn(value, *distribute_utils.select_replica_mirrored(i, args), **distribute_utils.select_replica_mirrored(i, kwargs))) return distribute_utils.update_regroup(self, updates, group) diff --git a/tensorflow/python/distribute/tpu_strategy_test.py b/tensorflow/python/distribute/tpu_strategy_test.py index 400b12112d6..4070336aae8 100644 --- a/tensorflow/python/distribute/tpu_strategy_test.py +++ b/tensorflow/python/distribute/tpu_strategy_test.py @@ -18,6 +18,8 @@ from __future__ import absolute_import from __future__ import division from __future__ import print_function +from absl.testing import parameterized + from tensorflow.python.data.ops import dataset_ops from tensorflow.python.distribute import distribute_lib from tensorflow.python.distribute import distribution_strategy_context @@ -64,14 +66,17 @@ def get_tpu_cluster_resolver(): return resolver -def get_tpu_strategy(): +def get_tpu_strategy(enable_packed_var=False): resolver = get_tpu_cluster_resolver() remote.connect_to_cluster(resolver) tpu_strategy_util.initialize_tpu_system(resolver) - return tpu_lib.TPUStrategy(resolver) + strategy = tpu_lib.TPUStrategy(resolver) + strategy._enable_packed_variable_in_eager_mode = enable_packed_var + return strategy -class TPUStrategyTest(test.TestCase): +# TPU tests which don't use TPUStrategy. +class TPUTest(test.TestCase): def test_multiple_initialize_system(self): resolver = get_tpu_cluster_resolver() @@ -82,177 +87,6 @@ class TPUStrategyTest(test.TestCase): tpu_strategy_util.initialize_tpu_system(resolver) self.assertRegex(str(mock_log.call_args), "already been initialized") - def test_sequential_experimental_runs(self): - resolver = get_tpu_cluster_resolver() - remote.connect_to_cluster(resolver) - topology = tpu_strategy_util.initialize_tpu_system(resolver) - # Computation replicated to all cores. - device_assignment = device_assignment_lib.DeviceAssignment.build( - topology, num_replicas=2) - strategy = tpu_lib.TPUStrategy( - resolver, device_assignment=device_assignment) - - # Computation on the 1st core. - device_assignment2 = device_assignment_lib.DeviceAssignment.build( - topology, num_replicas=1) - strategy2 = tpu_lib.TPUStrategy( - resolver, device_assignment=device_assignment2) - - def computation(x): - return math_ops.square(x) - - @def_function.function - def train_step(): - outputs = strategy.experimental_local_results( - strategy.run(computation, args=([2., 2.],))) - outputs2 = strategy2.run( - computation, args=([outputs[0]],)) - return outputs2 - - self.assertAllEqual([[16., 16.]], train_step()) - - def test_device_switch_case(self): - strategy = get_tpu_strategy() - with strategy.scope(): - a = variables.Variable(1) - - inference_iteration = variables.Variable(-1) - - def inference_fn(x, i): - return a + x + i - - @def_function.function - def run_inference(x): - - def do_inference(device, inference_fn, i): - with ops.device(device): - return inference_fn(x, i) - - branch_fns = { - 0: (lambda: do_inference("/device:TPU:0", inference_fn, 0)), - 1: (lambda: do_inference("/device:TPU:1", inference_fn, 1)), - } - branch_index = inference_iteration.assign_add(1, use_locking=True) % 2 - return control_flow_ops.switch_case(branch_index, branch_fns) - - self.assertAllEqual(2., run_inference(1)) # Use TPU core 0. - self.assertAllEqual(3., run_inference(1)) # Use TPU core 1. - - def test_recover_from_compilation_failures(self): - # TODO(b/148150981): Stop skipping this test once recovery works - # for non-local TPU. - if FLAGS.tpu: - self.skipTest("Recovery fails for non-local TPU, see b/148150981") - - # Disable automatic outside compilation. - config.set_soft_device_placement(False) - strategy = get_tpu_strategy() - - @def_function.function - def compilation_failure_run(): - - def computation(): - return random_ops.random_gamma([10], [0.5, 1.5]) - - return strategy.run(computation) - - with self.assertRaisesRegexp(errors.InvalidArgumentError, - "TPU compilation failed"): - compilation_failure_run() - - @def_function.function - def good_run(): - - def computation(): - return random_ops.random_normal([10]) - - return strategy.run(computation) - - good_run() - - def test_dynamic_shape_with_outside_compilation_failure(self): - # Enable automatic outside compilation. - config.set_soft_device_placement(True) - strategy = get_tpu_strategy() - dataset = dataset_ops.Dataset.from_tensors(("string", 1.0)).repeat().batch( - 2, drop_remainder=False) - dataset = strategy.experimental_distribute_dataset(dataset) - iterator = iter(dataset) - - @def_function.function - def train_fn(iterator): - - def step_fn(inputs): - _, inputs = inputs - return math_ops.reduce_sum(inputs) - - return strategy.experimental_local_results( - strategy.run(step_fn, args=(next(iterator),))) - - with self.assertRaisesRegex(errors.InternalError, "Compilation failure"): - logging.info(train_fn(iterator)) - - def test_computation_on_subset_cores(self): - resolver = get_tpu_cluster_resolver() - remote.connect_to_cluster(resolver) - topology = tpu_strategy_util.initialize_tpu_system(resolver) - all_core_strategy = tpu_lib.TPUStrategy(resolver) - - with all_core_strategy.scope(): - v = variables.Variable(0.0, - aggregation=variables.VariableAggregation.MEAN) - - # Computation on the 1st core. - device_assignment = device_assignment_lib.DeviceAssignment.build( - topology, num_replicas=1) - first_core_strategy = tpu_lib.TPUStrategy( - resolver, device_assignment=device_assignment) - - # Computation on the 2nd core. - device_assignment2 = device_assignment_lib.DeviceAssignment( - topology, [[[0, 0, 0, 1]]]) - second_core_strategy = tpu_lib.TPUStrategy( - resolver, device_assignment=device_assignment2) - - @def_function.function - def train_step(): - - def step_fn(): - return v + 1.0 - - all_core_strategy.run(step_fn) - r1 = first_core_strategy.run(step_fn) - r2 = second_core_strategy.run(step_fn) - return r1 + r2 - - train_step() - self.assertAllEqual(2., train_step()) - - def test_worker_devices_on_subset_cores(self): - resolver = get_tpu_cluster_resolver() - remote.connect_to_cluster(resolver) - topology = tpu_strategy_util.initialize_tpu_system(resolver) - - # Strategy for the 1st core. - device_assignment = device_assignment_lib.DeviceAssignment.build( - topology, num_replicas=1) - first_core_strategy = tpu_lib.TPUStrategy( - resolver, device_assignment=device_assignment) - - # Strategy for the 2nd core. - device_assignment2 = device_assignment_lib.DeviceAssignment( - topology, [[[0, 0, 0, 1]]]) - second_core_strategy = tpu_lib.TPUStrategy( - resolver, device_assignment=device_assignment2) - - self.assertLen(first_core_strategy.extended.worker_devices, 1) - self.assertEndsWith(first_core_strategy.extended.worker_devices[0], - "device:TPU:0") - - self.assertLen(second_core_strategy.extended.worker_devices, 1) - self.assertEndsWith(second_core_strategy.extended.worker_devices[0], - "device:TPU:1") - def test_tpu_tf_function_same_device(self): with ops.device("/device:TPU:0"): a = variables.Variable(1) @@ -288,8 +122,194 @@ class TPUStrategyTest(test.TestCase): result = bar() + 1 self.assertAllEqual(result, 2) - def test_control_output_in_while_body_fn(self): - strategy = get_tpu_strategy() + +@parameterized.named_parameters([("PackedVar", True), ("", False)]) +class TPUStrategyTest(test.TestCase, parameterized.TestCase): + + def test_sequential_experimental_runs(self, enable_packed_var): + resolver = get_tpu_cluster_resolver() + remote.connect_to_cluster(resolver) + topology = tpu_strategy_util.initialize_tpu_system(resolver) + # Computation replicated to all cores. + device_assignment = device_assignment_lib.DeviceAssignment.build( + topology, num_replicas=2) + strategy = tpu_lib.TPUStrategy( + resolver, device_assignment=device_assignment) + strategy._enable_packed_variable_in_eager_mode = enable_packed_var + + # Computation on the 1st core. + device_assignment2 = device_assignment_lib.DeviceAssignment.build( + topology, num_replicas=1) + strategy2 = tpu_lib.TPUStrategy( + resolver, device_assignment=device_assignment2) + + def computation(x): + return math_ops.square(x) + + @def_function.function + def train_step(): + outputs = strategy.experimental_local_results( + strategy.run(computation, args=([2., 2.],))) + outputs2 = strategy2.run( + computation, args=([outputs[0]],)) + return outputs2 + + self.assertAllEqual([[16., 16.]], train_step()) + + def test_device_switch_case(self, enable_packed_var): + strategy = get_tpu_strategy(enable_packed_var) + with strategy.scope(): + a = variables.Variable(1) + + inference_iteration = variables.Variable(-1) + + def inference_fn(x, i): + return a + x + i + + @def_function.function + def run_inference(x): + + def do_inference(device, inference_fn, i): + with ops.device(device): + return inference_fn(x, i) + + branch_fns = { + 0: (lambda: do_inference("/device:TPU:0", inference_fn, 0)), + 1: (lambda: do_inference("/device:TPU:1", inference_fn, 1)), + } + branch_index = inference_iteration.assign_add(1, use_locking=True) % 2 + return control_flow_ops.switch_case(branch_index, branch_fns) + + self.assertAllEqual(2., run_inference(1)) # Use TPU core 0. + self.assertAllEqual(3., run_inference(1)) # Use TPU core 1. + + def test_recover_from_compilation_failures(self, enable_packed_var): + # TODO(b/148150981): Stop skipping this test once recovery works + # for non-local TPU. + if FLAGS.tpu: + self.skipTest("Recovery fails for non-local TPU, see b/148150981") + + # Disable automatic outside compilation. + config.set_soft_device_placement(False) + strategy = get_tpu_strategy(enable_packed_var) + + @def_function.function + def compilation_failure_run(): + + def computation(): + return random_ops.random_gamma([10], [0.5, 1.5]) + + return strategy.run(computation) + + with self.assertRaisesRegex(errors.InvalidArgumentError, + "TPU compilation failed"): + compilation_failure_run() + + @def_function.function + def good_run(): + + def computation(): + return random_ops.random_normal([10]) + + return strategy.run(computation) + + good_run() + + def test_dynamic_shape_with_outside_compilation_failure( + self, enable_packed_var): + # Enable automatic outside compilation. + config.set_soft_device_placement(True) + strategy = get_tpu_strategy(enable_packed_var) + dataset = dataset_ops.Dataset.from_tensors(("string", 1.0)).repeat().batch( + 2, drop_remainder=False) + dataset = strategy.experimental_distribute_dataset(dataset) + iterator = iter(dataset) + + @def_function.function + def train_fn(iterator): + + def step_fn(inputs): + _, inputs = inputs + return math_ops.reduce_sum(inputs) + + return strategy.experimental_local_results( + strategy.run(step_fn, args=(next(iterator),))) + + with self.assertRaisesRegex(errors.InternalError, "Compilation failure"): + logging.info(train_fn(iterator)) + + def test_computation_on_subset_cores(self, enable_packed_var): + resolver = get_tpu_cluster_resolver() + remote.connect_to_cluster(resolver) + topology = tpu_strategy_util.initialize_tpu_system(resolver) + all_core_strategy = tpu_lib.TPUStrategy(resolver) + all_core_strategy._enable_packed_variable_in_eager_mode = enable_packed_var + + with all_core_strategy.scope(): + v = variables.Variable(0.0, + aggregation=variables.VariableAggregation.MEAN) + + # Computation on the 1st core. + device_assignment = device_assignment_lib.DeviceAssignment.build( + topology, num_replicas=1) + first_core_strategy = tpu_lib.TPUStrategy( + resolver, device_assignment=device_assignment) + first_core_strategy._enable_packed_variable_in_eager_mode = ( + enable_packed_var) + + # Computation on the 2nd core. + device_assignment2 = device_assignment_lib.DeviceAssignment( + topology, [[[0, 0, 0, 1]]]) + second_core_strategy = tpu_lib.TPUStrategy( + resolver, device_assignment=device_assignment2) + second_core_strategy._enable_packed_variable_in_eager_mode = ( + enable_packed_var) + + @def_function.function + def train_step(): + + def step_fn(): + return v + 1.0 + + all_core_strategy.run(step_fn) + r1 = first_core_strategy.run(step_fn) + r2 = second_core_strategy.run(step_fn) + return r1 + r2 + + train_step() + self.assertAllEqual(2., train_step()) + + def test_worker_devices_on_subset_cores(self, enable_packed_var): + resolver = get_tpu_cluster_resolver() + remote.connect_to_cluster(resolver) + topology = tpu_strategy_util.initialize_tpu_system(resolver) + + # Strategy for the 1st core. + device_assignment = device_assignment_lib.DeviceAssignment.build( + topology, num_replicas=1) + first_core_strategy = tpu_lib.TPUStrategy( + resolver, device_assignment=device_assignment) + first_core_strategy._enable_packed_variable_in_eager_mode = ( + enable_packed_var) + + # Strategy for the 2nd core. + device_assignment2 = device_assignment_lib.DeviceAssignment( + topology, [[[0, 0, 0, 1]]]) + second_core_strategy = tpu_lib.TPUStrategy( + resolver, device_assignment=device_assignment2) + second_core_strategy._enable_packed_variable_in_eager_mode = ( + enable_packed_var) + + self.assertLen(first_core_strategy.extended.worker_devices, 1) + self.assertEndsWith(first_core_strategy.extended.worker_devices[0], + "device:TPU:0") + + self.assertLen(second_core_strategy.extended.worker_devices, 1) + self.assertEndsWith(second_core_strategy.extended.worker_devices[0], + "device:TPU:1") + + def test_control_output_in_while_body_fn(self, enable_packed_var): + strategy = get_tpu_strategy(enable_packed_var) with strategy.scope(): v = variables.Variable( @@ -307,8 +327,8 @@ class TPUStrategyTest(test.TestCase): train_step() self.assertEqual(2.0, v.numpy()) - def test_cluster_in_graph_and_while_body_fn(self): - strategy = get_tpu_strategy() + def test_cluster_in_graph_and_while_body_fn(self, enable_packed_var): + strategy = get_tpu_strategy(enable_packed_var) @def_function.function def train_step(): @@ -328,8 +348,8 @@ class TPUStrategyTest(test.TestCase): sum_val = train_step().numpy().astype(float) self.assertEqual(sum_val, strategy.num_replicas_in_sync * 10) - def test_two_clusters_with_same_fn(self): - strategy = get_tpu_strategy() + def test_two_clusters_with_same_fn(self, enable_packed_var): + strategy = get_tpu_strategy(enable_packed_var) @def_function.function def foo(x): @@ -342,8 +362,8 @@ class TPUStrategyTest(test.TestCase): bar(1) - def test_using_external_variable_inside_tf_function(self): - strategy = get_tpu_strategy() + def test_using_external_variable_inside_tf_function(self, enable_packed_var): + strategy = get_tpu_strategy(enable_packed_var) dataset = dataset_ops.Dataset.range( strategy.num_replicas_in_sync * 2, output_type=dtypes.float32).batch(strategy.num_replicas_in_sync) @@ -364,8 +384,8 @@ class TPUStrategyTest(test.TestCase): strategy.experimental_local_results(train_step(next(input_iterator)))) # TODO(b/145574622): Remove this test once it is re-enabled in values_test.py. - def test_all_reduce_on_sync_on_read_variable(self): - strategy = get_tpu_strategy() + def test_all_reduce_on_sync_on_read_variable(self, enable_packed_var): + strategy = get_tpu_strategy(enable_packed_var) dataset = dataset_ops.Dataset.range( strategy.num_replicas_in_sync, output_type=dtypes.float32).batch( strategy.num_replicas_in_sync, drop_remainder=True) @@ -404,8 +424,8 @@ class TPUStrategyTest(test.TestCase): self.assertAllEqual((0.,), w.read_value()) # TODO(b/140633529): Re-enable the test. - def disable_test_experimental_run_output_on_device(self): - strategy = get_tpu_strategy() + def disable_test_experimental_run_output_on_device(self, enable_packed_var): + strategy = get_tpu_strategy(enable_packed_var) def computation(x): return math_ops.square(x) @@ -423,8 +443,8 @@ class TPUStrategyTest(test.TestCase): self.assertAllEqual("/job:localhost/replica:0/task:0/device:TPU:1", results[1].backing_device) - def test_composite_input(self): - strategy = get_tpu_strategy() + def test_composite_input(self, enable_packed_var): + strategy = get_tpu_strategy(enable_packed_var) if strategy.num_replicas_in_sync != 2: self.skipTest("Test assumes two replicas.") @@ -463,8 +483,9 @@ class TPUStrategyTest(test.TestCase): self.assertAllEqual(result, [[[0.0, 1.0], [3.0, 8.0]], [[0.0, 1.0], [3.0, 8.0]]]) - def test_composite_input_dynamic_shapes_outside_compilation(self): - strategy = get_tpu_strategy() + def test_composite_input_dynamic_shapes_outside_compilation( + self, enable_packed_var): + strategy = get_tpu_strategy(enable_packed_var) if strategy.num_replicas_in_sync != 2: self.skipTest("Test assumes two replicas.") @@ -506,11 +527,11 @@ class TPUStrategyTest(test.TestCase): result = sparse_lookup(dataset) self.assertAllEqual(result, [[0.0, 2.0], [1.5, 5.0]]) - def test_per_device_tracing_of_mirrored_variables(self): + def test_per_device_tracing_of_mirrored_variables(self, enable_packed_var): # Define trace_count as a list to avoid python scoping error trace_count = [0] - strategy = get_tpu_strategy() + strategy = get_tpu_strategy(enable_packed_var) with strategy.scope(): variable = variables.Variable(0.0) @@ -527,7 +548,10 @@ class TPUStrategyTest(test.TestCase): with strategy.scope(): update_variable.get_concrete_function() - self.assertEqual(trace_count[0], len(strategy.extended.worker_devices)) + self.assertLen(strategy.extended.worker_devices, trace_count[0]) + + +class TPUStrategyDataPrefetchTest(test.TestCase): def test_prefetch_to_device_default(self): strategy = get_tpu_strategy() diff --git a/tensorflow/python/distribute/tpu_values.py b/tensorflow/python/distribute/tpu_values.py index 40ab058ac7c..33885531966 100644 --- a/tensorflow/python/distribute/tpu_values.py +++ b/tensorflow/python/distribute/tpu_values.py @@ -24,6 +24,7 @@ from __future__ import print_function import contextlib +from tensorflow.python.distribute import packed_distributed_variable as packed from tensorflow.python.distribute import values from tensorflow.python.eager import context from tensorflow.python.eager import tape @@ -46,15 +47,27 @@ def _maybe_enter_graph(tensor): yield +@contextlib.contextmanager +def _maybe_on_device(var): + # Add a device scope for packed variables. + if isinstance(var, packed.PackedVarAndDevice): + with ops.device(var.device): + yield + else: + yield + + def _make_raw_assign_fn(raw_assign_fn): # pylint: disable=missing-docstring def assign_fn(var, value, use_locking=False, name=None, read_value=True): # pylint: disable=missing-docstring del use_locking # Unused. - with _maybe_enter_graph(var.handle): + handle = var.handle + with _maybe_enter_graph(handle), _maybe_on_device(var): op = raw_assign_fn( - var.handle, ops.convert_to_tensor(value, dtype=var.dtype), name=name) - + handle, + ops.convert_to_tensor(value, dtype=var.dtype), + name=name) with ops.control_dependencies([op]): return var._read_variable_op() if read_value else op # pylint: disable=protected-access @@ -97,23 +110,37 @@ class TPUVariableMixin(object): @property def handle(self): + """The handle by which this variable can be accessed.""" # If we're in a tpu.rewrite(), return the replicated handle. tpu_context = enclosing_tpu_context() if tpu_context is None or context.executing_eagerly(): return self._get_on_device_or_primary().handle else: - return tpu_context.get_replicated_var_handle(self._handle_id, - self._values, - self._is_mirrored()) + is_packed = self._packed_var is not None + val = self._values + if is_packed: + val = [self._packed_var] + + return tpu_context.get_replicated_var_handle(self._handle_id, val, + self._is_mirrored(), + is_packed) @property def device(self): return self.handle.device def _read_variable_op(self): + """Reads the value of this variable.""" if self.trainable: tape.variable_accessed(self) - return gen_resource_variable_ops.read_variable_op(self.handle, self.dtype) + + handle = self.handle + if getattr(handle, "is_packed", False): + # Add a device scope for a packed variable handle. + with ops.device(self._get_on_device_or_primary().device): + return gen_resource_variable_ops.read_variable_op(handle, self.dtype) + else: + return gen_resource_variable_ops.read_variable_op(handle, self.dtype) def read_value(self): if enclosing_tpu_context() is None: diff --git a/tensorflow/python/distribute/values.py b/tensorflow/python/distribute/values.py index 60b2ea4fe31..37643e03b18 100644 --- a/tensorflow/python/distribute/values.py +++ b/tensorflow/python/distribute/values.py @@ -472,6 +472,12 @@ class DistributedVariable(DistributedDelegate, variables_lib.Variable, # variable. self._var_policy = var_policy + @property + def _devices(self): + if self._packed_var is not None: + return tuple(d for d in self._packed_var.devices) + return tuple(v.device for v in self._values) + def is_initialized(self, name=None): """Identifies if all the component variables are initialized. @@ -482,6 +488,8 @@ class DistributedVariable(DistributedDelegate, variables_lib.Variable, The op that evaluates to True or False depending on if all the component variables are initialized. """ + if self._packed_var is not None: + return self._packed_var.is_initialized() result = self._primary.is_initialized() # We iterate through the list of values except the last one to allow us to # name the final `logical_and` op the same name that is passed by the user @@ -552,6 +560,10 @@ class DistributedVariable(DistributedDelegate, variables_lib.Variable, def aggregation(self): return self._aggregation + @property + def _packed_variable(self): + return self._packed_var + @property def handle(self): replica_id = values_util.get_current_replica_id_as_int() @@ -559,6 +571,8 @@ class DistributedVariable(DistributedDelegate, variables_lib.Variable, raise ValueError("`handle` is not available outside the replica context" " or a `tf.distribute.Strategy.update()` call.") else: + if self._packed_var is not None: + return self._packed_var.handle return self._values[replica_id].handle def eval(self, session=None): @@ -607,6 +621,33 @@ class DistributedVariable(DistributedDelegate, variables_lib.Variable, def _in_graph_mode(self): return self._primary._in_graph_mode # pylint: disable=protected-access + def _get_replica(self, replica_id): + """Returns the value on a device with the given replica_id.""" + if self._packed_var is not None: + return self._packed_var.on_device(self._devices[replica_id]) + return self._values[replica_id] + + def _get(self): + """Returns the value for the current device or raises a ValueError.""" + replica_id = values_util.get_current_replica_id_as_int() + if replica_id is None: + return self._get_cross_replica() + else: + return self._get_replica(replica_id) + + def _get_on_device_or_primary(self): + """Returns value in same replica or device if possible, else the _primary.""" + replica_id = values_util.get_current_replica_id_as_int() + if replica_id is None: + # Try to find a value on the current device. + current_device = device_util.canonicalize(device_util.current()) + for i, value in enumerate(self._values): + if device_util.canonicalize(value.device) == current_device: + return self._get_replica(i) + return self._get_replica(0) + else: + return self._get_replica(replica_id) + def read_value(self): with ds_context.enter_or_assert_strategy(self._distribute_strategy): return array_ops.identity(self._get()) @@ -778,7 +819,8 @@ class DistributedVariable(DistributedDelegate, variables_lib.Variable, if ds_context.in_cross_replica_context(): update_replica_id = distribute_lib.get_update_replica_id() if update_replica_id is not None: - return update_fn(self._values[update_replica_id], value, **kwargs) + replica_value = self._get_replica(update_replica_id) + return update_fn(replica_value, value, **kwargs) return self._update_cross_replica(update_fn, value, **kwargs) else: values_util.assert_replica_context(self.distribute_strategy) @@ -802,6 +844,7 @@ class DistributedVariable(DistributedDelegate, variables_lib.Variable, obj_map[v] = new_obj resource_map[v.handle] = new_obj.handle obj_map[self] = new_obj + resource_map[self.handle] = new_obj.handle resource_map[self] = new_obj.handle return obj_map, resource_map @@ -835,6 +878,12 @@ class _MirroredSaveable(saveable_object_util.ResourceVariableSaveable): def restore(self, restored_tensors, restored_shapes): """Restore the same value into all variables.""" tensor, = restored_tensors + packed_var = self._mirrored_variable._packed_variable # pylint: disable=protected-access + if packed_var is not None: + return control_flow_ops.group( + tuple( + values_util.assign_on_device(d, packed_var, tensor) + for d in packed_var.devices)) return control_flow_ops.group( tuple( values_util.assign_on_device(v.device, v, tensor) @@ -1013,7 +1062,7 @@ class SyncOnReadVariable(DistributedVariable): def _get_cross_replica(self): if self._aggregation == vs.VariableAggregation.ONLY_FIRST_REPLICA: - return self._primary + return self._get_replica(0) with ds_context.enter_or_assert_strategy(self._distribute_strategy): return self._distribute_strategy.reduce( diff --git a/tensorflow/python/distribute/values_test.py b/tensorflow/python/distribute/values_test.py index 0cb4d6ddd2a..d0e3eec22a8 100644 --- a/tensorflow/python/distribute/values_test.py +++ b/tensorflow/python/distribute/values_test.py @@ -42,7 +42,6 @@ from tensorflow.python.eager import context from tensorflow.python.eager import def_function from tensorflow.python.eager import test from tensorflow.python.framework import constant_op -from tensorflow.python.framework import device from tensorflow.python.framework import dtypes from tensorflow.python.framework import indexed_slices from tensorflow.python.framework import ops @@ -234,11 +233,11 @@ class DistributedValuesTest(test.TestCase, parameterized.TestCase): distribution=[ strategy_combinations.mirrored_strategy_with_gpu_and_cpu, strategy_combinations.tpu_strategy, + strategy_combinations.tpu_strategy_packed_var, # TODO(b/137795644): support CentralStroageStrategy # strategy_combinations.central_storage_strategy_with_two_gpus, ], - mode=["eager"] - )) + mode=["eager"])) def testMakeDistributedValueDefaultDevicePlacement(self, distribution): if not tf2.enabled(): self.skipTest("Only V2 is supported.") @@ -259,11 +258,11 @@ class DistributedValuesTest(test.TestCase, parameterized.TestCase): distribution=[ strategy_combinations.mirrored_strategy_with_gpu_and_cpu, strategy_combinations.tpu_strategy, + strategy_combinations.tpu_strategy_packed_var, # TODO(b/137795644): support CentralStroageStrategy # strategy_combinations.central_storage_strategy_with_two_gpus, ], - mode=["eager"] - )) + mode=["eager"])) def testMakeDistributedValueExplicitDevicePlacement(self, distribution): if not tf2.enabled(): self.skipTest("Only V2 is supported.") @@ -384,6 +383,16 @@ def _make_mirrored(): return mirrored +def mirrored_and_tpu_strategy_combinations(): + return combinations.combine( + distribution=[ + strategy_combinations.mirrored_strategy_with_gpu_and_cpu, + strategy_combinations.tpu_strategy, + strategy_combinations.tpu_strategy_packed_var, + ], + mode=["graph", "eager"]) + + class RegroupAndSelectDeviceTest(test.TestCase, parameterized.TestCase): def _is_per_replica(self, result, expected, klass=values.PerReplica): @@ -563,6 +572,7 @@ class RegroupAndSelectDeviceTest(test.TestCase, parameterized.TestCase): strategy_combinations.mirrored_strategy_with_one_cpu, strategy_combinations.mirrored_strategy_with_gpu_and_cpu, strategy_combinations.tpu_strategy, + strategy_combinations.tpu_strategy_packed_var, strategy_combinations.central_storage_strategy_with_two_gpus, ], synchronization=[ @@ -708,29 +718,40 @@ class DistributedVariableTest(test.TestCase, parameterized.TestCase): self.evaluate( distribution.experimental_local_results(distribution.run(assign))) - def testPackedVariable(self, distribution, synchronization, aggregation): + +@combinations.generate( + combinations.combine( + distribution=[ + strategy_combinations.mirrored_strategy_with_one_cpu, + strategy_combinations.tpu_strategy, + ], + mode=["eager"])) +class PackedDistributedVariableTest(test.TestCase, parameterized.TestCase): + + def testPackedVariable(self, distribution): with distribution.scope(): - v0 = variables_lib.Variable( - 0., synchronization=synchronization, aggregation=aggregation) - if not isinstance(v0, values.DistributedVariable): - self.skipTest("This test doesn't apply to non DistributedVariables") - - self.assertEqual(v0._packed_var, None) - - device_type = device.DeviceSpec.from_string(v0._devices[0]).device_type - for d in v0._devices: - if device.DeviceSpec.from_string(d).device_type != device_type: - self.skipTest("Packing variables on devices of different types " - "is not supported yet.") + v0 = variables_lib.Variable(0.) + self.assertIsNone(v0._packed_var) distribution._enable_packed_variable_in_eager_mode = True with distribution.scope(): - v1 = variables_lib.Variable( - 0., synchronization=synchronization, aggregation=aggregation) - if ops.executing_eagerly_outside_functions(): + v1 = variables_lib.Variable(0) self.assertIsInstance(v1._packed_var, packed.PackedDistributedVariable) - else: - self.assertEqual(v1._packed_var, None) + + devices = v1._devices + for i in range(1, len(devices)): + with distribute_lib.ReplicaContext(distribution, i): + v1.assign(i) + val = v1._get() + self.assertIsInstance(val, packed.PackedVarAndDevice) + self.assertEqual(val.device, devices[0]) + self.assertEqual(self.evaluate(val.read_value()), 0) + for i in range(0, len(devices)): + with distribute_lib.ReplicaContext(distribution, i): + val = v1._get() + self.assertIsInstance(val, packed.PackedVarAndDevice) + self.assertEqual(val.device, devices[i]) + self.assertEqual(self.evaluate(val.read_value()), i) class MirroredVariableTest(test.TestCase, parameterized.TestCase): @@ -920,6 +941,7 @@ class MirroredVariableTest(test.TestCase, parameterized.TestCase): distribution=[ strategy_combinations.mirrored_strategy_with_gpu_and_cpu, strategy_combinations.tpu_strategy, + strategy_combinations.tpu_strategy_packed_var, ], mode=["eager"])) def testAssignValueInReplicaContextWithoutAggregation(self, distribution): @@ -943,6 +965,7 @@ class MirroredVariableTest(test.TestCase, parameterized.TestCase): strategy_combinations.mirrored_strategy_with_one_cpu, strategy_combinations.mirrored_strategy_with_gpu_and_cpu, strategy_combinations.tpu_strategy, + strategy_combinations.tpu_strategy_packed_var, ], mode=["graph", "eager"])) def testValueInReplicaContext(self, distribution): @@ -968,6 +991,7 @@ class MirroredVariableTest(test.TestCase, parameterized.TestCase): strategy_combinations.mirrored_strategy_with_one_cpu, strategy_combinations.mirrored_strategy_with_gpu_and_cpu, strategy_combinations.tpu_strategy, + strategy_combinations.tpu_strategy_packed_var, ], mode=["graph", "eager"])) def testAssignOutOfScope(self, distribution): @@ -1041,6 +1065,7 @@ class MirroredVariableTest(test.TestCase, parameterized.TestCase): distribution=[ strategy_combinations.mirrored_strategy_with_gpu_and_cpu, strategy_combinations.tpu_strategy, + strategy_combinations.tpu_strategy_packed_var, ], mode=["eager"])) def testInitializedToSameValueInsideEagerRun(self, distribution): @@ -1066,6 +1091,7 @@ class MirroredVariableTest(test.TestCase, parameterized.TestCase): strategy_combinations.mirrored_strategy_with_one_cpu, strategy_combinations.mirrored_strategy_with_gpu_and_cpu, strategy_combinations.tpu_strategy, + strategy_combinations.tpu_strategy_packed_var, ], mode=["graph", "eager"])) def testAggregationOnlyFirstReplica(self, distribution): @@ -1093,6 +1119,7 @@ class MirroredVariableTest(test.TestCase, parameterized.TestCase): distribution=[ strategy_combinations.mirrored_strategy_with_gpu_and_cpu, strategy_combinations.tpu_strategy, + strategy_combinations.tpu_strategy_packed_var, ], mode=["eager"])) def testInitScope(self, distribution): @@ -1143,13 +1170,7 @@ class MirroredVariableTest(test.TestCase, parameterized.TestCase): distribution.experimental_local_results(distribution.run(add))) self.assertAllEqual([2, 2], per_replica_results) - @combinations.generate( - combinations.combine( - distribution=[ - strategy_combinations.mirrored_strategy_with_gpu_and_cpu, - strategy_combinations.tpu_strategy, - ], - mode=["graph", "eager"])) + @combinations.generate(mirrored_and_tpu_strategy_combinations()) def testAssignAdd(self, distribution): with distribution.scope(): v = variable_scope.variable( @@ -1456,15 +1477,6 @@ class SyncOnReadVariablePropertiesTest(test.TestCase): self.assertEqual(2., self.evaluate(add1(replica_local))) -def mirrored_and_tpu_strategy_combinations(): - return combinations.combine( - distribution=[ - strategy_combinations.mirrored_strategy_with_gpu_and_cpu, - strategy_combinations.tpu_strategy, - ], - mode=["graph", "eager"]) - - # TODO(b/144432582): Add variable aggregation type to combinations to simplify # tests. def strategy_and_run_tf_function_combinations(): @@ -1478,6 +1490,7 @@ def strategy_and_run_tf_function_combinations(): experimental_run_tf_function=[True, False]) + combinations.combine( distribution=[ strategy_combinations.tpu_strategy, + strategy_combinations.tpu_strategy_packed_var, ], mode=["graph", "eager"], experimental_run_tf_function=[True]) diff --git a/tensorflow/python/distribute/values_util.py b/tensorflow/python/distribute/values_util.py index ddb0d2d0401..5909bdd229e 100644 --- a/tensorflow/python/distribute/values_util.py +++ b/tensorflow/python/distribute/values_util.py @@ -61,8 +61,14 @@ def on_write_assign_sub(var, value, use_locking=False, name=None, def assign_on_each_device(var, assign_func, value, read_value): - update = control_flow_ops.group( - tuple(assign_func(v.device, v, value) for v in var._values)) # pylint: disable=protected-access + """Update the variable on each replica with the given assign_func and value.""" + if var._packed_variable is not None: # pylint: disable=protected-access + update = control_flow_ops.group( + tuple( + assign_func(d, var._packed_variable, value) for d in var._devices)) # pylint: disable=protected-access + else: + update = control_flow_ops.group( + tuple(assign_func(v.device, v, value) for v in var._values)) # pylint: disable=protected-access if not read_value: return update with ops.control_dependencies([update] if update else []): @@ -104,7 +110,7 @@ def on_read_assign_cross_replica(var, value, read_value=True): # TODO(anjs): Should this be over all the replicas in sync since we # call `reduce` on the variable during read? if var.aggregation == vs.VariableAggregation.SUM: - tensor = math_ops.cast(tensor / len(var._values), var.dtype) # pylint: disable=protected-access + tensor = math_ops.cast(tensor / len(var._devices), var.dtype) # pylint: disable=protected-access return assign_on_each_device(var, assign_on_device, tensor, read_value) diff --git a/tensorflow/python/tpu/tpu.py b/tensorflow/python/tpu/tpu.py index ce3aaa8a058..6f5f0bc26c2 100644 --- a/tensorflow/python/tpu/tpu.py +++ b/tensorflow/python/tpu/tpu.py @@ -298,7 +298,8 @@ class TPUReplicateContext(control_flow_ops.XLAControlFlowContext): self._pivot = pivot self._replicated_vars = {} - def get_replicated_var_handle(self, name, vars_, is_mirrored=False): + def get_replicated_var_handle(self, name, vars_, is_mirrored=False, + is_packed=False): """Returns a variable handle for replicated TPU variable 'var'. This is a method used by an experimental replicated variable implementation @@ -309,6 +310,7 @@ class TPUReplicateContext(control_flow_ops.XLAControlFlowContext): vars_: The replicated TPU variables. is_mirrored: Whether the variables are mirrored, which guarantees the values in each replica are always the same. + is_packed: Whether the replicated variables are packed into one variable. Returns: The handle of the TPU replicated input node. @@ -320,7 +322,7 @@ class TPUReplicateContext(control_flow_ops.XLAControlFlowContext): if handle is not None: return handle - if device_assignment is not None: + if device_assignment is not None and not is_packed: # Find a variable copy for each replica in the device assignment. # Note that the order of devices for replicas for the variable and the # device assignment might not match. @@ -356,7 +358,8 @@ class TPUReplicateContext(control_flow_ops.XLAControlFlowContext): graph._set_control_flow_context(self.outer_context) handle = tpu_ops.tpu_replicated_input([v.handle for v in replicated_vars], name=name + "/handle", - is_mirrored_variable=is_mirrored) + is_mirrored_variable=is_mirrored, + is_packed=is_packed) graph._set_control_flow_context(saved_context) # pylint: enable=protected-access self._replicated_vars[name] = handle From 62082d40720b56974436dd17625c247a5fce2a6b Mon Sep 17 00:00:00 2001 From: YoungSeok Yoon Date: Thu, 18 Jun 2020 20:16:19 -0700 Subject: [PATCH 267/867] Add build flags for objc libraries PiperOrigin-RevId: 317235962 Change-Id: I976ccd1ce3db49be3acac44f60b2dc44ed25d767 --- tensorflow/lite/experimental/objc/BUILD.apple | 7 +++++-- 1 file changed, 5 insertions(+), 2 deletions(-) diff --git a/tensorflow/lite/experimental/objc/BUILD.apple b/tensorflow/lite/experimental/objc/BUILD.apple index ff7e8fa58e9..09d4547813a 100644 --- a/tensorflow/lite/experimental/objc/BUILD.apple +++ b/tensorflow/lite/experimental/objc/BUILD.apple @@ -97,7 +97,7 @@ objc_library( "//tensorflow/lite:testdata/add.bin", "//tensorflow/lite:testdata/add_quantized.bin", ], - tags = TFL_DEFAULT_TAGS, + tags = TFL_DEFAULT_TAGS + ["builder_default_ios_x86_64"], deps = [ ":TensorFlowLite", ], @@ -135,7 +135,10 @@ objc_library( "apis", ], module_name = "TestApp", - tags = TFL_DEFAULT_TAGS + ["manual"], + tags = TFL_DEFAULT_TAGS + [ + "manual", + "builder_default_ios_x86_64", + ], deps = [ ":TensorFlowLite", ], From 85ad8031f60536361de71dd689c9d88848fefed6 Mon Sep 17 00:00:00 2001 From: Gaurav Jain Date: Thu, 18 Jun 2020 20:27:03 -0700 Subject: [PATCH 268/867] Expand dtype support for Neg PiperOrigin-RevId: 317237033 Change-Id: I59c5e45d469f7bf704976b66bc122aaac3982b5e --- .../mlir/tensorflow/ir/tf_generated_ops.td | 4 +-- tensorflow/core/kernels/BUILD | 3 ++- .../core/kernels/cwise_op_gpu_neg.cu.cc | 4 +-- .../{cwise_op_neg.cc => cwise_op_neg_1.cc} | 6 ++--- tensorflow/core/kernels/cwise_op_neg_2.cc | 26 +++++++++++++++++++ tensorflow/core/ops/math_ops.cc | 12 ++++----- .../kernel_tests/cwise_ops_unary_test.py | 6 +++++ 7 files changed, 46 insertions(+), 15 deletions(-) rename tensorflow/core/kernels/{cwise_op_neg.cc => cwise_op_neg_1.cc} (87%) create mode 100644 tensorflow/core/kernels/cwise_op_neg_2.cc diff --git a/tensorflow/compiler/mlir/tensorflow/ir/tf_generated_ops.td b/tensorflow/compiler/mlir/tensorflow/ir/tf_generated_ops.td index dcd083fc398..3b1f3eec699 100644 --- a/tensorflow/compiler/mlir/tensorflow/ir/tf_generated_ops.td +++ b/tensorflow/compiler/mlir/tensorflow/ir/tf_generated_ops.td @@ -6059,11 +6059,11 @@ I.e., \\(y = -x\\). }]; let arguments = (ins - TensorOf<[BF16, F16, F32, F64, I32, I64, TF_Complex128, TF_Complex64]>:$x + TensorOf<[BF16, F16, F32, F64, I16, I32, I64, I8, TF_Complex128, TF_Complex64]>:$x ); let results = (outs - TensorOf<[BF16, F16, F32, F64, I32, I64, TF_Complex128, TF_Complex64]>:$y + TensorOf<[BF16, F16, F32, F64, I16, I32, I64, I8, TF_Complex128, TF_Complex64]>:$y ); TF_DerivedOperandTypeAttr T = TF_DerivedOperandTypeAttr<0>; diff --git a/tensorflow/core/kernels/BUILD b/tensorflow/core/kernels/BUILD index ffe2a035591..279dff92c58 100644 --- a/tensorflow/core/kernels/BUILD +++ b/tensorflow/core/kernels/BUILD @@ -6802,7 +6802,8 @@ filegroup( "cwise_op_minimum.cc", "cwise_op_mul_1.cc", "cwise_op_mul_2.cc", - "cwise_op_neg.cc", + "cwise_op_neg_1.cc", + "cwise_op_neg_2.cc", "cwise_op_pow.cc", "cwise_op_real.cc", "cwise_op_reciprocal.cc", diff --git a/tensorflow/core/kernels/cwise_op_gpu_neg.cu.cc b/tensorflow/core/kernels/cwise_op_gpu_neg.cu.cc index ea1ca623560..4f7bb9b2075 100644 --- a/tensorflow/core/kernels/cwise_op_gpu_neg.cu.cc +++ b/tensorflow/core/kernels/cwise_op_gpu_neg.cu.cc @@ -19,8 +19,8 @@ limitations under the License. namespace tensorflow { namespace functor { -DEFINE_UNARY7(neg, Eigen::half, float, double, int32, int64, complex64, - complex128); +DEFINE_UNARY4(neg, int8, int16, int32, int64); +DEFINE_UNARY6(neg, Eigen::half, float, double, bfloat16, complex64, complex128); } // namespace functor } // namespace tensorflow diff --git a/tensorflow/core/kernels/cwise_op_neg.cc b/tensorflow/core/kernels/cwise_op_neg_1.cc similarity index 87% rename from tensorflow/core/kernels/cwise_op_neg.cc rename to tensorflow/core/kernels/cwise_op_neg_1.cc index f52cf6c8b91..18a7c61be90 100644 --- a/tensorflow/core/kernels/cwise_op_neg.cc +++ b/tensorflow/core/kernels/cwise_op_neg_1.cc @@ -16,8 +16,7 @@ limitations under the License. #include "tensorflow/core/kernels/cwise_ops_common.h" namespace tensorflow { -REGISTER8(UnaryOp, CPU, "Neg", functor::neg, float, Eigen::half, double, int32, - complex64, int64, complex128, bfloat16); +REGISTER4(UnaryOp, CPU, "Neg", functor::neg, int8, int16, int32, int64); #ifdef TENSORFLOW_USE_SYCL REGISTER3(UnaryOp, SYCL, "Neg", functor::neg, float, double, int64); @@ -30,8 +29,7 @@ REGISTER_KERNEL_BUILDER(Name("Neg") #endif // TENSORFLOW_USE_SYCL #if GOOGLE_CUDA || TENSORFLOW_USE_ROCM -REGISTER6(UnaryOp, GPU, "Neg", functor::neg, float, Eigen::half, double, int64, - complex64, complex128); +REGISTER3(UnaryOp, GPU, "Neg", functor::neg, int8, int16, int64); // A special GPU kernel for int32. // TODO(b/25387198): Also enable int32 in device memory. This kernel diff --git a/tensorflow/core/kernels/cwise_op_neg_2.cc b/tensorflow/core/kernels/cwise_op_neg_2.cc new file mode 100644 index 00000000000..5ea78ad665c --- /dev/null +++ b/tensorflow/core/kernels/cwise_op_neg_2.cc @@ -0,0 +1,26 @@ +/* Copyright 2020 The TensorFlow Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. +==============================================================================*/ + +#include "tensorflow/core/kernels/cwise_ops_common.h" + +namespace tensorflow { +REGISTER6(UnaryOp, CPU, "Neg", functor::neg, Eigen::half, float, double, + bfloat16, complex64, complex128); + +#if GOOGLE_CUDA || TENSORFLOW_USE_ROCM +REGISTER6(UnaryOp, GPU, "Neg", functor::neg, Eigen::half, float, double, + bfloat16, complex64, complex128); +#endif +} // namespace tensorflow diff --git a/tensorflow/core/ops/math_ops.cc b/tensorflow/core/ops/math_ops.cc index b81bb9d3afc..2a70f420260 100644 --- a/tensorflow/core/ops/math_ops.cc +++ b/tensorflow/core/ops/math_ops.cc @@ -201,12 +201,12 @@ REGISTER_OP("ComplexAbs") .SetShapeFn(shape_inference::UnchangedShape); // Declares cwise unary operations signature: 't -> 't -#define UNARY() \ - Input("x: T") \ - .Output("y: T") \ - .Attr( \ - "T: {bfloat16, half, float, double, int32, int64, complex64, " \ - "complex128}") \ +#define UNARY() \ + Input("x: T") \ + .Output("y: T") \ + .Attr( \ + "T: {bfloat16, half, float, double, int8, int16, int32, int64, " \ + "complex64, complex128}") \ .SetShapeFn(shape_inference::UnchangedShape) #define UNARY_REAL() \ diff --git a/tensorflow/python/kernel_tests/cwise_ops_unary_test.py b/tensorflow/python/kernel_tests/cwise_ops_unary_test.py index f4beaabc29a..df848a653d4 100644 --- a/tensorflow/python/kernel_tests/cwise_ops_unary_test.py +++ b/tensorflow/python/kernel_tests/cwise_ops_unary_test.py @@ -389,16 +389,22 @@ class UnaryOpTest(test.TestCase): 2).reshape(1, 3, 2).astype(dtypes_lib.bfloat16.as_numpy_dtype) self._compareCpu(x, np.abs, math_ops.abs) self._compareCpu(x, np.abs, _ABS) + self._compareBoth(x, np.negative, math_ops.negative) + self._compareBoth(x, np.negative, _NEG) def testInt8Basic(self): x = np.arange(-6, 6, 2).reshape(1, 3, 2).astype(np.int8) self._compareCpu(x, np.abs, math_ops.abs) self._compareCpu(x, np.abs, _ABS) + self._compareBoth(x, np.negative, math_ops.negative) + self._compareBoth(x, np.negative, _NEG) def testInt16Basic(self): x = np.arange(-6, 6, 2).reshape(1, 3, 2).astype(np.int16) self._compareCpu(x, np.abs, math_ops.abs) self._compareCpu(x, np.abs, _ABS) + self._compareBoth(x, np.negative, math_ops.negative) + self._compareBoth(x, np.negative, _NEG) def testInt32Basic(self): x = np.arange(-6, 6, 2).reshape(1, 3, 2).astype(np.int32) From 2a05589bd4f5e3042d1baf539e564d7ab9bd6287 Mon Sep 17 00:00:00 2001 From: Taehee Jeong Date: Thu, 18 Jun 2020 20:48:49 -0700 Subject: [PATCH 269/867] Add inference instruction for iOS PiperOrigin-RevId: 317239235 Change-Id: I55bd7e43bc286f34024ccfc27db61d28304a651d --- tensorflow/lite/g3doc/guide/inference.md | 425 +++++++++++++++-------- 1 file changed, 287 insertions(+), 138 deletions(-) diff --git a/tensorflow/lite/g3doc/guide/inference.md b/tensorflow/lite/g3doc/guide/inference.md index 5f3fba98cff..6e47d6d5190 100644 --- a/tensorflow/lite/g3doc/guide/inference.md +++ b/tensorflow/lite/g3doc/guide/inference.md @@ -7,9 +7,9 @@ inference with a TensorFlow Lite model, you must run it through an The interpreter uses a static graph ordering and a custom (less-dynamic) memory allocator to ensure minimal load, initialization, and execution latency. -This page describes how to access to the TensorFlow Lite interpreter and -perform an inference using C++, Java, and Python, plus links to other resources -for each [supported platform](#supported-platforms). +This page describes how to access to the TensorFlow Lite interpreter and perform +an inference using C++, Java, and Python, plus links to other resources for each +[supported platform](#supported-platforms). [TOC] @@ -17,31 +17,31 @@ for each [supported platform](#supported-platforms). TensorFlow Lite inference typically follows the following steps: -1. **Loading a model** +1. **Loading a model** - You must load the `.tflite` model into memory, which contains the model's - execution graph. + You must load the `.tflite` model into memory, which contains the model's + execution graph. -1. **Transforming data** +1. **Transforming data** - Raw input data for the model generally does not match the input data format - expected by the model. For example, you might need to resize an image or - change the image format to be compatible with the model. + Raw input data for the model generally does not match the input data format + expected by the model. For example, you might need to resize an image or + change the image format to be compatible with the model. -1. **Running inference** +1. **Running inference** - This step involves using the TensorFlow Lite API to execute the model. It - involves a few steps such as building the interpreter, and allocating - tensors, as described in the following sections. + This step involves using the TensorFlow Lite API to execute the model. It + involves a few steps such as building the interpreter, and allocating + tensors, as described in the following sections. -1. **Interpreting output** +1. **Interpreting output** - When you receive results from the model inference, you must interpret the - tensors in a meaningful way that's useful in your application. + When you receive results from the model inference, you must interpret the + tensors in a meaningful way that's useful in your application. - For example, a model might return only a list of probabilities. It's up to - you to map the probabilities to relevant categories and present it to your - end-user. + For example, a model might return only a list of probabilities. It's up to + you to map the probabilities to relevant categories and present it to your + end-user. ## Supported platforms @@ -54,8 +54,8 @@ should be no surprise that the APIs try to avoid unnecessary copies at the expense of convenience. Similarly, consistency with TensorFlow APIs was not an explicit goal and some variance between languages is to be expected. -Across all libraries, the TensorFlow Lite API enables you to load models, -feed inputs, and retrieve inference outputs. +Across all libraries, the TensorFlow Lite API enables you to load models, feed +inputs, and retrieve inference outputs. ### Android @@ -64,8 +64,8 @@ APIs. The Java APIs provide convenience and can be used directly within your Android Activity classes. The C++ APIs offer more flexibility and speed, but may require writing JNI wrappers to move data between Java and C++ layers. -See below for details about using C++ and Java, or -follow the [Android quickstart](android.md) for a tutorial and example code. +See below for details about using C++ and Java, or follow the +[Android quickstart](android.md) for a tutorial and example code. #### TensorFlow Lite Android wrapper code generator @@ -86,103 +86,36 @@ On iOS, TensorFlow Lite is available with native iOS libraries written in [Swift](https://www.tensorflow.org/code/tensorflow/lite/experimental/swift) and [Objective-C](https://www.tensorflow.org/code/tensorflow/lite/experimental/objc). +You can also use +[C API](https://www.tensorflow.org/code/tensorflow/lite/c/c_api.h) +directly in Objective-C codes. -This page doesn't include a discussion for about these languages, so you should -refer to the [iOS quickstart](ios.md) for a tutorial and example code. +See below for details about using Swift, Objective-C and C API, or follow the +[iOS quickstart](ios.md) for a tutorial and example code. ### Linux On Linux platforms (including [Raspberry Pi](build_rpi.md)), you can run -inferences using TensorFlow Lite APIs available in C++ and Python, as shown -in the following sections. +inferences using TensorFlow Lite APIs available in C++ and Python, as shown in +the following sections. +## Running a model -## Load and run a model in C++ +Running a TensorFlow Lite model involves a few simple steps: -Running a TensorFlow Lite model with C++ involves a few simple steps: - - 1. Load the model into memory as a `FlatBufferModel`. - 2. Build an `Interpreter` based on an existing `FlatBufferModel`. - 3. Set input tensor values. (Optionally resize input tensors if the - predefined sizes are not desired.) - 4. Invoke inference. - 5. Read output tensor values. - -The [`FlatBufferModel`]( -https://www.tensorflow.org/lite/api_docs/cc/class/tflite/flat-buffer-model.html) -class encapsulates a TensorFlow Lite model and you can -build it in a couple of different ways, depending on where the model is stored: - -```c++ -class FlatBufferModel { -  // Build a model based on a file. Return a nullptr in case of failure. -  static std::unique_ptr BuildFromFile( -      const char* filename, -      ErrorReporter* error_reporter); - -  // Build a model based on a pre-loaded flatbuffer. The caller retains -  // ownership of the buffer and should keep it alive until the returned object -  // is destroyed. Return a nullptr in case of failure. -  static std::unique_ptr BuildFromBuffer( -      const char* buffer, -      size_t buffer_size, -      ErrorReporter* error_reporter); -}; -``` - -Note: If TensorFlow Lite detects the presence of the [Android NNAPI]( -https://developer.android.com/ndk/guides/neuralnetworks), it will -automatically try to use shared memory to store the `FlatBufferModel`. - -Now that you have the model as a `FlatBufferModel` object, you can execute it -with an [`Interpreter`]( -https://www.tensorflow.org/lite/api_docs/cc/class/tflite/interpreter.html). -A single `FlatBufferModel` can be used -simultaneously by more than one `Interpreter`. - -Caution: The `FlatBufferModel` object must remain valid until -all instances of `Interpreter` using it have been destroyed. - -The important parts of the `Interpreter` API are shown in the -code snippet below. It should be noted that: - - * Tensors are represented by integers, in order to avoid string comparisons - (and any fixed dependency on string libraries). - * An interpreter must not be accessed from concurrent threads. - * Memory allocation for input and output tensors must be triggered - by calling `AllocateTensors()` right after resizing tensors. - -The simplest usage of TensorFlow Lite with C++ looks like this: - -```c++ -// Load the model -std::unique_ptr model = - tflite::FlatBufferModel::BuildFromFile(filename); - -// Build the interpreter -tflite::ops::builtin::BuiltinOpResolver resolver; -std::unique_ptr interpreter; -tflite::InterpreterBuilder(*model, resolver)(&interpreter); - -// Resize input tensors, if desired. -interpreter->AllocateTensors(); - -float* input = interpreter->typed_input_tensor(0); -// Fill `input`. - -interpreter->Invoke(); - -float* output = interpreter->typed_output_tensor(0); -``` - -For more example code, see [`minimal.cc`]( -https://github.com/tensorflow/tensorflow/blob/master/tensorflow/lite/examples/minimal/minimal.cc) -and [`label_image.cc`]( -https://github.com/tensorflow/tensorflow/blob/master/tensorflow/lite/examples/label_image/label_image.cc). +1. Load the model into memory. +2. Build an `Interpreter` based on an existing model. +3. Set input tensor values. (Optionally resize input tensors if the predefined + sizes are not desired.) +4. Invoke inference. +5. Read output tensor values. +Following sections describe how these steps can be done in each language. ## Load and run a model in Java +*Platform: Android* + The Java API for running an inference with TensorFlow Lite is primarily designed for use with Android, so it's available as an Android library dependency: `org.tensorflow:tensorflow-lite`. @@ -203,12 +136,12 @@ public Interpreter(@NotNull MappedByteBuffer mappedByteBuffer); ``` In both cases, you must provide a valid TensorFlow Lite model or the API throws -`IllegalArgumentException`. If you use `MappedByteBuffer` to -initialize an `Interpreter`, it must remain unchanged for the whole lifetime -of the `Interpreter`. +`IllegalArgumentException`. If you use `MappedByteBuffer` to initialize an +`Interpreter`, it must remain unchanged for the whole lifetime of the +`Interpreter`. -To then run an inference with the model, simply call `Interpreter.run()`. -For example: +To then run an inference with the model, simply call `Interpreter.run()`. For +example: ```java try (Interpreter interpreter = new Interpreter(file_of_a_tensorflowlite_model)) { @@ -228,9 +161,9 @@ In this case, each entry in `inputs` corresponds to an input tensor and output data. In both cases, the tensor indices should correspond to the values you gave to -the [TensorFlow Lite Converter](../convert/) when you created the model. -Be aware that the order of tensors in `input` must match the -order given to the TensorFlow Lite Converter. +the [TensorFlow Lite Converter](../convert/) when you created the model. Be +aware that the order of tensors in `input` must match the order given to the +TensorFlow Lite Converter. The `Interpreter` class also provides convenient functions for you to get the index of any model input or output using an operation name: @@ -250,8 +183,8 @@ resources must be released after use by: interpreter.close(); ``` -For an example project with Java, see the [Android image classification sample]( -https://github.com/tensorflow/examples/tree/master/lite/examples/image_classification/android). +For an example project with Java, see the +[Android image classification sample](https://github.com/tensorflow/examples/tree/master/lite/examples/image_classification/android). ### Supported data types (in Java) @@ -295,13 +228,231 @@ have dynamic outputs, where the shape of output tensors can vary depending on the input. There's no straightforward way of handling this with the existing Java inference API, but planned extensions will make this possible. +## Load and run a model in Swift + +*Platform: iOS* + +The +[Swift API](https://www.tensorflow.org/code/tensorflow/lite/experimental/swift) +is available in `TensorFlowLiteSwift` Pod from Cocoapods. + +First, you need to import `TensorFlowLite` module. + +```swift +import TensorFlowLite +``` + +```swift +// Getting model path +guard + let modelPath = Bundle.main.path(forResource: "model", ofType: "tflite") +else { + // Error handling... +} + +do { + // Initialize an interpreter with the model. + let interpreter = try Interpreter(modelPath: modelPath) + + // Allocate memory for the model's input `Tensor`s. + try interpreter.allocateTensors() + + let inputData: Data // Should be initialized + + // input data preparation... + + // Copy the input data to the input `Tensor`. + try self.interpreter.copy(inputData, toInputAt: 0) + + // Run inference by invoking the `Interpreter`. + try self.interpreter.invoke() + + // Get the output `Tensor` + let outputTensor = try self.interpreter.output(at: 0) + + // Copy output to `Data` to process the inference results. + let outputSize = outputTensor.shape.dimensions.reduce(1, {x, y in x * y}) + let outputData = + UnsafeMutableBufferPointer.allocate(capacity: outputSize) + outputTensor.data.copyBytes(to: outputData) + + if (error != nil) { /* Error handling... */ } +} catch error { + // Error handling... +} +``` + +## Load and run a model in Objective-C + +*Platform: iOS* + +The +[Objective-C API](https://www.tensorflow.org/code/tensorflow/lite/experimental/objc) +is available in `TensorFlowLiteObjC` Pod from Cocoapods. + +First, you need to import `TensorFlowLite` module. + +```objc +@import TensorFlowLite; +``` + +```objc +NSString *modelPath = [[NSBundle mainBundle] pathForResource:@"model" + ofType:@"tflite"]; +NSError *error; + +// Initialize an interpreter with the model. +TFLInterpreter *interpreter = [[TFLInterpreter alloc] initWithModelPath:modelPath + error:&error]; +if (error != nil) { /* Error handling... */ } + +// Allocate memory for the model's input `TFLTensor`s. +[interpreter allocateTensorsWithError:&error]; +if (error != nil) { /* Error handling... */ } + +NSMutableData *inputData; // Should be initialized +// input data preparation... + +// Copy the input data to the input `TFLTensor`. +[interpreter copyData:inputData toInputTensorAtIndex:0 error:&error]; +if (error != nil) { /* Error handling... */ } + +// Run inference by invoking the `TFLInterpreter`. +[interpreter invokeWithError:&error]; +if (error != nil) { /* Error handling... */ } + +// Get the output `TFLTensor` +TFLTensor *outputTensor = [interpreter outputTensorAtIndex:0 error:&error]; +if (error != nil) { /* Error handling... */ } + +// Copy output to `NSData` to process the inference results. +NSData *outputData = [outputTensor dataWithError:&error]; +if (error != nil) { /* Error handling... */ } +``` + +### Using C API in Objective-C code + +Currently Objective-C API does not support delegates. In order to use delegates +with Objective-C code, you need to directly call underlying +[C API](https://www.tensorflow.org/code/tensorflow/lite/c/c_api.h). + +```c +#include "tensorflow/lite/c/c_api.h" +``` + +```c +TfLiteModel* model = TfLiteModelCreateFromFile([modelPath UTF8String]); +TfLiteInterpreterOptions* options = TfLiteInterpreterOptionsCreate(); + +// Create the interpreter. +TfLiteInterpreter* interpreter = TfLiteInterpreterCreate(model, options); + +// Allocate tensors and populate the input tensor data. +TfLiteInterpreterAllocateTensors(interpreter); +TfLiteTensor* input_tensor = + TfLiteInterpreterGetInputTensor(interpreter, 0); +TfLiteTensorCopyFromBuffer(input_tensor, input.data(), + input.size() * sizeof(float)); + +// Execute inference. +TfLiteInterpreterInvoke(interpreter); + +// Extract the output tensor data. +const TfLiteTensor* output_tensor = +// TfLiteInterpreterGetOutputTensor(interpreter, 0); +TfLiteTensorCopyToBuffer(output_tensor, output.data(), + output.size() * sizeof(float)); + +// Dispose of the model and interpreter objects. +TfLiteInterpreterDelete(interpreter); +TfLiteInterpreterOptionsDelete(options); +TfLiteModelDelete(model); +``` + +## Load and run a model in C++ + +*Platforms: Android and Linux* + +In C++, the model is stored in +[`FlatBufferModel`](https://www.tensorflow.org/lite/api_docs/cc/class/tflite/flat-buffer-model.html) +class. It encapsulates a TensorFlow Lite model and you can build it in a couple +of different ways, depending on where the model is stored: + +```c++ +class FlatBufferModel { +  // Build a model based on a file. Return a nullptr in case of failure. +  static std::unique_ptr BuildFromFile( +      const char* filename, +      ErrorReporter* error_reporter); + +  // Build a model based on a pre-loaded flatbuffer. The caller retains +  // ownership of the buffer and should keep it alive until the returned object +  // is destroyed. Return a nullptr in case of failure. +  static std::unique_ptr BuildFromBuffer( +      const char* buffer, +      size_t buffer_size, +      ErrorReporter* error_reporter); +}; +``` + +Note: If TensorFlow Lite detects the presence of the +[Android NNAPI](https://developer.android.com/ndk/guides/neuralnetworks), it +will automatically try to use shared memory to store the `FlatBufferModel`. + +Now that you have the model as a `FlatBufferModel` object, you can execute it +with an +[`Interpreter`](https://www.tensorflow.org/lite/api_docs/cc/class/tflite/interpreter.html). +A single `FlatBufferModel` can be used simultaneously by more than one +`Interpreter`. + +Caution: The `FlatBufferModel` object must remain valid until all instances of +`Interpreter` using it have been destroyed. + +The important parts of the `Interpreter` API are shown in the code snippet +below. It should be noted that: + +* Tensors are represented by integers, in order to avoid string comparisons + (and any fixed dependency on string libraries). +* An interpreter must not be accessed from concurrent threads. +* Memory allocation for input and output tensors must be triggered by calling + `AllocateTensors()` right after resizing tensors. + +The simplest usage of TensorFlow Lite with C++ looks like this: + +```c++ +// Load the model +std::unique_ptr model = + tflite::FlatBufferModel::BuildFromFile(filename); + +// Build the interpreter +tflite::ops::builtin::BuiltinOpResolver resolver; +std::unique_ptr interpreter; +tflite::InterpreterBuilder(*model, resolver)(&interpreter); + +// Resize input tensors, if desired. +interpreter->AllocateTensors(); + +float* input = interpreter->typed_input_tensor(0); +// Fill `input`. + +interpreter->Invoke(); + +float* output = interpreter->typed_output_tensor(0); +``` + +For more example code, see +[`minimal.cc`](https://github.com/tensorflow/tensorflow/blob/master/tensorflow/lite/examples/minimal/minimal.cc) +and +[`label_image.cc`](https://github.com/tensorflow/tensorflow/blob/master/tensorflow/lite/examples/label_image/label_image.cc). ## Load and run a model in Python -The Python API for running an inference is provided in the `tf.lite` -module. From which, you mostly need only [`tf.lite.Interpreter`]( -https://www.tensorflow.org/api_docs/python/tf/lite/Interpreter) to load -a model and run an inference. +*Platform: Linux* + +The Python API for running an inference is provided in the `tf.lite` module. +From which, you mostly need only +[`tf.lite.Interpreter`](https://www.tensorflow.org/api_docs/python/tf/lite/Interpreter) +to load a model and run an inference. The following example shows how to use the Python interpreter to load a `.tflite` file and run inference with random input data: @@ -358,13 +509,12 @@ interpreter.allocate_tensors() # Continue to get tensors and so forth, as shown above... ``` -For more Python sample code, see [`label_image.py`]( -https://github.com/tensorflow/tensorflow/blob/master/tensorflow/lite/examples/python/label_image.py). +For more Python sample code, see +[`label_image.py`](https://github.com/tensorflow/tensorflow/blob/master/tensorflow/lite/examples/python/label_image.py). Tip: Run `help(tf.lite.Interpreter)` in the Python terminal to get detailed documentation about the interpreter. - ## Write a custom operator All TensorFlow Lite operators (both custom and builtin) are defined using a @@ -379,10 +529,10 @@ typedef struct { } TfLiteRegistration; ``` -Refer to `context.h` for details on `TfLiteContext` and `TfLiteNode`. The -former provides error reporting facilities and access to global objects, -including all the tensors. The latter allows implementations to access their -inputs and outputs. +Refer to `context.h` for details on `TfLiteContext` and `TfLiteNode`. The former +provides error reporting facilities and access to global objects, including all +the tensors. The latter allows implementations to access their inputs and +outputs. When the interpreter loads a model, it calls `init()` once for each node in the graph. A given `init()` will be called more than once if the op is used multiple @@ -403,9 +553,9 @@ implementations can access their state using `node->user_data`. Finally, each time inference runs, the interpreter traverses the graph calling `invoke()`, and here too the state is available as `node->user_data`. -Custom ops can be implemented in exactly the same way as builtin ops, by -defined those four functions and a global registration function that usually -looks like this: +Custom ops can be implemented in exactly the same way as builtin ops, by defined +those four functions and a global registration function that usually looks like +this: ```c++ namespace tflite { @@ -461,8 +611,7 @@ You can optionally register custom ops (before you pass the resolver to the resolver.AddOp("MY_CUSTOM_OP", Register_MY_CUSTOM_OP()); ``` -If the set of builtin ops is deemed to be too large, a new `OpResolver` could -be code-generated based on a given subset of ops, possibly only the ones -contained in a given model. This is the equivalent of TensorFlow's selective -registration (and a simple version of it is available in the `tools` -directory). +If the set of builtin ops is deemed to be too large, a new `OpResolver` could be +code-generated based on a given subset of ops, possibly only the ones contained +in a given model. This is the equivalent of TensorFlow's selective registration +(and a simple version of it is available in the `tools` directory). From 397494a2313aa51fe0b87b4e51d3a2349e4f8ecc Mon Sep 17 00:00:00 2001 From: "A. Unique TensorFlower" Date: Thu, 18 Jun 2020 20:51:42 -0700 Subject: [PATCH 270/867] Hoisting unconditional converts from conditional branch computations. PiperOrigin-RevId: 317239618 Change-Id: If3b16ff4f2bbcf38ee1ca51f5e8b187c58ab8e91 --- .../xla/service/conditional_code_motion.cc | 285 ++++++++++++++++-- .../xla/service/conditional_code_motion.h | 15 +- .../service/conditional_code_motion_test.cc | 101 ++++++- 3 files changed, 369 insertions(+), 32 deletions(-) diff --git a/tensorflow/compiler/xla/service/conditional_code_motion.cc b/tensorflow/compiler/xla/service/conditional_code_motion.cc index eecdcc851e9..6db4c3eb6d4 100644 --- a/tensorflow/compiler/xla/service/conditional_code_motion.cc +++ b/tensorflow/compiler/xla/service/conditional_code_motion.cc @@ -106,7 +106,6 @@ class BranchVisitor { boundaries_.emplace_back(operand, i, inst); continue; } - worklist_.push_back(operand); visited_.insert(operand); } @@ -197,6 +196,7 @@ bool WorthHoisting(HloInstruction* instruction) { case HloOpcode::kMultiply: case HloOpcode::kDivide: case HloOpcode::kTuple: + case HloOpcode::kSqrt: case HloOpcode::kGetTupleElement: return true; default: @@ -206,10 +206,11 @@ bool WorthHoisting(HloInstruction* instruction) { // Compare if the instructions to be visited at each branches are identical. bool InstructionWithinBranchIdentical( - const std::vector& instructions, bool is_layout_senstive) { + const std::vector& instructions, + bool is_layout_sensitive) { // Identical includes the shape of each operands are equal. auto eq_operand = [&](const HloInstruction* a, const HloInstruction* b) { - bool eq_operands = is_layout_senstive + bool eq_operands = is_layout_sensitive ? ShapeUtil::Equal(a->shape(), b->shape()) : ShapeUtil::Compatible(a->shape(), b->shape()); return eq_operands; @@ -233,7 +234,7 @@ bool InstructionWithinBranchIdentical( auto old_channel_id = instruction->channel_id(); instruction->set_channel_id(instructions[0]->channel_id()); bool eq_instructions = instructions[0]->Identical( - *instruction, eq_operand, eq_computations, is_layout_senstive); + *instruction, eq_operand, eq_computations, is_layout_sensitive); instruction->set_channel_id(old_channel_id); return eq_instructions; }); @@ -243,7 +244,7 @@ bool InstructionWithinBranchIdentical( [&](HloInstruction* instruction) { return instructions[0]->Identical( *instruction, eq_operand, eq_computations, - is_layout_senstive); + is_layout_sensitive); }); } @@ -354,12 +355,228 @@ Status RemoveInstructionFromComputation( return Status::OK(); } +// Identify converts to be hoisted/rematerialized out of the branch +// computations. +absl::flat_hash_set FindSpecialConverts(HloInstruction* old_root, + int branch_count, + HloInstruction* conditional, + bool is_layout_sensitive) { + absl::flat_hash_set kspecial_convert; + for (int64 operand_num = 0; operand_num < old_root->operand_count(); + ++operand_num) { + if (old_root->operand(operand_num)->opcode() != HloOpcode::kConvert) { + continue; + } + bool replica = true; + HloInstruction* kspecial_convert_candidate = + old_root->mutable_operand(operand_num); + // Check whether an identical candidate appears in other branches + for (int others = 1; others < branch_count; ++others) { + HloInstruction* others_root = + conditional->branch_computation(others)->root_instruction(); + bool eq_shape = + is_layout_sensitive + ? ShapeUtil::Equal(others_root->operand(operand_num)->shape(), + kspecial_convert_candidate->shape()) + : ShapeUtil::Compatible( + others_root->operand(operand_num)->shape(), + kspecial_convert_candidate->shape()); + if ((others_root->operand(operand_num)->opcode() == + HloOpcode::kConvert) && + eq_shape) { + // Nothing to be done. + } else { + replica = false; + break; + } + } + if (replica) { + kspecial_convert.insert(operand_num); + } + } + return kspecial_convert; +} + +// Restructuring the conditional instruction as follows: +// i.e., %result = conditional() becomes +// x = conditional() +// y.{0..n} = gte(x, {0..n}) +// z = tuple(y.0, y.1, ...y.n) +// Doing so ensures that we can accommodate the possible shape-change of the +// conditional when the instructions are hoisted. +Status RestructureConditionalInstruction(HloComputation* computation, + HloInstruction* conditional) { + HloInstruction* old_root = computation->root_instruction(); + std::vector new_operands; + int cur_index = 0; + for (; cur_index < ShapeUtil::TupleElementCount(conditional->shape()); + ++cur_index) { + new_operands.push_back( + computation->AddInstruction(HloInstruction::CreateGetTupleElement( + ShapeUtil::GetTupleElementShape(conditional->shape(), cur_index), + conditional, cur_index))); + } + HloInstruction* new_tuple = + computation->AddInstruction(HloInstruction::CreateTuple(new_operands)); + if (old_root == conditional) { + computation->set_root_instruction(new_tuple); + } else { + std::vector new_tuple_users; + for (auto conditional_user : conditional->users()) { + auto is_new_gte = absl::c_find_if( + new_operands, + [&](HloInstruction* instr) { return instr == conditional_user; }); + if (is_new_gte == new_operands.end()) { + new_tuple_users.push_back(conditional_user); + } + } + for (auto new_tuple_user : new_tuple_users) { + TF_RETURN_IF_ERROR( + conditional->ReplaceUseWith(new_tuple_user, new_tuple)); + } + } + VLOG(2) << "computation after root restructure:\n" << computation->ToString(); + return Status::OK(); +} + +StatusOr ConvertSpecialMove(HloInstruction* conditional, + bool is_layout_sensitive) { + int branch_count = conditional->branch_count(); + if (branch_count <= 0) { + return false; + } + + HloInstruction* old_root = + conditional->branch_computation(0)->root_instruction(); + if (old_root->opcode() != HloOpcode::kTuple) { + return false; + } else { + VLOG(2) << "BEFORE :" << conditional->parent()->parent()->ToString(); + // Identify the gte using `index'. + auto find_gte = [](const HloInstruction* conditional_result, + int64 index) -> HloInstruction* { + for (HloInstruction* instr : conditional_result->users()) { + if (instr->opcode() != HloOpcode::kGetTupleElement) { + return nullptr; + } + if (instr->tuple_index() == index) { + return instr; + } + } + return nullptr; + }; + + // Captures tuple indices refering to converts to be rematerialized/hoisted. + absl::flat_hash_set kspecial_convert = FindSpecialConverts( + old_root, branch_count, conditional, is_layout_sensitive); + + // Exit if we cannot find any converts to be hoisted. + if (kspecial_convert.empty()) { + return false; + } + + TF_RETURN_IF_ERROR( + RestructureConditionalInstruction(conditional->parent(), conditional)); + + for (int branch = 0; branch < branch_count; branch++) { + old_root = conditional->branch_computation(branch)->root_instruction(); + absl::flat_hash_map map_inst_to_tuple_index; + std::vector new_operands(old_root->operand_count()); + std::unordered_set to_hoist_set; + + for (int64 operand_num = 0; operand_num < old_root->operand_count(); + ++operand_num) { + map_inst_to_tuple_index[old_root->mutable_operand(operand_num)] = + operand_num; + } + for (int64 operand_num = 0; operand_num < old_root->operand_count(); + ++operand_num) { + HloInstruction* hoist = old_root->mutable_operand(operand_num); + if (!kspecial_convert.contains(operand_num)) { + new_operands[operand_num] = old_root->mutable_operand(operand_num); + continue; + } + + to_hoist_set.insert(hoist); + int64 new_tuple_count = old_root->operand_count(); + + // Replace the hoisted instr in the tuple with the operand/operands. + // We will replace at least one of the operands of the hoist at the + // tuple place; the rest will be added at the end. + bool inplace = true; + CHECK(!hoist->operands().empty()); + for (HloInstruction* prod : hoist->operands()) { + if (inplace) { + map_inst_to_tuple_index[prod] = map_inst_to_tuple_index[hoist]; + new_operands[map_inst_to_tuple_index[hoist]] = prod; + inplace = false; + } else { + map_inst_to_tuple_index[prod] = new_tuple_count++; + new_operands.push_back(prod); + } + } + } + + // Create the new root instruction. + HloComputation* cur_branch = conditional->branch_computation(branch); + HloInstruction* new_branch_root = + cur_branch->AddInstruction(HloInstruction::CreateTuple(new_operands)); + // The shape can vary since the operands to convert are now + // being returned through the branches' root. + cur_branch->set_root_instruction(new_branch_root, true /*new shape*/); + TF_CHECK_OK(cur_branch->RemoveInstruction(old_root)); + + // Only one of the branches needs to change the conditional->parent(). + if (branch != 0) { + continue; + } + HloComputation* conditional_parent = conditional->parent(); + HloInstruction* newconditional = + conditional_parent->AddInstruction(HloInstruction::CreateConditional( + cur_branch->root_instruction()->shape(), + conditional->mutable_operand(0), + absl::MakeSpan(conditional->branch_computations()), + absl::MakeSpan(conditional->operands()).subspan(1))); + // Ensure that all the users of conditional refer to the new one. + TF_RETURN_IF_ERROR( + conditional->ReplaceAllUsesWithDifferentShape(newconditional)); + TF_CHECK_OK(conditional_parent->RemoveInstruction(conditional)); + conditional = newconditional; + // Add the hoisted instructions in the parent. + for (HloInstruction* hoist : to_hoist_set) { + VLOG(2) << "Hoisting instruction:" << hoist->ToString(); + int64 hoist_index = map_inst_to_tuple_index[hoist]; + // Find out the gte that captured the hoisted instr result. + HloInstruction* gte_hoist = find_gte(conditional, hoist_index); + CHECK(gte_hoist != nullptr); + std::vector new_operands; + for (HloInstruction* op : hoist->operands()) { + HloInstruction* gte = conditional_parent->AddInstruction( + HloInstruction::CreateGetTupleElement( + op->shape(), conditional, map_inst_to_tuple_index[op])); + new_operands.push_back(gte); + } + HloInstruction* hoisted = conditional_parent->AddInstruction( + hoist->CloneWithNewOperands(hoist->shape(), new_operands)); + VLOG(2) << "Hoisted instruction in parent:" << hoisted->ToString(); + TF_RETURN_IF_ERROR(gte_hoist->ReplaceAllUsesWith(hoisted)); + TF_CHECK_OK(conditional_parent->RemoveInstruction(gte_hoist)); + } + // No need to explicitly delete a hoisted instruction since if its dead + // then the subsequent DCE will remove it. + } + } + VLOG(2) << "AFTER :" << conditional->parent()->parent()->ToString(); + return true; +} + // Hoist identical ops out of the conditional. The definition of identical // are the shape of the operands are identical and their properties are // identical. Will start from the root instruction of each branch and get // the identical ops to hoist. StatusOr MergeIdenticalElements(HloInstruction* conditional, bool is_layout_sensitive) { + VLOG(1) << " visiting conditional:" << conditional->ToString(); int branch_count = conditional->branch_count(); if (branch_count <= 0) { return false; @@ -399,7 +616,7 @@ StatusOr MergeIdenticalElements(HloInstruction* conditional, } } - if (visitors[0].HoistInstructionSize() <= 1) { + if (visitors[0].HoistInstructionSize() < 1) { return false; } @@ -442,7 +659,6 @@ StatusOr MergeIdenticalElements(HloInstruction* conditional, RemoveInstructionFromComputation(visitors[i].instructions_to_hoist(), conditional->branch_computation(i))); } - return true; } @@ -451,26 +667,55 @@ StatusOr MergeIdenticalElements(HloInstruction* conditional, StatusOr ConditionalCodeMotion::Run(HloModule* module) { bool changed = false; - // Gather all the conditional ops in our module. We do this ahead of time so - // we don't have to worry about mutating the lists of computations or - // instructions as we iterate. - std::vector conditional_ops; - for (auto* comp : module->MakeComputationPostOrder()) { - for (auto* instr : comp->MakeInstructionPostOrder()) { - if (instr->opcode() == HloOpcode::kConditional) { - conditional_ops.push_back(instr); + if (pursue_full_conditional_code_motion_) { + std::vector conditional_ops; + for (auto* comp : module->MakeComputationPostOrder()) { + for (auto* instr : comp->MakeInstructionPostOrder()) { + if (instr->opcode() == HloOpcode::kConditional) { + conditional_ops.push_back(instr); + } } } + + for (HloInstruction* conditional_op : conditional_ops) { + TF_ASSIGN_OR_RETURN( + bool result, + MergeIdenticalElements(conditional_op, is_layout_sensitive_)); + changed |= result; + } + + if (changed) { + HloPassPipeline subpipeline("after_conditional_code_motion"); + subpipeline.AddPass(); + subpipeline.AddPass(); + subpipeline.AddPass(); + TF_ASSIGN_OR_RETURN(bool cleanup_changed, subpipeline.Run(module)); + changed |= cleanup_changed; + } } - for (HloInstruction* conditional_op : conditional_ops) { - TF_ASSIGN_OR_RETURN(bool result, MergeIdenticalElements( - conditional_op, is_layout_sensitive_)); - changed |= result; + // handling convert rematerialization/hoisting + { + std::vector conditional_ops; + for (auto* comp : module->MakeComputationPostOrder()) { + for (auto* instr : comp->MakeInstructionPostOrder()) { + if (instr->opcode() == HloOpcode::kConditional) { + conditional_ops.push_back(instr); + } + } + } + for (HloInstruction* conditional_op : conditional_ops) { + TF_ASSIGN_OR_RETURN( + bool convert_result, + ConvertSpecialMove(conditional_op, is_layout_sensitive_)); + changed |= convert_result; + } } if (changed) { - HloPassPipeline subpipeline("after_conditional_code_motion"); + HloPassPipeline subpipeline( + "after_conditional_code_motion_after_convert_hoisting"); + subpipeline.AddPass(); subpipeline.AddPass(); subpipeline.AddPass(); TF_ASSIGN_OR_RETURN(bool cleanup_changed, subpipeline.Run(module)); diff --git a/tensorflow/compiler/xla/service/conditional_code_motion.h b/tensorflow/compiler/xla/service/conditional_code_motion.h index 1197a8b3620..95f02833e15 100644 --- a/tensorflow/compiler/xla/service/conditional_code_motion.h +++ b/tensorflow/compiler/xla/service/conditional_code_motion.h @@ -23,7 +23,11 @@ limitations under the License. namespace xla { -// HLO pass that moves identical ops out of conditional. +// ConditionalCodeMotion specializes in hoisting/rematerializing +// unconditional converts in the default mode. +// When pursue_full_conditional_code_motion_ is set to true, the +// full HLO pass moves identical ops out of a conditional in addition to moving +// converts. // - The definition of identical are the shape of the operands are identical // and their properties are identical. // - Currently, only some types of instructions is supported. @@ -35,13 +39,18 @@ class ConditionalCodeMotion : public HloModulePass { public: // If is_layout_sensitive is true, then the hoist process preserves layout // during identical comparison. Otherwise, layout is ignored. - explicit ConditionalCodeMotion(bool is_layout_sensitive = true) - : is_layout_sensitive_(is_layout_sensitive) {} + explicit ConditionalCodeMotion( + bool is_layout_sensitive = true, + bool pursue_full_conditional_code_motion = false) + : is_layout_sensitive_(is_layout_sensitive), + pursue_full_conditional_code_motion_( + pursue_full_conditional_code_motion) {} absl::string_view name() const override { return "conditional-code-motion"; } StatusOr Run(HloModule* module) override; private: const bool is_layout_sensitive_; + const bool pursue_full_conditional_code_motion_; }; } // namespace xla diff --git a/tensorflow/compiler/xla/service/conditional_code_motion_test.cc b/tensorflow/compiler/xla/service/conditional_code_motion_test.cc index 4a52303a42a..38b2b515fa0 100644 --- a/tensorflow/compiler/xla/service/conditional_code_motion_test.cc +++ b/tensorflow/compiler/xla/service/conditional_code_motion_test.cc @@ -38,7 +38,86 @@ namespace { using ConditionalCodeMotionTest = HloTestBase; namespace op = xla::testing::opcode_matchers; -TEST_F(ConditionalCodeMotionTest, DoNotMoveConvertOut) { +TEST_F(ConditionalCodeMotionTest, MoveSubsetTupleOut) { + absl::string_view hlo_string = + R"( +HloModule RemoveDotOpOut + +on_true { + %arg_tuple.1 = (f32[93184,4]{1,0}) parameter(0) + %get-tuple-element.1 = f32[93184,4]{1,0} get-tuple-element(%arg_tuple.1), index=0 + %reshape.8493 = f32[2,512,364]{2,1,0} reshape(f32[93184,4]{1,0} %get-tuple-element.1) + %convert.2894 = bf16[2,512,364]{2,1,0} convert(f32[2,512,364]{2,1,0} %reshape.8493) + ROOT %tuple.1 = ( bf16[2,512,364]{2,1,0}, f32[2,512,364]{2,1,0}) tuple(%convert.2894, %reshape.8493) +} + +on_false { + %arg_tuple.2 = (f32[93184,4]{1,0}) parameter(0) + %get-tuple-element.3 = f32[93184,4]{1,0} get-tuple-element(%arg_tuple.2), index=0 + %reshape.9717 = f32[2,512,364]{2,1,0} reshape(f32[93184,4]{1,0} %get-tuple-element.3) + %add = f32[2,512,364]{2,1,0} add(f32[2,512,364]{2,1,0} %reshape.9717, f32[2,512,364]{2,1,0} %reshape.9717) + %convert.3604 = bf16[2,512,364]{2,1,0} convert(f32[2,512,364]{2,1,0} %reshape.9717), metadata={op_type="Cast" op_name="gradients/Cast_125_grad/Cast"} + ROOT %tuple.2 = (bf16[2,512,364]{2,1,0}, f32[2,512,364]{2,1,0}) tuple(%convert.3604, %add) +} + +ENTRY main { + pred.1 = pred[] parameter(0) + arg_tuple.11 = (f32[93184,4]{1,0}) parameter(1) + arg_tuple.22 = (f32[93184,4]{1,0}) parameter(2) + conditional = (bf16[2,512,364]{2,1,0}, f32[2,512,364]{2,1,0}) conditional(pred.1, arg_tuple.11, arg_tuple.22), true_computation=on_true, false_computation=on_false + get-first-index = bf16[2,512,364]{2,1,0} get-tuple-element(conditional), index=0 + get-first-index.2 = f32[2,512,364]{2,1,0} get-tuple-element(conditional), index=1 + ROOT result = (bf16[2,512,364]{2,1,0}, f32[2,512,364]{2,1,0}) tuple(get-first-index, get-first-index.2) +} +)"; + auto module = ParseAndReturnVerifiedModule(hlo_string).ValueOrDie(); + ConditionalCodeMotion pass(true, true); + ASSERT_TRUE(pass.Run(&*module).ValueOrDie()); + + HloInstruction* root = module->entry_computation()->root_instruction(); + EXPECT_THAT(root, AllOf(op::Tuple(op::Convert(), op::GetTupleElement()))); +} + +TEST_F(ConditionalCodeMotionTest, MoveConvertOutConditionalRoot) { + absl::string_view hlo_string = + R"( +HloModule RemoveDotOpOut + +on_true { + %arg_tuple.1 = (f32[93184,4]{1,0}) parameter(0) + %get-tuple-element.1 = f32[93184,4]{1,0} get-tuple-element(%arg_tuple.1), index=0 + %reshape.8493 = f32[2,512,364]{2,1,0} reshape(f32[93184,4]{1,0} %get-tuple-element.1) + %add.8493 = f32[2,512,364]{2,1,0} add(f32[2,512,364]{2,1,0} %reshape.8493, f32[2,512,364]{2,1,0} %reshape.8493) + %convert.2894 = bf16[2,512,364]{2,1,0} convert(f32[2,512,364]{2,1,0} %add.8493) + ROOT %tuple.1 = ( bf16[2,512,364]{2,1,0}) tuple(%convert.2894) +} + +on_false { + %arg_tuple.2 = (f32[93184,4]{1,0}) parameter(0) + %get-tuple-element.3 = f32[93184,4]{1,0} get-tuple-element(%arg_tuple.2), index=0 + %reshape.9717 = f32[2,512,364]{2,1,0} reshape(f32[93184,4]{1,0} %get-tuple-element.3) + %add.8493 = f32[2,512,364]{2,1,0} add(f32[2,512,364]{2,1,0} %reshape.9717, f32[2,512,364]{2,1,0} %reshape.9717) + %sub.8493 = f32[2,512,364]{2,1,0} subtract(f32[2,512,364]{2,1,0} %add.8493, f32[2,512,364]{2,1,0} %reshape.9717) + %convert.3604 = bf16[2,512,364]{2,1,0} convert(f32[2,512,364]{2,1,0} %reshape.9717), metadata={op_type="Cast" op_name="gradients/Cast_125_grad/Cast"} + ROOT %tuple.2 = (bf16[2,512,364]{2,1,0}) tuple(%convert.3604) +} + +ENTRY main { + pred.1 = pred[] parameter(0) + arg_tuple.11 = (f32[93184,4]{1,0}) parameter(1) + arg_tuple.22 = (f32[93184,4]{1,0}) parameter(2) + ROOT conditional = (bf16[2,512,364]{2,1,0}) conditional(pred.1, arg_tuple.11, arg_tuple.22), true_computation=on_true, false_computation=on_false +} +)"; + auto module = ParseAndReturnVerifiedModule(hlo_string).ValueOrDie(); + ConditionalCodeMotion pass(true, true); + ASSERT_TRUE(pass.Run(&*module).ValueOrDie()); + + HloInstruction* root = module->entry_computation()->root_instruction(); + EXPECT_THAT(root, AllOf(op::Tuple(op::Convert()))); +} + +TEST_F(ConditionalCodeMotionTest, MoveConvertOut) { absl::string_view hlo_string = R"( HloModule RemoveDotOpOut @@ -65,12 +144,16 @@ ENTRY main { arg_tuple.22 = (f32[93184,4]{1,0}) parameter(2) conditional = (bf16[2,512,364]{2,1,0}) conditional(pred.1, arg_tuple.11, arg_tuple.22), true_computation=on_true, false_computation=on_false get-first-index = bf16[2,512,364]{2,1,0} get-tuple-element(conditional), index=0 - ROOT result = (bf16[2,512,364]{2,1,0}) tuple(get-first-index) + add.1 = bf16[2,512,364]{2,1,0} add(bf16[2,512,364]{2,1,0} get-first-index, bf16[2,512,364]{2,1,0} get-first-index) + ROOT result = (bf16[2,512,364]{2,1,0}) tuple(add.1) } )"; auto module = ParseAndReturnVerifiedModule(hlo_string).ValueOrDie(); - ConditionalCodeMotion pass; - ASSERT_FALSE(pass.Run(&*module).ValueOrDie()); + ConditionalCodeMotion pass(true, true); + ASSERT_TRUE(pass.Run(&*module).ValueOrDie()); + + HloInstruction* root = module->entry_computation()->root_instruction(); + EXPECT_THAT(root, AllOf(op::Tuple(op::Add(op::Convert(), op::Convert())))); } TEST_F(ConditionalCodeMotionTest, UserShareOperandCannotBeMoved) { @@ -123,7 +206,7 @@ ENTRY main { } )"; auto module = ParseAndReturnVerifiedModule(hlo_string).ValueOrDie(); - ConditionalCodeMotion pass; + ConditionalCodeMotion pass(true, true); ASSERT_TRUE(pass.Run(&*module).ValueOrDie()); const HloInstruction* conditional = @@ -181,7 +264,7 @@ ENTRY main { } )"; auto module = ParseAndReturnVerifiedModule(hlo_string).ValueOrDie(); - ConditionalCodeMotion pass; + ConditionalCodeMotion pass(true, true); ASSERT_TRUE(pass.Run(&*module).ValueOrDie()); const HloInstruction* conditional = FindInstruction(module.get(), "conditional"); @@ -245,7 +328,7 @@ ENTRY main { } )"; auto module = ParseAndReturnVerifiedModule(hlo_string).ValueOrDie(); - ConditionalCodeMotion pass; + ConditionalCodeMotion pass(true, true); ASSERT_TRUE(pass.Run(&*module).ValueOrDie()); const HloInstruction* conditional = @@ -317,7 +400,7 @@ ENTRY main { )"; auto module = ParseAndReturnVerifiedModule(hlo_string).ValueOrDie(); - ConditionalCodeMotion pass; + ConditionalCodeMotion pass(true, true); ASSERT_FALSE(pass.Run(&*module).ValueOrDie()); } @@ -390,7 +473,7 @@ ENTRY main { } )"; auto module = ParseAndReturnVerifiedModule(hlo_string).ValueOrDie(); - ConditionalCodeMotion pass; + ConditionalCodeMotion pass(true, true); ASSERT_TRUE(pass.Run(&*module).ValueOrDie()); const HloInstruction* conditional = FindInstruction(module.get(), "conditional"); From 0b0eef4031fa2674a2c5d32aa7570a82c3def6a8 Mon Sep 17 00:00:00 2001 From: Chao Mei Date: Thu, 18 Jun 2020 20:59:28 -0700 Subject: [PATCH 271/867] Move enabling xnnpack delegate to AllocateTensors to allow other delegates to be applied first. PiperOrigin-RevId: 317240424 Change-Id: I89b616f891f65f7cff6beedbf5c2a372f7456592 --- tensorflow/lite/interpreter.cc | 15 +++++++++++++-- tensorflow/lite/interpreter.h | 15 +++++++++++---- tensorflow/lite/interpreter_builder.cc | 17 ++++++----------- 3 files changed, 30 insertions(+), 17 deletions(-) diff --git a/tensorflow/lite/interpreter.cc b/tensorflow/lite/interpreter.cc index cae2ca7dde0..b49aa5031bf 100644 --- a/tensorflow/lite/interpreter.cc +++ b/tensorflow/lite/interpreter.cc @@ -86,8 +86,9 @@ TfLiteQuantization GetQuantizationFromLegacy( } // namespace Interpreter::Interpreter(ErrorReporter* error_reporter) - : error_reporter_(error_reporter ? error_reporter - : DefaultErrorReporter()) { + : error_reporter_(error_reporter ? error_reporter : DefaultErrorReporter()), + lazy_delegate_provider_( + TfLiteDelegatePtr(nullptr, [](TfLiteDelegate*) {})) { // TODO(b/128420794): Include the TFLite runtime version in the log. // Prod logging is useful for mobile platforms where scraping console logs is // critical for debugging. @@ -175,6 +176,16 @@ TfLiteStatus Interpreter::SetVariables(std::vector variables) { } TfLiteStatus Interpreter::AllocateTensors() { + // Apply the default delegate that TFLite will enable at this point to allow + // other user-level delegates to be applied first. + if (lazy_delegate_provider_) { + // The execution will fall back to default implementation if the XNNPACK + // delegate fails to be applied. Therefore, we ignore the return status + // here and let it fall through the rest of the code. + ModifyGraphWithDelegate(std::move(lazy_delegate_provider_)); + lazy_delegate_provider_.reset(); + } + return primary_subgraph().AllocateTensors(); } diff --git a/tensorflow/lite/interpreter.h b/tensorflow/lite/interpreter.h index 59cab6add6d..41377c4ce1f 100644 --- a/tensorflow/lite/interpreter.h +++ b/tensorflow/lite/interpreter.h @@ -347,10 +347,12 @@ class Interpreter { /// WARNING: Experimental interface, subject to change TfLiteStatus ReleaseNonPersistentMemory(); - /// Update allocations for all tensors. This will redim dependent tensors - /// using the input tensor dimensionality as given. This is relatively - /// expensive. If you know that your sizes are not changing, you need not call - /// this. Returns status of success or failure. + // Update allocations for all tensors. This will redim dependent tensors + // using the input tensor dimensionality as given. This is relatively + // expensive. This *must be* called after the interpreter has been created + // and before running inference (and accessing tensor buffers), and *must be* + // called again if (and only if) an input tensor is resized. Returns status of + // success or failure. TfLiteStatus AllocateTensors(); /// Invoke the interpreter (run the whole graph in dependency order). @@ -594,6 +596,11 @@ class Interpreter { // A map of resources. Owned by interpreter and shared by multiple subgraphs. resource::ResourceMap resources_; + + // Indicating a delegate that the TFLite interpreter will apply by default. + // A nullptr value means there's no delegate to be applied by default or the + // delegate has been applied and doesn't need to be applied again. + TfLiteDelegatePtr lazy_delegate_provider_; }; } // namespace impl diff --git a/tensorflow/lite/interpreter_builder.cc b/tensorflow/lite/interpreter_builder.cc index d73b298e595..4b491d41881 100644 --- a/tensorflow/lite/interpreter_builder.cc +++ b/tensorflow/lite/interpreter_builder.cc @@ -545,17 +545,7 @@ TfLiteStatus InterpreterBuilder::ParseTensors( TfLiteStatus InterpreterBuilder::ApplyDelegates(Interpreter* interpreter, int num_threads) { - // First, apply XNNPACK delegate if applicable. - if (num_fp32_tensors_ > 0) { - // The execution will fall back to default implementation if the XNNPACK - // delegate fails to be applied. Therefore, we ignore the return status - // here and let it fall through the rest of the code. - if (auto xnnpack_delegate = MaybeCreateXNNPACKDelegate(num_threads)) { - interpreter->ModifyGraphWithDelegate(std::move(xnnpack_delegate)); - } - } - - // Secondly, apply Flex delegate if applicable. + // Apply Flex delegate if applicable. if (has_flex_op_) { if (auto flex_delegate = AcquireFlexDelegate()) { return interpreter->ModifyGraphWithDelegate(std::move(flex_delegate)); @@ -672,6 +662,11 @@ TfLiteStatus InterpreterBuilder::operator()( modified_subgraph->SetVariables(std::move(variables)); } + if (num_fp32_tensors_ > 0) { + (*interpreter)->lazy_delegate_provider_ = + MaybeCreateXNNPACKDelegate(num_threads); + } + if (ApplyDelegates(interpreter->get(), num_threads) != kTfLiteOk) return cleanup_and_error(); From b5bb616121f1805c5ff5391daf00c86b6bcad1ae Mon Sep 17 00:00:00 2001 From: "A. Unique TensorFlower" Date: Thu, 18 Jun 2020 21:06:35 -0700 Subject: [PATCH 272/867] *** Reason for rollback *** CL316577397 prevents H2D prefetch overlapping for TF2 MLPerf Resnet model on V100x8 More details in b/159372996 With the rollback, the prefetching overlapping is back, and the training speed also recovered. *** Original change description *** Add DT_BOOL support to GPU variable ops This is a follow-on to PR #38848 & PR #39172 and resolves remaining ask in Issue #35994. The original PR tried to add many variable ops on the GPU including DT_BOOL. However, this caused testCondModifyBoolPred to fail and thus the DT_BOOL type was removed. The reason for the test failure is once DT_BOOL variables are supported on the GPU, we need to ensure the switch ops are also updated to not have ho... *** PiperOrigin-RevId: 317241338 Change-Id: Id7b7d79622e0537ccb677f081b487014ac4d2395 --- tensorflow/core/kernels/control_flow_ops.cc | 10 +++++----- tensorflow/core/kernels/variable_ops.cc | 3 ++- .../debug/lib/debug_graph_reconstruction_test.py | 6 +++--- tensorflow/python/ops/control_flow_ops_test.py | 6 +++--- 4 files changed, 13 insertions(+), 12 deletions(-) diff --git a/tensorflow/core/kernels/control_flow_ops.cc b/tensorflow/core/kernels/control_flow_ops.cc index 435de3c5954..c8e83b6f672 100644 --- a/tensorflow/core/kernels/control_flow_ops.cc +++ b/tensorflow/core/kernels/control_flow_ops.cc @@ -111,17 +111,15 @@ REGISTER_GPU_SWITCH(uint64); TF_CALL_variant(REGISTER_GPU_SWITCH); TF_CALL_uint32(REGISTER_GPU_SWITCH); TF_CALL_uint32(REGISTER_GPU_REF_SWITCH); -TF_CALL_bool(REGISTER_GPU_SWITCH); -TF_CALL_bool(REGISTER_GPU_REF_SWITCH); #undef REGISTER_CPU_SWITCH #undef REGISTER_CPU_REF_SWITCH #undef REGISTER_GPU_SWITCH #undef REGISTER_GPU_REF_SWITCH -// Special GPU kernels for int32, string & resource handles. Requiring all -// inputs and outputs to be in host memory. -// TODO(b/25387198): Also enable int32 in device memory. +// Special GPU kernels for int32 and string. +// TODO(b/25387198): Also enable int32 in device memory. This kernel +// registration requires all int32 inputs and outputs to be in host memory. #define REGISTER_GPU_HOST_KERNEL(type) \ REGISTER_KERNEL_BUILDER(Name("Switch") \ .Device(DEVICE_GPU) \ @@ -151,6 +149,8 @@ TF_CALL_bool(REGISTER_GPU_REF_SWITCH); REGISTER_GPU_HOST_KERNEL(int32); REGISTER_GPU_HOST_REF_KERNEL(int32); +REGISTER_GPU_HOST_KERNEL(bool); +REGISTER_GPU_HOST_REF_KERNEL(bool); REGISTER_GPU_HOST_KERNEL(tstring); REGISTER_GPU_HOST_REF_KERNEL(tstring); REGISTER_GPU_HOST_KERNEL(ResourceHandle); diff --git a/tensorflow/core/kernels/variable_ops.cc b/tensorflow/core/kernels/variable_ops.cc index ccd33e8c75a..6f5e0b94eca 100644 --- a/tensorflow/core/kernels/variable_ops.cc +++ b/tensorflow/core/kernels/variable_ops.cc @@ -252,7 +252,8 @@ TF_CALL_GPU_NUMBER_TYPES_NO_HALF(REGISTER_SYCL_KERNEL); TF_CALL_int64(REGISTER_GPU_KERNELS); TF_CALL_uint32(REGISTER_GPU_KERNELS); -TF_CALL_GPU_ALL_TYPES(REGISTER_GPU_KERNELS); +TF_CALL_GPU_NUMBER_TYPES(REGISTER_GPU_KERNELS); +TF_CALL_COMPLEX_TYPES(REGISTER_GPU_KERNELS); #undef REGISTER_GPU_KERNELS #endif // GOOGLE_CUDA || TENSORFLOW_USE_ROCM diff --git a/tensorflow/python/debug/lib/debug_graph_reconstruction_test.py b/tensorflow/python/debug/lib/debug_graph_reconstruction_test.py index b3baa6e7bc2..fb722efab4e 100644 --- a/tensorflow/python/debug/lib/debug_graph_reconstruction_test.py +++ b/tensorflow/python/debug/lib/debug_graph_reconstruction_test.py @@ -73,9 +73,9 @@ class ReconstructNonDebugGraphTest(test_util.TensorFlowTestCase): for attr_key in new_node.attr: if attr_key == "parallel_iterations": new_node.attr[attr_key].i = 1 - elif new_node.op == "Switch" or new_node.op == "Identity": - # We don't check the inputs to Switch or Identity ops as their inputs - # may be Send/Recv nodes. + elif new_node.op == "Switch": + # We don't check the inputs to Switch ops as their inputs may be + # Send/Recv nodes. del new_node.input[:] return output_graph_def diff --git a/tensorflow/python/ops/control_flow_ops_test.py b/tensorflow/python/ops/control_flow_ops_test.py index 3ca9bda82f2..9254695d988 100644 --- a/tensorflow/python/ops/control_flow_ops_test.py +++ b/tensorflow/python/ops/control_flow_ops_test.py @@ -396,10 +396,10 @@ class CondTest(test_util.TensorFlowTestCase): fn2=lambda: math_ops.add(y, 23)) self.assertEquals(self.evaluate(z), 24) - @test_util.run_v1_only("Exercises Ref variables") + @test_util.run_deprecated_v1 def testCondModifyBoolPred(self): - # We want to use the GPU here because we want to ensure that we can update - # a boolean ref variable on the GPU. + # This test in particular used to fail only when running in GPU, hence + # use_gpu=True. with test_util.use_gpu(): bool_var = variable_scope.get_variable( "bool_var", dtype=dtypes.bool, initializer=True) From cfbdd27fe3f2b904609e1551490a01640ae4fcac Mon Sep 17 00:00:00 2001 From: "A. Unique TensorFlower" Date: Thu, 18 Jun 2020 21:17:54 -0700 Subject: [PATCH 273/867] Update ops-related pbtxt files. PiperOrigin-RevId: 317242544 Change-Id: I36000fdb2d595b5006ea111105ece5ca6f537732 --- .../core/ops/compat/ops_history_v2/Acos.pbtxt | 29 +++++++++++++++++++ .../core/ops/compat/ops_history_v2/Asin.pbtxt | 29 +++++++++++++++++++ .../core/ops/compat/ops_history_v2/Atan.pbtxt | 29 +++++++++++++++++++ .../core/ops/compat/ops_history_v2/Inv.pbtxt | 29 +++++++++++++++++++ .../core/ops/compat/ops_history_v2/Neg.pbtxt | 29 +++++++++++++++++++ .../compat/ops_history_v2/Reciprocal.pbtxt | 29 +++++++++++++++++++ .../ops/compat/ops_history_v2/Round.pbtxt | 29 +++++++++++++++++++ .../ops/compat/ops_history_v2/Square.pbtxt | 29 +++++++++++++++++++ .../core/ops/compat/ops_history_v2/Tan.pbtxt | 29 +++++++++++++++++++ tensorflow/core/ops/ops.pbtxt | 18 ++++++++++++ 10 files changed, 279 insertions(+) diff --git a/tensorflow/core/ops/compat/ops_history_v2/Acos.pbtxt b/tensorflow/core/ops/compat/ops_history_v2/Acos.pbtxt index 3ed45186f6e..417dbfc7e7d 100644 --- a/tensorflow/core/ops/compat/ops_history_v2/Acos.pbtxt +++ b/tensorflow/core/ops/compat/ops_history_v2/Acos.pbtxt @@ -78,3 +78,32 @@ op { } } } +op { + name: "Acos" + input_arg { + name: "x" + type_attr: "T" + } + output_arg { + name: "y" + type_attr: "T" + } + attr { + name: "T" + type: "type" + allowed_values { + list { + type: DT_BFLOAT16 + type: DT_HALF + type: DT_FLOAT + type: DT_DOUBLE + type: DT_INT8 + type: DT_INT16 + type: DT_INT32 + type: DT_INT64 + type: DT_COMPLEX64 + type: DT_COMPLEX128 + } + } + } +} diff --git a/tensorflow/core/ops/compat/ops_history_v2/Asin.pbtxt b/tensorflow/core/ops/compat/ops_history_v2/Asin.pbtxt index 7df768f7c66..c799ff99169 100644 --- a/tensorflow/core/ops/compat/ops_history_v2/Asin.pbtxt +++ b/tensorflow/core/ops/compat/ops_history_v2/Asin.pbtxt @@ -78,3 +78,32 @@ op { } } } +op { + name: "Asin" + input_arg { + name: "x" + type_attr: "T" + } + output_arg { + name: "y" + type_attr: "T" + } + attr { + name: "T" + type: "type" + allowed_values { + list { + type: DT_BFLOAT16 + type: DT_HALF + type: DT_FLOAT + type: DT_DOUBLE + type: DT_INT8 + type: DT_INT16 + type: DT_INT32 + type: DT_INT64 + type: DT_COMPLEX64 + type: DT_COMPLEX128 + } + } + } +} diff --git a/tensorflow/core/ops/compat/ops_history_v2/Atan.pbtxt b/tensorflow/core/ops/compat/ops_history_v2/Atan.pbtxt index 86f0628ab53..4a80c7a751e 100644 --- a/tensorflow/core/ops/compat/ops_history_v2/Atan.pbtxt +++ b/tensorflow/core/ops/compat/ops_history_v2/Atan.pbtxt @@ -78,3 +78,32 @@ op { } } } +op { + name: "Atan" + input_arg { + name: "x" + type_attr: "T" + } + output_arg { + name: "y" + type_attr: "T" + } + attr { + name: "T" + type: "type" + allowed_values { + list { + type: DT_BFLOAT16 + type: DT_HALF + type: DT_FLOAT + type: DT_DOUBLE + type: DT_INT8 + type: DT_INT16 + type: DT_INT32 + type: DT_INT64 + type: DT_COMPLEX64 + type: DT_COMPLEX128 + } + } + } +} diff --git a/tensorflow/core/ops/compat/ops_history_v2/Inv.pbtxt b/tensorflow/core/ops/compat/ops_history_v2/Inv.pbtxt index ca208664617..0c191790030 100644 --- a/tensorflow/core/ops/compat/ops_history_v2/Inv.pbtxt +++ b/tensorflow/core/ops/compat/ops_history_v2/Inv.pbtxt @@ -168,3 +168,32 @@ op { } } } +op { + name: "Inv" + input_arg { + name: "x" + type_attr: "T" + } + output_arg { + name: "y" + type_attr: "T" + } + attr { + name: "T" + type: "type" + allowed_values { + list { + type: DT_BFLOAT16 + type: DT_HALF + type: DT_FLOAT + type: DT_DOUBLE + type: DT_INT8 + type: DT_INT16 + type: DT_INT32 + type: DT_INT64 + type: DT_COMPLEX64 + type: DT_COMPLEX128 + } + } + } +} diff --git a/tensorflow/core/ops/compat/ops_history_v2/Neg.pbtxt b/tensorflow/core/ops/compat/ops_history_v2/Neg.pbtxt index 77bb4a5872d..864d0257fe4 100644 --- a/tensorflow/core/ops/compat/ops_history_v2/Neg.pbtxt +++ b/tensorflow/core/ops/compat/ops_history_v2/Neg.pbtxt @@ -78,3 +78,32 @@ op { } } } +op { + name: "Neg" + input_arg { + name: "x" + type_attr: "T" + } + output_arg { + name: "y" + type_attr: "T" + } + attr { + name: "T" + type: "type" + allowed_values { + list { + type: DT_BFLOAT16 + type: DT_HALF + type: DT_FLOAT + type: DT_DOUBLE + type: DT_INT8 + type: DT_INT16 + type: DT_INT32 + type: DT_INT64 + type: DT_COMPLEX64 + type: DT_COMPLEX128 + } + } + } +} diff --git a/tensorflow/core/ops/compat/ops_history_v2/Reciprocal.pbtxt b/tensorflow/core/ops/compat/ops_history_v2/Reciprocal.pbtxt index 5ea1abe4c9c..7e03554871a 100644 --- a/tensorflow/core/ops/compat/ops_history_v2/Reciprocal.pbtxt +++ b/tensorflow/core/ops/compat/ops_history_v2/Reciprocal.pbtxt @@ -78,3 +78,32 @@ op { } } } +op { + name: "Reciprocal" + input_arg { + name: "x" + type_attr: "T" + } + output_arg { + name: "y" + type_attr: "T" + } + attr { + name: "T" + type: "type" + allowed_values { + list { + type: DT_BFLOAT16 + type: DT_HALF + type: DT_FLOAT + type: DT_DOUBLE + type: DT_INT8 + type: DT_INT16 + type: DT_INT32 + type: DT_INT64 + type: DT_COMPLEX64 + type: DT_COMPLEX128 + } + } + } +} diff --git a/tensorflow/core/ops/compat/ops_history_v2/Round.pbtxt b/tensorflow/core/ops/compat/ops_history_v2/Round.pbtxt index 4f59b21afd5..c5685dc6143 100644 --- a/tensorflow/core/ops/compat/ops_history_v2/Round.pbtxt +++ b/tensorflow/core/ops/compat/ops_history_v2/Round.pbtxt @@ -78,3 +78,32 @@ op { } } } +op { + name: "Round" + input_arg { + name: "x" + type_attr: "T" + } + output_arg { + name: "y" + type_attr: "T" + } + attr { + name: "T" + type: "type" + allowed_values { + list { + type: DT_BFLOAT16 + type: DT_HALF + type: DT_FLOAT + type: DT_DOUBLE + type: DT_INT8 + type: DT_INT16 + type: DT_INT32 + type: DT_INT64 + type: DT_COMPLEX64 + type: DT_COMPLEX128 + } + } + } +} diff --git a/tensorflow/core/ops/compat/ops_history_v2/Square.pbtxt b/tensorflow/core/ops/compat/ops_history_v2/Square.pbtxt index 4d07faf4fd0..6af75b3ddc1 100644 --- a/tensorflow/core/ops/compat/ops_history_v2/Square.pbtxt +++ b/tensorflow/core/ops/compat/ops_history_v2/Square.pbtxt @@ -78,3 +78,32 @@ op { } } } +op { + name: "Square" + input_arg { + name: "x" + type_attr: "T" + } + output_arg { + name: "y" + type_attr: "T" + } + attr { + name: "T" + type: "type" + allowed_values { + list { + type: DT_BFLOAT16 + type: DT_HALF + type: DT_FLOAT + type: DT_DOUBLE + type: DT_INT8 + type: DT_INT16 + type: DT_INT32 + type: DT_INT64 + type: DT_COMPLEX64 + type: DT_COMPLEX128 + } + } + } +} diff --git a/tensorflow/core/ops/compat/ops_history_v2/Tan.pbtxt b/tensorflow/core/ops/compat/ops_history_v2/Tan.pbtxt index 7dc7f84fd38..80e0b1e22c4 100644 --- a/tensorflow/core/ops/compat/ops_history_v2/Tan.pbtxt +++ b/tensorflow/core/ops/compat/ops_history_v2/Tan.pbtxt @@ -78,3 +78,32 @@ op { } } } +op { + name: "Tan" + input_arg { + name: "x" + type_attr: "T" + } + output_arg { + name: "y" + type_attr: "T" + } + attr { + name: "T" + type: "type" + allowed_values { + list { + type: DT_BFLOAT16 + type: DT_HALF + type: DT_FLOAT + type: DT_DOUBLE + type: DT_INT8 + type: DT_INT16 + type: DT_INT32 + type: DT_INT64 + type: DT_COMPLEX64 + type: DT_COMPLEX128 + } + } + } +} diff --git a/tensorflow/core/ops/ops.pbtxt b/tensorflow/core/ops/ops.pbtxt index 1f1cf7444fb..dbd91c91b65 100644 --- a/tensorflow/core/ops/ops.pbtxt +++ b/tensorflow/core/ops/ops.pbtxt @@ -216,6 +216,8 @@ op { type: DT_HALF type: DT_FLOAT type: DT_DOUBLE + type: DT_INT8 + type: DT_INT16 type: DT_INT32 type: DT_INT64 type: DT_COMPLEX64 @@ -2333,6 +2335,8 @@ op { type: DT_HALF type: DT_FLOAT type: DT_DOUBLE + type: DT_INT8 + type: DT_INT16 type: DT_INT32 type: DT_INT64 type: DT_COMPLEX64 @@ -2646,6 +2650,8 @@ op { type: DT_HALF type: DT_FLOAT type: DT_DOUBLE + type: DT_INT8 + type: DT_INT16 type: DT_INT32 type: DT_INT64 type: DT_COMPLEX64 @@ -19442,6 +19448,8 @@ op { type: DT_HALF type: DT_FLOAT type: DT_DOUBLE + type: DT_INT8 + type: DT_INT16 type: DT_INT32 type: DT_INT64 type: DT_COMPLEX64 @@ -25498,6 +25506,8 @@ op { type: DT_HALF type: DT_FLOAT type: DT_DOUBLE + type: DT_INT8 + type: DT_INT16 type: DT_INT32 type: DT_INT64 type: DT_COMPLEX64 @@ -35191,6 +35201,8 @@ op { type: DT_HALF type: DT_FLOAT type: DT_DOUBLE + type: DT_INT8 + type: DT_INT16 type: DT_INT32 type: DT_INT64 type: DT_COMPLEX64 @@ -40686,6 +40698,8 @@ op { type: DT_HALF type: DT_FLOAT type: DT_DOUBLE + type: DT_INT8 + type: DT_INT16 type: DT_INT32 type: DT_INT64 type: DT_COMPLEX64 @@ -48071,6 +48085,8 @@ op { type: DT_HALF type: DT_FLOAT type: DT_DOUBLE + type: DT_INT8 + type: DT_INT16 type: DT_INT32 type: DT_INT64 type: DT_COMPLEX64 @@ -50832,6 +50848,8 @@ op { type: DT_HALF type: DT_FLOAT type: DT_DOUBLE + type: DT_INT8 + type: DT_INT16 type: DT_INT32 type: DT_INT64 type: DT_COMPLEX64 From b7caba2c42285a8e1cb875bec1664d5b0e6c65e9 Mon Sep 17 00:00:00 2001 From: Ashwin Murthy Date: Thu, 18 Jun 2020 22:10:41 -0700 Subject: [PATCH 274/867] Update RNN conversion tflite g3doc This uses the content from the blog post/dogfood announcement email PiperOrigin-RevId: 317248288 Change-Id: I210c64bd54c70aa5b68742d59d6d36fa154e856c --- tensorflow/lite/g3doc/convert/rnn.md | 240 +++++++++++++++++++-------- 1 file changed, 167 insertions(+), 73 deletions(-) diff --git a/tensorflow/lite/g3doc/convert/rnn.md b/tensorflow/lite/g3doc/convert/rnn.md index 52bc287c151..734992c0904 100644 --- a/tensorflow/lite/g3doc/convert/rnn.md +++ b/tensorflow/lite/g3doc/convert/rnn.md @@ -1,99 +1,193 @@ -# Convert RNN models +# TensorFlow RNN conversion to TensorFlow Lite -The TensorFlow Lite interpreter currently implements a subset of TensorFlow -operations, meaning some model architectures cannot immediately be converted due -to missing operations. +## Overview -Some RNN-based architectures are affected by this. The following document -outlines the current state of play and provides strategies for converting RNN -models. +TensorFlow Lite supports converting TensorFlow RNN models to TensorFlow Lite’s +fused LSTM operators. Fused operators exist to maximize the performance of their +underlying kernel implementations, as well as provide a higher level interface +to define complex transformations like quantizatization. -## Currently supported +Since there are many variants of RNN APIs in TensorFlow, our approach has been +two fold: -Currently, RNN models using -[`tf.compat.v1.nn.static_rnn`](https://www.tensorflow.org/api_docs/python/tf/nn/static_rnn) -can be converted successfully as long as no `sequence_length` is specified. +1. Provide **native support for standard TensorFlow RNN APIs** like Keras LSTM. + This is the recommended option. +1. Provide an **interface** **into the conversion infrastructure for** + **user-defined** **RNN implementations** to plug in and get converted to + TensorFlow Lite. We provide a couple of out of box examples of such + conversion using lingvo’s + [LSTMCellSimple](https://github.com/tensorflow/tensorflow/blob/master/tensorflow/compiler/mlir/lite/transforms/prepare_composite_functions_tf.cc#L123) + and + [LayerNormalizedLSTMCellSimple](https://github.com/tensorflow/tensorflow/blob/master/tensorflow/compiler/mlir/lite/utils/lstm_utils.cc#L519) + RNN interfaces. -The following `tf.compat.v1.nn.rnn_cell` operations work with -`tf.compat.v1.nn.static_rnn`: +## Converter API -* [tf.compat.v1.nn.rnn_cell.LSTMCell](https://www.tensorflow.org/api_docs/python/tf/nn/rnn_cell/LSTMCell) -* [tf.compat.v1.nn.rnn_cell.RNNCell](https://www.tensorflow.org/api_docs/python/tf/nn/rnn_cell/RNNCell) -* [tf.compat.v1.nn.rnn_cell.GRUCell](https://www.tensorflow.org/api_docs/python/tf/nn/rnn_cell/GRUCell) -* [tf.compat.v1.nn.rnn_cell.BasicLSTMCell](https://www.tensorflow.org/api_docs/python/tf/nn/rnn_cell/BasicLSTMCell) -* [tf.compat.v1.nn.rnn_cell.BasicRNNCell](https://www.tensorflow.org/api_docs/python/tf/nn/rnn_cell/BasicRNNCell) +Currently this feature is available through the +[tf-nightly](https://pypi.org/project/tf-nightly/) pip or from head. This will +be available in the TensorFlow 2.3 release. -In addition, TensorFlow Lite provides some experimental drop-in replacements for -RNN operations that enable dynamic RNN architectures with TensorFlow Lite. +This conversion functionality is available when converting to TensorFlow Lite +via a SavedModel or from the Keras model directly. See example usages. -Drop-in replacements are available for the following: +### From saved model -* [tf.compat.v1.nn.dynamic_rnn](https://www.tensorflow.org/api_docs/python/tf/nn/dynamic_rnn) -* [tf.compat.v1.nn.bidirectional_dynamic_rnn](https://www.tensorflow.org/api_docs/python/tf/nn/bidirectional_dynamic_rnn) -* [tf.compat.v1.nn.rnn_cell.RNNCell](https://www.tensorflow.org/api_docs/python/tf/nn/rnn_cell/RNNCell) -* [tf.compat.v1.nn.rnn_cell.LSTMCell](https://www.tensorflow.org/api_docs/python/tf/nn/rnn_cell/LSTMCell) +``` +# build a saved model. Here concrete_function is the exported function +# corresponding to the TensorFlow model containing one or more +# Keras LSTM layers. +saved_model, saved_model_dir = build_saved_model_lstm(...) +saved_model.save(saved_model_dir, save_format="tf", signatures=concrete_func) -## Not currently supported +# Convert the model. +converter = TFLiteConverter.from_saved_model(saved_model_dir) +tflite_model = converter.convert() +``` -TensorFlow Lite does not currently support -[Control Flow](https://www.tensorflow.org/api_docs/cc/group/control-flow-ops) -operations. This means that, unless one of the conversion strategies discussed -in the next section are employed, models built with the following TensorFlow -functions will not convert successfully: +### From Keras model -* [tf.compat.v1.nn.static_rnn](https://www.tensorflow.org/api_docs/python/tf/nn/static_rnn) - where a `sequence_length` is specified -* [tf.compat.v1.nn.dynamic_rnn](https://www.tensorflow.org/api_docs/python/tf/nn/dynamic_rnn) -* [tf.compat.v1.nn.bidirectional_dynamic_rnn](https://www.tensorflow.org/api_docs/python/tf/nn/bidirectional_dynamic_rnn) +``` +# build a Keras model +keras_model = build_keras_lstm(...) -Note: TensorFlow Lite plans to implement all required Control Flow operations by -the end of 2019. At this point, all RNN architectures will convert successfully. +# Convert the model. +converter = TFLiteConverter.from_keras_model(keras_model) +tflite_model = converter.convert() -## Conversion strategies +``` -To convert an RNN model that uses the functions specified above, you will have -to modify its architecture and retrain it. The following strategies can be used. +## Example -### 1. Refactoring +Keras LSTM to TensorFlow Lite +[Colab](https://colab.research.google.com/github/tensorflow/tensorflow/blob/master/tensorflow/lite/examples/experimental_new_converter/Keras_LSTM_fusion_Codelab.ipynb) +illustrates the end to end usage with the TensorFlow Lite interpreter. -The simplest approach, if possible, is to refactor the model architecture to use -[tf.compat.v1.nn.static_rnn](https://www.tensorflow.org/api_docs/python/tf/nn/static_rnn) -without `sequence_length`. +## TensorFlow RNNs APIs supported -### 2. Drop-in replacements that use op hints and fused ops +### Keras LSTM conversion (recommended) -TensorFlow Lite provides the some experimental drop-in replacements for RNN -operations that enable dynamic RNN architectures with TensorFlow Lite. Using -[OpHints](https://www.tensorflow.org/lite/guide/ops_custom#converting_tensorflow_models_to_convert_graphs), -they run normally during training, but are substituted with special fused ops -when run by the Lite interpreter. +We support out-of-the-box conversion of Keras LSTM to TensorFlow Lite. For +details on how this works please refer to the +[Keras LSTM interface](https://colab.sandbox.google.com/github/tensorflow/tensorflow/blob/master/tensorflow/lite/examples/experimental_new_converter/Keras_LSTM_fusion_Codelab.ipynb) +and to the conversion logic +[here](https://github.com/tensorflow/tensorflow/blob/master/tensorflow/compiler/mlir/lite/utils/lstm_utils.cc#L627). -The following drop-in replacements are available: +Also important is to highlight the TensorFlow Lite’s LSTM contract with respect +to the Keras operation definition: -* [tf.compat.v1.lite.experimental.nn.dynamic_rnn](https://github.com/tensorflow/tensorflow/blob/master/tensorflow/lite/experimental/examples/lstm/rnn.py#L41) - * replacement for tf.nn.dynamic_rnn -* [tf.compat.v1.lite.experimental.nn.bidirectional_dynamic_rnn](https://github.com/tensorflow/tensorflow/blob/master/tensorflow/lite/experimental/examples/lstm/rnn.py#L279) - * replacement for tf.nn.bidirectional_dynamic_rnn -* [tf.compat.v1.lite.experimental.nn.TfLiteRNNCell](https://github.com/tensorflow/tensorflow/blob/master/tensorflow/lite/experimental/examples/lstm/rnn_cell.py#L39) - * replacement for tf.nn.rnn_cell.RNNCell -* [tf.compat.v1.lite.experimental.nn.TfLiteLSTMCell](https://github.com/tensorflow/tensorflow/blob/master/tensorflow/lite/experimental/examples/lstm/rnn_cell.py#L159) - * replacement for tf.nn.rnn_cell.LSTMCell +1. The dimension 0 of the input tensor is the batch size. +1. The dimension 0 of the recurrent\_weight tensor is the number of outputs. +1. The **weight** and **recurrent\_kernel** tensors are transposed. +1. The transposed weight, transposed recurrent\_kernel and bias tensors are + split into 4 equal sized tensors along the dimension 0. These correspond to + **input gate, forget gate, cell, and output gate**. -Note: These replacements must be used together. For example, if you are using -`tf.compat.v1.lite.experimental.nn.dynamic_rnn`, you must combine it with -`tf.compat.v1.lite.experimental.nn.TfLiteRNNCell` instead of using -`tf.compat.v1.nn.rnn_cell.RNNCell`. +#### Keras LSTM Variants -Instead of -[tf.compat.v1.nn.rnn_cell.MultiRNNCell](https://www.tensorflow.org/api_docs/python/tf/nn/rnn_cell/MultiRNNCell), -you should use -[tf.compat.v1.keras.layers.StackedRNNCells](https://www.tensorflow.org/api_docs/python/tf/keras/layers/StackedRNNCells). +##### Time major -For a tutorial on using these replacements, see -[TensorFlow Lite LSTM ops API](https://github.com/tensorflow/tensorflow/blob/master/tensorflow/lite/experimental/examples/lstm/g3doc/README.md). +Users may choose time-major or no time-major. Keras LSTM adds a time-major +attribute in the function def attributes. For Unidirectional sequence LSTM, we +can simply map to unidirecional\_sequence\_lstm's +[time major attribute](https://github.com/tensorflow/tensorflow/blob/master/tensorflow/compiler/mlir/lite/ir/tfl_ops.td#L3508). -For a Colab demonstrating these classes, refer to -[TensorFlowLite_LSTM_Keras_Tutorial](https://github.com/tensorflow/tensorflow/blob/master/tensorflow/lite/experimental/examples/lstm/TensorFlowLite_LSTM_Keras_Tutorial.ipynb). +##### BiDirectional LSTM -Note: There is no replacement available for -[tf.compat.v1.nn.rnn_cell.GRUCell](https://www.tensorflow.org/api_docs/python/tf/nn/rnn_cell/GRUCell). +Bidirectional LSTM can be implemented with two Keras LSTM layers, one for +forward and one for backward, see examples +[here](https://github.com/tensorflow/tensorflow/blob/master/tensorflow/python/keras/layers/wrappers.py#L381). +Once we see the go\_backward attribute, we recognize it as backward LSTM, then +we group forward & backward LSTM together. **This is future work.** Currently, +this creates two UnidirectionalSequenceLSTM operators in the TensorFlow Lite +model. + +### User-defined LSTM conversion examples + +TensorFlow Lite also provides a way to convert user defined LSTM +implementations. Here we use Lingvo’s LSTM as an example of how that can be +implemented. For details please refer to the +[lingvo.LSTMCellSimple interface](https://github.com/tensorflow/lingvo/blob/master/lingvo/core/rnn_cell.py#L230) +and the conversion logic +[here](https://github.com/tensorflow/tensorflow/blob/master/tensorflow/compiler/mlir/lite/transforms/prepare_composite_functions_tf.cc#L123). +We also provide an example for another of Lingvo’s LSTM definitions in +[lingvo.LayerNormalizedLSTMCellSimple interface](https://github.com/tensorflow/lingvo/blob/master/lingvo/core/rnn_cell.py#L1179) +and its convertion logic +[here](https://github.com/tensorflow/tensorflow/blob/master/tensorflow/compiler/mlir/lite/transforms/prepare_composite_functions_tf.cc#L130). + +## “Bring your own TensorFlow RNN” to TensorFlow Lite + +If a user's RNN interface is different from the standard supported ones, there +are a couple of options: + +**Option 1:** Write adapter code in TensorFlow python to adapt the RNN interface +to the Keras RNN interface. This means a tf.function with +[tf\_implements annotation](https://github.com/tensorflow/community/pull/113) on +the generated RNN interface’s function that is identical to the one generated by +the Keras LSTM layer. After this, the same conversion API used for Keras LSTM +will work. + +**Option 2:** If the above is not possible (e.g. the Keras LSTM is missing some +functionality that is currently exposed by TensorFlow Lite’s fused LSTM op like +layer normalization), then extend the TensorFlow Lite converter by writing +custom conversion code and plug it into the prepare-composite-functions +MLIR-pass +[here](https://github.com/tensorflow/tensorflow/blob/master/tensorflow/compiler/mlir/lite/transforms/prepare_composite_functions_tf.cc#L108). +The function’s interface should be treated like an API contract and should +contain the arguments needed to convert to fused TensorFlow Lite LSTM +operators - i.e. input, bias, weights, projection, layer normalization, etc. It +is preferable for the tensors passed as arguments to this function to have known +rank (i.e. RankedTensorType in MLIR). This makes it much easier to write +conversion code that can assume these tensors as RankedTensorType and helps +transform them to ranked tensors corresponding to the fused TensorFlow Lite +operator’s operands. + +A complete example of such conversion flow is Lingvo’s LSTMCellSimple to +TensorFlow Lite conversion. + +The LSTMCellSimple in Lingvo is defined +[here](https://github.com/tensorflow/lingvo/blob/master/lingvo/core/rnn_cell.py#L230). +Models trained with this LSTM cell can be converted to TensorFlow Lite as +follows: + +1. Wrap all uses of LSTMCellSimple in a tf.function with a tf\_implements + annotation that is labelled as such (e.g. lingvo.LSTMCellSimple would be a + good annotation name here). Make sure the tf.function that is generated + matches the interface of the function expected in the conversion code. This + is a contract between the model author adding the annotation and the + conversion code. +1. Extend the prepare-composite-functions pass to plug in a custom composite op + to TensorFlow Lite fused LSTM op conversion. See + [LSTMCellSimple](https://github.com/tensorflow/tensorflow/blob/master/tensorflow/compiler/mlir/lite/transforms/prepare_composite_functions_tf.cc#L123) + conversion code. + + The conversion contract: + +1. **Weight** and **projection** tensors are transposed. + +1. The **{input, recurrent}** to **{cell, input gate, forget gate, output + gate}** are extracted by slicing the transposed weight tensor. + +1. The **{bias}** to **{cell, input gate, forget gate, output gate}** are + extracted by slicing the bias tensor. + +1. The **projection** is extracted by slicing the transposed projection tensor. + +1. Similar conversion is written for + [LayerNormalizedLSTMCellSimple](https://github.com/tensorflow/tensorflow/blob/master/tensorflow/compiler/mlir/lite/utils/lstm_utils.cc#L519). + +1. The rest of the TensorFlow Lite conversion infrastructure, including all the + [MLIR passes](https://github.com/tensorflow/tensorflow/blob/master/tensorflow/compiler/mlir/lite/tf_tfl_passes.cc#L58) + defined as well as the final export to TensorFlow Lite flatbuffer can be + reused. + +## Known issues/limitations + +1. Currently there is support only for converting stateless Keras LSTM (default + behavior in Keras). Stateful Keras LSTM conversion is future work. +1. It is still possible to model a stateful Keras LSTM layer using the + underlying stateless Keras LSTM layer and managing the state explicitly in + the user program. Such a TensorFlow program can still be converted to + TensorFlow Lite using the feature being described here. +1. Bidirectional LSTM is currently modelled as two UnidirectionalSequenceLSTM + operators in TensorFlow Lite. This will be replaced with a single + BidirectionalSequenceLSTM op. From 158d4be42d7aea11a395d2f79483ac93289e1bb8 Mon Sep 17 00:00:00 2001 From: Xinyi Wang Date: Thu, 18 Jun 2020 22:15:44 -0700 Subject: [PATCH 275/867] Add get_next_as_optional method for a distributed iterator The function is called on a distributed iterator and returns an `Optional` that contains the next value, the PerReplica input, from Distributed iterator or no value if this `iterator` has reached the end of the sequence. PiperOrigin-RevId: 317248910 Change-Id: Ide217da1aff1d62f8d0d8f43423be2d859d933d3 --- .../custom_training_loop_input_test.py | 49 ++++++++++- .../python/distribute/distribute_lib.py | 4 + tensorflow/python/distribute/input_lib.py | 60 +++++++++++++ .../python/distribute/input_lib_test.py | 86 +++++++++++++------ ...low.distribute.-distributed-iterator.pbtxt | 4 + 5 files changed, 177 insertions(+), 26 deletions(-) diff --git a/tensorflow/python/distribute/custom_training_loop_input_test.py b/tensorflow/python/distribute/custom_training_loop_input_test.py index e4f782810dd..5660b5839ce 100644 --- a/tensorflow/python/distribute/custom_training_loop_input_test.py +++ b/tensorflow/python/distribute/custom_training_loop_input_test.py @@ -30,6 +30,7 @@ from tensorflow.python.distribute import strategy_combinations from tensorflow.python.eager import def_function from tensorflow.python.eager import test from tensorflow.python.framework import constant_op +from tensorflow.python.framework import dtypes from tensorflow.python.framework import errors from tensorflow.python.framework import ops from tensorflow.python.ops import array_ops @@ -136,8 +137,52 @@ class InputIterationTest(test.TestCase, parameterized.TestCase, @combinations.generate( combinations.combine( - distribution=strategy_combinations.tpu_strategies, - mode=["eager"])) + distribution=strategy_combinations.all_strategies, mode=["eager"])) + def testGetNextAsOptional(self, distribution): + data = [5., 6., 7., 8.] + dataset = get_dataset_from_tensor_slices(data).batch(2) + dist_dataset = distribution.experimental_distribute_dataset(dataset) + iterator = iter(dist_dataset) + + def train_step(data): + return math_ops.square(data) + + @def_function.function + def run(iterator): + return distribution.experimental_local_results( + distribution.run( + train_step, args=(iterator.get_next_as_optional().get_value(),))) + + self.assert_equal_flattened([[25., 36.]], [run(iterator)]) + + @combinations.generate( + combinations.combine( + distribution=strategy_combinations.all_strategies, mode=["eager"])) + def testGetNextAsOptionalExampleUsage(self, distribution): + global_batch_size = 2 + steps_per_loop = 6 + dataset = dataset_ops.Dataset.range( + 8, output_type=dtypes.int32).batch(global_batch_size) + distributed_iterator = iter( + distribution.experimental_distribute_dataset(dataset)) + + @def_function.function + def train_fn(distributed_iterator): + + def step_fn(x): + return x + + for _ in math_ops.range(steps_per_loop): + optional_data = distributed_iterator.get_next_as_optional() + if not optional_data.has_value(): + break + distribution.run(step_fn, args=(optional_data.get_value(),)) + + train_fn(distributed_iterator) + + @combinations.generate( + combinations.combine( + distribution=strategy_combinations.tpu_strategies, mode=["eager"])) def testFullEagerTPU(self, distribution): dataset = get_dataset_from_tensor_slices([5., 6., 7., 8.]).batch(2) diff --git a/tensorflow/python/distribute/distribute_lib.py b/tensorflow/python/distribute/distribute_lib.py index b6a89463426..ec0b911ebe0 100644 --- a/tensorflow/python/distribute/distribute_lib.py +++ b/tensorflow/python/distribute/distribute_lib.py @@ -200,6 +200,7 @@ import six from tensorflow.python.autograph.core import ag_ctx as autograph_ctx from tensorflow.python.autograph.impl import api as autograph from tensorflow.python.data.ops import dataset_ops +from tensorflow.python.data.ops import iterator_ops from tensorflow.python.distribute import collective_util from tensorflow.python.distribute import device_util from tensorflow.python.distribute import distribution_strategy_context @@ -2879,6 +2880,9 @@ class _DefaultDistributionExtended(StrategyExtendedV1): def get_next(self): return self._iterator.get_next() + def get_next_as_optional(self): + return iterator_ops.get_next_as_optional(self._iterator) + @deprecated(None, "Use the iterator's `initializer` property instead.") def initialize(self): """Initialize underlying iterators. diff --git a/tensorflow/python/distribute/input_lib.py b/tensorflow/python/distribute/input_lib.py index ff468af7f87..e4a362a92c6 100644 --- a/tensorflow/python/distribute/input_lib.py +++ b/tensorflow/python/distribute/input_lib.py @@ -29,6 +29,7 @@ from tensorflow.python.data.experimental.ops import batching from tensorflow.python.data.experimental.ops import distribute from tensorflow.python.data.ops import dataset_ops from tensorflow.python.data.ops import multi_device_iterator_ops +from tensorflow.python.data.ops import optional_ops from tensorflow.python.distribute import device_util from tensorflow.python.distribute import distribute_utils from tensorflow.python.distribute import distribution_strategy_context @@ -235,6 +236,40 @@ class DistributedIteratorInterface(collections.Iterator, raise NotImplementedError( "DistributedIterator.element_spec() must be implemented in descendants") + def get_next_as_optional(self): + """Returns a `tf.experimental.Optional` that contains the next value for all replicas. + + If the `tf.distribute.DistributedIterator` has reached the end of the + sequence, the returned `tf.experimental.Optional` will have no value. + + Example usage: + + >>> strategy = tf.distribute.MirroredStrategy() + >>> global_batch_size = 2 + >>> steps_per_loop = 2 + >>> dataset = tf.data.Dataset.range(10).batch(global_batch_size) + >>> distributed_iterator = iter( + ... strategy.experimental_distribute_dataset(dataset)) + >>> def step_fn(x): + ... return x + >>> @tf.function + ... def train_fn(distributed_iterator): + ... for _ in tf.range(steps_per_loop): + ... optional_data = distributed_iterator.get_next_as_optional() + ... if not optional_data.has_value(): + ... break + ... tf.print(strategy.run(step_fn, args=(optional_data.get_value(),))) + >>> train_fn(distributed_iterator) + ... # ([0 1],) + ... # ([2 3],) + + Returns: + An `tf.experimental.Optional` object representing the next value from the + `tf.distribute.DistributedIterator` (if it has one) or no value. + """ + raise NotImplementedError( + "get_next_as_optional() not implemented in descendants") + @tf_export("distribute.DistributedDataset", v1=[]) class DistributedDatasetInterface(collections.Iterable, @@ -622,6 +657,31 @@ class DistributedIteratorBase(DistributedIteratorInterface): def __iter__(self): return self + def get_next_as_optional(self): + global_has_value, replicas = _get_next_as_optional(self, self._strategy) + + def return_none(): + return optional_ops.Optional.empty(self._element_spec) + + def return_value(replicas): + """Wraps the inputs for replicas in an `tf.experimental.Optional`.""" + results = [] + for i, worker in enumerate(self._input_workers.worker_devices): + with ops.device(worker): + devices = self._input_workers.compute_devices_for_worker(i) + for j, device in enumerate(devices): + with ops.device(device): + result = replicas[i][j] + results.append(result) + replicas = results + + return optional_ops.Optional.from_value( + distribute_utils.regroup(replicas)) + + return control_flow_ops.cond(global_has_value, + lambda: return_value(replicas), + lambda: return_none()) # pylint: disable=unnecessary-lambda + def get_next(self, name=None): """Returns the next input from the iterator for all replicas.""" if not self._enable_get_next_as_optional: diff --git a/tensorflow/python/distribute/input_lib_test.py b/tensorflow/python/distribute/input_lib_test.py index ff4436c4c8c..7f02d0121d0 100644 --- a/tensorflow/python/distribute/input_lib_test.py +++ b/tensorflow/python/distribute/input_lib_test.py @@ -185,38 +185,76 @@ class DistributedIteratorTestBase(test.TestCase): if not ops.executing_eagerly_outside_functions(): evaluate(control_flow_ops.group(iterator.initializer)) - for expected_value in expected_values: - next_element = iterator.get_next() - computed_value = evaluate( - [distribute_utils.select_replica(r, next_element) - for r in range(len(devices))]) - self.assertEqual(len(expected_value), len(computed_value)) - for i in range(len(expected_value)): - self.assertAllEqual(expected_value[i], computed_value[i]) + def test_get_next(iterator): + for expected_value in expected_values: + next_element = iterator.get_next() + computed_value = evaluate([ + distribute_utils.select_replica(r, next_element) + for r in range(len(devices)) + ]) - with self.assertRaises(errors.OutOfRangeError): - next_element = iterator.get_next() - evaluate( - [distribute_utils.select_replica(r, next_element) - for r in range(len(devices))]) + self.assertEqual(len(expected_value), len(computed_value)) + for i in range(len(expected_value)): + self.assertAllEqual(expected_value[i], computed_value[i]) - # After re-initializing the iterator, should be able to iterate again. - if not ops.executing_eagerly_outside_functions(): - evaluate(control_flow_ops.group(iterator.initializer)) + with self.assertRaises(errors.OutOfRangeError): + next_element = iterator.get_next() + evaluate([ + distribute_utils.select_replica(r, next_element) + for r in range(len(devices)) + ]) + + # After re-initializing the iterator, should be able to iterate again. + if not ops.executing_eagerly_outside_functions(): + evaluate(control_flow_ops.group(iterator.initializer)) + else: + if api_type == "wrap_into_iterator": + self.skipTest("unsupported test combination") + else: + iterator = iter(dataset) + + for expected_value in expected_values: + next_element = iterator.get_next() + computed_value = evaluate([ + distribute_utils.select_replica(r, next_element) + for r in range(len(devices)) + ]) + self.assertEqual(len(expected_value), len(computed_value)) + for i in range(len(expected_value)): + self.assertAllEqual(expected_value[i], computed_value[i]) + + def test_get_next_as_optional(iterator): + for expected_value in expected_values: + next_element = iterator.get_next_as_optional() + computed_value = evaluate([ + distribute_utils.select_replica(r, next_element.get_value()) + for r in range(len(devices)) + ]) + + self.assertEqual(len(expected_value), len(computed_value)) + for i in range(len(expected_value)): + self.assertAllEqual(expected_value[i], computed_value[i]) + + next_element = iterator.get_next_as_optional() + self.assertFalse(self.evaluate(next_element.has_value())) + with self.assertRaises(errors.InvalidArgumentError): + evaluate([ + distribute_utils.select_replica(r, next_element.get_value()) + for r in range(len(devices)) + ]) + + test_get_next(iterator) + + # re-initializing the iterator + if not tf2.enabled(): + self.skipTest("Not testing get_next_as_optional in TF1") else: if api_type == "wrap_into_iterator": self.skipTest("unsupported test combination") else: iterator = iter(dataset) - for expected_value in expected_values: - next_element = iterator.get_next() - computed_value = evaluate( - [distribute_utils.select_replica(r, next_element) - for r in range(len(devices))]) - self.assertEqual(len(expected_value), len(computed_value)) - for i in range(len(expected_value)): - self.assertAllEqual(expected_value[i], computed_value[i]) + test_get_next_as_optional(iterator) if iteration_type == "for_loop" and context.executing_eagerly(): actual_values = [] diff --git a/tensorflow/tools/api/golden/v2/tensorflow.distribute.-distributed-iterator.pbtxt b/tensorflow/tools/api/golden/v2/tensorflow.distribute.-distributed-iterator.pbtxt index f712d9058b9..47899cc4188 100644 --- a/tensorflow/tools/api/golden/v2/tensorflow.distribute.-distributed-iterator.pbtxt +++ b/tensorflow/tools/api/golden/v2/tensorflow.distribute.-distributed-iterator.pbtxt @@ -13,4 +13,8 @@ tf_class { name: "get_next" argspec: "args=[\'self\'], varargs=None, keywords=None, defaults=None" } + member_method { + name: "get_next_as_optional" + argspec: "args=[\'self\'], varargs=None, keywords=None, defaults=None" + } } From 539e9cb3a22793aad5d2df885e016f43b81a6a9f Mon Sep 17 00:00:00 2001 From: Meghna Natraj Date: Thu, 18 Jun 2020 22:33:44 -0700 Subject: [PATCH 276/867] Update quantization docs to use TFLiteConverter.from_saved_model() API instead of .from_keras_model() API PiperOrigin-RevId: 317251205 Change-Id: Ia8166decfa76327e3fd44871b194ffcae0f049f8 --- .../lite/g3doc/performance/post_training_quantization.md | 8 ++++---- 1 file changed, 4 insertions(+), 4 deletions(-) diff --git a/tensorflow/lite/g3doc/performance/post_training_quantization.md b/tensorflow/lite/g3doc/performance/post_training_quantization.md index ac584dd4c1c..dcf251e6d3d 100644 --- a/tensorflow/lite/g3doc/performance/post_training_quantization.md +++ b/tensorflow/lite/g3doc/performance/post_training_quantization.md @@ -34,7 +34,7 @@ weights from floating point to integer, which has 8-bits of precision: 
 import tensorflow as tf
-converter = tf.lite.TFLiteConverter.from_keras_model(keras_model)
+converter = tf.lite.TFLiteConverter.from_saved_model(saved_model_dir)
 converter.optimizations = [tf.lite.Optimize.DEFAULT]
 tflite_quant_model = converter.convert()
@@ -68,7 +68,7 @@ the following steps: 
 import tensorflow as tf
-converter = tf.lite.TFLiteConverter.from_keras_model(keras_model)
+converter = tf.lite.TFLiteConverter.from_saved_model(saved_model_dir)
 converter.optimizations = [tf.lite.Optimize.DEFAULT]
 def representative_dataset_gen():
   for _ in range(num_calibration_steps):
@@ -96,7 +96,7 @@ the following steps:
 
 
 import tensorflow as tf
-converter = tf.lite.TFLiteConverter.from_keras_model(keras_model)
+converter = tf.lite.TFLiteConverter.from_saved_model(saved_model_dir)
 converter.optimizations = [tf.lite.Optimize.DEFAULT]
 def representative_dataset_gen():
   for _ in range(num_calibration_steps):
@@ -120,7 +120,7 @@ quantization of weights, use the following steps:
 
 
 import tensorflow as tf
-converter = tf.lite.TFLiteConverter.from_keras_model(keras_model)
+converter = tf.lite.TFLiteConverter.from_saved_model(saved_model_dir)
 converter.optimizations = [tf.lite.Optimize.DEFAULT]
 converter.target_spec.supported_types = [tf.float16]
 tflite_quant_model = converter.convert()

From 8e654afea4adba36b94b0f7a3d33a23e788612e0 Mon Sep 17 00:00:00 2001
From: "A. Unique TensorFlower" 
Date: Thu, 18 Jun 2020 23:35:25 -0700
Subject: [PATCH 277/867] tf.numpy: Improve ndarray.__getitem__ to match numpy
 semantics.

PiperOrigin-RevId: 317256717
Change-Id: Ie89b81689f96242e3e9b01568e13937b80aaffc7
---
 .../python/ops/numpy_ops/np_array_ops.py      | 261 ++++++++++++++++--
 tensorflow/python/ops/numpy_ops/np_arrays.py  | 137 ---------
 2 files changed, 245 insertions(+), 153 deletions(-)

diff --git a/tensorflow/python/ops/numpy_ops/np_array_ops.py b/tensorflow/python/ops/numpy_ops/np_array_ops.py
index 906e53c556d..47236d45561 100644
--- a/tensorflow/python/ops/numpy_ops/np_array_ops.py
+++ b/tensorflow/python/ops/numpy_ops/np_array_ops.py
@@ -20,12 +20,15 @@ from __future__ import division
 from __future__ import print_function
 
 import math
+import numbers
+from typing import Sequence
 import numpy as np
 import six
 
 from tensorflow.python.framework import constant_op
 from tensorflow.python.framework import dtypes
 from tensorflow.python.framework import ops
+from tensorflow.python.framework import tensor_shape
 from tensorflow.python.ops import array_ops
 from tensorflow.python.ops import clip_ops
 from tensorflow.python.ops import control_flow_ops
@@ -164,9 +167,11 @@ def full_like(a, fill_value, dtype=None, order='K', subok=True, shape=None):  #
 @np_utils.np_doc_only(np.array)
 def array(val, dtype=None, copy=True, ndmin=0):  # pylint: disable=redefined-outer-name
   """Since Tensors are immutable, a copy is made only if val is placed on a
+
   different device than the current one. Even if `copy` is False, a new Tensor
   may need to be built to satisfy `dtype` and `ndim`. This is used only if `val`
-  is an ndarray or a Tensor."""  # pylint:disable=g-docstring-missing-newline
+  is an ndarray or a Tensor.
+  """  # pylint:disable=g-docstring-missing-newline
   if dtype:
     dtype = np_utils.result_type(dtype)
   if isinstance(val, np_arrays.ndarray):
@@ -215,6 +220,8 @@ def array(val, dtype=None, copy=True, ndmin=0):  # pylint: disable=redefined-out
   result_t = np_utils.cond(
       np_utils.greater(ndmin, ndims), true_fn, lambda: result_t)
   return np_arrays.tensor_to_ndarray(result_t)
+
+
 # pylint: enable=g-short-docstring-punctuation,g-no-space-after-docstring-summary,g-doc-return-or-yield,g-doc-args
 
 
@@ -1446,14 +1453,13 @@ def take_along_axis(arr, indices, axis):  # pylint: disable=missing-docstring
   # broadcast.
   arr_shape_original = array_ops.shape(arr)
   indices_shape_original = array_ops.shape(indices)
-  arr_shape = array_ops.tensor_scatter_update(
-      arr_shape_original, [[axis]], [1])
-  indices_shape = array_ops.tensor_scatter_update(
-      indices_shape_original, [[axis]], [1])
-  broadcasted_shape = array_ops.broadcast_dynamic_shape(
-      arr_shape, indices_shape)
-  arr_shape = array_ops.tensor_scatter_update(
-      broadcasted_shape, [[axis]], [arr_shape_original[axis]])
+  arr_shape = array_ops.tensor_scatter_update(arr_shape_original, [[axis]], [1])
+  indices_shape = array_ops.tensor_scatter_update(indices_shape_original,
+                                                  [[axis]], [1])
+  broadcasted_shape = array_ops.broadcast_dynamic_shape(arr_shape,
+                                                        indices_shape)
+  arr_shape = array_ops.tensor_scatter_update(broadcasted_shape, [[axis]],
+                                              [arr_shape_original[axis]])
   indices_shape = array_ops.tensor_scatter_update(
       broadcasted_shape, [[axis]], [indices_shape_original[axis]])
   arr = array_ops.broadcast_to(arr, arr_shape)
@@ -1468,10 +1474,10 @@ def take_along_axis(arr, indices, axis):  # pylint: disable=missing-docstring
   swapaxes_ = lambda t: swapaxes(np_utils.tensor_to_ndarray(t), axis, -1).data
 
   dont_move_axis_to_end = math_ops.equal(axis, rank - 1)
-  arr = np_utils.cond(
-      dont_move_axis_to_end, lambda: arr, lambda: swapaxes_(arr))
-  indices = np_utils.cond(
-      dont_move_axis_to_end, lambda: indices, lambda: swapaxes_(indices))
+  arr = np_utils.cond(dont_move_axis_to_end, lambda: arr,
+                      lambda: swapaxes_(arr))
+  indices = np_utils.cond(dont_move_axis_to_end, lambda: indices,
+                          lambda: swapaxes_(indices))
 
   arr_shape = array_ops.shape(arr)
   arr = array_ops.reshape(arr, [-1, arr_shape[-1]])
@@ -1481,8 +1487,231 @@ def take_along_axis(arr, indices, axis):  # pylint: disable=missing-docstring
 
   result = array_ops.gather(arr, indices, batch_dims=1)
   result = array_ops.reshape(result, indices_shape)
-  result = np_utils.cond(
-      dont_move_axis_to_end, lambda: result, lambda: swapaxes_(result))
+  result = np_utils.cond(dont_move_axis_to_end, lambda: result,
+                         lambda: swapaxes_(result))
   result.set_shape(possible_result_shape)
 
-  return  np_utils.tensor_to_ndarray(result)
+  return np_utils.tensor_to_ndarray(result)
+
+
+_SLICE_ERORR = (
+    'only integers, slices (`:`), ellipsis (`...`), '
+    'numpy.newaxis (`None`) and integer or boolean arrays are valid indices')
+
+
+def _as_index(idx, need_scalar=True):
+  """Helper function to parse idx as an index.
+
+  Args:
+    idx: index
+    need_scalar: If idx needs to be a scalar value.
+
+  Returns:
+    A pair, (indx, bool). First one is the parsed index and can be a tensor,
+    or scalar integer / Dimension. Second one is True if rank is known to be 0.
+
+  Raises:
+    IndexError: For incorrect indices.
+  """
+  if isinstance(idx, (numbers.Integral, tensor_shape.Dimension)):
+    return idx, True
+  data = asarray(idx).data
+  if data.dtype == dtypes.bool:
+    if data.shape.ndims != 1:
+      # TODO(agarwal): handle higher rank boolean masks.
+      raise NotImplementedError('Need rank 1 for bool index %s' % idx)
+    data = array_ops.where_v2(data)
+    data = array_ops.reshape(data, [-1])
+  if need_scalar and data.shape.rank not in (None, 0):
+    raise IndexError(_SLICE_ERORR + ', got {!r}'.format(idx))
+  np_dtype = data.dtype.as_numpy_dtype
+  if not np.issubdtype(np_dtype, np.integer):
+    raise IndexError(_SLICE_ERORR + ', got {!r}'.format(idx))
+  if data.dtype not in (dtypes.int64, dtypes.int32):
+    # TF slicing can only handle int32/int64. So we need to cast.
+    promoted_dtype = np.promote_types(np.int32, np_dtype)
+    if promoted_dtype == np.int32:
+      data = math_ops.cast(data, dtypes.int32)
+    elif promoted_dtype == np.int64:
+      data = math_ops.cast(data, dtypes.int64)
+    else:
+      raise IndexError(_SLICE_ERORR + ', got {!r}'.format(idx))
+  return data, data.shape.rank == 0
+
+
+def _slice_helper(tensor, slice_spec):
+  """Helper function for __getitem__."""
+  begin, end, strides = [], [], []
+  new_axis_mask, shrink_axis_mask = 0, 0
+  begin_mask, end_mask = 0, 0
+  ellipsis_mask = 0
+  advanced_indices = []
+  shrink_indices = []
+  for index, s in enumerate(slice_spec):
+    if isinstance(s, slice):
+      if s.start is not None:
+        begin.append(_as_index(s.start)[0])
+      else:
+        begin.append(0)
+        begin_mask |= (1 << index)
+      if s.stop is not None:
+        end.append(_as_index(s.stop)[0])
+      else:
+        end.append(0)
+        end_mask |= (1 << index)
+      if s.step is not None:
+        strides.append(_as_index(s.step)[0])
+      else:
+        strides.append(1)
+    elif s is Ellipsis:
+      begin.append(0)
+      end.append(0)
+      strides.append(1)
+      ellipsis_mask |= (1 << index)
+    elif s is array_ops.newaxis:
+      begin.append(0)
+      end.append(0)
+      strides.append(1)
+      new_axis_mask |= (1 << index)
+    else:
+      s, is_scalar = _as_index(s, False)
+      if is_scalar:
+        begin.append(s)
+        end.append(s + 1)
+        strides.append(1)
+        shrink_axis_mask |= (1 << index)
+        shrink_indices.append(index)
+      else:
+        begin.append(0)
+        end.append(0)
+        strides.append(1)
+        begin_mask |= (1 << index)
+        end_mask |= (1 << index)
+        advanced_indices.append((index, s, ellipsis_mask != 0))
+
+  # stack possibly involves no tensors, so we must use op_scope correct graph.
+  with ops.name_scope(
+      None,
+      'strided_slice', [tensor] + begin + end + strides,
+      skip_on_eager=False) as name:
+    if begin:
+      packed_begin, packed_end, packed_strides = (array_ops.stack(begin),
+                                                  array_ops.stack(end),
+                                                  array_ops.stack(strides))
+      if (packed_begin.dtype == dtypes.int64 or
+          packed_end.dtype == dtypes.int64 or
+          packed_strides.dtype == dtypes.int64):
+        if packed_begin.dtype != dtypes.int64:
+          packed_begin = math_ops.cast(packed_begin, dtypes.int64)
+        if packed_end.dtype != dtypes.int64:
+          packed_end = math_ops.cast(packed_end, dtypes.int64)
+        if packed_strides.dtype != dtypes.int64:
+          packed_strides = math_ops.cast(packed_strides, dtypes.int64)
+    else:
+      var_empty = constant_op.constant([], dtype=dtypes.int32)
+      packed_begin = packed_end = packed_strides = var_empty
+    # TODO(agarwal): set_shape on tensor to set rank.
+    tensor = array_ops.strided_slice(
+        tensor,
+        packed_begin,
+        packed_end,
+        packed_strides,
+        begin_mask=begin_mask,
+        end_mask=end_mask,
+        shrink_axis_mask=shrink_axis_mask,
+        new_axis_mask=new_axis_mask,
+        ellipsis_mask=ellipsis_mask,
+        name=name)
+    if not advanced_indices:
+      return tensor
+    advanced_indices_map = {}
+    for index, data, had_ellipsis in advanced_indices:
+      if had_ellipsis:
+        num_shrink = len([x for x in shrink_indices if x > index])
+        dim = index - len(slice_spec) + num_shrink
+      else:
+        num_shrink = len([x for x in shrink_indices if x < index])
+        dim = index - num_shrink
+      advanced_indices_map[dim] = data
+    dims = sorted(advanced_indices_map.keys())
+    dims_contiguous = True
+    if len(dims) > 1:
+      if dims[0] < 0 and dims[-1] >= 0:  # not all same sign
+        dims_contiguous = False
+      else:
+        for i in range(len(dims) - 1):
+          if dims[i] + 1 != dims[i + 1]:
+            dims_contiguous = False
+            break
+    indices = [advanced_indices_map[x] for x in dims]
+    indices = [x.data for x in _promote_dtype(*indices)]
+    indices = np_utils.tf_broadcast(*indices)
+    stacked_indices = array_ops.stack(indices, axis=-1)
+    if not dims_contiguous:
+      tensor = moveaxis(tensor, dims, range(len(dims))).data
+      tensor_shape_prefix = array_ops.shape(
+          tensor, out_type=stacked_indices.dtype)[:len(dims)]
+      stacked_indices = array_ops.where_v2(
+          stacked_indices < 0, stacked_indices + tensor_shape_prefix,
+          stacked_indices)
+      return array_ops.gather_nd(tensor, stacked_indices)
+    # Note that gather_nd does not support gathering from inside the array.
+    # To avoid shuffling data back and forth, we transform the indices and
+    # do a gather instead.
+    rank = np_utils._maybe_static(array_ops.rank(tensor))  # pylint: disable=protected-access
+    dims = [(x + rank if x < 0 else x) for x in dims]
+    shape_tensor = array_ops.shape(tensor, out_type=stacked_indices.dtype)
+    dim_sizes = array_ops.gather(shape_tensor, dims)
+    if len(dims) == 1:
+      stacked_indices = indices[0]
+    stacked_indices = array_ops.where_v2(stacked_indices < 0,
+                                         stacked_indices + dim_sizes,
+                                         stacked_indices)
+    axis = dims[0]
+    if len(dims) > 1:
+      index_scaling = math_ops.cumprod(
+          dim_sizes, reverse=True, exclusive=True)
+      stacked_indices = math_ops.tensordot(
+          stacked_indices, index_scaling, axes=1)
+      flat_shape = array_ops.concat(
+          [shape_tensor[:axis], [-1], shape_tensor[axis + len(dims):]],
+          axis=0)
+      tensor = array_ops.reshape(tensor, flat_shape)
+
+    return array_ops.gather(tensor, stacked_indices, axis=axis)
+
+
+def _as_spec_tuple(slice_spec):
+  """Convert slice_spec to tuple."""
+  if isinstance(slice_spec,
+                Sequence) and not isinstance(slice_spec, np.ndarray):
+    is_index = True
+    for s in slice_spec:
+      if s is None or s is Ellipsis or isinstance(s, (Sequence, slice)):
+        is_index = False
+        break
+      elif isinstance(s, (np_arrays.ndarray, np.ndarray)) and s.ndim != 0:
+        is_index = False
+        break
+    if not is_index:
+      return tuple(slice_spec)
+  return (slice_spec,)
+
+
+def _getitem(self, slice_spec):
+  """Implementation of ndarray.__getitem__."""
+  if (isinstance(slice_spec, bool) or (isinstance(slice_spec, ops.Tensor) and
+                                       slice_spec.dtype == dtypes.bool) or
+      (isinstance(slice_spec, (np.ndarray, np_arrays.ndarray)) and
+       slice_spec.dtype == np.bool)):
+    return np_utils.tensor_to_ndarray(
+        array_ops.boolean_mask(tensor=self.data, mask=slice_spec))
+
+  if not isinstance(slice_spec, tuple):
+    slice_spec = _as_spec_tuple(slice_spec)
+
+  result_t = _slice_helper(self.data, slice_spec)
+  return np_utils.tensor_to_ndarray(result_t)
+
+
+setattr(np_arrays.ndarray, '__getitem__', _getitem)
diff --git a/tensorflow/python/ops/numpy_ops/np_arrays.py b/tensorflow/python/ops/numpy_ops/np_arrays.py
index 8bec8a469a2..88bf4e7499a 100644
--- a/tensorflow/python/ops/numpy_ops/np_arrays.py
+++ b/tensorflow/python/ops/numpy_ops/np_arrays.py
@@ -20,138 +20,17 @@ from __future__ import absolute_import
 from __future__ import division
 from __future__ import print_function
 
-import numbers
 import numpy as np
 import six
 
 from tensorflow.python.framework import composite_tensor
-from tensorflow.python.framework import constant_op
 from tensorflow.python.framework import dtypes
 from tensorflow.python.framework import ops
-from tensorflow.python.framework import tensor_shape
 from tensorflow.python.framework import tensor_spec
 from tensorflow.python.framework import type_spec
 from tensorflow.python.ops import array_ops
 from tensorflow.python.ops import math_ops
 from tensorflow.python.ops.numpy_ops import np_dtypes
-from tensorflow.python.util import nest
-
-
-_SLICE_TYPE_ERROR = (
-    'Only integers, slices (`:`), ellipsis (`...`), '
-    'tf.newaxis (`None`) and scalar tf.int32/tf.int64 tensors are valid '
-    'indices')
-
-_SUPPORTED_SLICE_DTYPES = (dtypes.int32, dtypes.int32_ref, dtypes.int64,
-                           dtypes.int64_ref)
-
-
-def _check_index(idx):
-  """Check if a given value is a valid index into a tensor."""
-  if isinstance(idx, (numbers.Integral, tensor_shape.Dimension)):
-    return
-
-  # Optimistic check. Assumptions:
-  # * any object with a dtype is supported
-  # * any object with a dtype has a sizeable shape attribute.
-  dtype = getattr(idx, 'dtype', None)
-  if (dtype is None or dtypes.as_dtype(dtype) not in _SUPPORTED_SLICE_DTYPES or
-      idx.shape and len(idx.shape) == 1):
-    # TODO(slebedev): IndexError seems more appropriate here, but it
-    # will break `_slice_helper` contract.
-    raise TypeError(_SLICE_TYPE_ERROR + ', got {!r}'.format(idx))
-
-
-def _is_undefined_dimension(d):
-  return isinstance(d, tensor_shape.Dimension) and d.value is None
-
-
-def _slice_helper(tensor, slice_spec, var=None):
-  """Copied from array_ops._slice_helper, will be merged back later."""
-  if isinstance(slice_spec, bool) or \
-  (isinstance(slice_spec, ops.Tensor) and slice_spec.dtype == dtypes.bool) or \
-  (isinstance(slice_spec, np.ndarray) and slice_spec.dtype == bool):
-    return array_ops.boolean_mask(tensor=tensor, mask=slice_spec)
-
-  if not isinstance(slice_spec, (list, tuple)):
-    slice_spec = [slice_spec]
-
-  begin, end, strides = [], [], []
-  index = 0
-
-  new_axis_mask, shrink_axis_mask = 0, 0
-  begin_mask, end_mask = 0, 0
-  ellipsis_mask = 0
-  for s in slice_spec:
-    if isinstance(s, slice):
-      if s.start is not None and not _is_undefined_dimension(s.start):
-        _check_index(s.start)
-        begin.append(s.start)
-      else:
-        begin.append(0)
-        begin_mask |= (1 << index)
-      if s.stop is not None and not _is_undefined_dimension(s.stop):
-        _check_index(s.stop)
-        end.append(s.stop)
-      else:
-        end.append(0)
-        end_mask |= (1 << index)
-      if s.step is not None and not _is_undefined_dimension(s.step):
-        _check_index(s.step)
-        strides.append(s.step)
-      else:
-        strides.append(1)
-    elif s is Ellipsis:
-      begin.append(0)
-      end.append(0)
-      strides.append(1)
-      ellipsis_mask |= (1 << index)
-    elif s is array_ops.newaxis:
-      begin.append(0)
-      end.append(0)
-      strides.append(1)
-      new_axis_mask |= (1 << index)
-    else:
-      _check_index(s)
-      begin.append(s)
-      end.append(s + 1)
-      strides.append(1)
-      shrink_axis_mask |= (1 << index)
-    index += 1
-
-  # stack possibly involves no tensors, so we must use op_scope correct graph.
-  with ops.name_scope(
-      None,
-      'strided_slice', [tensor] + begin + end + strides,
-      skip_on_eager=False) as name:
-    if begin:
-      packed_begin, packed_end, packed_strides = (array_ops.stack(begin),
-                                                  array_ops.stack(end),
-                                                  array_ops.stack(strides))
-      if (packed_begin.dtype == dtypes.int64 or
-          packed_end.dtype == dtypes.int64 or
-          packed_strides.dtype == dtypes.int64):
-        if packed_begin.dtype != dtypes.int64:
-          packed_begin = math_ops.cast(packed_begin, dtypes.int64)
-        if packed_end.dtype != dtypes.int64:
-          packed_end = math_ops.cast(packed_end, dtypes.int64)
-        if packed_strides.dtype != dtypes.int64:
-          packed_strides = math_ops.cast(packed_strides, dtypes.int64)
-    else:
-      var_empty = constant_op.constant([], dtype=dtypes.int32)
-      packed_begin = packed_end = packed_strides = var_empty
-    return array_ops.strided_slice(
-        tensor,
-        packed_begin,
-        packed_end,
-        packed_strides,
-        begin_mask=begin_mask,
-        end_mask=end_mask,
-        shrink_axis_mask=shrink_axis_mask,
-        new_axis_mask=new_axis_mask,
-        ellipsis_mask=ellipsis_mask,
-        var=var,
-        name=name)
 
 
 def convert_to_tensor(value, dtype=None, dtype_hint=None):
@@ -361,22 +240,6 @@ class ndarray(composite_tensor.CompositeTensor):  # pylint: disable=invalid-name
   def __bool__(self):
     return self.__nonzero__()
 
-  def __getitem__(self, slice_spec):
-    # TODO(srbs): Need to support better indexing.
-    def _gettensor(x):
-      if isinstance(x, ndarray):
-        x = x.data
-      if isinstance(x, ops.Tensor) and x.dtype not in (
-          dtypes.int32, dtypes.int64):
-        # Currently _slice_helper will only work with int32/int64 tensors, but
-        # type inference by numpy can create {u,}int{8,16}, so just cast.
-        x = math_ops.cast(x, dtypes.int32)
-      return x
-    slice_spec = nest.map_structure(_gettensor, slice_spec)
-
-    result_t = _slice_helper(self.data, slice_spec)
-    return tensor_to_ndarray(result_t)
-
   def __iter__(self):
     if not isinstance(self.data, ops.EagerTensor):
       raise TypeError('Iteration over symbolic tensor is not allowed')

From e972c5572634efd188696038e9241b75cdcd69bc Mon Sep 17 00:00:00 2001
From: Gaurav Jain 
Date: Fri, 19 Jun 2020 00:07:20 -0700
Subject: [PATCH 278/867] Add uint32 & uint64 to TF_CALL_INTEGRAL_TYPES

Both uint32 & uint64 had been omitted from TF_CALL_INTEGRAL_TYPES due to
suggested concerns of size bloat. In reality it seems that the size
increase is only around 2MB. Further, this fixes #39649 since we are no
longer inadvertently using the XLA_CPU device to perform tf.reduce_mean.

PiperOrigin-RevId: 317259372
Change-Id: Iacf75eaedce198fbef4bd9fd59b6fefa584cbf34
---
 tensorflow/core/framework/register_types.h    | 21 +++++---------
 tensorflow/core/framework/types.cc            |  5 ----
 tensorflow/core/kernels/BUILD                 |  2 ++
 tensorflow/core/kernels/concat_lib_cpu.cc     |  2 --
 tensorflow/core/kernels/concat_op.cc          |  2 --
 tensorflow/core/kernels/constant_op.cc        |  1 -
 tensorflow/core/kernels/control_flow_ops.cc   |  5 ----
 .../core/kernels/data/dataset_test_base.cc    |  2 --
 tensorflow/core/kernels/dense_update_ops.cc   |  1 -
 .../core/kernels/dynamic_partition_op.cc      |  2 --
 tensorflow/core/kernels/fill_functor.cc       |  5 +++-
 tensorflow/core/kernels/gather_op.cc          |  2 --
 tensorflow/core/kernels/identity_op.cc        |  1 -
 tensorflow/core/kernels/ragged_gather_op.cc   |  2 --
 .../kernels/ragged_tensor_from_variant_op.cc  |  2 --
 .../kernels/ragged_tensor_to_tensor_op.cc     |  2 --
 .../kernels/ragged_tensor_to_variant_op.cc    |  2 --
 .../core/kernels/resource_variable_ops.cc     |  1 -
 tensorflow/core/kernels/split_lib_cpu.cc      |  1 -
 tensorflow/core/kernels/split_op.cc           |  1 -
 tensorflow/core/kernels/strided_slice_op.cc   |  2 --
 .../core/kernels/strided_slice_op_impl.h      |  2 --
 tensorflow/core/kernels/topk_op.cc            |  2 --
 .../core/kernels/topk_op_gpu_uint32.cu.cc     | 28 +++++++++++++++++++
 .../core/kernels/topk_op_gpu_uint64.cu.cc     | 28 +++++++++++++++++++
 tensorflow/core/util/batch_util.cc            |  8 ------
 .../core/util/saved_tensor_slice_util.h       |  2 ++
 27 files changed, 71 insertions(+), 63 deletions(-)
 create mode 100644 tensorflow/core/kernels/topk_op_gpu_uint32.cu.cc
 create mode 100644 tensorflow/core/kernels/topk_op_gpu_uint64.cu.cc

diff --git a/tensorflow/core/framework/register_types.h b/tensorflow/core/framework/register_types.h
index bc3e5e1743b..0cf6536e8c2 100644
--- a/tensorflow/core/framework/register_types.h
+++ b/tensorflow/core/framework/register_types.h
@@ -153,16 +153,9 @@ limitations under the License.
 #endif  // defined(IS_MOBILE_PLATFORM)  - end of TF_CALL_type defines
 
 // Defines for sets of types.
-
-// TODO(b/111604096): Add uint32 and uint64 to TF_CALL_INTEGRAL_TYPES.
-//
-// The uint32 and uint64 types were introduced in 10/2017 to be used via XLA and
-// thus were not included in TF_CALL_INTEGRAL_TYPES. Including them in
-// TF_CALL_INTEGRAL_TYPES should only happen after evaluating the effect on the
-// TF binary size and performance.
-#define TF_CALL_INTEGRAL_TYPES(m)                                      \
-  TF_CALL_int64(m) TF_CALL_int32(m) TF_CALL_uint16(m) TF_CALL_int16(m) \
-      TF_CALL_uint8(m) TF_CALL_int8(m)
+#define TF_CALL_INTEGRAL_TYPES(m)                                       \
+  TF_CALL_uint64(m) TF_CALL_int64(m) TF_CALL_uint32(m) TF_CALL_int32(m) \
+      TF_CALL_uint16(m) TF_CALL_int16(m) TF_CALL_uint8(m) TF_CALL_int8(m)
 
 #define TF_CALL_FLOAT_TYPES(m) \
   TF_CALL_half(m) TF_CALL_bfloat16(m) TF_CALL_float(m) TF_CALL_double(m)
@@ -174,10 +167,10 @@ limitations under the License.
 #define TF_CALL_REAL_NUMBER_TYPES_NO_BFLOAT16(m) \
   TF_CALL_INTEGRAL_TYPES(m) TF_CALL_half(m) TF_CALL_float(m) TF_CALL_double(m)
 
-#define TF_CALL_REAL_NUMBER_TYPES_NO_INT32(m)                              \
-  TF_CALL_half(m) TF_CALL_bfloat16(m) TF_CALL_float(m) TF_CALL_double(m)   \
-      TF_CALL_int64(m) TF_CALL_uint16(m) TF_CALL_int16(m) TF_CALL_uint8(m) \
-          TF_CALL_int8(m)
+#define TF_CALL_REAL_NUMBER_TYPES_NO_INT32(m)                                \
+  TF_CALL_half(m) TF_CALL_bfloat16(m) TF_CALL_float(m) TF_CALL_double(m)     \
+      TF_CALL_uint64(m) TF_CALL_int64(m) TF_CALL_uint32(m) TF_CALL_uint16(m) \
+          TF_CALL_int16(m) TF_CALL_uint8(m) TF_CALL_int8(m)
 
 #define TF_CALL_COMPLEX_TYPES(m) TF_CALL_complex64(m) TF_CALL_complex128(m)
 
diff --git a/tensorflow/core/framework/types.cc b/tensorflow/core/framework/types.cc
index 97eaec98ffe..d6455e012d0 100644
--- a/tensorflow/core/framework/types.cc
+++ b/tensorflow/core/framework/types.cc
@@ -238,11 +238,6 @@ int DataTypeSize(DataType dt) {
     TF_CALL_qint16(CASE);
     TF_CALL_quint16(CASE);
 
-    // uint32 and uint64 aren't included in TF_CALL_POD_TYPES because we
-    // don't want to define kernels for them at this stage to avoid binary
-    // bloat.
-    TF_CALL_uint32(CASE);
-    TF_CALL_uint64(CASE);
     default:
       return 0;
   }
diff --git a/tensorflow/core/kernels/BUILD b/tensorflow/core/kernels/BUILD
index 279dff92c58..97f974c6af4 100644
--- a/tensorflow/core/kernels/BUILD
+++ b/tensorflow/core/kernels/BUILD
@@ -4900,7 +4900,9 @@ tf_kernel_library(
         "topk_op_gpu_double.cu.cc",
         "topk_op_gpu_float.cu.cc",
         "topk_op_gpu_half.cu.cc",
+        "topk_op_gpu_uint64.cu.cc",
         "topk_op_gpu_int64.cu.cc",
+        "topk_op_gpu_uint32.cu.cc",
         "topk_op_gpu_int32.cu.cc",
         "topk_op_gpu_int16.cu.cc",
         "topk_op_gpu_uint16.cu.cc",
diff --git a/tensorflow/core/kernels/concat_lib_cpu.cc b/tensorflow/core/kernels/concat_lib_cpu.cc
index da73d3d2c56..1dec589d3ff 100644
--- a/tensorflow/core/kernels/concat_lib_cpu.cc
+++ b/tensorflow/core/kernels/concat_lib_cpu.cc
@@ -116,8 +116,6 @@ REGISTER(qint8)
 REGISTER(quint16)
 REGISTER(qint16)
 REGISTER(qint32)
-REGISTER(uint32)
-REGISTER(uint64)
 
 #if defined(IS_MOBILE_PLATFORM) && !defined(SUPPORT_SELECTIVE_REGISTRATION) && \
     !defined(__ANDROID_TYPES_FULL__)
diff --git a/tensorflow/core/kernels/concat_op.cc b/tensorflow/core/kernels/concat_op.cc
index be3e9a67c5f..d3f3a04f33b 100644
--- a/tensorflow/core/kernels/concat_op.cc
+++ b/tensorflow/core/kernels/concat_op.cc
@@ -208,8 +208,6 @@ REGISTER_CONCAT(qint8);
 REGISTER_CONCAT(quint16);
 REGISTER_CONCAT(qint16);
 REGISTER_CONCAT(qint32);
-REGISTER_CONCAT(uint32);
-REGISTER_CONCAT(uint64);
 
 #undef REGISTER_CONCAT
 
diff --git a/tensorflow/core/kernels/constant_op.cc b/tensorflow/core/kernels/constant_op.cc
index 4bcbc076446..dc178d17d49 100644
--- a/tensorflow/core/kernels/constant_op.cc
+++ b/tensorflow/core/kernels/constant_op.cc
@@ -211,7 +211,6 @@ TF_CALL_ALL_TYPES(REGISTER_CPU_KERNEL);
 // the conversion from uint8 to quint8.
 REGISTER_KERNEL(CPU, quint8);
 REGISTER_KERNEL(CPU, quint16);
-REGISTER_KERNEL(CPU, uint32);
 #undef REGISTER_CPU_KERNEL
 
 #ifdef TENSORFLOW_USE_SYCL
diff --git a/tensorflow/core/kernels/control_flow_ops.cc b/tensorflow/core/kernels/control_flow_ops.cc
index c8e83b6f672..accb2c59540 100644
--- a/tensorflow/core/kernels/control_flow_ops.cc
+++ b/tensorflow/core/kernels/control_flow_ops.cc
@@ -101,16 +101,12 @@ TF_CALL_ALL_TYPES(REGISTER_CPU_SWITCH);
 TF_CALL_ALL_TYPES(REGISTER_CPU_REF_SWITCH);
 TF_CALL_QUANTIZED_TYPES(REGISTER_CPU_SWITCH);
 TF_CALL_QUANTIZED_TYPES(REGISTER_CPU_REF_SWITCH);
-REGISTER_CPU_SWITCH(uint64);
 
 TF_CALL_NUMBER_TYPES_NO_INT32(REGISTER_GPU_SWITCH);
 TF_CALL_QUANTIZED_TYPES(REGISTER_GPU_SWITCH);
 TF_CALL_NUMBER_TYPES_NO_INT32(REGISTER_GPU_REF_SWITCH);
 TF_CALL_QUANTIZED_TYPES(REGISTER_GPU_REF_SWITCH);
-REGISTER_GPU_SWITCH(uint64);
 TF_CALL_variant(REGISTER_GPU_SWITCH);
-TF_CALL_uint32(REGISTER_GPU_SWITCH);
-TF_CALL_uint32(REGISTER_GPU_REF_SWITCH);
 
 #undef REGISTER_CPU_SWITCH
 #undef REGISTER_CPU_REF_SWITCH
@@ -311,7 +307,6 @@ TF_CALL_QUANTIZED_TYPES(REGISTER_GPU_KERNEL);
 TF_CALL_QUANTIZED_TYPES(REGISTER_GPU_REF_KERNEL);
 REGISTER_GPU_KERNEL(bool);
 REGISTER_GPU_REF_KERNEL(bool);
-REGISTER_GPU_KERNEL(uint64);
 TF_CALL_variant(REGISTER_GPU_KERNEL);
 
 #undef REGISTER_GPU_KERNEL
diff --git a/tensorflow/core/kernels/data/dataset_test_base.cc b/tensorflow/core/kernels/data/dataset_test_base.cc
index b91ab9b733c..e41e35be1e9 100644
--- a/tensorflow/core/kernels/data/dataset_test_base.cc
+++ b/tensorflow/core/kernels/data/dataset_test_base.cc
@@ -220,8 +220,6 @@ Status DatasetOpsTestBase::ExpectEqual(const Tensor& a, const Tensor& b) {
     break;
     TF_CALL_NUMBER_TYPES(CASE);
     TF_CALL_tstring(CASE);
-    TF_CALL_uint32(CASE);
-    TF_CALL_uint64(CASE);
     // TODO(feihugis): figure out how to support variant tensors.
 #undef CASE
     default:
diff --git a/tensorflow/core/kernels/dense_update_ops.cc b/tensorflow/core/kernels/dense_update_ops.cc
index 55e4cd7606a..71235fca143 100644
--- a/tensorflow/core/kernels/dense_update_ops.cc
+++ b/tensorflow/core/kernels/dense_update_ops.cc
@@ -98,7 +98,6 @@ typedef Eigen::SyclDevice SYCLDevice;
 
 TF_CALL_ALL_TYPES(REGISTER_KERNELS);
 // uint32 not included in ALL_TYPES
-TF_CALL_uint32(REGISTER_KERNELS);
 TF_CALL_QUANTIZED_TYPES(REGISTER_KERNELS);
 // quint16 not included in QUANTIZIED_TYPES
 TF_CALL_quint16(REGISTER_KERNELS);
diff --git a/tensorflow/core/kernels/dynamic_partition_op.cc b/tensorflow/core/kernels/dynamic_partition_op.cc
index 90ed71dccce..95af19c4c48 100644
--- a/tensorflow/core/kernels/dynamic_partition_op.cc
+++ b/tensorflow/core/kernels/dynamic_partition_op.cc
@@ -164,8 +164,6 @@ class DynamicPartitionOp : public DynamicPartitionOp_Shared {
       DynamicPartitionOp)
 
 TF_CALL_ALL_TYPES(REGISTER_DYNAMIC_PARTITION);
-// For partitioning fingerprints.
-TF_CALL_uint64(REGISTER_DYNAMIC_PARTITION);
 #undef REGISTER_DYNAMIC_PARTITION
 
 }  // namespace tensorflow
diff --git a/tensorflow/core/kernels/fill_functor.cc b/tensorflow/core/kernels/fill_functor.cc
index 10dd3df1915..174a4e45a79 100644
--- a/tensorflow/core/kernels/fill_functor.cc
+++ b/tensorflow/core/kernels/fill_functor.cc
@@ -45,6 +45,8 @@ DEFINE_SETZERO_CPU(Eigen::half);
 DEFINE_SETZERO_CPU(bfloat16);
 DEFINE_SETZERO_CPU(float);
 DEFINE_SETZERO_CPU(double);
+DEFINE_SETZERO_CPU(uint32);
+DEFINE_SETZERO_CPU(uint64);
 DEFINE_SETZERO_CPU(uint8);
 DEFINE_SETZERO_CPU(int8);
 DEFINE_SETZERO_CPU(uint16);
@@ -96,6 +98,8 @@ DEFINE_SETONE_CPU(Eigen::half);
 DEFINE_SETONE_CPU(bfloat16);
 DEFINE_SETONE_CPU(float);
 DEFINE_SETONE_CPU(double);
+DEFINE_SETONE_CPU(uint32);
+DEFINE_SETONE_CPU(uint64);
 DEFINE_SETONE_CPU(uint8);
 DEFINE_SETONE_CPU(int8);
 DEFINE_SETONE_CPU(uint16);
@@ -137,7 +141,6 @@ struct FillFunctor {
 TF_CALL_ALL_TYPES(DEFINE_FILL_CPU);
 DEFINE_FILL_CPU(quint8);
 DEFINE_FILL_CPU(quint16);
-DEFINE_FILL_CPU(uint32);
 #undef DEFINE_FILL_CPU
 
 #ifdef TENSORFLOW_USE_SYCL
diff --git a/tensorflow/core/kernels/gather_op.cc b/tensorflow/core/kernels/gather_op.cc
index 6d493a5f2ea..948567e019a 100644
--- a/tensorflow/core/kernels/gather_op.cc
+++ b/tensorflow/core/kernels/gather_op.cc
@@ -211,8 +211,6 @@ TF_CALL_ALL_TYPES(REGISTER_GATHER_CPU);
 TF_CALL_QUANTIZED_TYPES(REGISTER_GATHER_CPU);
 TF_CALL_quint16(REGISTER_GATHER_CPU);
 TF_CALL_qint16(REGISTER_GATHER_CPU);
-TF_CALL_uint32(REGISTER_GATHER_CPU);
-TF_CALL_uint64(REGISTER_GATHER_CPU);
 
 #undef REGISTER_GATHER_CPU
 
diff --git a/tensorflow/core/kernels/identity_op.cc b/tensorflow/core/kernels/identity_op.cc
index fd94df9a768..daa8a1ddb25 100644
--- a/tensorflow/core/kernels/identity_op.cc
+++ b/tensorflow/core/kernels/identity_op.cc
@@ -122,7 +122,6 @@ REGISTER_SYCL_HOST_KERNEL(bool);
 
 TF_CALL_NUMBER_TYPES_NO_INT32(REGISTER_GPU_KERNEL);
 REGISTER_GPU_KERNEL(Variant);
-TF_CALL_uint32(REGISTER_GPU_KERNEL);
 
 #undef REGISTER_GPU_KERNEL
 
diff --git a/tensorflow/core/kernels/ragged_gather_op.cc b/tensorflow/core/kernels/ragged_gather_op.cc
index 88c0d1ebd69..3bf82cba050 100644
--- a/tensorflow/core/kernels/ragged_gather_op.cc
+++ b/tensorflow/core/kernels/ragged_gather_op.cc
@@ -296,8 +296,6 @@ TF_CALL_tstring(REGISTER_CPU_KERNEL);
 TF_CALL_QUANTIZED_TYPES(REGISTER_CPU_KERNEL);
 TF_CALL_quint16(REGISTER_CPU_KERNEL);
 TF_CALL_qint16(REGISTER_CPU_KERNEL);
-TF_CALL_uint32(REGISTER_CPU_KERNEL);
-TF_CALL_uint64(REGISTER_CPU_KERNEL);
 #undef REGISTER_CPU_KERNEL
 #undef REGISTER_CPU_KERNEL_WITH_INDEX_TYPE
 
diff --git a/tensorflow/core/kernels/ragged_tensor_from_variant_op.cc b/tensorflow/core/kernels/ragged_tensor_from_variant_op.cc
index f83bcb38c6c..ad0712e6fd0 100644
--- a/tensorflow/core/kernels/ragged_tensor_from_variant_op.cc
+++ b/tensorflow/core/kernels/ragged_tensor_from_variant_op.cc
@@ -308,8 +308,6 @@ TF_CALL_tstring(REGISTER_KERNELS);
 TF_CALL_QUANTIZED_TYPES(REGISTER_KERNELS);
 TF_CALL_quint16(REGISTER_KERNELS);
 TF_CALL_qint16(REGISTER_KERNELS);
-TF_CALL_uint32(REGISTER_KERNELS);
-TF_CALL_uint64(REGISTER_KERNELS);
 #undef REGISTER_KERNELS
 #undef REGISTER_KERNELS_WITH_SPLIT_TYPE
 }  // namespace tensorflow
diff --git a/tensorflow/core/kernels/ragged_tensor_to_tensor_op.cc b/tensorflow/core/kernels/ragged_tensor_to_tensor_op.cc
index d729c43f25a..9ae5d7ffbdc 100644
--- a/tensorflow/core/kernels/ragged_tensor_to_tensor_op.cc
+++ b/tensorflow/core/kernels/ragged_tensor_to_tensor_op.cc
@@ -561,8 +561,6 @@ TF_CALL_string(REGISTER_CPU_KERNEL);
 TF_CALL_QUANTIZED_TYPES(REGISTER_CPU_KERNEL);
 TF_CALL_quint16(REGISTER_CPU_KERNEL);
 TF_CALL_qint16(REGISTER_CPU_KERNEL);
-TF_CALL_uint32(REGISTER_CPU_KERNEL);
-TF_CALL_uint64(REGISTER_CPU_KERNEL);
 
 #undef REGISTER_CPU_KERNEL
 
diff --git a/tensorflow/core/kernels/ragged_tensor_to_variant_op.cc b/tensorflow/core/kernels/ragged_tensor_to_variant_op.cc
index 7a5ae1c6240..64c372b005e 100644
--- a/tensorflow/core/kernels/ragged_tensor_to_variant_op.cc
+++ b/tensorflow/core/kernels/ragged_tensor_to_variant_op.cc
@@ -213,8 +213,6 @@ TF_CALL_tstring(REGISTER_KERNELS);
 TF_CALL_QUANTIZED_TYPES(REGISTER_KERNELS);
 TF_CALL_quint16(REGISTER_KERNELS);
 TF_CALL_qint16(REGISTER_KERNELS);
-TF_CALL_uint32(REGISTER_KERNELS);
-TF_CALL_uint64(REGISTER_KERNELS);
 #undef REGISTER_KERNELS
 #undef REGISTER_KERNELS_WITH_SPLIT_TYPE
 }  // namespace tensorflow
diff --git a/tensorflow/core/kernels/resource_variable_ops.cc b/tensorflow/core/kernels/resource_variable_ops.cc
index 0fc1d53749f..79a64cb9219 100644
--- a/tensorflow/core/kernels/resource_variable_ops.cc
+++ b/tensorflow/core/kernels/resource_variable_ops.cc
@@ -512,7 +512,6 @@ class AssignVariableOp : public OpKernel {
 
 TF_CALL_ALL_TYPES(REGISTER_KERNELS);
 TF_CALL_QUANTIZED_TYPES(REGISTER_KERNELS);
-TF_CALL_uint32(REGISTER_KERNELS);
 #undef REGISTER_KERNELS
 
 #if GOOGLE_CUDA || TENSORFLOW_USE_ROCM
diff --git a/tensorflow/core/kernels/split_lib_cpu.cc b/tensorflow/core/kernels/split_lib_cpu.cc
index 0cb0a94d498..a3060e4e90d 100644
--- a/tensorflow/core/kernels/split_lib_cpu.cc
+++ b/tensorflow/core/kernels/split_lib_cpu.cc
@@ -43,7 +43,6 @@ void Split::operator()(
 
 TF_CALL_ALL_TYPES(DEFINE_CPU_KERNELS)
 DEFINE_CPU_KERNELS(quint8)
-DEFINE_CPU_KERNELS(uint64)
 
 #ifdef TENSORFLOW_USE_SYCL
 template 
diff --git a/tensorflow/core/kernels/split_op.cc b/tensorflow/core/kernels/split_op.cc
index f09740c6198..08575f01f67 100644
--- a/tensorflow/core/kernels/split_op.cc
+++ b/tensorflow/core/kernels/split_op.cc
@@ -404,7 +404,6 @@ class SplitOpSYCL : public SplitOpBase {
 
 TF_CALL_ALL_TYPES(REGISTER_SPLIT);
 REGISTER_SPLIT(quint8);
-REGISTER_SPLIT(uint64);
 
 #undef REGISTER_SPLIT
 
diff --git a/tensorflow/core/kernels/strided_slice_op.cc b/tensorflow/core/kernels/strided_slice_op.cc
index ccc1984bb98..b4099213303 100644
--- a/tensorflow/core/kernels/strided_slice_op.cc
+++ b/tensorflow/core/kernels/strided_slice_op.cc
@@ -440,8 +440,6 @@ class StridedSliceAssignOp : public OpKernel {
                           StridedSliceAssignOp)
 
 TF_CALL_ALL_TYPES(REGISTER_STRIDED_SLICE);
-TF_CALL_uint32(REGISTER_STRIDED_SLICE);
-TF_CALL_uint64(REGISTER_STRIDED_SLICE);
 
 #undef REGISTER_STRIDED_SLICE
 
diff --git a/tensorflow/core/kernels/strided_slice_op_impl.h b/tensorflow/core/kernels/strided_slice_op_impl.h
index 1ae959b7b3f..5ce1d773e33 100644
--- a/tensorflow/core/kernels/strided_slice_op_impl.h
+++ b/tensorflow/core/kernels/strided_slice_op_impl.h
@@ -287,8 +287,6 @@ TF_CALL_GPU_ALL_TYPES(DECLARE_FOR_N_GPU);
 #endif  // END GOOGLE_CUDA || TENSORFLOW_USE_ROCM
 
 TF_CALL_ALL_TYPES(DECLARE_FOR_N_CPU);
-TF_CALL_uint32(DECLARE_FOR_N_CPU);
-TF_CALL_uint64(DECLARE_FOR_N_CPU);
 
 #ifdef TENSORFLOW_USE_SYCL
 #define PREVENT_FOR_N_SYCL(T) \
diff --git a/tensorflow/core/kernels/topk_op.cc b/tensorflow/core/kernels/topk_op.cc
index c555b42f005..50325b7bcfe 100644
--- a/tensorflow/core/kernels/topk_op.cc
+++ b/tensorflow/core/kernels/topk_op.cc
@@ -258,7 +258,6 @@ namespace functor {
 
 TF_CALL_GPU_NUMBER_TYPES(DECLARE_GPU_SPEC);
 TF_CALL_INTEGRAL_TYPES(DECLARE_GPU_SPEC);
-TF_CALL_uint32(DECLARE_GPU_SPEC);
 
 #undef DECLARE_GPU_SPEC
 
@@ -276,7 +275,6 @@ TF_CALL_uint32(DECLARE_GPU_SPEC);
 
 TF_CALL_GPU_NUMBER_TYPES(REGISTER_KERNELS);
 TF_CALL_INTEGRAL_TYPES(REGISTER_KERNELS);
-TF_CALL_uint32(REGISTER_KERNELS)
 #undef REGISTER_KERNELS
 
 #endif  // end GOOGLE_CUDA
diff --git a/tensorflow/core/kernels/topk_op_gpu_uint32.cu.cc b/tensorflow/core/kernels/topk_op_gpu_uint32.cu.cc
new file mode 100644
index 00000000000..16e2e0e9420
--- /dev/null
+++ b/tensorflow/core/kernels/topk_op_gpu_uint32.cu.cc
@@ -0,0 +1,28 @@
+/* Copyright 2020 The TensorFlow Authors. All Rights Reserved.
+
+Licensed under the Apache License, Version 2.0 (the "License");
+you may not use this file except in compliance with the License.
+You may obtain a copy of the License at
+
+    http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software
+distributed under the License is distributed on an "AS IS" BASIS,
+WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+See the License for the specific language governing permissions and
+limitations under the License.
+==============================================================================*/
+
+#if GOOGLE_CUDA
+#define EIGEN_USE_GPU
+
+#include "tensorflow/core/kernels/topk_op.h"
+#include "tensorflow/core/kernels/topk_op_gpu.h"
+
+namespace tensorflow {
+using Eigen::GpuDevice;
+
+template struct functor::TopKFunctor;
+}  // namespace tensorflow
+
+#endif  // GOOGLE_CUDA
diff --git a/tensorflow/core/kernels/topk_op_gpu_uint64.cu.cc b/tensorflow/core/kernels/topk_op_gpu_uint64.cu.cc
new file mode 100644
index 00000000000..895247a63a2
--- /dev/null
+++ b/tensorflow/core/kernels/topk_op_gpu_uint64.cu.cc
@@ -0,0 +1,28 @@
+/* Copyright 2020 The TensorFlow Authors. All Rights Reserved.
+
+Licensed under the Apache License, Version 2.0 (the "License");
+you may not use this file except in compliance with the License.
+You may obtain a copy of the License at
+
+    http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software
+distributed under the License is distributed on an "AS IS" BASIS,
+WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+See the License for the specific language governing permissions and
+limitations under the License.
+==============================================================================*/
+
+#if GOOGLE_CUDA
+#define EIGEN_USE_GPU
+
+#include "tensorflow/core/kernels/topk_op.h"
+#include "tensorflow/core/kernels/topk_op_gpu.h"
+
+namespace tensorflow {
+using Eigen::GpuDevice;
+
+template struct functor::TopKFunctor;
+}  // namespace tensorflow
+
+#endif  // GOOGLE_CUDA
diff --git a/tensorflow/core/util/batch_util.cc b/tensorflow/core/util/batch_util.cc
index b88c365ced0..e03188b04da 100644
--- a/tensorflow/core/util/batch_util.cc
+++ b/tensorflow/core/util/batch_util.cc
@@ -182,8 +182,6 @@ Status CopyElementToSlice(Tensor element, Tensor* parent, int64 index) {
   switch (element.dtype()) {
     TF_CALL_ALL_TYPES(HANDLE_TYPE);
     TF_CALL_QUANTIZED_TYPES(HANDLE_TYPE);
-    TF_CALL_uint32(HANDLE_TYPE);
-    TF_CALL_uint64(HANDLE_TYPE);
 #undef HANDLE_TYPE
     default:
       return errors::Unimplemented("CopyElementToSlice Unhandled data type: ",
@@ -207,8 +205,6 @@ Status CopySliceToElement(const Tensor& parent, Tensor* element, int64 index) {
   switch (parent.dtype()) {
     TF_CALL_ALL_TYPES(HANDLE_TYPE);
     TF_CALL_QUANTIZED_TYPES(HANDLE_TYPE);
-    TF_CALL_uint32(HANDLE_TYPE);
-    TF_CALL_uint64(HANDLE_TYPE);
 #undef HANDLE_TYPE
     default:
       return errors::Unimplemented("CopySliceToElement Unhandled data type: ",
@@ -280,8 +276,6 @@ Status CopyContiguousSlices(const Tensor& src, int64 src_offset,
   switch (src.dtype()) {
     TF_CALL_ALL_TYPES(HANDLE_TYPE);
     TF_CALL_QUANTIZED_TYPES(HANDLE_TYPE);
-    TF_CALL_uint32(HANDLE_TYPE);
-    TF_CALL_uint64(HANDLE_TYPE);
 #undef HANDLE_TYPE
     default:
       return errors::Unimplemented("CopyContiguousSlices unhandled data type: ",
@@ -308,8 +302,6 @@ Status MaybeMoveSliceToElement(Tensor* parent, Tensor* element, int64 index) {
   switch (parent->dtype()) {
     TF_CALL_ALL_TYPES(HANDLE_TYPE);
     TF_CALL_QUANTIZED_TYPES(HANDLE_TYPE);
-    TF_CALL_uint32(HANDLE_TYPE);
-    TF_CALL_uint64(HANDLE_TYPE);
 #undef HANDLE_TYPE
     default:
       return errors::Unimplemented(
diff --git a/tensorflow/core/util/saved_tensor_slice_util.h b/tensorflow/core/util/saved_tensor_slice_util.h
index 09b9235b711..1f9768f5163 100644
--- a/tensorflow/core/util/saved_tensor_slice_util.h
+++ b/tensorflow/core/util/saved_tensor_slice_util.h
@@ -116,7 +116,9 @@ TENSOR_PROTO_EXTRACT_TYPE(double, double, double);
 TENSOR_PROTO_EXTRACT_TYPE_COMPLEX(complex64, scomplex, float);
 TENSOR_PROTO_EXTRACT_TYPE_COMPLEX(complex128, dcomplex, double);
 TENSOR_PROTO_EXTRACT_TYPE(int32, int, int32);
+TENSOR_PROTO_EXTRACT_TYPE(uint32, uint32, uint32);
 TENSOR_PROTO_EXTRACT_TYPE(int64, int64, protobuf_int64);
+TENSOR_PROTO_EXTRACT_TYPE(uint64, uint64, protobuf_uint64);
 TENSOR_PROTO_EXTRACT_TYPE(uint16, int, int32);
 TENSOR_PROTO_EXTRACT_TYPE(uint8, int, int32);
 TENSOR_PROTO_EXTRACT_TYPE(int8, int, int32);

From 9f20b156bc7862fb621756fd5d6744255b1f3735 Mon Sep 17 00:00:00 2001
From: George Karpenkov 
Date: Fri, 19 Jun 2020 01:13:01 -0700
Subject: [PATCH 279/867] [XLA:GPU] [NFC] Clarify the precondition for the fast
 reduction emitter

PiperOrigin-RevId: 317266013
Change-Id: I384acac279f0db53f195d5b43318c38c87a1739c
---
 tensorflow/compiler/xla/service/gpu/ir_emission_utils.cc | 5 +++++
 1 file changed, 5 insertions(+)

diff --git a/tensorflow/compiler/xla/service/gpu/ir_emission_utils.cc b/tensorflow/compiler/xla/service/gpu/ir_emission_utils.cc
index b97aa3651c6..01bcf456f75 100644
--- a/tensorflow/compiler/xla/service/gpu/ir_emission_utils.cc
+++ b/tensorflow/compiler/xla/service/gpu/ir_emission_utils.cc
@@ -226,6 +226,11 @@ bool IsReductionFromOrToContiguousDimensions(const HloInstruction& reduce) {
       dims_to_keep.push_back(dim);
     }
   }
+
+  // We support fast codegen for three cases:
+  // 1) Row reduction: (K, R)
+  // 2) Column reduction: (K, R, K)
+  // 3) "Batched" row reduction: (R, K, R)
   if (!LayoutUtil::AreDimensionsConsecutive(input->shape().layout(),
                                             dims_to_keep) &&
       !LayoutUtil::AreDimensionsConsecutive(input->shape().layout(),

From 051d1b70f5f0636316e2651630f0ade554f192c0 Mon Sep 17 00:00:00 2001
From: Stefano Galarraga 
Date: Fri, 19 Jun 2020 01:21:24 -0700
Subject: [PATCH 280/867] Fix NNAPI delegation error on models with MAX/MIN
 operations with scalar quantized operators

PiperOrigin-RevId: 317266736
Change-Id: Ieed8a77685d4ca0d51389b5976addf0de167cfcf
---
 tensorflow/lite/delegates/nnapi/nnapi_delegate.cc |  2 ++
 tensorflow/lite/kernels/maximum_minimum_test.cc   | 11 +++++++++++
 2 files changed, 13 insertions(+)

diff --git a/tensorflow/lite/delegates/nnapi/nnapi_delegate.cc b/tensorflow/lite/delegates/nnapi/nnapi_delegate.cc
index a3a3f9fda4d..1c35ee370c2 100644
--- a/tensorflow/lite/delegates/nnapi/nnapi_delegate.cc
+++ b/tensorflow/lite/delegates/nnapi/nnapi_delegate.cc
@@ -160,6 +160,8 @@ bool IsScalarInputSupported(int builtin_code) {
     case kTfLiteBuiltinLess:
     case kTfLiteBuiltinLessEqual:
     case kTfLiteBuiltinPow:
+    case kTfLiteBuiltinMaximum:
+    case kTfLiteBuiltinMinimum:
       return true;
     default:
       return false;
diff --git a/tensorflow/lite/kernels/maximum_minimum_test.cc b/tensorflow/lite/kernels/maximum_minimum_test.cc
index 2c036e369bd..803fe91c460 100644
--- a/tensorflow/lite/kernels/maximum_minimum_test.cc
+++ b/tensorflow/lite/kernels/maximum_minimum_test.cc
@@ -190,6 +190,17 @@ TEST(MaximumOpTest, Int32WithBroadcastTest_ScalarY) {
                      data1, data2, {1, 0, -1, -2, 2, 2}, /*is_constant=*/true);
 }
 
+TEST(MaximumOpTest, Int8WithBroadcastTest_ScalarY) {
+  std::initializer_list data1 = {1, 0, -1, -2, 3, 11};
+  std::initializer_list data2 = {2};
+  TestModel(BuiltinOperator_MAXIMUM, {TensorType_INT8, {3, 1, 2}},
+                    {TensorType_INT8, {}}, {TensorType_INT8, {3, 1, 2}}, data1,
+                    data2, {2, 2, 2, 2, 3, 11}, /*is_constant=*/true);
+  TestModel(BuiltinOperator_MINIMUM, {TensorType_INT8, {3, 1, 2}},
+                    {TensorType_INT8, {}}, {TensorType_INT8, {3, 1, 2}}, data1,
+                    data2, {1, 0, -1, -2, 2, 2}, /*is_constant=*/true);
+}
+
 TEST(MaxMinOpTest, Int8Test8D) {
   std::initializer_list data1 = {1, 0, 2, 11, 2, 23};
   std::initializer_list data2 = {0, 0, 1, 12, 123, 1};

From e51b17f4582183a216d3a47450117c5e8cdd387d Mon Sep 17 00:00:00 2001
From: Adrian Kuegel 
Date: Fri, 19 Jun 2020 01:49:14 -0700
Subject: [PATCH 281/867] Add a small test to cover the mlir generated Tanh GPU
 kernel.

This test is a first step towards being able to ensure that we don't
accidentally break the kernel generation.

PiperOrigin-RevId: 317269120
Change-Id: Iad6bdd7ab7e9fb819a478c947ba6294a191f1099
---
 tensorflow/core/kernels/BUILD                 | 19 +++++
 .../mlir_generated_op_gpu_tanh_test.cc        | 85 +++++++++++++++++++
 2 files changed, 104 insertions(+)
 create mode 100644 tensorflow/core/kernels/mlir_generated_op_gpu_tanh_test.cc

diff --git a/tensorflow/core/kernels/BUILD b/tensorflow/core/kernels/BUILD
index 97f974c6af4..0b7a092033b 100644
--- a/tensorflow/core/kernels/BUILD
+++ b/tensorflow/core/kernels/BUILD
@@ -4168,6 +4168,25 @@ tf_kernel_library(
     ]),
 )
 
+tf_cuda_cc_test(
+    name = "mlir_generated_op_gpu_tanh_test",
+    size = "small",
+    srcs = if_mlir_generated_gpu_kernels_enabled(["mlir_generated_op_gpu_tanh_test.cc"]),
+    tags = tf_cuda_tests_tags() + ["no_rocm"],
+    deps = [
+        ":cwise_op",
+        ":ops_testutil",
+        "//tensorflow/core:framework",
+        "//tensorflow/core:framework_internal",
+        "//tensorflow/core:tensorflow",
+        "//tensorflow/core:test",
+        "//tensorflow/core:test_main",
+        "//tensorflow/core:testlib",
+        "//tensorflow/core/common_runtime:device",
+        "//tensorflow/core/common_runtime:device_factory",
+    ],
+)
+
 tf_kernel_library(
     name = "nextafter_op",
     prefix = "nextafter_op",
diff --git a/tensorflow/core/kernels/mlir_generated_op_gpu_tanh_test.cc b/tensorflow/core/kernels/mlir_generated_op_gpu_tanh_test.cc
new file mode 100644
index 00000000000..39c1d709b1e
--- /dev/null
+++ b/tensorflow/core/kernels/mlir_generated_op_gpu_tanh_test.cc
@@ -0,0 +1,85 @@
+/* Copyright 2020 The TensorFlow Authors. All Rights Reserved.
+
+Licensed under the Apache License, Version 2.0 (the "License");
+you may not use this file except in compliance with the License.
+You may obtain a copy of the License at
+
+    http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software
+distributed under the License is distributed on an "AS IS" BASIS,
+WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+See the License for the specific language governing permissions and
+limitations under the License.
+==============================================================================*/
+
+#include 
+#include 
+#include 
+
+#include "tensorflow/core/common_runtime/device.h"
+#include "tensorflow/core/common_runtime/device_factory.h"
+#include "tensorflow/core/framework/fake_input.h"
+#include "tensorflow/core/framework/node_def_builder.h"
+#include "tensorflow/core/framework/tensor.h"
+#include "tensorflow/core/kernels/ops_testutil.h"
+#include "tensorflow/core/lib/core/status_test_util.h"
+#include "tensorflow/core/platform/test.h"
+
+namespace tensorflow {
+namespace {
+
+class MlirGeneratedOpGpuTanhTest : public OpsTestBase {
+ protected:
+  void SetUp() override {
+    std::unique_ptr device_gpu(
+        tensorflow::DeviceFactory::NewDevice("GPU", {},
+                                             "/job:a/replica:0/task:0"));
+    SetDevice(tensorflow::DEVICE_GPU, std::move(device_gpu));
+  }
+  template 
+  void RunTanhOp(std::initializer_list input) {
+    TensorShape shape({2, 7});
+    TF_ASSERT_OK(NodeDefBuilder("tanh_op", "Tanh")
+                     .Input(FakeInput(DataTypeToEnum::v()))
+                     .Attr("T", DataTypeToEnum::v())
+                     .Finalize(node_def()));
+
+    TF_ASSERT_OK(InitOp());
+    AddInputFromArray(shape, input);
+    TF_ASSERT_OK(RunOpKernel());
+
+    Tensor expected_tensor(allocator(), DataTypeToEnum::value, shape);
+    std::vector expected;
+    expected.reserve(input.size());
+    for (const T& inp : input) {
+      expected.push_back(static_cast(std::tanh(static_cast(inp))));
+    }
+    test::FillValues(&expected_tensor, expected);
+    test::ExpectClose(expected_tensor, *GetOutput(0));
+  }
+};
+
+TEST_F(MlirGeneratedOpGpuTanhTest, TanhFloat) {
+  RunTanhOp({-18.0f, -9.0f, -1e-6f, -0.0f, 0.0f, 1e-6, 0.1f, 0.2f, 0.3f,
+                    0.5f, 0.7f, 0.9f, 9.0f, 18.0f});
+}
+
+TEST_F(MlirGeneratedOpGpuTanhTest, TanhDouble) {
+  RunTanhOp({-18.0, -9.0, -1e-6, -0.0, 0.0, 1e-6, 0.1, 0.2, 0.3, 0.5,
+                     0.7, 0.9, 9.0, 18.0});
+}
+
+TEST_F(MlirGeneratedOpGpuTanhTest, TanhHalf) {
+  RunTanhOp(
+      {static_cast(-18.0), static_cast(-9.0),
+       static_cast(-1e-6), static_cast(-0.0),
+       static_cast(0.0), static_cast(1e-6),
+       static_cast(0.1), static_cast(0.2),
+       static_cast(0.3), static_cast(0.5),
+       static_cast(0.7), static_cast(0.9),
+       static_cast(9.0), static_cast(18.0)});
+}
+
+}  // namespace
+}  // end namespace tensorflow

From 772433a2a2120d0aefc6c3628c6254d5a1aaf19d Mon Sep 17 00:00:00 2001
From: YoungSeok Yoon 
Date: Fri, 19 Jun 2020 01:58:58 -0700
Subject: [PATCH 282/867] Add flag for using optimized TFLite CPU kernels on
 iOS

This adds new experimental flags to the interpreter options of TFLite Obj-C and
Swift APIs, which can be used for opting in to a set of highly optimized
floating point kernels provided via the XNNPACK delegate. The flags can be used
as follows.

Obj-C:

    TFLInterpreterOptions *options = [[TFLInterpreterOptions alloc] init];
    options.useXNNPACK = YES;
    NSError *error;
    TFLInterpreter *interpreter =
        [[TFLInterpreter alloc] initWithModelPath:@"model/path"
                                          options:options
                                            error:&error];

Swift:

    var options = InterpreterOptions()
    options.isXNNPackEnabled = true
    var interpreter = try Interpreter(modelPath: "model/path", options: options)

PiperOrigin-RevId: 317270012
Change-Id: I82aae43c3de13ab08af3c70513e2a458e807b0f1
---
 tensorflow/lite/delegates/xnnpack/BUILD       |  4 ++
 tensorflow/lite/experimental/ios/BUILD.apple  | 18 +++++
 tensorflow/lite/experimental/objc/BUILD.apple |  1 +
 .../objc/TensorFlowLiteObjC-nightly.podspec   |  1 +
 .../objc/TensorFlowLiteObjC.podspec           |  1 +
 .../objc/TensorFlowLiteObjC.podspec.template  |  1 +
 .../objc/apis/TFLInterpreterOptions.h         | 21 ++++++
 .../objc/sources/TFLInterpreter.mm            | 15 +++++
 .../objc/tests/TFLInterpreterOptionsTests.m   |  9 +++
 .../swift/Sources/Interpreter.swift           | 67 ++++++++++++++++---
 .../swift/Tests/InterpreterTests.swift        | 62 +++++++++++------
 11 files changed, 171 insertions(+), 29 deletions(-)

diff --git a/tensorflow/lite/delegates/xnnpack/BUILD b/tensorflow/lite/delegates/xnnpack/BUILD
index 97e6aea2a6b..eaf7d8f6f03 100644
--- a/tensorflow/lite/delegates/xnnpack/BUILD
+++ b/tensorflow/lite/delegates/xnnpack/BUILD
@@ -14,6 +14,10 @@ EMSCRIPTEN_LINKOPTS = [
     "-s TOTAL_MEMORY=134217728",
 ]
 
+exports_files([
+    "xnnpack_delegate.h",
+])
+
 cc_library(
     name = "xnnpack_delegate",
     srcs = ["xnnpack_delegate.cc"],
diff --git a/tensorflow/lite/experimental/ios/BUILD.apple b/tensorflow/lite/experimental/ios/BUILD.apple
index 1a85b604f9b..7a40ca7b8e7 100644
--- a/tensorflow/lite/experimental/ios/BUILD.apple
+++ b/tensorflow/lite/experimental/ios/BUILD.apple
@@ -18,10 +18,26 @@ sh_binary(
     ],
 )
 
+# When the static framework is built with bazel, the all header files are moved
+# to the "Headers" directory with no header path prefixes. This auxiliary rule
+# is used for stripping the path prefix to the "common.h" file included by the
+# "xnnpack_delegate.h" header.
+genrule(
+    name = "strip_xnnpack_include_hdr",
+    srcs = ["//tensorflow/lite/delegates/xnnpack:xnnpack_delegate.h"],
+    outs = ["xnnpack_delegate.h"],
+    cmd = """
+    sed 's|#include ".*common.h"|#include "common.h"|'\
+    "$(location //tensorflow/lite/delegates/xnnpack:xnnpack_delegate.h)"\
+    > "$@"
+    """,
+)
+
 # bazel build -c opt --config=ios_fat //tensorflow/lite/experimental/ios:TensorFlowLiteC_framework
 tflite_ios_static_framework(
     name = "TensorFlowLiteC_framework",
     hdrs = [
+        ":xnnpack_delegate.h",
         "//tensorflow/lite/c:c_api.h",
         "//tensorflow/lite/c:common.h",
     ],
@@ -105,6 +121,7 @@ cc_library(
     hdrs = [
         "//tensorflow/lite/c:c_api.h",
         "//tensorflow/lite/c:common.h",
+        "//tensorflow/lite/delegates/xnnpack:xnnpack_delegate.h",
     ],
     tags = [
         "nobuilder",
@@ -112,6 +129,7 @@ cc_library(
     ],
     deps = [
         "//tensorflow/lite/c:c_api",
+        "//tensorflow/lite/delegates/xnnpack:xnnpack_delegate",
     ],
 )
 
diff --git a/tensorflow/lite/experimental/objc/BUILD.apple b/tensorflow/lite/experimental/objc/BUILD.apple
index 09d4547813a..d26d90c46a1 100644
--- a/tensorflow/lite/experimental/objc/BUILD.apple
+++ b/tensorflow/lite/experimental/objc/BUILD.apple
@@ -64,6 +64,7 @@ objc_library(
     visibility = ios_visibility_whitelist(),
     deps = [
         "//tensorflow/lite/c:c_api",
+        "//tensorflow/lite/delegates/xnnpack:xnnpack_delegate",
     ],
     alwayslink = 1,
 )
diff --git a/tensorflow/lite/experimental/objc/TensorFlowLiteObjC-nightly.podspec b/tensorflow/lite/experimental/objc/TensorFlowLiteObjC-nightly.podspec
index e039fb57114..eed0f087f44 100644
--- a/tensorflow/lite/experimental/objc/TensorFlowLiteObjC-nightly.podspec
+++ b/tensorflow/lite/experimental/objc/TensorFlowLiteObjC-nightly.podspec
@@ -26,6 +26,7 @@ Pod::Spec.new do |s|
     objc_dir + '{apis,sources}/*.{h,m,mm}',
     tfl_dir + 'c/c_api.h',
     tfl_dir + 'c/common.h',
+    tfl_dir + 'delegates/xnnpack/xnnpack_delegate.h',
   ]
   s.module_map = objc_dir + 'apis/framework.modulemap'
   s.dependency 'TensorFlowLiteC', "~> #{s.version}"
diff --git a/tensorflow/lite/experimental/objc/TensorFlowLiteObjC.podspec b/tensorflow/lite/experimental/objc/TensorFlowLiteObjC.podspec
index c673cfad759..5817619a58f 100644
--- a/tensorflow/lite/experimental/objc/TensorFlowLiteObjC.podspec
+++ b/tensorflow/lite/experimental/objc/TensorFlowLiteObjC.podspec
@@ -26,6 +26,7 @@ Pod::Spec.new do |s|
     objc_dir + '{apis,sources}/*.{h,m,mm}',
     tfl_dir + 'c/c_api.h',
     tfl_dir + 'c/common.h',
+    tfl_dir + 'delegates/xnnpack/xnnpack_delegate.h',
   ]
   s.module_map = objc_dir + 'apis/framework.modulemap'
   s.dependency 'TensorFlowLiteC', "#{s.version}"
diff --git a/tensorflow/lite/experimental/objc/TensorFlowLiteObjC.podspec.template b/tensorflow/lite/experimental/objc/TensorFlowLiteObjC.podspec.template
index fc9e10e4a2c..4ab5753e016 100644
--- a/tensorflow/lite/experimental/objc/TensorFlowLiteObjC.podspec.template
+++ b/tensorflow/lite/experimental/objc/TensorFlowLiteObjC.podspec.template
@@ -26,6 +26,7 @@ Pod::Spec.new do |s|
     objc_dir + '{apis,sources}/*.{h,m,mm}',
     tfl_dir + 'c/c_api.h',
     tfl_dir + 'c/common.h',
+    tfl_dir + 'delegates/xnnpack/xnnpack_delegate.h',
   ]
   s.module_map = objc_dir + 'apis/framework.modulemap'
   s.dependency 'TensorFlowLiteC', '~> 0.0.1-nightly'
diff --git a/tensorflow/lite/experimental/objc/apis/TFLInterpreterOptions.h b/tensorflow/lite/experimental/objc/apis/TFLInterpreterOptions.h
index 6461fbf0178..d7dbb2bd970 100644
--- a/tensorflow/lite/experimental/objc/apis/TFLInterpreterOptions.h
+++ b/tensorflow/lite/experimental/objc/apis/TFLInterpreterOptions.h
@@ -25,6 +25,27 @@ NS_ASSUME_NONNULL_BEGIN
  */
 @property(nonatomic) NSUInteger numberOfThreads;
 
+/**
+ * Experimental: Enable an optimized set of floating point CPU kernels (provided by XNNPACK).
+ *
+ * Enabling this flag will enable use of a new, highly optimized set of CPU kernels provided via the
+ * XNNPACK delegate. Currently, this is restricted to a subset of floating point operations.
+ * Eventually, we plan to enable this by default, as it can provide significant performance benefits
+ * for many classes of floating point models. See
+ * https://github.com/tensorflow/tensorflow/blob/master/tensorflow/lite/delegates/xnnpack/README.md
+ * for more details.
+ *
+ * Things to keep in mind when enabling this flag:
+ *
+ *     * Startup time and resize time may increase.
+ *     * Baseline memory consumption may increase.
+ *     * Compatibility with other delegates (e.g., GPU) has not been fully validated.
+ *     * Quantized models will not see any benefit.
+ *
+ * WARNING: This is an experimental interface that is subject to change.
+ */
+@property(nonatomic) BOOL useXNNPACK;
+
 /**
  * Initializes a new instance of `TFLInterpreterOptions`.
  *
diff --git a/tensorflow/lite/experimental/objc/sources/TFLInterpreter.mm b/tensorflow/lite/experimental/objc/sources/TFLInterpreter.mm
index 94031ee5428..34dd119885d 100644
--- a/tensorflow/lite/experimental/objc/sources/TFLInterpreter.mm
+++ b/tensorflow/lite/experimental/objc/sources/TFLInterpreter.mm
@@ -23,6 +23,7 @@
 #import "tensorflow/lite/experimental/objc/apis/TFLTensor.h"
 
 #include "tensorflow/lite/c/c_api.h"
+#include "tensorflow/lite/delegates/xnnpack/xnnpack_delegate.h"
 
 NS_ASSUME_NONNULL_BEGIN
 
@@ -45,6 +46,9 @@ static void TFLInterpreterErrorReporter(void *user_data, const char *format, va_
 /** TfLiteInterpreter backed by C API. */
 @property(nonatomic, nullable) TfLiteInterpreter *interpreter;
 
+/** TfLiteDelegate backed by C API. */
+@property(nonatomic, nullable) TfLiteDelegate *xnnpack_delegate;
+
 @end
 
 @implementation TFLInterpreter
@@ -53,6 +57,7 @@ static void TFLInterpreterErrorReporter(void *user_data, const char *format, va_
 
 - (void)dealloc {
   TfLiteInterpreterDelete(_interpreter);
+  TfLiteXNNPackDelegateDelete(_xnnpack_delegate);
 }
 
 #pragma mark - Public
@@ -104,6 +109,16 @@ static void TFLInterpreterErrorReporter(void *user_data, const char *format, va_
       }
       TfLiteInterpreterOptionsSetErrorReporter(cOptions, TFLInterpreterErrorReporter, nullptr);
 
+      if (options.useXNNPACK) {
+        TfLiteXNNPackDelegateOptions xnnpack_options = TfLiteXNNPackDelegateOptionsDefault();
+        if (options.numberOfThreads > 0) {
+          xnnpack_options.num_threads = (int32_t)options.numberOfThreads;
+        }
+
+        _xnnpack_delegate = TfLiteXNNPackDelegateCreate(&xnnpack_options);
+        TfLiteInterpreterOptionsAddDelegate(cOptions, _xnnpack_delegate);
+      }
+
       _interpreter = TfLiteInterpreterCreate(model, cOptions);
       if (_interpreter == nullptr) {
         [TFLErrorUtil saveInterpreterErrorWithCode:TFLInterpreterErrorCodeFailedToCreateInterpreter
diff --git a/tensorflow/lite/experimental/objc/tests/TFLInterpreterOptionsTests.m b/tensorflow/lite/experimental/objc/tests/TFLInterpreterOptionsTests.m
index 00b800d6af9..286cba98b49 100644
--- a/tensorflow/lite/experimental/objc/tests/TFLInterpreterOptionsTests.m
+++ b/tensorflow/lite/experimental/objc/tests/TFLInterpreterOptionsTests.m
@@ -32,6 +32,7 @@ NS_ASSUME_NONNULL_BEGIN
   TFLInterpreterOptions *options = [[TFLInterpreterOptions alloc] init];
   XCTAssertNotNil(options);
   XCTAssertEqual(options.numberOfThreads, 0);
+  XCTAssertFalse(options.useXNNPACK);
 }
 
 - (void)testSetNumberOfThread {
@@ -44,6 +45,14 @@ NS_ASSUME_NONNULL_BEGIN
   XCTAssertEqual(options.numberOfThreads, 3);
 }
 
+- (void)testUseXNNPACK {
+  TFLInterpreterOptions *options = [[TFLInterpreterOptions alloc] init];
+  options.useXNNPACK = YES;
+  XCTAssertTrue(options.useXNNPACK);
+  options.useXNNPACK = NO;
+  XCTAssertFalse(options.useXNNPACK);
+}
+
 @end
 
 NS_ASSUME_NONNULL_END
diff --git a/tensorflow/lite/experimental/swift/Sources/Interpreter.swift b/tensorflow/lite/experimental/swift/Sources/Interpreter.swift
index b83c36c4e1d..3567822208d 100644
--- a/tensorflow/lite/experimental/swift/Sources/Interpreter.swift
+++ b/tensorflow/lite/experimental/swift/Sources/Interpreter.swift
@@ -39,6 +39,9 @@ public final class Interpreter {
   /// The underlying `TfLiteInterpreter` C pointer.
   private var cInterpreter: CInterpreter?
 
+  /// The underlying `TfLiteDelegate` C pointer for XNNPACK delegate.
+  private var cXNNPackDelegate: Delegate.CDelegate?
+
   /// Creates a new instance with the given values.
   ///
   /// - Parameters:
@@ -78,6 +81,14 @@ public final class Interpreter {
       )
     }
     delegates?.forEach { TfLiteInterpreterOptionsAddDelegate(cInterpreterOptions, $0.cDelegate) }
+
+    // Configure the XNNPack delegate after the other delegates explicitly added by the user.
+    options.map {
+      if $0.isXNNPackEnabled {
+        configureXNNPack(options: $0, cInterpreterOptions: cInterpreterOptions)
+      }
+    }
+
     guard let cInterpreter = TfLiteInterpreterCreate(model.cModel, cInterpreterOptions) else {
       throw InterpreterError.failedToCreateInterpreter
     }
@@ -86,6 +97,7 @@ public final class Interpreter {
 
   deinit {
     TfLiteInterpreterDelete(cInterpreter)
+    TfLiteXNNPackDelegateDelete(cXNNPackDelegate)
   }
 
   /// Invokes the interpreter to perform inference from the loaded graph.
@@ -201,12 +213,13 @@ public final class Interpreter {
     guard case 0...maxIndex = index else {
       throw InterpreterError.invalidTensorIndex(index: index, maxIndex: maxIndex)
     }
-    guard TfLiteInterpreterResizeInputTensor(
-      cInterpreter,
-      Int32(index),
-      shape.int32Dimensions,
-      Int32(shape.rank)
-    ) == kTfLiteOk
+    guard
+      TfLiteInterpreterResizeInputTensor(
+        cInterpreter,
+        Int32(index),
+        shape.int32Dimensions,
+        Int32(shape.rank)
+      ) == kTfLiteOk
     else {
       throw InterpreterError.failedToResizeInputTensor(index: index)
     }
@@ -236,11 +249,11 @@ public final class Interpreter {
     }
 
     #if swift(>=5.0)
-    let status = data.withUnsafeBytes {
-      TfLiteTensorCopyFromBuffer(cTensor, $0.baseAddress, data.count)
-    }
+      let status = data.withUnsafeBytes {
+        TfLiteTensorCopyFromBuffer(cTensor, $0.baseAddress, data.count)
+      }
     #else
-    let status = data.withUnsafeBytes { TfLiteTensorCopyFromBuffer(cTensor, $0, data.count) }
+      let status = data.withUnsafeBytes { TfLiteTensorCopyFromBuffer(cTensor, $0, data.count) }
     #endif  // swift(>=5.0)
     guard status == kTfLiteOk else { throw InterpreterError.failedToCopyDataToInputTensor }
     return try input(at: index)
@@ -256,6 +269,18 @@ public final class Interpreter {
       throw InterpreterError.failedToAllocateTensors
     }
   }
+
+  // MARK: - Private
+
+  private func configureXNNPack(options: Options, cInterpreterOptions: OpaquePointer) {
+    var cXNNPackOptions = TfLiteXNNPackDelegateOptionsDefault()
+    if let threadCount = options.threadCount, threadCount > 0 {
+      cXNNPackOptions.num_threads = Int32(threadCount)
+    }
+
+    cXNNPackDelegate = TfLiteXNNPackDelegateCreate(&cXNNPackOptions)
+    TfLiteInterpreterOptionsAddDelegate(cInterpreterOptions, cXNNPackDelegate)
+  }
 }
 
 extension Interpreter {
@@ -265,6 +290,28 @@ extension Interpreter {
     /// indicating that the `Interpreter` will decide the number of threads to use.
     public var threadCount: Int? = nil
 
+    /// Indicates whether an optimized set of floating point CPU kernels, provided by XNNPACK, is
+    /// enabled.
+    ///
+    /// - Experiment:
+    /// Enabling this flag will enable use of a new, highly optimized set of CPU kernels provided
+    /// via the XNNPACK delegate. Currently, this is restricted to a subset of floating point
+    /// operations. Eventually, we plan to enable this by default, as it can provide significant
+    /// performance benefits for many classes of floating point models. See
+    /// https://github.com/tensorflow/tensorflow/blob/master/tensorflow/lite/delegates/xnnpack/README.md
+    /// for more details.
+    ///
+    /// - Important:
+    /// Things to keep in mind when enabling this flag:
+    ///
+    ///     * Startup time and resize time may increase.
+    ///     * Baseline memory consumption may increase.
+    ///     * Compatibility with other delegates (e.g., GPU) has not been fully validated.
+    ///     * Quantized models will not see any benefit.
+    ///
+    /// - Warning: This is an experimental interface that is subject to change.
+    public var isXNNPackEnabled: Bool = false
+
     /// Creates a new instance with the default values.
     public init() {}
   }
diff --git a/tensorflow/lite/experimental/swift/Tests/InterpreterTests.swift b/tensorflow/lite/experimental/swift/Tests/InterpreterTests.swift
index 8d0140279af..67d8120df4d 100644
--- a/tensorflow/lite/experimental/swift/Tests/InterpreterTests.swift
+++ b/tensorflow/lite/experimental/swift/Tests/InterpreterTests.swift
@@ -142,10 +142,12 @@ class InterpreterTests: XCTestCase {
   }
 
   func testResizeInputTensorAtIndexToShape_ThrowsInvalidIndex() {
-    XCTAssertThrowsError(try interpreter.resizeInput(
-      at: AddModel.invalidIndex,
-      to: [2, 2, 3]
-    )) { error in
+    XCTAssertThrowsError(
+      try interpreter.resizeInput(
+        at: AddModel.invalidIndex,
+        to: [2, 2, 3]
+      )
+    ) { error in
       let maxIndex = AddModel.inputTensorCount - 1
       self.assertEqualErrors(
         actual: error,
@@ -162,10 +164,12 @@ class InterpreterTests: XCTestCase {
   }
 
   func testCopyDataToInputTensorAtIndex_ThrowsInvalidIndex() {
-    XCTAssertThrowsError(try interpreter.copy(
-      AddModel.inputData,
-      toInputAt: AddModel.invalidIndex
-    )) { error in
+    XCTAssertThrowsError(
+      try interpreter.copy(
+        AddModel.inputData,
+        toInputAt: AddModel.invalidIndex
+      )
+    ) { error in
       let maxIndex = AddModel.inputTensorCount - 1
       self.assertEqualErrors(
         actual: error,
@@ -178,10 +182,12 @@ class InterpreterTests: XCTestCase {
     try interpreter.resizeInput(at: AddModel.validIndex, to: AddModel.shape)
     try interpreter.allocateTensors()
     let invalidData = Data(count: AddModel.dataCount - 1)
-    XCTAssertThrowsError(try interpreter.copy(
-      invalidData,
-      toInputAt: AddModel.validIndex
-    )) { error in
+    XCTAssertThrowsError(
+      try interpreter.copy(
+        invalidData,
+        toInputAt: AddModel.validIndex
+      )
+    ) { error in
       self.assertEqualErrors(
         actual: error,
         expected: .invalidTensorDataCount(provided: invalidData.count, required: AddModel.dataCount)
@@ -223,12 +229,20 @@ class InterpreterOptionsTests: XCTestCase {
   func testInitWithDefaultValues() {
     let options = Interpreter.Options()
     XCTAssertNil(options.threadCount)
+    XCTAssertFalse(options.isXNNPackEnabled)
   }
 
   func testInitWithCustomValues() {
     var options = Interpreter.Options()
+
     options.threadCount = 2
     XCTAssertEqual(options.threadCount, 2)
+
+    options.isXNNPackEnabled = false
+    XCTAssertFalse(options.isXNNPackEnabled)
+
+    options.isXNNPackEnabled = true
+    XCTAssertTrue(options.isXNNPackEnabled)
   }
 
   func testEquatable() {
@@ -242,6 +256,15 @@ class InterpreterOptionsTests: XCTestCase {
 
     options2.threadCount = 3
     XCTAssertNotEqual(options1, options2)
+
+    options2.threadCount = 2
+    XCTAssertEqual(options1, options2)
+
+    options2.isXNNPackEnabled = true
+    XCTAssertNotEqual(options1, options2)
+
+    options1.isXNNPackEnabled = true
+    XCTAssertEqual(options1, options2)
   }
 }
 
@@ -326,14 +349,15 @@ extension Array {
   init?(unsafeData: Data) {
     guard unsafeData.count % MemoryLayout.stride == 0 else { return nil }
     #if swift(>=5.0)
-    self = unsafeData.withUnsafeBytes { .init($0.bindMemory(to: Element.self)) }
+      self = unsafeData.withUnsafeBytes { .init($0.bindMemory(to: Element.self)) }
     #else
-    self = unsafeData.withUnsafeBytes {
-      .init(UnsafeBufferPointer(
-        start: $0,
-        count: unsafeData.count / MemoryLayout.stride
-      ))
-    }
+      self = unsafeData.withUnsafeBytes {
+        .init(
+          UnsafeBufferPointer(
+            start: $0,
+            count: unsafeData.count / MemoryLayout.stride
+          ))
+      }
     #endif  // swift(>=5.0)
   }
 }

From 4b0a6f818fa8e3c38fd0cf68d9a647f82cf6c93a Mon Sep 17 00:00:00 2001
From: "A. Unique TensorFlower" 
Date: Fri, 19 Jun 2020 02:01:50 -0700
Subject: [PATCH 283/867] Update GraphDef version to 437.

PiperOrigin-RevId: 317270285
Change-Id: Ib8d1e6dbb565c01d2bdf0304a03be1c1eebbde41
---
 tensorflow/core/public/version.h | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/tensorflow/core/public/version.h b/tensorflow/core/public/version.h
index 546d86e58fa..9a79fc1eddf 100644
--- a/tensorflow/core/public/version.h
+++ b/tensorflow/core/public/version.h
@@ -108,7 +108,7 @@ limitations under the License.
 
 #define TF_GRAPH_DEF_VERSION_MIN_PRODUCER 0
 #define TF_GRAPH_DEF_VERSION_MIN_CONSUMER 0
-#define TF_GRAPH_DEF_VERSION 436  // Updated: 2020/6/18
+#define TF_GRAPH_DEF_VERSION 437  // Updated: 2020/6/19
 
 // Checkpoint compatibility versions (the versions field in SavedSliceMeta).
 //

From d41d28120e6aac1efbad27523a78cd254434dc4e Mon Sep 17 00:00:00 2001
From: "A. Unique TensorFlower" 
Date: Fri, 19 Jun 2020 02:02:01 -0700
Subject: [PATCH 284/867] compat: Update forward compatibility horizon to
 2020-06-19

PiperOrigin-RevId: 317270310
Change-Id: Idc8188172496af9f2494c580cdab27558b16e4a8
---
 tensorflow/python/compat/compat.py | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/tensorflow/python/compat/compat.py b/tensorflow/python/compat/compat.py
index 32545ac8463..22988d26cfc 100644
--- a/tensorflow/python/compat/compat.py
+++ b/tensorflow/python/compat/compat.py
@@ -33,7 +33,7 @@ from tensorflow.python.util.tf_export import tf_export
 # This value changes every day with an automatic CL. It can be modified in code
 # via `forward_compatibility_horizon()` or with the environment variable
 # TF_FORWARD_COMPATIBILITY_DELTA_DAYS, which is added to the compatibility date.
-_FORWARD_COMPATIBILITY_HORIZON = datetime.date(2020, 6, 18)
+_FORWARD_COMPATIBILITY_HORIZON = datetime.date(2020, 6, 19)
 _FORWARD_COMPATIBILITY_DELTA_DAYS_VAR_NAME = "TF_FORWARD_COMPATIBILITY_DELTA_DAYS"
 _FORWARD_COMPATIBILITY_DATE_NUMBER = None
 

From 9d1ec55aed0a4d9baf7302974fefe08546bfad25 Mon Sep 17 00:00:00 2001
From: "A. Unique TensorFlower" 
Date: Fri, 19 Jun 2020 02:04:36 -0700
Subject: [PATCH 285/867] Integrate LLVM at
 https://github.com/llvm/llvm-project/commit/7f0d7f326316

PiperOrigin-RevId: 317270655
Change-Id: Ic80ab697da45212c8d58bcda989e5ee0a330b565
---
 .../mlir/xla/transforms/hlo_legalize_to_lhlo.cc     | 13 +++++++------
 .../compiler/mlir/xla/transforms/legalize_tf.cc     |  4 ++--
 .../mlir/xla/transforms/lhlo_legalize_to_gpu.cc     |  2 +-
 .../xla/transforms/lhlo_legalize_to_llvm_pass.cc    |  2 +-
 .../transforms/lhlo_legalize_to_parallel_loops.cc   |  2 +-
 .../mlir/xla/transforms/xla_legalize_to_linalg.cc   |  4 ++--
 .../xla/service/mlir_gpu/kernel_lowering.cc         |  2 +-
 tensorflow/workspace.bzl                            |  4 ++--
 8 files changed, 17 insertions(+), 16 deletions(-)

diff --git a/tensorflow/compiler/mlir/xla/transforms/hlo_legalize_to_lhlo.cc b/tensorflow/compiler/mlir/xla/transforms/hlo_legalize_to_lhlo.cc
index 1cfe0c12e20..a11b08e0ea6 100644
--- a/tensorflow/compiler/mlir/xla/transforms/hlo_legalize_to_lhlo.cc
+++ b/tensorflow/compiler/mlir/xla/transforms/hlo_legalize_to_lhlo.cc
@@ -389,10 +389,13 @@ struct HloLegalizeToLhlo
     target.addLegalOp();
     target.addLegalOp();
     target.addIllegalDialect();
+
+    BufferAssignmentTypeConverter converter;
     target.addDynamicallyLegalOp([&](FuncOp op) {
       auto inputs = op.getType().getInputs();
-      return std::all_of(inputs.begin(), inputs.end(),
-                         [](Type input) { return input.isa(); });
+      return llvm::all_of(inputs,
+                          [](Type input) { return input.isa(); }) &&
+             converter.isLegal(&op.getBody());
     });
     target.addDynamicallyLegalOp([&](mlir::ReturnOp returnOp) {
       return std::all_of(returnOp.operand_type_begin(),
@@ -401,8 +404,7 @@ struct HloLegalizeToLhlo
     });
 
     auto module = getOperation();
-    BufferAssignmentTypeConverter converter;
-    module.walk([&](FuncOp func) {
+    module.walk([&](FuncOp func) -> WalkResult {
       BufferAssignmentPlacer bufferAssignment(func);
       OwningRewritePatternList patterns;
       populateHLOToLHLOConversionPattern(func.getContext(), &bufferAssignment,
@@ -418,8 +420,7 @@ struct HloLegalizeToLhlo
             /*allowMemrefFunctionResults=*/false>(&context, &bufferAssignment,
                                                   &converter, &patterns);
       }
-      return WalkResult(
-          applyPartialConversion(func, target, patterns, &converter));
+      return applyPartialConversion(func, target, patterns);
     });
   }
 
diff --git a/tensorflow/compiler/mlir/xla/transforms/legalize_tf.cc b/tensorflow/compiler/mlir/xla/transforms/legalize_tf.cc
index b7cad554043..1788cd1b270 100644
--- a/tensorflow/compiler/mlir/xla/transforms/legalize_tf.cc
+++ b/tensorflow/compiler/mlir/xla/transforms/legalize_tf.cc
@@ -5238,8 +5238,8 @@ LogicalResult legalizeTF(Operation *op, bool allow_partial_conversion,
     // Fully qualify ReturnOp here as xla_hlo dialect also defines a ReturnOp.
     target.addLegalOp();
     DenseSet nonlegalized_ops;
-    LogicalResult result = applyPartialConversion(
-        op, target, patterns, /*converter=*/nullptr, &nonlegalized_ops);
+    LogicalResult result =
+        applyPartialConversion(op, target, patterns, &nonlegalized_ops);
     // In order to enforce that the conversion result is fully converted,
     // fail if there are any nonlegalized ops in the set.
     if (failed(result) || !nonlegalized_ops.empty()) {
diff --git a/tensorflow/compiler/mlir/xla/transforms/lhlo_legalize_to_gpu.cc b/tensorflow/compiler/mlir/xla/transforms/lhlo_legalize_to_gpu.cc
index f0eb3cc1a0f..c23b8b49268 100644
--- a/tensorflow/compiler/mlir/xla/transforms/lhlo_legalize_to_gpu.cc
+++ b/tensorflow/compiler/mlir/xla/transforms/lhlo_legalize_to_gpu.cc
@@ -177,7 +177,7 @@ struct LhloLegalizeToGpu : public PassWrapper {
     target.addIllegalOp();
     auto func = getFunction();
     patterns.insert(func.getContext());
-    if (failed(applyPartialConversion(func, target, patterns, nullptr))) {
+    if (failed(applyPartialConversion(func, target, patterns))) {
       signalPassFailure();
     }
   }
diff --git a/tensorflow/compiler/mlir/xla/transforms/lhlo_legalize_to_llvm_pass.cc b/tensorflow/compiler/mlir/xla/transforms/lhlo_legalize_to_llvm_pass.cc
index 9b809049290..63265c4a7e7 100644
--- a/tensorflow/compiler/mlir/xla/transforms/lhlo_legalize_to_llvm_pass.cc
+++ b/tensorflow/compiler/mlir/xla/transforms/lhlo_legalize_to_llvm_pass.cc
@@ -43,7 +43,7 @@ class TestLhloToLLVMPass
     target.addLegalOp();
     target.addIllegalDialect();
 
-    if (failed(applyFullConversion(m, target, patterns, &converter))) {
+    if (failed(applyFullConversion(m, target, patterns))) {
       signalPassFailure();
     }
   }
diff --git a/tensorflow/compiler/mlir/xla/transforms/lhlo_legalize_to_parallel_loops.cc b/tensorflow/compiler/mlir/xla/transforms/lhlo_legalize_to_parallel_loops.cc
index b3112d49103..65962c5b7a5 100644
--- a/tensorflow/compiler/mlir/xla/transforms/lhlo_legalize_to_parallel_loops.cc
+++ b/tensorflow/compiler/mlir/xla/transforms/lhlo_legalize_to_parallel_loops.cc
@@ -711,7 +711,7 @@ struct LhloLegalizeToParallelLoops
     target.addIllegalOp();
 
-    if (failed(applyPartialConversion(func, target, patterns, nullptr))) {
+    if (failed(applyPartialConversion(func, target, patterns))) {
       signalPassFailure();
     }
   }
diff --git a/tensorflow/compiler/mlir/xla/transforms/xla_legalize_to_linalg.cc b/tensorflow/compiler/mlir/xla/transforms/xla_legalize_to_linalg.cc
index ad78a01100b..8a2f8ce7d04 100644
--- a/tensorflow/compiler/mlir/xla/transforms/xla_legalize_to_linalg.cc
+++ b/tensorflow/compiler/mlir/xla/transforms/xla_legalize_to_linalg.cc
@@ -867,7 +867,7 @@ struct LhloLegalizeToLinalg
 
     auto func = getFunction();
     populateLHLOToLinalgConversionPattern(func.getContext(), &patterns);
-    if (failed(applyPartialConversion(func, target, patterns, nullptr))) {
+    if (failed(applyPartialConversion(func, target, patterns))) {
       signalPassFailure();
     }
   }
@@ -882,7 +882,7 @@ struct HloLegalizeToLinalg
 
     auto func = getFunction();
     xla_hlo::populateHLOToLinalgConversionPattern(func.getContext(), &patterns);
-    if (failed(applyPartialConversion(func, target, patterns, nullptr))) {
+    if (failed(applyPartialConversion(func, target, patterns))) {
       signalPassFailure();
     }
   }
diff --git a/tensorflow/compiler/xla/service/mlir_gpu/kernel_lowering.cc b/tensorflow/compiler/xla/service/mlir_gpu/kernel_lowering.cc
index 9d5b52df010..ecd1308be4b 100644
--- a/tensorflow/compiler/xla/service/mlir_gpu/kernel_lowering.cc
+++ b/tensorflow/compiler/xla/service/mlir_gpu/kernel_lowering.cc
@@ -552,7 +552,7 @@ class LowerToNVVMPass
     // TODO(csigg): Remove once we support replacing non-root ops.
     target.addLegalOp<::mlir::gpu::GPUModuleOp, ::mlir::gpu::ModuleEndOp,
                       ::mlir::gpu::YieldOp>();
-    if (failed(mlir::applyFullConversion(m, target, patterns, &converter))) {
+    if (failed(mlir::applyFullConversion(m, target, patterns))) {
       signalPassFailure();
     }
   }
diff --git a/tensorflow/workspace.bzl b/tensorflow/workspace.bzl
index 52c573628ac..27eca0ee54f 100755
--- a/tensorflow/workspace.bzl
+++ b/tensorflow/workspace.bzl
@@ -710,8 +710,8 @@ def tf_repositories(path_prefix = "", tf_repo_name = ""):
     )
 
     # Check out LLVM and MLIR from llvm-project.
-    LLVM_COMMIT = "92d8ad02e92fed3884169ba5d98056fe4fa5660d"
-    LLVM_SHA256 = "a4995ace7ddaef0c49293dc65771f58ef1fea96ebe1f39aa0a2d6d75d07f6cc7"
+    LLVM_COMMIT = "7f0d7f32631648acf48bc23047635ab5e2058a1a"
+    LLVM_SHA256 = "2f1dbae231b3b8f9c67d6a4f578c8ce29f3aa2831313b34c40ff2edb4014476a"
     LLVM_URLS = [
         "https://storage.googleapis.com/mirror.tensorflow.org/github.com/llvm/llvm-project/archive/{commit}.tar.gz".format(commit = LLVM_COMMIT),
         "https://github.com/llvm/llvm-project/archive/{commit}.tar.gz".format(commit = LLVM_COMMIT),

From c662daf4891a1e6efe64797615c3bd2bebedc5f5 Mon Sep 17 00:00:00 2001
From: Smit Hinsu 
Date: Fri, 19 Jun 2020 02:22:43 -0700
Subject: [PATCH 286/867] Override CustomCall in MlirHloBuilder

Also, enable mlir bridge for image ops compilers test. ResizeBilinear op
lowering usese CustomCall in case of TPU lowerings.

PiperOrigin-RevId: 317272443
Change-Id: I134c828cdc76552a0cbfdeb7c65532aa986314e2
---
 .../compiler/mlir/xla/ir/mlir_hlo_builder.cc  | 16 ++++++++++++
 .../compiler/mlir/xla/ir/mlir_hlo_builder.h   |  6 +++++
 .../xla/transforms/legalize_tf_with_tf2xla.cc |  8 ++++++
 tensorflow/compiler/tests/BUILD               |  1 +
 tensorflow/compiler/xla/client/xla_builder.cc | 26 ++++++++++++++-----
 tensorflow/compiler/xla/client/xla_builder.h  |  8 ++++++
 6 files changed, 58 insertions(+), 7 deletions(-)

diff --git a/tensorflow/compiler/mlir/xla/ir/mlir_hlo_builder.cc b/tensorflow/compiler/mlir/xla/ir/mlir_hlo_builder.cc
index 21b1ac5f0ea..3c11d8e590d 100644
--- a/tensorflow/compiler/mlir/xla/ir/mlir_hlo_builder.cc
+++ b/tensorflow/compiler/mlir/xla/ir/mlir_hlo_builder.cc
@@ -132,6 +132,22 @@ StatusOr MlirHloBuilder::FftInternal(
   return MakeXlaOp(op);
 }
 
+StatusOr MlirHloBuilder::CustomCallInternal(
+    const string& call_target_name, absl::Span operands,
+    const Shape& shape, const string& opaque,
+    absl::optional> operand_shapes_with_layout) {
+  if (operand_shapes_with_layout.has_value())
+    return Unimplemented(
+        "CustomCall doesn't support operands shapes with layout");
+  TF_ASSIGN_OR_RETURN(mlir::Type ty, ConvertShapeToType(
+                                         shape, builder_));
+  auto op = builder_.create(
+      loc_, ty, GetValues(operands), builder_.getStringAttr(call_target_name),
+      /*has_side_effect=*/builder_.getBoolAttr(false),
+      builder_.getStringAttr(opaque));
+  return MakeXlaOp(op);
+}
+
 StatusOr MlirHloBuilder::ReduceInternal(
     const Shape& shape, absl::Span all_operands,
     const XlaComputation& computation,
diff --git a/tensorflow/compiler/mlir/xla/ir/mlir_hlo_builder.h b/tensorflow/compiler/mlir/xla/ir/mlir_hlo_builder.h
index 4b28c32db99..4d7d93af7a7 100644
--- a/tensorflow/compiler/mlir/xla/ir/mlir_hlo_builder.h
+++ b/tensorflow/compiler/mlir/xla/ir/mlir_hlo_builder.h
@@ -124,6 +124,12 @@ class MlirHloBuilder : public XlaBuilder {
                               FftType fft_type,
                               absl::Span fft_length) override;
 
+  StatusOr CustomCallInternal(const string& call_target_name,
+                                     absl::Span operands,
+                                     const Shape& shape, const string& opaque,
+                                     absl::optional>
+                                         operand_shapes_with_layout) override;
+
   StatusOr ReduceInternal(
       const Shape& shape, absl::Span all_operands,
       const XlaComputation& computation,
diff --git a/tensorflow/compiler/mlir/xla/transforms/legalize_tf_with_tf2xla.cc b/tensorflow/compiler/mlir/xla/transforms/legalize_tf_with_tf2xla.cc
index ef79c8868bb..8f96f4d1305 100644
--- a/tensorflow/compiler/mlir/xla/transforms/legalize_tf_with_tf2xla.cc
+++ b/tensorflow/compiler/mlir/xla/transforms/legalize_tf_with_tf2xla.cc
@@ -88,6 +88,9 @@ static bool IsOpWhitelisted(Operation* op) {
     TypeID::get(),
     TypeID::get(),
     TypeID::get(),
+    TypeID::get(),
+    TypeID::get(),
+    TypeID::get(),
     TypeID::get(),
     TypeID::get(),
     TypeID::get(),
@@ -127,6 +130,7 @@ static bool IsOpWhitelisted(Operation* op) {
     TypeID::get(),
     TypeID::get(),
     TypeID::get(),
+    TypeID::get(),
     TypeID::get(),
     TypeID::get(),
     TypeID::get(),
@@ -157,10 +161,14 @@ static bool IsOpWhitelisted(Operation* op) {
     TypeID::get(),
     TypeID::get(),
     TypeID::get(),
+    TypeID::get(),
     TypeID::get(),
     TypeID::get(),
     TypeID::get(),
     TypeID::get(),
+    TypeID::get(),
+    TypeID::get(),
+    TypeID::get(),
     TypeID::get(),
     TypeID::get(),
     TypeID::get(),
diff --git a/tensorflow/compiler/tests/BUILD b/tensorflow/compiler/tests/BUILD
index b574622efce..034ec82de10 100644
--- a/tensorflow/compiler/tests/BUILD
+++ b/tensorflow/compiler/tests/BUILD
@@ -770,6 +770,7 @@ tf_xla_py_test(
     size = "small",
     timeout = "long",
     srcs = ["image_ops_test.py"],
+    enable_mlir_bridge = True,
     python_version = "PY3",
     shard_count = 10,
     tags = [
diff --git a/tensorflow/compiler/xla/client/xla_builder.cc b/tensorflow/compiler/xla/client/xla_builder.cc
index c7b6a7f9491..03ae23ea18b 100644
--- a/tensorflow/compiler/xla/client/xla_builder.cc
+++ b/tensorflow/compiler/xla/client/xla_builder.cc
@@ -1564,16 +1564,12 @@ XlaOp XlaBuilder::CustomCall(
     const Shape& shape, const string& opaque,
     absl::optional> operand_shapes_with_layout) {
   return ReportErrorOrReturn([&]() -> StatusOr {
-    HloInstructionProto instr;
     if (absl::StartsWith(call_target_name, "$")) {
       return InvalidArgument(
           "Invalid custom_call_target \"%s\": Call targets that start with '$' "
           "are reserved for internal use.",
           call_target_name);
     }
-    *instr.mutable_shape() = shape.ToProto();
-    instr.set_custom_call_target(call_target_name);
-    instr.set_backend_config(opaque);
     if (operand_shapes_with_layout.has_value()) {
       if (!LayoutUtil::HasLayout(shape)) {
         return InvalidArgument(
@@ -1586,7 +1582,6 @@ XlaOp XlaBuilder::CustomCall(
             "with constrained layout; given %d shapes, expected %d",
             operand_shapes_with_layout->size(), operands.size());
       }
-      instr.set_constrain_layout(true);
       int64 operand_num = 0;
       for (const Shape& operand_shape : *operand_shapes_with_layout) {
         if (!LayoutUtil::HasLayout(operand_shape)) {
@@ -1595,14 +1590,31 @@ XlaOp XlaBuilder::CustomCall(
               "constrained layout.",
               operand_num);
         }
-        *instr.add_operand_shapes_with_layout() = operand_shape.ToProto();
         ++operand_num;
       }
     }
-    return AddInstruction(std::move(instr), HloOpcode::kCustomCall, operands);
+    return CustomCallInternal(call_target_name, operands, shape, opaque,
+                              operand_shapes_with_layout);
   });
 }
 
+StatusOr XlaBuilder::CustomCallInternal(
+    const string& call_target_name, absl::Span operands,
+    const Shape& shape, const string& opaque,
+    absl::optional> operand_shapes_with_layout) {
+  HloInstructionProto instr;
+  *instr.mutable_shape() = shape.ToProto();
+  instr.set_custom_call_target(call_target_name);
+  instr.set_backend_config(opaque);
+  if (operand_shapes_with_layout.has_value()) {
+    instr.set_constrain_layout(true);
+    for (const Shape& operand_shape : *operand_shapes_with_layout) {
+      *instr.add_operand_shapes_with_layout() = operand_shape.ToProto();
+    }
+  }
+  return AddInstruction(std::move(instr), HloOpcode::kCustomCall, operands);
+}
+
 XlaOp XlaBuilder::CustomCall(
     const string& call_target_name, absl::Span operands,
     const XlaComputation& computation, const Shape& shape, const string& opaque,
diff --git a/tensorflow/compiler/xla/client/xla_builder.h b/tensorflow/compiler/xla/client/xla_builder.h
index b8af180b83e..3fc26747468 100644
--- a/tensorflow/compiler/xla/client/xla_builder.h
+++ b/tensorflow/compiler/xla/client/xla_builder.h
@@ -527,6 +527,14 @@ class XlaBuilder {
       const Shape& shape_with_layout, const string& opaque,
       absl::optional> operand_shapes_with_layout);
 
+  // Internal version of CustomCall without computation that doesn't do op
+  // specific error handling and expects arguments to be legal. CustomCall
+  // method above calls this method after error handling.
+  virtual StatusOr CustomCallInternal(
+      const string& call_target_name, absl::Span operands,
+      const Shape& shape_with_layout, const string& opaque,
+      absl::optional> operand_shapes_with_layout);
+
   XlaOp CustomCall(
       const string& call_target_name, absl::Span operands,
       const XlaComputation& computation, const Shape& shape_with_layout,

From 8ad5bc80e71921c3c2530d93d3856ba59e524c60 Mon Sep 17 00:00:00 2001
From: Lukas Geiger 
Date: Fri, 19 Jun 2020 11:50:24 +0200
Subject: [PATCH 287/867] Remove unnecessary assert

---
 .../mixed_precision/experimental/autocast_variable_test.py      | 2 --
 1 file changed, 2 deletions(-)

diff --git a/tensorflow/python/keras/mixed_precision/experimental/autocast_variable_test.py b/tensorflow/python/keras/mixed_precision/experimental/autocast_variable_test.py
index 2fa7c103258..9036109af96 100644
--- a/tensorflow/python/keras/mixed_precision/experimental/autocast_variable_test.py
+++ b/tensorflow/python/keras/mixed_precision/experimental/autocast_variable_test.py
@@ -372,7 +372,6 @@ class AutoCastVariableTest(test.TestCase, parameterized.TestCase):
         # Variable should be increased, despite it appearing to be the same
         # float16 value.
         self.evaluate(x.assign(1. + small_tensor))
-        self.assertEqual(1. + small_val, self.evaluate(x._variable))
         self.assertEqual(1., self.evaluate(x.value()))
       self.assertEqual(1. + small_val, self.evaluate(x))
 
@@ -380,7 +379,6 @@ class AutoCastVariableTest(test.TestCase, parameterized.TestCase):
       with ops.get_default_graph()._enable_auto_casting_variables(
           dtypes.float16):
         self.evaluate(x.assign_add(small_tensor))
-        self.assertEqual(1. + small_val, self.evaluate(x._variable))
         self.assertEqual(1., self.evaluate(x.value()))
       self.assertEqual(1. + small_val, self.evaluate(x))
 

From dc8d42922b9ff89e717f130515c968186ec4504c Mon Sep 17 00:00:00 2001
From: Lukas Geiger 
Date: Fri, 19 Jun 2020 12:33:01 +0200
Subject: [PATCH 288/867] Remove unnecessary control_dependencies

---
 .../mixed_precision/experimental/autocast_variable.py | 11 +++--------
 1 file changed, 3 insertions(+), 8 deletions(-)

diff --git a/tensorflow/python/keras/mixed_precision/experimental/autocast_variable.py b/tensorflow/python/keras/mixed_precision/experimental/autocast_variable.py
index ca6420f0c0b..b60100c7b48 100644
--- a/tensorflow/python/keras/mixed_precision/experimental/autocast_variable.py
+++ b/tensorflow/python/keras/mixed_precision/experimental/autocast_variable.py
@@ -190,25 +190,20 @@ class AutoCastVariable(variables.Variable, core.Tensor):
 
   def _apply_assign_update(
       self, update_fn, value, use_locking=None, name=None, read_value=True):
-    if not read_value:
-      return update_fn(value, use_locking, name, read_value)
-
     if context.executing_eagerly() or ops.inside_function():
       assign_op = update_fn(value, use_locking, name, False)
-      with ops.control_dependencies([assign_op]):
-        return self
+      return self if read_value else assign_op
 
     # Fallback to wrapping the returned variable in graph mode if possible
     assign_var = update_fn(value, use_locking, name, read_value)
-    if resource_variable_ops.is_resource_variable(assign_var):
+    if read_value and resource_variable_ops.is_resource_variable(assign_var):
       return create_autocast_variable(assign_var)
     return assign_var
 
   def _apply_update(self, update_fn, *args, **kwargs):
     update_var = update_fn(*args, **kwargs)
     if context.executing_eagerly() or ops.inside_function():
-      with ops.control_dependencies([update_var]):
-        return self
+      return self
 
     # Fallback to wrapping the returned variable in graph mode if possible
     if resource_variable_ops.is_resource_variable(update_var):

From b58e6000457f26c7a53a9b945642fbe2baddbf20 Mon Sep 17 00:00:00 2001
From: Alexander Belyaev 
Date: Fri, 19 Jun 2020 04:25:24 -0700
Subject: [PATCH 289/867] [XLA][MLIR] Enable xla_hlo.ReshapeOp ->
 xla_lhlo.ReshapeOp conversion.

PiperOrigin-RevId: 317284676
Change-Id: Ia845183efcfabe77f6eb66d8c56dcbfc82653982
---
 tensorflow/compiler/mlir/xla/transforms/hlo_legalize_to_lhlo.cc | 1 +
 tensorflow/compiler/mlir/xla/transforms/map_hlo_to_lhlo_op.h    | 1 +
 2 files changed, 2 insertions(+)

diff --git a/tensorflow/compiler/mlir/xla/transforms/hlo_legalize_to_lhlo.cc b/tensorflow/compiler/mlir/xla/transforms/hlo_legalize_to_lhlo.cc
index a11b08e0ea6..446f2aae833 100644
--- a/tensorflow/compiler/mlir/xla/transforms/hlo_legalize_to_lhlo.cc
+++ b/tensorflow/compiler/mlir/xla/transforms/hlo_legalize_to_lhlo.cc
@@ -464,6 +464,7 @@ void populateHLOToLHLOConversionPattern(
       HloToLhloOpConverter,
       HloToLhloOpConverter,
       HloToLhloOpConverter,
+      HloToLhloOpConverter,
       HloToLhloOpConverter,
       HloToLhloOpConverter,
       HloToLhloOpConverter,
diff --git a/tensorflow/compiler/mlir/xla/transforms/map_hlo_to_lhlo_op.h b/tensorflow/compiler/mlir/xla/transforms/map_hlo_to_lhlo_op.h
index 4b9397795a1..8d5f27474a5 100644
--- a/tensorflow/compiler/mlir/xla/transforms/map_hlo_to_lhlo_op.h
+++ b/tensorflow/compiler/mlir/xla/transforms/map_hlo_to_lhlo_op.h
@@ -61,6 +61,7 @@ MAP_HLO_TO_LHLO(MulOp);
 MAP_HLO_TO_LHLO(NegOp);
 MAP_HLO_TO_LHLO(RealOp);
 MAP_HLO_TO_LHLO(ReduceOp);
+MAP_HLO_TO_LHLO(ReshapeOp);
 MAP_HLO_TO_LHLO(RemOp);
 MAP_HLO_TO_LHLO(RsqrtOp);
 MAP_HLO_TO_LHLO(SelectOp);

From 7b1a726ec3ab2d831167d32b9b820bcf14aba6f7 Mon Sep 17 00:00:00 2001
From: Devi Sandeep Endluri 
Date: Fri, 19 Jun 2020 06:23:39 -0500
Subject: [PATCH 290/867] Ensure there are test samples for imdb dataset, when
 maxlen is low

With the current imdb.load_data(), the following results are seen
for different values of maxlen.

	load_data                (len(x_train), len(x_test))
------------------------------------------------------------
imdb.load_data(maxlen=50)    -->    (1035, 0)
imdb.load_data(maxlen=100)   -->    (5736, 0)
imdb.load_data(maxlen=200)   -->    (25000, 3913)
imdb.load_data()             -->    (25000, 25000)

Analysis: We can observe that when maxlen is low, the number
of test samples can be 0. This is because the train and test data is
concatenated, then the samples with length > maxlen are removed, and
the first 25,000 are considered as training data.

Fix: This can be fixed when data can be filtered first to remove the
ones with length > maxlen, and then concatenate to process further.
The following are the results after the fix.

     fixed load_data              (len(x_train), len(x_test))
------------------------------------------------------------
imdb.load_data(maxlen=50)    -->    (477, 558)
imdb.load_data(maxlen=100)   -->    (2773, 2963)
imdb.load_data(maxlen=200)   -->    (14244, 14669)
imdb.load_data()             -->    (25000, 25000)
---
 tensorflow/python/keras/datasets/imdb.py | 18 +++++++++++-------
 1 file changed, 11 insertions(+), 7 deletions(-)

diff --git a/tensorflow/python/keras/datasets/imdb.py b/tensorflow/python/keras/datasets/imdb.py
index 37403228edf..e359d691a5d 100644
--- a/tensorflow/python/keras/datasets/imdb.py
+++ b/tensorflow/python/keras/datasets/imdb.py
@@ -124,20 +124,24 @@ def load_data(path='imdb.npz',
   x_test = x_test[indices]
   labels_test = labels_test[indices]
 
-  xs = np.concatenate([x_train, x_test])
-  labels = np.concatenate([labels_train, labels_test])
-
   if start_char is not None:
-    xs = [[start_char] + [w + index_from for w in x] for x in xs]
+    x_train = [[start_char] + [w + index_from for w in x] for x in x_train]
+    x_test = [[start_char] + [w + index_from for w in x] for x in x_test]
   elif index_from:
-    xs = [[w + index_from for w in x] for x in xs]
+    x_train = [[w + index_from for w in x] for x in x_train]
+    x_test = [[w + index_from for w in x] for x in x_test]
 
   if maxlen:
-    xs, labels = _remove_long_seq(maxlen, xs, labels)
-    if not xs:
+    x_train, labels_train = _remove_long_seq(maxlen, x_train, labels_train)
+    x_test, labels_test = _remove_long_seq(maxlen, x_test, labels_test)
+    if not x_train or not x_test:
       raise ValueError('After filtering for sequences shorter than maxlen=' +
                        str(maxlen) + ', no sequence was kept. '
                        'Increase maxlen.')
+
+  xs = np.concatenate([x_train, x_test])
+  labels = np.concatenate([labels_train, labels_test])
+
   if not num_words:
     num_words = max(max(x) for x in xs)
 

From 6ddd920c44540284ab28cc7cdf2faef18b85ff11 Mon Sep 17 00:00:00 2001
From: Xinan Jiang 
Date: Fri, 19 Jun 2020 20:51:30 +0800
Subject: [PATCH 291/867] [MLIR][XLA] Modify for FileCheck commands

---
 tensorflow/compiler/xla/service/mlir_gpu/tests/gather.hlo | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/tensorflow/compiler/xla/service/mlir_gpu/tests/gather.hlo b/tensorflow/compiler/xla/service/mlir_gpu/tests/gather.hlo
index 99faa319bf6..8a24daae0f4 100644
--- a/tensorflow/compiler/xla/service/mlir_gpu/tests/gather.hlo
+++ b/tensorflow/compiler/xla/service/mlir_gpu/tests/gather.hlo
@@ -11,5 +11,5 @@ ENTRY %Gather (x: f32[100,10], y: s64[4,6]) -> f32[4,6,10] {
 
 // CHECK: func @gather(%[[ARG0:.*]]: [[TYPE0:.*]], %[[ARG1:.*]]: [[TYPE1:.*]], %[[RESULT:.*]]: [[RTYPE:.*]]) {
 // CHECK:   "xla_lhlo.gather"(%[[ARG0]], %[[ARG1]], %[[RESULT]])
-// CHECK:   {collapsed_slice_dims = dense<0> : tensor<1xi64>, index_vector_dim = 2 : i64, offset_dims = dense<2> : tensor<1xi64>, slice_sizes = dense<[1, 10]> : tensor<2xi64>, start_index_map = dense<0> : tensor<1xi64>} : ([[TYPE0]], [[TYPE1]], [[RTYPE]]) -> ()
+// CHECK-SAME:   {collapsed_slice_dims = dense<0> : tensor<1xi64>, index_vector_dim = 2 : i64, offset_dims = dense<2> : tensor<1xi64>, slice_sizes = dense<[1, 10]> : tensor<2xi64>, start_index_map = dense<0> : tensor<1xi64>} : ([[TYPE0]], [[TYPE1]], [[RTYPE]]) -> ()
 // CHECK: }

From 42579858f9cda701c7c69d4a1f89035f0a68b258 Mon Sep 17 00:00:00 2001
From: Xinan Jiang 
Date: Fri, 19 Jun 2020 20:56:08 +0800
Subject: [PATCH 292/867] [MLIR][XLA] Add GatherOp to HLO to LHLO converters

---
 tensorflow/compiler/mlir/xla/transforms/hlo_legalize_to_lhlo.cc | 1 +
 tensorflow/compiler/mlir/xla/transforms/map_hlo_to_lhlo_op.h    | 1 +
 2 files changed, 2 insertions(+)

diff --git a/tensorflow/compiler/mlir/xla/transforms/hlo_legalize_to_lhlo.cc b/tensorflow/compiler/mlir/xla/transforms/hlo_legalize_to_lhlo.cc
index 1cfe0c12e20..524965cfacd 100644
--- a/tensorflow/compiler/mlir/xla/transforms/hlo_legalize_to_lhlo.cc
+++ b/tensorflow/compiler/mlir/xla/transforms/hlo_legalize_to_lhlo.cc
@@ -453,6 +453,7 @@ void populateHLOToLHLOConversionPattern(
       HloToLhloOpConverter,
       HloToLhloOpConverter,
       HloToLhloOpConverter,
+      HloToLhloOpConverter,
       HloToLhloOpConverter,
       HloToLhloOpConverter,
       HloToLhloOpConverter,
diff --git a/tensorflow/compiler/mlir/xla/transforms/map_hlo_to_lhlo_op.h b/tensorflow/compiler/mlir/xla/transforms/map_hlo_to_lhlo_op.h
index 4b9397795a1..5e3d1cb9302 100644
--- a/tensorflow/compiler/mlir/xla/transforms/map_hlo_to_lhlo_op.h
+++ b/tensorflow/compiler/mlir/xla/transforms/map_hlo_to_lhlo_op.h
@@ -52,6 +52,7 @@ MAP_HLO_TO_LHLO(CosOp);
 MAP_HLO_TO_LHLO(DivOp);
 MAP_HLO_TO_LHLO(DotOp);
 MAP_HLO_TO_LHLO(ExpOp);
+MAP_HLO_TO_LHLO(GatherOp);
 MAP_HLO_TO_LHLO(ImagOp);
 MAP_HLO_TO_LHLO(IotaOp);
 MAP_HLO_TO_LHLO(LogOp);

From 0c7e61d6608e29324357fb5f79a9b925281521a0 Mon Sep 17 00:00:00 2001
From: Hanhan Wang 
Date: Fri, 19 Jun 2020 07:09:13 -0700
Subject: [PATCH 293/867] Remove the canonicalize pattern for folding a pad op
 into the following conv op.

Basically rolledback for cl/305641881, the pattern could hurt performance
because the operation can't be fully tiled in Linalg transformation. In this
context, not everyone wants this pattern, so remove it from canonicalize
patterns.

PiperOrigin-RevId: 317302072
Change-Id: I19aa64e14eecccfd738ad3f775f3670974bc68f9
---
 tensorflow/compiler/mlir/xla/ir/hlo_ops.cc    | 56 ----------------
 tensorflow/compiler/mlir/xla/ir/hlo_ops.td    |  2 -
 .../compiler/mlir/xla/tests/canonicalize.mlir | 65 -------------------
 .../mlir/xla/transforms/canonicalize.td       | 51 ---------------
 4 files changed, 174 deletions(-)

diff --git a/tensorflow/compiler/mlir/xla/ir/hlo_ops.cc b/tensorflow/compiler/mlir/xla/ir/hlo_ops.cc
index d7950919883..e0fa1da93b8 100644
--- a/tensorflow/compiler/mlir/xla/ir/hlo_ops.cc
+++ b/tensorflow/compiler/mlir/xla/ir/hlo_ops.cc
@@ -106,53 +106,6 @@ DenseIntElementsAttr BuildSliceLimits(DenseIntElementsAttr start_indices,
   return GetI64ElementsAttr(slice_limits, builder);
 }
 
-// Returns the padding value of the given position. If padding_attr is a
-// nullptr, returns 0.
-static int64_t GetPaddingValue(DenseIntElementsAttr padding_attr,
-                               ArrayRef index) {
-  if (!padding_attr) return 0;
-  return padding_attr.getValue(index);
-}
-
-static bool IsOnlyPaddingSpatialDims(Value lhs,
-                                     ConvDimensionNumbers dimension_numbers,
-                                     DenseIntElementsAttr edge_padding_low,
-                                     DenseIntElementsAttr edge_padding_high) {
-  const int64_t batch_dim = dimension_numbers.input_batch_dimension().getInt();
-  const int64_t feature_dim =
-      dimension_numbers.input_feature_dimension().getInt();
-  if (edge_padding_low.getValue(batch_dim) ||
-      edge_padding_high.getValue(batch_dim))
-    return false;
-  if (edge_padding_low.getValue(feature_dim) ||
-      edge_padding_high.getValue(feature_dim))
-    return false;
-  return true;
-}
-
-DenseIntElementsAttr BuildConvPaddingAttrs(
-    DenseIntElementsAttr edge_padding_low,
-    DenseIntElementsAttr edge_padding_high, DenseIntElementsAttr padding_attr,
-    ConvDimensionNumbers dimension_numbers, Builder* builder) {
-  SmallVector padding_low, padding_high;
-  for (const auto& dim : dimension_numbers.input_spatial_dimensions()) {
-    unsigned i = dim.getZExtValue();
-    padding_low.push_back(edge_padding_low.getValue(i));
-    padding_high.push_back(edge_padding_high.getValue(i));
-  }
-
-  int rank = padding_low.size();
-  SmallVector padding;
-  for (unsigned i = 0, e = rank; i < e; ++i) {
-    padding.push_back(GetPaddingValue(padding_attr, {i, 0}) + padding_low[i]);
-    padding.push_back(GetPaddingValue(padding_attr, {i, 1}) + padding_high[i]);
-  }
-  // padding_attr.getType() doesn't work because it is an optional attribute,
-  // which can be a nullptr.
-  auto type = RankedTensorType::get({rank, 2}, builder->getIntegerType(64));
-  return DenseIntElementsAttr::get(type, padding);
-}
-
 #include "tensorflow/compiler/mlir/xla/transforms/generated_canonicalize.inc"
 }  // namespace
 
@@ -2153,14 +2106,5 @@ LogicalResult deriveShapeFromFirstOperand(
   return success();
 }
 
-//===----------------------------------------------------------------------===//
-// ConvOp
-//===----------------------------------------------------------------------===//
-
-void ConvOp::getCanonicalizationPatterns(OwningRewritePatternList& results,
-                                         MLIRContext* context) {
-  results.insert(context);
-}
-
 }  // namespace xla_hlo
 }  // namespace mlir
diff --git a/tensorflow/compiler/mlir/xla/ir/hlo_ops.td b/tensorflow/compiler/mlir/xla/ir/hlo_ops.td
index b1745c73fbf..f92d1c5b85c 100644
--- a/tensorflow/compiler/mlir/xla/ir/hlo_ops.td
+++ b/tensorflow/compiler/mlir/xla/ir/hlo_ops.td
@@ -929,8 +929,6 @@ def HLO_ConvOp : HLO_Op<"convolution", [NoSideEffect]>, BASE_HLO_ConvOp {
   );
 
   let results = (outs HLO_Tensor);
-
-  let hasCanonicalizer = 1;
 }
 
 def HLO_CopyOp: HLO_Op<"copy", [NoSideEffect, SameOperandsAndResultType]>, BASE_HLO_CopyOp {
diff --git a/tensorflow/compiler/mlir/xla/tests/canonicalize.mlir b/tensorflow/compiler/mlir/xla/tests/canonicalize.mlir
index ef0f8c4d200..1954c3344df 100644
--- a/tensorflow/compiler/mlir/xla/tests/canonicalize.mlir
+++ b/tensorflow/compiler/mlir/xla/tests/canonicalize.mlir
@@ -415,71 +415,6 @@ func @fold_copy(%arg : tensor<1x4xf32>) -> tensor<1x4xf32> {
   return %0 : tensor<1x4xf32>
 }
 
-// CHECK-LABEL: func @fold_pad_into_conv_f32
-func @fold_pad_into_conv_f32(%arg0 : tensor<1x32x32x3xf32>,
-                         %arg1 : tensor<7x7x3x64xf32>)
-    -> tensor<1x16x16x64xf32> {
-  //  CHECK-NOT: xla_hlo.pad
-  //      CHECK: xla_hlo.convolution
-  // CHECK-SAME: padding = dense<3> : tensor<2x2xi64>
-  %0 = xla_hlo.constant dense<0.000000e+00> : tensor
-  %1 = "xla_hlo.pad"(%arg0, %0) {
-    edge_padding_high = dense<[0, 3, 3, 0]> : tensor<4xi64>,
-    edge_padding_low = dense<[0, 3, 3, 0]> : tensor<4xi64>,
-    interior_padding = dense<0> : tensor<4xi64>
-  } : (tensor<1x32x32x3xf32>, tensor) -> tensor<1x38x38x3xf32>
-  %2 = "xla_hlo.convolution"(%1, %arg1) {
-    batch_group_count = 1 : i64,
-    dimension_numbers = {
-      input_batch_dimension = 0 : i64,
-      input_feature_dimension = 3 : i64,
-      input_spatial_dimensions = dense<[1, 2]> : tensor<2xi64>,
-      kernel_input_feature_dimension = 2 : i64,
-      kernel_output_feature_dimension = 3 : i64,
-      kernel_spatial_dimensions = dense<[0, 1]> : tensor<2xi64>,
-      output_batch_dimension = 0 : i64,
-      output_feature_dimension = 3 : i64,
-      output_spatial_dimensions = dense<[1, 2]> : tensor<2xi64>
-    },
-    feature_group_count = 1 : i64,
-    padding = dense<0> : tensor<2x2xi64>,
-    window_strides = dense<2> : tensor<2xi64>
-  } : (tensor<1x38x38x3xf32>, tensor<7x7x3x64xf32>) -> tensor<1x16x16x64xf32>
-  return %2 : tensor<1x16x16x64xf32>
-}
-
-// CHECK-LABEL: func @fold_pad_into_conv_i32
-func @fold_pad_into_conv_i32(%arg0 : tensor<1x32x32x3xi32>,
-                         %arg1 : tensor<7x7x3x64xi32>)
-    -> tensor<1x16x16x64xi32> {
-  //  CHECK-NOT: xla_hlo.pad
-  //      CHECK: xla_hlo.convolution
-  // CHECK-SAME: padding = dense<3> : tensor<2x2xi64>
-  %0 = xla_hlo.constant dense<0> : tensor
-  %1 = "xla_hlo.pad"(%arg0, %0) {
-    edge_padding_high = dense<[0, 3, 3, 0]> : tensor<4xi64>,
-    edge_padding_low = dense<[0, 3, 3, 0]> : tensor<4xi64>,
-    interior_padding = dense<0> : tensor<4xi64>
-  } : (tensor<1x32x32x3xi32>, tensor) -> tensor<1x38x38x3xi32>
-  %2 = "xla_hlo.convolution"(%1, %arg1) {
-    batch_group_count = 1 : i64,
-    dimension_numbers = {
-      input_batch_dimension = 0 : i64,
-      input_feature_dimension = 3 : i64,
-      input_spatial_dimensions = dense<[1, 2]> : tensor<2xi64>,
-      kernel_input_feature_dimension = 2 : i64,
-      kernel_output_feature_dimension = 3 : i64,
-      kernel_spatial_dimensions = dense<[0, 1]> : tensor<2xi64>,
-      output_batch_dimension = 0 : i64,
-      output_feature_dimension = 3 : i64,
-      output_spatial_dimensions = dense<[1, 2]> : tensor<2xi64>
-    },
-    feature_group_count = 1 : i64,
-    window_strides = dense<2> : tensor<2xi64>
-  } : (tensor<1x38x38x3xi32>, tensor<7x7x3x64xi32>) -> tensor<1x16x16x64xi32>
-  return %2 : tensor<1x16x16x64xi32>
-}
-
 // CHECK-LABEL: func @dynamic_reshape_not_actually_dynamic
 func @dynamic_reshape_not_actually_dynamic(%arg0: tensor<4xf32>, %shape: tensor<2xindex>) -> tensor<4x1xf32> {
   // CHECK: xla_hlo.reshape
diff --git a/tensorflow/compiler/mlir/xla/transforms/canonicalize.td b/tensorflow/compiler/mlir/xla/transforms/canonicalize.td
index b788cb80380..c319551d92a 100644
--- a/tensorflow/compiler/mlir/xla/transforms/canonicalize.td
+++ b/tensorflow/compiler/mlir/xla/transforms/canonicalize.td
@@ -28,54 +28,3 @@ def UnaryEinsumToEinsum : Pat<
   (HLO_UnaryEinsumOp $operand, $equation),
   (HLO_EinsumOp (HLO_ConstOp (GetScalarOfType<1> $operand)),
                 $operand, (UnaryToBinaryEinsumEq $equation))>;
-
-//===----------------------------------------------------------------------===//
-// Conv op patterns.
-//===----------------------------------------------------------------------===//
-
-def IsZero : Attr() &&"
-  "$_self.cast().isSplat() &&"
-  "$_self.cast().getSplatValue()"
-  ".getValue().isZero()) ||"
-  "($_self.isa() &&"
-  "$_self.cast().isSplat() &&"
-  "$_self.cast().getSplatValue()"
-  ".getInt() == 0)">>;
-
-def IsOnlyPaddingSpatialDims
-  : Constraint>;
-
-def BuildConvPaddingAttrs : NativeCodeCall<
-  "BuildConvPaddingAttrs($0, $1, $2, $3, &$_builder)">;
-
-def FoldPadIntoConv : Pat<
-  (HLO_ConvOp
-    (HLO_PadOp $lhs,
-      (HLO_ConstOp IsZero:$padding_value),
-      $edge_padding_low,
-      $edge_padding_high,
-      IsZero:$interior_padding),
-    $rhs,
-    $window_strides,
-    $padding,
-    $lhs_dilation,
-    $rhs_dilation,
-    $dimension_numbers,
-    $feature_group_count,
-    $batch_group_count,
-    $precision_config),
-  (HLO_ConvOp
-    $lhs,
-    $rhs,
-    $window_strides,
-    (BuildConvPaddingAttrs $edge_padding_low, $edge_padding_high, $padding,
-      $dimension_numbers),
-    $lhs_dilation,
-    $rhs_dilation,
-    $dimension_numbers,
-    $feature_group_count,
-    $batch_group_count,
-    $precision_config),
-    [(IsOnlyPaddingSpatialDims $lhs, $dimension_numbers, $edge_padding_low,
-      $edge_padding_high)]>;

From 9e7d5ef6f25e436fffae03597838294d872404f0 Mon Sep 17 00:00:00 2001
From: "T.J. Alumbaugh" 
Date: Fri, 19 Jun 2020 07:29:02 -0700
Subject: [PATCH 294/867] Full int8 quantization BatchMatMul

PiperOrigin-RevId: 317304259
Change-Id: Icf96d9d129db30b965e36f5c8befd27762b173b2
---
 tensorflow/compiler/mlir/lite/ir/tfl_ops.td   |   6 +-
 tensorflow/lite/kernels/batch_matmul.cc       |  92 ++++++++++-
 tensorflow/lite/kernels/batch_matmul_test.cc  | 156 ++++++++++++++++--
 .../kernels/internal/optimized/batch_matmul.h | 106 ++++++++++++
 .../kernels/internal/reference/batch_matmul.h |  93 +++++++++++
 tensorflow/lite/kernels/register.cc           |   4 +-
 .../lite/tools/optimize/operator_property.cc  |   6 +
 .../lite/tools/versioning/op_version.cc       |   2 +
 .../lite/tools/versioning/runtime_version.cc  |   1 +
 9 files changed, 439 insertions(+), 27 deletions(-)

diff --git a/tensorflow/compiler/mlir/lite/ir/tfl_ops.td b/tensorflow/compiler/mlir/lite/ir/tfl_ops.td
index 509c13ae161..33281cc58fb 100644
--- a/tensorflow/compiler/mlir/lite/ir/tfl_ops.td
+++ b/tensorflow/compiler/mlir/lite/ir/tfl_ops.td
@@ -953,14 +953,14 @@ in the batch dimensions and broadcasting.
   }];
 
   let arguments = (ins
-    TFL_TensorOf<[F32]>:$x,
-    TFL_TensorOf<[F32]>:$y,
+    TFL_TensorOf<[F32, QI8]>:$x,
+    TFL_TensorOf<[F32, QI8]>:$y,
     DefaultValuedAttr:$adj_x,
     DefaultValuedAttr:$adj_y
   );
 
    let results = (outs
-    TFL_TensorOf<[F32]>:$output
+    TFL_TensorOf<[F32, QI8]>:$output
   );
 
   let hasOptions = 1;
diff --git a/tensorflow/lite/kernels/batch_matmul.cc b/tensorflow/lite/kernels/batch_matmul.cc
index 8bc23c9c94a..a414a226504 100644
--- a/tensorflow/lite/kernels/batch_matmul.cc
+++ b/tensorflow/lite/kernels/batch_matmul.cc
@@ -19,6 +19,7 @@ limitations under the License.
 
 #include 
 #include 
+#include 
 
 #include "tensorflow/lite/c/builtin_op_data.h"
 #include "tensorflow/lite/c/common.h"
@@ -52,6 +53,14 @@ enum KernelType {
 };
 
 struct OpData {
+  // The scaling factor from input to output (aka the 'real multiplier') can
+  // be represented as a fixed point multiplier plus a left shift.
+  int32_t output_multiplier;
+  int output_shift;
+  // The range of the fused activation layer. For example for kNone and
+  // uint8_t these would be 0 and 255.
+  int32_t output_activation_min;
+  int32_t output_activation_max;
   // The index of the temporary tensors where we store transposed LHS/RHS.
   int scratch_tensor_index;
   bool rhs_transposed;
@@ -274,6 +283,7 @@ TfLiteStatus Prepare(TfLiteContext* context, TfLiteNode* node) {
 
   OpContext op_context(context, node);
   TF_LITE_ENSURE_OK(context, InitializeTemporaries(context, node, &op_context));
+  OpData* op_data = reinterpret_cast(node->user_data);
 
   bool adj_x = op_context.params->adj_x;
   bool adj_y = op_context.params->adj_y;
@@ -282,7 +292,24 @@ TfLiteStatus Prepare(TfLiteContext* context, TfLiteNode* node) {
   const TfLiteTensor* rhs_data = GetInput(context, node, kInputRHSTensor);
   TfLiteTensor* output = GetOutput(context, node, kOutputTensor);
 
-  TF_LITE_ENSURE_TYPES_EQ(context, lhs_data->type, kTfLiteFloat32);
+  // Note that quantized inference requires that all tensors have their
+  // parameters set. This is usually done during quantized training.
+  if (lhs_data->type == kTfLiteInt8) {
+    double real_multiplier = 0.0;
+    TF_LITE_ENSURE_STATUS(GetQuantizedConvolutionMultipler(
+        context, lhs_data, rhs_data, output, &real_multiplier));
+    int exponent;
+    QuantizeMultiplier(real_multiplier, &op_data->output_multiplier, &exponent);
+    op_data->output_shift = exponent;
+    // BatchMatMul has no fused activation functions. Therefore, set
+    // output activation min and max to min and max of int8_t type,
+    // respecitvely.
+    op_data->output_activation_min = std::numeric_limits::min();
+    op_data->output_activation_max = std::numeric_limits::max();
+  }
+
+  TF_LITE_ENSURE(context, lhs_data->type == kTfLiteFloat32 ||
+                              lhs_data->type == kTfLiteInt8);
   TF_LITE_ENSURE(context, rhs_data->type == kTfLiteFloat32 ||
                               rhs_data->type == kTfLiteInt8);
   // Support dimensions between 2 and 4, inclusive.
@@ -433,6 +460,41 @@ TfLiteStatus EvalHybrid(TfLiteContext* context, TfLiteNode* node, OpData* data,
   return kTfLiteOk;
 }
 
+template 
+TfLiteStatus EvalInt8(TfLiteContext* context, const OpData* data,
+                      const RuntimeShape& lhs_shape, const TfLiteTensor* lhs,
+                      const RuntimeShape& rhs_shape, const TfLiteTensor* rhs,
+                      const RuntimeShape& output_shape, TfLiteTensor* output) {
+  // Reuse params struct from FullyConnected Op.
+  FullyConnectedParams op_params;
+  int32_t input_offset = -lhs->params.zero_point;
+  int32_t filter_offset = -rhs->params.zero_point;
+  int32_t output_offset = output->params.zero_point;
+  op_params.input_offset = input_offset;
+  op_params.weights_offset = filter_offset;
+  op_params.output_offset = output_offset;
+  op_params.output_multiplier = data->output_multiplier;
+  op_params.output_shift = data->output_shift;
+  op_params.quantized_activation_min = data->output_activation_min;
+  op_params.quantized_activation_max = data->output_activation_max;
+  op_params.lhs_cacheable = IsConstantTensor(lhs);
+  op_params.rhs_cacheable = IsConstantTensor(rhs);
+
+  if (kernel_type == kReference) {
+    reference_ops::BatchMatMul(op_params, rhs_shape, GetTensorData(rhs),
+                               lhs_shape, GetTensorData(lhs),
+                               GetTensorShape(output),
+                               GetTensorData(output));
+  } else {
+    optimized_ops::BatchMatMul(op_params, rhs_shape, GetTensorData(rhs),
+                               lhs_shape, GetTensorData(lhs),
+                               GetTensorShape(output),
+                               GetTensorData(output),
+                               CpuBackendContext::GetFromContext(context));
+  }
+  return kTfLiteOk;
+}
+
 template 
 TfLiteStatus EvalQuantized(TfLiteContext* context, TfLiteNode* node,
                            OpData* data, const RuntimeShape& lhs_shape,
@@ -448,25 +510,39 @@ TfLiteStatus EvalQuantized(TfLiteContext* context, TfLiteNode* node,
     return EvalHybrid(
         context, node, data, lhs_shape, lhs, rhs_shape, rhs, input_quantized,
         scaling_factors, accum_scratch, row_sums, input_offsets, output);
+  } else if (lhs->type == kTfLiteInt8) {
+    return EvalInt8(context, data, lhs_shape, lhs, rhs_shape, rhs,
+                                 GetTensorShape(output), output);
   } else {
-    TF_LITE_KERNEL_LOG(context,
-                       "Currently only hybrid quantization is supported.\n");
+    TF_LITE_KERNEL_LOG(
+        context, "Currently only hybrid and int8 quantization is supported.\n");
     return kTfLiteError;
   }
   return kTfLiteOk;
 }
 
-TfLiteTensor* GetRhs(TfLiteContext* context, TfLiteNode* node,
-                     const TfLiteTensor* rhs) {
+TfLiteTensor* GetTempRhs(TfLiteContext* context, TfLiteNode* node,
+                         const TfLiteTensor* rhs) {
   TfLiteTensor* transposed_rhs = GetTemporary(context, node, 1);
   if (rhs->type == kTfLiteInt8) {
-    // Get the quantization params from the weights tensors.
+    // Get the quantization params from the RHS tensor.
     transposed_rhs->params.scale = rhs->params.scale;
     transposed_rhs->params.zero_point = rhs->params.zero_point;
   }
   return transposed_rhs;
 }
 
+TfLiteTensor* GetTempLhs(TfLiteContext* context, TfLiteNode* node,
+                         const TfLiteTensor* lhs) {
+  TfLiteTensor* transposed_lhs = GetTemporary(context, node, 0);
+  if (lhs->type == kTfLiteInt8) {
+    // Get the quantization params from the LHS tensor.
+    transposed_lhs->params.scale = lhs->params.scale;
+    transposed_lhs->params.zero_point = lhs->params.zero_point;
+  }
+  return transposed_lhs;
+}
+
 // Perform a batch matrix multiply on
 // LHS <..., A, B>  X  RHS<..., B, C>
 // where the leading dimensions of LHS and RHS obey broadcasting rules
@@ -491,8 +567,8 @@ TfLiteStatus Eval(TfLiteContext* context, TfLiteNode* node) {
   bool adj_y = op_context.params->adj_y;
   bool adj_x = op_context.params->adj_x;
 
-  const TfLiteTensor* rhs_tensor = adj_y ? rhs : GetRhs(context, node, rhs);
-  const TfLiteTensor* lhs_tensor = adj_x ? GetTemporary(context, node, 0) : lhs;
+  const TfLiteTensor* rhs_tensor = adj_y ? rhs : GetTempRhs(context, node, rhs);
+  const TfLiteTensor* lhs_tensor = adj_x ? GetTempLhs(context, node, lhs) : lhs;
   if (!adj_y) {
     // TODO(b/154760341) Constant tensors should already be transposed, but
     // we transpose once if necessary for now.
diff --git a/tensorflow/lite/kernels/batch_matmul_test.cc b/tensorflow/lite/kernels/batch_matmul_test.cc
index 5e52479f49b..98df8ebe3db 100644
--- a/tensorflow/lite/kernels/batch_matmul_test.cc
+++ b/tensorflow/lite/kernels/batch_matmul_test.cc
@@ -24,8 +24,19 @@ limitations under the License.
 #include "tensorflow/lite/schema/schema_generated.h"
 
 namespace tflite {
+
+namespace ops {
+namespace builtin {
+
+TfLiteRegistration* Register_BATCH_MATMUL_REF();
+TfLiteRegistration* Register_BATCH_MATMUL_GENERIC_OPTIMIZED();
+
+}  // namespace builtin
+}  // namespace ops
+
 namespace {
 
+using ::testing::ElementsAre;
 using ::testing::ElementsAreArray;
 
 template 
@@ -53,7 +64,20 @@ class BatchMatMulOpModel : public SingleOpModel {
   int output_id_;
 };
 
-TEST(BatchMatMulOpModelTest, Float32Test_Simple) {
+const auto kKernelMap = new std::map({
+    {"Reference", ops::builtin::Register_BATCH_MATMUL_REF()},
+    {"GenericOptimized",
+     ops::builtin::Register_BATCH_MATMUL_GENERIC_OPTIMIZED()},
+});
+
+class BatchMatMulOpTest : public SingleOpTest {
+ protected:
+  const std::map& GetKernelMap() override {
+    return *kKernelMap;
+  }
+};
+
+TEST_P(BatchMatMulOpTest, Float32Test_Simple) {
   BatchMatMulOpModel model({TensorType_FLOAT32, {1, 2, 3}},
                                   {TensorType_FLOAT32, {1, 3, 4}});
   model.PopulateTensor(model.lhs(), {1, 2, 3, 4, 5, 6});
@@ -65,7 +89,7 @@ TEST(BatchMatMulOpModelTest, Float32Test_Simple) {
   EXPECT_THAT(model.GetOutputShape(), ElementsAreArray({1, 2, 4}));
 }
 
-TEST(BatchMatMulOpModelTest, Float32Test_SimpleRHSAdjoint) {
+TEST_P(BatchMatMulOpTest, Float32Test_SimpleRHSAdjoint) {
   BatchMatMulOpModel model({TensorType_FLOAT32, {1, 2, 3}},
                                   {TensorType_FLOAT32, {1, 4, 3}}, false, true);
   model.PopulateTensor(model.lhs(), {1, 2, 3, 4, 5, 6});
@@ -77,7 +101,7 @@ TEST(BatchMatMulOpModelTest, Float32Test_SimpleRHSAdjoint) {
   EXPECT_THAT(model.GetOutputShape(), ElementsAreArray({1, 2, 4}));
 }
 
-TEST(BatchMatMulOpModelTest, Float32Test_SimpleLHSAdjoint) {
+TEST_P(BatchMatMulOpTest, Float32Test_SimpleLHSAdjoint) {
   BatchMatMulOpModel model({TensorType_FLOAT32, {1, 3, 2}},
                                   {TensorType_FLOAT32, {1, 3, 4}}, true, false);
   model.PopulateTensor(model.lhs(), {1, 4, 2, 5, 3, 6});
@@ -89,7 +113,7 @@ TEST(BatchMatMulOpModelTest, Float32Test_SimpleLHSAdjoint) {
   EXPECT_THAT(model.GetOutputShape(), ElementsAreArray({1, 2, 4}));
 }
 
-TEST(BatchMatMulOpModelTest, Float32Test_BatchSizeTwo) {
+TEST_P(BatchMatMulOpTest, Float32Test_BatchSizeTwo) {
   BatchMatMulOpModel model({TensorType_FLOAT32, {2, 2, 3}},
                                   {TensorType_FLOAT32, {2, 3, 4}});
   model.PopulateTensor(model.lhs(),
@@ -105,7 +129,7 @@ TEST(BatchMatMulOpModelTest, Float32Test_BatchSizeTwo) {
   EXPECT_THAT(model.GetOutputShape(), ElementsAreArray({2, 2, 4}));
 }
 
-TEST(BatchMatMulOpModelTest, Float32Test_Broadcast) {
+TEST_P(BatchMatMulOpTest, Float32Test_Broadcast) {
   BatchMatMulOpModel model({TensorType_FLOAT32, {2, 2, 3}},
                                   {TensorType_FLOAT32, {3, 4}});
   model.PopulateTensor(model.lhs(),
@@ -121,7 +145,7 @@ TEST(BatchMatMulOpModelTest, Float32Test_Broadcast) {
   EXPECT_THAT(model.GetOutputShape(), ElementsAreArray({2, 2, 4}));
 }
 
-TEST(BatchMatMulOpModelTest, Float32Test_BroadcastLHSAdjoint) {
+TEST_P(BatchMatMulOpTest, Float32Test_BroadcastLHSAdjoint) {
   BatchMatMulOpModel model({TensorType_FLOAT32, {2, 3, 2}},
                                   {TensorType_FLOAT32, {3, 4}}, true, false);
   model.PopulateTensor(model.lhs(),
@@ -137,7 +161,7 @@ TEST(BatchMatMulOpModelTest, Float32Test_BroadcastLHSAdjoint) {
   EXPECT_THAT(model.GetOutputShape(), ElementsAreArray({2, 2, 4}));
 }
 
-TEST(BatchMatMulOpModelTest, Float32Test_Broadcast2) {
+TEST_P(BatchMatMulOpTest, Float32Test_Broadcast2) {
   BatchMatMulOpModel model({TensorType_FLOAT32, {2, 1, 3, 2}},
                                   {TensorType_FLOAT32, {3, 2, 4}});
   model.PopulateTensor(model.lhs(),
@@ -161,7 +185,7 @@ TEST(BatchMatMulOpModelTest, Float32Test_Broadcast2) {
   EXPECT_THAT(model.GetOutputShape(), ElementsAreArray({2, 3, 3, 4}));
 }
 
-TEST(BatchMatMulOpModelTest, Float32Test_Broadcast2LHSAdjoint) {
+TEST_P(BatchMatMulOpTest, Float32Test_Broadcast2LHSAdjoint) {
   BatchMatMulOpModel model({TensorType_FLOAT32, {2, 1, 2, 3}},
                                   {TensorType_FLOAT32, {3, 2, 4}}, true, false);
   model.PopulateTensor(model.lhs(),
@@ -185,7 +209,7 @@ TEST(BatchMatMulOpModelTest, Float32Test_Broadcast2LHSAdjoint) {
   EXPECT_THAT(model.GetOutputShape(), ElementsAreArray({2, 3, 3, 4}));
 }
 
-TEST(BatchMatMulOpModelTest, Float32Test_Broadcast2RHSAdjoint) {
+TEST_P(BatchMatMulOpTest, Float32Test_Broadcast2RHSAdjoint) {
   BatchMatMulOpModel model({TensorType_FLOAT32, {2, 1, 3, 2}},
                                   {TensorType_FLOAT32, {3, 4, 2}}, false, true);
   model.PopulateTensor(model.lhs(),
@@ -208,7 +232,7 @@ TEST(BatchMatMulOpModelTest, Float32Test_Broadcast2RHSAdjoint) {
   EXPECT_THAT(model.GetOutputShape(), ElementsAreArray({2, 3, 3, 4}));
 }
 
-TEST(BatchMatMulOpModelTest, Float32Test_Broadcast2BothAdjoint) {
+TEST_P(BatchMatMulOpTest, Float32Test_Broadcast2BothAdjoint) {
   BatchMatMulOpModel model({TensorType_FLOAT32, {2, 1, 2, 3}},
                                   {TensorType_FLOAT32, {3, 4, 2}}, true, true);
   model.PopulateTensor(model.lhs(),
@@ -231,7 +255,7 @@ TEST(BatchMatMulOpModelTest, Float32Test_Broadcast2BothAdjoint) {
   EXPECT_THAT(model.GetOutputShape(), ElementsAreArray({2, 3, 3, 4}));
 }
 
-TEST(BatchMatMulOpModelTest, Float32Test_BroadcastFromRHS) {
+TEST_P(BatchMatMulOpTest, Float32Test_BroadcastFromRHS) {
   BatchMatMulOpModel model({TensorType_FLOAT32, {4, 5}},
                                   {TensorType_FLOAT32, {3, 1, 5, 2}});
   model.PopulateTensor(
@@ -251,6 +275,10 @@ TEST(BatchMatMulOpModelTest, Float32Test_BroadcastFromRHS) {
   EXPECT_THAT(model.GetOutputShape(), ElementsAreArray({3, 1, 4, 2}));
 }
 
+INSTANTIATE_TEST_SUITE_P(
+    BatchMatMulOpTest, BatchMatMulOpTest,
+    ::testing::ValuesIn(SingleOpTest::GetKernelTags(*kKernelMap)));
+
 // In the hybrid model the weights are quantized int8. But the input
 // and output are expected to be in float precision.
 class HybridAsymmetricBatchMatMulOpModel : public SingleOpModel {
@@ -304,7 +332,14 @@ class HybridAsymmetricBatchMatMulOpModel : public SingleOpModel {
   int input_size_;
 };
 
-TEST(HybridAsymmetricBatchMatMulOpTest, SimpleTestQuantizedInt8) {
+class HybridAsymmetricBatchMatMulOpTest : public SingleOpTest {
+ protected:
+  const std::map& GetKernelMap() override {
+    return *kKernelMap;
+  }
+};
+
+TEST_P(HybridAsymmetricBatchMatMulOpTest, SimpleTestQuantizedInt8) {
   HybridAsymmetricBatchMatMulOpModel m(
       /*units=*/3, /*batches=*/2,
       /*lhs=*/{TensorType_FLOAT32, {2, 10}},
@@ -335,7 +370,7 @@ TEST(HybridAsymmetricBatchMatMulOpTest, SimpleTestQuantizedInt8) {
   EXPECT_THAT(m.GetOutputShape(), ElementsAreArray({2, 3}));
 }
 
-TEST(HybridAsymmetricBatchMatMulOpTest, QuantizedInt8BroadcastWeights) {
+TEST_P(HybridAsymmetricBatchMatMulOpTest, QuantizedInt8BroadcastWeights) {
   HybridAsymmetricBatchMatMulOpModel m(
       /*units=*/3, /*batches=*/2,
       /*lhs=*/{TensorType_FLOAT32, {2, 2, 10}},
@@ -366,7 +401,7 @@ TEST(HybridAsymmetricBatchMatMulOpTest, QuantizedInt8BroadcastWeights) {
   EXPECT_THAT(m.GetOutputShape(), ElementsAreArray({2, 2, 3}));
 }
 
-TEST(HybridAsymmetricBatchMatMulOpTest, QuantizedInt8BroadcastBigWeights) {
+TEST_P(HybridAsymmetricBatchMatMulOpTest, QuantizedInt8BroadcastBigWeights) {
   HybridAsymmetricBatchMatMulOpModel m(
       /*units=*/9, /*batches=*/2,
       /*lhs=*/{TensorType_FLOAT32, {2, 2, 10}},
@@ -401,7 +436,7 @@ TEST(HybridAsymmetricBatchMatMulOpTest, QuantizedInt8BroadcastBigWeights) {
   EXPECT_THAT(m.GetOutputShape(), ElementsAreArray({2, 2, 9}));
 }
 
-TEST(HybridAsymmetricBatchMatMulOpTest, QuantizedInt8BroadcastInputs) {
+TEST_P(HybridAsymmetricBatchMatMulOpTest, QuantizedInt8BroadcastInputs) {
   HybridAsymmetricBatchMatMulOpModel m(
       /*units=*/3, /*batches=*/2,
       /*lhs=*/{TensorType_FLOAT32, {2, 10}},
@@ -431,5 +466,96 @@ TEST(HybridAsymmetricBatchMatMulOpTest, QuantizedInt8BroadcastInputs) {
   EXPECT_THAT(m.GetOutputShape(), ElementsAreArray({2, 2, 3}));
 }
 
+INSTANTIATE_TEST_SUITE_P(
+    HybridAsymmetricBatchMatMulOpTest, HybridAsymmetricBatchMatMulOpTest,
+    ::testing::ValuesIn(SingleOpTest::GetKernelTags(*kKernelMap)));
+
+class QuantizedBatchMatMulOpModel : public SingleOpModel {
+ public:
+  QuantizedBatchMatMulOpModel(int units, int batches, const TensorData& lhs,
+                              const TensorData& output = {TensorType_INT8},
+                              bool adj_x = false, bool adj_y = false)
+      : units_(units), batches_(batches) {
+    int total_input_size = 1;
+    for (size_t i = 0; i < lhs.shape.size(); ++i) {
+      total_input_size *= lhs.shape[i];
+    }
+    input_size_ = total_input_size / batches_;
+
+    lhs_id_ = AddInput(lhs);
+    rhs_id_ = AddInput({lhs.type, {input_size_, units_}, lhs.min, lhs.max});
+
+    output_id_ = AddOutput(output);
+
+    SetBuiltinOp(BuiltinOperator_BATCH_MATMUL,
+                 BuiltinOptions_BatchMatMulOptions,
+                 CreateBatchMatMulOptions(builder_, adj_x, adj_y).Union());
+    BuildInterpreter({GetShape(lhs_id_), GetShape(rhs_id_)});
+  }
+
+  template 
+  void SetWeights(const std::vector& data) {
+    QuantizeAndPopulate(rhs_id_, data);
+  }
+
+  template 
+  void SetInput(const std::vector& data) {
+    QuantizeAndPopulate(lhs_id_, data);
+  }
+
+  template 
+  std::vector GetOutput() {
+    return ExtractVector(output_id_);
+  }
+
+  template 
+  std::vector GetDequantizedOutput() {
+    return Dequantize(ExtractVector(output_id_), GetScale(output_id_),
+                         GetZeroPoint(output_id_));
+  }
+
+ protected:
+  int lhs_id_;
+  int rhs_id_;
+  int output_id_;
+  int units_;
+  int batches_;
+  int input_size_;
+};
+
+class QuantizedBatchMatMulOpTest : public SingleOpTest {
+ protected:
+  const std::map& GetKernelMap() override {
+    return *kKernelMap;
+  }
+};
+
+TEST_P(QuantizedBatchMatMulOpTest, SimpleTestQuantizedInt8) {
+  QuantizedBatchMatMulOpModel m(
+      /*units=*/3, /*batches*/ 2,
+      /*lhs=*/{TensorType_INT8, {2, 10}, -63.5, 64},
+      /*output=*/{TensorType_INT8, {}, -127, 128});
+
+  m.SetWeights({
+      1, 1, 1, 2, 2, 2, 3, 3, 3, 4, 4, 4, 5,  5,  5,
+      6, 6, 6, 7, 7, 7, 8, 8, 8, 9, 9, 9, 10, 10, 10,
+  });
+
+  m.SetInput({
+      1, 2, 3, 4, 5, 6, 7, 8,  -9, -10,  // b = 0
+      1, 2, 3, 4, 5, 6, 7, -8, 9,  -10,  // b = 1
+  });
+
+  m.Invoke();
+
+  EXPECT_THAT(m.GetDequantizedOutput(),
+              ElementsAreArray(ArrayFloatNear({23, 23, 23, 57, 57, 57})));
+  EXPECT_THAT(m.GetOutput(), ElementsAre(22, 22, 22, 56, 56, 56));
+}
+
+INSTANTIATE_TEST_SUITE_P(
+    QuantizedBatchMatMulOpTest, QuantizedBatchMatMulOpTest,
+    ::testing::ValuesIn(SingleOpTest::GetKernelTags(*kKernelMap)));
+
 }  // namespace
 }  // namespace tflite
diff --git a/tensorflow/lite/kernels/internal/optimized/batch_matmul.h b/tensorflow/lite/kernels/internal/optimized/batch_matmul.h
index 24b5012304f..5e622154d60 100644
--- a/tensorflow/lite/kernels/internal/optimized/batch_matmul.h
+++ b/tensorflow/lite/kernels/internal/optimized/batch_matmul.h
@@ -272,6 +272,112 @@ inline void BatchMatMul(const RuntimeShape& lhs_shape, const int8_t* lhs_data,
   }
 }
 
+inline void BatchMatMul(const FullyConnectedParams& params,
+                        const RuntimeShape& lhs_shape, const int8_t* lhs_data,
+                        const RuntimeShape& rhs_shape, const int8_t* rhs_data,
+                        const RuntimeShape& output_shape, int8_t* output_data,
+                        CpuBackendContext* context) {
+  using ::tflite::cpu_backend_gemm::Gemm;
+  using ::tflite::cpu_backend_gemm::GemmParams;
+  using ::tflite::cpu_backend_gemm::MatrixParams;
+
+  const RuntimeShape extended_lhs_shape =
+      RuntimeShape::ExtendedShape(5, lhs_shape);
+  const RuntimeShape extended_rhs_shape =
+      RuntimeShape::ExtendedShape(5, rhs_shape);
+
+  // Determine which dimension is the broadcast dimension.
+  auto broadcast_dim = [](int lhs_dim, int rhs_dim) {
+    if (lhs_dim == rhs_dim) return lhs_dim;
+    if (lhs_dim == 1) return rhs_dim;
+    TFLITE_DCHECK_EQ(rhs_dim, 1);
+    return lhs_dim;
+  };
+
+  // Compute the "extent" for iterating on this dimension.
+  // If we are broadcasting, then don't advance (i.e return 0).
+  auto extent = [](const RuntimeShape& shape, int x) {
+    if (shape.Dims(x) == 1) {
+      return 0;
+    }
+    int prod = 1;
+    for (int i = x + 1; i < shape.DimensionsCount(); ++i) {
+      prod *= shape.Dims(i);
+    }
+    return prod;
+  };
+
+  const int batch_dim0 =
+      broadcast_dim(extended_lhs_shape.Dims(0), extended_rhs_shape.Dims(0));
+  const int batch_dim1 =
+      broadcast_dim(extended_lhs_shape.Dims(1), extended_rhs_shape.Dims(1));
+  const int batch_dim2 =
+      broadcast_dim(extended_lhs_shape.Dims(2), extended_rhs_shape.Dims(2));
+
+  const int lhs_ext0 = extent(extended_lhs_shape, 0);
+  const int lhs_ext1 = extent(extended_lhs_shape, 1);
+  const int lhs_ext2 = extent(extended_lhs_shape, 2);
+  const int rhs_ext0 = extent(extended_rhs_shape, 0);
+  const int rhs_ext1 = extent(extended_rhs_shape, 1);
+  const int rhs_ext2 = extent(extended_rhs_shape, 2);
+
+  // Set params for each matrix multiply.
+  const int lhs_rows = extended_lhs_shape.Dims(3);
+  const int rhs_cols = extended_rhs_shape.Dims(4);
+  const int accum_depth = extended_lhs_shape.Dims(4);
+
+  const int32 input_offset = params.input_offset;
+  const int32 filter_offset = params.weights_offset;
+  const int32 output_offset = params.output_offset;
+  const int32 output_multiplier = params.output_multiplier;
+  const int output_shift = params.output_shift;
+  const int32 output_activation_min = params.quantized_activation_min;
+  const int32 output_activation_max = params.quantized_activation_max;
+  TFLITE_DCHECK_LE(output_activation_min, output_activation_max);
+
+  MatrixParams lhs_params;
+  lhs_params.order = cpu_backend_gemm::Order::kRowMajor;
+  lhs_params.rows = lhs_rows;
+  lhs_params.cols = accum_depth;
+  lhs_params.zero_point = -filter_offset;
+
+  MatrixParams rhs_params;
+  rhs_params.order = cpu_backend_gemm::Order::kColMajor;
+  rhs_params.rows = accum_depth;
+  rhs_params.cols = rhs_cols;
+  rhs_params.zero_point = -input_offset;
+
+  MatrixParams dst_params;
+  dst_params.order = cpu_backend_gemm::Order::kColMajor;
+  dst_params.rows = lhs_rows;
+  dst_params.cols = rhs_cols;
+  dst_params.zero_point = output_offset;
+
+  for (int b0 = 0; b0 < batch_dim0; ++b0) {
+    const int8_t* lhs_ptr0 = lhs_data + (b0 * lhs_ext0);
+    const int8_t* rhs_ptr0 = rhs_data + (b0 * rhs_ext0);
+    for (int b1 = 0; b1 < batch_dim1; ++b1) {
+      const int8_t* lhs_ptr1 = lhs_ptr0 + b1 * lhs_ext1;
+      const int8_t* rhs_ptr1 = rhs_ptr0 + b1 * rhs_ext1;
+      for (int b2 = 0; b2 < batch_dim2; ++b2) {
+        const int8_t* lhs_ptr2 = lhs_ptr1 + b2 * lhs_ext2;
+        const int8_t* rhs_ptr2 = rhs_ptr1 + b2 * rhs_ext2;
+        int8_t* out_ptr = output_data + ((b0 * batch_dim1 * batch_dim2) +
+                                         b1 * batch_dim2 + b2) *
+                                            lhs_rows * rhs_cols;
+
+        GemmParams gemm_params;
+        gemm_params.clamp_min = output_activation_min;
+        gemm_params.clamp_max = output_activation_max;
+        gemm_params.multiplier_fixedpoint = output_multiplier;
+        gemm_params.multiplier_exponent = output_shift;
+        cpu_backend_gemm::Gemm(lhs_params, lhs_ptr2, rhs_params, rhs_ptr2,
+                               dst_params, out_ptr, gemm_params, context);
+      }
+    }
+  }
+}
+
 }  // namespace optimized_ops
 }  // namespace tflite
 
diff --git a/tensorflow/lite/kernels/internal/reference/batch_matmul.h b/tensorflow/lite/kernels/internal/reference/batch_matmul.h
index 1394bd9da64..05caefaca5d 100644
--- a/tensorflow/lite/kernels/internal/reference/batch_matmul.h
+++ b/tensorflow/lite/kernels/internal/reference/batch_matmul.h
@@ -217,6 +217,99 @@ inline void BatchMatMul(const RuntimeShape& lhs_shape, const int8_t* lhs_data,
   }
 }
 
+inline void BatchMatMul(const FullyConnectedParams& params,
+                        const RuntimeShape& lhs_shape, const int8_t* lhs_data,
+                        const RuntimeShape& rhs_shape, const int8_t* rhs_data,
+                        const RuntimeShape& output_shape, int8_t* output_data) {
+  const RuntimeShape extended_lhs_shape =
+      RuntimeShape::ExtendedShape(5, lhs_shape);
+  const RuntimeShape extended_rhs_shape =
+      RuntimeShape::ExtendedShape(5, rhs_shape);
+
+  // Determine which dimension is the broadcast dimension.
+  auto broadcast_dim = [](int lhs_dim, int rhs_dim) {
+    if (lhs_dim == rhs_dim) return lhs_dim;
+    if (lhs_dim == 1) return rhs_dim;
+    TFLITE_DCHECK_EQ(rhs_dim, 1);
+    return lhs_dim;
+  };
+
+  // Compute the "extent" for iterating on this dimension.
+  // If we are broadcasting, then don't advance (i.e return 0).
+  auto extent = [](const RuntimeShape& shape, int x) {
+    if (shape.Dims(x) == 1) {
+      return 0;
+    }
+    int prod = 1;
+    for (int i = x + 1; i < shape.DimensionsCount(); ++i) {
+      prod *= shape.Dims(i);
+    }
+    return prod;
+  };
+
+  const int batch_dim0 =
+      broadcast_dim(extended_lhs_shape.Dims(0), extended_rhs_shape.Dims(0));
+  const int batch_dim1 =
+      broadcast_dim(extended_lhs_shape.Dims(1), extended_rhs_shape.Dims(1));
+  const int batch_dim2 =
+      broadcast_dim(extended_lhs_shape.Dims(2), extended_rhs_shape.Dims(2));
+
+  const int lhs_ext0 = extent(extended_lhs_shape, 0);
+  const int lhs_ext1 = extent(extended_lhs_shape, 1);
+  const int lhs_ext2 = extent(extended_lhs_shape, 2);
+  const int rhs_ext0 = extent(extended_rhs_shape, 0);
+  const int rhs_ext1 = extent(extended_rhs_shape, 1);
+  const int rhs_ext2 = extent(extended_rhs_shape, 2);
+
+  // Set params for each matrix multiply.
+  const int lhs_rows = extended_lhs_shape.Dims(3);
+  const int rhs_cols = extended_rhs_shape.Dims(4);
+  const int accum_depth = extended_lhs_shape.Dims(4);
+
+  const int32 input_offset = params.input_offset;
+  const int32 filter_offset = params.weights_offset;
+  const int32 output_offset = params.output_offset;
+  const int32 output_multiplier = params.output_multiplier;
+  const int output_shift = params.output_shift;
+  const int32 output_activation_min = params.quantized_activation_min;
+  const int32 output_activation_max = params.quantized_activation_max;
+  TFLITE_DCHECK_LE(output_activation_min, output_activation_max);
+
+  for (int b0 = 0; b0 < batch_dim0; ++b0) {
+    const int8_t* lhs_ptr0 = lhs_data + (b0 * lhs_ext0);
+    const int8_t* rhs_ptr0 = rhs_data + (b0 * rhs_ext0);
+    for (int b1 = 0; b1 < batch_dim1; ++b1) {
+      const int8_t* lhs_ptr1 = lhs_ptr0 + b1 * lhs_ext1;
+      const int8_t* rhs_ptr1 = rhs_ptr0 + b1 * rhs_ext1;
+      for (int b2 = 0; b2 < batch_dim2; ++b2) {
+        const int8_t* lhs_ptr2 = lhs_ptr1 + b2 * lhs_ext2;
+        const int8_t* rhs_ptr2 = rhs_ptr1 + b2 * rhs_ext2;
+        int8_t* out_ptr = output_data + ((b0 * batch_dim1 * batch_dim2) +
+                                         b1 * batch_dim2 + b2) *
+                                            lhs_rows * rhs_cols;
+
+        for (int j = 0; j < rhs_cols; ++j) {
+          for (int i = 0; i < lhs_rows; ++i) {
+            int32_t total = 0;
+            for (int k = 0; k < accum_depth; ++k) {
+              int32 lhs_val = lhs_ptr2[accum_depth * i + k];
+              int32 rhs_val = rhs_ptr2[accum_depth * j + k];
+              total += (lhs_val + filter_offset) * (rhs_val + input_offset);
+            }
+            total = MultiplyByQuantizedMultiplier(total, output_multiplier,
+                                                  output_shift);
+            total += output_offset;
+            total = std::max(total, output_activation_min);
+            total = std::min(total, output_activation_max);
+            const int idx = lhs_rows * j + i;
+            out_ptr[idx] = static_cast(total);
+          }
+        }
+      }
+    }
+  }
+}
+
 }  // namespace reference_ops
 }  // namespace tflite
 
diff --git a/tensorflow/lite/kernels/register.cc b/tensorflow/lite/kernels/register.cc
index 90688a2aa1f..c3a4aaad16d 100644
--- a/tensorflow/lite/kernels/register.cc
+++ b/tensorflow/lite/kernels/register.cc
@@ -289,7 +289,9 @@ BuiltinOpResolver::BuiltinOpResolver() {
   AddBuiltin(BuiltinOperator_SCATTER_ND, Register_SCATTER_ND());
   AddBuiltin(BuiltinOperator_DENSIFY, Register_DENSIFY());
   AddBuiltin(BuiltinOperator_SEGMENT_SUM, Register_SEGMENT_SUM());
-  AddBuiltin(BuiltinOperator_BATCH_MATMUL, Register_BATCH_MATMUL());
+  AddBuiltin(BuiltinOperator_BATCH_MATMUL, Register_BATCH_MATMUL(),
+             /* min_version = */ 1,
+             /* max_version = */ 2);
   AddCustom("NumericVerify", tflite::ops::custom::Register_NUMERIC_VERIFY());
   // TODO(andrewharp, ahentz): Move these somewhere more appropriate so that
   // custom ops aren't always included by default.
diff --git a/tensorflow/lite/tools/optimize/operator_property.cc b/tensorflow/lite/tools/optimize/operator_property.cc
index 8a0cbca29e2..f2cb98ef31a 100644
--- a/tensorflow/lite/tools/optimize/operator_property.cc
+++ b/tensorflow/lite/tools/optimize/operator_property.cc
@@ -88,6 +88,12 @@ OperatorProperty GetOperatorProperty(const ModelT* model, int subgraph_index,
       property.restrict_same_input_output_scale = true;
       property.version = 2;
       break;
+    case BuiltinOperator_BATCH_MATMUL: {
+      property.inputs = {{0, {}}, {1, {}}};
+      property.outputs = {{0, {}}};
+      property.version = 2;
+      break;
+    }
     case BuiltinOperator_BATCH_TO_SPACE_ND:
     case BuiltinOperator_SPACE_TO_BATCH_ND:
     case BuiltinOperator_SPACE_TO_DEPTH:
diff --git a/tensorflow/lite/tools/versioning/op_version.cc b/tensorflow/lite/tools/versioning/op_version.cc
index 118e2d420f8..a97b9da47f1 100644
--- a/tensorflow/lite/tools/versioning/op_version.cc
+++ b/tensorflow/lite/tools/versioning/op_version.cc
@@ -24,6 +24,7 @@ limitations under the License.
 #include "absl/strings/str_split.h"
 #include "tensorflow/core/platform/logging.h"
 #include "tensorflow/lite/kernels/internal/compatibility.h"
+#include "tensorflow/lite/schema/schema_generated.h"
 
 namespace tflite {
 namespace {
@@ -518,6 +519,7 @@ int GetBuiltinOperatorVersion(const OpSignature& op_sig) {
     case BuiltinOperator_LESS:
     case BuiltinOperator_LESS_EQUAL:
     case BuiltinOperator_SELECT:
+    case BuiltinOperator_BATCH_MATMUL:
       if (op_sig.input_types.at(0) == TensorType_INT8) {
         return 2;
       }
diff --git a/tensorflow/lite/tools/versioning/runtime_version.cc b/tensorflow/lite/tools/versioning/runtime_version.cc
index 92a7001606f..36976354685 100644
--- a/tensorflow/lite/tools/versioning/runtime_version.cc
+++ b/tensorflow/lite/tools/versioning/runtime_version.cc
@@ -58,6 +58,7 @@ std::string FindMinimumRuntimeVersionForOp(tflite::BuiltinOperator op_code,
               {{BuiltinOperator_AVERAGE_POOL_2D, 2}, "1.14.0"},
               {{BuiltinOperator_AVERAGE_POOL_2D, 3}, kPendingReleaseVersion},
               {{BuiltinOperator_BATCH_MATMUL, 1}, kPendingReleaseVersion},
+              {{BuiltinOperator_BATCH_MATMUL, 2}, kPendingReleaseVersion},
               {{BuiltinOperator_CONV_2D, 1}, "1.5.0"},
               {{BuiltinOperator_CONV_2D, 2}, "1.14.0"},
               {{BuiltinOperator_CONV_2D, 3}, "1.14.0"},

From 07c54454eec55c1279c243a3c148eeee81b41ed5 Mon Sep 17 00:00:00 2001
From: Tamara Norman 
Date: Fri, 19 Jun 2020 08:48:51 -0700
Subject: [PATCH 295/867] Add an option such that the cached host_value can be
 discarded

PiperOrigin-RevId: 317315157
Change-Id: I9d7145390a526003069321c7e04794e139a53c09
---
 tensorflow/compiler/xla/pjrt/pjrt_client.cc | 6 +++++-
 tensorflow/compiler/xla/pjrt/pjrt_client.h  | 8 ++++++--
 2 files changed, 11 insertions(+), 3 deletions(-)

diff --git a/tensorflow/compiler/xla/pjrt/pjrt_client.cc b/tensorflow/compiler/xla/pjrt/pjrt_client.cc
index 46f592100c9..b4f0363e69a 100644
--- a/tensorflow/compiler/xla/pjrt/pjrt_client.cc
+++ b/tensorflow/compiler/xla/pjrt/pjrt_client.cc
@@ -1077,13 +1077,17 @@ Status PjRtBuffer::CopyToHostAsync() {
   return Status::OK();
 }
 
-StatusOr> PjRtBuffer::ToLiteral() {
+StatusOr> PjRtBuffer::ToLiteral(
+    const bool discard_cached_copy) {
   tensorflow::profiler::TraceMe traceme("PjRtBuffer::ToLiteral");
   TF_RETURN_IF_ERROR(CopyToHostAsync());
   std::shared_ptr host_value;
   {
     absl::MutexLock lock(&mu_);
     host_value = host_value_;
+    if (discard_cached_copy) {
+      host_value_ = nullptr;
+    }
   }
   if (host_value == nullptr) {
     return InvalidArgument("ToLiteral called on invalid buffer");
diff --git a/tensorflow/compiler/xla/pjrt/pjrt_client.h b/tensorflow/compiler/xla/pjrt/pjrt_client.h
index 754eb19bec6..8f74e6244d6 100644
--- a/tensorflow/compiler/xla/pjrt/pjrt_client.h
+++ b/tensorflow/compiler/xla/pjrt/pjrt_client.h
@@ -478,8 +478,12 @@ class PjRtBuffer {
 
   // Returns the buffer's value as an XLA Literal. If the value has previously
   // been prefetched to the host, then returns the prefetched version, otherwise
-  // copies the buffer to the host. Blocks until the value is ready.
-  StatusOr> ToLiteral();
+  // copies the buffer to the host. Blocks until the value is ready. If
+  // `discard_cached_copy` is true then buffer will no longer keep hold of a
+  // cached copy of the literal (i.e. The reference to the host value will be
+  // removed.)
+  StatusOr> ToLiteral(
+      bool discard_cached_copy = false);
 
   // Initiates a copy of the buffer to the host. Does not block waiting for
   // the transfer to complete. The value can be retrieved by a later call to

From 16cb89bd7b40fc816aa7440f62b443ca480bcf6d Mon Sep 17 00:00:00 2001
From: "A. Unique TensorFlower" 
Date: Fri, 19 Jun 2020 09:15:04 -0700
Subject: [PATCH 296/867] Qualify uses of std::string

PiperOrigin-RevId: 317319501
Change-Id: Ib75a31ad89fa1a6bda81450f2ab5ba07d7338ada
---
 tensorflow/lite/toco/tflite/export.cc         | 58 +++++++++----------
 tensorflow/lite/toco/tflite/export.h          | 12 ++--
 tensorflow/lite/toco/tflite/export_test.cc    | 56 ++++++++++--------
 tensorflow/lite/toco/tflite/import.cc         | 18 +++---
 tensorflow/lite/toco/tflite/import.h          |  6 +-
 tensorflow/lite/toco/tflite/import_test.cc    |  6 +-
 tensorflow/lite/toco/tflite/op_version.cc     |  8 +--
 .../lite/toco/tflite/op_version_test.cc       | 32 +++++-----
 tensorflow/lite/toco/tflite/operator.cc       | 37 ++++++------
 tensorflow/lite/toco/tflite/operator.h        | 12 ++--
 tensorflow/lite/toco/tflite/operator_test.cc  |  8 +--
 tensorflow/lite/toco/tflite/types.cc          |  6 +-
 12 files changed, 133 insertions(+), 126 deletions(-)

diff --git a/tensorflow/lite/toco/tflite/export.cc b/tensorflow/lite/toco/tflite/export.cc
index d72a902001d..d109ab875b5 100644
--- a/tensorflow/lite/toco/tflite/export.cc
+++ b/tensorflow/lite/toco/tflite/export.cc
@@ -52,7 +52,7 @@ using ::tflite::Tensor;
 namespace {
 
 // Check if a TensorFlow Op is a control flow op by its name.
-bool IsControlFlowOp(const string& tensorflow_op) {
+bool IsControlFlowOp(const std::string& tensorflow_op) {
   // Technically this is equivalent to `::tensorflow::Node::IsControlFlow()`.
   // It requires to construct a `::tensorflow::Graph` to use that helper
   // function, so we simply hardcode the list of control flow ops here.
@@ -68,7 +68,7 @@ bool IsControlFlowOp(const string& tensorflow_op) {
 }
 
 // Check if a TensorFlow Op is unsupported by the Flex runtime.
-bool IsUnsupportedFlexOp(const string& tensorflow_op) {
+bool IsUnsupportedFlexOp(const std::string& tensorflow_op) {
   if (IsControlFlowOp(tensorflow_op)) {
     return true;
   }
@@ -82,14 +82,14 @@ bool IsUnsupportedFlexOp(const string& tensorflow_op) {
 }
 
 // Map from operator name to TF Lite enum value, for all builtins.
-const std::map& GetBuiltinOpsMap() {
-  static std::map* builtin_ops = nullptr;
+const std::map& GetBuiltinOpsMap() {
+  static std::map* builtin_ops = nullptr;
   if (builtin_ops == nullptr) {
-    builtin_ops = new std::map();
+    builtin_ops = new std::map();
 
     for (int i = BuiltinOperator_MIN; i <= BuiltinOperator_MAX; ++i) {
       BuiltinOperator op = static_cast(i);
-      string name = EnumNameBuiltinOperator(op);
+      std::string name = EnumNameBuiltinOperator(op);
       if (op != BuiltinOperator_CUSTOM && !name.empty()) {
         (*builtin_ops)[name] = op;
       }
@@ -99,10 +99,10 @@ const std::map& GetBuiltinOpsMap() {
 }
 
 void WriteModelToString(const flatbuffers::FlatBufferBuilder& builder,
-                        string* file_contents) {
+                        std::string* file_contents) {
   const uint8_t* buffer = builder.GetBufferPointer();
   int size = builder.GetSize();
-  *file_contents = string(reinterpret_cast(buffer), size);
+  *file_contents = std::string(reinterpret_cast(buffer), size);
 }
 
 }  // Anonymous namespace.
@@ -115,7 +115,7 @@ OperatorKey::OperatorKey(
     bool enable_select_tf_ops) {
   // Get the op name (by Toco definition).
   const ::toco::Operator& op = *op_signature.op;
-  string name = HelpfulOperatorTypeName(op);
+  std::string name = HelpfulOperatorTypeName(op);
 
   bool is_builtin = false;
   const auto& builtin_ops = GetBuiltinOpsMap();
@@ -146,7 +146,7 @@ OperatorKey::OperatorKey(
       is_flex_op_ = true;
       flex_tensorflow_op_ = tensorflow_op;
       custom_code_ =
-          string(::tflite::kFlexCustomCodePrefix) + flex_tensorflow_op_;
+          std::string(::tflite::kFlexCustomCodePrefix) + flex_tensorflow_op_;
     } else {
       custom_code_ = tensorflow_op;
     }
@@ -158,7 +158,7 @@ OperatorKey::OperatorKey(
     is_flex_op_ = true;
     flex_tensorflow_op_ = name;
     custom_code_ =
-        string(::tflite::kFlexCustomCodePrefix) + flex_tensorflow_op_;
+        std::string(::tflite::kFlexCustomCodePrefix) + flex_tensorflow_op_;
   } else {
     // If Flex is disabled or the original TensorFlow NodeDef isn't available,
     // we produce a custom op. This gives developers a chance to implement
@@ -175,7 +175,7 @@ OperatorKey::OperatorKey(
 
 void LoadTensorsMap(const Model& model, TensorsMap* tensors_map) {
   // First find a list of unique array names.
-  std::set names;
+  std::set names;
   for (const auto& array_pair : model.GetArrayMap()) {
     names.insert(array_pair.first);
   }
@@ -218,7 +218,7 @@ Offset>> ExportTensors(
   std::map> ordered_tensors;
 
   for (const auto& array_pair : model.GetArrayMap()) {
-    const string& tensor_name = array_pair.first;
+    const std::string& tensor_name = array_pair.first;
     const toco::Array& array = *array_pair.second;
 
     int buffer_index = buffers_to_write->size();
@@ -283,7 +283,7 @@ Offset> ExportOutputTensors(
     const Model& model, const details::TensorsMap& tensors_map,
     FlatBufferBuilder* builder) {
   std::vector outputs;
-  for (const string& output : model.flags.output_arrays()) {
+  for (const std::string& output : model.flags.output_arrays()) {
     outputs.push_back(tensors_map.at(output));
   }
   return builder->CreateVector(outputs);
@@ -295,10 +295,10 @@ Offset>> ExportOperatorCodes(
     const details::OperatorsMap& operators_map, FlatBufferBuilder* builder,
     const ExportParams& params) {
   // Map from operator name to TF Lite enum value, for all builtins.
-  std::map builtin_ops;
+  std::map builtin_ops;
   for (int i = BuiltinOperator_MIN; i <= BuiltinOperator_MAX; ++i) {
     BuiltinOperator op = static_cast(i);
-    string name = EnumNameBuiltinOperator(op);
+    std::string name = EnumNameBuiltinOperator(op);
     if (op != BuiltinOperator_CUSTOM && !name.empty()) {
       builtin_ops[name] = op;
     }
@@ -349,13 +349,13 @@ Offset>> ExportOperators(
   std::vector> op_vector;
   for (const auto& op : model.operators) {
     std::vector inputs;
-    for (const string& input : op->inputs) {
+    for (const std::string& input : op->inputs) {
       // -1 is the ID for optional tensor in TFLite output
       int id = model.IsOptionalArray(input) ? -1 : tensors_map.at(input);
       inputs.push_back(id);
     }
     std::vector outputs;
-    for (const string& output : op->outputs) {
+    for (const std::string& output : op->outputs) {
       outputs.push_back(tensors_map.at(output));
     }
     const toco::OperatorSignature op_signature = {op.get(), &model};
@@ -428,15 +428,15 @@ Offset>> ExportBuffers(
   return builder->CreateVector(buffer_vector);
 }
 
-tensorflow::Status Export(const Model& model, string* output_file_contents,
+tensorflow::Status Export(const Model& model, std::string* output_file_contents,
                           const ExportParams& params) {
   const auto ops_by_type = BuildOperatorByTypeMap(params.enable_select_tf_ops);
   return Export(model, output_file_contents, params, ops_by_type);
 }
 
-void ParseControlFlowErrors(std::set* custom_ops,
-                            std::vector* error_msgs) {
-  std::set unsupported_control_flow_ops;
+void ParseControlFlowErrors(std::set* custom_ops,
+                            std::vector* error_msgs) {
+  std::set unsupported_control_flow_ops;
   // Check if unsupported ops contains control flow ops. It's impossible
   // to implement these ops as custom ops at the moment.
   for (const auto& op : *custom_ops) {
@@ -471,10 +471,10 @@ void ExportModelVersionBuffer(
 }
 
 tensorflow::Status Export(
-    const Model& model, string* output_file_contents,
+    const Model& model, std::string* output_file_contents,
     const ExportParams& params,
     const std::map>& ops_by_type) {
-  for (const string& input_array : model.GetInvalidInputArrays()) {
+  for (const std::string& input_array : model.GetInvalidInputArrays()) {
     if (model.HasArray(input_array)) {
       return tensorflow::errors::InvalidArgument(
           absl::StrCat("Placeholder ", input_array,
@@ -509,11 +509,11 @@ tensorflow::Status Export(
   }
 
   // The set of used builtin ops.
-  std::set builtin_ops;
+  std::set builtin_ops;
   // The set of custom ops (not including Flex ops).
-  std::set custom_ops;
+  std::set custom_ops;
   // The set of Flex ops which are not supported.
-  std::set unsupported_flex_ops;
+  std::set unsupported_flex_ops;
 
   for (const auto& it : operators_map) {
     const details::OperatorKey& key = it.first;
@@ -540,7 +540,7 @@ tensorflow::Status Export(
                "40-tflite-op-request.md\n and pasting the following:\n\n";
       };
 
-      std::vector error_msgs;
+      std::vector error_msgs;
       ParseControlFlowErrors(&custom_ops, &error_msgs);
 
       // Remove ExpandDims and ReorderAxes from unimplemented list unless they
@@ -549,7 +549,7 @@ tensorflow::Status Export(
       // transformation is unable to run because the output shape is not
       // defined. This causes unnecessary confusion during model conversion
       // time.
-      std::set custom_ops_final;
+      std::set custom_ops_final;
       for (const auto& op_type : custom_ops) {
         if (op_type != "ReorderAxes" && op_type != "ExpandDims") {
           custom_ops_final.insert(op_type);
diff --git a/tensorflow/lite/toco/tflite/export.h b/tensorflow/lite/toco/tflite/export.h
index 3af77ffcf43..64f7c7b128f 100644
--- a/tensorflow/lite/toco/tflite/export.h
+++ b/tensorflow/lite/toco/tflite/export.h
@@ -35,19 +35,19 @@ struct ExportParams {
 
 // Transform the given tf.mini model into a TF Lite flatbuffer and deposit the
 // result in the given string.
-tensorflow::Status Export(const Model& model, string* output_file_contents,
+tensorflow::Status Export(const Model& model, std::string* output_file_contents,
                           const ExportParams& params);
 
 // Export API with custom TFLite operator mapping.
 tensorflow::Status Export(
-    const Model& model, string* output_file_contents,
+    const Model& model, std::string* output_file_contents,
     const ExportParams& params,
     const std::map>& ops_by_type);
 
 // This is for backward-compatibility.
 // TODO(ycling): Remove the deprecated entry functions.
 inline void Export(const Model& model, bool allow_custom_ops,
-                   bool quantize_weights, string* output_file_contents) {
+                   bool quantize_weights, std::string* output_file_contents) {
   ExportParams params;
   params.allow_custom_ops = allow_custom_ops;
   params.quantize_weights =
@@ -60,7 +60,7 @@ inline void Export(const Model& model, bool allow_custom_ops,
 // TODO(ycling): Remove the deprecated entry functions.
 inline void Export(
     const Model& model, bool allow_custom_ops, bool quantize_weights,
-    string* output_file_contents,
+    std::string* output_file_contents,
     const std::map>& ops_by_type) {
   ExportParams params;
   params.allow_custom_ops = allow_custom_ops;
@@ -72,7 +72,7 @@ inline void Export(
 
 // This is for backward-compatibility.
 // TODO(ycling): Remove the deprecated entry functions.
-inline void Export(const Model& model, string* output_file_contents) {
+inline void Export(const Model& model, std::string* output_file_contents) {
   ExportParams params;
   params.allow_custom_ops = true;
   auto status = Export(model, output_file_contents, params);
@@ -82,7 +82,7 @@ inline void Export(const Model& model, string* output_file_contents) {
 namespace details {
 
 // A map from tensor name to its final position in the TF Lite buffer.
-using TensorsMap = std::unordered_map;
+using TensorsMap = std::unordered_map;
 
 // A key to identify an operator.
 // Only when `type` is `kUnsupported`, `custom_code` is filled to
diff --git a/tensorflow/lite/toco/tflite/export_test.cc b/tensorflow/lite/toco/tflite/export_test.cc
index 19b77543c66..ed347a28d51 100644
--- a/tensorflow/lite/toco/tflite/export_test.cc
+++ b/tensorflow/lite/toco/tflite/export_test.cc
@@ -34,13 +34,13 @@ using ::testing::HasSubstr;
 class ExportTest : public ::testing::Test {
  protected:
   void ResetOperators() { input_model_.operators.clear(); }
-  void AddTensorsByName(std::initializer_list names) {
-    for (const string& name : names) {
+  void AddTensorsByName(std::initializer_list names) {
+    for (const std::string& name : names) {
       input_model_.GetOrCreateArray(name);
     }
   }
-  void AddOperatorsByName(std::initializer_list names) {
-    for (const string& name : names) {
+  void AddOperatorsByName(std::initializer_list names) {
+    for (const std::string& name : names) {
       if (name == "Conv") {
         auto* op = new ConvOperator;
         op->padding.type = PaddingType::kSame;
@@ -153,14 +153,15 @@ class ExportTest : public ::testing::Test {
   }
 
   tensorflow::Status ExportAndReturnStatus(const ExportParams& params) {
-    string result;
+    std::string result;
     return Export(input_model_, &result, params);
   }
 
-  std::vector ExportAndSummarizeOperators(const ExportParams& params) {
-    std::vector names;
+  std::vector ExportAndSummarizeOperators(
+      const ExportParams& params) {
+    std::vector names;
 
-    string result;
+    std::string result;
     auto status = Export(input_model_, &result, params);
     if (!status.ok()) {
       LOG(INFO) << status.error_message();
@@ -171,10 +172,12 @@ class ExportTest : public ::testing::Test {
 
     for (const ::tflite::OperatorCode* opcode : *model->operator_codes()) {
       if (opcode->builtin_code() != ::tflite::BuiltinOperator_CUSTOM) {
-        names.push_back(string("builtin:") + ::tflite::EnumNameBuiltinOperator(
-                                                 opcode->builtin_code()));
+        names.push_back(
+            std::string("builtin:") +
+            ::tflite::EnumNameBuiltinOperator(opcode->builtin_code()));
       } else {
-        names.push_back(string("custom:") + opcode->custom_code()->c_str());
+        names.push_back(std::string("custom:") +
+                        opcode->custom_code()->c_str());
       }
     }
 
@@ -185,7 +188,7 @@ class ExportTest : public ::testing::Test {
       const ExportParams& params) {
     std::vector indices;
 
-    string result;
+    std::string result;
     if (!Export(input_model_, &result, params).ok()) return indices;
     auto* model = ::tflite::GetModel(result.data());
 
@@ -257,7 +260,7 @@ TEST_F(ExportTest, ExportMinRuntime) {
   params.enable_select_tf_ops = false;
   params.quantize_weights = QuantizedBufferType::NONE;
 
-  string output;
+  std::string output;
   auto status = Export(input_model_, &output, params);
   auto* model = ::tflite::GetModel(output.data());
   EXPECT_EQ(model->metadata()->size(), 1);
@@ -265,7 +268,8 @@ TEST_F(ExportTest, ExportMinRuntime) {
   auto buf = model->metadata()->Get(0)->buffer();
   auto* buffer = (*model->buffers())[buf];
   auto* array = buffer->data();
-  string version(reinterpret_cast(array->data()), array->size());
+  std::string version(reinterpret_cast(array->data()),
+                      array->size());
   EXPECT_EQ(version, "1.6.0");
 }
 
@@ -275,7 +279,7 @@ TEST_F(ExportTest, ExportEmptyMinRuntime) {
   ExportParams params;
   params.allow_custom_ops = true;
 
-  string output;
+  std::string output;
   auto status = Export(input_model_, &output, params);
   auto* model = ::tflite::GetModel(output.data());
   EXPECT_EQ(model->metadata()->size(), 1);
@@ -283,7 +287,8 @@ TEST_F(ExportTest, ExportEmptyMinRuntime) {
   auto buf = model->metadata()->Get(0)->buffer();
   auto* buffer = (*model->buffers())[buf];
   auto* array = buffer->data();
-  string version(reinterpret_cast(array->data()), array->size());
+  std::string version(reinterpret_cast(array->data()),
+                      array->size());
   EXPECT_EQ(version, "");
 }
 
@@ -296,7 +301,7 @@ TEST_F(ExportTest, UnsupportedControlFlowErrors) {
   // The model contains control flow ops which are not convertible, so we should
   // check the returned error message.
 
-  string output;
+  std::string output;
   const auto ops_by_type = BuildOperatorByTypeMap();
   auto status = Export(input_model_, &output, params, ops_by_type);
   EXPECT_EQ(status.error_message(),
@@ -318,7 +323,7 @@ TEST_F(ExportTest, UnsupportedOpsAndNeedEnableFlex) {
   params.allow_custom_ops = false;
   params.enable_select_tf_ops = false;
 
-  string output;
+  std::string output;
   const auto ops_by_type = BuildOperatorByTypeMap();
   auto status = Export(input_model_, &output, params, ops_by_type);
   EXPECT_EQ(
@@ -348,7 +353,7 @@ TEST_F(ExportTest, UnsupportedOpsNeedCustomImplementation) {
   params.allow_custom_ops = false;
   params.enable_select_tf_ops = true;
 
-  string output;
+  std::string output;
   const auto ops_by_type = BuildOperatorByTypeMap();
   auto status = Export(input_model_, &output, params, ops_by_type);
   EXPECT_EQ(
@@ -378,7 +383,7 @@ TEST_F(ExportTest, UnsupportedControlFlowAndCustomOpsErrors) {
   // The model contains control flow ops which are not convertible, so we should
   // check the returned error message.
 
-  string output;
+  std::string output;
   const auto ops_by_type = BuildOperatorByTypeMap();
   auto status = Export(input_model_, &output, params, ops_by_type);
   EXPECT_EQ(
@@ -407,11 +412,11 @@ TEST_F(ExportTest, UnsupportedControlFlowAndCustomOpsErrors) {
 TEST_F(ExportTest, QuantizeWeights) {
   // Sanity check for quantize_weights parameter.
   BuildQuantizableTestModel();
-  string unquantized_result;
+  std::string unquantized_result;
   Export(input_model_, true, /*quantize_weights*/ false, &unquantized_result);
 
   BuildQuantizableTestModel();
-  string quantized_result;
+  std::string quantized_result;
   Export(input_model_, true, /*quantize_weights*/ true, &quantized_result);
 
   // The quantized models should be smaller.
@@ -443,12 +448,13 @@ class OpSetsTest : public ExportTest {
     }
   }
 
-  std::vector ImportExport(std::initializer_list op_names) {
+  std::vector ImportExport(
+      std::initializer_list op_names) {
     ResetOperators();
     if (!import_all_ops_as_unsupported_) {
       AddOperatorsByName(op_names);
     } else {
-      for (const string& name : op_names) {
+      for (const std::string& name : op_names) {
         auto* op = new TensorFlowUnsupportedOperator;
         op->tensorflow_op = name;
         input_model_.operators.emplace_back(op);
@@ -644,7 +650,7 @@ TEST_F(VersionedOpExportTest, Export) {
   AddConvOp(false);
   AddConvOp(true);
 
-  string result;
+  std::string result;
   const auto ops_by_type = BuildFakeOperatorByTypeMap();
   Export(input_model_, true, false, &result, ops_by_type);
 
diff --git a/tensorflow/lite/toco/tflite/import.cc b/tensorflow/lite/toco/tflite/import.cc
index 0f3dd48652e..136aa4ffaa8 100644
--- a/tensorflow/lite/toco/tflite/import.cc
+++ b/tensorflow/lite/toco/tflite/import.cc
@@ -99,7 +99,7 @@ void ImportTensors(const ::tflite::Model& input_model, Model* model) {
 
 void ImportOperators(
     const ::tflite::Model& input_model,
-    const std::map>& ops_by_name,
+    const std::map>& ops_by_name,
     const details::TensorsTable& tensors_table,
     const details::OperatorsTable& operators_table, Model* model) {
   // TODO(aselle): add support for multiple subgraphs.
@@ -112,12 +112,12 @@ void ImportOperators(
       LOG(FATAL) << "Index " << index << " must be between zero and "
                  << operators_table.size();
     }
-    string opname = operators_table.at(index);
+    std::string opname = operators_table.at(index);
 
     // Find and use the appropriate operator deserialization factory.
     std::unique_ptr new_op = nullptr;
     if (ops_by_name.count(opname) == 0) {
-      string effective_opname = "TENSORFLOW_UNSUPPORTED";
+      std::string effective_opname = "TENSORFLOW_UNSUPPORTED";
       if (ops_by_name.count(effective_opname) == 0) {
         LOG(FATAL) << "Internal logic error: TENSORFLOW_UNSUPPORTED not found.";
       }
@@ -147,10 +147,10 @@ void ImportOperators(
       auto input_index = inputs->Get(i);
       // input_index == -1 indicates optional tensor.
       if (input_index != -1) {
-        const string& input_name = tensors_table.at(input_index);
+        const std::string& input_name = tensors_table.at(input_index);
         op->inputs.push_back(input_name);
       } else {
-        const string& tensor_name =
+        const std::string& tensor_name =
             toco::AvailableArrayName(*model, "OptionalTensor");
         model->CreateOptionalArray(tensor_name);
         op->inputs.push_back(tensor_name);
@@ -159,7 +159,7 @@ void ImportOperators(
     auto outputs = input_op->outputs();
     for (int i = 0; i < outputs->Length(); i++) {
       auto output_index = outputs->Get(i);
-      const string& output_name = tensors_table.at(output_index);
+      const std::string& output_name = tensors_table.at(output_index);
       op->outputs.push_back(output_name);
     }
   }
@@ -173,7 +173,7 @@ void ImportIOTensors(const ModelFlags& model_flags,
     auto inputs = (*input_model.subgraphs())[0]->inputs();
     if (inputs) {
       for (int input : *inputs) {
-        const string& input_name = tensors_table.at(input);
+        const std::string& input_name = tensors_table.at(input);
         model->flags.add_input_arrays()->set_name(input_name);
       }
     }
@@ -184,7 +184,7 @@ void ImportIOTensors(const ModelFlags& model_flags,
     auto outputs = (*input_model.subgraphs())[0]->outputs();
     if (outputs) {
       for (int output : *outputs) {
-        const string& output_name = tensors_table.at(output);
+        const std::string& output_name = tensors_table.at(output);
         model->flags.add_output_arrays(output_name);
       }
     }
@@ -199,7 +199,7 @@ bool Verify(const void* buf, size_t len) {
 }  // namespace
 
 std::unique_ptr Import(const ModelFlags& model_flags,
-                              const string& input_file_contents) {
+                              const std::string& input_file_contents) {
   ::tflite::AlwaysTrueResolver r;
   if (!::tflite::Verify(input_file_contents.data(), input_file_contents.size(),
                         r, ::tflite::DefaultErrorReporter())) {
diff --git a/tensorflow/lite/toco/tflite/import.h b/tensorflow/lite/toco/tflite/import.h
index f5de3b53b5b..bac55aae8b9 100644
--- a/tensorflow/lite/toco/tflite/import.h
+++ b/tensorflow/lite/toco/tflite/import.h
@@ -24,17 +24,17 @@ namespace tflite {
 
 // Parse the given string as TF Lite flatbuffer and return a new tf.mini model.
 std::unique_ptr Import(const ModelFlags &model_flags,
-                              const string &input_file_contents);
+                              const std::string &input_file_contents);
 
 namespace details {
 
 // The names of all tensors found in a TF Lite model.
-using TensorsTable = std::vector;
+using TensorsTable = std::vector;
 
 // The names of all operators found in TF Lite model. If the operator is
 // builtin, the string representation of the corresponding enum value is used
 // as name.
-using OperatorsTable = std::vector;
+using OperatorsTable = std::vector;
 
 void LoadTensorsTable(const ::tflite::Model &input_model,
                       TensorsTable *tensors_table);
diff --git a/tensorflow/lite/toco/tflite/import_test.cc b/tensorflow/lite/toco/tflite/import_test.cc
index b00c4124d83..6163ebab45b 100644
--- a/tensorflow/lite/toco/tflite/import_test.cc
+++ b/tensorflow/lite/toco/tflite/import_test.cc
@@ -134,9 +134,9 @@ class ImportTest : public ::testing::Test {
 
     input_model_ = ::tflite::GetModel(builder_.GetBufferPointer());
   }
-  string InputModelAsString() {
-    return string(reinterpret_cast(builder_.GetBufferPointer()),
-                  builder_.GetSize());
+  std::string InputModelAsString() {
+    return std::string(reinterpret_cast(builder_.GetBufferPointer()),
+                       builder_.GetSize());
   }
   flatbuffers::FlatBufferBuilder builder_;
   const ::tflite::Model* input_model_ = nullptr;
diff --git a/tensorflow/lite/toco/tflite/op_version.cc b/tensorflow/lite/toco/tflite/op_version.cc
index cf127a9f459..efa53c69cae 100644
--- a/tensorflow/lite/toco/tflite/op_version.cc
+++ b/tensorflow/lite/toco/tflite/op_version.cc
@@ -29,7 +29,7 @@ namespace tflite {
 
 // Deprecated and please register new ops/versions in
 // tflite/tools/versioning/op_version.cc".
-string GetMinimumRuntimeVersionForModel(const Model& model) {
+std::string GetMinimumRuntimeVersionForModel(const Model& model) {
   // Use this as the placeholder string if a particular op is not yet included
   // in any Tensorflow's RC/Final release source package. Once that op is
   // included in the release, please update this with the real version string.
@@ -37,8 +37,8 @@ string GetMinimumRuntimeVersionForModel(const Model& model) {
   // A map from the version key of an op to its minimum runtime version.
   // For example, {{kAveragePool, 1}, "1.5.0"},  means the 1st version of
   // AveragePool requires a minimum TF Lite runtime version '1.5.0`.
-  static const std::map, string>* op_version_map =
-      new std::map, string>({
+  static const std::map, std::string>*
+      op_version_map = new std::map, std::string>({
           {{OperatorType::kAveragePool, 1}, "1.5.0"},
           {{OperatorType::kAveragePool, 2}, "1.14.0"},
           {{OperatorType::kAveragePool, 3}, kPendingReleaseOpVersion},
@@ -253,7 +253,7 @@ string GetMinimumRuntimeVersionForModel(const Model& model) {
       tflite::BuildOperatorByTypeMap(false /*enable_select_tf_ops=*/);
   OperatorSignature op_signature;
   op_signature.model = &model;
-  string model_min_version;
+  std::string model_min_version;
   for (const auto& op : model.operators) {
     if (op_types_map.find(op->type) == op_types_map.end()) continue;
     op_signature.op = op.get();
diff --git a/tensorflow/lite/toco/tflite/op_version_test.cc b/tensorflow/lite/toco/tflite/op_version_test.cc
index 14b086471b7..8466fc35ad7 100644
--- a/tensorflow/lite/toco/tflite/op_version_test.cc
+++ b/tensorflow/lite/toco/tflite/op_version_test.cc
@@ -27,9 +27,9 @@ TEST(OpVersionTest, MinimumVersionForSameOpVersions) {
   Model model;
   // Float convolutional kernel is introduced since '1.5.0'.
   std::unique_ptr conv(new ConvOperator());
-  const string conv_input = "conv_input";
-  const string conv_filter = "conv_filter";
-  const string conv_output = "conv_output";
+  const std::string conv_input = "conv_input";
+  const std::string conv_filter = "conv_filter";
+  const std::string conv_output = "conv_output";
   conv->inputs.push_back(conv_input);
   conv->inputs.push_back(conv_filter);
   conv->outputs.push_back(conv_output);
@@ -44,8 +44,8 @@ TEST(OpVersionTest, MinimumVersionForSameOpVersions) {
 
   // Float softmax kernel is introduced since '1.5.0'.
   std::unique_ptr softmax(new SoftmaxOperator());
-  const string softmax_input = "softmax_input";
-  const string softmax_output = "softmax_output";
+  const std::string softmax_input = "softmax_input";
+  const std::string softmax_output = "softmax_output";
   softmax->inputs.push_back(softmax_input);
   softmax->outputs.push_back(softmax_output);
   array_map[softmax_input] = std::unique_ptr(new Array);
@@ -60,9 +60,9 @@ TEST(OpVersionTest, MinimumVersionForMultipleOpVersions) {
   Model model;
   // Dilated DepthWiseConvolution is introduced since '1.12.0'.
   std::unique_ptr conv(new DepthwiseConvOperator());
-  const string conv_input = "conv_input";
-  const string conv_filter = "conv_filter";
-  const string conv_output = "conv_output";
+  const std::string conv_input = "conv_input";
+  const std::string conv_filter = "conv_filter";
+  const std::string conv_output = "conv_output";
   conv->inputs.push_back(conv_input);
   conv->inputs.push_back(conv_filter);
   conv->outputs.push_back(conv_output);
@@ -77,10 +77,10 @@ TEST(OpVersionTest, MinimumVersionForMultipleOpVersions) {
   // FullyConnected op with kShuffled4x16Int8 weight format is introduced from
   // '1.10.0'.
   std::unique_ptr fc(new FullyConnectedOperator());
-  const string fc_input = "fc_input";
-  const string fc_weights = "fc_weights";
-  const string fc_bias = "fc_bias";
-  const string fc_output = "fc_output";
+  const std::string fc_input = "fc_input";
+  const std::string fc_weights = "fc_weights";
+  const std::string fc_bias = "fc_bias";
+  const std::string fc_output = "fc_output";
   fc->inputs.push_back(fc_input);
   fc->inputs.push_back(fc_weights);
   fc->inputs.push_back(fc_bias);
@@ -121,10 +121,10 @@ TEST(OpVersionTest, MinimumVersionForMixedOpVersions) {
   // FullyConnected op with kShuffled4x16Int8 weight format is introduced from
   // '1.10.0'.
   std::unique_ptr fc(new FullyConnectedOperator());
-  const string fc_input = "fc_input";
-  const string fc_weights = "fc_weights";
-  const string fc_bias = "fc_bias";
-  const string fc_output = "fc_output";
+  const std::string fc_input = "fc_input";
+  const std::string fc_weights = "fc_weights";
+  const std::string fc_bias = "fc_bias";
+  const std::string fc_output = "fc_output";
   fc->inputs.push_back(fc_input);
   fc->inputs.push_back(fc_weights);
   fc->inputs.push_back(fc_bias);
diff --git a/tensorflow/lite/toco/tflite/operator.cc b/tensorflow/lite/toco/tflite/operator.cc
index fee10a19787..be539cf6054 100644
--- a/tensorflow/lite/toco/tflite/operator.cc
+++ b/tensorflow/lite/toco/tflite/operator.cc
@@ -238,7 +238,7 @@ class SpaceToBatchND
                    TocoOperator* op) const override {}
 
   int GetVersion(const OperatorSignature& op_signature) const override {
-    const string& input_name = op_signature.op->inputs[0];
+    const std::string& input_name = op_signature.op->inputs[0];
     const Array& input_array = op_signature.model->GetArray(input_name);
     ::tflite::OpSignature op_sig =
         GetVersioningOpSig(builtin_op(), op_signature);
@@ -268,8 +268,8 @@ class Sub : public BuiltinOperatorinputs[0];
-    const string& input2_name = op_signature.op->inputs[1];
+    const std::string& input1_name = op_signature.op->inputs[0];
+    const std::string& input2_name = op_signature.op->inputs[1];
     const Array& input1_array = op_signature.model->GetArray(input1_name);
     const Array& input2_array = op_signature.model->GetArray(input2_name);
     ::tflite::OpSignature op_sig =
@@ -305,8 +305,8 @@ class Div : public BuiltinOperatorinputs[0];
-    const string& input2_name = op_signature.op->inputs[1];
+    const std::string& input1_name = op_signature.op->inputs[0];
+    const std::string& input2_name = op_signature.op->inputs[1];
     const Array& input1_array = op_signature.model->GetArray(input1_name);
     const Array& input2_array = op_signature.model->GetArray(input2_name);
     ::tflite::OpSignature op_sig =
@@ -339,7 +339,7 @@ class BatchToSpaceND
                    TocoOperator* op) const override {}
 
   int GetVersion(const OperatorSignature& op_signature) const override {
-    const string& input_name = op_signature.op->inputs[0];
+    const std::string& input_name = op_signature.op->inputs[0];
     const Array& input_array = op_signature.model->GetArray(input_name);
     ::tflite::OpSignature op_sig =
         GetVersioningOpSig(builtin_op(), op_signature);
@@ -662,9 +662,9 @@ class Mul : public BuiltinOperatorinputs[0];
-    const string& input2_name = op_signature.op->inputs[1];
-    const string& output_name = op_signature.op->outputs[0];
+    const std::string& input1_name = op_signature.op->inputs[0];
+    const std::string& input2_name = op_signature.op->inputs[1];
+    const std::string& output_name = op_signature.op->outputs[0];
     const Array& input1_array = op_signature.model->GetArray(input1_name);
     const Array& input2_array = op_signature.model->GetArray(input2_name);
     const Array& output_array = op_signature.model->GetArray(output_name);
@@ -1440,7 +1440,7 @@ class Unpack : public BuiltinOperatorinputs[0];
+    const std::string& input_name = op_signature.op->inputs[0];
     const Array& input_array = op_signature.model->GetArray(input_name);
     // If the op take int8/uint8 input, it is version 2.
     if (input_array.data_type == ArrayDataType::kInt8 ||
@@ -1577,7 +1577,7 @@ class Where : public BuiltinOperator WriteFlexOpOptions(
-    const string& tensorflow_node_def) {
+    const std::string& tensorflow_node_def) {
   auto fbb = absl::make_unique();
 
   ::tensorflow::NodeDef node_def;
@@ -1597,7 +1597,7 @@ std::unique_ptr WriteFlexOpOptions(
 
 class TensorFlowUnsupported : public BaseOperator {
  public:
-  TensorFlowUnsupported(const string& name, OperatorType type,
+  TensorFlowUnsupported(const std::string& name, OperatorType type,
                         bool enable_select_tf_ops)
       : BaseOperator(name, type), enable_select_tf_ops_(enable_select_tf_ops) {}
 
@@ -1676,7 +1676,7 @@ class TensorFlowUnsupported : public BaseOperator {
         case tensorflow::AttrValue::kList:
           if (attr.list().s_size() > 0) {
             auto start = fbb->StartVector(key);
-            for (const string& v : attr.list().s()) {
+            for (const std::string& v : attr.list().s()) {
               fbb->Add(v);
             }
             fbb->EndVector(start, /*typed=*/true, /*fixed=*/false);
@@ -1736,10 +1736,11 @@ class TensorFlowUnsupported : public BaseOperator {
           break;
         case flexbuffers::FBT_BOOL:
           (*attr)[key].set_b(value.AsBool());
-          if (string(key) == "_output_quantized") {
+          if (std::string(key) == "_output_quantized") {
             op->quantized = value.AsBool();
           }
-          if (string(key) == "_support_output_type_float_in_quantized_op") {
+          if (std::string(key) ==
+              "_support_output_type_float_in_quantized_op") {
             op->support_output_type_float_in_quantized_op = value.AsBool();
           }
           break;
@@ -2095,9 +2096,9 @@ std::map> BuildOperatorByTypeMap(
   return result;
 }
 
-std::map> BuildOperatorByNameMap(
+std::map> BuildOperatorByNameMap(
     bool enable_select_tf_ops) {
-  std::map> result;
+  std::map> result;
 
   std::vector> ops =
       BuildOperatorList(enable_select_tf_ops);
@@ -2109,7 +2110,7 @@ std::map> BuildOperatorByNameMap(
 }
 
 bool ShouldExportAsFlexOp(bool enable_select_tf_ops,
-                          const string& tensorflow_op_name) {
+                          const std::string& tensorflow_op_name) {
   // If Flex ops aren't allow at all, simply return false.
   if (!enable_select_tf_ops) {
     return false;
diff --git a/tensorflow/lite/toco/tflite/operator.h b/tensorflow/lite/toco/tflite/operator.h
index 19d92145e0c..fb79b97f46e 100644
--- a/tensorflow/lite/toco/tflite/operator.h
+++ b/tensorflow/lite/toco/tflite/operator.h
@@ -30,7 +30,7 @@ class BaseOperator;
 // Return a map contained all know TF Lite Operators, keyed by their names.
 // TODO(ycling): The pattern to propagate parameters (e.g. enable_select_tf_ops)
 // is ugly here. Consider refactoring.
-std::map> BuildOperatorByNameMap(
+std::map> BuildOperatorByNameMap(
     bool enable_select_tf_ops = false);
 
 // Return a map contained all know TF Lite Operators, keyed by the type of
@@ -41,7 +41,7 @@ std::map> BuildOperatorByTypeMap(
 // Write the custom option FlexBuffer with a serialized TensorFlow NodeDef
 // for a Flex op.
 std::unique_ptr WriteFlexOpOptions(
-    const string& tensorflow_node_def);
+    const std::string& tensorflow_node_def);
 
 // These are the flatbuffer types for custom and builtin options.
 using CustomOptions = flatbuffers::Vector;
@@ -71,11 +71,11 @@ struct Options {
 class BaseOperator {
  public:
   // Build an operator with the given TF Lite name and tf.mini type.
-  BaseOperator(const string& name, OperatorType type)
+  BaseOperator(const std::string& name, OperatorType type)
       : name_(name), type_(type) {}
   virtual ~BaseOperator() = default;
 
-  string name() const { return name_; }
+  std::string name() const { return name_; }
   OperatorType type() const { return type_; }
 
   // Given a tf.mini operator, create the corresponding flatbuffer options and
@@ -111,7 +111,7 @@ class BaseOperator {
   }
 
  private:
-  string name_;
+  std::string name_;
   OperatorType type_;
 };
 
@@ -123,7 +123,7 @@ class BaseOperator {
 // Helper function to determine if a unsupported TensorFlow op should be
 // exported as an Flex op or a regular custom op.
 bool ShouldExportAsFlexOp(bool enable_select_tf_ops,
-                          const string& tensorflow_op_name);
+                          const std::string& tensorflow_op_name);
 
 }  // namespace tflite
 
diff --git a/tensorflow/lite/toco/tflite/operator_test.cc b/tensorflow/lite/toco/tflite/operator_test.cc
index a4fe01e4afd..cb466fef079 100644
--- a/tensorflow/lite/toco/tflite/operator_test.cc
+++ b/tensorflow/lite/toco/tflite/operator_test.cc
@@ -30,8 +30,8 @@ namespace {
 class OperatorTest : public ::testing::Test {
  protected:
   // Return the operator for the given name and type.
-  const BaseOperator& GetOperator(const string& name, OperatorType type) {
-    using OpsByName = std::map>;
+  const BaseOperator& GetOperator(const std::string& name, OperatorType type) {
+    using OpsByName = std::map>;
     using OpsByType = std::map>;
 
     static auto* by_name = new OpsByName(BuildOperatorByNameMap());
@@ -86,7 +86,7 @@ class OperatorTest : public ::testing::Test {
   // Verify serialization and deserialization of simple operators (those
   // that don't have any configuration parameters).
   template 
-  void CheckSimpleOperator(const string& name, OperatorType type) {
+  void CheckSimpleOperator(const std::string& name, OperatorType type) {
     Options options;
     auto output_toco_op =
         SerializeAndDeserialize(GetOperator(name, type), T(), &options);
@@ -99,7 +99,7 @@ class OperatorTest : public ::testing::Test {
   }
 
   template 
-  void CheckReducerOperator(const string& name, OperatorType type) {
+  void CheckReducerOperator(const std::string& name, OperatorType type) {
     T op;
 
     op.keep_dims = false;
diff --git a/tensorflow/lite/toco/tflite/types.cc b/tensorflow/lite/toco/tflite/types.cc
index 96cad557baf..9d4ab8434d1 100644
--- a/tensorflow/lite/toco/tflite/types.cc
+++ b/tensorflow/lite/toco/tflite/types.cc
@@ -25,7 +25,7 @@ DataBuffer::FlatBufferOffset CopyStringToBuffer(
     const Array& array, flatbuffers::FlatBufferBuilder* builder) {
   const auto& src_data = array.GetBuffer().data;
   ::tflite::DynamicBuffer dyn_buffer;
-  for (const string& str : src_data) {
+  for (const std::string& str : src_data) {
     dyn_buffer.AddString(str.c_str(), str.length());
   }
   char* tensor_buffer;
@@ -58,12 +58,12 @@ DataBuffer::FlatBufferOffset CopyBuffer(
 
 void CopyStringFromBuffer(const ::tflite::Buffer& buffer, Array* array) {
   auto* src_data = reinterpret_cast(buffer.data()->data());
-  std::vector* dst_data =
+  std::vector* dst_data =
       &array->GetMutableBuffer().data;
   int32_t num_strings = ::tflite::GetStringCount(src_data);
   for (int i = 0; i < num_strings; i++) {
     ::tflite::StringRef str_ref = ::tflite::GetString(src_data, i);
-    string this_str(str_ref.str, str_ref.len);
+    std::string this_str(str_ref.str, str_ref.len);
     dst_data->push_back(this_str);
   }
 }

From 642ad434d8315561ef9cc02cc9157436fe9c0f72 Mon Sep 17 00:00:00 2001
From: Jingyue Wu 
Date: Fri, 19 Jun 2020 09:50:56 -0700
Subject: [PATCH 297/867] Fix exports_files.

cl/317237033 replaced cwise_op_neg.cc with cwise_op_neg_1.cc and
cwise_op_neg_2.cc.

PiperOrigin-RevId: 317325461
Change-Id: Ib44ff36474b7e55d9e84ff737ea82b9dac46b9f9
---
 tensorflow/core/kernels/BUILD | 3 ++-
 1 file changed, 2 insertions(+), 1 deletion(-)

diff --git a/tensorflow/core/kernels/BUILD b/tensorflow/core/kernels/BUILD
index 0b7a092033b..e2ff5aed283 100644
--- a/tensorflow/core/kernels/BUILD
+++ b/tensorflow/core/kernels/BUILD
@@ -8802,7 +8802,8 @@ exports_files([
     "cwise_op_mod.cc",
     "cwise_op_mul_1.cc",
     "cwise_op_mul_2.cc",
-    "cwise_op_neg.cc",
+    "cwise_op_neg_1.cc",
+    "cwise_op_neg_2.cc",
     "cwise_op_not_equal_to_1.cc",
     "cwise_op_not_equal_to_2.cc",
     "cwise_op_round.cc",

From c3cc3c40b08a37535a281e1d7a5fd7d3d802aac6 Mon Sep 17 00:00:00 2001
From: Mihai Maruseac 
Date: Fri, 19 Jun 2020 09:55:49 -0700
Subject: [PATCH 298/867] Move fuzzers for TF ops to own subdir. Trim some
 dependencies.

This duplicates some of the BUILD dependency tree to go around the need to link huge bottleneck dependencies (such as `//tensorflow/core:framework`). Until TF can use `cc_shared_library` in a stable way (and all support in Bazel exists), we will need to use the duplicated tree for fuzzing.

PiperOrigin-RevId: 317326319
Change-Id: I1493e3ae7340298971fe15bd3702b63657f9bf9f
---
 tensorflow/core/framework/BUILD               |   1 +
 tensorflow/security/fuzzing/BUILD             |  14 --
 tensorflow/security/fuzzing/op_fuzzing/BUILD  |  39 +++++
 .../fuzzing/op_fuzzing/fuzz_session.h         | 156 ++++++++++++++++++
 .../fuzzing/{ => op_fuzzing}/identity_fuzz.cc |   2 +-
 5 files changed, 197 insertions(+), 15 deletions(-)
 create mode 100644 tensorflow/security/fuzzing/op_fuzzing/BUILD
 create mode 100644 tensorflow/security/fuzzing/op_fuzzing/fuzz_session.h
 rename tensorflow/security/fuzzing/{ => op_fuzzing}/identity_fuzz.cc (95%)

diff --git a/tensorflow/core/framework/BUILD b/tensorflow/core/framework/BUILD
index 52f15dcb5c2..d47c74a629d 100644
--- a/tensorflow/core/framework/BUILD
+++ b/tensorflow/core/framework/BUILD
@@ -719,6 +719,7 @@ tf_cuda_library(
     visibility = [
         "//tensorflow/core:__pkg__",
         "//tensorflow/core/util:__pkg__",
+        "//tensorflow/security/fuzzing:__subpackages__",
     ],
     deps = [
         ":allocation_description_proto_cc",
diff --git a/tensorflow/security/fuzzing/BUILD b/tensorflow/security/fuzzing/BUILD
index 9b5aeec2d36..6b6c8275275 100644
--- a/tensorflow/security/fuzzing/BUILD
+++ b/tensorflow/security/fuzzing/BUILD
@@ -18,17 +18,3 @@ tf_fuzz_target(
         "//tensorflow/core/platform:status",
     ],
 )
-
-# A trivial fuzzer with no pre-specified corpus.
-# TODO(mihaimaruseac): Move fuzz_session and the op fuzzers to a subdirectory
-tf_fuzz_target(
-    name = "identity_fuzz",
-    srcs = ["identity_fuzz.cc"],
-    deps = [
-        "//tensorflow/cc:cc_ops",
-        "//tensorflow/core/kernels/fuzzing:fuzz_session",
-        # Needed only to transitiviely link dependencies
-        "//tensorflow/cc:scope",
-        "//tensorflow/core:core_cpu",
-    ],
-)
diff --git a/tensorflow/security/fuzzing/op_fuzzing/BUILD b/tensorflow/security/fuzzing/op_fuzzing/BUILD
new file mode 100644
index 00000000000..aacd2f16cc4
--- /dev/null
+++ b/tensorflow/security/fuzzing/op_fuzzing/BUILD
@@ -0,0 +1,39 @@
+# Fuzzing TensorFlow ops..
+# Most ops have a similar set of dependencies and a similar fuzzing
+# infrastructure. Hence, we gather everything in one single place.
+# Note that these fuzzers cover a large part of TF, they are not granular.
+
+load(
+    "//tensorflow/security/fuzzing:tf_fuzzing.bzl",
+    "tf_fuzz_target",
+)
+
+package(
+    licenses = ["notice"],  # Apache 2.0
+)
+
+# Since all ops need to have a graph created before being fuzzed, we define
+# this header-only library to handle the needed plumbing.
+cc_library(
+    name = "fuzz_session",
+    hdrs = ["fuzz_session.h"],
+    deps = [
+        "//tensorflow/cc:scope",
+        "//tensorflow/core:core_cpu_base",
+        "//tensorflow/core:session_options",
+        "//tensorflow/core/common_runtime:direct_session_internal",
+        "//tensorflow/core/framework:tensor",
+        "//tensorflow/core/platform:status",
+    ],
+)
+
+# A trivial fuzzer with no pre-specified corpus.
+tf_fuzz_target(
+    name = "identity_fuzz",
+    srcs = ["identity_fuzz.cc"],
+    deps = [
+        ":fuzz_session",
+        "//tensorflow/cc:cc_ops",
+        "//tensorflow/core/kernels:array",
+    ],
+)
diff --git a/tensorflow/security/fuzzing/op_fuzzing/fuzz_session.h b/tensorflow/security/fuzzing/op_fuzzing/fuzz_session.h
new file mode 100644
index 00000000000..575212b3b86
--- /dev/null
+++ b/tensorflow/security/fuzzing/op_fuzzing/fuzz_session.h
@@ -0,0 +1,156 @@
+/* Copyright 2016 Google Inc. All Rights Reserved.
+
+Licensed under the Apache License, Version 2.0 (the "License");
+you may not use this file except in compliance with the License.
+You may obtain a copy of the License at
+
+    http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software
+distributed under the License is distributed on an "AS IS" BASIS,
+WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+See the License for the specific language governing permissions and
+limitations under the License.
+==============================================================================*/
+#ifndef TENSORFLOW_SECURITY_FUZZING_OP_FUZZING_FUZZ_SESSION_H_
+#define TENSORFLOW_SECURITY_FUZZING_OP_FUZZING_FUZZ_SESSION_H_
+
+#include 
+#include 
+#include 
+#include 
+
+#include "tensorflow/cc/framework/scope.h"
+#include "tensorflow/core/framework/tensor.h"
+#include "tensorflow/core/platform/status.h"
+#include "tensorflow/core/public/session.h"
+#include "tensorflow/core/public/session_options.h"
+
+// Standard invoking function macro to dispatch to a fuzzer class.
+#define STANDARD_TF_FUZZ_FUNCTION(FuzzerClass)                              \
+  extern "C" int LLVMFuzzerTestOneInput(const uint8_t* data, size_t size) { \
+    static FuzzerClass* fuzzer = new FuzzerClass();                         \
+    return fuzzer->Fuzz(data, size);                                        \
+  }
+
+// Standard builder for hooking one placeholder to one op.
+#define SINGLE_INPUT_OP_BUILDER(dtype, opName)                          \
+  void BuildGraph(const Scope& scope) override {                        \
+    auto op_node =                                                      \
+        tensorflow::ops::Placeholder(scope.WithOpName("input"), dtype); \
+    (void)tensorflow::ops::opName(scope.WithOpName("output"), op_node); \
+  }
+
+namespace tensorflow {
+namespace fuzzing {
+
+// Create a TensorFlow session using a specific GraphDef created
+// by BuildGraph(), and make it available for fuzzing.
+// Users must override BuildGraph and FuzzImpl to specify
+// (1) which operations are being fuzzed; and
+// (2) How to translate the uint8_t* buffer from the fuzzer
+//     to a Tensor or Tensors that are semantically appropriate
+//     for the op under test.
+// For the simple cases of testing a single op that takes a single
+// input Tensor, use the SINGLE_INPUT_OP_BUILDER(dtype, opName) macro in place
+// of defining BuildGraphDef.
+//
+// Typical use:
+// class FooFuzzer : public FuzzSession {
+//   SINGLE_INPUT_OP_BUILDER(DT_INT8, Identity);
+//   void FuzzImpl(const uint8_t* data, size_t size) {
+//      ... convert data and size to a Tensor, pass it to:
+//      RunInputs({{"input", input_tensor}});
+//
+class FuzzSession {
+ public:
+  FuzzSession() : initialized_(false) {}
+  virtual ~FuzzSession() {}
+
+  // Constructs a Graph using the supplied Scope.
+  // By convention, the graph should have inputs named "input1", ...
+  // "inputN", and one output node, named "output".
+  // Users of FuzzSession should override this method to create their graph.
+  virtual void BuildGraph(const Scope& scope) = 0;
+
+  // Implements the logic that converts an opaque byte buffer
+  // from the fuzzer to Tensor inputs to the graph.  Users must override.
+  virtual void FuzzImpl(const uint8_t* data, size_t size) = 0;
+
+  // Initializes the FuzzSession.  Not safe for multithreading.
+  // Separate init function because the call to virtual BuildGraphDef
+  // can't be put into the constructor.
+  Status InitIfNeeded() {
+    if (initialized_) {
+      return Status::OK();
+    }
+    initialized_ = true;
+
+    Scope root = Scope::DisabledShapeInferenceScope().ExitOnError();
+    SessionOptions options;
+    session_ = std::unique_ptr(NewSession(options));
+
+    BuildGraph(root);
+
+    GraphDef graph_def;
+    TF_CHECK_OK(root.ToGraphDef(&graph_def));
+
+    Status status = session_->Create(graph_def);
+    if (!status.ok()) {
+      // This is FATAL, because this code is designed to fuzz an op
+      // within a session.  Failure to create the session means we
+      // can't send any data to the op.
+      LOG(FATAL) << "Could not create session: " << status.error_message();
+    }
+    return status;
+  }
+
+  // Runs the TF session by pulling on the "output" node, attaching
+  // the supplied input_tensor to the input node(s), and discarding
+  // any returned output.
+  // Note: We are ignoring Status from Run here since fuzzers don't need to
+  // check it (as that will slow them down and printing/logging is useless).
+  void RunInputs(const std::vector >& inputs) {
+    RunInputsWithStatus(inputs).IgnoreError();
+  }
+
+  // Same as RunInputs but don't ignore status
+  Status RunInputsWithStatus(
+      const std::vector >& inputs) {
+    return session_->Run(inputs, {}, {"output"}, nullptr);
+  }
+
+  // Dispatches to FuzzImpl;  small amount of sugar to keep the code
+  // of the per-op fuzzers tiny.
+  int Fuzz(const uint8_t* data, size_t size) {
+    Status status = InitIfNeeded();
+    TF_CHECK_OK(status) << "Fuzzer graph initialization failed: "
+                        << status.error_message();
+    // No return value from fuzzing:  Success is defined as "did not
+    // crash".  The actual application results are irrelevant.
+    FuzzImpl(data, size);
+    return 0;
+  }
+
+ private:
+  bool initialized_;
+  std::unique_ptr session_;
+};
+
+// A specialized fuzz implementation for ops that take
+// a single string.  Caller must still define the op
+// to plumb by overriding BuildGraph or using
+// a plumbing macro.
+class FuzzStringInputOp : public FuzzSession {
+  void FuzzImpl(const uint8_t* data, size_t size) final {
+    Tensor input_tensor(tensorflow::DT_STRING, TensorShape({}));
+    input_tensor.scalar()() =
+        string(reinterpret_cast(data), size);
+    RunInputs({{"input", input_tensor}});
+  }
+};
+
+}  // end namespace fuzzing
+}  // end namespace tensorflow
+
+#endif  // TENSORFLOW_SECURITY_FUZZING_OP_FUZZING_FUZZ_SESSION_H_
diff --git a/tensorflow/security/fuzzing/identity_fuzz.cc b/tensorflow/security/fuzzing/op_fuzzing/identity_fuzz.cc
similarity index 95%
rename from tensorflow/security/fuzzing/identity_fuzz.cc
rename to tensorflow/security/fuzzing/op_fuzzing/identity_fuzz.cc
index 4c1049d381b..a63c35b45e2 100644
--- a/tensorflow/security/fuzzing/identity_fuzz.cc
+++ b/tensorflow/security/fuzzing/op_fuzzing/identity_fuzz.cc
@@ -14,7 +14,7 @@ limitations under the License.
 ==============================================================================*/
 
 #include "tensorflow/cc/ops/standard_ops.h"
-#include "tensorflow/core/kernels/fuzzing/fuzz_session.h"
+#include "tensorflow/security/fuzzing/op_fuzzing/fuzz_session.h"
 
 namespace tensorflow {
 namespace fuzzing {

From 8e305b9cec35ccf9821c2fd2a82194e328d98704 Mon Sep 17 00:00:00 2001
From: Denisa Roberts 
Date: Wed, 6 May 2020 11:29:37 -0400
Subject: [PATCH 299/867] Add Qr Grad for wide matrices

---
 tensorflow/python/kernel_tests/qr_op_test.py |  5 +--
 tensorflow/python/ops/linalg_grad.py         | 47 +++++++++++++++-----
 2 files changed, 38 insertions(+), 14 deletions(-)

diff --git a/tensorflow/python/kernel_tests/qr_op_test.py b/tensorflow/python/kernel_tests/qr_op_test.py
index d5337c183a6..0c291dbd940 100644
--- a/tensorflow/python/kernel_tests/qr_op_test.py
+++ b/tensorflow/python/kernel_tests/qr_op_test.py
@@ -278,14 +278,13 @@ if __name__ == "__main__":
                                     use_static_shape))
 
   # TODO(pfau): Get working with complex types.
-  # TODO(pfau): Get working with full_matrices when rows != cols
-  # TODO(pfau): Get working when rows < cols
+  # TODO(pfau): Get working with full_matrices when rows > cols
   # TODO(pfau): Get working with shapeholders (dynamic shapes)
   for full_matrices in False, True:
     for dtype in np.float32, np.float64:
       for rows in 1, 2, 5, 10:
         for cols in 1, 2, 5, 10:
-          if rows == cols or (not full_matrices and rows > cols):
+          if rows <= cols or (not full_matrices and rows > cols):
             for batch_dims in [(), (3,)] + [(3, 2)] * (max(rows, cols) < 10):
               shape = batch_dims + (rows, cols)
               name = "%s_%s_full_%s" % (dtype.__name__,
diff --git a/tensorflow/python/ops/linalg_grad.py b/tensorflow/python/ops/linalg_grad.py
index 437e28e7e6b..5ec372430ba 100644
--- a/tensorflow/python/ops/linalg_grad.py
+++ b/tensorflow/python/ops/linalg_grad.py
@@ -493,15 +493,10 @@ def _QrGrad(op, dq, dr):
   if (r.shape.ndims is None or r.shape.as_list()[-2] is None or
       r.shape.as_list()[-1] is None):
     raise NotImplementedError("QrGrad not implemented with dynamic shapes.")
-  if r.shape.dims[-2].value != r.shape.dims[-1].value:
+  if (r.shape.dims[-2].value > r.shape.dims[-1].value and
+      q.shape.dims[-2].value == q.shape.dims[-1].value):
     raise NotImplementedError("QrGrad not implemented when ncols > nrows "
-                              "or full_matrices is true and ncols != nrows.")
-
-  qdq = math_ops.matmul(q, dq, adjoint_a=True)
-  qdq_ = qdq - _linalg.adjoint(qdq)
-  rdr = math_ops.matmul(r, dr, adjoint_b=True)
-  rdr_ = rdr - _linalg.adjoint(rdr)
-  tril = array_ops.matrix_band_part(qdq_ + rdr_, -1, 0)
+                              "and full_matrices is true.")
 
   def _TriangularSolve(x, r):
     """Equiv to matmul(x, adjoint(matrix_inverse(r))) if r is upper-tri."""
@@ -509,9 +504,39 @@ def _QrGrad(op, dq, dr):
         linalg_ops.matrix_triangular_solve(
             r, _linalg.adjoint(x), lower=False, adjoint=False))
 
-  grad_a = math_ops.matmul(q, dr + _TriangularSolve(tril, r))
-  grad_b = _TriangularSolve(dq - math_ops.matmul(q, qdq), r)
-  return grad_a + grad_b
+  def _QrGradSquareAndDeepMatrices(q, r, dq, dr):
+    """Gradient for matrix orders num_rows >= num_cols
+    and full_matrices is false.
+    """
+    qdq = math_ops.matmul(q, dq, adjoint_a=True)
+    qdq_ = qdq - _linalg.adjoint(qdq)
+    rdr = math_ops.matmul(r, dr, adjoint_b=True)
+    rdr_ = rdr - _linalg.adjoint(rdr)
+    tril = array_ops.matrix_band_part(qdq_ + rdr_, -1, 0)
+
+    grad_a = math_ops.matmul(q, dr + _TriangularSolve(tril, r))
+    grad_b = _TriangularSolve(dq - math_ops.matmul(q, qdq), r)
+    return grad_a + grad_b
+
+  num_rows, num_cols = q.shape.dims[-2].value, r.shape.dims[-1]
+
+  if num_rows >= num_cols:
+    return _QrGradSquareAndDeepMatrices(q, r, dq, dr)
+
+  # Partition a = [x, y], r = [u, v] and reduce to the square case
+  a = op.inputs[0]
+  y = a[..., :, num_rows:]
+  u = r[..., :, :num_rows]
+  dv = dr[..., :, num_rows:]
+  du = dr[..., :, :num_rows]
+  dy = math_ops.matmul(q, dv)
+  dx = _QrGradSquareAndDeepMatrices(q,
+                                    u,
+                                    dq + math_ops.matmul(y,
+                                                         dv,
+                                                         adjoint_b=True),
+                                    du)
+  return array_ops.concat([dx, dy], axis=-1)
 
 
 @ops.RegisterGradient("MatrixSolve")

From 390b259dfc8407484533c1cb61ca3515ed5166e8 Mon Sep 17 00:00:00 2001
From: Gabriel Rasskin 
Date: Fri, 19 Jun 2020 10:16:47 -0700
Subject: [PATCH 300/867] Fix unused variable issue with fuzzing methods

---
 tensorflow/security/fuzzing/status_group_fuzz.cc | 7 ++++---
 1 file changed, 4 insertions(+), 3 deletions(-)

diff --git a/tensorflow/security/fuzzing/status_group_fuzz.cc b/tensorflow/security/fuzzing/status_group_fuzz.cc
index bc80cd72bc9..989e1c9d1cb 100644
--- a/tensorflow/security/fuzzing/status_group_fuzz.cc
+++ b/tensorflow/security/fuzzing/status_group_fuzz.cc
@@ -54,9 +54,10 @@ extern "C" int LLVMFuzzerTestOneInput(const uint8_t *data, size_t size) {
     }
   }
 
-  sg.as_summary_status();
-  sg.as_concatenated_status();
-  sg.AttachLogMessages();
+  // Ignore warnings that these values are unused
+  sg.as_summary_status().IgnoreError();;
+  sg.as_concatenated_status().IgnoreError();;
+  sg.AttachLogMessages().IgnoreError();;
 
   return 0;
 }

From 40b9713f6459abd043248c55b1b06ebf60712961 Mon Sep 17 00:00:00 2001
From: Gabriel Rasskin 
Date: Fri, 19 Jun 2020 10:17:22 -0700
Subject: [PATCH 301/867] Doubling syntax

---
 tensorflow/security/fuzzing/status_group_fuzz.cc | 6 +++---
 1 file changed, 3 insertions(+), 3 deletions(-)

diff --git a/tensorflow/security/fuzzing/status_group_fuzz.cc b/tensorflow/security/fuzzing/status_group_fuzz.cc
index 989e1c9d1cb..002785734bb 100644
--- a/tensorflow/security/fuzzing/status_group_fuzz.cc
+++ b/tensorflow/security/fuzzing/status_group_fuzz.cc
@@ -55,9 +55,9 @@ extern "C" int LLVMFuzzerTestOneInput(const uint8_t *data, size_t size) {
   }
 
   // Ignore warnings that these values are unused
-  sg.as_summary_status().IgnoreError();;
-  sg.as_concatenated_status().IgnoreError();;
-  sg.AttachLogMessages().IgnoreError();;
+  sg.as_summary_status().IgnoreError();
+  sg.as_concatenated_status().IgnoreError();
+  sg.AttachLogMessages().IgnoreError();
 
   return 0;
 }

From 4d751f9da4914d6e1fc7aafe2f5f27e0f96830d6 Mon Sep 17 00:00:00 2001
From: Katherine Wu 
Date: Fri, 19 Jun 2020 10:12:17 -0700
Subject: [PATCH 302/867] Roll forward of cl/316247127: Make sure compiled
 metrics are accessible after loading from H5 or SavedModel.

PiperOrigin-RevId: 317329395
Change-Id: I33578515f36aa0ba227e75bda52966d493a4bebb
---
 .../python/keras/engine/compile_utils.py      |  14 +--
 tensorflow/python/keras/engine/functional.py  |  10 +-
 tensorflow/python/keras/engine/training.py    |   7 +-
 tensorflow/python/keras/metrics.py            |   5 +-
 tensorflow/python/keras/saving/hdf5_format.py |   1 +
 .../python/keras/saving/hdf5_format_test.py   | 114 +++++++-----------
 .../python/keras/saving/saved_model/load.py   |   1 +
 .../python/keras/saving/saving_utils.py       |  14 +++
 8 files changed, 79 insertions(+), 87 deletions(-)

diff --git a/tensorflow/python/keras/engine/compile_utils.py b/tensorflow/python/keras/engine/compile_utils.py
index 3858639f024..ba7ce624090 100644
--- a/tensorflow/python/keras/engine/compile_utils.py
+++ b/tensorflow/python/keras/engine/compile_utils.py
@@ -37,7 +37,7 @@ class Container(object):
   def __init__(self, output_names=None):
     self._output_names = output_names
 
-  def _build(self, y_pred):
+  def build(self, y_pred):
     if self._output_names is None:
       # In Subclass API, output names like 'output_1' are used for
       # `Metric` names.
@@ -131,9 +131,9 @@ class LossesContainer(Container):
     ]
     return [self._loss_metric] + per_output_metrics
 
-  def _build(self, y_pred):
+  def build(self, y_pred):
     """One-time setup of loss objects."""
-    super(LossesContainer, self)._build(y_pred)
+    super(LossesContainer, self).build(y_pred)
 
     self._losses = self._maybe_broadcast_to_outputs(y_pred, self._losses)
     self._losses = self._conform_to_outputs(y_pred, self._losses)
@@ -184,7 +184,7 @@ class LossesContainer(Container):
     sample_weight = self._conform_to_outputs(y_pred, sample_weight)
 
     if not self._built:
-      self._build(y_pred)
+      self.build(y_pred)
 
     y_pred = nest.flatten(y_pred)
     y_true = nest.flatten(y_true)
@@ -295,9 +295,9 @@ class MetricsContainer(Container):
       return []
     return self._metrics_in_order
 
-  def _build(self, y_pred, y_true):
+  def build(self, y_pred, y_true):
     """One-time setup of metric objects."""
-    super(MetricsContainer, self)._build(y_pred)
+    super(MetricsContainer, self).build(y_pred)
 
     self._metrics = self._maybe_broadcast_to_outputs(y_pred, self._metrics)
     self._metrics = self._conform_to_outputs(y_pred, self._metrics)
@@ -385,7 +385,7 @@ class MetricsContainer(Container):
     sample_weight = self._conform_to_outputs(y_pred, sample_weight)
 
     if not self._built:
-      self._build(y_pred, y_true)
+      self.build(y_pred, y_true)
 
     y_pred = nest.flatten(y_pred)
     y_true = nest.flatten(y_true) if y_true is not None else []
diff --git a/tensorflow/python/keras/engine/functional.py b/tensorflow/python/keras/engine/functional.py
index 0ef4840b651..0612d70044d 100644
--- a/tensorflow/python/keras/engine/functional.py
+++ b/tensorflow/python/keras/engine/functional.py
@@ -1007,10 +1007,12 @@ def _map_subgraph_network(inputs, outputs):
 
 def _should_skip_first_node(layer):
   """Returns True if the first layer node should not be saved or loaded."""
-  # Networks start with a pre-existing node linking their input to output.
-  # For a sequential model, it is first created with _is_graph_network = False,
-  # we have to keep the _is_graph_network check here.
-  return isinstance(layer, Functional) and layer._is_graph_network
+  # Networks that are constructed with an Input layer/shape start with a
+  # pre-existing node linking their input to output. This node is excluded from
+  # the network config.
+  return (isinstance(layer, Functional) and
+          # Filter out Sequential models without an input shape.
+          isinstance(layer._layers[0], input_layer_module.InputLayer))
 
 
 def _deserialize_keras_tensors(kwargs, layer_map):
diff --git a/tensorflow/python/keras/engine/training.py b/tensorflow/python/keras/engine/training.py
index ccd184a8bc4..a0ebec4f95e 100644
--- a/tensorflow/python/keras/engine/training.py
+++ b/tensorflow/python/keras/engine/training.py
@@ -436,7 +436,6 @@ class Model(base_layer.Layer, version_utils.ModelVersionSelector):
                            'Instead, in order to instantiate and build your '
                            'model, `call` your model on real tensor data (of '
                            'the correct dtype).')
-
     super(Model, self).build(input_shape)
 
   def call(self, inputs, training=None, mask=None):
@@ -2417,6 +2416,12 @@ class Model(base_layer.Layer, version_utils.ModelVersionSelector):
 
     self._saved_model_inputs_spec = specs
 
+    # Store the input shapes
+    if (self.__class__.__name__ == 'Sequential' and
+        self._build_input_shape is None):
+      self._build_input_shape = nest.map_structure(
+          lambda x: None if x is None else x.shape, specs)
+
   def _assert_weights_created(self):
     """Asserts that all the weights for the model have been created.
 
diff --git a/tensorflow/python/keras/metrics.py b/tensorflow/python/keras/metrics.py
index a67755b9333..7f40423595b 100644
--- a/tensorflow/python/keras/metrics.py
+++ b/tensorflow/python/keras/metrics.py
@@ -630,10 +630,9 @@ class MeanMetricWrapper(Mean):
   def from_config(cls, config):
     # Note that while MeanMetricWrapper itself isn't public, objects of this
     # class may be created and added to the model by calling model.compile.
+    fn = config.pop('fn', None)
     if cls is MeanMetricWrapper:
-      fn = get(config.pop('fn'))
-      return cls(fn, **config)
-
+      return cls(get(fn), **config)
     return super(MeanMetricWrapper, cls).from_config(config)
 
 
diff --git a/tensorflow/python/keras/saving/hdf5_format.py b/tensorflow/python/keras/saving/hdf5_format.py
index 800d609fe99..01a5e12e4c6 100644
--- a/tensorflow/python/keras/saving/hdf5_format.py
+++ b/tensorflow/python/keras/saving/hdf5_format.py
@@ -192,6 +192,7 @@ def load_model_from_hdf5(filepath, custom_objects=None, compile=True):  # pylint
       # Compile model.
       model.compile(**saving_utils.compile_args_from_training_config(
           training_config, custom_objects))
+      saving_utils.try_build_compiled_arguments(model)
 
       # Set optimizer weights.
       if 'optimizer_weights' in f:
diff --git a/tensorflow/python/keras/saving/hdf5_format_test.py b/tensorflow/python/keras/saving/hdf5_format_test.py
index 757385a25ea..b079bf8cac8 100644
--- a/tensorflow/python/keras/saving/hdf5_format_test.py
+++ b/tensorflow/python/keras/saving/hdf5_format_test.py
@@ -26,10 +26,12 @@ from absl.testing import parameterized
 import numpy as np
 
 from tensorflow.python import keras
+from tensorflow.python import tf2
 from tensorflow.python.eager import context
 from tensorflow.python.framework import constant_op
 from tensorflow.python.framework import dtypes
 from tensorflow.python.framework import ops
+from tensorflow.python.framework import test_util
 from tensorflow.python.keras import combinations
 from tensorflow.python.keras import keras_parameterized
 from tensorflow.python.keras import optimizers
@@ -368,48 +370,54 @@ class TestWeightSavingAndLoading(test.TestCase, parameterized.TestCase):
 
 
 @keras_parameterized.run_with_all_saved_model_formats
-class TestWholeModelSaving(test.TestCase, parameterized.TestCase):
+class TestWholeModelSaving(keras_parameterized.TestCase):
 
   def _save_model_dir(self, dirname='saved_model'):
     temp_dir = self.get_temp_dir()
     self.addCleanup(shutil.rmtree, temp_dir, ignore_errors=True)
     return os.path.join(temp_dir, dirname)
 
-  def _assert_same_weights(self, model, loaded_model,
-                           original_optimizer_has_iterations_variable=True):
-    """Checks that the loaded weighs are the same as the original weights.
+  def _assert_same_weights_and_metrics(self, model, loaded_model):
+    """Checks that the loaded weights and metrics are the same as the original.
 
     Args:
       model: original model
       loaded_model: loaded model
-      original_optimizer_has_iterations_variable: If the original optimizer
-        uses an iterations variable. The loaded model will have a v2
-        optimizer, which always contains an iterations variable. So when
-        comparing the weights, the first variable in the loaded optimizer
-        weights may need to be ignored.
     """
     self.assertAllClose(model.weights, loaded_model.weights)
+
     if loaded_model.optimizer:
       if testing_utils.get_save_format() == 'tf':
         # TODO(b/153110928): Keras TF format doesn't restore optimizer weights
         # currently.
         return
-      if original_optimizer_has_iterations_variable:
-        self.assertAllClose(model.optimizer.weights,
-                            loaded_model.optimizer.weights)
-      else:
-        self.assertAllClose(model.optimizer.weights,
-                            loaded_model.optimizer.weights[1:])
+      self.assertAllClose(model.optimizer.weights,
+                          loaded_model.optimizer.weights)
 
-  def test_sequential_model_saving(self):
+    # In V1/Graph mode, the model isn't built, so the metrics are not loaded
+    # immediately (requires model to be called on some data before building
+    # metrics).
+    check_metrics = tf2.enabled() and context.executing_eagerly()
+
+    if check_metrics:
+      self.assertAllEqual([m.name for m in model.metrics],
+                          [m.name for m in loaded_model.metrics])
+
+  @keras_parameterized.run_with_all_model_types
+  @keras_parameterized.run_all_keras_modes
+  def test_save_and_load(self):
     saved_model_dir = self._save_model_dir()
     save_format = testing_utils.get_save_format()
 
+    if save_format == 'h5' and testing_utils.get_model_type() == 'subclass':
+      return  # HDF5 format currently does not allow saving classed models.
+
     with self.cached_session():
-      model = keras.models.Sequential()
-      model.add(keras.layers.Dense(2, input_shape=(3,)))
-      model.add(keras.layers.RepeatVector(3))
-      model.add(keras.layers.TimeDistributed(keras.layers.Dense(3)))
+      model = testing_utils.get_model_from_layers(
+          [keras.layers.Dense(2),
+           keras.layers.RepeatVector(3),
+           keras.layers.TimeDistributed(keras.layers.Dense(3))],
+          input_shape=(3,))
       model.compile(
           loss=keras.losses.MSE,
           optimizer=keras.optimizer_v2.rmsprop.RMSprop(lr=0.0001),
@@ -432,43 +440,35 @@ class TestWholeModelSaving(test.TestCase, parameterized.TestCase):
       out = model.predict(x)
       keras.models.save_model(model, saved_model_dir, save_format=save_format)
 
-      new_model = keras.models.load_model(saved_model_dir)
-      self._assert_same_weights(model, new_model)
+      loaded_model = keras.models.load_model(saved_model_dir)
+      self._assert_same_weights_and_metrics(model, loaded_model)
 
-      out2 = new_model.predict(x)
+      out2 = loaded_model.predict(x)
       self.assertAllClose(out, out2, atol=1e-05)
 
-      # test that new updates are the same with both models
-      model.train_on_batch(x, y)
-      new_model.train_on_batch(x, y)
-
       eval_out = model.evaluate(x, y)
-      eval_out2 = new_model.evaluate(x, y)
+      eval_out2 = loaded_model.evaluate(x, y)
       self.assertArrayNear(eval_out, eval_out2, 0.001)
 
-      out = model.predict(x)
-      out2 = new_model.predict(x)
-      # The model has been trained on two batches. So the tolerance is larger.
-      self.assertAllClose(out, out2, atol=0.01)
-
+  @test_util.run_in_graph_and_eager_modes
   def test_sequential_model_saving_without_input_shape(self):
     saved_model_dir = self._save_model_dir()
     save_format = testing_utils.get_save_format()
-    with ops.Graph().as_default(), self.cached_session():
+    with self.cached_session():
       model = keras.models.Sequential()
       model.add(keras.layers.Dense(2))
       model.add(keras.layers.RepeatVector(3))
       model.add(keras.layers.TimeDistributed(keras.layers.Dense(3)))
       model.compile(
           loss=keras.losses.MSE,
-          optimizer=keras.optimizers.RMSprop(lr=0.0001),
+          optimizer='rmsprop',
           metrics=[
               keras.metrics.categorical_accuracy,
-              keras.metrics.CategoricalAccuracy()
+              keras.metrics.CategoricalAccuracy(name='cat_acc')
           ],
           weighted_metrics=[
               keras.metrics.categorical_accuracy,
-              keras.metrics.CategoricalAccuracy()
+              keras.metrics.CategoricalAccuracy(name='cat_acc2')
           ],
           sample_weight_mode='temporal')
       x = np.random.random((1, 3))
@@ -479,12 +479,13 @@ class TestWholeModelSaving(test.TestCase, parameterized.TestCase):
       model.save(saved_model_dir, save_format=save_format)
 
       new_model = keras.models.load_model(saved_model_dir)
-      self._assert_same_weights(
-          model, new_model, original_optimizer_has_iterations_variable=False)
+
+      self._assert_same_weights_and_metrics(model, new_model)
 
       out2 = new_model.predict(x)
       self.assertAllClose(out, out2, atol=1e-05)
 
+  @test_util.run_in_graph_and_eager_modes
   def test_sequential_model_saving_without_compile(self):
     saved_model_dir = self._save_model_dir()
     save_format = testing_utils.get_save_format()
@@ -501,7 +502,7 @@ class TestWholeModelSaving(test.TestCase, parameterized.TestCase):
       keras.models.save_model(model, saved_model_dir, save_format=save_format)
 
       new_model = keras.models.load_model(saved_model_dir)
-      self._assert_same_weights(model, new_model)
+      self._assert_same_weights_and_metrics(model, new_model)
 
       out2 = new_model.predict(x)
       self.assertAllClose(out, out2, atol=1e-05)
@@ -535,42 +536,11 @@ class TestWholeModelSaving(test.TestCase, parameterized.TestCase):
           saved_model_dir,
           custom_objects={'CustomOp': CustomOp,
                           'custom_loss': custom_loss})
-      self._assert_same_weights(model, new_model)
+      self._assert_same_weights_and_metrics(model, new_model)
 
       out2 = new_model.predict(x)
       self.assertAllClose(out, out2, atol=1e-05)
 
-  def test_functional_model_saving(self):
-    saved_model_dir = self._save_model_dir()
-    save_format = testing_utils.get_save_format()
-    with ops.Graph().as_default(), self.cached_session():
-      inputs = keras.layers.Input(shape=(3,))
-      x = keras.layers.Dense(2)(inputs)
-      output = keras.layers.Dense(3)(x)
-
-      model = keras.models.Model(inputs, output)
-      model.compile(
-          loss=keras.losses.MSE,
-          optimizer=keras.optimizers.RMSprop(lr=0.0001),
-          metrics=[
-              keras.metrics.categorical_accuracy,
-              keras.metrics.CategoricalAccuracy()
-          ],
-          weighted_metrics=[
-              keras.metrics.categorical_accuracy,
-              keras.metrics.CategoricalAccuracy()
-          ])
-      x = np.random.random((1, 3))
-      y = np.random.random((1, 3))
-      model.train_on_batch(x, y)
-
-      out = model.predict(x)
-      keras.models.save_model(model, saved_model_dir, save_format=save_format)
-      model = keras.models.load_model(saved_model_dir)
-
-      out2 = model.predict(x)
-      self.assertAllClose(out, out2, atol=1e-05)
-
   def test_saving_without_compilation(self):
     saved_model_dir = self._save_model_dir()
     save_format = testing_utils.get_save_format()
diff --git a/tensorflow/python/keras/saving/saved_model/load.py b/tensorflow/python/keras/saving/saved_model/load.py
index a378c1b98e7..0b55e30c27b 100644
--- a/tensorflow/python/keras/saving/saved_model/load.py
+++ b/tensorflow/python/keras/saving/saved_model/load.py
@@ -129,6 +129,7 @@ def load(path, compile=True, options=None):  # pylint: disable=redefined-builtin
     if training_config is not None:
       model.compile(**saving_utils.compile_args_from_training_config(
           training_config))
+      saving_utils.try_build_compiled_arguments(model)
     else:
       logging.warning('No training configuration found in save file, so the '
                       'model was *not* compiled. Compile it manually.')
diff --git a/tensorflow/python/keras/saving/saving_utils.py b/tensorflow/python/keras/saving/saving_utils.py
index 3c9c33531bf..9fdf81cae2a 100644
--- a/tensorflow/python/keras/saving/saving_utils.py
+++ b/tensorflow/python/keras/saving/saving_utils.py
@@ -27,6 +27,7 @@ from tensorflow.python.keras import losses
 from tensorflow.python.keras import optimizers
 from tensorflow.python.keras.engine import base_layer_utils
 from tensorflow.python.keras.utils import generic_utils
+from tensorflow.python.keras.utils import version_utils
 from tensorflow.python.keras.utils.io_utils import ask_to_proceed_with_overwrite
 from tensorflow.python.platform import tf_logging as logging
 from tensorflow.python.util import nest
@@ -307,3 +308,16 @@ def _enforce_names_consistency(specs):
   if name_inconsistency:
     specs = nest.map_structure(_clear_name, specs)
   return specs
+
+
+def try_build_compiled_arguments(model):
+  if (not version_utils.is_v1_layer_or_model(model) and
+      model.outputs is not None):
+    try:
+      model.compiled_loss.build(model.outputs)
+      model.compiled_metrics.build(model.outputs, model.outputs)
+    except:  # pylint: disable=bare-except
+      logging.warning(
+          'Compiled the loaded model, but the compiled metrics have yet to '
+          'be built. `model.compile_metrics` will be empty until you train '
+          'or evaluate the model.')

From f1f5ed68595a56357d92985466d7e0687b23303e Mon Sep 17 00:00:00 2001
From: Mihai Maruseac 
Date: Fri, 19 Jun 2020 17:22:51 +0000
Subject: [PATCH 303/867] Update
 tensorflow/security/fuzzing/status_group_fuzz.cc

---
 tensorflow/security/fuzzing/status_group_fuzz.cc | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/tensorflow/security/fuzzing/status_group_fuzz.cc b/tensorflow/security/fuzzing/status_group_fuzz.cc
index 002785734bb..a560766410a 100644
--- a/tensorflow/security/fuzzing/status_group_fuzz.cc
+++ b/tensorflow/security/fuzzing/status_group_fuzz.cc
@@ -57,7 +57,7 @@ extern "C" int LLVMFuzzerTestOneInput(const uint8_t *data, size_t size) {
   // Ignore warnings that these values are unused
   sg.as_summary_status().IgnoreError();
   sg.as_concatenated_status().IgnoreError();
-  sg.AttachLogMessages().IgnoreError();
+  sg.AttachLogMessages();
 
   return 0;
 }

From d00691f7aaa954bfc9194f62e729f8b41de899af Mon Sep 17 00:00:00 2001
From: Scott Zhu 
Date: Fri, 19 Jun 2020 10:15:24 -0700
Subject: [PATCH 304/867] Fork the keras related mirrored_strategy_test to
 keras/distribute.

PiperOrigin-RevId: 317329998
Change-Id: I7dc55499cb0409129729696c286862a6b6d574aa
---
 tensorflow/python/distribute/BUILD            |  1 -
 .../distribute/mirrored_strategy_test.py      | 47 ----------
 tensorflow/python/keras/distribute/BUILD      | 23 +++++
 .../distribute/mirrored_strategy_test.py      | 89 +++++++++++++++++++
 4 files changed, 112 insertions(+), 48 deletions(-)
 create mode 100644 tensorflow/python/keras/distribute/mirrored_strategy_test.py

diff --git a/tensorflow/python/distribute/BUILD b/tensorflow/python/distribute/BUILD
index 0062705126f..38c5550be16 100644
--- a/tensorflow/python/distribute/BUILD
+++ b/tensorflow/python/distribute/BUILD
@@ -1486,7 +1486,6 @@ cuda_py_test(
         "//tensorflow/python/autograph/core:test_lib",
         "//tensorflow/python/eager:context",
         "//tensorflow/python/eager:test",
-        "//tensorflow/python/keras/layers",
     ],
 )
 
diff --git a/tensorflow/python/distribute/mirrored_strategy_test.py b/tensorflow/python/distribute/mirrored_strategy_test.py
index 950b6f2446b..e6414b2704a 100644
--- a/tensorflow/python/distribute/mirrored_strategy_test.py
+++ b/tensorflow/python/distribute/mirrored_strategy_test.py
@@ -22,7 +22,6 @@ import json
 import sys
 
 from absl.testing import parameterized
-import numpy as np
 
 from tensorflow.core.protobuf import config_pb2
 from tensorflow.python import tf2
@@ -50,16 +49,12 @@ from tensorflow.python.framework import func_graph
 from tensorflow.python.framework import ops
 from tensorflow.python.framework import tensor_shape
 from tensorflow.python.framework import tensor_util
-from tensorflow.python.keras.engine import training as keras_training
-from tensorflow.python.keras.layers import core as keras_core
 from tensorflow.python.ops import array_ops
 from tensorflow.python.ops import control_flow_ops
 from tensorflow.python.ops import gradients
 from tensorflow.python.ops import math_ops
 from tensorflow.python.ops import variable_scope
 from tensorflow.python.ops import variables
-from tensorflow.python.training import gradient_descent
-from tensorflow.python.training import optimizer as optimizer_lib
 from tensorflow.python.training import server_lib
 
 
@@ -988,22 +983,6 @@ class MockModel(object):
     return x
 
 
-class MiniModel(keras_training.Model):
-  """Minimal model for mnist.
-
-  Useful for testing and debugging on slow TPU simulators.
-  """
-
-  def __init__(self):
-    super(MiniModel, self).__init__(name="")
-    self.fc = keras_core.Dense(1, name="fc", kernel_initializer="ones",
-                               bias_initializer="ones")
-
-  def call(self, inputs, training=True):
-    inputs = array_ops.ones([1, 10])
-    return self.fc(inputs)
-
-
 @combinations.generate(
     combinations.combine(
         distribution=[
@@ -1116,32 +1095,6 @@ class MirroredStrategyDefunTest(test.TestCase):
     expected_result = values.PerReplica((5.0 * 1.25, 3.0 * 1.25))
     self._call_and_check(distribution, fn1, [factors], expected_result, [fn1])
 
-  def testTrain(self, distribution):
-    with distribution.scope():
-      mock_model = MiniModel()
-      mock_model.call = function.defun(mock_model.call)
-
-      def loss_fn(ctx):
-        del ctx
-        return mock_model(array_ops.ones([1, 10]))
-
-      gradients_fn = backprop.implicit_grad(loss_fn)
-      gradients_fn = optimizer_lib.get_filtered_grad_fn(gradients_fn)
-      grads_and_vars = distribution.extended.call_for_each_replica(
-          gradients_fn, args=(None,))
-
-      optimizer = gradient_descent.GradientDescentOptimizer(0.25)
-      update_ops = optimizer._distributed_apply(distribution, grads_and_vars)  # pylint: disable=protected-access
-
-      if not context.executing_eagerly():
-        self.evaluate(variables.global_variables_initializer())
-        self.evaluate(update_ops)
-
-      updated_var_values = self.evaluate(mock_model.variables)
-      # All variables start at 1.0 and get two updates of 0.25.
-      self.assertAllEqual(0.5 * np.ones([10, 1]), updated_var_values[0])
-      self.assertAllEqual([0.5], updated_var_values[1])
-
 
 @combinations.generate(
     combinations.combine(
diff --git a/tensorflow/python/keras/distribute/BUILD b/tensorflow/python/keras/distribute/BUILD
index ddf274f299f..247e655621c 100644
--- a/tensorflow/python/keras/distribute/BUILD
+++ b/tensorflow/python/keras/distribute/BUILD
@@ -324,6 +324,29 @@ cuda_py_test(
     ],
 )
 
+cuda_py_test(
+    name = "mirrored_strategy_test",
+    srcs = ["mirrored_strategy_test.py"],
+    python_version = "PY3",
+    tags = [
+        "multi_and_single_gpu",
+        "no_windows_gpu",  # TODO(b/130551176)
+    ],
+    deps = [
+        "//tensorflow/python:array_ops",
+        "//tensorflow/python:training_lib",
+        "//tensorflow/python:variables",
+        "//tensorflow/python/distribute:combinations",
+        "//tensorflow/python/distribute:strategy_combinations",
+        "//tensorflow/python/eager:backprop",
+        "//tensorflow/python/eager:context",
+        "//tensorflow/python/eager:function",
+        "//tensorflow/python/eager:test",
+        "//tensorflow/python/keras/engine",
+        "//tensorflow/python/keras/layers:core",
+    ],
+)
+
 cuda_py_test(
     name = "multi_worker_test",
     srcs = ["multi_worker_test.py"],
diff --git a/tensorflow/python/keras/distribute/mirrored_strategy_test.py b/tensorflow/python/keras/distribute/mirrored_strategy_test.py
new file mode 100644
index 00000000000..2844af8cc3a
--- /dev/null
+++ b/tensorflow/python/keras/distribute/mirrored_strategy_test.py
@@ -0,0 +1,89 @@
+# Copyright 2018 The TensorFlow Authors. All Rights Reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+# ==============================================================================
+"""Tests for MirroredStrategy."""
+
+from __future__ import absolute_import
+from __future__ import division
+from __future__ import print_function
+
+import numpy as np
+
+from tensorflow.python.distribute import combinations
+from tensorflow.python.distribute import strategy_combinations
+from tensorflow.python.eager import backprop
+from tensorflow.python.eager import context
+from tensorflow.python.eager import function
+from tensorflow.python.eager import test
+from tensorflow.python.keras.engine import training as keras_training
+from tensorflow.python.keras.layers import core as keras_core
+from tensorflow.python.ops import array_ops
+from tensorflow.python.ops import variables
+from tensorflow.python.training import gradient_descent
+from tensorflow.python.training import optimizer as optimizer_lib
+
+
+class MiniModel(keras_training.Model):
+  """Minimal model for mnist.
+
+  Useful for testing and debugging on slow TPU simulators.
+  """
+
+  def __init__(self):
+    super(MiniModel, self).__init__(name="")
+    self.fc = keras_core.Dense(1, name="fc", kernel_initializer="ones",
+                               bias_initializer="ones")
+
+  def call(self, inputs, training=True):
+    inputs = array_ops.ones([1, 10])
+    return self.fc(inputs)
+
+
+@combinations.generate(
+    combinations.combine(
+        distribution=[
+            strategy_combinations.mirrored_strategy_with_gpu_and_cpu,
+        ],
+        mode=["graph", "eager"]))
+class MirroredStrategyDefunTest(test.TestCase):
+
+  def testTrain(self, distribution):
+    with distribution.scope():
+      mock_model = MiniModel()
+      mock_model.call = function.defun(mock_model.call)
+
+      def loss_fn(ctx):
+        del ctx
+        return mock_model(array_ops.ones([1, 10]))
+
+      gradients_fn = backprop.implicit_grad(loss_fn)
+      gradients_fn = optimizer_lib.get_filtered_grad_fn(gradients_fn)
+      grads_and_vars = distribution.extended.call_for_each_replica(
+          gradients_fn, args=(None,))
+
+      optimizer = gradient_descent.GradientDescentOptimizer(0.25)
+      update_ops = optimizer._distributed_apply(distribution, grads_and_vars)  # pylint: disable=protected-access
+
+      if not context.executing_eagerly():
+        self.evaluate(variables.global_variables_initializer())
+        self.evaluate(update_ops)
+
+      updated_var_values = self.evaluate(mock_model.variables)
+      # All variables start at 1.0 and get two updates of 0.25.
+      self.assertAllEqual(0.5 * np.ones([10, 1]), updated_var_values[0])
+      self.assertAllEqual([0.5], updated_var_values[1])
+
+
+if __name__ == "__main__":
+  test.main()

From 57f9d638c00083e864aaa7e4f8114c0dd3ba479c Mon Sep 17 00:00:00 2001
From: Bixia Zheng 
Date: Fri, 19 Jun 2020 10:17:18 -0700
Subject: [PATCH 305/867] [TF:TRT] Add a prefix to the warning messages from
 TF-TRT.

Add LOG_WARNING_WITH_PREFIX to common/utils.h. Replace the use of LOG(WARNING)
with this new macro.

PiperOrigin-RevId: 317330336
Change-Id: Ife0aa0347dd72f6eb0f8805af4d46a7d4cb099ea
---
 tensorflow/compiler/tf2tensorrt/BUILD         | 14 ++++++++
 .../compiler/tf2tensorrt/common/utils.h       | 35 +++++++++++++++++++
 .../tf2tensorrt/convert/convert_graph.cc      | 30 +++++++++-------
 .../tf2tensorrt/convert/convert_nodes.cc      | 27 +++++++-------
 .../tf2tensorrt/kernels/trt_engine_op.cc      | 25 +++++++------
 .../compiler/tf2tensorrt/segment/segment.cc   | 10 +++---
 .../compiler/tf2tensorrt/utils/py_utils.cc    |  8 +++--
 .../compiler/tf2tensorrt/utils/trt_logger.cc  |  3 +-
 8 files changed, 108 insertions(+), 44 deletions(-)
 create mode 100644 tensorflow/compiler/tf2tensorrt/common/utils.h

diff --git a/tensorflow/compiler/tf2tensorrt/BUILD b/tensorflow/compiler/tf2tensorrt/BUILD
index 4a8599e29f6..368cb5af2ed 100644
--- a/tensorflow/compiler/tf2tensorrt/BUILD
+++ b/tensorflow/compiler/tf2tensorrt/BUILD
@@ -79,6 +79,15 @@ tf_cuda_cc_test(
     ]),
 )
 
+cc_library(
+    name = "common_utils",
+    hdrs = ["common/utils.h"],
+    copts = tf_copts(),
+    deps = [
+        "//tensorflow/core/platform:logging",
+    ] + if_tensorrt([":tensorrt_lib"]),
+)
+
 cc_library(
     name = "trt_op_kernels",
     srcs = [
@@ -95,6 +104,7 @@ cc_library(
         ":trt_plugins",
         ":trt_resources",
         ":utils",
+        ":common_utils",
         "@com_google_absl//absl/memory",
         "@com_google_absl//absl/strings",
         "@local_config_cuda//cuda:cuda_headers",
@@ -240,6 +250,7 @@ tf_cuda_library(
     hdrs = ["utils/trt_logger.h"],
     visibility = ["//visibility:public"],
     deps = [
+        ":common_utils",
         ":logger_registry",
         "//tensorflow/core:lib_proto_parsing",
     ] + if_tensorrt([":tensorrt_lib"]),
@@ -375,6 +386,7 @@ tf_cuda_library(
         "convert/trt_optimization_pass.h",
     ],
     deps = [
+        ":common_utils",
         ":logger_registry",
         ":segment",
         ":trt_allocator",
@@ -488,6 +500,7 @@ cc_library(
     ],
     copts = tf_copts(),
     deps = [
+        ":common_utils",
         "//tensorflow/core:graph",
         "//tensorflow/core:lib",
         "//tensorflow/core:lib_internal",
@@ -575,6 +588,7 @@ cc_library(
     hdrs = ["utils/py_utils.h"],
     copts = tf_copts(),
     deps = if_tensorrt([
+        ":common_utils",
         ":tensorrt_lib",
         "//tensorflow/stream_executor/platform:dso_loader",
     ]),
diff --git a/tensorflow/compiler/tf2tensorrt/common/utils.h b/tensorflow/compiler/tf2tensorrt/common/utils.h
new file mode 100644
index 00000000000..9ab0145e1ec
--- /dev/null
+++ b/tensorflow/compiler/tf2tensorrt/common/utils.h
@@ -0,0 +1,35 @@
+/* Copyright 2020 The TensorFlow Authors. All Rights Reserved.
+
+Licensed under the Apache License, Version 2.0 (the "License");
+you may not use this file except in compliance with the License.
+You may obtain a copy of the License at
+
+    http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software
+distributed under the License is distributed on an "AS IS" BASIS,
+WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+See the License for the specific language governing permissions and
+limitations under the License.
+==============================================================================*/
+
+#ifndef TENSORFLOW_COMPILER_TF2TENSORRT_COMMON_UTILS_H_
+#define TENSORFLOW_COMPILER_TF2TENSORRT_COMMON_UTILS_H_
+
+#if GOOGLE_CUDA
+#if GOOGLE_TENSORRT
+
+#include "tensorflow/core/platform/logging.h"
+
+namespace tensorflow {
+namespace tensorrt {
+
+#define LOG_WARNING_WITH_PREFIX LOG(WARNING) << "TF-TRT Warning: "
+
+}  // namespace tensorrt
+}  // namespace tensorflow
+
+#endif
+#endif
+
+#endif  // TENSORFLOW_COMPILER_TF2TENSORRT_COMMON_UTILS_H_
diff --git a/tensorflow/compiler/tf2tensorrt/convert/convert_graph.cc b/tensorflow/compiler/tf2tensorrt/convert/convert_graph.cc
index 414d27477bc..1c51d51f1c9 100644
--- a/tensorflow/compiler/tf2tensorrt/convert/convert_graph.cc
+++ b/tensorflow/compiler/tf2tensorrt/convert/convert_graph.cc
@@ -25,6 +25,7 @@ limitations under the License.
 #include 
 
 #include "absl/strings/str_cat.h"
+#include "tensorflow/compiler/tf2tensorrt/common/utils.h"
 #include "tensorflow/compiler/tf2tensorrt/convert/convert_nodes.h"
 #include "tensorflow/compiler/tf2tensorrt/convert/logger_registry.h"
 #include "tensorflow/compiler/tf2tensorrt/convert/utils.h"
@@ -276,8 +277,9 @@ Status GetEngineInfo(const Graph* g,
   if (segment_devices.size() == 1) {
     info->device = *segment_devices.begin();
   } else if (segment_devices.size() > 1) {
-    LOG(WARNING) << "Detected multiple (" << segment_devices.size()
-                 << ") devices for the segment. Picking first one to continue.";
+    LOG_WARNING_WITH_PREFIX
+        << "Detected multiple (" << segment_devices.size()
+        << ") devices for the segment. Picking first one to continue.";
     info->device = *segment_devices.begin();
   } else {
     TfGpuId tf_gpu_id;
@@ -663,7 +665,7 @@ std::pair GetDeviceAndAllocator(const ConversionParams& params,
       StrAppend(&msg, engine.device, "': ");
       for (auto d : devices) StrAppend(&msg, d->name(), ", ");
       StrAppend(&msg, ". Will get the allocator from first one.");
-      LOG(WARNING) << msg;
+      LOG_WARNING_WITH_PREFIX << msg;
     }
     AllocatorAttributes alloc_attr;
     cuda_device_id = devices[0]->tensorflow_gpu_device_info()->gpu_id;
@@ -671,8 +673,8 @@ std::pair GetDeviceAndAllocator(const ConversionParams& params,
     VLOG(1) << "Using allocator " << dev_allocator->Name()
             << " and cuda_device_id " << cuda_device_id;
   } else {
-    LOG(WARNING) << "Cluster is set but device '" << engine.device
-                 << "' is not found in the cluster";
+    LOG_WARNING_WITH_PREFIX << "Cluster is set but device '" << engine.device
+                            << "' is not found in the cluster";
   }
   return std::make_pair(cuda_device_id, dev_allocator);
 }
@@ -770,8 +772,8 @@ Status ConvertAfterShapes(const ConversionParams& params) {
     Status status = GetEngineInfo(&graph, static_graph_properties, curr_segment,
                                   node_map, reverse_topo_order, &curr_engine);
     if (!status.ok()) {
-      LOG(WARNING) << "Failed to get engine info for segment " << t << ": "
-                   << status;
+      LOG_WARNING_WITH_PREFIX << "Failed to get engine info for segment " << t
+                              << ": " << status;
       continue;
     }
     curr_engine.precision_mode = params.precision_mode;
@@ -784,8 +786,9 @@ Status ConvertAfterShapes(const ConversionParams& params) {
                                             &graph, curr_engine.engine_name);
 
     if (!status.ok()) {
-      LOG(WARNING) << "Failed to register segment graphdef to the library " << t
-                   << ": " << status;
+      LOG_WARNING_WITH_PREFIX
+          << "Failed to register segment graphdef to the library " << t << ": "
+          << status;
       continue;
     }
 
@@ -836,7 +839,8 @@ Status ConvertAfterShapes(const ConversionParams& params) {
       alloc.reset(new TRTDeviceAllocator(device_alloc.second));
     } else {
       // Setting allocator as nullptr should get revert to the cudamalloc
-      LOG(WARNING) << "Can't identify the cuda device. Running on device 0 ";
+      LOG_WARNING_WITH_PREFIX
+          << "Can't identify the cuda device. Running on device 0 ";
     }
     cudaSetDevice(cuda_device_id);
     auto status =
@@ -850,9 +854,9 @@ Status ConvertAfterShapes(const ConversionParams& params) {
       LOG(INFO) << "Replaced " << msg << ".";
     } else {
       // Graph is not modified.
-      LOG(WARNING) << "Cannot replace " << msg
-                   << " reason: " << status.error_message()
-                   << " (keeping original segment).";
+      LOG_WARNING_WITH_PREFIX << "Cannot replace " << msg
+                              << " reason: " << status.error_message()
+                              << " (keeping original segment).";
     }
     if (VLOG_IS_ON(1)) {
       msg = "Segment consists of nodes: ";
diff --git a/tensorflow/compiler/tf2tensorrt/convert/convert_nodes.cc b/tensorflow/compiler/tf2tensorrt/convert/convert_nodes.cc
index 28b27959afc..96cec556942 100644
--- a/tensorflow/compiler/tf2tensorrt/convert/convert_nodes.cc
+++ b/tensorflow/compiler/tf2tensorrt/convert/convert_nodes.cc
@@ -31,6 +31,7 @@ limitations under the License.
 #include "absl/strings/str_cat.h"
 #include "absl/strings/str_format.h"
 #include "absl/strings/string_view.h"
+#include "tensorflow/compiler/tf2tensorrt/common/utils.h"
 #include "tensorflow/compiler/tf2tensorrt/convert/utils.h"
 #include "tensorflow/compiler/tf2tensorrt/utils/trt_logger.h"
 #include "tensorflow/compiler/tf2tensorrt/utils/trt_shape_optimization_profiles.h"
@@ -1214,15 +1215,16 @@ static void InitializeTrtPlugins(nvinfer1::ILogger* trt_logger) {
   nvinfer1::IPluginCreator* const* trt_plugin_creator_list =
       getPluginRegistry()->getPluginCreatorList(&num_trt_plugins);
   if (!trt_plugin_creator_list) {
-    LOG(WARNING) << "Can not find any TensorRT plugins in registry.";
+    LOG_WARNING_WITH_PREFIX << "Can not find any TensorRT plugins in registry.";
   } else {
     VLOG(1) << "Found the following " << num_trt_plugins
             << " TensorRT plugins in registry:";
     for (int i = 0; i < num_trt_plugins; ++i) {
       if (!trt_plugin_creator_list[i]) {
-        LOG(WARNING) << "TensorRT plugin at index " << i
-                     << " is not accessible (null pointer returned by "
-                        "getPluginCreatorList for this plugin)";
+        LOG_WARNING_WITH_PREFIX
+            << "TensorRT plugin at index " << i
+            << " is not accessible (null pointer returned by "
+               "getPluginCreatorList for this plugin)";
       } else {
         VLOG(1) << "  " << trt_plugin_creator_list[i]->getPluginName();
       }
@@ -1827,9 +1829,9 @@ void Converter::MaybeApplyQuantizationRanges() {
       // are tensors which are created internally by TF-TRT. The ranges for
       // these unnamed ITensors are always inferred from user provided ranges,
       // thus there will also be a warning for the range(s) the user missed.
-      LOG(WARNING) << "Quantization range was not found for "
-                   << tensor->getName() << ". "
-                   << "Setting invalid quantization range.";
+      LOG_WARNING_WITH_PREFIX << "Quantization range was not found for "
+                              << tensor->getName() << ". "
+                              << "Setting invalid quantization range.";
       // Set the range to something unusable so the engine will fail if it
       // tries to actually use the tensor's range.
       tensor->setDynamicRange(0, 0);
@@ -4898,10 +4900,11 @@ Status ConvertFusedBatchNorm(OpConverterParams* params) {
     // Trying to use batchnorm in training mode is a very common problem.
     // Because the error message will only be printed in VLOG(1) by the
     // segmenter, we issue a special warning so that users will actually see it.
-    LOG(WARNING) << node_def.op() << " only supports is_training=false. If you "
-                 << "are using Keras, please call "
-                 << "keras.backend.set_learning_phase(0) before constructing "
-                 << "your model. At " << node_def.name();
+    LOG_WARNING_WITH_PREFIX
+        << node_def.op() << " only supports is_training=false. If you "
+        << "are using Keras, please call "
+        << "keras.backend.set_learning_phase(0) before constructing "
+        << "your model. At " << node_def.name();
     return errors::Unimplemented(node_def.op(),
                                  " only supports is_training=false, at ",
                                  node_def.name());
@@ -6039,7 +6042,7 @@ Status ConvertGraphDefToEngine(
         const string error_message =
             StrCat("Validation failed for ", node_name, " and input slot ",
                    slot_number, ": ", status.error_message());
-        LOG(WARNING) << error_message;
+        LOG_WARNING_WITH_PREFIX << error_message;
         return Status(status.code(), error_message);
       }
       VLOG(2) << "Adding engine input tensor " << node_name << " with shape "
diff --git a/tensorflow/compiler/tf2tensorrt/kernels/trt_engine_op.cc b/tensorflow/compiler/tf2tensorrt/kernels/trt_engine_op.cc
index d9b8e198f4f..ac4a331041d 100644
--- a/tensorflow/compiler/tf2tensorrt/kernels/trt_engine_op.cc
+++ b/tensorflow/compiler/tf2tensorrt/kernels/trt_engine_op.cc
@@ -20,6 +20,7 @@ limitations under the License.
 #include "absl/strings/ascii.h"
 #include "absl/strings/str_cat.h"
 #include "absl/strings/string_view.h"
+#include "tensorflow/compiler/tf2tensorrt/common/utils.h"
 #include "tensorflow/compiler/tf2tensorrt/convert/convert_nodes.h"
 #include "tensorflow/compiler/tf2tensorrt/convert/utils.h"
 #include "tensorflow/compiler/tf2tensorrt/utils/trt_allocator.h"
@@ -613,8 +614,8 @@ void TRTEngineOp::ComputeAsync(OpKernelContext* ctx,
   }
   Status stat = ExecuteTrtEngine(ctx, engine_context, trt_context_idx);
   if (!stat.ok()) {
-    LOG(WARNING) << "Failed to execute engine: " << stat
-                 << " Retrying with native segment for " << name();
+    LOG_WARNING_WITH_PREFIX << "Failed to execute engine: " << stat
+                            << " Retrying with native segment for " << name();
     // Release any outputs that are allocated, ExecuteNativeSegment will
     // re-allocate them and fail if they are currently allocated.
     for (int i = 0; i < ctx->num_outputs(); i++) {
@@ -727,9 +728,9 @@ StatusOr> TRTEngineOp::BuildEngine(
       calibrator, &engine, use_calibration, use_implicit_batch_, nullptr,
       &cache_resource->profiles_);
   if (!status.ok()) {
-    LOG(WARNING) << "Engine creation for " << name() << " failed. "
-                 << "The native segment will be used instead. "
-                 << "Reason: " << status;
+    LOG_WARNING_WITH_PREFIX << "Engine creation for " << name() << " failed. "
+                            << "The native segment will be used instead. "
+                            << "Reason: " << status;
     // Store an empty engine in the cache for these input shapes so we don't try
     // to build the same failing engine again.
     cache_resource->cache_.emplace(input_concrete_shapes,
@@ -791,8 +792,9 @@ StatusOr> TRTEngineOp::GetEngine(
               FunctionDefToGraphDef(func_handle_, lib, &segment_graph_def_);
         }
         if (!status.ok()) {
-          LOG(WARNING) << "Getting segment graph for " << name() << " failed. "
-                       << "Reason: " << status;
+          LOG_WARNING_WITH_PREFIX << "Getting segment graph for " << name()
+                                  << " failed. "
+                                  << "Reason: " << status;
         }
       }
       auto result = BuildEngine(input_concrete_shapes, batch_size,
@@ -851,10 +853,11 @@ StatusOr> TRTEngineOp::GetEngine(
   // If cache does not have a compatible engine then create a new engine.
   if (engine_contexts == nullptr) {
     if (!allow_build_at_runtime_) {
-      LOG(WARNING) << "Found no engine in cache matching input shapes. "
-                   << "Not building a new engine because "
-                   << "allow_build_at_runtime=False. "
-                   << "The native segment will be used instead.";
+      LOG_WARNING_WITH_PREFIX
+          << "Found no engine in cache matching input shapes. "
+          << "Not building a new engine because "
+          << "allow_build_at_runtime=False. "
+          << "The native segment will be used instead.";
       // Store an empty engine in the cache for these input shapes so we don't
       // try to build the same failing engine again.
       cache.emplace(input_concrete_shapes, absl::make_unique());
diff --git a/tensorflow/compiler/tf2tensorrt/segment/segment.cc b/tensorflow/compiler/tf2tensorrt/segment/segment.cc
index 749335f1b09..32e30006f58 100644
--- a/tensorflow/compiler/tf2tensorrt/segment/segment.cc
+++ b/tensorflow/compiler/tf2tensorrt/segment/segment.cc
@@ -22,6 +22,7 @@ limitations under the License.
 
 #include "absl/strings/str_cat.h"
 #include "absl/strings/str_format.h"
+#include "tensorflow/compiler/tf2tensorrt/common/utils.h"
 #include "tensorflow/compiler/tf2tensorrt/segment/union_find.h"
 #include "tensorflow/core/common_runtime/graph_constructor.h"
 #include "tensorflow/core/graph/algorithm.h"
@@ -748,9 +749,10 @@ Status SegmentGraph(const Graph* tf_graph,
         exclude_node(status.error_message());
       } else if (tftrt_op_blacklist.count(node->tf_node()->type_string())) {
         // WARNING verbosity since the user explicitly requests this behavior.
-        LOG(WARNING) << "Blacklisted as TF-TRT candidate, "
-                     << "(Op type: " << node->tf_node()->type_string() << "), "
-                     << "(Op name: " << node->name() << ")";
+        LOG_WARNING_WITH_PREFIX
+            << "Blacklisted as TF-TRT candidate, "
+            << "(Op type: " << node->tf_node()->type_string() << "), "
+            << "(Op name: " << node->name() << ")";
         exclude_node("Blacklisted with the env var TF_TRT_OP_BLACKLIST");
       } else {
         VLOG(2) << "Accepted as a TF-TRT candidate, "
@@ -1038,7 +1040,7 @@ Status SegmentGraph(const Graph* tf_graph,
       for (const auto& dev : dev_itr->second) {
         StrAppend(&s, dev, ", ");
       }
-      LOG(WARNING) << s;
+      LOG_WARNING_WITH_PREFIX << s;
     }
 
     segments->emplace_back(segment_nodes);
diff --git a/tensorflow/compiler/tf2tensorrt/utils/py_utils.cc b/tensorflow/compiler/tf2tensorrt/utils/py_utils.cc
index 885f58cd70c..a8e24aa8983 100644
--- a/tensorflow/compiler/tf2tensorrt/utils/py_utils.cc
+++ b/tensorflow/compiler/tf2tensorrt/utils/py_utils.cc
@@ -16,6 +16,7 @@ limitations under the License.
 #include "tensorflow/compiler/tf2tensorrt/utils/py_utils.h"
 
 #if GOOGLE_CUDA && GOOGLE_TENSORRT
+#include "tensorflow/compiler/tf2tensorrt/common/utils.h"
 #include "tensorflow/stream_executor/platform/dso_loader.h"
 #include "third_party/tensorrt/NvInfer.h"
 #endif
@@ -27,9 +28,10 @@ bool IsGoogleTensorRTEnabled() {
 #if GOOGLE_CUDA && GOOGLE_TENSORRT
   auto handle_or = se::internal::DsoLoader::TryDlopenTensorRTLibraries();
   if (!handle_or.ok()) {
-    LOG(WARNING) << "Cannot dlopen some TensorRT libraries. If you would like "
-                    "to use Nvidia GPU with TensorRT, please make sure the "
-                    "missing libraries mentioned above are installed properly.";
+    LOG_WARNING_WITH_PREFIX
+        << "Cannot dlopen some TensorRT libraries. If you would like "
+           "to use Nvidia GPU with TensorRT, please make sure the "
+           "missing libraries mentioned above are installed properly.";
     return false;
   } else {
     return true;
diff --git a/tensorflow/compiler/tf2tensorrt/utils/trt_logger.cc b/tensorflow/compiler/tf2tensorrt/utils/trt_logger.cc
index 6bb6f1f9dd8..193687ebc8c 100644
--- a/tensorflow/compiler/tf2tensorrt/utils/trt_logger.cc
+++ b/tensorflow/compiler/tf2tensorrt/utils/trt_logger.cc
@@ -17,6 +17,7 @@ limitations under the License.
 
 #if GOOGLE_CUDA
 #if GOOGLE_TENSORRT
+#include "tensorflow/compiler/tf2tensorrt/common/utils.h"
 #include "tensorflow/compiler/tf2tensorrt/convert/logger_registry.h"
 #include "tensorflow/core/platform/logging.h"
 
@@ -35,7 +36,7 @@ void Logger::log(Severity severity, const char* msg) {
       break;
     }
     case Severity::kWARNING: {
-      LOG(WARNING) << name_ << " " << msg;
+      LOG_WARNING_WITH_PREFIX << name_ << " " << msg;
       break;
     }
     case Severity::kERROR: {

From ef1cabc7a8fb05f4c66b21aa2409a74e22635e45 Mon Sep 17 00:00:00 2001
From: Thomas Joerg 
Date: Fri, 19 Jun 2020 10:18:32 -0700
Subject: [PATCH 306/867] [XLA:GPU] Split reduce ops with large but
 non-consecutive reduction dimensions.

PiperOrigin-RevId: 317330616
Change-Id: Icdcf320b233479c2f74c5b40ee4c8d9a73a6088a
---
 tensorflow/compiler/xla/service/gpu/BUILD     |  28 ++++
 .../compiler/xla/service/gpu/gpu_compiler.cc  |   2 +
 .../xla/service/gpu/reduction_splitter.cc     | 117 +++++++++++++++
 .../xla/service/gpu/reduction_splitter.h      |  49 ++++++
 .../service/gpu/reduction_splitter_test.cc    | 140 ++++++++++++++++++
 .../reduction_degenerate_dim_remover_test.cc  |   1 +
 .../tests/reduction_layout_normalizer_test.cc |   1 +
 7 files changed, 338 insertions(+)
 create mode 100644 tensorflow/compiler/xla/service/gpu/reduction_splitter.cc
 create mode 100644 tensorflow/compiler/xla/service/gpu/reduction_splitter.h
 create mode 100644 tensorflow/compiler/xla/service/gpu/reduction_splitter_test.cc

diff --git a/tensorflow/compiler/xla/service/gpu/BUILD b/tensorflow/compiler/xla/service/gpu/BUILD
index 0eb82128159..472d2117a2c 100644
--- a/tensorflow/compiler/xla/service/gpu/BUILD
+++ b/tensorflow/compiler/xla/service/gpu/BUILD
@@ -1174,6 +1174,7 @@ cc_library(
         ":reduction_degenerate_dim_remover",
         ":reduction_dimension_grouper",
         ":reduction_layout_normalizer",
+        ":reduction_splitter",
         ":stream_assignment",
         ":stream_executor_util",
         ":target_constants",
@@ -1819,6 +1820,33 @@ cc_library(
     ],
 )
 
+cc_library(
+    name = "reduction_splitter",
+    srcs = ["reduction_splitter.cc"],
+    hdrs = ["reduction_splitter.h"],
+    deps = [
+        ":ir_emission_utils",
+        "//tensorflow/compiler/xla:shape_util",
+        "//tensorflow/compiler/xla/service:hlo",
+        "//tensorflow/compiler/xla/service:hlo_pass",
+    ],
+)
+
+tf_cc_test(
+    name = "reduction_splitter_test",
+    srcs = ["reduction_splitter_test.cc"],
+    deps = [
+        ":reduction_splitter",
+        "//tensorflow/compiler/xla:shape_util",
+        "//tensorflow/compiler/xla:test",
+        "//tensorflow/compiler/xla:test_helpers",
+        "//tensorflow/compiler/xla/service:hlo_matchers",
+        "//tensorflow/compiler/xla/service:hlo_parser",
+        "//tensorflow/compiler/xla/tests:hlo_test_base",
+        "//tensorflow/compiler/xla/tests:xla_internal_test_main",
+    ],
+)
+
 cc_library(
     name = "reduction_layout_normalizer",
     srcs = ["reduction_layout_normalizer.cc"],
diff --git a/tensorflow/compiler/xla/service/gpu/gpu_compiler.cc b/tensorflow/compiler/xla/service/gpu/gpu_compiler.cc
index cddbee92874..156cb112285 100644
--- a/tensorflow/compiler/xla/service/gpu/gpu_compiler.cc
+++ b/tensorflow/compiler/xla/service/gpu/gpu_compiler.cc
@@ -65,6 +65,7 @@ limitations under the License.
 #include "tensorflow/compiler/xla/service/gpu/reduction_degenerate_dim_remover.h"
 #include "tensorflow/compiler/xla/service/gpu/reduction_dimension_grouper.h"
 #include "tensorflow/compiler/xla/service/gpu/reduction_layout_normalizer.h"
+#include "tensorflow/compiler/xla/service/gpu/reduction_splitter.h"
 #include "tensorflow/compiler/xla/service/gpu/stream_assignment.h"
 #include "tensorflow/compiler/xla/service/gpu/stream_executor_util.h"
 #include "tensorflow/compiler/xla/service/gpu/target_constants.h"
@@ -371,6 +372,7 @@ Status GpuCompiler::OptimizeHloPostLayoutAssignment(
   pipeline.AddPass();
   pipeline.AddPass();
   pipeline.AddPass();
+  pipeline.AddPass>();
 
   // The LayoutAssignment pass may leave behind kCopy instructions which are
   // duplicate or NOPs, so remove them with algebraic simplification and CSE.
diff --git a/tensorflow/compiler/xla/service/gpu/reduction_splitter.cc b/tensorflow/compiler/xla/service/gpu/reduction_splitter.cc
new file mode 100644
index 00000000000..b68213ec35f
--- /dev/null
+++ b/tensorflow/compiler/xla/service/gpu/reduction_splitter.cc
@@ -0,0 +1,117 @@
+/* Copyright 2019 The TensorFlow Authors. All Rights Reserved.
+
+Licensed under the Apache License, Version 2.0 (the "License");
+you may not use this file except in compliance with the License.
+You may obtain a copy of the License at
+
+    http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software
+distributed under the License is distributed on an "AS IS" BASIS,
+WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+See the License for the specific language governing permissions and
+limitations under the License.
+==============================================================================*/
+
+#include "tensorflow/compiler/xla/service/gpu/reduction_splitter.h"
+
+#include 
+
+#include "tensorflow/compiler/xla/service/dfs_hlo_visitor_with_default.h"
+#include "tensorflow/compiler/xla/service/gpu/ir_emission_utils.h"
+#include "tensorflow/compiler/xla/service/hlo_instructions.h"
+#include "tensorflow/compiler/xla/shape_util.h"
+
+namespace xla {
+namespace gpu {
+
+class ReductionSplitterVisitor : public DfsHloRewriteVisitor {
+ public:
+  Status HandleReduce(HloInstruction *reduce) override {
+    VLOG(4) << "Input: " << reduce->ToString();
+
+    // Reductions with contiguous dimensions are lowered to efficient code. No
+    // need to split such ops.
+    if (IsReductionFromOrToContiguousDimensions(*reduce)) {
+      return Status::OK();
+    }
+    if (reduce->dimensions().size() < 2) {
+      return Status::OK();
+    }
+    if (!reduce->shape().IsArray()) {
+      // TODO(cheshire): Handle variadic reduction.
+      return Status::OK();
+    }
+
+    HloInstruction *operand = reduce->mutable_operand(0);
+    const Shape &shape = operand->shape();
+    CHECK(shape == LayoutUtil::GetWithDefaultLayout(shape))
+        << "Default layout should be enforced on reduction operand";
+    // Verify that contiguous dimensions have been grouped by the
+    // ReductionDimensionGrouper pass.
+    for (int64 i = 0; i < reduce->dimensions().size(); ++i) {
+      for (int64 j = i + 1; j < reduce->dimensions().size(); ++j) {
+        CHECK(abs(reduce->dimensions(i) - reduce->dimensions(j)) > 1)
+            << "Reduction dimensions must not be consecutive";
+      }
+    }
+
+    // The reduce op has non-contiguous dimensions. Look for the dimension with
+    // the largest shape dimension. Reducing along this dimension first will
+    // reduce the output size most effectively.
+    int64 max_shape_dim = 0;
+    int64 max_reduce_dim = 0;
+    const auto &input_shape = reduce->operand(0)->shape();
+    for (int64 i = 0; i < reduce->dimensions().size(); ++i) {
+      if (input_shape.dimensions(reduce->dimensions(i)) > max_shape_dim) {
+        max_reduce_dim = reduce->dimensions(i);
+        max_shape_dim = input_shape.dimensions(max_reduce_dim);
+      }
+    }
+    // TODO(tjoerg): Run microbenchmarks to tune this threshold.
+    if (max_shape_dim < 128) {
+      return Status::OK();
+    }
+
+    // Split the reduction into a pre-reduction and a final reduction.
+    VLOG(3) << "Splitting reduction " << reduce->name() << " at dimension "
+            << max_reduce_dim;
+    std::vector pre_reduce_dims;
+    pre_reduce_dims.push_back(max_reduce_dim);
+    std::vector pre_reduce_shape_dims(input_shape.dimensions().begin(),
+                                             input_shape.dimensions().end());
+    pre_reduce_shape_dims.erase(pre_reduce_shape_dims.begin() + max_reduce_dim);
+    Shape pre_reduce_shape = ShapeUtil::MakeShape(
+        reduce->shape().element_type(), pre_reduce_shape_dims);
+    std::unique_ptr pre_reduce = HloInstruction::CreateReduce(
+        pre_reduce_shape, reduce->mutable_operand(0),
+        reduce->mutable_operand(1), pre_reduce_dims, reduce->to_apply());
+    pre_reduce->set_metadata(reduce->metadata());
+
+    std::vector final_reduce_dims(reduce->dimensions().begin(),
+                                         reduce->dimensions().end());
+    final_reduce_dims.erase(
+        std::remove(final_reduce_dims.begin(), final_reduce_dims.end(),
+                    max_reduce_dim),
+        final_reduce_dims.end());
+    for (int64 i = 0; i < final_reduce_dims.size(); ++i) {
+      if (final_reduce_dims[i] > max_reduce_dim) {
+        final_reduce_dims[i]--;
+      }
+    }
+    std::unique_ptr final_reduce = HloInstruction::CreateReduce(
+        reduce->shape(),
+        reduce->parent()->AddInstruction(std::move(pre_reduce)),
+        reduce->mutable_operand(1), final_reduce_dims, reduce->to_apply());
+    return ReplaceWithNewInstruction(reduce, std::move(final_reduce));
+  }
+};
+
+StatusOr ReductionSplitter::Run(HloModule *module) {
+  TF_ASSIGN_OR_RETURN(bool changed,
+                      ReductionSplitterVisitor().RunOnModule(module));
+  return changed;
+}
+
+}  // namespace gpu
+}  // namespace xla
diff --git a/tensorflow/compiler/xla/service/gpu/reduction_splitter.h b/tensorflow/compiler/xla/service/gpu/reduction_splitter.h
new file mode 100644
index 00000000000..f161b579eb8
--- /dev/null
+++ b/tensorflow/compiler/xla/service/gpu/reduction_splitter.h
@@ -0,0 +1,49 @@
+/* Copyright 2019 The TensorFlow Authors. All Rights Reserved.
+
+Licensed under the Apache License, Version 2.0 (the "License");
+you may not use this file except in compliance with the License.
+You may obtain a copy of the License at
+
+    http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software
+distributed under the License is distributed on an "AS IS" BASIS,
+WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+See the License for the specific language governing permissions and
+limitations under the License.
+==============================================================================*/
+#ifndef TENSORFLOW_COMPILER_XLA_SERVICE_GPU_REDUCTION_SPLITTER_H_
+#define TENSORFLOW_COMPILER_XLA_SERVICE_GPU_REDUCTION_SPLITTER_H_
+
+#include "tensorflow/compiler/xla/service/hlo_module.h"
+#include "tensorflow/compiler/xla/service/hlo_pass_interface.h"
+
+namespace xla {
+namespace gpu {
+
+// Splits a reduce op into two consecutive reduce ops if
+// * the reduce dimensions are not contiguous and
+// * at least one reduce dimension is large (i.e. corresponds to a large input
+//   shape dimension).
+//
+// Reductions with non-contiguous dimensions are emitted as simple element-wise
+// loops. This is inefficient when reducing large input shape dimensions.
+// Splitting such reductions allows using more efficient reduction emitters.
+//
+// This pass splits reduce ops into two consecutive reduce ops. Run it to a
+// fixpoint to split reduce ops along multiple large dimensions.
+//
+// Precondition: ReductionDimensionGrouper has been run and adjacent reduce
+// dimentsions have been grouped. Reduction layouts have been normalized.
+
+class ReductionSplitter : public HloModulePass {
+ public:
+  absl::string_view name() const override { return "reduction-splitter"; }
+
+  StatusOr Run(HloModule* module) override;
+};
+
+}  // namespace gpu
+}  // namespace xla
+
+#endif  // TENSORFLOW_COMPILER_XLA_SERVICE_GPU_REDUCTION_SPLITTER_H_
diff --git a/tensorflow/compiler/xla/service/gpu/reduction_splitter_test.cc b/tensorflow/compiler/xla/service/gpu/reduction_splitter_test.cc
new file mode 100644
index 00000000000..1be55b84204
--- /dev/null
+++ b/tensorflow/compiler/xla/service/gpu/reduction_splitter_test.cc
@@ -0,0 +1,140 @@
+/* Copyright 2016 The TensorFlow Authors. All Rights Reserved.
+
+Licensed under the Apache License, Version 2.0 (the "License");
+you may not use this file except in compliance with the License.
+You may obtain a copy of the License at
+
+    http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software
+distributed under the License is distributed on an "AS IS" BASIS,
+WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+See the License for the specific language governing permissions and
+limitations under the License.
+==============================================================================*/
+
+#include "tensorflow/compiler/xla/service/gpu/reduction_splitter.h"
+
+#include "tensorflow/compiler/xla/service/hlo_matchers.h"
+#include "tensorflow/compiler/xla/service/hlo_parser.h"
+#include "tensorflow/compiler/xla/shape_util.h"
+#include "tensorflow/compiler/xla/test.h"
+#include "tensorflow/compiler/xla/test_helpers.h"
+#include "tensorflow/compiler/xla/tests/hlo_test_base.h"
+
+namespace xla {
+namespace gpu {
+namespace {
+
+namespace op = xla::testing::opcode_matchers;
+
+class ReductionSplitterTest : public HloTestBase {};
+
+TEST_F(ReductionSplitterTest, SplitReductionAtDimensionTwo) {
+  auto module = ParseAndReturnVerifiedModule(R"(
+  HloModule test
+
+  add_computation {
+    x = f32[] parameter(0)
+    y = f32[] parameter(1)
+    ROOT add = f32[] add(x, y)
+  }
+
+  ENTRY entry_computation {
+    param_0 = f16[6,16,512,64]{3,2,1,0} parameter(0)
+    transpose.1781 = f16[6,512,16,64]{3,1,2,0} transpose(param_0), dimensions={0,2,1,3}
+    convert.6986 = f32[6,512,16,64]{3,1,2,0} convert(transpose.1781)
+    bitcast.2136 = f32[6,16,512,64]{3,2,1,0} bitcast(convert.6986)
+    constant_11111 = f32[] constant(0)
+    ROOT reduce.982 = f32[16,64]{1,0} reduce(bitcast.2136, constant_11111), dimensions={0,2}, to_apply=add_computation
+  }
+  )")
+                    .ValueOrDie();
+  ASSERT_TRUE(ReductionSplitter().Run(module.get()).ValueOrDie());
+  SCOPED_TRACE(module->ToString());
+  const HloInstruction* root_reduction =
+      module->entry_computation()->root_instruction();
+  ASSERT_THAT(root_reduction, op::Reduce(op::Reduce(), op::Constant()));
+
+  auto* pre_reduction = root_reduction->operand(0);
+  EXPECT_THAT(pre_reduction->dimensions(), std::vector({2}));
+  EXPECT_THAT(pre_reduction->shape(), ShapeUtil::MakeShape(F32, {6, 16, 64}));
+  EXPECT_THAT(root_reduction->dimensions(), std::vector({0}));
+  EXPECT_THAT(root_reduction->shape(), ShapeUtil::MakeShape(F32, {16, 64}));
+}
+
+TEST_F(ReductionSplitterTest, SplitReductionAtDimensionZero) {
+  auto module = ParseAndReturnVerifiedModule(R"(
+  HloModule test
+
+  add_computation {
+    x = f32[] parameter(0)
+    y = f32[] parameter(1)
+    ROOT add = f32[] add(x, y)
+  }
+
+  ENTRY entry_computation {
+    param_0 = f32[1024,16,512,64,128]{4,3,2,1,0} parameter(0)
+    constant_11111 = f32[] constant(0)
+    ROOT reduce.982 = f32[16,64]{1,0} reduce(param_0, constant_11111), dimensions={2,0,4}, to_apply=add_computation
+  }
+  )")
+                    .ValueOrDie();
+  ASSERT_TRUE(ReductionSplitter().Run(module.get()).ValueOrDie());
+  SCOPED_TRACE(module->ToString());
+  const HloInstruction* root_reduction =
+      module->entry_computation()->root_instruction();
+  ASSERT_THAT(root_reduction, op::Reduce(op::Reduce(), op::Constant()));
+
+  auto* pre_reduction = root_reduction->operand(0);
+  EXPECT_THAT(pre_reduction->dimensions(), std::vector({0}));
+  EXPECT_THAT(pre_reduction->shape(),
+              ShapeUtil::MakeShape(F32, {16, 512, 64, 128}));
+  EXPECT_THAT(root_reduction->dimensions(), std::vector({1, 3}));
+  EXPECT_THAT(root_reduction->shape(), ShapeUtil::MakeShape(F32, {16, 64}));
+}
+
+TEST_F(ReductionSplitterTest, DontSplitReductionWithSmallDimensions) {
+  auto module = ParseAndReturnVerifiedModule(R"(
+  HloModule test
+
+  add_computation {
+    x = f32[] parameter(0)
+    y = f32[] parameter(1)
+    ROOT add = f32[] add(x, y)
+  }
+
+  ENTRY entry_computation {
+    param_0 = f32[8,1024,8]{2,1,0} parameter(0)
+    constant_11111 = f32[] constant(0)
+    ROOT reduce.982 = f32[1024]{0} reduce(param_0, constant_11111), dimensions={2,0}, to_apply=add_computation
+  }
+  )")
+                    .ValueOrDie();
+  EXPECT_FALSE(ReductionSplitter().Run(module.get()).ValueOrDie());
+}
+
+TEST_F(ReductionSplitterTest, DontSplitReductionsWithContiguousDimensions) {
+  auto module = ParseAndReturnVerifiedModule(R"(
+  HloModule test
+
+  add_computation {
+    x = f32[] parameter(0)
+    y = f32[] parameter(1)
+    ROOT add = f32[] add(x, y)
+  }
+
+  ENTRY entry_computation {
+    param_0 = f32[128,128,64,128]{3,2,1,0} parameter(0)
+    constant_11111 = f32[] constant(0)
+    // The dimenstions to keep (1 and 2) are contiguous.
+    ROOT reduce.982 = f32[128,64]{1,0} reduce(param_0, constant_11111), dimensions={3,0}, to_apply=add_computation
+  }
+  )")
+                    .ValueOrDie();
+  EXPECT_FALSE(ReductionSplitter().Run(module.get()).ValueOrDie());
+}
+
+}  // namespace
+}  // namespace gpu
+}  // namespace xla
diff --git a/tensorflow/compiler/xla/service/gpu/tests/reduction_degenerate_dim_remover_test.cc b/tensorflow/compiler/xla/service/gpu/tests/reduction_degenerate_dim_remover_test.cc
index 2c5e704d7c2..92f558ee98d 100644
--- a/tensorflow/compiler/xla/service/gpu/tests/reduction_degenerate_dim_remover_test.cc
+++ b/tensorflow/compiler/xla/service/gpu/tests/reduction_degenerate_dim_remover_test.cc
@@ -37,6 +37,7 @@ class ReductionDegenerateDimRemoverTest : public GpuCodegenTest {
     DebugOptions debug_options = GpuCodegenTest::GetDebugOptionsForTest();
     debug_options.add_xla_disable_hlo_passes("reduction-layout-normalizer");
     debug_options.add_xla_disable_hlo_passes("reduction-dimension-grouper");
+    debug_options.add_xla_disable_hlo_passes("reduction-splitter");
     debug_options.add_xla_disable_hlo_passes("gpu-tree-reduction-rewriter");
     return debug_options;
   }
diff --git a/tensorflow/compiler/xla/service/gpu/tests/reduction_layout_normalizer_test.cc b/tensorflow/compiler/xla/service/gpu/tests/reduction_layout_normalizer_test.cc
index d06385480e5..b65c2842320 100644
--- a/tensorflow/compiler/xla/service/gpu/tests/reduction_layout_normalizer_test.cc
+++ b/tensorflow/compiler/xla/service/gpu/tests/reduction_layout_normalizer_test.cc
@@ -33,6 +33,7 @@ class ReductionLayoutNormalizerTest : public GpuCodegenTest {
   DebugOptions GetDebugOptionsForTest() override {
     DebugOptions debug_options = GpuCodegenTest::GetDebugOptionsForTest();
     debug_options.add_xla_disable_hlo_passes("reduction-dimension-grouper");
+    debug_options.add_xla_disable_hlo_passes("reduction-splitter");
     debug_options.add_xla_disable_hlo_passes("layout-assignment");
     debug_options.add_xla_disable_hlo_passes("gpu-tree-reduction-rewriter");
     return debug_options;

From e0780ef031fc27f4f2a71d745004d859a711c90a Mon Sep 17 00:00:00 2001
From: Reed Wanderman-Milne 
Date: Fri, 19 Jun 2020 10:21:14 -0700
Subject: [PATCH 307/867] Fix fp16 FusedBatchNorm CPU crash if batch dimension
 is 1.

The GPU kernel outputs NaNs for the variance in this case, which is also incorrect, but better than crashing.

PiperOrigin-RevId: 317331280
Change-Id: Iea4e5a3337625796c50244e51d7ccb4b89f4c3e4
---
 tensorflow/core/kernels/redux_functor.h       |  6 ++--
 .../python/ops/nn_fused_batchnorm_test.py     | 28 +++++++++++++++++--
 2 files changed, 30 insertions(+), 4 deletions(-)

diff --git a/tensorflow/core/kernels/redux_functor.h b/tensorflow/core/kernels/redux_functor.h
index 30038c62dbd..e07fa5364f1 100644
--- a/tensorflow/core/kernels/redux_functor.h
+++ b/tensorflow/core/kernels/redux_functor.h
@@ -57,7 +57,8 @@ struct ReduceOuterDimensions {
     if (1 == outer_dim) {
       // Nothing to do but passing input to output.
       output->template flat() =
-          input.template flat().reshape(output_dims);
+          input.template flat().template cast().reshape(
+              output_dims);
       return;
     }
 
@@ -226,7 +227,8 @@ struct ReduceMiddleDimensions {
     if ((1 == inner_dim * outer_dim)) {
       // Nothing to do.
       output->template flat() =
-          input.template flat().reshape(output_dims);
+          input.template flat().template cast().reshape(
+              output_dims);
       return;
     } else if (1 == inner_dim) {
       // Equivalent to ReduceOuterDimensions.
diff --git a/tensorflow/python/ops/nn_fused_batchnorm_test.py b/tensorflow/python/ops/nn_fused_batchnorm_test.py
index 5497325f6c0..1742a919216 100644
--- a/tensorflow/python/ops/nn_fused_batchnorm_test.py
+++ b/tensorflow/python/ops/nn_fused_batchnorm_test.py
@@ -375,9 +375,10 @@ class BatchNormalizationTest(test.TestCase):
     self.assertLess(err_grad_x_2, err_tolerance)
     self.assertLess(err_grad_scale, err_tolerance)
 
-  def _runtests(self, x_shape, is_training, gradient_test=False):
+  def _runtests(self, x_shape, is_training, gradient_test=False,
+                cpu_only=False):
     use_gpu_vals = [False]
-    if test.is_gpu_available(cuda_only=True):
+    if test.is_gpu_available(cuda_only=True) and not cpu_only:
       use_gpu_vals += [True]
     factors = [1.0, 0.6]
     for dtype in [np.float16, np.float32]:
@@ -438,6 +439,11 @@ class BatchNormalizationTest(test.TestCase):
     x_shape = [0, 131, 127, 6]
     self._runtests(x_shape, False)
 
+  def testInferenceShape6(self):
+    x_shape = [1, 1, 1, 1]
+    # GPU kernel doesn't properly handle case where non-channel dimensions are 1
+    self._runtests(x_shape, False, cpu_only=True)
+
   def testTrainingShape1(self):
     x_shape = [1, 1, 6, 1]
     self._runtests(x_shape, True)
@@ -459,6 +465,11 @@ class BatchNormalizationTest(test.TestCase):
     x_shape = [0, 131, 127, 6]
     self._runtests(x_shape, True)
 
+  def testTrainingShape6(self):
+    x_shape = [1, 1, 1, 1]
+    # GPU kernel doesn't properly handle case where non-channel dimensions are 1
+    self._runtests(x_shape, True, cpu_only=True)
+
   @test_util.run_deprecated_v1
   def testBatchNormGradInferenceShape1(self):
     x_shape = [1, 1, 6, 1]
@@ -485,6 +496,13 @@ class BatchNormalizationTest(test.TestCase):
     x_shape = [0, 7, 11, 4]
     self._runtests(x_shape, is_training=False, gradient_test=True)
 
+  @test_util.run_deprecated_v1
+  def testBatchNormGradInferenceShape6(self):
+    x_shape = [1, 1, 1, 1]
+    # GPU kernel doesn't properly handle case where non-channel dimensions are 1
+    self._runtests(x_shape, is_training=False, gradient_test=True,
+                   cpu_only=True)
+
   @test_util.run_deprecated_v1
   def testBatchNormGradTrainingShape1(self):
     x_shape = [1, 1, 6, 1]
@@ -511,6 +529,12 @@ class BatchNormalizationTest(test.TestCase):
     x_shape = [0, 7, 11, 4]
     self._runtests(x_shape, is_training=True, gradient_test=True)
 
+  @test_util.run_deprecated_v1
+  def testBatchNormGradTrainingShape6(self):
+    x_shape = [1, 1, 1, 1]
+    # GPU kernel doesn't properly handle case where non-channel dimensions are 1
+    self._runtests(x_shape, is_training=True, gradient_test=True, cpu_only=True)
+
   def _testBatchNormGradGrad(self, config):
     shape = config['shape']
     err_tolerance = config['err_tolerance']

From 94f241379a351da7cb7a180967915beec48d14ef Mon Sep 17 00:00:00 2001
From: Vo Van Nghia 
Date: Fri, 19 Jun 2020 20:42:39 +0700
Subject: [PATCH 308/867] Add gcs_filesystem_test

---
 .../experimental/filesystem/plugins/gcs/BUILD | 24 +++++++++-
 .../filesystem/plugins/gcs/gcs_filesystem.cc  | 10 ++++-
 .../plugins/gcs/gcs_filesystem_test.cc        | 45 +++++++++++++++++++
 3 files changed, 77 insertions(+), 2 deletions(-)
 create mode 100644 tensorflow/c/experimental/filesystem/plugins/gcs/gcs_filesystem_test.cc

diff --git a/tensorflow/c/experimental/filesystem/plugins/gcs/BUILD b/tensorflow/c/experimental/filesystem/plugins/gcs/BUILD
index 43f3a507f0b..3a65824cd7c 100644
--- a/tensorflow/c/experimental/filesystem/plugins/gcs/BUILD
+++ b/tensorflow/c/experimental/filesystem/plugins/gcs/BUILD
@@ -1,5 +1,5 @@
 # Experimental gcs filesystem plugin.
-load("//tensorflow:tensorflow.bzl", "get_win_copts", "tf_cc_shared_object")
+load("//tensorflow:tensorflow.bzl", "get_win_copts", "tf_cc_shared_object", "tf_cc_test")
 
 package(
     licenses = ["notice"],  # Apache 2.0
@@ -42,3 +42,25 @@ cc_library(
         "//tensorflow/c:env",
     ],
 )
+
+tf_cc_test(
+    name = "gcs_filesystem_test",
+    srcs = [
+        "gcs_filesystem.cc",
+        "gcs_filesystem_test.cc",
+    ],
+    local_defines = ["TF_GCS_FILESYSTEM_TEST"],
+    tags = [
+        "manual",
+    ],
+    deps = [
+        "gcs_helper",
+        "//tensorflow/c:tf_status",
+        "//tensorflow/c/experimental/filesystem:filesystem_interface",
+        "@com_github_googlecloudplatform_google_cloud_cpp//:storage_client",
+        "@com_google_absl//absl/strings",
+    ] + [
+        "//tensorflow/c:tf_status_helper",
+        "@com_google_googletest//:gtest",
+    ],
+)
diff --git a/tensorflow/c/experimental/filesystem/plugins/gcs/gcs_filesystem.cc b/tensorflow/c/experimental/filesystem/plugins/gcs/gcs_filesystem.cc
index 24d85f359ef..f459710ddc6 100644
--- a/tensorflow/c/experimental/filesystem/plugins/gcs/gcs_filesystem.cc
+++ b/tensorflow/c/experimental/filesystem/plugins/gcs/gcs_filesystem.cc
@@ -24,6 +24,14 @@ limitations under the License.
 #include "tensorflow/c/experimental/filesystem/plugins/gcs/gcs_helper.h"
 #include "tensorflow/c/tf_status.h"
 
+#ifdef TF_GCS_FILESYSTEM_TEST
+// For testing purpose, we expose some functions.
+#define TF_STATIC
+#else
+// Otherwise, we don't expose any symbol.
+#define TF_STATIC static
+#endif
+
 // Implementation of a filesystem for GCS environments.
 // This filesystem will support `gs://` URI schemes.
 namespace gcs = google::cloud::storage;
@@ -122,7 +130,7 @@ namespace tf_read_only_memory_region {
 namespace tf_gcs_filesystem {
 
 // TODO(vnvo2409): Add lazy-loading and customizing parameters.
-static void Init(TF_Filesystem* filesystem, TF_Status* status) {
+TF_STATIC void Init(TF_Filesystem* filesystem, TF_Status* status) {
   google::cloud::StatusOr client =
       gcs::Client::CreateDefaultClient();
   if (!client) {
diff --git a/tensorflow/c/experimental/filesystem/plugins/gcs/gcs_filesystem_test.cc b/tensorflow/c/experimental/filesystem/plugins/gcs/gcs_filesystem_test.cc
new file mode 100644
index 00000000000..eb0fbfc33f7
--- /dev/null
+++ b/tensorflow/c/experimental/filesystem/plugins/gcs/gcs_filesystem_test.cc
@@ -0,0 +1,45 @@
+#include "gtest/gtest.h"
+#include "tensorflow/c/experimental/filesystem/filesystem_interface.h"
+#include "tensorflow/c/tf_status_helper.h"
+
+#define ASSERT_TF_OK(x) ASSERT_EQ(TF_OK, TF_GetCode(x))
+
+// Forward declaration
+namespace tf_gcs_filesystem {
+void Init(TF_Filesystem* filesystem, TF_Status* status);
+}
+
+namespace tensorflow {
+namespace {
+
+class GCSFilesystemTest : public ::testing::Test {
+ public:
+  void SetUp() override {
+    status = TF_NewStatus();
+    filesystem = new TF_Filesystem;
+    tf_gcs_filesystem::Init(filesystem, status);
+    ASSERT_TF_OK(status) << "Can not initialize filesystem. "
+                         << TF_Message(status);
+  }
+  void TearDown() override {
+    TF_DeleteStatus(status);
+    // TODO(vnvo2409): Add filesystem cleanup
+    delete filesystem;
+  }
+
+ protected:
+  TF_Filesystem* filesystem;
+  TF_Status* status;
+};
+
+// We have to add this test here because there must be at least one test.
+// This test will be removed in the future.
+TEST_F(GCSFilesystemTest, TestInit) { ASSERT_TF_OK(status); }
+
+}  // namespace
+}  // namespace tensorflow
+
+GTEST_API_ int main(int argc, char** argv) {
+  ::testing::InitGoogleTest(&argc, argv);
+  return RUN_ALL_TESTS();
+}

From fbf407383c93774d10bd7c45cd66788a070b0e07 Mon Sep 17 00:00:00 2001
From: Nick Kreeger 
Date: Fri, 19 Jun 2020 10:40:30 -0700
Subject: [PATCH 309/867] Reduce the size of TfLiteTensor for the TF Micro
 runtime.

This change uses the existing micro-specific build flag (TF_LITE_STATIC_MEMORY) to reduce the size of TfLiteTensor. In this build setting, only the minimum number of fields required for preparing and initializing a model in TFLM are used. This build define is opt-in only for internal builds and continues to be enabled by default in Makefile builds./

All TFLM internal targets can be built with this flag by adding '--copt=-DTF_LITE_STATIC_MEMORY'.

This change reduces the sizeof(TfLiteTensor) to 64 bytes (64bit systems) down from 112 bytes (64 bit systems).

TfLiteTensor struct reduced by 1.75x (~43% reduction)
Tail allocation reduced by: 2,592kb (~12.5% reduction)
Total allocation reduced by: 2,592kb (~12% reduction)

Optimized results from memory_arena_threshold_test:
Keyword Model:
--------------
[RecordingMicroAllocator] Arena allocation total 18448 bytes
[RecordingMicroAllocator] Arena allocation head 672 bytes
[RecordingMicroAllocator] Arena allocation tail 17776 bytes
[RecordingMicroAllocator] 'TfLiteTensor struct' used 3456 bytes with alignment overhead (requested 3456 bytes for 54 tensors)
[RecordingMicroAllocator] 'TfLiteTensor quantization data' used 1728 bytes with alignment overhead (requested 1728 bytes for 108 allocations)
[RecordingMicroAllocator] 'TfLiteTensor variable buffer data' used 10240 bytes with alignment overhead (requested 10240 bytes for 7 allocations)
[RecordingMicroAllocator] 'NodeAndRegistration struct' used 1200 bytes with alignment overhead (requested 1200 bytes for 15 NodeAndRegistration structs)
[RecordingMicroAllocator] 'Operator runtime data' used 148 bytes with alignment overhead (requested 148 bytes for 13 OpData structs)

Test Conv Model:
----------------
[RecordingMicroAllocator] Arena allocation total 10960 bytes
[RecordingMicroAllocator] Arena allocation head 7744 bytes
[RecordingMicroAllocator] Arena allocation tail 3216 bytes
[RecordingMicroAllocator] 'TfLiteTensor struct' used 960 bytes with alignment overhead (requested 960 bytes for 15 tensors)
[RecordingMicroAllocator] 'TfLiteTensor quantization data' used 768 bytes with alignment overhead (requested 752 bytes for 24 allocations)
[RecordingMicroAllocator] 'TfLiteTensor variable buffer data' used 0 bytes with alignment overhead (requested 0 bytes for 0 allocations)
[RecordingMicroAllocator] 'NodeAndRegistration struct' used 560 bytes with alignment overhead (requested 560 bytes for 7 NodeAndRegistration structs)
[RecordingMicroAllocator] 'Operator runtime data' used 136 bytes with alignment overhead (requested 136 bytes for 5 OpData structs)

PiperOrigin-RevId: 317335359
Change-Id: Ic3d4d2c3e62249f072ece8f621f9ef94eaa28589
---
 tensorflow/lite/c/common.h                    | 46 +++++++++++++++++++
 .../lite/micro/memory_arena_threshold_test.cc | 19 ++++++--
 .../lite/micro/micro_interpreter_test.cc      |  2 +-
 tensorflow/lite/micro/tools/make/Makefile     |  2 +
 .../benchmark/experimental/c/c_api_types.h    | 46 +++++++++++++++++++
 5 files changed, 111 insertions(+), 4 deletions(-)

diff --git a/tensorflow/lite/c/common.h b/tensorflow/lite/c/common.h
index 15823784d12..9093e5d50ad 100644
--- a/tensorflow/lite/c/common.h
+++ b/tensorflow/lite/c/common.h
@@ -375,6 +375,7 @@ typedef struct TfLiteSparsity {
 
 // An tensor in the interpreter system which is a wrapper around a buffer of
 // data including a dimensionality (or NULL if not currently defined).
+#ifndef TF_LITE_STATIC_MEMORY
 typedef struct TfLiteTensor {
   // The data type specification for data stored in `data`. This affects
   // what member of `data` union should be used.
@@ -439,6 +440,51 @@ typedef struct TfLiteTensor {
   // `dims_signature` contains [1, -1, -1, 3]).
   const TfLiteIntArray* dims_signature;
 } TfLiteTensor;
+#else
+// Specific reduced TfLiteTensor struct for TF Micro runtime. This struct
+// contains only the minimum fields required to initialize and prepare a micro
+// inference graph. The fields in this struct have been ordered from
+// largest-to-smallest for optimal struct sizeof.
+//
+// NOTE: This flag is opt-in only at compile time.
+typedef struct TfLiteTensor {
+  // TODO(b/155784997): Consider consolidating these quantization fields:
+  // Quantization information. Replaces params field above.
+  TfLiteQuantization quantization;
+
+  // Quantization information.
+  TfLiteQuantizationParams params;
+
+  // A union of data pointers. The appropriate type should be used for a typed
+  // tensor based on `type`.
+  TfLitePtrUnion data;
+
+  // A pointer to a structure representing the dimensionality interpretation
+  // that the buffer should have. NOTE: the product of elements of `dims`
+  // and the element datatype size should be equal to `bytes` below.
+  TfLiteIntArray* dims;
+
+  // The number of bytes required to store the data of this Tensor. I.e.
+  // (bytes of each element) * dims[0] * ... * dims[n-1].  For example, if
+  // type is kTfLiteFloat32 and dims = {3, 2} then
+  // bytes = sizeof(float) * 3 * 2 = 4 * 3 * 2 = 24.
+  size_t bytes;
+
+  // The data type specification for data stored in `data`. This affects
+  // what member of `data` union should be used.
+  TfLiteType type;
+
+  // How memory is mapped
+  //  kTfLiteMmapRo: Memory mapped read only.
+  //  i.e. weights
+  //  kTfLiteArenaRw: Arena allocated read write memory
+  //  (i.e. temporaries, outputs).
+  TfLiteAllocationType allocation_type;
+
+  // True if the tensor is a variable.
+  bool is_variable;
+} TfLiteTensor;
+#endif  // TF_LITE_STATIC_MEMORY
 
 #ifndef TF_LITE_STATIC_MEMORY
 // Free data memory of tensor `t`.
diff --git a/tensorflow/lite/micro/memory_arena_threshold_test.cc b/tensorflow/lite/micro/memory_arena_threshold_test.cc
index 4f49b57112a..b45de85a21b 100644
--- a/tensorflow/lite/micro/memory_arena_threshold_test.cc
+++ b/tensorflow/lite/micro/memory_arena_threshold_test.cc
@@ -41,9 +41,17 @@ constexpr int kKeywordModelNodeAndRegistrationCount = 15;
 
 // NOTE: These values are measured on x86-64:
 // TODO(b/158651472): Consider auditing these values on non-64 bit systems.
+//
+// Run this test with '--copt=-DTF_LITE_MICRO_OPTIMIZED_RUNTIME' to get
+// optimized memory runtime values:
+#ifdef TF_LITE_STATIC_MEMORY
+constexpr int kKeywordModelTotalSize = 18448;
+constexpr int kKeywordModelTailSize = 17776;
+#else
 constexpr int kKeywordModelTotalSize = 21040;
-constexpr int kKeywordModelHeadSize = 672;
 constexpr int kKeywordModelTailSize = 20368;
+#endif
+constexpr int kKeywordModelHeadSize = 672;
 constexpr int kKeywordModelTfLiteTensorVariableBufferDataSize = 10240;
 constexpr int kKeywordModelTfLiteTensorQuantizationDataSize = 1728;
 constexpr int kKeywordModelOpRuntimeDataSize = 148;
@@ -56,9 +64,14 @@ constexpr int kTestConvModelNodeAndRegistrationCount = 7;
 
 // NOTE: These values are measured on x86-64:
 // TODO(b/158651472): Consider auditing these values on non-64 bit systems.
+#ifdef TF_LITE_STATIC_MEMORY
+constexpr int kTestConvModelTotalSize = 10960;
+constexpr int kTestConvModelTailSize = 3216;
+#else
 constexpr int kTestConvModelTotalSize = 11680;
-constexpr int kTestConvModelHeadSize = 7744;
 constexpr int kTestConvModelTailSize = 3936;
+#endif
+constexpr int kTestConvModelHeadSize = 7744;
 constexpr int kTestConvModelTfLiteTensorQuantizationDataSize = 768;
 constexpr int kTestConvModelOpRuntimeDataSize = 136;
 
@@ -81,7 +94,7 @@ void EnsureAllocatedSizeThreshold(const char* allocation_type, size_t actual,
     TF_LITE_MICRO_EXPECT_NEAR(actual, expected, kAllocationThreshold);
     if (actual != expected) {
       TF_LITE_REPORT_ERROR(micro_test::reporter,
-                           "%s threshold failed: %ld != %ld", allocation_type,
+                           "%s threshold failed: %d != %d", allocation_type,
                            actual, expected);
     }
   } else {
diff --git a/tensorflow/lite/micro/micro_interpreter_test.cc b/tensorflow/lite/micro/micro_interpreter_test.cc
index f54c212b573..c577d8cb513 100644
--- a/tensorflow/lite/micro/micro_interpreter_test.cc
+++ b/tensorflow/lite/micro/micro_interpreter_test.cc
@@ -284,7 +284,7 @@ TF_LITE_MICRO_TEST(TestIncompleteInitializationAllocationsWithSmallArena) {
 
   tflite::testing::MockOpResolver mock_resolver;
   // 1kb is too small for the ComplexMockModel:
-  constexpr size_t allocator_buffer_size = 1048;
+  constexpr size_t allocator_buffer_size = 500;
   uint8_t allocator_buffer[allocator_buffer_size];
 
   tflite::RecordingMicroAllocator* allocator =
diff --git a/tensorflow/lite/micro/tools/make/Makefile b/tensorflow/lite/micro/tools/make/Makefile
index 8b6cba06a0b..a75c59b05c9 100644
--- a/tensorflow/lite/micro/tools/make/Makefile
+++ b/tensorflow/lite/micro/tools/make/Makefile
@@ -75,6 +75,8 @@ TEST_SCRIPT := tensorflow/lite/micro/testing/test_linux_binary.sh
 MICROLITE_LIBS := -lm
 
 # TODO(b/150240249): Add in -fno-rtti once that works for the Xtensa toolchain.
+# TODO(b/159155203): Consider TF_LITE_STATIC_MEMORY to align more with the fact
+# this flag is for an optimized micro runtime.
 CXXFLAGS := -std=c++11 -DTF_LITE_STATIC_MEMORY
 CCFLAGS  := -std=c11   -DTF_LITE_STATIC_MEMORY
 ARFLAGS := -r
diff --git a/tensorflow/lite/tools/benchmark/experimental/c/c_api_types.h b/tensorflow/lite/tools/benchmark/experimental/c/c_api_types.h
index 15823784d12..9093e5d50ad 100644
--- a/tensorflow/lite/tools/benchmark/experimental/c/c_api_types.h
+++ b/tensorflow/lite/tools/benchmark/experimental/c/c_api_types.h
@@ -375,6 +375,7 @@ typedef struct TfLiteSparsity {
 
 // An tensor in the interpreter system which is a wrapper around a buffer of
 // data including a dimensionality (or NULL if not currently defined).
+#ifndef TF_LITE_STATIC_MEMORY
 typedef struct TfLiteTensor {
   // The data type specification for data stored in `data`. This affects
   // what member of `data` union should be used.
@@ -439,6 +440,51 @@ typedef struct TfLiteTensor {
   // `dims_signature` contains [1, -1, -1, 3]).
   const TfLiteIntArray* dims_signature;
 } TfLiteTensor;
+#else
+// Specific reduced TfLiteTensor struct for TF Micro runtime. This struct
+// contains only the minimum fields required to initialize and prepare a micro
+// inference graph. The fields in this struct have been ordered from
+// largest-to-smallest for optimal struct sizeof.
+//
+// NOTE: This flag is opt-in only at compile time.
+typedef struct TfLiteTensor {
+  // TODO(b/155784997): Consider consolidating these quantization fields:
+  // Quantization information. Replaces params field above.
+  TfLiteQuantization quantization;
+
+  // Quantization information.
+  TfLiteQuantizationParams params;
+
+  // A union of data pointers. The appropriate type should be used for a typed
+  // tensor based on `type`.
+  TfLitePtrUnion data;
+
+  // A pointer to a structure representing the dimensionality interpretation
+  // that the buffer should have. NOTE: the product of elements of `dims`
+  // and the element datatype size should be equal to `bytes` below.
+  TfLiteIntArray* dims;
+
+  // The number of bytes required to store the data of this Tensor. I.e.
+  // (bytes of each element) * dims[0] * ... * dims[n-1].  For example, if
+  // type is kTfLiteFloat32 and dims = {3, 2} then
+  // bytes = sizeof(float) * 3 * 2 = 4 * 3 * 2 = 24.
+  size_t bytes;
+
+  // The data type specification for data stored in `data`. This affects
+  // what member of `data` union should be used.
+  TfLiteType type;
+
+  // How memory is mapped
+  //  kTfLiteMmapRo: Memory mapped read only.
+  //  i.e. weights
+  //  kTfLiteArenaRw: Arena allocated read write memory
+  //  (i.e. temporaries, outputs).
+  TfLiteAllocationType allocation_type;
+
+  // True if the tensor is a variable.
+  bool is_variable;
+} TfLiteTensor;
+#endif  // TF_LITE_STATIC_MEMORY
 
 #ifndef TF_LITE_STATIC_MEMORY
 // Free data memory of tensor `t`.

From 1332b9365e403baf448358465e6ffd3e5368e614 Mon Sep 17 00:00:00 2001
From: Rahul Joshi 
Date: Fri, 19 Jun 2020 10:42:32 -0700
Subject: [PATCH 310/867] [NFC] Adopt FuncOp::isPublic() and friends

PiperOrigin-RevId: 317335755
Change-Id: I6f1b25798150bb3a255d4572831d3be4747d28c7
---
 tensorflow/compiler/mlir/tensorflow/ir/tf_saved_model.cc      | 4 ++--
 .../mlir/tensorflow/transforms/stack_ops_decomposition.cc     | 2 +-
 .../tensorflow/transforms/tensor_array_ops_decomposition.cc   | 2 +-
 .../tensorflow/transforms/tensor_list_ops_decomposition.cc    | 2 +-
 4 files changed, 5 insertions(+), 5 deletions(-)

diff --git a/tensorflow/compiler/mlir/tensorflow/ir/tf_saved_model.cc b/tensorflow/compiler/mlir/tensorflow/ir/tf_saved_model.cc
index d59532fef65..ef248379d2e 100644
--- a/tensorflow/compiler/mlir/tensorflow/ir/tf_saved_model.cc
+++ b/tensorflow/compiler/mlir/tensorflow/ir/tf_saved_model.cc
@@ -231,12 +231,12 @@ static LogicalResult VerifySavedModelModule(
   for (auto func : module.getOps()) {
     const bool is_exported = IsExported(func);
 
-    if (is_exported && func.getVisibility() != FuncOp::Visibility::Public) {
+    if (is_exported && !func.isPublic()) {
       return func.emitError()
              << "exported function @" << func.getName() << " should be public";
     }
 
-    if (!is_exported && func.getVisibility() == FuncOp::Visibility::Public) {
+    if (!is_exported && func.isPublic()) {
       return func.emitError() << "non-exported function @" << func.getName()
                               << " should be private";
     }
diff --git a/tensorflow/compiler/mlir/tensorflow/transforms/stack_ops_decomposition.cc b/tensorflow/compiler/mlir/tensorflow/transforms/stack_ops_decomposition.cc
index c349c2b4c3e..734a7d04a86 100644
--- a/tensorflow/compiler/mlir/tensorflow/transforms/stack_ops_decomposition.cc
+++ b/tensorflow/compiler/mlir/tensorflow/transforms/stack_ops_decomposition.cc
@@ -343,7 +343,7 @@ LogicalResult HandlePartitionedCallOp(
   }
   llvm::SmallDenseMap callee_map;
   FuncOp lowered_callee = callee;
-  if (callee.getVisibility() != SymbolTable::Visibility::Private) {
+  if (!callee.isPrivate()) {
     // Clone non-private callee in case of signature change.
     lowered_callee = callee.clone();
     lowered_callee.setVisibility(SymbolTable::Visibility::Private);
diff --git a/tensorflow/compiler/mlir/tensorflow/transforms/tensor_array_ops_decomposition.cc b/tensorflow/compiler/mlir/tensorflow/transforms/tensor_array_ops_decomposition.cc
index cfeb2b1f031..a9e1243714e 100644
--- a/tensorflow/compiler/mlir/tensorflow/transforms/tensor_array_ops_decomposition.cc
+++ b/tensorflow/compiler/mlir/tensorflow/transforms/tensor_array_ops_decomposition.cc
@@ -759,7 +759,7 @@ LogicalResult HandlePartitionedCallOp(
     return it->getSecond().accumulate_on_write;
   };
   FuncOp lowered_callee = callee;
-  if (callee.getVisibility() != SymbolTable::Visibility::Private) {
+  if (!callee.isPrivate()) {
     // Clone non-private callee in case of signature change.
     lowered_callee = callee.clone();
     lowered_callee.setVisibility(SymbolTable::Visibility::Private);
diff --git a/tensorflow/compiler/mlir/tensorflow/transforms/tensor_list_ops_decomposition.cc b/tensorflow/compiler/mlir/tensorflow/transforms/tensor_list_ops_decomposition.cc
index 9733bfe2290..b118ab6c6c9 100644
--- a/tensorflow/compiler/mlir/tensorflow/transforms/tensor_list_ops_decomposition.cc
+++ b/tensorflow/compiler/mlir/tensorflow/transforms/tensor_list_ops_decomposition.cc
@@ -322,7 +322,7 @@ LogicalResult HandlePartitionedCallOp(
   // Rewrite the callee.
   llvm::SmallDenseMap callee_map;
   FuncOp lowered_callee = callee;
-  if (callee.getVisibility() != SymbolTable::Visibility::Private) {
+  if (!callee.isPrivate()) {
     // Clone non-private callee in case of signature change.
     lowered_callee = callee.clone();
     lowered_callee.setVisibility(SymbolTable::Visibility::Private);

From 89de554b69d9b74896445a41250943d97e3b9e77 Mon Sep 17 00:00:00 2001
From: Scott Zhu 
Date: Fri, 19 Jun 2020 10:43:39 -0700
Subject: [PATCH 311/867] Fork keras related mirror_variable_test to
 keras/distribute.

PiperOrigin-RevId: 317335959
Change-Id: Ie34b8ddecdd137926959868157ab7ade7c72e1b9
---
 tensorflow/python/distribute/BUILD            |   1 -
 .../distribute/mirrored_variable_test.py      |  34 ------
 tensorflow/python/keras/distribute/BUILD      |  22 ++++
 .../distribute/mirrored_variable_test.py      | 106 ++++++++++++++++++
 4 files changed, 128 insertions(+), 35 deletions(-)
 create mode 100644 tensorflow/python/keras/distribute/mirrored_variable_test.py

diff --git a/tensorflow/python/distribute/BUILD b/tensorflow/python/distribute/BUILD
index 38c5550be16..f0f3766afe1 100644
--- a/tensorflow/python/distribute/BUILD
+++ b/tensorflow/python/distribute/BUILD
@@ -1511,7 +1511,6 @@ cuda_py_test(
         "//tensorflow/python:variable_scope",
         "//tensorflow/python/eager:context",
         "//tensorflow/python/eager:test",
-        "//tensorflow/python/keras/layers",
     ],
 )
 
diff --git a/tensorflow/python/distribute/mirrored_variable_test.py b/tensorflow/python/distribute/mirrored_variable_test.py
index 6623422b45f..df32a6babea 100644
--- a/tensorflow/python/distribute/mirrored_variable_test.py
+++ b/tensorflow/python/distribute/mirrored_variable_test.py
@@ -18,7 +18,6 @@ from __future__ import absolute_import
 from __future__ import division
 from __future__ import print_function
 
-from tensorflow.python.data.ops import dataset_ops
 from tensorflow.python.distribute import collective_all_reduce_strategy
 from tensorflow.python.distribute import combinations
 from tensorflow.python.distribute import distribution_strategy_context as ds_context
@@ -32,7 +31,6 @@ from tensorflow.python.framework import constant_op
 from tensorflow.python.framework import dtypes
 from tensorflow.python.framework import func_graph
 from tensorflow.python.framework import ops
-from tensorflow.python.keras.layers import core
 from tensorflow.python.ops import array_ops
 from tensorflow.python.ops import math_ops
 from tensorflow.python.ops import rnn
@@ -208,38 +206,6 @@ class MirroredVariableCreationTest(test.TestCase):
       # The resulting mirrored variable will use the name from the first device.
       self.assertEqual("foo_0:0", result.name)
 
-  def testWithLayers(self, distribution):
-
-    def model_fn(features):
-
-      layer1 = core.Dense(1)
-      layer1(features)
-      layer2 = core.Dense(1)
-      layer2(features)
-      # We rely on names and orders to make sure replica references the same
-      # MirroredVariable. Uniquifying names may involve global states,
-      # merge_call switches threads so we need to test things work after
-      # merge_call.
-      ds_context.get_replica_context().merge_call(lambda _: _)
-      layer3 = core.Dense(1)
-      layer3(features)
-      return [(layer1.kernel, layer1.bias), (layer2.kernel, layer2.bias),
-              (layer3.kernel, layer3.bias)]
-
-    iterator = distribution.make_input_fn_iterator(
-        lambda _: dataset_ops.Dataset.from_tensors([[1.]]).repeat(10))
-    self.evaluate(iterator.initializer)
-    features = iterator.get_next()
-
-    with distribution.scope():
-      result = distribution.extended.call_for_each_replica(
-          model_fn, args=(features,))
-      for kernel, bias in result:
-        self.assertIsInstance(kernel, values.MirroredVariable)
-        self.assertAllDifferent(distribution.experimental_local_results(kernel))
-        self.assertIsInstance(bias, values.MirroredVariable)
-        self.assertAllDifferent(distribution.experimental_local_results(kernel))
-
   def testWithVariableAndVariableScope(self, distribution):
 
     def model_fn():
diff --git a/tensorflow/python/keras/distribute/BUILD b/tensorflow/python/keras/distribute/BUILD
index 247e655621c..c6a8f2c5f91 100644
--- a/tensorflow/python/keras/distribute/BUILD
+++ b/tensorflow/python/keras/distribute/BUILD
@@ -347,6 +347,28 @@ cuda_py_test(
     ],
 )
 
+cuda_py_test(
+    name = "mirrored_variable_test",
+    srcs = ["mirrored_variable_test.py"],
+    python_version = "PY3",
+    tags = [
+        "guitar",
+        "multi_and_single_gpu",
+    ],
+    deps = [
+        "//tensorflow/python:config",
+        "//tensorflow/python/data/ops:dataset_ops",
+        "//tensorflow/python/distribute:collective_all_reduce_strategy",
+        "//tensorflow/python/distribute:combinations",
+        "//tensorflow/python/distribute:distribute_lib",
+        "//tensorflow/python/distribute:strategy_combinations",
+        "//tensorflow/python/distribute:values",
+        "//tensorflow/python/eager:context",
+        "//tensorflow/python/eager:test",
+        "//tensorflow/python/keras/layers:core",
+    ],
+)
+
 cuda_py_test(
     name = "multi_worker_test",
     srcs = ["multi_worker_test.py"],
diff --git a/tensorflow/python/keras/distribute/mirrored_variable_test.py b/tensorflow/python/keras/distribute/mirrored_variable_test.py
new file mode 100644
index 00000000000..0edfa4806f2
--- /dev/null
+++ b/tensorflow/python/keras/distribute/mirrored_variable_test.py
@@ -0,0 +1,106 @@
+# Copyright 2019 The TensorFlow Authors. All Rights Reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+# ==============================================================================
+"""Test MirroredVariable in MirroredStrategy and MultiWorkerMirroredStrategy."""
+
+from __future__ import absolute_import
+from __future__ import division
+from __future__ import print_function
+
+from tensorflow.python.data.ops import dataset_ops
+from tensorflow.python.distribute import collective_all_reduce_strategy
+from tensorflow.python.distribute import combinations
+from tensorflow.python.distribute import distribution_strategy_context as ds_context
+from tensorflow.python.distribute import strategy_combinations
+from tensorflow.python.distribute import values
+from tensorflow.python.eager import context
+from tensorflow.python.eager import test
+from tensorflow.python.framework import config
+from tensorflow.python.keras.layers import core
+
+
+def _mimic_two_cpus():
+  cpus = config.list_physical_devices("CPU")
+
+  config.set_logical_device_configuration(cpus[0], [
+      context.LogicalDeviceConfiguration(),
+      context.LogicalDeviceConfiguration(),
+  ])
+
+
+@combinations.generate(
+    combinations.combine(
+        distribution=[
+            strategy_combinations.mirrored_strategy_with_gpu_and_cpu,
+            combinations.NamedDistribution(
+                "Collective2CPUs",
+                # pylint: disable=g-long-lambda
+                lambda: collective_all_reduce_strategy.
+                CollectiveAllReduceStrategy._from_local_devices((
+                    "/device:CPU:0", "/device:CPU:1")),
+                required_gpus=0)
+        ],
+        mode=["graph", "eager"]))
+class MirroredVariableCreationTest(test.TestCase):
+  """Base class that tests mirrored variable creator.
+
+  Currently it assumes all strategy objects have two replicas.
+  """
+
+  @classmethod
+  def setUpClass(cls):
+    _mimic_two_cpus()
+
+  def assertAllDifferent(self, objs):
+    for i in range(len(objs)):
+      for j in range(len(objs)):
+        if i == j:
+          continue
+        self.assertIsNot(objs[i], objs[j])
+
+  def testWithLayers(self, distribution):
+
+    def model_fn(features):
+
+      layer1 = core.Dense(1)
+      layer1(features)
+      layer2 = core.Dense(1)
+      layer2(features)
+      # We rely on names and orders to make sure replica references the same
+      # MirroredVariable. Uniquifying names may involve global states,
+      # merge_call switches threads so we need to test things work after
+      # merge_call.
+      ds_context.get_replica_context().merge_call(lambda _: _)
+      layer3 = core.Dense(1)
+      layer3(features)
+      return [(layer1.kernel, layer1.bias), (layer2.kernel, layer2.bias),
+              (layer3.kernel, layer3.bias)]
+
+    iterator = distribution.make_input_fn_iterator(
+        lambda _: dataset_ops.Dataset.from_tensors([[1.]]).repeat(10))
+    self.evaluate(iterator.initializer)
+    features = iterator.get_next()
+
+    with distribution.scope():
+      result = distribution.extended.call_for_each_replica(
+          model_fn, args=(features,))
+      for kernel, bias in result:
+        self.assertIsInstance(kernel, values.MirroredVariable)
+        self.assertAllDifferent(distribution.experimental_local_results(kernel))
+        self.assertIsInstance(bias, values.MirroredVariable)
+        self.assertAllDifferent(distribution.experimental_local_results(kernel))
+
+
+if __name__ == "__main__":
+  test.main()

From 8088eddf203220799cd51ede0142329fa4665d3c Mon Sep 17 00:00:00 2001
From: Akshay Modi 
Date: Fri, 19 Jun 2020 10:58:32 -0700
Subject: [PATCH 312/867] Let tensorflow op take precedence when doing "ndarray
  tensor"

Also add a few more interop tests.

PiperOrigin-RevId: 317339113
Change-Id: Ic28fab7abefea681e1e8d840b8e4cf4f98b63f1e
---
 tensorflow/python/ops/numpy_ops/np_arrays.py  |  3 +-
 .../python/ops/numpy_ops/np_interop_test.py   | 50 ++++++++++++++++++
 .../python/ops/numpy_ops/np_math_ops.py       | 52 +++++++++++--------
 3 files changed, 82 insertions(+), 23 deletions(-)

diff --git a/tensorflow/python/ops/numpy_ops/np_arrays.py b/tensorflow/python/ops/numpy_ops/np_arrays.py
index 88bf4e7499a..65e8273375f 100644
--- a/tensorflow/python/ops/numpy_ops/np_arrays.py
+++ b/tensorflow/python/ops/numpy_ops/np_arrays.py
@@ -262,7 +262,8 @@ class ndarray(composite_tensor.CompositeTensor):  # pylint: disable=invalid-name
     """
     return np.asarray(self.data, dtype)
 
-  __array_priority__ = 110
+  # NOTE: we currently prefer interop with TF to allow TF to take precedence.
+  __array_priority__ = 90
 
   def __index__(self):
     """Returns a python scalar.
diff --git a/tensorflow/python/ops/numpy_ops/np_interop_test.py b/tensorflow/python/ops/numpy_ops/np_interop_test.py
index 052949dff9d..f52d3dae78b 100644
--- a/tensorflow/python/ops/numpy_ops/np_interop_test.py
+++ b/tensorflow/python/ops/numpy_ops/np_interop_test.py
@@ -19,12 +19,18 @@ from __future__ import division
 from __future__ import print_function
 
 
+import numpy as onp
+
+
 from tensorflow.python.eager import backprop
 from tensorflow.python.eager import def_function
+from tensorflow.python.framework import constant_op
 from tensorflow.python.framework import ops
 from tensorflow.python.ops import control_flow_ops
+from tensorflow.python.ops import math_ops
 from tensorflow.python.ops.numpy_ops import np_array_ops
 from tensorflow.python.ops.numpy_ops import np_arrays
+from tensorflow.python.ops.numpy_ops import np_math_ops
 from tensorflow.python.platform import test
 
 
@@ -88,6 +94,50 @@ class InteropTest(test.TestCase):
     self.assertEqual(10000, fn()[0])
     self.assertEqual(10000, def_function.function(fn)()[0])
 
+  def testTensorTFNPArrayInterop(self):
+    arr = np_array_ops.asarray(0.)
+    t = constant_op.constant(10.)
+
+    arr_plus_t = arr + t
+    t_plus_arr = t + arr
+
+    self.assertIsInstance(arr_plus_t, ops.Tensor)
+    self.assertIsInstance(t_plus_arr, ops.Tensor)
+    self.assertEqual(10., arr_plus_t.numpy())
+    self.assertEqual(10., t_plus_arr.numpy())
+
+  def testTensorTFNPOp(self):
+    t = constant_op.constant(10.)
+
+    sq = np_math_ops.square(t)
+    self.assertIsInstance(sq, np_arrays.ndarray)
+    self.assertEqual(100., sq)
+
+  def testTFNPArrayTFOpInterop(self):
+    arr = np_array_ops.asarray(10.)
+
+    # TODO(nareshmodi): Test more ops.
+    sq = math_ops.square(arr)
+    self.assertIsInstance(sq, ops.Tensor)
+    self.assertEqual(100., sq.numpy())
+
+  def testTFNPArrayNPOpInterop(self):
+    arr = np_array_ops.asarray([10.])
+
+    # TODO(nareshmodi): Test more ops.
+    sq = onp.square(arr)
+    self.assertIsInstance(sq, onp.ndarray)
+    self.assertEqual(100., sq[0])
+
+    # TODO(nareshmodi): Fails since the autopacking code doesn't use
+    # nest.flatten.
+#   def testAutopacking(self):
+#     arr1 = np_array_ops.asarray(1.)
+#     arr2 = np_array_ops.asarray(2.)
+#     arr3 = np_array_ops.asarray(3.)
+#     t = ops.convert_to_tensor_v2([arr1, arr2, arr3])
+
+#     self.assertEqual(t.numpy(), [1., 2., 3.])
 
 if __name__ == '__main__':
   ops.enable_eager_execution()
diff --git a/tensorflow/python/ops/numpy_ops/np_math_ops.py b/tensorflow/python/ops/numpy_ops/np_math_ops.py
index abfd9087ffd..361bfb50dec 100644
--- a/tensorflow/python/ops/numpy_ops/np_math_ops.py
+++ b/tensorflow/python/ops/numpy_ops/np_math_ops.py
@@ -917,29 +917,37 @@ def diff(a, n=1, axis=-1):  # pylint: disable=missing-function-docstring
   return _scalar(f, a)
 
 
-def _flip_args(f):
+def _wrap(f, reverse=False):
+  """Wraps binary ops so they can be added as operator overloads on ndarray."""
 
   def _f(a, b):
-    return f(b, a)
+    if reverse:
+      a, b = b, a
+
+    if getattr(b, '__array_priority__',
+               0) > np_arrays.ndarray.__array_priority__:
+      return NotImplemented
+
+    return f(a, b)
 
   return _f
 
 
 setattr(np_arrays.ndarray, '__abs__', absolute)
-setattr(np_arrays.ndarray, '__floordiv__', floor_divide)
-setattr(np_arrays.ndarray, '__rfloordiv__', _flip_args(floor_divide))
-setattr(np_arrays.ndarray, '__mod__', mod)
-setattr(np_arrays.ndarray, '__rmod__', _flip_args(mod))
-setattr(np_arrays.ndarray, '__add__', add)
-setattr(np_arrays.ndarray, '__radd__', _flip_args(add))
-setattr(np_arrays.ndarray, '__sub__', subtract)
-setattr(np_arrays.ndarray, '__rsub__', _flip_args(subtract))
-setattr(np_arrays.ndarray, '__mul__', multiply)
-setattr(np_arrays.ndarray, '__rmul__', _flip_args(multiply))
-setattr(np_arrays.ndarray, '__pow__', power)
-setattr(np_arrays.ndarray, '__rpow__', _flip_args(power))
-setattr(np_arrays.ndarray, '__truediv__', true_divide)
-setattr(np_arrays.ndarray, '__rtruediv__', _flip_args(true_divide))
+setattr(np_arrays.ndarray, '__floordiv__', _wrap(floor_divide))
+setattr(np_arrays.ndarray, '__rfloordiv__', _wrap(floor_divide, True))
+setattr(np_arrays.ndarray, '__mod__', _wrap(mod))
+setattr(np_arrays.ndarray, '__rmod__', _wrap(mod, True))
+setattr(np_arrays.ndarray, '__add__', _wrap(add))
+setattr(np_arrays.ndarray, '__radd__', _wrap(add, True))
+setattr(np_arrays.ndarray, '__sub__', _wrap(subtract))
+setattr(np_arrays.ndarray, '__rsub__', _wrap(subtract, True))
+setattr(np_arrays.ndarray, '__mul__', _wrap(multiply))
+setattr(np_arrays.ndarray, '__rmul__', _wrap(multiply, True))
+setattr(np_arrays.ndarray, '__pow__', _wrap(power))
+setattr(np_arrays.ndarray, '__rpow__', _wrap(power, True))
+setattr(np_arrays.ndarray, '__truediv__', _wrap(true_divide))
+setattr(np_arrays.ndarray, '__rtruediv__', _wrap(true_divide, True))
 
 
 def _comparison(tf_fun, x1, x2, cast_bool_to_int=False):
@@ -1031,12 +1039,12 @@ def logical_not(x):
 
 
 setattr(np_arrays.ndarray, '__invert__', logical_not)
-setattr(np_arrays.ndarray, '__lt__', less)
-setattr(np_arrays.ndarray, '__le__', less_equal)
-setattr(np_arrays.ndarray, '__gt__', greater)
-setattr(np_arrays.ndarray, '__ge__', greater_equal)
-setattr(np_arrays.ndarray, '__eq__', equal)
-setattr(np_arrays.ndarray, '__ne__', not_equal)
+setattr(np_arrays.ndarray, '__lt__', _wrap(less))
+setattr(np_arrays.ndarray, '__le__', _wrap(less_equal))
+setattr(np_arrays.ndarray, '__gt__', _wrap(greater))
+setattr(np_arrays.ndarray, '__ge__', _wrap(greater_equal))
+setattr(np_arrays.ndarray, '__eq__', _wrap(equal))
+setattr(np_arrays.ndarray, '__ne__', _wrap(not_equal))
 
 
 @np_utils.np_doc(np.linspace)

From 9152edc1f09e7420771628fd55ffef4b35b1cbc3 Mon Sep 17 00:00:00 2001
From: Bixia Zheng 
Date: Fri, 19 Jun 2020 11:04:49 -0700
Subject: [PATCH 313/867] [TF:TRT] Add flag
 TF_TRT_ALLOW_ENGINE_NATIVE_SEGMENT_EXECUTION.

The default value of the flag is True. When the flag value is false, the
bridge will report an error when the native segment of a TRTEngineOp is
executed.

Add test cases.

PiperOrigin-RevId: 317340632
Change-Id: Iacded09b38e63442bbd93076a079d385fb8a77e6
---
 .../tf2tensorrt/kernels/trt_engine_op.cc      | 34 ++++++++++++++++---
 .../compiler/tensorrt/trt_convert_test.py     | 33 +++++++++++++++++-
 2 files changed, 62 insertions(+), 5 deletions(-)

diff --git a/tensorflow/compiler/tf2tensorrt/kernels/trt_engine_op.cc b/tensorflow/compiler/tf2tensorrt/kernels/trt_engine_op.cc
index ac4a331041d..98d199ca9ab 100644
--- a/tensorflow/compiler/tf2tensorrt/kernels/trt_engine_op.cc
+++ b/tensorflow/compiler/tf2tensorrt/kernels/trt_engine_op.cc
@@ -45,6 +45,7 @@ limitations under the License.
 #include "tensorflow/core/platform/stream_executor.h"
 #include "tensorflow/core/platform/thread_annotations.h"
 #include "tensorflow/core/platform/types.h"
+#include "tensorflow/core/util/env_var.h"
 #include "tensorflow/stream_executor/lib/statusor.h"
 
 #if GOOGLE_CUDA
@@ -521,6 +522,17 @@ Status TRTEngineOp::VerifyInputShapes(
   return Status::OK();
 }
 
+static bool AllowEngineNativeSegmentExecution() {
+  bool value;
+  Status status =
+      ReadBoolFromEnvVar("TF_TRT_ALLOW_ENGINE_NATIVE_SEGMENT_EXECUTION",
+                         /*default_value=*/true, &value);
+  if (!status.ok()) {
+    LOG(ERROR) << status;
+  }
+  return value;
+}
+
 void TRTEngineOp::ComputeAsync(OpKernelContext* ctx,
                                AsyncOpKernel::DoneCallback done) {
   auto helper = new AsyncHelper(done);
@@ -605,17 +617,31 @@ void TRTEngineOp::ComputeAsync(OpKernelContext* ctx,
 
   EngineContext* engine_context = status.ValueOrDie().first;
   int trt_context_idx = status.ValueOrDie().second;
+  auto may_execute_native_segment = [&] {
+    if (!AllowEngineNativeSegmentExecution()) {
+      ctx->CtxFailure(
+          errors::Aborted("User disallowed engine native segment execution"));
+      return false;
+    }
+    return true;
+  };
   if (!engine_context->cuda_engine) {
-    VLOG(1) << "Engine retrieval for input shapes: "
-            << TensorShapeUtils::ShapeListString(input_concrete_shapes)
-            << " failed. Running native segment for " << name();
-    ExecuteNativeSegment(ctx, helper);
+    LOG_WARNING_WITH_PREFIX
+        << "Engine retrieval for input shapes: "
+        << TensorShapeUtils::ShapeListString(input_concrete_shapes)
+        << " failed. Running native segment for " << name();
+    if (may_execute_native_segment()) {
+      ExecuteNativeSegment(ctx, helper);
+    }
     return;
   }
   Status stat = ExecuteTrtEngine(ctx, engine_context, trt_context_idx);
   if (!stat.ok()) {
     LOG_WARNING_WITH_PREFIX << "Failed to execute engine: " << stat
                             << " Retrying with native segment for " << name();
+    if (!may_execute_native_segment()) {
+      return;
+    }
     // Release any outputs that are allocated, ExecuteNativeSegment will
     // re-allocate them and fail if they are currently allocated.
     for (int i = 0; i < ctx->num_outputs(); i++) {
diff --git a/tensorflow/python/compiler/tensorrt/trt_convert_test.py b/tensorflow/python/compiler/tensorrt/trt_convert_test.py
index df21e93f836..05ff6fcaebe 100644
--- a/tensorflow/python/compiler/tensorrt/trt_convert_test.py
+++ b/tensorflow/python/compiler/tensorrt/trt_convert_test.py
@@ -439,6 +439,7 @@ class TrtConvertTest(test_util.TensorFlowTestCase, parameterized.TestCase):
       self,
       input_saved_model_dir,
       input_saved_model_signature_key=_SAVED_MODEL_SIGNATURE_KEY,
+      max_workspace_size_bytes=10 << 20,  # Use a smaller workspace.
       precision_mode=trt_convert.TrtPrecisionMode.FP32,
       is_dynamic_op=True,
       maximum_cached_engines=2):
@@ -446,7 +447,7 @@ class TrtConvertTest(test_util.TensorFlowTestCase, parameterized.TestCase):
         input_saved_model_dir=input_saved_model_dir,
         input_saved_model_signature_key=input_saved_model_signature_key,
         conversion_params=trt_convert.DEFAULT_TRT_CONVERSION_PARAMS._replace(
-            max_workspace_size_bytes=10 << 20,  # Use a smaller workspace.
+            max_workspace_size_bytes=max_workspace_size_bytes,
             precision_mode=precision_mode,
             is_dynamic_op=is_dynamic_op,
             maximum_cached_engines=maximum_cached_engines))
@@ -924,6 +925,36 @@ class TrtConvertTest(test_util.TensorFlowTestCase, parameterized.TestCase):
         # to fall back to TF function.
         self._TestRun(sess, 2)
 
+  @test_util.run_v2_only
+  def testTrtGraphConverter_AllowEngineNativeSegmentExecution(self):
+    if not is_tensorrt_enabled():
+      return
+
+    np_input1, np_input2 = self._RandomInput([4, 1, 1])
+
+    # Create a model and save it.
+    input_saved_model_dir = self.mkdtemp()
+    root = self._GetModelForV2()
+    save.save(root, input_saved_model_dir,
+              {_SAVED_MODEL_SIGNATURE_KEY: root.run})
+
+    def _InputFn():
+      yield np_input1, np_input2
+
+    # Run TRT conversion and request an unreasonably large workspace.
+    converter = self._CreateConverterV2(
+        input_saved_model_dir, max_workspace_size_bytes=10 << 40)
+    converter.convert()
+
+    os.environ["TF_TRT_ALLOW_ENGINE_NATIVE_SEGMENT_EXECUTION"] = "False"
+    with self.assertRaisesRegex(
+        errors.AbortedError,
+        r"User disallowed engine native segment execution"):
+      converter.build(input_fn=_InputFn)
+
+    os.environ["TF_TRT_ALLOW_ENGINE_NATIVE_SEGMENT_EXECUTION"] = "True"
+    converter.build(input_fn=_InputFn)
+
   @test_util.run_v2_only
   def testBackwardCompatibility(self):
     """Load and execute a model that was saved in TF2.0."""

From 1129f21360b24fd957a248b86ddefa311d9f5658 Mon Sep 17 00:00:00 2001
From: Robert David 
Date: Fri, 19 Jun 2020 11:16:53 -0700
Subject: [PATCH 314/867] Consistently call LSTM's gate biases in variable
 names _gate_bias.

PiperOrigin-RevId: 317343158
Change-Id: I385ddbad6c1283b84574b2ec0b523ce9f88a4cd3
---
 .../kernels/bidirectional_sequence_lstm.cc    | 20 ++--
 .../bidirectional_sequence_lstm_test.cc       | 48 +++++-----
 tensorflow/lite/kernels/lstm.cc               | 56 ++++++-----
 tensorflow/lite/kernels/lstm_eval.cc          | 64 +++++++------
 tensorflow/lite/kernels/lstm_eval.h           |  8 +-
 tensorflow/lite/kernels/lstm_eval_test.cc     | 93 ++++++++++---------
 tensorflow/lite/kernels/lstm_test.cc          | 70 +++++++-------
 .../lite/kernels/optional_tensor_test.cc      |  8 +-
 .../kernels/unidirectional_sequence_lstm.cc   | 18 ++--
 .../unidirectional_sequence_lstm_test.cc      | 32 +++----
 .../calibration/builtin_logging_ops/lstm.cc   | 13 +--
 11 files changed, 224 insertions(+), 206 deletions(-)

diff --git a/tensorflow/lite/kernels/bidirectional_sequence_lstm.cc b/tensorflow/lite/kernels/bidirectional_sequence_lstm.cc
index 439fc94afad..fd60fe573ef 100644
--- a/tensorflow/lite/kernels/bidirectional_sequence_lstm.cc
+++ b/tensorflow/lite/kernels/bidirectional_sequence_lstm.cc
@@ -318,11 +318,11 @@ TfLiteStatus CheckLstmTensorDimensionsAndTypes(
   TF_LITE_ENSURE_EQ(context, forget_gate_bias->dims->data[0], n_cell);
   TF_LITE_ENSURE_TYPES_EQ(context, forget_gate_bias->type, kTfLiteFloat32);
 
-  const TfLiteTensor* cell_bias =
+  const TfLiteTensor* cell_gate_bias =
       GetInput(context, node, cell_gate_bias_tensor);
-  TF_LITE_ENSURE_EQ(context, cell_bias->dims->size, 1);
-  TF_LITE_ENSURE_EQ(context, cell_bias->dims->data[0], n_cell);
-  TF_LITE_ENSURE_EQ(context, cell_bias->type, kTfLiteFloat32);
+  TF_LITE_ENSURE_EQ(context, cell_gate_bias->dims->size, 1);
+  TF_LITE_ENSURE_EQ(context, cell_gate_bias->dims->data[0], n_cell);
+  TF_LITE_ENSURE_EQ(context, cell_gate_bias->type, kTfLiteFloat32);
 
   const TfLiteTensor* output_gate_bias =
       GetInput(context, node, output_gate_bias_tensor);
@@ -886,7 +886,7 @@ TfLiteStatus Eval(TfLiteContext* context, TfLiteNode* node) {
       GetOptionalInputTensor(context, node, kFwInputGateBiasTensor);
   const TfLiteTensor* fw_forget_gate_bias =
       GetInput(context, node, kFwForgetGateBiasTensor);
-  const TfLiteTensor* fw_cell_bias =
+  const TfLiteTensor* fw_cell_gate_bias =
       GetInput(context, node, kFwCellGateBiasTensor);
   const TfLiteTensor* fw_output_gate_bias =
       GetInput(context, node, kFwOutputGateBiasTensor);
@@ -934,7 +934,7 @@ TfLiteStatus Eval(TfLiteContext* context, TfLiteNode* node) {
       GetOptionalInputTensor(context, node, kBwInputGateBiasTensor);
   const TfLiteTensor* bw_forget_gate_bias =
       GetInput(context, node, kBwForgetGateBiasTensor);
-  const TfLiteTensor* bw_cell_bias =
+  const TfLiteTensor* bw_cell_gate_bias =
       GetInput(context, node, kBwCellGateBiasTensor);
   const TfLiteTensor* bw_output_gate_bias =
       GetInput(context, node, kBwOutputGateBiasTensor);
@@ -1029,7 +1029,7 @@ TfLiteStatus Eval(TfLiteContext* context, TfLiteNode* node) {
           /*output_layer_norm_coefficients=*/nullptr, real_aux_input,
           fw_aux_input_to_input_weights, fw_aux_input_to_forget_weights,
           fw_aux_input_to_cell_weights, fw_aux_input_to_output_weights,
-          fw_input_gate_bias, fw_forget_gate_bias, fw_cell_bias,
+          fw_input_gate_bias, fw_forget_gate_bias, fw_cell_gate_bias,
           fw_output_gate_bias, fw_projection_weights, fw_projection_bias,
           &lstm_params,
           /*forward_sequence=*/true, time_major, /*output_offset=*/0,
@@ -1049,7 +1049,7 @@ TfLiteStatus Eval(TfLiteContext* context, TfLiteNode* node) {
           /*output_layer_norm_coefficients=*/nullptr, real_aux_input,
           bw_aux_input_to_input_weights, bw_aux_input_to_forget_weights,
           bw_aux_input_to_cell_weights, bw_aux_input_to_output_weights,
-          bw_input_gate_bias, bw_forget_gate_bias, bw_cell_bias,
+          bw_input_gate_bias, bw_forget_gate_bias, bw_cell_gate_bias,
           bw_output_gate_bias, bw_projection_weights, bw_projection_bias,
           &lstm_params,
           /*forward_sequence=*/false, time_major, bw_output_offset,
@@ -1099,7 +1099,7 @@ TfLiteStatus Eval(TfLiteContext* context, TfLiteNode* node) {
           /*output_layer_norm_coefficients=*/nullptr, real_aux_input,
           fw_aux_input_to_input_weights, fw_aux_input_to_forget_weights,
           fw_aux_input_to_cell_weights, fw_aux_input_to_output_weights,
-          fw_input_gate_bias, fw_forget_gate_bias, fw_cell_bias,
+          fw_input_gate_bias, fw_forget_gate_bias, fw_cell_gate_bias,
           fw_output_gate_bias, fw_projection_weights, fw_projection_bias,
           &lstm_params,
           /*forward_sequence=*/true, /*time_major=*/true, /*output_offset=*/0,
@@ -1125,7 +1125,7 @@ TfLiteStatus Eval(TfLiteContext* context, TfLiteNode* node) {
           /*output_layer_norm_coefficients=*/nullptr, real_aux_input,
           bw_aux_input_to_input_weights, bw_aux_input_to_forget_weights,
           bw_aux_input_to_cell_weights, bw_aux_input_to_output_weights,
-          bw_input_gate_bias, bw_forget_gate_bias, bw_cell_bias,
+          bw_input_gate_bias, bw_forget_gate_bias, bw_cell_gate_bias,
           bw_output_gate_bias, bw_projection_weights, bw_projection_bias,
           &lstm_params,
           /*forward_sequence=*/false, /*time_major=*/true, bw_output_offset,
diff --git a/tensorflow/lite/kernels/bidirectional_sequence_lstm_test.cc b/tensorflow/lite/kernels/bidirectional_sequence_lstm_test.cc
index 3a52de130e3..778751aa04b 100644
--- a/tensorflow/lite/kernels/bidirectional_sequence_lstm_test.cc
+++ b/tensorflow/lite/kernels/bidirectional_sequence_lstm_test.cc
@@ -89,7 +89,7 @@ class BidirectionalLSTMOpModel : public SingleOpModel {
       fw_input_gate_bias_ = AddInput(TensorType_FLOAT32);
     }
     fw_forget_gate_bias_ = AddInput(TensorType_FLOAT32);
-    fw_cell_bias_ = AddInput(TensorType_FLOAT32);
+    fw_cell_gate_bias_ = AddInput(TensorType_FLOAT32);
     fw_output_gate_bias_ = AddInput(TensorType_FLOAT32);
 
     if (use_projection_weights) {
@@ -144,7 +144,7 @@ class BidirectionalLSTMOpModel : public SingleOpModel {
       bw_input_gate_bias_ = AddInput(TensorType_FLOAT32);
     }
     bw_forget_gate_bias_ = AddInput(TensorType_FLOAT32);
-    bw_cell_bias_ = AddInput(TensorType_FLOAT32);
+    bw_cell_gate_bias_ = AddInput(TensorType_FLOAT32);
     bw_output_gate_bias_ = AddInput(TensorType_FLOAT32);
 
     if (use_projection_weights) {
@@ -288,8 +288,8 @@ class BidirectionalLSTMOpModel : public SingleOpModel {
   }
 
   void SetCellBias(const std::vector& f) {
-    PopulateTensor(fw_cell_bias_, f);
-    PopulateTensor(bw_cell_bias_, f);
+    PopulateTensor(fw_cell_gate_bias_, f);
+    PopulateTensor(bw_cell_gate_bias_, f);
   }
 
   void SetOutputGateBias(const std::vector& f) {
@@ -364,7 +364,7 @@ class BidirectionalLSTMOpModel : public SingleOpModel {
 
   int fw_input_gate_bias_;
   int fw_forget_gate_bias_;
-  int fw_cell_bias_;
+  int fw_cell_gate_bias_;
   int fw_output_gate_bias_;
 
   int fw_projection_weights_;
@@ -386,7 +386,7 @@ class BidirectionalLSTMOpModel : public SingleOpModel {
 
   int bw_input_gate_bias_;
   int bw_forget_gate_bias_;
-  int bw_cell_bias_;
+  int bw_cell_gate_bias_;
   int bw_output_gate_bias_;
 
   int bw_projection_weights_;
@@ -467,7 +467,7 @@ TEST_P(LSTMOpTest, BlackBoxTestNoCifgNoPeepholeNoProjectionNoClipping) {
 
           {n_cell},  // input_gate_bias tensor
           {n_cell},  // forget_gate_bias tensor
-          {n_cell},  // cell_bias tensor
+          {n_cell},  // cell_gate_bias tensor
           {n_cell},  // output_gate_bias tensor
 
           {0, 0},  // projection_weight tensor
@@ -490,7 +490,7 @@ TEST_P(LSTMOpTest, BlackBoxTestNoCifgNoPeepholeNoProjectionNoClipping) {
 
           {n_cell},  // input_gate_bias tensor
           {n_cell},  // forget_gate_bias tensor
-          {n_cell},  // cell_bias tensor
+          {n_cell},  // cell_gate_bias tensor
           {n_cell},  // output_gate_bias tensor
 
           {0, 0},  // projection_weight tensor
@@ -633,7 +633,7 @@ TEST_P(LSTMOpTest, BlackBoxTestMergedOutput) {
 
           {n_cell},  // input_gate_bias tensor
           {n_cell},  // forget_gate_bias tensor
-          {n_cell},  // cell_bias tensor
+          {n_cell},  // cell_gate_bias tensor
           {n_cell},  // output_gate_bias tensor
 
           {0, 0},  // projection_weight tensor
@@ -656,7 +656,7 @@ TEST_P(LSTMOpTest, BlackBoxTestMergedOutput) {
 
           {n_cell},  // input_gate_bias tensor
           {n_cell},  // forget_gate_bias tensor
-          {n_cell},  // cell_bias tensor
+          {n_cell},  // cell_gate_bias tensor
           {n_cell},  // output_gate_bias tensor
 
           {0, 0},  // projection_weight tensor
@@ -796,7 +796,7 @@ TEST(LSTMOpTest, BlackBoxTestNoCifgNoPeepholeNoProjectionNoClippingReverse) {
 
           {n_cell},  // input_gate_bias tensor
           {n_cell},  // forget_gate_bias tensor
-          {n_cell},  // cell_bias tensor
+          {n_cell},  // cell_gate_bias tensor
           {n_cell},  // output_gate_bias tensor
 
           {0, 0},  // projection_weight tensor
@@ -819,7 +819,7 @@ TEST(LSTMOpTest, BlackBoxTestNoCifgNoPeepholeNoProjectionNoClippingReverse) {
 
           {n_cell},  // input_gate_bias tensor
           {n_cell},  // forget_gate_bias tensor
-          {n_cell},  // cell_bias tensor
+          {n_cell},  // cell_gate_bias tensor
           {n_cell},  // output_gate_bias tensor
 
           {0, 0},  // projection_weight tensor
@@ -956,7 +956,7 @@ TEST(LSTMOpTest, BlackBoxTestWithCifgWithPeepholeNoProjectionNoClipping) {
 
           {0},       // input_gate_bias tensor
           {n_cell},  // forget_gate_bias tensor
-          {n_cell},  // cell_bias tensor
+          {n_cell},  // cell_gate_bias tensor
           {n_cell},  // output_gate_bias tensor
 
           {0, 0},  // projection_weight tensor
@@ -978,7 +978,7 @@ TEST(LSTMOpTest, BlackBoxTestWithCifgWithPeepholeNoProjectionNoClipping) {
 
           {0},       // input_gate_bias tensor
           {n_cell},  // forget_gate_bias tensor
-          {n_cell},  // cell_bias tensor
+          {n_cell},  // cell_gate_bias tensor
           {n_cell},  // output_gate_bias tensor
 
           {0, 0},  // projection_weight tensor
@@ -1107,7 +1107,7 @@ TEST(LSTMOpTest,
 
           {0},       // input_gate_bias tensor
           {n_cell},  // forget_gate_bias tensor
-          {n_cell},  // cell_bias tensor
+          {n_cell},  // cell_gate_bias tensor
           {n_cell},  // output_gate_bias tensor
 
           {0, 0},  // projection_weight tensor
@@ -1129,7 +1129,7 @@ TEST(LSTMOpTest,
 
           {0},       // input_gate_bias tensor
           {n_cell},  // forget_gate_bias tensor
-          {n_cell},  // cell_bias tensor
+          {n_cell},  // cell_gate_bias tensor
           {n_cell},  // output_gate_bias tensor
 
           {0, 0},  // projection_weight tensor
@@ -1258,7 +1258,7 @@ TEST(LSTMOpTest, BlackBoxTestWithPeepholeWithProjectionNoClipping) {
 
           {n_cell},  // input_gate_bias tensor
           {n_cell},  // forget_gate_bias tensor
-          {n_cell},  // cell_bias tensor
+          {n_cell},  // cell_gate_bias tensor
           {n_cell},  // output_gate_bias tensor
 
           {n_output, n_cell},  // projection_weight tensor
@@ -1280,7 +1280,7 @@ TEST(LSTMOpTest, BlackBoxTestWithPeepholeWithProjectionNoClipping) {
 
           {n_cell},  // input_gate_bias tensor
           {n_cell},  // forget_gate_bias tensor
-          {n_cell},  // cell_bias tensor
+          {n_cell},  // cell_gate_bias tensor
           {n_cell},  // output_gate_bias tensor
 
           {n_output, n_cell},  // projection_weight tensor
@@ -1961,7 +1961,7 @@ TEST(LSTMOpTest, BlackBoxTestWithPeepholeWithProjectionNoClippingBatchMajor) {
 
           {n_cell},  // input_gate_bias tensor
           {n_cell},  // forget_gate_bias tensor
-          {n_cell},  // cell_bias tensor
+          {n_cell},  // cell_gate_bias tensor
           {n_cell},  // output_gate_bias tensor
 
           {n_output, n_cell},  // projection_weight tensor
@@ -1983,7 +1983,7 @@ TEST(LSTMOpTest, BlackBoxTestWithPeepholeWithProjectionNoClippingBatchMajor) {
 
           {n_cell},  // input_gate_bias tensor
           {n_cell},  // forget_gate_bias tensor
-          {n_cell},  // cell_bias tensor
+          {n_cell},  // cell_gate_bias tensor
           {n_cell},  // output_gate_bias tensor
 
           {n_output, n_cell},  // projection_weight tensor
@@ -2667,7 +2667,7 @@ TEST_P(LSTMOpTest, BlackBoxTestWithAuxInputZeroAuxWeight) {
 
           {n_cell},  // input_gate_bias tensor
           {n_cell},  // forget_gate_bias tensor
-          {n_cell},  // cell_bias tensor
+          {n_cell},  // cell_gate_bias tensor
           {n_cell},  // output_gate_bias tensor
 
           {0, 0},  // projection_weight tensor
@@ -2690,7 +2690,7 @@ TEST_P(LSTMOpTest, BlackBoxTestWithAuxInputZeroAuxWeight) {
 
           {n_cell},  // input_gate_bias tensor
           {n_cell},  // forget_gate_bias tensor
-          {n_cell},  // cell_bias tensor
+          {n_cell},  // cell_gate_bias tensor
           {n_cell},  // output_gate_bias tensor
 
           {0, 0},  // projection_weight tensor
@@ -2841,7 +2841,7 @@ TEST_P(LSTMOpTest, BlackBoxTestWithAuxInput) {
 
           {n_cell},  // input_gate_bias tensor
           {n_cell},  // forget_gate_bias tensor
-          {n_cell},  // cell_bias tensor
+          {n_cell},  // cell_gate_bias tensor
           {n_cell},  // output_gate_bias tensor
 
           {0, 0},  // projection_weight tensor
@@ -2864,7 +2864,7 @@ TEST_P(LSTMOpTest, BlackBoxTestWithAuxInput) {
 
           {n_cell},  // input_gate_bias tensor
           {n_cell},  // forget_gate_bias tensor
-          {n_cell},  // cell_bias tensor
+          {n_cell},  // cell_gate_bias tensor
           {n_cell},  // output_gate_bias tensor
 
           {0, 0},  // projection_weight tensor
diff --git a/tensorflow/lite/kernels/lstm.cc b/tensorflow/lite/kernels/lstm.cc
index aa6a112a022..b941f2237ca 100644
--- a/tensorflow/lite/kernels/lstm.cc
+++ b/tensorflow/lite/kernels/lstm.cc
@@ -407,7 +407,8 @@ TfLiteStatus PopulateQuantizedLstmParams8x8_8(
       GetOptionalInputTensor(context, node, kInputGateBiasTensor);
   const TfLiteTensor* forget_gate_bias =
       GetInput(context, node, kForgetGateBiasTensor);
-  const TfLiteTensor* cell_bias = GetInput(context, node, kCellGateBiasTensor);
+  const TfLiteTensor* cell_gate_bias =
+      GetInput(context, node, kCellGateBiasTensor);
   const TfLiteTensor* output_gate_bias =
       GetInput(context, node, kOutputGateBiasTensor);
 
@@ -446,10 +447,10 @@ TfLiteStatus PopulateQuantizedLstmParams8x8_8(
   int16_t* layer_norm_forget_weight_ptr = nullptr;
   int16_t* layer_norm_cell_weight_ptr = nullptr;
   int16_t* layer_norm_output_weight_ptr = nullptr;
-  int32_t* input_bias_ptr = nullptr;
-  int32_t* forget_bias_ptr = nullptr;
-  int32_t* cell_bias_ptr = nullptr;
-  int32_t* output_bias_ptr = nullptr;
+  int32_t* input_gate_bias_ptr = nullptr;
+  int32_t* forget_gate_bias_ptr = nullptr;
+  int32_t* cell_gate_bias_ptr = nullptr;
+  int32_t* output_gate_bias_ptr = nullptr;
   int32_t* proj_bias_ptr = nullptr;
   int16_t* cell_ptr = nullptr;
   int8_t* output_state_ptr = nullptr;
@@ -497,7 +498,7 @@ TfLiteStatus PopulateQuantizedLstmParams8x8_8(
   if (!use_cifg) {
     input_to_input_weight_ptr = input_to_input_weights->data.int8;
     recurrent_to_input_weight_ptr = recurrent_to_input_weights->data.int8;
-    input_bias_ptr = input_gate_bias->data.i32;
+    input_gate_bias_ptr = input_gate_bias->data.i32;
     input_to_input_weight_scale = input_to_input_weights->params.scale;
     recurrent_to_input_weight_scale = recurrent_to_input_weights->params.scale;
   }
@@ -547,9 +548,9 @@ TfLiteStatus PopulateQuantizedLstmParams8x8_8(
   recurrent_to_cell_weight_scale = recurrent_to_cell_weights->params.scale;
   recurrent_to_output_weight_ptr = recurrent_to_output_weights->data.int8;
   recurrent_to_output_weight_scale = recurrent_to_output_weights->params.scale;
-  forget_bias_ptr = forget_gate_bias->data.i32;
-  cell_bias_ptr = cell_bias->data.i32;
-  output_bias_ptr = output_gate_bias->data.i32;
+  forget_gate_bias_ptr = forget_gate_bias->data.i32;
+  cell_gate_bias_ptr = cell_gate_bias->data.i32;
+  output_gate_bias_ptr = output_gate_bias->data.i32;
   output_state_ptr = output_state->data.int8;
   cell_ptr = cell_state->data.i16;
   input_scale = input->params.scale;
@@ -875,13 +876,14 @@ TfLiteStatus CheckInputTensorDimensions(TfLiteContext* context,
     TF_LITE_ENSURE_TYPES_EQ(context, forget_gate_bias->type, kTfLiteFloat32);
   }
 
-  const TfLiteTensor* cell_bias = GetInput(context, node, kCellGateBiasTensor);
-  TF_LITE_ENSURE_EQ(context, cell_bias->dims->size, 1);
-  TF_LITE_ENSURE_EQ(context, cell_bias->dims->data[0], n_cell);
+  const TfLiteTensor* cell_gate_bias =
+      GetInput(context, node, kCellGateBiasTensor);
+  TF_LITE_ENSURE_EQ(context, cell_gate_bias->dims->size, 1);
+  TF_LITE_ENSURE_EQ(context, cell_gate_bias->dims->data[0], n_cell);
   if (is_integer) {
-    TF_LITE_ENSURE_TYPES_EQ(context, cell_bias->type, kTfLiteInt32);
+    TF_LITE_ENSURE_TYPES_EQ(context, cell_gate_bias->type, kTfLiteInt32);
   } else {
-    TF_LITE_ENSURE_TYPES_EQ(context, cell_bias->type, kTfLiteFloat32);
+    TF_LITE_ENSURE_TYPES_EQ(context, cell_gate_bias->type, kTfLiteFloat32);
   }
 
   const TfLiteTensor* output_gate_bias =
@@ -1526,7 +1528,8 @@ TfLiteStatus Eval(TfLiteContext* context, TfLiteNode* node) {
       GetOptionalInputTensor(context, node, kInputGateBiasTensor);
   const TfLiteTensor* forget_gate_bias =
       GetInput(context, node, kForgetGateBiasTensor);
-  const TfLiteTensor* cell_bias = GetInput(context, node, kCellGateBiasTensor);
+  const TfLiteTensor* cell_gate_bias =
+      GetInput(context, node, kCellGateBiasTensor);
   const TfLiteTensor* output_gate_bias =
       GetInput(context, node, kOutputGateBiasTensor);
 
@@ -1560,8 +1563,9 @@ TfLiteStatus Eval(TfLiteContext* context, TfLiteNode* node) {
           /*aux_input_to_forget_weights=*/nullptr,
           /*aux_input_to_cell_weights=*/nullptr,
           /*aux_input_to_output_weights=*/nullptr, input_gate_bias,
-          forget_gate_bias, cell_bias, output_gate_bias, projection_weights,
-          projection_bias, params, /*forward_sequence=*/true,
+          forget_gate_bias, cell_gate_bias, output_gate_bias,
+          projection_weights, projection_bias, params,
+          /*forward_sequence=*/true,
           /*time_major=*/true,
           /*output_offset=*/0, scratch_buffer, output_state, cell_state,
           output);
@@ -1603,8 +1607,9 @@ TfLiteStatus Eval(TfLiteContext* context, TfLiteNode* node) {
             /*aux_input_to_forget_weights=*/nullptr,
             /*aux_input_to_cell_weights=*/nullptr,
             /*aux_input_to_output_weights=*/nullptr, input_gate_bias,
-            forget_gate_bias, cell_bias, output_gate_bias, projection_weights,
-            projection_bias, params, /*forward_sequence=*/true,
+            forget_gate_bias, cell_gate_bias, output_gate_bias,
+            projection_weights, projection_bias, params,
+            /*forward_sequence=*/true,
             /*time_major=*/true, /*output_offset=*/0, scratch_buffer,
             scaling_factors, prod_scaling_factors, recovered_cell_weights,
             input_quantized,
@@ -1631,10 +1636,11 @@ TfLiteStatus Eval(TfLiteContext* context, TfLiteNode* node) {
               cell_to_output_weights, input_layer_norm_coefficients,
               forget_layer_norm_coefficients, cell_layer_norm_coefficients,
               output_layer_norm_coefficients, input_gate_bias, forget_gate_bias,
-              cell_bias, output_gate_bias, projection_weights, projection_bias,
-              params, &op_data->integer_lstm_param, output_state, cell_state,
-              output, scratch0, scratch1, scratch2, scratch3, scratch4,
-              scratch5, CpuBackendContext::GetFromContext(context));
+              cell_gate_bias, output_gate_bias, projection_weights,
+              projection_bias, params, &op_data->integer_lstm_param,
+              output_state, cell_state, output, scratch0, scratch1, scratch2,
+              scratch3, scratch4, scratch5,
+              CpuBackendContext::GetFromContext(context));
         } else {
           TfLiteTensor* scratch0 = GetTemporary(context, node, /*index=*/0);
           TfLiteTensor* scratch1 = GetTemporary(context, node, /*index=*/1);
@@ -1653,8 +1659,8 @@ TfLiteStatus Eval(TfLiteContext* context, TfLiteNode* node) {
               cell_to_output_weights, input_layer_norm_coefficients,
               forget_layer_norm_coefficients, cell_layer_norm_coefficients,
               output_layer_norm_coefficients, input_gate_bias, forget_gate_bias,
-              cell_bias, output_gate_bias, projection_weights, projection_bias,
-              params, output_state, cell_state, output,
+              cell_gate_bias, output_gate_bias, projection_weights,
+              projection_bias, params, output_state, cell_state, output,
               &op_data->integer_lstm_param, scratch0, scratch1, scratch2,
               scratch3, scratch4, scratch5, scratch6, scratch7);
           return kTfLiteOk;
diff --git a/tensorflow/lite/kernels/lstm_eval.cc b/tensorflow/lite/kernels/lstm_eval.cc
index f45d46762bf..9bdbfa9d48d 100644
--- a/tensorflow/lite/kernels/lstm_eval.cc
+++ b/tensorflow/lite/kernels/lstm_eval.cc
@@ -942,10 +942,10 @@ inline void LstmStepHybrid(
 //   effective_proj_scale_b                  - optional
 //
 // Gate biases of size 'n_cell':
-//   input_bias_ptr                 - optional
-//   forget_bias_ptr
+//   input_gate_bias_ptr                 - optional
+//   forget_gate_bias_ptr
 //   cell_gate_bias_ptr
-//   output_bias_ptr
+//   output_gate_bias_ptr
 //
 // Layer norm coefficients of size 'n_cell', representing diagonal matrices.
 //   layer_norm_input_weight_ptr    - optional
@@ -1031,8 +1031,8 @@ inline void LstmStepInteger(
     int32_t layer_norm_cell_scale_b,
     const int16_t* layer_norm_output_weight_ptr,
     int32_t layer_norm_output_scale_a, int32_t layer_norm_output_scale_b,
-    const int32_t* input_bias_ptr, const int32_t* forget_bias_ptr,
-    const int32_t* cell_gate_bias_ptr, const int32_t* output_bias_ptr,
+    const int32_t* input_gate_bias_ptr, const int32_t* forget_gate_bias_ptr,
+    const int32_t* cell_gate_bias_ptr, const int32_t* output_gate_bias_ptr,
     int16_t quantized_cell_clip, int8_t quantized_proj_clip, int32_t cell_scale,
     int32_t input_variance_guard, int32_t forget_variance_guard,
     int32_t cell_variance_guard, int32_t output_variance_guard,
@@ -1098,7 +1098,7 @@ inline void LstmStepInteger(
 
   if (use_layer_norm) {
     tensor_utils::ApplyLayerNorm(
-        scratch_1_ptr, layer_norm_forget_weight_ptr, forget_bias_ptr,
+        scratch_1_ptr, layer_norm_forget_weight_ptr, forget_gate_bias_ptr,
         layer_norm_forget_scale_a, layer_norm_forget_scale_b,
         forget_variance_guard, n_batch, n_cell, scratch_1_ptr);
   }
@@ -1149,7 +1149,7 @@ inline void LstmStepInteger(
 
     if (use_layer_norm) {
       tensor_utils::ApplyLayerNorm(
-          scratch_0_ptr, layer_norm_input_weight_ptr, input_bias_ptr,
+          scratch_0_ptr, layer_norm_input_weight_ptr, input_gate_bias_ptr,
           layer_norm_input_scale_a, layer_norm_input_scale_b,
           input_variance_guard, n_batch, n_cell, scratch_0_ptr);
     }
@@ -1190,7 +1190,7 @@ inline void LstmStepInteger(
 
   if (use_layer_norm) {
     tensor_utils::ApplyLayerNorm(
-        scratch_3_ptr, layer_norm_output_weight_ptr, output_bias_ptr,
+        scratch_3_ptr, layer_norm_output_weight_ptr, output_gate_bias_ptr,
         layer_norm_output_scale_a, layer_norm_output_scale_b,
         output_variance_guard, n_batch, n_cell, scratch_3_ptr);
   }
@@ -1268,10 +1268,10 @@ inline void LstmStepInteger(
 //   effective_proj_scale_b                  - optional
 //
 // Gate biases of size 'n_cell':
-//   input_bias_ptr                 - optional
-//   forget_bias_ptr
+//   input_gate_bias_ptr                 - optional
+//   forget_gate_bias_ptr
 //   cell_gate_bias_ptr
-//   output_bias_ptr
+//   output_gate_bias_ptr
 //
 // Layer norm coefficients of size 'n_cell', representing diagonal matrices.
 //   layer_norm_input_weight_ptr    - optional
@@ -1358,8 +1358,8 @@ void LstmStepInteger(
     int32_t layer_norm_cell_scale_b,
     const int16_t* layer_norm_output_weight_ptr,
     int32_t layer_norm_output_scale_a, int32_t layer_norm_output_scale_b,
-    const int32_t* input_bias_ptr, const int32_t* forget_bias_ptr,
-    const int32_t* cell_gate_bias_ptr, const int32_t* output_bias_ptr,
+    const int32_t* input_gate_bias_ptr, const int32_t* forget_gate_bias_ptr,
+    const int32_t* cell_gate_bias_ptr, const int32_t* output_gate_bias_ptr,
     const int32_t* proj_bias_ptr, const TfLiteLSTMParams* params,
     const int32_t* intermediate_scale_a, const int32_t* intermediate_scale_b,
     const int32_t* intermediate_zp, int16_t quantized_cell_clip,
@@ -1391,7 +1391,8 @@ void LstmStepInteger(
   // Forget gate layer norm.
   tensor_utils::ApplyLayerNormFloat(
       scratch2, layer_norm_forget_weight_ptr, layer_norm_forget_scale_a,
-      layer_norm_forget_scale_b, forget_bias_ptr, n_batch, n_cell, scratch2);
+      layer_norm_forget_scale_b, forget_gate_bias_ptr, n_batch, n_cell,
+      scratch2);
 
   // Forget gate sigmoid.
   tensor_utils::ApplySigmoidFloat(scratch2, n_batch, n_cell, scratch2);
@@ -1444,7 +1445,8 @@ void LstmStepInteger(
   // Output gate with layer norm.
   tensor_utils::ApplyLayerNormFloat(
       scratch4, layer_norm_output_weight_ptr, layer_norm_output_scale_a,
-      layer_norm_output_scale_b, output_bias_ptr, n_batch, n_cell, scratch4);
+      layer_norm_output_scale_b, output_gate_bias_ptr, n_batch, n_cell,
+      scratch4);
 
   // Output gate sigmoid.
   tensor_utils::ApplySigmoidFloat(scratch4, n_batch, n_cell, scratch4);
@@ -1512,7 +1514,7 @@ TfLiteStatus EvalFloat(
     const TfLiteTensor* aux_input_to_cell_weights,
     const TfLiteTensor* aux_input_to_output_weights,
     const TfLiteTensor* input_gate_bias, const TfLiteTensor* forget_gate_bias,
-    const TfLiteTensor* cell_bias, const TfLiteTensor* output_gate_bias,
+    const TfLiteTensor* cell_gate_bias, const TfLiteTensor* output_gate_bias,
     const TfLiteTensor* projection_weights, const TfLiteTensor* projection_bias,
     const TfLiteLSTMParams* params, bool forward_sequence, bool time_major,
     int output_offset, TfLiteTensor* scratch_buffer, TfLiteTensor* output_state,
@@ -1595,7 +1597,7 @@ TfLiteStatus EvalFloat(
           GetTensorData(output_layer_norm_coefficients),
           GetTensorData(input_gate_bias),
           GetTensorData(forget_gate_bias),
-          GetTensorData(cell_bias),
+          GetTensorData(cell_gate_bias),
           GetTensorData(output_gate_bias),
           GetTensorData(projection_weights),
           GetTensorData(projection_bias), params, n_batch, n_cell,
@@ -1656,7 +1658,7 @@ TfLiteStatus EvalFloat(
             GetTensorData(output_layer_norm_coefficients),
             GetTensorData(input_gate_bias),
             GetTensorData(forget_gate_bias),
-            GetTensorData(cell_bias),
+            GetTensorData(cell_gate_bias),
             GetTensorData(output_gate_bias),
             GetTensorData(projection_weights),
             GetTensorData(projection_bias), params, /*n_batch=*/1,
@@ -1693,7 +1695,7 @@ TfLiteStatus EvalHybrid(
     const TfLiteTensor* aux_input_to_cell_weights,
     const TfLiteTensor* aux_input_to_output_weights,
     const TfLiteTensor* input_gate_bias, const TfLiteTensor* forget_gate_bias,
-    const TfLiteTensor* cell_bias, const TfLiteTensor* output_gate_bias,
+    const TfLiteTensor* cell_gate_bias, const TfLiteTensor* output_gate_bias,
     const TfLiteTensor* projection_weights, const TfLiteTensor* projection_bias,
     const TfLiteLSTMParams* params, bool forward_sequence, bool time_major,
     int output_offset, TfLiteTensor* scratch_buffer,
@@ -1802,7 +1804,7 @@ TfLiteStatus EvalHybrid(
           GetTensorData(output_layer_norm_coefficients),
           GetTensorData(input_gate_bias),
           GetTensorData(forget_gate_bias),
-          GetTensorData(cell_bias),
+          GetTensorData(cell_gate_bias),
           GetTensorData(output_gate_bias),
           GetTensorData(projection_weights),
           GetTensorScale(projection_weights),
@@ -1888,7 +1890,7 @@ TfLiteStatus EvalHybrid(
             GetTensorData(output_layer_norm_coefficients),
             GetTensorData(input_gate_bias),
             GetTensorData(forget_gate_bias),
-            GetTensorData(cell_bias),
+            GetTensorData(cell_gate_bias),
             GetTensorData(output_gate_bias),
             GetTensorData(projection_weights),
             GetTensorScale(projection_weights),
@@ -1930,7 +1932,7 @@ TfLiteStatus EvalInteger8x8_16(
     const TfLiteTensor* cell_layer_norm_coefficients,
     const TfLiteTensor* output_layer_norm_coefficients,
     const TfLiteTensor* input_gate_bias, const TfLiteTensor* forget_gate_bias,
-    const TfLiteTensor* cell_bias, const TfLiteTensor* output_gate_bias,
+    const TfLiteTensor* cell_gate_bias, const TfLiteTensor* output_gate_bias,
     const TfLiteTensor* projection_weights, const TfLiteTensor* projection_bias,
     const TfLiteLSTMParams* params,
     const lstm_eval::IntegerLstmParameter* integer_lstm_param,
@@ -2020,7 +2022,7 @@ TfLiteStatus EvalInteger8x8_16(
         integer_lstm_param->layer_norm_output_scale_b,
         GetTensorData(input_gate_bias),
         GetTensorData(forget_gate_bias),
-        GetTensorData(cell_bias),
+        GetTensorData(cell_gate_bias),
         GetTensorData(output_gate_bias),
         integer_lstm_param->quantized_cell_clip,
         integer_lstm_param->quantized_proj_clip, integer_lstm_param->cell_scale,
@@ -2065,7 +2067,7 @@ TfLiteStatus EvalInteger8x8_8(
     const TfLiteTensor* cell_layer_norm_coefficients,
     const TfLiteTensor* output_layer_norm_coefficients,
     const TfLiteTensor* input_gate_bias, const TfLiteTensor* forget_gate_bias,
-    const TfLiteTensor* cell_bias, const TfLiteTensor* output_gate_bias,
+    const TfLiteTensor* cell_gate_bias, const TfLiteTensor* output_gate_bias,
     const TfLiteTensor* projection_weights, const TfLiteTensor* projection_bias,
     const TfLiteLSTMParams* params, TfLiteTensor* output_state,
     TfLiteTensor* cell_state, TfLiteTensor* output,
@@ -2120,10 +2122,12 @@ TfLiteStatus EvalInteger8x8_8(
       GetTensorData(cell_layer_norm_coefficients);
   const int16_t* layer_norm_output_weight_ptr =
       GetTensorData(output_layer_norm_coefficients);
-  const int32_t* input_bias_ptr = GetTensorData(input_gate_bias);
-  const int32_t* forget_bias_ptr = GetTensorData(forget_gate_bias);
-  const int32_t* cell_gate_bias_ptr = GetTensorData(cell_bias);
-  const int32_t* output_bias_ptr = GetTensorData(output_gate_bias);
+  const int32_t* input_gate_bias_ptr = GetTensorData(input_gate_bias);
+  const int32_t* forget_gate_bias_ptr =
+      GetTensorData(forget_gate_bias);
+  const int32_t* cell_gate_bias_ptr = GetTensorData(cell_gate_bias);
+  const int32_t* output_gate_bias_ptr =
+      GetTensorData(output_gate_bias);
   const int32_t* proj_bias_ptr = GetTensorData(projection_bias);
   int16_t* cell_ptr = GetTensorData(cell_state);
   int8_t* output_state_ptr = GetTensorData(output_state);
@@ -2209,8 +2213,8 @@ TfLiteStatus EvalInteger8x8_8(
         integer_lstm_param->layer_norm_output_scale_a,
         integer_lstm_param->layer_norm_output_scale_b,
 
-        input_bias_ptr, forget_bias_ptr, cell_gate_bias_ptr, output_bias_ptr,
-        proj_bias_ptr,
+        input_gate_bias_ptr, forget_gate_bias_ptr, cell_gate_bias_ptr,
+        output_gate_bias_ptr, proj_bias_ptr,
 
         params, integer_lstm_param->intermediate_scale_a,
         integer_lstm_param->intermediate_scale_b,
diff --git a/tensorflow/lite/kernels/lstm_eval.h b/tensorflow/lite/kernels/lstm_eval.h
index 3c9b4bccf42..9b3bd0c54ec 100644
--- a/tensorflow/lite/kernels/lstm_eval.h
+++ b/tensorflow/lite/kernels/lstm_eval.h
@@ -117,7 +117,7 @@ TfLiteStatus EvalFloat(
     const TfLiteTensor* aux_input_to_cell_weights,
     const TfLiteTensor* aux_input_to_output_weights,
     const TfLiteTensor* input_gate_bias, const TfLiteTensor* forget_gate_bias,
-    const TfLiteTensor* cell_bias, const TfLiteTensor* output_gate_bias,
+    const TfLiteTensor* cell_gate_bias, const TfLiteTensor* output_gate_bias,
     const TfLiteTensor* projection_weights, const TfLiteTensor* projection_bias,
     const TfLiteLSTMParams* params, bool forward_sequence, bool time_major,
     int output_offset, TfLiteTensor* scratch_buffer, TfLiteTensor* output_state,
@@ -145,7 +145,7 @@ TfLiteStatus EvalHybrid(
     const TfLiteTensor* aux_input_to_cell_weights,
     const TfLiteTensor* aux_input_to_output_weights,
     const TfLiteTensor* input_gate_bias, const TfLiteTensor* forget_gate_bias,
-    const TfLiteTensor* cell_bias, const TfLiteTensor* output_gate_bias,
+    const TfLiteTensor* cell_gate_bias, const TfLiteTensor* output_gate_bias,
     const TfLiteTensor* projection_weights, const TfLiteTensor* projection_bias,
     const TfLiteLSTMParams* params, bool forward_sequence, bool time_major,
     int output_offset, TfLiteTensor* scratch_buffer,
@@ -174,7 +174,7 @@ TfLiteStatus EvalInteger8x8_16(
     const TfLiteTensor* cell_layer_norm_coefficients,
     const TfLiteTensor* output_layer_norm_coefficients,
     const TfLiteTensor* input_gate_bias, const TfLiteTensor* forget_gate_bias,
-    const TfLiteTensor* cell_bias, const TfLiteTensor* output_gate_bias,
+    const TfLiteTensor* cell_gate_bias, const TfLiteTensor* output_gate_bias,
     const TfLiteTensor* projection_weights, const TfLiteTensor* projection_bias,
     const TfLiteLSTMParams* params,
     const lstm_eval::IntegerLstmParameter* integer_lstm_param,
@@ -200,7 +200,7 @@ TfLiteStatus EvalInteger8x8_8(
     const TfLiteTensor* cell_layer_norm_coefficients,
     const TfLiteTensor* output_layer_norm_coefficients,
     const TfLiteTensor* input_gate_bias, const TfLiteTensor* forget_gate_bias,
-    const TfLiteTensor* cell_bias, const TfLiteTensor* output_gate_bias,
+    const TfLiteTensor* cell_gate_bias, const TfLiteTensor* output_gate_bias,
     const TfLiteTensor* projection_weights, const TfLiteTensor* projection_bias,
     const TfLiteLSTMParams* params, TfLiteTensor* output_state,
     TfLiteTensor* cell_state, TfLiteTensor* output,
diff --git a/tensorflow/lite/kernels/lstm_eval_test.cc b/tensorflow/lite/kernels/lstm_eval_test.cc
index baf2e5e83df..78459117859 100644
--- a/tensorflow/lite/kernels/lstm_eval_test.cc
+++ b/tensorflow/lite/kernels/lstm_eval_test.cc
@@ -113,10 +113,10 @@ class BaseLstmParam {
     TfLiteIntArrayFree(layer_norm_forget_tensor_.dims);
     TfLiteIntArrayFree(layer_norm_cell_tensor_.dims);
     TfLiteIntArrayFree(layer_norm_output_tensor_.dims);
-    TfLiteIntArrayFree(input_bias_tensor_.dims);
-    TfLiteIntArrayFree(forget_bias_tensor_.dims);
-    TfLiteIntArrayFree(cell_bias_tensor_.dims);
-    TfLiteIntArrayFree(output_bias_tensor_.dims);
+    TfLiteIntArrayFree(input_gate_bias_tensor_.dims);
+    TfLiteIntArrayFree(forget_gate_bias_tensor_.dims);
+    TfLiteIntArrayFree(cell_gate_bias_tensor_.dims);
+    TfLiteIntArrayFree(output_gate_bias_tensor_.dims);
     TfLiteIntArrayFree(projection_tensor_.dims);
     TfLiteIntArrayFree(projection_bias_tensor_.dims);
     TfLiteIntArrayFree(activation_tensor_.dims);
@@ -275,17 +275,17 @@ class BaseLstmParam {
   std::vector layer_norm_output_size_ = {n_cell_};
   TfLiteTensor layer_norm_output_tensor_;
 
-  std::vector input_bias_size_ = {n_cell_};
-  TfLiteTensor input_bias_tensor_;
+  std::vector input_gate_bias_size_ = {n_cell_};
+  TfLiteTensor input_gate_bias_tensor_;
 
-  std::vector forget_bias_size_ = {n_cell_};
-  TfLiteTensor forget_bias_tensor_;
+  std::vector forget_gate_bias_size_ = {n_cell_};
+  TfLiteTensor forget_gate_bias_tensor_;
 
-  std::vector cell_bias_size_ = {n_cell_};
-  TfLiteTensor cell_bias_tensor_;
+  std::vector cell_gate_bias_size_ = {n_cell_};
+  TfLiteTensor cell_gate_bias_tensor_;
 
-  std::vector output_bias_size_ = {n_cell_};
-  TfLiteTensor output_bias_tensor_;
+  std::vector output_gate_bias_size_ = {n_cell_};
+  TfLiteTensor output_gate_bias_tensor_;
 
   // projection_weights.
   std::vector projection_ = {
@@ -350,24 +350,28 @@ class QuantizedLstmParam : public BaseLstmParam {
     return &layer_norm_output_tensor_;
   }
   TfLiteTensor* GetInputBias() {
-    PackWeightToTensor(&input_bias_tensor_, input_bias_, input_bias_size_);
-    input_bias_tensor_.data.i32 = input_bias_.data();
-    return &input_bias_tensor_;
+    PackWeightToTensor(&input_gate_bias_tensor_, input_gate_bias_,
+                       input_gate_bias_size_);
+    input_gate_bias_tensor_.data.i32 = input_gate_bias_.data();
+    return &input_gate_bias_tensor_;
   }
   TfLiteTensor* GetForgetBias() {
-    PackWeightToTensor(&forget_bias_tensor_, forget_bias_, forget_bias_size_);
-    forget_bias_tensor_.data.i32 = forget_bias_.data();
-    return &forget_bias_tensor_;
+    PackWeightToTensor(&forget_gate_bias_tensor_, forget_gate_bias_,
+                       forget_gate_bias_size_);
+    forget_gate_bias_tensor_.data.i32 = forget_gate_bias_.data();
+    return &forget_gate_bias_tensor_;
   }
   TfLiteTensor* GetCellBias() {
-    PackWeightToTensor(&cell_bias_tensor_, cell_bias_, cell_bias_size_);
-    cell_bias_tensor_.data.i32 = cell_bias_.data();
-    return &cell_bias_tensor_;
+    PackWeightToTensor(&cell_gate_bias_tensor_, cell_gate_bias_,
+                       cell_gate_bias_size_);
+    cell_gate_bias_tensor_.data.i32 = cell_gate_bias_.data();
+    return &cell_gate_bias_tensor_;
   }
   TfLiteTensor* GetOutputBias() {
-    PackWeightToTensor(&output_bias_tensor_, output_bias_, output_bias_size_);
-    output_bias_tensor_.data.i32 = output_bias_.data();
-    return &output_bias_tensor_;
+    PackWeightToTensor(&output_gate_bias_tensor_, output_gate_bias_,
+                       output_gate_bias_size_);
+    output_gate_bias_tensor_.data.i32 = output_gate_bias_.data();
+    return &output_gate_bias_tensor_;
   }
   TfLiteTensor* GetProjectionBias() {
     PackWeightToTensor(&projection_bias_tensor_, projection_bias_,
@@ -539,22 +543,22 @@ class QuantizedLstmParam : public BaseLstmParam {
   };
 
   // input_gate_bias.
-  std::vector input_bias_ = {
+  std::vector input_gate_bias_ = {
       16, 4, 5, 6, 1, 1, 3, 4, -5, 6,  //
   };
 
   // forget_gate_bias.
-  std::vector forget_bias_ = {
+  std::vector forget_gate_bias_ = {
       16, 4, 5, 6, 1, 1, 3, 4, -5, 6,  //
   };
 
-  // cell_bias.
-  std::vector cell_bias_ = {
+  // cell_gate_bias.
+  std::vector cell_gate_bias_ = {
       16, 4, 5, 6, 1, 1, 3, 4, -5, 6,  //
   };
 
   // output_gate_bias.
-  std::vector output_bias_ = {
+  std::vector output_gate_bias_ = {
       16, 4, 5, 6, 1, 1, 3, 4, -5, 6,  //
   };
 
@@ -711,27 +715,28 @@ class HybridLstmParam : public BaseLstmParam {
     return &accum_scratch_tensor_;
   }
   TfLiteTensor* GetInputBias() {
-    PackWeightToTensor(&input_bias_tensor_, input_float_bias_,
-                       input_bias_size_);
-    input_bias_tensor_.data.f = input_float_bias_.data();
-    return &input_bias_tensor_;
+    PackWeightToTensor(&input_gate_bias_tensor_, input_float_bias_,
+                       input_gate_bias_size_);
+    input_gate_bias_tensor_.data.f = input_float_bias_.data();
+    return &input_gate_bias_tensor_;
   }
   TfLiteTensor* GetForgetBias() {
-    PackWeightToTensor(&forget_bias_tensor_, forget_float_bias_,
-                       forget_bias_size_);
-    forget_bias_tensor_.data.f = forget_float_bias_.data();
-    return &forget_bias_tensor_;
+    PackWeightToTensor(&forget_gate_bias_tensor_, forget_float_bias_,
+                       forget_gate_bias_size_);
+    forget_gate_bias_tensor_.data.f = forget_float_bias_.data();
+    return &forget_gate_bias_tensor_;
   }
   TfLiteTensor* GetCellBias() {
-    PackWeightToTensor(&cell_bias_tensor_, cell_float_bias_, cell_bias_size_);
-    cell_bias_tensor_.data.f = cell_float_bias_.data();
-    return &cell_bias_tensor_;
+    PackWeightToTensor(&cell_gate_bias_tensor_, cell_float_bias_,
+                       cell_gate_bias_size_);
+    cell_gate_bias_tensor_.data.f = cell_float_bias_.data();
+    return &cell_gate_bias_tensor_;
   }
   TfLiteTensor* GetOutputBias() {
-    PackWeightToTensor(&output_bias_tensor_, output_float_bias_,
-                       output_bias_size_);
-    output_bias_tensor_.data.f = output_float_bias_.data();
-    return &output_bias_tensor_;
+    PackWeightToTensor(&output_gate_bias_tensor_, output_float_bias_,
+                       output_gate_bias_size_);
+    output_gate_bias_tensor_.data.f = output_float_bias_.data();
+    return &output_gate_bias_tensor_;
   }
   TfLiteTensor* GetProjectionBias() {
     PackWeightToTensor(&projection_bias_tensor_, projection_float_bias_,
diff --git a/tensorflow/lite/kernels/lstm_test.cc b/tensorflow/lite/kernels/lstm_test.cc
index f8594f9adf0..a9023dce371 100644
--- a/tensorflow/lite/kernels/lstm_test.cc
+++ b/tensorflow/lite/kernels/lstm_test.cc
@@ -89,7 +89,7 @@ class LSTMOpModel : public SingleOpModel {
       input_gate_bias_ = AddInput(TensorType_FLOAT32);
     }
     forget_gate_bias_ = AddInput(TensorType_FLOAT32);
-    cell_bias_ = AddInput(TensorType_FLOAT32);
+    cell_gate_bias_ = AddInput(TensorType_FLOAT32);
     output_gate_bias_ = AddInput(TensorType_FLOAT32);
 
     if (use_projection_weights) {
@@ -211,7 +211,7 @@ class LSTMOpModel : public SingleOpModel {
   }
 
   void SetCellBias(const std::vector& f) {
-    PopulateTensor(cell_bias_, f);
+    PopulateTensor(cell_gate_bias_, f);
   }
 
   void SetOutputGateBias(const std::vector& f) {
@@ -261,7 +261,7 @@ class LSTMOpModel : public SingleOpModel {
 
   int input_gate_bias_;
   int forget_gate_bias_;
-  int cell_bias_;
+  int cell_gate_bias_;
   int output_gate_bias_;
 
   int projection_weights_;
@@ -498,7 +498,7 @@ TEST_F(NoCifgNoPeepholeNoProjectionNoClippingLstmTest, LstmBlackBoxTest) {
 
                        {n_cell},  // input_gate_bias tensor
                        {n_cell},  // forget_gate_bias tensor
-                       {n_cell},  // cell_bias tensor
+                       {n_cell},  // cell_gate_bias tensor
                        {n_cell},  // output_gate_bias tensor
 
                        {0, 0},  // projection_weight tensor
@@ -545,7 +545,7 @@ TEST_F(NoCifgNoPeepholeNoProjectionNoClippingOmittedLayerNormLstmTest,
 
                        {n_cell},  // input_gate_bias tensor
                        {n_cell},  // forget_gate_bias tensor
-                       {n_cell},  // cell_bias tensor
+                       {n_cell},  // cell_gate_bias tensor
                        {n_cell},  // output_gate_bias tensor
 
                        {0, 0},  // projection_weight tensor
@@ -601,7 +601,7 @@ TEST_P(NoCifgNoPeepholeNoProjectionNoClippingLstmTest,
 
                        {n_cell},  // input_gate_bias tensor
                        {n_cell},  // forget_gate_bias tensor
-                       {n_cell},  // cell_bias tensor
+                       {n_cell},  // cell_gate_bias tensor
                        {n_cell},  // output_gate_bias tensor
 
                        {0, 0},  // projection_weight tensor
@@ -652,7 +652,7 @@ TEST_P(NoCifgNoPeepholeNoProjectionNoClippingLstmInt8Test,
 
                        {n_cell},  // input_gate_bias tensor
                        {n_cell},  // forget_gate_bias tensor
-                       {n_cell},  // cell_bias tensor
+                       {n_cell},  // cell_gate_bias tensor
                        {n_cell},  // output_gate_bias tensor
 
                        {0, 0},  // projection_weight tensor
@@ -743,7 +743,7 @@ TEST_F(CifgNoPeepholeNoProjectionNoClippingLstmTest, LstmBlackBoxTest) {
 
                        {0},       // input_gate_bias tensor
                        {n_cell},  // forget_gate_bias tensor
-                       {n_cell},  // cell_bias tensor
+                       {n_cell},  // cell_gate_bias tensor
                        {n_cell},  // output_gate_bias tensor
 
                        {0, 0},  // projection_weight tensor
@@ -791,7 +791,7 @@ TEST_P(CifgNoPeepholeNoProjectionNoClippingLstmTest,
 
                        {0},       // input_gate_bias tensor
                        {n_cell},  // forget_gate_bias tensor
-                       {n_cell},  // cell_bias tensor
+                       {n_cell},  // cell_gate_bias tensor
                        {n_cell},  // output_gate_bias tensor
 
                        {0, 0},  // projection_weight tensor
@@ -840,7 +840,7 @@ TEST_P(CifgNoPeepholeNoProjectionNoClippingLstmInt8Test,
 
                        {0},       // input_gate_bias tensor
                        {n_cell},  // forget_gate_bias tensor
-                       {n_cell},  // cell_bias tensor
+                       {n_cell},  // cell_gate_bias tensor
                        {n_cell},  // output_gate_bias tensor
 
                        {0, 0},  // projection_weight tensor
@@ -1481,7 +1481,7 @@ TEST_F(NoCifgPeepholeProjectionNoClippingLstmTest, LstmBlackBoxTest) {
 
                        {n_cell},  // input_gate_bias tensor
                        {n_cell},  // forget_gate_bias tensor
-                       {n_cell},  // cell_bias tensor
+                       {n_cell},  // cell_gate_bias tensor
                        {n_cell},  // output_gate_bias tensor
 
                        {n_output, n_cell},  // projection_weight tensor
@@ -1528,7 +1528,7 @@ TEST_P(NoCifgPeepholeProjectionNoClippingLstmTest,
 
                        {n_cell},  // input_gate_bias tensor
                        {n_cell},  // forget_gate_bias tensor
-                       {n_cell},  // cell_bias tensor
+                       {n_cell},  // cell_gate_bias tensor
                        {n_cell},  // output_gate_bias tensor
 
                        {n_output, n_cell},  // projection_weight tensor
@@ -1577,7 +1577,7 @@ TEST_P(NoCifgPeepholeProjectionNoClippingLstmInt8Test,
 
                        {n_cell},  // input_gate_bias tensor
                        {n_cell},  // forget_gate_bias tensor
-                       {n_cell},  // cell_bias tensor
+                       {n_cell},  // cell_gate_bias tensor
                        {n_cell},  // output_gate_bias tensor
 
                        {n_output, n_cell},  // projection_weight tensor
@@ -1689,7 +1689,7 @@ TEST_F(NoCifgPeepholeProjectionNoClippingLayerNormLstmTest,
 
           {n_cell},  // input_gate_bias tensor
           {n_cell},  // forget_gate_bias tensor
-          {n_cell},  // cell_bias tensor
+          {n_cell},  // cell_gate_bias tensor
           {n_cell},  // output_gate_bias tensor
 
           {n_output, n_cell},  // projection_weight tensor
@@ -1760,7 +1760,7 @@ TEST_P(NoCifgPeepholeProjectionNoClippingLayerNormLstmTest,
 
           {n_cell},  // input_gate_bias tensor
           {n_cell},  // forget_gate_bias tensor
-          {n_cell},  // cell_bias tensor
+          {n_cell},  // cell_gate_bias tensor
           {n_cell},  // output_gate_bias tensor
 
           {n_output, n_cell},  // projection_weight tensor
@@ -1833,7 +1833,7 @@ TEST_P(NoCifgPeepholeProjectionNoClippingLayerNormLstmInt8Test,
 
           {n_cell},  // input_gate_bias tensor
           {n_cell},  // forget_gate_bias tensor
-          {n_cell},  // cell_bias tensor
+          {n_cell},  // cell_gate_bias tensor
           {n_cell},  // output_gate_bias tensor
 
           {n_output, n_cell},  // projection_weight tensor
@@ -1947,7 +1947,7 @@ TEST_F(CifgPeepholeProjectionNoClippingLayerNormLstmTest,
 
           {0},       // input_gate_bias tensor
           {n_cell},  // forget_gate_bias tensor
-          {n_cell},  // cell_bias tensor
+          {n_cell},  // cell_gate_bias tensor
           {n_cell},  // output_gate_bias tensor
 
           {n_output, n_cell},  // projection_weight tensor
@@ -2018,7 +2018,7 @@ TEST_P(CifgPeepholeProjectionNoClippingLayerNormLstmTest,
 
           {0},       // input_gate_bias tensor
           {n_cell},  // forget_gate_bias tensor
-          {n_cell},  // cell_bias tensor
+          {n_cell},  // cell_gate_bias tensor
           {n_cell},  // output_gate_bias tensor
 
           {n_output, n_cell},  // projection_weight tensor
@@ -2090,7 +2090,7 @@ TEST_P(CifgPeepholeProjectionNoClippingLayerNormLstmInt8Test,
 
           {0},       // input_gate_bias tensor
           {n_cell},  // forget_gate_bias tensor
-          {n_cell},  // cell_bias tensor
+          {n_cell},  // cell_gate_bias tensor
           {n_cell},  // output_gate_bias tensor
 
           {n_output, n_cell},  // projection_weight tensor
@@ -2195,8 +2195,8 @@ class LSTMIntegerOpModel : public SingleOpModel {
     }
     forget_gate_bias_ = AddInput({TensorType_INT32, input_shapes[13],
                                   ranges[13].first, ranges[13].second});
-    cell_bias_ = AddInput({TensorType_INT32, input_shapes[14], ranges[14].first,
-                           ranges[14].second});
+    cell_gate_bias_ = AddInput({TensorType_INT32, input_shapes[14],
+                                ranges[14].first, ranges[14].second});
     output_gate_bias_ = AddInput({TensorType_INT32, input_shapes[15],
                                   ranges[15].first, ranges[15].second});
 
@@ -2330,7 +2330,7 @@ class LSTMIntegerOpModel : public SingleOpModel {
   }
 
   void SetCellBias(const std::vector& f) {
-    QuantizeAndPopulate(cell_bias_, f);
+    QuantizeAndPopulate(cell_gate_bias_, f);
   }
 
   void SetOutputGateBias(const std::vector& f) {
@@ -2379,7 +2379,7 @@ class LSTMIntegerOpModel : public SingleOpModel {
 
   int input_gate_bias_;
   int forget_gate_bias_;
-  int cell_bias_;
+  int cell_gate_bias_;
   int output_gate_bias_;
 
   int projection_weights_;
@@ -2473,7 +2473,7 @@ TEST(LSTMIntegerOpModel, NoCifgYesLayerNormNoYesProjectionNoPeephole) {
 
       {n_cell},  // input_gate_bias tensor
       {n_cell},  // forget_gate_bias tensor
-      {n_cell},  // cell_bias tensor
+      {n_cell},  // cell_gate_bias tensor
       {n_cell},  // output_gate_bias tensor
 
       {n_output, n_cell},  // projection_weight tensor
@@ -2507,7 +2507,7 @@ TEST(LSTMIntegerOpModel, NoCifgYesLayerNormNoYesProjectionNoPeephole) {
 
       {-100, 100},  // input_gate_bias tensor
       {-100, 100},  // forget_gate_bias tensor
-      {-100, 100},  // cell_bias tensor
+      {-100, 100},  // cell_gate_bias tensor
       {-100, 100},  // output_gate_bias tensor
 
       {-0.5, 0.5},  // projection_weight tensor
@@ -2675,7 +2675,7 @@ TEST(LSTMIntegerOpModel, NoCifgYesLayerNormNoYesProjectionYesPeephole) {
 
       {n_cell},  // input_gate_bias tensor
       {n_cell},  // forget_gate_bias tensor
-      {n_cell},  // cell_bias tensor
+      {n_cell},  // cell_gate_bias tensor
       {n_cell},  // output_gate_bias tensor
 
       {n_output, n_cell},  // projection_weight tensor
@@ -2709,7 +2709,7 @@ TEST(LSTMIntegerOpModel, NoCifgYesLayerNormNoYesProjectionYesPeephole) {
 
       {-100, 100},  // input_gate_bias tensor
       {-100, 80},   // forget_gate_bias tensor
-      {-100, 100},  // cell_bias tensor
+      {-100, 100},  // cell_gate_bias tensor
       {-100, 100},  // output_gate_bias tensor
 
       {-0.5, 0.5},  // projection_weight tensor
@@ -2869,8 +2869,8 @@ class LSTMIntegerOpModel8x8_8 : public SingleOpModel {
     }
     forget_gate_bias_ = AddInput({TensorType_INT32, input_shapes[13],
                                   ranges[13].first, ranges[13].second});
-    cell_bias_ = AddInput({TensorType_INT32, input_shapes[14], ranges[14].first,
-                           ranges[14].second});
+    cell_gate_bias_ = AddInput({TensorType_INT32, input_shapes[14],
+                                ranges[14].first, ranges[14].second});
     output_gate_bias_ = AddInput({TensorType_INT32, input_shapes[15],
                                   ranges[15].first, ranges[15].second});
 
@@ -3004,7 +3004,7 @@ class LSTMIntegerOpModel8x8_8 : public SingleOpModel {
   }
 
   void SetCellBias(const std::vector& f) {
-    QuantizeAndPopulate(cell_bias_, f);
+    QuantizeAndPopulate(cell_gate_bias_, f);
   }
 
   void SetOutputGateBias(const std::vector& f) {
@@ -3053,7 +3053,7 @@ class LSTMIntegerOpModel8x8_8 : public SingleOpModel {
 
   int input_gate_bias_;
   int forget_gate_bias_;
-  int cell_bias_;
+  int cell_gate_bias_;
   int output_gate_bias_;
 
   int projection_weights_;
@@ -3148,7 +3148,7 @@ TEST(LSTMIntegerOpModel8x8_8, CifgYesLayerNormNoYesProjectionNoPeephole) {
 
       {0},       // input_gate_bias tensor
       {n_cell},  // forget_gate_bias tensor
-      {n_cell},  // cell_bias tensor
+      {n_cell},  // cell_gate_bias tensor
       {n_cell},  // output_gate_bias tensor
 
       {n_output, n_cell},  // projection_weight tensor
@@ -3182,7 +3182,7 @@ TEST(LSTMIntegerOpModel8x8_8, CifgYesLayerNormNoYesProjectionNoPeephole) {
 
       {-100, 100},  // input_gate_bias tensor
       {-100, 100},  // forget_gate_bias tensor
-      {-100, 100},  // cell_bias tensor
+      {-100, 100},  // cell_gate_bias tensor
       {-100, 100},  // output_gate_bias tensor
 
       {-0.5, 0.5},  // projection_weight tensor
@@ -3303,7 +3303,7 @@ TEST(LSTMOpModel, InvalidTypeTest) {
 
                        {n_cell},  // input_gate_bias tensor
                        {n_cell},  // forget_gate_bias tensor
-                       {n_cell},  // cell_bias tensor
+                       {n_cell},  // cell_gate_bias tensor
                        {n_cell},  // output_gate_bias tensor
 
                        {0, 0},  // projection_weight tensor
@@ -3338,7 +3338,7 @@ TEST(LSTMOpModel, InvalidTypeTest) {
 
                        {n_cell},  // input_gate_bias tensor
                        {n_cell},  // forget_gate_bias tensor
-                       {n_cell},  // cell_bias tensor
+                       {n_cell},  // cell_gate_bias tensor
                        {n_cell},  // output_gate_bias tensor
 
                        {0, 0},  // projection_weight tensor
diff --git a/tensorflow/lite/kernels/optional_tensor_test.cc b/tensorflow/lite/kernels/optional_tensor_test.cc
index 26d619276aa..9e83c74da8d 100644
--- a/tensorflow/lite/kernels/optional_tensor_test.cc
+++ b/tensorflow/lite/kernels/optional_tensor_test.cc
@@ -78,7 +78,7 @@ class LSTMOpModel : public SingleOpModel {
       input_gate_bias_ = AddInput(TensorType_FLOAT32);
     }
     forget_gate_bias_ = AddInput(TensorType_FLOAT32);
-    cell_bias_ = AddInput(TensorType_FLOAT32);
+    cell_gate_bias_ = AddInput(TensorType_FLOAT32);
     output_gate_bias_ = AddInput(TensorType_FLOAT32);
 
     if (use_projection_weights) {
@@ -161,7 +161,7 @@ class LSTMOpModel : public SingleOpModel {
   }
 
   void SetCellBias(std::initializer_list f) {
-    PopulateTensor(cell_bias_, f);
+    PopulateTensor(cell_gate_bias_, f);
   }
 
   void SetOutputGateBias(std::initializer_list f) {
@@ -209,7 +209,7 @@ class LSTMOpModel : public SingleOpModel {
 
   int input_gate_bias_;
   int forget_gate_bias_;
-  int cell_bias_;
+  int cell_gate_bias_;
   int output_gate_bias_;
 
   int projection_weights_;
@@ -256,7 +256,7 @@ TEST(LSTMOpTest, BlackBoxTestWithCifgWithPeepholeNoProjectionNoClipping) {
 
                        {0},       // input_gate_bias tensor
                        {n_cell},  // forget_gate_bias tensor
-                       {n_cell},  // cell_bias tensor
+                       {n_cell},  // cell_gate_bias tensor
                        {n_cell},  // output_gate_bias tensor
 
                        {0, 0},  // projection_weight tensor
diff --git a/tensorflow/lite/kernels/unidirectional_sequence_lstm.cc b/tensorflow/lite/kernels/unidirectional_sequence_lstm.cc
index f1c0f9d42a6..0b2cba72369 100644
--- a/tensorflow/lite/kernels/unidirectional_sequence_lstm.cc
+++ b/tensorflow/lite/kernels/unidirectional_sequence_lstm.cc
@@ -179,10 +179,10 @@ TfLiteStatus CheckInputTensorDimensions(TfLiteContext* context,
   TF_LITE_ENSURE_EQ(context, forget_gate_bias->dims->size, 1);
   TF_LITE_ENSURE_EQ(context, forget_gate_bias->dims->data[0], n_cell);
 
-  const TfLiteTensor* cell_bias =
+  const TfLiteTensor* cell_gate_bias =
       GetInput(context, node, lstm::full::kCellGateBiasTensor);
-  TF_LITE_ENSURE_EQ(context, cell_bias->dims->size, 1);
-  TF_LITE_ENSURE_EQ(context, cell_bias->dims->data[0], n_cell);
+  TF_LITE_ENSURE_EQ(context, cell_gate_bias->dims->size, 1);
+  TF_LITE_ENSURE_EQ(context, cell_gate_bias->dims->data[0], n_cell);
 
   const TfLiteTensor* output_gate_bias =
       GetInput(context, node, lstm::full::kOutputGateBiasTensor);
@@ -546,7 +546,7 @@ TfLiteStatus Eval(TfLiteContext* context, TfLiteNode* node) {
       GetOptionalInputTensor(context, node, lstm::full::kInputGateBiasTensor);
   const TfLiteTensor* forget_gate_bias =
       GetInput(context, node, lstm::full::kForgetGateBiasTensor);
-  const TfLiteTensor* cell_bias =
+  const TfLiteTensor* cell_gate_bias =
       GetInput(context, node, lstm::full::kCellGateBiasTensor);
   const TfLiteTensor* output_gate_bias =
       GetInput(context, node, lstm::full::kOutputGateBiasTensor);
@@ -611,8 +611,9 @@ TfLiteStatus Eval(TfLiteContext* context, TfLiteNode* node) {
           /*aux_input_to_forget_weights=*/nullptr,
           /*aux_input_to_cell_weights=*/nullptr,
           /*aux_input_to_output_weights=*/nullptr, input_gate_bias,
-          forget_gate_bias, cell_bias, output_gate_bias, projection_weights,
-          projection_bias, &lstm_params, /*forward_sequence=*/true, time_major,
+          forget_gate_bias, cell_gate_bias, output_gate_bias,
+          projection_weights, projection_bias, &lstm_params,
+          /*forward_sequence=*/true, time_major,
           /*output_offset=*/0, scratch_buffer, output_state, cell_state,
           output);
     }
@@ -648,8 +649,9 @@ TfLiteStatus Eval(TfLiteContext* context, TfLiteNode* node) {
           /*aux_input_to_forget_weights=*/nullptr,
           /*aux_input_to_cell_weights=*/nullptr,
           /*aux_input_to_output_weights=*/nullptr, input_gate_bias,
-          forget_gate_bias, cell_bias, output_gate_bias, projection_weights,
-          projection_bias, &lstm_params, /*forward_sequence=*/true, time_major,
+          forget_gate_bias, cell_gate_bias, output_gate_bias,
+          projection_weights, projection_bias, &lstm_params,
+          /*forward_sequence=*/true, time_major,
           /*output_offset=*/0, scratch_buffer, scaling_factors,
           prod_scaling_factors, recovered_cell_weights, input_quantized,
           /*aux_input_quantized=*/nullptr, output_state_quantized,
diff --git a/tensorflow/lite/kernels/unidirectional_sequence_lstm_test.cc b/tensorflow/lite/kernels/unidirectional_sequence_lstm_test.cc
index ec20d76ae2e..74584ec9e85 100644
--- a/tensorflow/lite/kernels/unidirectional_sequence_lstm_test.cc
+++ b/tensorflow/lite/kernels/unidirectional_sequence_lstm_test.cc
@@ -85,7 +85,7 @@ class UnidirectionalLSTMOpModel : public SingleOpModel {
       input_gate_bias_ = AddInput(TensorType_FLOAT32);
     }
     forget_gate_bias_ = AddInput(TensorType_FLOAT32);
-    cell_bias_ = AddInput(TensorType_FLOAT32);
+    cell_gate_bias_ = AddInput(TensorType_FLOAT32);
     output_gate_bias_ = AddInput(TensorType_FLOAT32);
 
     if (use_projection_weights) {
@@ -187,7 +187,7 @@ class UnidirectionalLSTMOpModel : public SingleOpModel {
   }
 
   void SetCellBias(const std::vector& f) {
-    PopulateTensor(cell_bias_, f);
+    PopulateTensor(cell_gate_bias_, f);
   }
 
   void SetOutputGateBias(const std::vector& f) {
@@ -249,7 +249,7 @@ class UnidirectionalLSTMOpModel : public SingleOpModel {
 
   int input_gate_bias_;
   int forget_gate_bias_;
-  int cell_bias_;
+  int cell_gate_bias_;
   int output_gate_bias_;
 
   int projection_weights_;
@@ -530,7 +530,7 @@ TEST_F(NoCifgNoPeepholeNoProjectionNoClippingUnidirectionalLstmTest,
 
           {n_cell},  // input_gate_bias tensor
           {n_cell},  // forget_gate_bias tensor
-          {n_cell},  // cell_bias tensor
+          {n_cell},  // cell_gate_bias tensor
           {n_cell},  // output_gate_bias tensor
 
           {0, 0},  // projection_weight tensor
@@ -592,7 +592,7 @@ TEST_F(NoCifgNoPeepholeNoProjectionNoClippingUnidirectionalLstmTest,
 
           {n_cell},  // input_gate_bias tensor
           {n_cell},  // forget_gate_bias tensor
-          {n_cell},  // cell_bias tensor
+          {n_cell},  // cell_gate_bias tensor
           {n_cell},  // output_gate_bias tensor
 
           {0, 0},  // projection_weight tensor
@@ -658,7 +658,7 @@ TEST_P(NoCifgNoPeepholeNoProjectionNoClippingUnidirectionalLstmTest,
 
           {n_cell},  // input_gate_bias tensor
           {n_cell},  // forget_gate_bias tensor
-          {n_cell},  // cell_bias tensor
+          {n_cell},  // cell_gate_bias tensor
           {n_cell},  // output_gate_bias tensor
 
           {0, 0},  // projection_weight tensor
@@ -721,7 +721,7 @@ TEST_P(NoCifgNoPeepholeNoProjectionNoClippingUnidirectionalLstmTest,
 
           {n_cell},  // input_gate_bias tensor
           {n_cell},  // forget_gate_bias tensor
-          {n_cell},  // cell_bias tensor
+          {n_cell},  // cell_gate_bias tensor
           {n_cell},  // output_gate_bias tensor
 
           {0, 0},  // projection_weight tensor
@@ -833,7 +833,7 @@ TEST_F(CifgPeepholeNoProjectionNoClippingUnidirectionalLstmTest,
 
           {0},       // input_gate_bias tensor
           {n_cell},  // forget_gate_bias tensor
-          {n_cell},  // cell_bias tensor
+          {n_cell},  // cell_gate_bias tensor
           {n_cell},  // output_gate_bias tensor
 
           {0, 0},  // projection_weight tensor
@@ -894,7 +894,7 @@ TEST_P(CifgPeepholeNoProjectionNoClippingUnidirectionalLstmTest,
 
           {0},       // input_gate_bias tensor
           {n_cell},  // forget_gate_bias tensor
-          {n_cell},  // cell_bias tensor
+          {n_cell},  // cell_gate_bias tensor
           {n_cell},  // output_gate_bias tensor
 
           {0, 0},  // projection_weight tensor
@@ -957,7 +957,7 @@ TEST_P(CifgPeepholeNoProjectionNoClippingUnidirectionalLstmTest,
 
           {0},       // input_gate_bias tensor
           {n_cell},  // forget_gate_bias tensor
-          {n_cell},  // cell_bias tensor
+          {n_cell},  // cell_gate_bias tensor
           {n_cell},  // output_gate_bias tensor
 
           {0, 0},  // projection_weight tensor
@@ -1619,7 +1619,7 @@ TEST_F(NoCifgPeepholeProjectionClippingUnidirectionalLstmTest,
 
           {n_cell},  // input_gate_bias tensor
           {n_cell},  // forget_gate_bias tensor
-          {n_cell},  // cell_bias tensor
+          {n_cell},  // cell_gate_bias tensor
           {n_cell},  // output_gate_bias tensor
 
           {n_output, n_cell},  // projection_weight tensor
@@ -1688,7 +1688,7 @@ TEST_P(NoCifgPeepholeProjectionClippingUnidirectionalLstmTest,
 
           {n_cell},  // input_gate_bias tensor
           {n_cell},  // forget_gate_bias tensor
-          {n_cell},  // cell_bias tensor
+          {n_cell},  // cell_gate_bias tensor
           {n_cell},  // output_gate_bias tensor
 
           {n_output, n_cell},  // projection_weight tensor
@@ -1759,7 +1759,7 @@ TEST_P(NoCifgPeepholeProjectionClippingUnidirectionalLstmTest,
 
           {n_cell},  // input_gate_bias tensor
           {n_cell},  // forget_gate_bias tensor
-          {n_cell},  // cell_bias tensor
+          {n_cell},  // cell_gate_bias tensor
           {n_cell},  // output_gate_bias tensor
 
           {n_output, n_cell},  // projection_weight tensor
@@ -2430,7 +2430,7 @@ TEST_F(NoCifgPeepholeProjectionAndBiasClippingUnidirectionalLstmTest,
 
           {n_cell},  // input_gate_bias tensor
           {n_cell},  // forget_gate_bias tensor
-          {n_cell},  // cell_bias tensor
+          {n_cell},  // cell_gate_bias tensor
           {n_cell},  // output_gate_bias tensor
 
           {n_output, n_cell},  // projection_weight tensor
@@ -2636,7 +2636,7 @@ TEST_F(CifgPeepholeNoProjectionNoClippingLayerNormUnidirectionalLstmTest,
 
           {0},       // input_gate_bias tensor
           {n_cell},  // forget_gate_bias tensor
-          {n_cell},  // cell_bias tensor
+          {n_cell},  // cell_gate_bias tensor
           {n_cell},  // output_gate_bias tensor
 
           {0, 0},  // projection_weight tensor
@@ -2707,7 +2707,7 @@ TEST_F(CifgPeepholeNoProjectionNoClippingUnidirectionalLstmTest,
 
           {0},       // input_gate_bias tensor
           {n_cell},  // forget_gate_bias tensor
-          {n_cell},  // cell_bias tensor
+          {n_cell},  // cell_gate_bias tensor
           {n_cell},  // output_gate_bias tensor
 
           {0, 0},  // projection_weight tensor
diff --git a/tensorflow/lite/tools/optimize/calibration/builtin_logging_ops/lstm.cc b/tensorflow/lite/tools/optimize/calibration/builtin_logging_ops/lstm.cc
index 88ea7c1d591..50138442c25 100644
--- a/tensorflow/lite/tools/optimize/calibration/builtin_logging_ops/lstm.cc
+++ b/tensorflow/lite/tools/optimize/calibration/builtin_logging_ops/lstm.cc
@@ -299,7 +299,7 @@ TfLiteStatus EvalFloat(
     const TfLiteTensor* aux_input_to_cell_weights,
     const TfLiteTensor* aux_input_to_output_weights,
     const TfLiteTensor* input_gate_bias, const TfLiteTensor* forget_gate_bias,
-    const TfLiteTensor* cell_bias, const TfLiteTensor* output_gate_bias,
+    const TfLiteTensor* cell_gate_bias, const TfLiteTensor* output_gate_bias,
     const TfLiteTensor* projection_weights, const TfLiteTensor* projection_bias,
     const TfLiteLSTMParams* params, bool forward_sequence, bool time_major,
     int output_offset, TfLiteTensor* scratch_buffer, TfLiteTensor* output_state,
@@ -384,7 +384,7 @@ TfLiteStatus EvalFloat(
           GetTensorData(output_layer_norm_coefficients),
           GetTensorData(input_gate_bias),
           GetTensorData(forget_gate_bias),
-          GetTensorData(cell_bias),
+          GetTensorData(cell_gate_bias),
           GetTensorData(output_gate_bias),
           GetTensorData(projection_weights),
           GetTensorData(projection_bias), params, n_batch, n_cell,
@@ -446,7 +446,7 @@ TfLiteStatus EvalFloat(
             GetTensorData(output_layer_norm_coefficients),
             GetTensorData(input_gate_bias),
             GetTensorData(forget_gate_bias),
-            GetTensorData(cell_bias),
+            GetTensorData(cell_gate_bias),
             GetTensorData(output_gate_bias),
             GetTensorData(projection_weights),
             GetTensorData(projection_bias), params, /*n_batch=*/1,
@@ -527,7 +527,7 @@ TfLiteStatus lstm_eval(TfLiteContext* context, TfLiteNode* node, Logger* logger,
       context, node, ops::builtin::lstm::full::kInputGateBiasTensor);
   const TfLiteTensor* forget_gate_bias =
       GetInput(context, node, ops::builtin::lstm::full::kForgetGateBiasTensor);
-  const TfLiteTensor* cell_bias =
+  const TfLiteTensor* cell_gate_bias =
       GetInput(context, node, ops::builtin::lstm::full::kCellGateBiasTensor);
   const TfLiteTensor* output_gate_bias =
       GetInput(context, node, ops::builtin::lstm::full::kOutputGateBiasTensor);
@@ -570,8 +570,9 @@ TfLiteStatus lstm_eval(TfLiteContext* context, TfLiteNode* node, Logger* logger,
           /*aux_input_to_forget_weights=*/nullptr,
           /*aux_input_to_cell_weights=*/nullptr,
           /*aux_input_to_output_weights=*/nullptr, input_gate_bias,
-          forget_gate_bias, cell_bias, output_gate_bias, projection_weights,
-          projection_bias, params, /*forward_sequence=*/true,
+          forget_gate_bias, cell_gate_bias, output_gate_bias,
+          projection_weights, projection_bias, params,
+          /*forward_sequence=*/true,
           /*time_major=*/true,
           /*output_offset=*/0, scratch_buffer, output_state, cell_state, output,
           logger, intermediate_tensor_indexes, error_reporter);

From 7cd6c3115badaf79fff2b8809cb4a0b49a5f9c7c Mon Sep 17 00:00:00 2001
From: Denisa Roberts 
Date: Fri, 19 Jun 2020 14:26:51 -0400
Subject: [PATCH 315/867] Allow gradient access to QR input

---
 tensorflow/python/eager/pywrap_gradient_exclusions.cc | 3 +--
 1 file changed, 1 insertion(+), 2 deletions(-)

diff --git a/tensorflow/python/eager/pywrap_gradient_exclusions.cc b/tensorflow/python/eager/pywrap_gradient_exclusions.cc
index 7da45e36118..7e9f0b16334 100644
--- a/tensorflow/python/eager/pywrap_gradient_exclusions.cc
+++ b/tensorflow/python/eager/pywrap_gradient_exclusions.cc
@@ -50,7 +50,7 @@ auto OpGradientInfoInit(const T &a) {
 
 absl::optional> OpGradientUnusedInputIndices(
     const tensorflow::string &op_name) {
-  static std::array a = {{
+  static std::array a = {{
       {"Acosh"},
       {"AllToAll", 1, {0}},
       {"ApproximateEqual"},
@@ -222,7 +222,6 @@ absl::optional> OpGradientUnusedInputIndices(
       {"PlaceholderWithDefault"},
       {"PopulationCount"},
       {"PreventGradient"},
-      {"Qr"},
       {"QuantizeAndDequantize"},
       {"QuantizeAndDequantizeV2"},
       {"QuantizeAndDequantizeV3"},

From 9aea03e98d2ef5416dff09f721d6d519ea83e74f Mon Sep 17 00:00:00 2001
From: "A. Unique TensorFlower" 
Date: Fri, 19 Jun 2020 11:39:56 -0700
Subject: [PATCH 316/867] Add StatefulRandom op to TF MLIR.

PiperOrigin-RevId: 317347849
Change-Id: I19b7b1b3157d065cc556cce7aba72b79ed9b776f
---
 .../mlir/tensorflow/ir/tf_generated_ops.td    | 26 +++++++++++++++++++
 1 file changed, 26 insertions(+)

diff --git a/tensorflow/compiler/mlir/tensorflow/ir/tf_generated_ops.td b/tensorflow/compiler/mlir/tensorflow/ir/tf_generated_ops.td
index 3b1f3eec699..dcef99e6971 100644
--- a/tensorflow/compiler/mlir/tensorflow/ir/tf_generated_ops.td
+++ b/tensorflow/compiler/mlir/tensorflow/ir/tf_generated_ops.td
@@ -9190,6 +9190,32 @@ def TF_StackV2Op : TF_Op<"StackV2", []> {
   );
 }
 
+def TF_StatelessRandomUniformOp : TF_Op<"StatelessRandomUniform", [NoSideEffect]> {
+  let summary = [{
+Outputs deterministic pseudorandom random values from a uniform distribution.
+  }];
+
+  let description = [{
+The generated values follow a uniform distribution in the range `[0, 1)`. The
+lower bound 0 is included in the range, while the upper bound 1 is excluded.
+
+The outputs are a deterministic function of `shape` and `seed`.
+  }];
+
+  let arguments = (ins
+    TF_I32OrI64Tensor:$shape,
+    TF_I32OrI64Tensor:$seed
+  );
+
+  let results = (outs
+    TF_FpTensor:$output
+  );
+
+  TF_DerivedOperandTypeAttr T = TF_DerivedOperandTypeAttr<0>;
+  TF_DerivedOperandTypeAttr Tseed = TF_DerivedOperandTypeAttr<1>;
+  TF_DerivedResultTypeAttr dtype = TF_DerivedResultTypeAttr<0>;
+}
+
 def TF_StopGradientOp : TF_Op<"StopGradient", [NoSideEffect, TF_AllTypesMatch<["input", "output"]>]> {
   let summary = "Stops gradient computation.";
 

From 02d312b25edc2d8999640d7fbe2a833c2eaab0d8 Mon Sep 17 00:00:00 2001
From: Ran Chen 
Date: Fri, 19 Jun 2020 11:44:48 -0700
Subject: [PATCH 317/867] *** Reason for rollback ***

break pip test

PiperOrigin-RevId: 317348830
Change-Id: Ia6968b2cd3bf001f6afbc4487234d88123d3d0a8
---
 .../python/distribute/multi_process_runner.py | 205 ++----------------
 .../distribute/multi_process_runner_test.py   |  69 +-----
 2 files changed, 25 insertions(+), 249 deletions(-)

diff --git a/tensorflow/python/distribute/multi_process_runner.py b/tensorflow/python/distribute/multi_process_runner.py
index ce36287a9da..8699e59b410 100644
--- a/tensorflow/python/distribute/multi_process_runner.py
+++ b/tensorflow/python/distribute/multi_process_runner.py
@@ -67,7 +67,8 @@ except ImportError:
 # exception stack trace info is stored in exc_info to pass on to parent process
 # to be re-raised.
 _ProcessStatusInfo = collections.namedtuple(
-    '_ProcessStatusInfo', ['is_successful', 'exc_info', 'return_value'])
+    '_ProcessStatusInfo',
+    ['task_type', 'is_successful', 'exc_info', 'return_value'])
 
 # Information returned from a successful MultiProcessRunner run.
 MultiProcessRunnerResult = collections.namedtuple('MultiProcessRunnerResult',
@@ -123,7 +124,6 @@ class MultiProcessRunner(object):
                stream_stdout=True,
                list_stdout=False,
                use_dill_for_args=True,
-               daemon=False,
                args=None,
                kwargs=None):
     """Creates a multi-process runner.
@@ -157,7 +157,6 @@ class MultiProcessRunner(object):
       use_dill_for_args: Whether to use dill to pickle `args` and `kwargs`. dill
         can pickle more objects, but doesn't work with types in
         `multiprocessing` library like `Mutex`.
-      daemon: Whether to start processes as daemons.
       args: Positional arguments to be sent to functions run on processes.
       kwargs: Keyword arguments to be sent to functions run on processes.
 
@@ -189,7 +188,6 @@ class MultiProcessRunner(object):
     self._list_stdout = list_stdout
     self._dependence_on_chief = True
     self._use_dill_for_args = use_dill_for_args
-    self._daemon = daemon
     self._args = args or ()
     self._kwargs = kwargs or {}
 
@@ -270,8 +268,7 @@ class MultiProcessRunner(object):
         test_env=test_env,
         target=_ProcFunc(),
         args=(resources, test_env, proc_func, args, kwargs,
-              self._use_dill_for_args),
-        daemon=self._daemon)
+              self._use_dill_for_args))
     p.start()
     self._processes[(task_type, task_id)] = p
     self._outstanding_subprocess_count += 1
@@ -571,6 +568,7 @@ class _ProcFunc(object):
         time.sleep(0.1)
     self._resources.process_status_queue.put(
         _ProcessStatusInfo(
+            task_type=task_type,
             is_successful=True,
             exc_info=None,
             return_value=None))
@@ -630,9 +628,17 @@ class _ProcFunc(object):
     if test_env.v2_enabled:
       v2_compat.enable_v2_behavior()
 
-    with self._runtime_mode(test_env.executing_eagerly):
-      info = _run_contained(proc_func, args, kwargs)
-      self._resources.process_status_queue.put(info)
+    try:
+      with self._runtime_mode(test_env.executing_eagerly):
+        return_value = proc_func(*args, **kwargs)
+        is_successful = True
+        exc_info = None
+
+    except Exception:  # pylint: disable=broad-except
+      # Capture all exceptions to be reported to parent process.
+      return_value = None
+      is_successful = False
+      exc_info = sys.exc_info()
 
       # Re-raise the exception in addition to reporting it to the parent
       # process, so that even if `--test_timeout` flag is set and the
@@ -641,181 +647,16 @@ class _ProcFunc(object):
       # instead of silently suppressing the error due to early bazel
       # timeout. Raising an error in the subprocess produces stack trace in
       # the log, but the program continues running.
-      if not info.is_successful:
-        six.reraise(*info.exc_info)
+      raise
 
-      self._close_streaming()
-
-
-class MultiProcessPoolRunner(object):
-  """A utility class to start a process pool to simulate a cluster.
-
-  It's similar to MultiProcessRunner, but uses a pool of processes to avoid the
-  expensive initialization cost of Tensorflow.
-  """
-
-  def __init__(self, cluster_spec, initializer=None):
-    """Creates a multi-process pool runner.
-
-    Args:
-      cluster_spec: Dict for cluster spec. The following is an example of
-        cluster with three workers.
-        {"worker": ["worker0.example.com:2222",
-                    "worker1.example.com:2222",
-                    "worker2.example.com:2222"]}
-      initializer: a callable to called at the startup of worker processes.
-
-    Raises:
-      RuntimeError: if `multi_process_runner.test_main()` is not called.
-      ValueError: if there are more than one chief in the `cluster_spec`.
-    """
-    self._cluster_spec = cluster_spec
-    self._initializer = initializer
-    self._conn = {}
-    self._runner = None
-
-  def __del__(self):
-    self._reset()
-
-  def _reset(self):
-    for conn in self._conn.values():
-      conn.close()
-    self._conn = {}
-    if self._runner is not None:
-      self._runner.join()
-      self._runner = None
-
-  def _start(self):
-    """Starts the worker pool."""
-    # We need different arguments for different processes so we're passing a
-    # no-op proc_func here and use start_single_process instead.
-    #
-    # We also need to start the process pool as daemon, so that they don't block
-    # the program from exiting. Note that __del__ may not get called when
-    # there's an exception. The user may also store a pool runner in a global
-    # object to share across test cases
-    self._runner = MultiProcessRunner(
-        proc_func=lambda: None,
-        cluster_spec=self._cluster_spec,
-        use_dill_for_args=False,
-        daemon=True)
-    if self._initializer:
-      initializer = dill.dumps(self._initializer, dill.HIGHEST_PROTOCOL)
-    else:
-      initializer = None
-    for task_type, addresses in self._cluster_spec.items():
-      for task_id, _ in enumerate(addresses):
-        conn1, conn2 = multiprocessing.Pipe(duplex=True)
-        self._conn[(task_type, task_id)] = conn1
-        self._runner.start_single_process(
-            task_type,
-            task_id,
-            proc_func=_pool_runner_worker,
-            args=(initializer, conn2))
-
-  def run(self, proc_func, args=None, kwargs=None):
-    """Runs `proc_func` with `args` and `kwargs` on all jobs.
-
-    Args:
-      proc_func: The function to be run.
-      args: Optional positional arguments to be supplied in `proc_func`.
-      kwargs: Optional keyword arguments to be supplied in `proc_func`.
-
-    Returns:
-      A list of return values.
-    """
-    if self._runner is None:
-      self._start()
-
-    # Since we start the processes as daemon they're going to be killed by
-    # SIGTERM when the program exits. We only turn on streaming during run() to
-    # avoid printing the stacktrace caused by the SIGTERM.
-    self._runner._stream_stdout = True  # pylint: disable=protected-access
-
-    try:
-      proc_func = dill.dumps(proc_func, dill.HIGHEST_PROTOCOL)
-      for conn in self._conn.values():
-        conn.send((proc_func, args or [], kwargs or {}))
-
-      process_statuses = []
-      for (task_type, task_id), conn in self._conn.items():
-        logging.info('Waiting for the result from %s-%d', task_type, task_id)
-        try:
-          process_statuses.append(conn.recv())
-        except EOFError:
-          # This shouldn't happen due to exceptions in proc_func. This usually
-          # means bugs in the runner.
-          self._reset()
-          raise RuntimeError('Unexpected EOF. Worker process may have died. '
-                             'Please report a bug')
-
-      return_values = []
-      for process_status in process_statuses:
-        assert isinstance(process_status, _ProcessStatusInfo)
-        if not process_status.is_successful:
-          six.reraise(*process_status.exc_info)
-        if process_status.return_value is not None:
-          return_values.append(process_status.return_value)
-
-      return return_values
     finally:
-      self._runner._stream_stdout = False  # pylint: disable=protected-access
-
-
-def _pool_runner_worker(initializer, conn):
-  """Function that runs on the workers in a pool.
-
-  It listens for callables to run and returns the result until `conn` is closed.
-  It captures the exceptions during executing the callable and return it through
-  `conn`.
-
-  Args:
-    initializer: A callable to execute during startup.
-    conn: A multiprocessing.Connection object to listen for tasks and send
-      results.
-  """
-  if initializer:
-    initializer = dill.loads(initializer)
-    initializer()
-  while True:
-    try:
-      proc_func, args, kwargs = conn.recv()
-    except EOFError:
-      break
-    proc_func = dill.loads(proc_func)
-    info = _run_contained(proc_func, args, kwargs)
-    sys.stdout.flush()
-    sys.stderr.flush()
-    conn.send(info)
-
-
-def _run_contained(proc_func, args, kwargs):
-  """Runs `proc_func` with `args` and `kwargs`.
-
-  The function returns _ProcessStatusInfo which captures the return value and
-  the exception.
-
-  Args:
-    proc_func: The function to be run.
-    args: Optional positional arguments to be supplied in `proc_func`.
-    kwargs: Optional keyword arguments to be supplied in `proc_func`.
-
-  Returns:
-    a _ProcessStatusInfo.
-  """
-  try:
-    return_value = proc_func(*args, **kwargs)
-    is_successful = True
-    exc_info = None
-  except Exception:  # pylint: disable=broad-except
-    return_value = None
-    is_successful = False
-    exc_info = sys.exc_info()
-  finally:
-    return _ProcessStatusInfo(  # pylint: disable=lost-exception
-        is_successful=is_successful,
-        exc_info=exc_info,
-        return_value=return_value)
+      info = _ProcessStatusInfo(
+          task_type=test_env.task_type,
+          is_successful=is_successful,
+          exc_info=exc_info,
+          return_value=return_value)
+      self._resources.process_status_queue.put(info)
+      self._close_streaming()
 
 
 class SubprocessTimeoutError(RuntimeError):
diff --git a/tensorflow/python/distribute/multi_process_runner_test.py b/tensorflow/python/distribute/multi_process_runner_test.py
index d76ef5a5a3c..aeba43b6b7c 100644
--- a/tensorflow/python/distribute/multi_process_runner_test.py
+++ b/tensorflow/python/distribute/multi_process_runner_test.py
@@ -22,8 +22,6 @@ import json
 import os
 import threading
 import time
-import unittest
-
 from absl import logging
 
 from tensorflow.python.distribute import multi_process_runner
@@ -47,7 +45,7 @@ def proc_func_that_adds_simple_return_data():
   return 'dummy_data'
 
 
-def proc_func_that_returns_args_and_kwargs(*args, **kwargs):
+def proc_func_that_return_args_and_kwargs(*args, **kwargs):
   return list(args) + list(kwargs.items())
 
 
@@ -55,20 +53,6 @@ def proc_func_with_barrier():
   return multi_process_runner.barrier()
 
 
-def proc_func_that_returns_pid():
-  return os.getpid()
-
-
-V = None
-
-
-def proc_func_that_sets_global(val):
-  global V
-  old_val = V
-  V = val
-  return old_val
-
-
 class MultiProcessRunnerTest(test.TestCase):
 
   def _worker_idx(self):
@@ -111,7 +95,7 @@ class MultiProcessRunnerTest(test.TestCase):
 
   def test_multi_process_runner_args_passed_correctly(self):
     return_value = multi_process_runner.run(
-        proc_func_that_returns_args_and_kwargs,
+        proc_func_that_return_args_and_kwargs,
         multi_worker_test_base.create_cluster_spec(num_workers=1),
         args=('a', 'b'),
         kwargs={
@@ -341,54 +325,5 @@ class MultiProcessRunnerTest(test.TestCase):
                 for line in list_to_assert))
 
 
-class MultiProcessPoolRunnerTest(test.TestCase):
-
-  def test_same_process_across_runs(self):
-    cluster_spec = multi_worker_test_base.create_cluster_spec(num_workers=2)
-    runner = multi_process_runner.MultiProcessPoolRunner(cluster_spec)
-    pid = runner.run(proc_func_that_returns_pid)
-    for _ in range(3):
-      self.assertAllEqual(runner.run(proc_func_that_returns_pid), pid)
-
-  def test_exceptions_in_sub_process(self):
-    cluster_spec = multi_worker_test_base.create_cluster_spec(num_workers=2)
-    runner = multi_process_runner.MultiProcessPoolRunner(cluster_spec)
-    pid = runner.run(proc_func_that_returns_pid)
-    with self.assertRaisesRegexp(ValueError, 'This is an error.'):
-      runner.run(proc_func_that_errors)
-    self.assertAllEqual(runner.run(proc_func_that_returns_pid), pid)
-
-  def test_tf_config(self):
-    cluster_spec = multi_worker_test_base.create_cluster_spec(
-        has_chief=True, num_workers=2)
-    runner = multi_process_runner.MultiProcessPoolRunner(cluster_spec)
-    result = runner.run(proc_func_that_adds_task_type_in_return_data)
-
-    job_count_dict = {'worker': 2, 'chief': 1}
-    for data in result:
-      job_count_dict[data] -= 1
-
-    self.assertEqual(job_count_dict['worker'], 0)
-    self.assertEqual(job_count_dict['chief'], 0)
-
-  @unittest.expectedFailure
-  def test_exception_in_main_process(self):
-    # When there's an exception in the main process, __del__() is not called.
-    # This test is to verify MultiProcessPoolRunner can cope with __del__() not
-    # being called.
-    cluster_spec = multi_worker_test_base.create_cluster_spec(
-        has_chief=True, num_workers=2)
-    runner = multi_process_runner.MultiProcessPoolRunner(cluster_spec)
-    runner.run(proc_func_that_returns_pid)
-    raise ValueError('failure')
-
-  def test_initializer(self):
-    cluster_spec = multi_worker_test_base.create_cluster_spec(num_workers=2)
-    runner = multi_process_runner.MultiProcessPoolRunner(
-        cluster_spec, initializer=lambda: proc_func_that_sets_global(1))
-    result = runner.run(proc_func_that_sets_global, args=(2,))
-    self.assertAllEqual(result, [1, 1])
-
-
 if __name__ == '__main__':
   multi_process_runner.test_main()

From 2417a15bf186cde76b601a53324ebbc2c9193124 Mon Sep 17 00:00:00 2001
From: "A. Unique TensorFlower" 
Date: Fri, 19 Jun 2020 11:46:26 -0700
Subject: [PATCH 318/867] Clean up post fusion using DCE.

PiperOrigin-RevId: 317349120
Change-Id: I479d9967323d86e924315d2b1302bafd01ed4151
---
 tensorflow/compiler/xla/service/BUILD                  | 1 +
 tensorflow/compiler/xla/service/multi_output_fusion.cc | 5 +++++
 2 files changed, 6 insertions(+)

diff --git a/tensorflow/compiler/xla/service/BUILD b/tensorflow/compiler/xla/service/BUILD
index acd35cbc153..2fd457e8e47 100644
--- a/tensorflow/compiler/xla/service/BUILD
+++ b/tensorflow/compiler/xla/service/BUILD
@@ -1679,6 +1679,7 @@ cc_library(
     hdrs = ["multi_output_fusion.h"],
     deps = [
         ":hlo",
+        ":hlo_dce",
         ":hlo_pass",
         ":hlo_reachability",
         "//tensorflow/compiler/xla:debug_options_flags",
diff --git a/tensorflow/compiler/xla/service/multi_output_fusion.cc b/tensorflow/compiler/xla/service/multi_output_fusion.cc
index b95b27d6291..a21cec538d1 100644
--- a/tensorflow/compiler/xla/service/multi_output_fusion.cc
+++ b/tensorflow/compiler/xla/service/multi_output_fusion.cc
@@ -17,6 +17,7 @@ limitations under the License.
 
 #include "absl/container/flat_hash_set.h"
 #include "tensorflow/compiler/xla/debug_options_flags.h"
+#include "tensorflow/compiler/xla/service/hlo_dce.h"
 #include "tensorflow/compiler/xla/service/hlo_instruction.h"
 #include "tensorflow/compiler/xla/service/hlo_opcode.h"
 #include "tensorflow/compiler/xla/service/hlo_reachability.h"
@@ -126,6 +127,10 @@ StatusOr MultiOutputFusion::Run(HloModule* module) {
   candidates_index_.clear();
   all_fusion_candidates_.clear();
   reachability_.reset();
+  if (changed) {
+    HloDCE dce;
+    TF_RETURN_IF_ERROR(dce.Run(module).status());
+  }
   return changed;
 }
 

From f805153a25b00d12072bd728e91bb1621bfcf1b1 Mon Sep 17 00:00:00 2001
From: "A. Unique TensorFlower" 
Date: Fri, 19 Jun 2020 11:48:22 -0700
Subject: [PATCH 319/867] Fix regression that omitted -DNDEBUG in opt builds
 after toolchain refactoring.

Not adding -DNDEBUG leads to enabling assertions which makes LLVM often 10x slower.

PiperOrigin-RevId: 317349500
Change-Id: I52df6ab5013ad5a02101dd96f18220e054c3e94c
---
 third_party/gpus/crosstool/cc_toolchain_config.bzl.tpl | 6 +++++-
 1 file changed, 5 insertions(+), 1 deletion(-)

diff --git a/third_party/gpus/crosstool/cc_toolchain_config.bzl.tpl b/third_party/gpus/crosstool/cc_toolchain_config.bzl.tpl
index 082ed950b04..eb320a94201 100644
--- a/third_party/gpus/crosstool/cc_toolchain_config.bzl.tpl
+++ b/third_party/gpus/crosstool/cc_toolchain_config.bzl.tpl
@@ -583,7 +583,11 @@ def _features(cpu, compiler, ctx):
                     ),
                 ],
             ),
-            feature(name = "opt"),
+            feature(name = "disable-assertions"),
+            feature(
+                name = "opt",
+                implies = ["disable-assertions"],
+            ),
             feature(name = "fastbuild"),
             feature(name = "dbg"),
             feature(name = "supports_dynamic_linker", enabled = True),

From f53e1aac6576b3f5e575f870da1fbdb286094444 Mon Sep 17 00:00:00 2001
From: Lu Wang 
Date: Fri, 19 Jun 2020 11:50:15 -0700
Subject: [PATCH 320/867] Stamp the version number for all metadata components
 simultaneously

PiperOrigin-RevId: 317349859
Change-Id: Ica912c3fb310889185c026e6d73ce4c69a9f0505
---
 .../lite/experimental/support/metadata/BUILD  | 24 ++++++++++-
 .../support/metadata/build_defs.bzl           | 43 +++++++++++++++++++
 .../experimental/support/metadata/cc/BUILD    | 13 +++++-
 .../metadata/cc/metadata_parser.h.template    | 28 ++++++++++++
 .../support/metadata/cc/test/BUILD            |  9 ++++
 .../metadata/cc/test/metadata_parser_test.cc  | 33 ++++++++++++++
 .../metadata/cc/test/metadata_version_test.cc |  2 +-
 .../experimental/support/metadata/java/BUILD  |  8 +++-
 .../support/metadata/MetadataExtractor.java   |  8 +---
 .../lite/support/metadata/MetadataParser.java | 27 ++++++++++++
 .../metadata/metadata_parser.py.template      | 26 +++++++++++
 .../support/metadata/metadata_parser_test.py  | 38 ++++++++++++++++
 .../support/metadata/metadata_schema.fbs      | 29 ++++++++-----
 13 files changed, 267 insertions(+), 21 deletions(-)
 create mode 100644 tensorflow/lite/experimental/support/metadata/build_defs.bzl
 create mode 100644 tensorflow/lite/experimental/support/metadata/cc/metadata_parser.h.template
 create mode 100644 tensorflow/lite/experimental/support/metadata/cc/test/metadata_parser_test.cc
 create mode 100644 tensorflow/lite/experimental/support/metadata/java/src/java/org/tensorflow/lite/support/metadata/MetadataParser.java
 create mode 100644 tensorflow/lite/experimental/support/metadata/metadata_parser.py.template
 create mode 100644 tensorflow/lite/experimental/support/metadata/metadata_parser_test.py

diff --git a/tensorflow/lite/experimental/support/metadata/BUILD b/tensorflow/lite/experimental/support/metadata/BUILD
index 4621c8c55d2..ba410d914c7 100644
--- a/tensorflow/lite/experimental/support/metadata/BUILD
+++ b/tensorflow/lite/experimental/support/metadata/BUILD
@@ -1,5 +1,6 @@
 load("//tensorflow:tensorflow.bzl", "py_test")
 load("@flatbuffers//:build_defs.bzl", "flatbuffer_android_library", "flatbuffer_cc_library", "flatbuffer_java_library", "flatbuffer_py_library")
+load("//tensorflow/lite/experimental/support/metadata:build_defs.bzl", "stamp_metadata_parser_version")
 
 package(
     default_visibility = [
@@ -51,9 +52,19 @@ flatbuffer_android_library(
     custom_package = "org.tensorflow.lite.support.metadata.schema",
 )
 
+# TODO(b/157813075): move the metadata python library to metadata/python/ when migrating to the new repo.
+stamp_metadata_parser_version(
+    name = "metadata_parser_py",
+    srcs = ["metadata_parser.py.template"],
+    outs = ["metadata_parser.py"],
+)
+
 py_library(
     name = "metadata",
-    srcs = ["metadata.py"],
+    srcs = [
+        "metadata.py",
+        ":metadata_parser_py",
+    ],
     data = [
         "//tensorflow/lite/experimental/support/metadata:metadata_schema.fbs",
     ],
@@ -89,3 +100,14 @@ py_test(
         "@six_archive//:six",
     ],
 )
+
+py_test(
+    name = "metadata_parser_test",
+    srcs = ["metadata_parser_test.py"],
+    python_version = "PY3",
+    srcs_version = "PY2AND3",
+    deps = [
+        ":metadata",
+        "//tensorflow/python:client_testlib",
+    ],
+)
diff --git a/tensorflow/lite/experimental/support/metadata/build_defs.bzl b/tensorflow/lite/experimental/support/metadata/build_defs.bzl
new file mode 100644
index 00000000000..3ea945770e0
--- /dev/null
+++ b/tensorflow/lite/experimental/support/metadata/build_defs.bzl
@@ -0,0 +1,43 @@
+"""Build rules to generate metadata schema versions."""
+
+METADATA_SCHEMA_FILE = "//tensorflow/lite/experimental/support/metadata:metadata_schema.fbs"
+
+def stamp_metadata_parser_version(
+        name,
+        srcs,
+        outs):
+    """Stamps the latest metadata parser version into the srcs files.
+
+    Replaces all the occurrences of "{LATEST_METADATA_PARSER_VERSION}" in the
+    srcs files with the metadata schema version extracted from
+    METADATA_SCHEMA_FILE and then outputs the generated file into outs,
+    respectively. The number of srcs files needs to match the number of outs
+    files.
+
+    Args:
+        name: Rule name. (required)
+        srcs: List of source files. (required)
+        outs: List of output files. (required)
+    """
+    if len(srcs) != len(outs):
+        fail(("The number of srcs files (%d) does not match that of the outs" +
+              " files (%d).") %
+             (len(srcs), len(outs)))
+
+    for i in range(0, len(srcs)):
+        native.genrule(
+            name = "%s_file%d" % (name, i),
+            srcs = [srcs[i]],
+            outs = [outs[i]],
+            tools = [METADATA_SCHEMA_FILE],
+            # Gets the metadata schema version from the file, and stamps it
+            # into the srcs file.
+            cmd = "version=$$(sed -n -e '/Schema Semantic version/ s/.*\\: *//p' $(location %s));" %
+                  METADATA_SCHEMA_FILE +
+                  'sed "s/{LATEST_METADATA_PARSER_VERSION}/$$version/" $< > $@',
+        )
+
+    native.filegroup(
+        name = name,
+        srcs = outs,
+    )
diff --git a/tensorflow/lite/experimental/support/metadata/cc/BUILD b/tensorflow/lite/experimental/support/metadata/cc/BUILD
index 832e2edb56d..8febc7a2237 100644
--- a/tensorflow/lite/experimental/support/metadata/cc/BUILD
+++ b/tensorflow/lite/experimental/support/metadata/cc/BUILD
@@ -1,12 +1,23 @@
+load("//tensorflow/lite/experimental/support/metadata:build_defs.bzl", "stamp_metadata_parser_version")
+
 package(
     default_visibility = ["//tensorflow/lite/experimental/support:users"],
     licenses = ["notice"],  # Apache 2.0
 )
 
+stamp_metadata_parser_version(
+    name = "metadata_parser_h",
+    srcs = ["metadata_parser.h.template"],
+    outs = ["metadata_parser.h"],
+)
+
 cc_library(
     name = "metadata_version",
     srcs = ["metadata_version.cc"],
-    hdrs = ["metadata_version.h"],
+    hdrs = [
+        "metadata_version.h",
+        ":metadata_parser_h",
+    ],
     deps = [
         "//tensorflow/lite/c:common",
         "//tensorflow/lite/experimental/support/metadata:metadata_schema_cc",
diff --git a/tensorflow/lite/experimental/support/metadata/cc/metadata_parser.h.template b/tensorflow/lite/experimental/support/metadata/cc/metadata_parser.h.template
new file mode 100644
index 00000000000..dfb62d0de81
--- /dev/null
+++ b/tensorflow/lite/experimental/support/metadata/cc/metadata_parser.h.template
@@ -0,0 +1,28 @@
+/* Copyright 2020 The TensorFlow Authors. All Rights Reserved.
+
+Licensed under the Apache License, Version 2.0 (the "License");
+you may not use this file except in compliance with the License.
+You may obtain a copy of the License at
+
+    http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software
+distributed under the License is distributed on an "AS IS" BASIS,
+WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+See the License for the specific language governing permissions and
+limitations under the License.
+==============================================================================*/
+#ifndef TENSORFLOW_LITE_EXPERIMENTAL_SUPPORT_METADATA_CC_METADATA_PARSER_H_
+#define TENSORFLOW_LITE_EXPERIMENTAL_SUPPORT_METADATA_CC_METADATA_PARSER_H_
+
+namespace tflite {
+namespace metadata {
+
+// The version of the metadata parser that this metadata versioning library is
+// depending on.
+inline constexpr char kMatadataParserVersion[] = "{LATEST_METADATA_PARSER_VERSION}";
+
+}  // namespace metadata
+}  // namespace tflite
+
+#endif  // TENSORFLOW_LITE_EXPERIMENTAL_SUPPORT_METADATA_CC_METADATA_PARSER_H_
diff --git a/tensorflow/lite/experimental/support/metadata/cc/test/BUILD b/tensorflow/lite/experimental/support/metadata/cc/test/BUILD
index fd829124c73..f9d78567d70 100644
--- a/tensorflow/lite/experimental/support/metadata/cc/test/BUILD
+++ b/tensorflow/lite/experimental/support/metadata/cc/test/BUILD
@@ -13,3 +13,12 @@ cc_test(
         "@flatbuffers",
     ],
 )
+
+cc_test(
+    name = "metadata_parser_test",
+    srcs = ["metadata_parser_test.cc"],
+    deps = [
+        "//tensorflow/lite/experimental/support/metadata/cc:metadata_version",
+        "@com_google_googletest//:gtest_main",
+    ],
+)
diff --git a/tensorflow/lite/experimental/support/metadata/cc/test/metadata_parser_test.cc b/tensorflow/lite/experimental/support/metadata/cc/test/metadata_parser_test.cc
new file mode 100644
index 00000000000..af7b8791fe8
--- /dev/null
+++ b/tensorflow/lite/experimental/support/metadata/cc/test/metadata_parser_test.cc
@@ -0,0 +1,33 @@
+/* Copyright 2020 The TensorFlow Authors. All Rights Reserved.
+
+Licensed under the Apache License, Version 2.0 (the "License");
+you may not use this file except in compliance with the License.
+You may obtain a copy of the License at
+
+    http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software
+distributed under the License is distributed on an "AS IS" BASIS,
+WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+See the License for the specific language governing permissions and
+limitations under the License.
+==============================================================================*/
+#include "tensorflow/lite/experimental/support/metadata/cc/metadata_parser.h"
+
+#include 
+#include 
+
+namespace tflite {
+namespace metadata {
+namespace {
+
+using ::testing::MatchesRegex;
+
+TEST(MetadataParserTest, MatadataParserVersionIsWellFormed) {
+  // Validates that the version is well-formed (x.y.z).
+  EXPECT_THAT(kMatadataParserVersion, MatchesRegex("[0-9]+\\.[0-9]+\\.[0-9]+"));
+}
+
+}  // namespace
+}  // namespace metadata
+}  // namespace tflite
diff --git a/tensorflow/lite/experimental/support/metadata/cc/test/metadata_version_test.cc b/tensorflow/lite/experimental/support/metadata/cc/test/metadata_version_test.cc
index 02ecfdbd232..03f4d3bf28b 100644
--- a/tensorflow/lite/experimental/support/metadata/cc/test/metadata_version_test.cc
+++ b/tensorflow/lite/experimental/support/metadata/cc/test/metadata_version_test.cc
@@ -44,7 +44,7 @@ TEST(MetadataVersionTest,
                                             builder.GetSize(), &min_version),
             kTfLiteOk);
   // Validates that the version is well-formed (x.y.z).
-  EXPECT_THAT(min_version, MatchesRegex("[0-9]*\\.[0-9]*\\.[0-9]"));
+  EXPECT_THAT(min_version, MatchesRegex("[0-9]+\\.[0-9]+\\.[0-9]+"));
 }
 
 TEST(MetadataVersionTest,
diff --git a/tensorflow/lite/experimental/support/metadata/java/BUILD b/tensorflow/lite/experimental/support/metadata/java/BUILD
index c208752ae24..00d10bcca56 100644
--- a/tensorflow/lite/experimental/support/metadata/java/BUILD
+++ b/tensorflow/lite/experimental/support/metadata/java/BUILD
@@ -9,9 +9,13 @@ package(
     licenses = ["notice"],  # Apache 2.0
 )
 
+METADATA_SRCS = glob(
+    ["src/java/org/tensorflow/lite/support/metadata/**/*.java"],
+)
+
 android_library(
     name = "tensorflow-lite-support-metadata",
-    srcs = glob(["src/java/org/tensorflow/lite/support/metadata/**/*.java"]),
+    srcs = METADATA_SRCS,
     manifest = "AndroidManifest.xml",
     deps = [
         "//tensorflow/lite/experimental/support/metadata:metadata_schema_fbs_android",
@@ -22,7 +26,7 @@ android_library(
 
 java_library(
     name = "tensorflow-lite-support-metadata-lib",
-    srcs = glob(["src/java/org/tensorflow/lite/support/metadata/**/*.java"]),
+    srcs = METADATA_SRCS,
     javacopts = JAVACOPTS,
     resource_jars = [
         "//tensorflow/lite/experimental/support/metadata:libmetadata_schema_java.jar",
diff --git a/tensorflow/lite/experimental/support/metadata/java/src/java/org/tensorflow/lite/support/metadata/MetadataExtractor.java b/tensorflow/lite/experimental/support/metadata/java/src/java/org/tensorflow/lite/support/metadata/MetadataExtractor.java
index 9da5b59cf46..9bf5ae93138 100644
--- a/tensorflow/lite/experimental/support/metadata/java/src/java/org/tensorflow/lite/support/metadata/MetadataExtractor.java
+++ b/tensorflow/lite/experimental/support/metadata/java/src/java/org/tensorflow/lite/support/metadata/MetadataExtractor.java
@@ -52,10 +52,6 @@ import org.tensorflow.lite.support.metadata.schema.TensorMetadata;
  * MetadataExtractor} omits subgraph index as an input in its methods.
  */
 public class MetadataExtractor {
-  // TODO(b/156539454): remove the hardcode versioning number and populate the version through
-  // genrule.
-  /** The version of the metadata parser that this {@link MetadataExtractor} library depends on. */
-  public static final String METADATA_PARSER_VERSION = "1.0.1";
 
   /** The helper class to load metadata from TFLite model FlatBuffer. */
   private final ModelInfo modelInfo;
@@ -85,7 +81,7 @@ public class MetadataExtractor {
         System.err.printf(
             " Some fields in the metadata belong to a future schema. The minimum parser"
                 + " version required is %s, but the version of the current metadata parser is %s",
-            metadataInfo.getMininumParserVersion(), METADATA_PARSER_VERSION);
+            metadataInfo.getMininumParserVersion(), MetadataParser.VERSION);
       }
 
       checkArgument(
@@ -290,7 +286,7 @@ public class MetadataExtractor {
     if (minVersion == null) {
       return true;
     }
-    return compareVersions(minVersion, METADATA_PARSER_VERSION) <= 0;
+    return compareVersions(minVersion, MetadataParser.VERSION) <= 0;
   }
 
   /**
diff --git a/tensorflow/lite/experimental/support/metadata/java/src/java/org/tensorflow/lite/support/metadata/MetadataParser.java b/tensorflow/lite/experimental/support/metadata/java/src/java/org/tensorflow/lite/support/metadata/MetadataParser.java
new file mode 100644
index 00000000000..195a330462b
--- /dev/null
+++ b/tensorflow/lite/experimental/support/metadata/java/src/java/org/tensorflow/lite/support/metadata/MetadataParser.java
@@ -0,0 +1,27 @@
+/* Copyright 2020 The TensorFlow Authors. All Rights Reserved.
+
+Licensed under the Apache License, Version 2.0 (the "License");
+you may not use this file except in compliance with the License.
+You may obtain a copy of the License at
+
+    http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software
+distributed under the License is distributed on an "AS IS" BASIS,
+WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+See the License for the specific language governing permissions and
+limitations under the License.
+==============================================================================*/
+
+package org.tensorflow.lite.support.metadata;
+
+/** Information about the metadata parser that this metadata extractor library is depending on. */
+public final class MetadataParser {
+  /**
+   * The version of the metadata parser that this metadata extractor library is depending on. The
+   * value should match the value of "Schema Semantic version" in metadata_schema.fbs.
+   */
+  public static final String VERSION = "1.0.1";
+
+  private MetadataParser() {}
+}
diff --git a/tensorflow/lite/experimental/support/metadata/metadata_parser.py.template b/tensorflow/lite/experimental/support/metadata/metadata_parser.py.template
new file mode 100644
index 00000000000..a41ac06969c
--- /dev/null
+++ b/tensorflow/lite/experimental/support/metadata/metadata_parser.py.template
@@ -0,0 +1,26 @@
+# Copyright 2020 The TensorFlow Authors. All Rights Reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+# ==============================================================================
+"""Information about the metadata parser that this python library depends on."""
+
+from __future__ import absolute_import
+from __future__ import division
+from __future__ import print_function
+
+
+class MetadataParser(object):
+  """Information about the metadata parser."""
+
+  # The version of the metadata parser.
+  VERSION = "{LATEST_METADATA_PARSER_VERSION}"
diff --git a/tensorflow/lite/experimental/support/metadata/metadata_parser_test.py b/tensorflow/lite/experimental/support/metadata/metadata_parser_test.py
new file mode 100644
index 00000000000..3b1d19278cd
--- /dev/null
+++ b/tensorflow/lite/experimental/support/metadata/metadata_parser_test.py
@@ -0,0 +1,38 @@
+# Copyright 2020 The TensorFlow Authors. All Rights Reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+# ==============================================================================
+"""Tests for tensorflow.lite.experimental.support.metadata.metadata_parser."""
+
+from __future__ import absolute_import
+from __future__ import division
+from __future__ import print_function
+
+import re
+
+from tensorflow.lite.experimental.support.metadata import metadata_parser
+from tensorflow.python.framework import test_util
+from tensorflow.python.platform import test
+
+
+class MetadataParserTest(test_util.TensorFlowTestCase):
+
+  def test_version_wellFormedSemanticVersion(self):
+    # Validates that the version is well-formed (x.y.z).
+    self.assertTrue(
+        re.match('[0-9]+\\.[0-9]+\\.[0-9]+',
+                 metadata_parser.MetadataParser.VERSION))
+
+
+if __name__ == '__main__':
+  test.main()
diff --git a/tensorflow/lite/experimental/support/metadata/metadata_schema.fbs b/tensorflow/lite/experimental/support/metadata/metadata_schema.fbs
index 53c26b3e079..a88225f1960 100644
--- a/tensorflow/lite/experimental/support/metadata/metadata_schema.fbs
+++ b/tensorflow/lite/experimental/support/metadata/metadata_schema.fbs
@@ -29,18 +29,31 @@ namespace tflite;
 // generate the model interface. It is recommended to fill in at least those
 // enties to boost the codegen performance.
 
-// LINT.IfChange
-
-// The Metadata schema is versioned by the Semantic versioning number, which
-// tracks the schema changes according to the Semantic versioning rules.
+// The Metadata schema is versioned by the Semantic versioning number, such as
+// MAJOR.MINOR.PATCH. It tracks the schema changes according to the rules below:
+//  * Bump up the MAJOR number when making potentially backwards incompatible
+//    changes. It must be incremented if the new changes break the backwards
+//    compatibility. It may also include minor and patch level changes as
+//    needed. The true backwards compatibility is indicated by the file
+//    identifier.
+//  * Bump up the MINOR number when making backwards compatible updates for
+//    major features, such as supporting new content types or adding new
+//    processing units.
+//  * Bump up the PATCH number when making small backwards compatible changes,
+//    such as adding a new fields or deprecating certain fields (not deleting
+//    them).
 //
 // ModelMetadata.min_parser_version indicates the minimum necessary metadata
 // parser version to fully understand all fields in a given metadata flatbuffer.
 //
-// New fields and types will have associated comments with the schema version for
-// which they were added.
+// New fields and types will have associated comments with the schema version
+// for which they were added.
 //
+// LINT.IfChange
 // Schema Semantic version: 1.0.1
+// LINT.ThenChange(//tensorflow/lite/experimental/\
+//.    support/metadata/java/src/java/org/tensorflow/lite/support/metadata/\
+//.    MetadataParser.java)
 
 // This indicates the flatbuffer compatibility. The number will bump up when a
 // break change is applied to the schema, such as removing fields or adding new
@@ -53,10 +66,6 @@ file_identifier "M001";
 // File extension of any written files.
 file_extension "tflitemeta";
 
-// LINT.ThenChange(//tensorflow/lite/experimental/\
-//     /support/metadata/java/src/java/org/tensorflow/lite/support/metadata/\
-//     MetadataExtractor.java)
-
 // LINT.IfChange
 enum AssociatedFileType : byte {
   UNKNOWN = 0,

From a8456eae42648e81f8a8f75c793ad53473bebc70 Mon Sep 17 00:00:00 2001
From: Yunxing Dai 
Date: Fri, 19 Jun 2020 11:59:51 -0700
Subject: [PATCH 321/867] Check the value shape of set dimension size in shape
 inference.

PiperOrigin-RevId: 317351738
Change-Id: Ia185e7745753711ca9ebf657522ef7422c9696ca
---
 tensorflow/compiler/xla/client/xla_builder.cc |  6 +++--
 .../compiler/xla/service/hlo_verifier.cc      |  3 ++-
 .../compiler/xla/service/shape_inference.cc   |  7 +++++-
 .../compiler/xla/service/shape_inference.h    |  5 ++++-
 .../xla/service/shape_inference_test.cc       | 22 +++++++++++++++++++
 5 files changed, 38 insertions(+), 5 deletions(-)

diff --git a/tensorflow/compiler/xla/client/xla_builder.cc b/tensorflow/compiler/xla/client/xla_builder.cc
index 03ae23ea18b..56e9aba6112 100644
--- a/tensorflow/compiler/xla/client/xla_builder.cc
+++ b/tensorflow/compiler/xla/client/xla_builder.cc
@@ -2762,9 +2762,11 @@ XlaOp XlaBuilder::SetDimensionSize(XlaOp operand, XlaOp val, int64 dimension) {
   return ReportErrorOrReturn([&]() -> StatusOr {
     HloInstructionProto instr;
     TF_ASSIGN_OR_RETURN(const Shape* operand_shape, GetShapePtr(operand));
+    TF_ASSIGN_OR_RETURN(const Shape* val_shape, GetShapePtr(val));
 
-    TF_ASSIGN_OR_RETURN(Shape shape, ShapeInference::InferSetDimensionSizeShape(
-                                         *operand_shape, dimension));
+    TF_ASSIGN_OR_RETURN(Shape shape,
+                        ShapeInference::InferSetDimensionSizeShape(
+                            *operand_shape, *val_shape, dimension));
     // Setting an op's dynamic dimension to the static size is a noop.
     TF_ASSIGN_OR_RETURN(const HloInstructionProto* val_proto,
                         LookUpInstruction(val));
diff --git a/tensorflow/compiler/xla/service/hlo_verifier.cc b/tensorflow/compiler/xla/service/hlo_verifier.cc
index 4661b8fd9e3..d8baebd6fdd 100644
--- a/tensorflow/compiler/xla/service/hlo_verifier.cc
+++ b/tensorflow/compiler/xla/service/hlo_verifier.cc
@@ -1123,7 +1123,8 @@ Status ShapeVerifier::HandleGetDimensionSize(HloInstruction* get_size) {
 Status ShapeVerifier::HandleSetDimensionSize(HloInstruction* set_size) {
   return CheckShape(set_size,
                     ShapeInference::InferSetDimensionSizeShape(
-                        set_size->operand(0)->shape(), set_size->dimension()));
+                        set_size->operand(0)->shape(),
+                        set_size->operand(1)->shape(), set_size->dimension()));
 }
 
 Status ShapeVerifier::CheckShape(const HloInstruction* instruction,
diff --git a/tensorflow/compiler/xla/service/shape_inference.cc b/tensorflow/compiler/xla/service/shape_inference.cc
index 75a80747c1d..bb4a38ded1e 100644
--- a/tensorflow/compiler/xla/service/shape_inference.cc
+++ b/tensorflow/compiler/xla/service/shape_inference.cc
@@ -2248,12 +2248,17 @@ ShapeInference::InferDegenerateDimensionBroadcastShape(HloOpcode operation,
 }
 
 /* static */ StatusOr ShapeInference::InferSetDimensionSizeShape(
-    const Shape& shape, int64 dimension) {
+    const Shape& shape, const Shape& val_shape, int64 dimension) {
   if (dimension < 0 || dimension >= shape.rank()) {
     return InvalidArgument("SetDimensionSize dimension out of bounds: %d.",
                            dimension);
   }
 
+  if (val_shape.rank() != 0 || val_shape.element_type() != S32) {
+    return InvalidArgument(
+        "SetDimensionSize's value has to be S32 scalar, got %s",
+        val_shape.ToString());
+  }
   // TODO(b/119580730): Remove this restriction when very large dimension size
   // is needed.
   if (shape.dimensions(dimension) > std::numeric_limits::max()) {
diff --git a/tensorflow/compiler/xla/service/shape_inference.h b/tensorflow/compiler/xla/service/shape_inference.h
index 2cb5930d098..d47d96ab52d 100644
--- a/tensorflow/compiler/xla/service/shape_inference.h
+++ b/tensorflow/compiler/xla/service/shape_inference.h
@@ -303,10 +303,13 @@ class ShapeInference {
       const Shape& updates_shape, const ProgramShape& to_apply_shape,
       const ScatterDimensionNumbers& scatter_dim_numbers);
 
+  // Helper that validates the given input shape to GetDimensionSize.
   static StatusOr InferGetDimensionSizeShape(const Shape& shape,
                                                     int64 dimension);
 
-  static StatusOr InferSetDimensionSizeShape(const Shape& shape,
+  // Helper that validates the given input shape to SetDimensionSize.
+  static StatusOr InferSetDimensionSizeShape(const Shape& operand_shape,
+                                                    const Shape& val_shape,
                                                     int64 dimension);
 
   // Helper function for creating a Window proto from user-supplied data.
diff --git a/tensorflow/compiler/xla/service/shape_inference_test.cc b/tensorflow/compiler/xla/service/shape_inference_test.cc
index b5ecf6e583e..916d3ab15c8 100644
--- a/tensorflow/compiler/xla/service/shape_inference_test.cc
+++ b/tensorflow/compiler/xla/service/shape_inference_test.cc
@@ -1365,6 +1365,28 @@ TEST_F(ShapeInferenceTest, DotWithTwoContractingDimsPasses) {
   EXPECT_TRUE(ShapeUtil::Equal(inferred_status.ValueOrDie(), output_shape));
 }
 
+TEST_F(ShapeInferenceTest, ErrorSetDimensionSize) {
+  Shape arg_shape = ShapeUtil::MakeShape(F32, {5, 3});
+  Shape val_shape = ShapeUtil::MakeShape(S32, {1});
+  auto inferred_status = ShapeInference::InferSetDimensionSizeShape(
+      arg_shape, val_shape, /*dimension=*/0);
+
+  EXPECT_FALSE(inferred_status.ok());
+  EXPECT_THAT(inferred_status.status().error_message(),
+              HasSubstr("value has to be S32 scalar"));
+}
+
+TEST_F(ShapeInferenceTest, ErrorSetDimensionSizeWrongType) {
+  Shape arg_shape = ShapeUtil::MakeShape(F32, {5, 3});
+  Shape val_shape = ShapeUtil::MakeShape(U32, {});
+  auto inferred_status = ShapeInference::InferSetDimensionSizeShape(
+      arg_shape, val_shape, /*dimension=*/0);
+
+  EXPECT_FALSE(inferred_status.ok());
+  EXPECT_THAT(inferred_status.status().error_message(),
+              HasSubstr("value has to be S32 scalar"));
+}
+
 // BatchMatMul with different batch dimension sizes fails.
 TEST_F(ShapeInferenceTest, DotWithMismatchedBatchDimSizesFails) {
   Shape lhs_shape = ShapeUtil::MakeShape(F32, {2, 11, 3});

From 2229ae89c927b46355a15e8af22365d24afc25bf Mon Sep 17 00:00:00 2001
From: Jiho Choi 
Date: Fri, 19 Jun 2020 12:06:36 -0700
Subject: [PATCH 322/867] Use group_id as step_id.

PiperOrigin-RevId: 317353238
Change-Id: If52b2b4872c92d3f65af8f6ce1651e8c6da7dae7
---
 tensorflow/core/profiler/convert/BUILD        |  1 +
 .../convert/xplane_to_memory_profile.cc       | 59 +++++++------------
 .../convert/xplane_to_memory_profile.h        |  1 +
 .../convert/xplane_to_memory_profile_test.cc  |  3 +-
 .../profiler/protobuf/memory_profile.proto    |  4 +-
 5 files changed, 25 insertions(+), 43 deletions(-)

diff --git a/tensorflow/core/profiler/convert/BUILD b/tensorflow/core/profiler/convert/BUILD
index abf0176bf6f..06594b1aeaf 100644
--- a/tensorflow/core/profiler/convert/BUILD
+++ b/tensorflow/core/profiler/convert/BUILD
@@ -525,6 +525,7 @@ tf_cc_test(
         "//tensorflow/core:test_main",
         "//tensorflow/core/profiler/protobuf:memory_profile_proto_cc",
         "//tensorflow/core/profiler/protobuf:xplane_proto_cc",
+        "//tensorflow/core/profiler/utils:group_events",
         "//tensorflow/core/profiler/utils:xplane_builder",
         "//tensorflow/core/profiler/utils:xplane_schema",
         "//tensorflow/core/profiler/utils:xplane_test_utils",
diff --git a/tensorflow/core/profiler/convert/xplane_to_memory_profile.cc b/tensorflow/core/profiler/convert/xplane_to_memory_profile.cc
index d039ca8da32..d7104c2bbf5 100644
--- a/tensorflow/core/profiler/convert/xplane_to_memory_profile.cc
+++ b/tensorflow/core/profiler/convert/xplane_to_memory_profile.cc
@@ -42,6 +42,8 @@ namespace profiler {
 
 namespace {
 
+constexpr int64 kInvalidStepId = -1;
+
 // Index of the time-sorted memory_profile_snapshots list, and the
 // MemoryActivityMetadata proto it contains.
 using IndexMetaPair = std::pair;
@@ -63,7 +65,7 @@ struct ActivityMetadata {
   int64 allocation_bytes = 0;
   uint64 address = 0;
   absl::string_view tf_op_name;
-  int64 step_id = -1;
+  int64 step_id = kInvalidStepId;
   absl::string_view region_type;
   int64 data_type = 0;
   absl::string_view tensor_shape;
@@ -129,7 +131,6 @@ void UpdateProfileSummary(const AggregationStats& stats, int64 time_offset_ps,
 MemoryProfile GenerateMemoryProfile(const XPlane* host_trace) {
   XPlaneVisitor plane = CreateTfXPlaneVisitor(host_trace);
   MemoryProfile memory_profile;
-  auto* step_count = memory_profile.mutable_step_count();
   // Iterate over all XEvents in the XPlane, and add the XStats to a new
   // MemoryProfileSnapshot if the EventType is kMemoryAllocation or
   // kMemoryDeallocation.
@@ -181,9 +182,8 @@ MemoryProfile GenerateMemoryProfile(const XPlane* host_trace) {
           case StatType::kTfOp:
             metadata.tf_op_name = stat.StrOrRefValue();
             break;
-          case StatType::kStepId:
+          case StatType::kGroupId:
             metadata.step_id = stat.IntValue();
-            if (metadata.step_id != 0) (*step_count)[metadata.step_id]++;
             break;
           case StatType::kRegionType:
             metadata.region_type = stat.StrOrRefValue();
@@ -214,40 +214,21 @@ MemoryProfile GenerateMemoryProfile(const XPlane* host_trace) {
   return memory_profile;
 }
 
-// Sequentialize step ids for the memory profile.
-void UpdateStepId(const tensorflow::protobuf::Map<
-                      tensorflow::protobuf_int64 /*orig_step_id*/,
-                      tensorflow::protobuf_int64 /*count*/>& step_count,
-                  PerAllocatorMemoryProfile* memory_profile) {
-  // Map from original random step id to sequential step id.
-  absl::flat_hash_map step_map;
-  constexpr int kUnknownStep = -2;
-  constexpr double kStepFilterRatio = 0.1;  // Magic number for filtering.
-  tensorflow::protobuf_int64 max_step_count = 0;
-  for (const auto& step_and_count : step_count) {
-    max_step_count = std::max(max_step_count, step_and_count.second);
-  }
-  // Filter out noisy and incomplete original step ids.
-  for (const auto& step_and_count : step_count) {
-    if (static_cast(step_and_count.second) / max_step_count >
-        kStepFilterRatio) {
-      step_map[step_and_count.first] = kUnknownStep;
-    }
-  }
-
-  // Update the step ids in memory_profile for this allocator.
-  int64 step_id = -1;
+// Fix invalid step ids of snapshots at the beginning/end of the profile or at
+// the step boundaries. The snapshots with invalid step ids at the beginning get
+// 0 for their step ids. Those at the step boundaries or at the end get the
+// previous snapshot's step id + 1.
+void UpdateStepId(PerAllocatorMemoryProfile* memory_profile) {
+  int64 last_valid_step_id = -1;
+  // Snapshots are already sorted in time.
   for (auto& snapshot : *memory_profile->mutable_memory_profile_snapshots()) {
     DCHECK(snapshot.has_activity_metadata());
-    // Convert the random step id to sequential step id.
-    int64 orig_step_id = snapshot.activity_metadata().step_id();
-    if (step_map.contains(orig_step_id) &&
-        step_map[orig_step_id] == kUnknownStep) {
-      step_map[orig_step_id] = ++step_id;
+    if (snapshot.mutable_activity_metadata()->step_id() == kInvalidStepId) {
+      snapshot.mutable_activity_metadata()->set_step_id(last_valid_step_id + 1);
+    } else {
+      last_valid_step_id = snapshot.mutable_activity_metadata()->step_id();
     }
-    snapshot.mutable_activity_metadata()->set_step_id(step_id);
   }
-  VLOG(2) << "Max sequential step id in profile: " << step_id;
 }
 
 // Update the MemoryActivityMetadata for each deallocation event by copying from
@@ -481,14 +462,14 @@ void ProcessMemoryProfileProto(int64 max_num_snapshots,
       return a.time_offset_ps() < b.time_offset_ps();
     });
 
-    UpdateStepId(memory_profile->step_count(), allocator_memory_profile);
+    UpdateStepId(allocator_memory_profile);
     UpdateDeallocation(allocator_memory_profile);
 
-    int64 peak_bytes_profile = allocator_memory_profile->profile_summary()
-                                   .peak_stats()
-                                   .peak_bytes_in_use();
     int64 peak_step_id =
-        GetPeakMemoryStep(peak_bytes_profile, allocator_memory_profile);
+        GetPeakMemoryStep(allocator_memory_profile->profile_summary()
+                              .peak_stats()
+                              .peak_bytes_in_use(),
+                          allocator_memory_profile);
     ProcessActiveAllocations(peak_step_id, allocator_memory_profile);
     SampleSnapshots(max_num_snapshots, snapshots);
   }
diff --git a/tensorflow/core/profiler/convert/xplane_to_memory_profile.h b/tensorflow/core/profiler/convert/xplane_to_memory_profile.h
index 873ac800aa5..6eddaeeec71 100644
--- a/tensorflow/core/profiler/convert/xplane_to_memory_profile.h
+++ b/tensorflow/core/profiler/convert/xplane_to_memory_profile.h
@@ -25,6 +25,7 @@ namespace profiler {
 
 // Process the host threads XPlane and generate MemoryProfile result; at most
 // max_num_snapshots will be displayed on the UI.
+// REQUIRED: host_plane should have been grouped by calling GroupTfEvents().
 MemoryProfile ConvertXPlaneToMemoryProfile(const XPlane& host_plane,
                                            int64 max_num_snapshots = 1000);
 
diff --git a/tensorflow/core/profiler/convert/xplane_to_memory_profile_test.cc b/tensorflow/core/profiler/convert/xplane_to_memory_profile_test.cc
index 5ddcbcfc75d..c334318dcfe 100644
--- a/tensorflow/core/profiler/convert/xplane_to_memory_profile_test.cc
+++ b/tensorflow/core/profiler/convert/xplane_to_memory_profile_test.cc
@@ -20,6 +20,7 @@ limitations under the License.
 #include "tensorflow/core/platform/types.h"
 #include "tensorflow/core/profiler/protobuf/memory_profile.pb.h"
 #include "tensorflow/core/profiler/protobuf/xplane.pb.h"
+#include "tensorflow/core/profiler/utils/group_events.h"
 #include "tensorflow/core/profiler/utils/xplane_builder.h"
 #include "tensorflow/core/profiler/utils/xplane_schema.h"
 #include "tensorflow/core/profiler/utils/xplane_test_utils.h"
@@ -84,11 +85,11 @@ TEST(ConvertXPlaneToMemoryProfile, OneAllocatorMultiActivitiesTest) {
                 {StatType::kRegionType, "temp"},
                 {StatType::kTensorShapes, "[1, 2]"}});
 
+  tensorflow::profiler::GroupTfEvents(&space, nullptr);
   MemoryProfile memory_profile = ConvertXPlaneToMemoryProfile(*host_plane);
   EXPECT_EQ(memory_profile.memory_profile_per_allocator().size(), 1);
   EXPECT_EQ(memory_profile.num_hosts(), 1);
   EXPECT_EQ(memory_profile.memory_ids_size(), 1);
-  EXPECT_EQ(memory_profile.step_count().size(), 1);
   EXPECT_EQ(memory_profile.memory_profile_per_allocator().begin()->first,
             "GPU_0_bfc");
   const auto& allocator_memory_profile =
diff --git a/tensorflow/core/profiler/protobuf/memory_profile.proto b/tensorflow/core/profiler/protobuf/memory_profile.proto
index 7a5272c60b2..4d492a56255 100644
--- a/tensorflow/core/profiler/protobuf/memory_profile.proto
+++ b/tensorflow/core/profiler/protobuf/memory_profile.proto
@@ -122,7 +122,5 @@ message MemoryProfile {
   // Ids for profiled memory allocators, used to populate memory selection list
   // at front end.
   repeated string memory_ids = 3;
-  // Map of original random int64 step id to the count of memory activity events
-  // assigned with it.
-  map step_count = 4;
+  reserved 4;
 }

From 2ec0214b48878b94763c4f41e095b4579d78b58f Mon Sep 17 00:00:00 2001
From: Kuangyuan Chen 
Date: Fri, 19 Jun 2020 12:13:11 -0700
Subject: [PATCH 323/867] Internal rollback for saved model importer.

PiperOrigin-RevId: 317354577
Change-Id: I6fdc05921b2f1d85b34b1751ed00f65525d45ad3
---
 tensorflow/compiler/mlir/tensorflow/BUILD     |   4 +-
 .../mlir/tensorflow/ir/tf_saved_model.cc      |  25 ---
 .../mlir/tensorflow/ir/tf_saved_model_ops.td  |  24 ---
 .../tests/tf_saved_model/common_v1.py         |   1 -
 .../tests/tf_saved_model/hash_table_v1.py     |  92 -----------
 .../tensorflow/tests/tf_saved_model_ops.mlir  |   5 -
 .../tests/tf_saved_model_ops_invalid.mlir     |  33 ----
 .../mlir/tensorflow/translate/import_model.cc | 149 +++++-------------
 8 files changed, 43 insertions(+), 290 deletions(-)
 delete mode 100644 tensorflow/compiler/mlir/tensorflow/tests/tf_saved_model/hash_table_v1.py

diff --git a/tensorflow/compiler/mlir/tensorflow/BUILD b/tensorflow/compiler/mlir/tensorflow/BUILD
index 7c0d427e87b..bcabb13d301 100644
--- a/tensorflow/compiler/mlir/tensorflow/BUILD
+++ b/tensorflow/compiler/mlir/tensorflow/BUILD
@@ -661,9 +661,7 @@ cc_library(
         ":tensorflow_types",
         ":translate_utils",
         "//tensorflow/cc/saved_model:bundle_v2",
-        "//tensorflow/cc/saved_model:constants",
         "//tensorflow/cc/saved_model:loader_lite",
-        "//tensorflow/cc/saved_model:loader_util",
         "//tensorflow/compiler/jit:shape_inference_helpers",
         "//tensorflow/compiler/mlir:op_or_arg_name_mapper",
         "//tensorflow/compiler/tf2xla:functionalize_control_flow",
@@ -675,7 +673,6 @@ cc_library(
         "//tensorflow/core:lib",
         "//tensorflow/core:protos_all_cc",
         "//tensorflow/core/grappler/utils:transitive_fanin",
-        "//tensorflow/core/platform:protobuf_internal",
         "//tensorflow/core/platform:types",
         "//tensorflow/stream_executor/lib",
         "@com_google_absl//absl/algorithm:container",
@@ -685,6 +682,7 @@ cc_library(
         "@com_google_absl//absl/strings",
         "@com_google_absl//absl/types:optional",
         "@llvm-project//llvm:Support",
+        "@llvm-project//mlir:Analysis",
         "@llvm-project//mlir:IR",
         "@llvm-project//mlir:Pass",
         "@llvm-project//mlir:StandardOps",
diff --git a/tensorflow/compiler/mlir/tensorflow/ir/tf_saved_model.cc b/tensorflow/compiler/mlir/tensorflow/ir/tf_saved_model.cc
index ef248379d2e..7db0eed7713 100644
--- a/tensorflow/compiler/mlir/tensorflow/ir/tf_saved_model.cc
+++ b/tensorflow/compiler/mlir/tensorflow/ir/tf_saved_model.cc
@@ -76,23 +76,6 @@ static LogicalResult Verify(GlobalTensorOp global_tensor) {
   return success();
 }
 
-static LogicalResult Verify(SessionInitializerOp session_initializer) {
-  mlir::SymbolTable symbol_table(
-      session_initializer.getParentOfType());
-
-  auto init_func_op =
-      symbol_table.lookup(session_initializer.initializer());
-  if (!init_func_op)
-    return session_initializer.emitOpError()
-           << "the initializer function does not exist";
-
-  if (!init_func_op.getType().getResults().empty())
-    return session_initializer.emitOpError()
-           << "the initializer function should have no output";
-
-  return success();
-}
-
 #define GET_OP_CLASSES
 #include "tensorflow/compiler/mlir/tensorflow/ir/tf_saved_model.cc.inc"
 
@@ -249,14 +232,6 @@ static LogicalResult VerifySavedModelModule(
       }
     }
   }
-
-  auto session_initializers = module.getOps();
-  if (std::distance(session_initializers.begin(), session_initializers.end()) >
-      1) {
-    return (*++session_initializers.begin()).emitError()
-           << "there must be no more than one session_initializer op";
-  }
-
   SymbolTable symbol_table(module);
   auto symbol_uses = SymbolTable::getSymbolUses(&module.getBodyRegion());
   if (!symbol_uses.hasValue()) {
diff --git a/tensorflow/compiler/mlir/tensorflow/ir/tf_saved_model_ops.td b/tensorflow/compiler/mlir/tensorflow/ir/tf_saved_model_ops.td
index 497f4d90cb9..4431a160edf 100644
--- a/tensorflow/compiler/mlir/tensorflow/ir/tf_saved_model_ops.td
+++ b/tensorflow/compiler/mlir/tensorflow/ir/tf_saved_model_ops.td
@@ -128,28 +128,4 @@ def TfSavedModel_GlobalTensorOp : TfSavedModel_Op<"global_tensor"> {
   let verifier = [{ return Verify(*this); }];
 }
 
-def TfSavedModel_SessionInitializerOp: TfSavedModel_Op<"session_initializer"> {
-  let summary = "Initializes TensorFlow session state.";
-  let description = [{
-    Represents a session initializer function initializes TensorFlow session
-    state. It is used to initialize resources in the saved model before calling
-    any exported functions. There must be no more than one session initializer
-    in a saved model.
-
-    The `initializer` represents the initialization function. The function have
-    no output and this function should be only called once.
-
-    This is used, for example, to initialize hash tables stored in resources and
-    accessed by resource name (rather than as resource handles or bound inputs
-    which is how `global_tensor`s are referenced)
-  }];
-
-  let arguments = (ins
-    FlatSymbolRefAttr:$initializer
-  );
-
-
-  let verifier = [{ return Verify(*this); }];
-}
-
 #endif // SAVED_MODEL_DIALECT
diff --git a/tensorflow/compiler/mlir/tensorflow/tests/tf_saved_model/common_v1.py b/tensorflow/compiler/mlir/tensorflow/tests/tf_saved_model/common_v1.py
index 51ccbeb1fbd..7171f63bb05 100644
--- a/tensorflow/compiler/mlir/tensorflow/tests/tf_saved_model/common_v1.py
+++ b/tensorflow/compiler/mlir/tensorflow/tests/tf_saved_model/common_v1.py
@@ -84,7 +84,6 @@ def do_test(signature_def_map, show_debug_info=False):
     builder.add_meta_graph_and_variables(
         sess, [tf.saved_model.tag_constants.SERVING],
         signature_def_map,
-        main_op=tf.tables_initializer(),
         strip_default_attrs=True)
     builder.save()
 
diff --git a/tensorflow/compiler/mlir/tensorflow/tests/tf_saved_model/hash_table_v1.py b/tensorflow/compiler/mlir/tensorflow/tests/tf_saved_model/hash_table_v1.py
deleted file mode 100644
index 64847434b82..00000000000
--- a/tensorflow/compiler/mlir/tensorflow/tests/tf_saved_model/hash_table_v1.py
+++ /dev/null
@@ -1,92 +0,0 @@
-# Copyright 2019 The TensorFlow Authors. All Rights Reserved.
-#
-# Licensed under the Apache License, Version 2.0 (the "License");
-# you may not use this file except in compliance with the License.
-# You may obtain a copy of the License at
-#
-#     http://www.apache.org/licenses/LICENSE-2.0
-#
-# Unless required by applicable law or agreed to in writing, software
-# distributed under the License is distributed on an "AS IS" BASIS,
-# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-# See the License for the specific language governing permissions and
-# limitations under the License.
-# ==============================================================================
-
-# RUN: %p/hash_table_v1 | FileCheck %s
-
-# pylint: disable=missing-docstring,line-too-long
-from __future__ import absolute_import
-from __future__ import division
-from __future__ import print_function
-
-import tensorflow.compat.v1 as tf
-from tensorflow.compiler.mlir.tensorflow.tests.tf_saved_model import common_v1
-
-# Verify that the tf.versions attribute exists. It is difficult to enforce
-# contents, since the version numbers change over time. The conversion logic
-# itself is verified in the common graphdef converter, so here just assert
-# it is being invoked.
-# CHECK: module
-# CHECK-SAME: tf.versions
-# CHECK-SAME: bad_consumers
-# CHECK-SAME: min_consumer
-# CHECK-SAME: producer
-
-# CHECK: "tf_saved_model.session_initializer"() {initializer = [[init:@.*]]} : () -> ()
-# CHECK: "tf_saved_model.global_tensor"()
-
-# CHECK:      func {{@[a-zA-Z_0-9]+}}(
-# CHECK-SAME: [[ARG0:%.*]]: tensor
-# CHECK-SAME: [[ARG1:%.*]]: tensor ()
-
   // Representation for constants: (immutable) global tensor.
   // CHECK: tf_saved_model.global_tensor
   "tf_saved_model.global_tensor"() {
diff --git a/tensorflow/compiler/mlir/tensorflow/tests/tf_saved_model_ops_invalid.mlir b/tensorflow/compiler/mlir/tensorflow/tests/tf_saved_model_ops_invalid.mlir
index f04e1a60b36..7287fcf66c8 100644
--- a/tensorflow/compiler/mlir/tensorflow/tests/tf_saved_model_ops_invalid.mlir
+++ b/tensorflow/compiler/mlir/tensorflow/tests/tf_saved_model_ops_invalid.mlir
@@ -261,39 +261,6 @@ module attributes {tf_saved_model.semantics} {
 
 // -----
 
-module attributes {tf_saved_model.semantics} {
-
-  // expected-error@+1 {{the initializer function does not exist}}
-  "tf_saved_model.session_initializer"() { initializer = @init } : () -> ()
-}
-
-// -----
-
-module attributes {tf_saved_model.semantics} {
-
-  // expected-error@+1 {{the initializer function should have no output}}
-  "tf_saved_model.session_initializer"() { initializer = @init } : () -> ()
-  func @init() -> tensor<1xf32> attributes {sym_visibility = "private"} {
-    %0 = "tf.Const"() {value = dense<[1.0]> : tensor<1xf32> } : () -> tensor<1xf32>
-    return %0 : tensor<1xf32>
-  }
-}
-
-// -----
-
-module attributes {tf_saved_model.semantics} {
-
-  "tf_saved_model.session_initializer"() { initializer = @init } : () -> ()
-  // expected-error@+1 {{there must be no more than one session_initializer op}}
-  "tf_saved_model.session_initializer"() { initializer = @init } : () -> ()
-  func @init() -> tensor<1xf32> attributes {sym_visibility = "private"} {
-    %0 = "tf.Const"() {value = dense<[1.0]> : tensor<1xf32> } : () -> tensor<1xf32>
-    return %0 : tensor<1xf32>
-  }
-}
-
-// -----
-
 module attributes {tf_saved_model.semantics} {
 
   // expected-error@+1 {{exported function @f should be public}}
diff --git a/tensorflow/compiler/mlir/tensorflow/translate/import_model.cc b/tensorflow/compiler/mlir/tensorflow/translate/import_model.cc
index 3cff4217215..820d0ce31fb 100644
--- a/tensorflow/compiler/mlir/tensorflow/translate/import_model.cc
+++ b/tensorflow/compiler/mlir/tensorflow/translate/import_model.cc
@@ -60,8 +60,6 @@ limitations under the License.
 #include "mlir/IR/Types.h"  // from @llvm-project
 #include "mlir/IR/Verifier.h"  // from @llvm-project
 #include "mlir/Pass/PassManager.h"  // from @llvm-project
-#include "tensorflow/cc/saved_model/constants.h"
-#include "tensorflow/cc/saved_model/loader_util.h"
 #include "tensorflow/compiler/jit/shape_inference_helpers.h"
 #include "tensorflow/compiler/mlir/op_or_arg_name_mapper.h"
 #include "tensorflow/compiler/mlir/tensorflow/ir/tf_attributes.h"
@@ -101,7 +99,6 @@ limitations under the License.
 #include "tensorflow/core/lib/core/errors.h"
 #include "tensorflow/core/lib/strings/str_util.h"
 #include "tensorflow/core/platform/protobuf.h"
-#include "tensorflow/core/platform/protobuf_internal.h"
 #include "tensorflow/core/platform/types.h"
 #include "tensorflow/core/protobuf/graph_debug_info.pb.h"
 #include "tensorflow/core/protobuf/meta_graph.pb.h"
@@ -119,7 +116,6 @@ using mlir::NamedAttrList;
 using mlir::TensorType;
 using mlir::TF::VarHandleOp;
 using mlir::tf_saved_model::GlobalTensorOp;
-using mlir::tf_saved_model::SessionInitializerOp;
 using stream_executor::port::StatusOr;
 
 namespace {
@@ -2959,13 +2955,6 @@ void SortSavedModelModule(mlir::ModuleOp module) {
     named_global_tensor.global_tensor.getOperation()->moveBefore(
         &module.getBody()->front());
   }
-
-  auto initializers = module.getOps();
-  if (!initializers.empty()) {
-    (*initializers.begin())
-        .getOperation()
-        ->moveBefore(&module.getBody()->front());
-  }
 }
 
 Status CreateSavedModelIR(
@@ -3252,29 +3241,17 @@ class SavedModelSignatureDefImporter {
                                  absl::Span exported_names,
                                  mlir::MLIRContext* context)
       : bundle_(bundle),
-        flib_def_(OpRegistry::Global(), graph_def().library()),
-        debug_info_(),
         exported_names_(exported_names),
-        module_(mlir::ModuleOp::create(mlir::UnknownLoc::get(context))) {
-    // debug_info might not be loaded with loader_lite.
-    if (bundle_.debug_info != nullptr) debug_info_ = *bundle_.debug_info;
-  }
+        module_(mlir::ModuleOp::create(mlir::UnknownLoc::get(context))) {}
 
   // Converts the SavedModel to the SavedModel dialect. Creates an MLIR function
   // for each signature.
   StatusOr ConvertSignatures();
-  Status ConvertSignature(const std::string& sig_def_key,
-                          const SignatureDef& signature_def);
-
-  // Converts the initialization graph in the SavedModel to an MLIR function.
-  Status ConvertInitializer();
-
-  // Converts a graph with feeds and fetches to an MLIR function.
-  StatusOr ConvertGraph(
-      const std::string& name,
-      const std::vector>& inputs,
-      const std::vector>& outputs,
-      const std::vector control_outputs);
+  Status ConvertSignature(const GraphDef& graphdef,
+                          const std::string& sig_def_key,
+                          const SignatureDef& signature_def,
+                          const GraphDebugInfo& debug_info,
+                          const FunctionLibraryDefinition& flib_def);
 
   // Creates GlobalTensorOp for each variable and moves each VarHandle op to
   // the enclosing function's arguments.
@@ -3296,62 +3273,18 @@ class SavedModelSignatureDefImporter {
   GraphImportConfig::InputArrays ParseInputArrays(
       const std::vector>& inputs);
 
-  const GraphDef& graph_def() const {
-    return bundle_.meta_graph_def.graph_def();
-  }
-  const FunctionLibraryDefinition& flib_def() const { return flib_def_; }
-  const GraphDebugInfo& debug_info() const { return debug_info_; }
-
   const SavedModelBundle& bundle_;
-  FunctionLibraryDefinition flib_def_;
-  GraphDebugInfo debug_info_;
   absl::Span exported_names_;
   mlir::OwningModuleRef module_;
 };
 
-Status SavedModelSignatureDefImporter::ConvertInitializer() {
-  std::vector asset_file_defs;
-  TF_RETURN_IF_ERROR(
-      internal::GetAssetFileDefs(bundle_.meta_graph_def, &asset_file_defs));
-
-  if (!asset_file_defs.empty())
-    return errors::Unimplemented(
-        absl::StrCat("Assets are not supported in signaturedef importer"));
-
-  std::string init_node_name;
-  TF_RETURN_IF_ERROR(
-      internal::GetInitOp("", bundle_.meta_graph_def, &init_node_name));
-
-  if (init_node_name.empty()) return Status::OK();
-
-  TF_ASSIGN_OR_RETURN(auto sub_module,
-                      ConvertGraph(init_node_name, {}, {}, {init_node_name}));
-
-  mlir::SymbolTable symbol_table(*sub_module);
-
-  auto init_func_op = symbol_table.lookup(init_node_name);
-
-  init_func_op.removeAttr("tf.entry_function");
-
-  mlir::OpBuilder builder(module_->getBodyRegion());
-
-  builder.create(
-      module_->getLoc(), builder.getSymbolRefAttr(init_func_op.getName()));
-
-  // Move the converted functions to top level MLIR module.
-  auto* block = module_->getBody();
-  auto* sub_block = sub_module->getBody();
-  block->getOperations().splice(
-      mlir::Block::iterator(block->getTerminator()), sub_block->getOperations(),
-      sub_block->begin(), mlir::Block::iterator(sub_block->getTerminator()));
-
-  return Status::OK();
-}
-
 StatusOr
 SavedModelSignatureDefImporter::ConvertSignatures() {
   const auto& signatures = bundle_.GetSignatures();
-  PopulateTfVersions(module_.get(), graph_def().versions());
+  const auto& graphdef = bundle_.meta_graph_def.graph_def();
+  PopulateTfVersions(module_.get(), graphdef.versions());
+
+  FunctionLibraryDefinition flib_def(OpRegistry::Global(), graphdef.library());
 
   // debug_info might not be loaded with loader_lite.
   GraphDebugInfo debug_info;
@@ -3374,10 +3307,9 @@ SavedModelSignatureDefImporter::ConvertSignatures() {
       continue;
     }
 
-    TF_RETURN_IF_ERROR(ConvertSignature(sig_def_key, signature_def));
+    TF_RETURN_IF_ERROR(ConvertSignature(graphdef, sig_def_key, signature_def,
+                                        debug_info, flib_def));
   }
-
-  TF_RETURN_IF_ERROR(ConvertInitializer());
   TF_RETURN_IF_ERROR(LiftVariables());
 
   mlir::OpBuilder builder(module_->getBodyRegion());
@@ -3388,32 +3320,10 @@ SavedModelSignatureDefImporter::ConvertSignatures() {
   return std::move(module_);
 }
 
-StatusOr SavedModelSignatureDefImporter::ConvertGraph(
-    const std::string& name,
-    const std::vector>& inputs,
-    const std::vector>& outputs,
-    const std::vector control_outputs) {
-  GraphImportConfig specs;
-  specs.prune_unused_nodes = true;
-  specs.inputs = ParseInputArrays(inputs);
-  for (auto& output : outputs) specs.outputs.push_back(output.second.name());
-  specs.control_outputs = control_outputs;
-
-  // Convert sub-graphdef to sub-graph.
-  GraphConstructorOptions options;
-  options.allow_internal_ops = true;
-  options.add_default_attributes = true;
-  Graph graph(OpRegistry::Global());
-
-  TF_RETURN_IF_ERROR(ConvertGraphDefToGraph(options, graph_def(), &graph));
-
-  // Convert sub-graph to MLIR module.true
-  return GraphDefImporter::Convert(module_->getContext(), graph, debug_info(),
-                                   flib_def(), specs, name);
-}
-
 Status SavedModelSignatureDefImporter::ConvertSignature(
-    const std::string& sig_def_key, const SignatureDef& signature_def) {
+    const GraphDef& graphdef, const std::string& sig_def_key,
+    const SignatureDef& signature_def, const GraphDebugInfo& debug_info,
+    const FunctionLibraryDefinition& flib_def) {
   // Create local vectors for the input and output and sort them to be
   // deterministic. We don't want anyone to really depend on the order, client
   // should lookup argument/result mapping by attribute name.
@@ -3429,9 +3339,34 @@ Status SavedModelSignatureDefImporter::ConvertSignature(
     return lhs.first.size() < rhs.first.size() || lhs.first > rhs.first;
   });
 
+  GraphImportConfig specs;
+  specs.prune_unused_nodes = true;
+  specs.inputs = ParseInputArrays(inputs);
+  for (auto& output : outputs) specs.outputs.push_back(output.second.name());
+
+  // Remove unused nodes and create sub-graphdef.
+  GraphDef sub_graph_def;
+  TF_RETURN_IF_ERROR(tensorflow::grappler::SetTransitiveFaninGraph(
+      graphdef, &sub_graph_def,
+      /*terminal_nodes=*/{specs.outputs.begin(), specs.outputs.end()}));
+
+  // Set the function library definitions in the pruned graphdef.
+  *sub_graph_def.mutable_library() = flib_def.ToProto();
+
+  // Convert sub-graphdef to sub-graph.
+  GraphConstructorOptions options;
+  options.allow_internal_ops = true;
+  options.add_default_attributes = true;
+  Graph sub_graph(OpRegistry::Global());
+
+  TF_RETURN_IF_ERROR(
+      ConvertGraphDefToGraph(options, sub_graph_def, &sub_graph));
+
   // Convert sub-graph to MLIR module.
-  TF_ASSIGN_OR_RETURN(auto sub_module,
-                      ConvertGraph(sig_def_key, inputs, outputs, {}));
+  TF_ASSIGN_OR_RETURN(
+      auto sub_module,
+      GraphDefImporter::Convert(module_->getContext(), sub_graph, debug_info,
+                                flib_def, specs, sig_def_key));
   mlir::OpBuilder builder(sub_module->getBodyRegion());
 
   // Find the FuncOp which corresponds to current SignatureDef.

From 08a99df84299b4c55f40800c148f55773f3d6198 Mon Sep 17 00:00:00 2001
From: Jonah Kohn <51345541+jonah-kohn@users.noreply.github.com>
Date: Fri, 19 Jun 2020 12:34:16 -0700
Subject: [PATCH 324/867] Implement format_dtype as intended.

---
 tensorflow/python/keras/utils/vis_utils.py | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/tensorflow/python/keras/utils/vis_utils.py b/tensorflow/python/keras/utils/vis_utils.py
index 21de8014c2a..49407d92ce3 100644
--- a/tensorflow/python/keras/utils/vis_utils.py
+++ b/tensorflow/python/keras/utils/vis_utils.py
@@ -197,7 +197,7 @@ def model_to_dot(model,
           else:
             return str(dtype)
           
-      label = '%s|%s' % (label, dtype)
+      label = '%s|%s' % (label, format_dtype(dtype))
 
     # Rebuild the label as a table including input/output shapes.
     if show_shapes:

From eeb6ccf2cf72a758db2ef41c925b03998229b3a8 Mon Sep 17 00:00:00 2001
From: Jonah Kohn <51345541+jonah-kohn@users.noreply.github.com>
Date: Fri, 19 Jun 2020 12:48:52 -0700
Subject: [PATCH 325/867] Correctly access layer attributes.

---
 tensorflow/python/keras/utils/vis_utils.py | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/tensorflow/python/keras/utils/vis_utils.py b/tensorflow/python/keras/utils/vis_utils.py
index 49407d92ce3..aa2a4c978cb 100644
--- a/tensorflow/python/keras/utils/vis_utils.py
+++ b/tensorflow/python/keras/utils/vis_utils.py
@@ -197,7 +197,7 @@ def model_to_dot(model,
           else:
             return str(dtype)
           
-      label = '%s|%s' % (label, format_dtype(dtype))
+      label = '%s|%s' % (label, format_dtype(layer.dtype))
 
     # Rebuild the label as a table including input/output shapes.
     if show_shapes:

From 0869ff0af5cacc9d9ea85f6cfc0027166b964f4e Mon Sep 17 00:00:00 2001
From: Robert David 
Date: Fri, 19 Jun 2020 12:50:42 -0700
Subject: [PATCH 326/867] Unabbreviate "proj" to "projection" a few places, to
 be consistent over the codebase.

PiperOrigin-RevId: 317361865
Change-Id: I72fedd0bda16a5668fb7b880128f36d2595f0042
---
 tensorflow/lite/kernels/lstm.cc      | 20 +++++++--------
 tensorflow/lite/kernels/lstm_eval.cc | 38 +++++++++++++++-------------
 2 files changed, 30 insertions(+), 28 deletions(-)

diff --git a/tensorflow/lite/kernels/lstm.cc b/tensorflow/lite/kernels/lstm.cc
index b941f2237ca..803fbba4eae 100644
--- a/tensorflow/lite/kernels/lstm.cc
+++ b/tensorflow/lite/kernels/lstm.cc
@@ -182,7 +182,7 @@ TfLiteStatus PopulateQuantizedLstmParams8x8_16(
   float input_to_output_weight_scale = default_scale;
   float recurrent_to_output_weight_scale = default_scale;
   float cell_to_output_weight_scale = default_scale;
-  float proj_weight_scale = default_scale;
+  float projection_weight_scale = default_scale;
   float layer_norm_input_scale = default_scale;
   float layer_norm_forget_scale = default_scale;
   float layer_norm_cell_scale = default_scale;
@@ -229,7 +229,7 @@ TfLiteStatus PopulateQuantizedLstmParams8x8_16(
   }
 
   if (use_projection) {
-    proj_weight_scale = projection_weights->params.scale;
+    projection_weight_scale = projection_weights->params.scale;
   }
   output_state_scale = output_state->params.scale;
 
@@ -276,7 +276,7 @@ TfLiteStatus PopulateQuantizedLstmParams8x8_16(
       std::pow(2, -15) / intermediate_scale[4] * std::pow(2, -15);
 
   effective_proj_scale =
-      proj_weight_scale * intermediate_scale[4] / output_state_scale;
+      projection_weight_scale * intermediate_scale[4] / output_state_scale;
 
   if (use_peephole) {
     if (!use_cifg) {
@@ -442,7 +442,7 @@ TfLiteStatus PopulateQuantizedLstmParams8x8_8(
   int8_t* input_to_output_weight_ptr = nullptr;
   int8_t* recurrent_to_output_weight_ptr = nullptr;
   int8_t* cell_to_output_weight_ptr = nullptr;
-  int8_t* proj_weight_ptr = nullptr;
+  int8_t* projection_weight_ptr = nullptr;
   int16_t* layer_norm_input_weight_ptr = nullptr;
   int16_t* layer_norm_forget_weight_ptr = nullptr;
   int16_t* layer_norm_cell_weight_ptr = nullptr;
@@ -451,7 +451,7 @@ TfLiteStatus PopulateQuantizedLstmParams8x8_8(
   int32_t* forget_gate_bias_ptr = nullptr;
   int32_t* cell_gate_bias_ptr = nullptr;
   int32_t* output_gate_bias_ptr = nullptr;
-  int32_t* proj_bias_ptr = nullptr;
+  int32_t* projection_bias_ptr = nullptr;
   int16_t* cell_ptr = nullptr;
   int8_t* output_state_ptr = nullptr;
 
@@ -469,7 +469,7 @@ TfLiteStatus PopulateQuantizedLstmParams8x8_8(
   float input_to_output_weight_scale = default_scale;
   float recurrent_to_output_weight_scale = default_scale;
   float cell_to_output_weight_scale = default_scale;
-  float proj_weight_scale = default_scale;
+  float projection_weight_scale = default_scale;
   float layer_norm_input_scale = default_scale;
   float layer_norm_forget_scale = default_scale;
   float layer_norm_cell_scale = default_scale;
@@ -528,10 +528,10 @@ TfLiteStatus PopulateQuantizedLstmParams8x8_8(
   }
 
   if (use_projection) {
-    proj_weight_ptr = projection_weights->data.int8;
-    proj_weight_scale = projection_weights->params.scale;
+    projection_weight_ptr = projection_weights->data.int8;
+    projection_weight_scale = projection_weights->params.scale;
     if (projection_bias) {
-      proj_bias_ptr = projection_bias->data.i32;
+      projection_bias_ptr = projection_bias->data.i32;
     }
   }
   output_state_scale = output_state->params.scale;
@@ -593,7 +593,7 @@ TfLiteStatus PopulateQuantizedLstmParams8x8_8(
                                         output_state_scale /
                                         intermediate_scale[11];
   effective_proj_scale =
-      proj_weight_scale * std::pow(2, -15) / output_state_scale;
+      projection_weight_scale * std::pow(2, -15) / output_state_scale;
 
   if (use_peephole) {
     if (!use_cifg) {
diff --git a/tensorflow/lite/kernels/lstm_eval.cc b/tensorflow/lite/kernels/lstm_eval.cc
index 9bdbfa9d48d..f38fdc95f3e 100644
--- a/tensorflow/lite/kernels/lstm_eval.cc
+++ b/tensorflow/lite/kernels/lstm_eval.cc
@@ -919,7 +919,7 @@ inline void LstmStepHybrid(
 //   cell_to_output_weights              - optional
 //
 // Quantized projection weights of size 'n_output * n_cell'
-//   proj_weight_ptr                     - optional
+//   projection_weight_ptr                     - optional
 //
 // Weight scales (scalars) for each of the weights above.
 //   effective_input_to_input_scale_a    - optional
@@ -1019,10 +1019,10 @@ inline void LstmStepInteger(
     int32_t effective_cell_to_forget_scale_b,
     const int16_t* cell_to_output_weight_ptr,
     int32_t effective_cell_to_output_scale_a,
-    int32_t effective_cell_to_output_scale_b, const int8_t* proj_weight_ptr,
-    int32_t effective_proj_scale_a, int32_t effective_proj_scale_b,
-    int32_t hidden_zp, int32_t effective_hidden_scale_a,
-    int32_t effective_hidden_scale_b,
+    int32_t effective_cell_to_output_scale_b,
+    const int8_t* projection_weight_ptr, int32_t effective_proj_scale_a,
+    int32_t effective_proj_scale_b, int32_t hidden_zp,
+    int32_t effective_hidden_scale_a, int32_t effective_hidden_scale_b,
     const int16_t* layer_norm_input_weight_ptr,
     int32_t layer_norm_input_scale_a, int32_t layer_norm_input_scale_b,
     const int16_t* layer_norm_forget_weight_ptr,
@@ -1055,7 +1055,7 @@ inline void LstmStepInteger(
   const bool use_cifg = (input_to_input_weight_ptr == nullptr);
   const bool use_peephole = (cell_to_output_weight_ptr != nullptr);
   const bool use_layer_norm = (layer_norm_forget_weight_ptr != nullptr);
-  const bool use_projection = (proj_weight_ptr != nullptr);
+  const bool use_projection = (projection_weight_ptr != nullptr);
 
   // Check for nullptrs.
   TFLITE_DCHECK(input_to_forget_effective_bias);
@@ -1208,7 +1208,7 @@ inline void LstmStepInteger(
   if (use_projection) {
     std::fill_n(output_ptr, n_batch * n_output, 0);
     tensor_utils::MatrixBatchVectorMultiplyAccumulate(
-        scratch_4_ptr, projection_effective_bias, proj_weight_ptr,
+        scratch_4_ptr, projection_effective_bias, projection_weight_ptr,
         effective_proj_scale_a, effective_proj_scale_b, n_batch, n_cell,
         n_output, output_state_zp, scratch_5_ptr, output_ptr, context);
     if (quantized_proj_clip > 0) {
@@ -1245,7 +1245,7 @@ inline void LstmStepInteger(
 //   cell_to_output_weights              - optional
 //
 // Quantized projection weights of size 'n_output * n_cell'
-//   proj_weight_ptr                     - optional
+//   projection_weight_ptr                     - optional
 //
 // Weight scales (scalars) for each of the weights above.
 //   effective_input_to_input_scale_a    - optional
@@ -1348,9 +1348,9 @@ void LstmStepInteger(
     int32_t effective_cell_to_forget_scale_b,
     const int8_t* cell_to_output_weight_ptr,
     int32_t effective_cell_to_output_scale_a,
-    int32_t effective_cell_to_output_scale_b, const int8_t* proj_weight_ptr,
-    int32_t effective_proj_scale_a, int32_t effective_proj_scale_b,
-    const int16_t* layer_norm_input_weight_ptr,
+    int32_t effective_cell_to_output_scale_b,
+    const int8_t* projection_weight_ptr, int32_t effective_proj_scale_a,
+    int32_t effective_proj_scale_b, const int16_t* layer_norm_input_weight_ptr,
     int32_t layer_norm_input_scale_a, int32_t layer_norm_input_scale_b,
     const int16_t* layer_norm_forget_weight_ptr,
     int32_t layer_norm_forget_scale_a, int32_t layer_norm_forget_scale_b,
@@ -1360,7 +1360,7 @@ void LstmStepInteger(
     int32_t layer_norm_output_scale_a, int32_t layer_norm_output_scale_b,
     const int32_t* input_gate_bias_ptr, const int32_t* forget_gate_bias_ptr,
     const int32_t* cell_gate_bias_ptr, const int32_t* output_gate_bias_ptr,
-    const int32_t* proj_bias_ptr, const TfLiteLSTMParams* params,
+    const int32_t* projection_bias_ptr, const TfLiteLSTMParams* params,
     const int32_t* intermediate_scale_a, const int32_t* intermediate_scale_b,
     const int32_t* intermediate_zp, int16_t quantized_cell_clip,
     int8_t quantized_proj_clip, int n_batch, int n_cell, int n_input,
@@ -1476,8 +1476,9 @@ void LstmStepInteger(
 
   // Projection.
   tensor_utils::MatrixBatchVectorMultiply(
-      scratch3, proj_weight_ptr, effective_proj_scale_a, effective_proj_scale_b,
-      proj_bias_ptr, n_batch, n_cell, n_output, output_state_zp, output_ptr);
+      scratch3, projection_weight_ptr, effective_proj_scale_a,
+      effective_proj_scale_b, projection_bias_ptr, n_batch, n_cell, n_output,
+      output_state_zp, output_ptr);
 
   // Projection clipping.
   if (quantized_proj_clip > 0) {
@@ -2113,7 +2114,8 @@ TfLiteStatus EvalInteger8x8_8(
       GetTensorData(recurrent_to_output_weights);
   const int8_t* cell_to_output_weight_ptr =
       GetTensorData(cell_to_output_weights);
-  const int8_t* proj_weight_ptr = GetTensorData(projection_weights);
+  const int8_t* projection_weight_ptr =
+      GetTensorData(projection_weights);
   const int16_t* layer_norm_input_weight_ptr =
       GetTensorData(input_layer_norm_coefficients);
   const int16_t* layer_norm_forget_weight_ptr =
@@ -2128,7 +2130,7 @@ TfLiteStatus EvalInteger8x8_8(
   const int32_t* cell_gate_bias_ptr = GetTensorData(cell_gate_bias);
   const int32_t* output_gate_bias_ptr =
       GetTensorData(output_gate_bias);
-  const int32_t* proj_bias_ptr = GetTensorData(projection_bias);
+  const int32_t* projection_bias_ptr = GetTensorData(projection_bias);
   int16_t* cell_ptr = GetTensorData(cell_state);
   int8_t* output_state_ptr = GetTensorData(output_state);
   int8_t* output_ptr = nullptr;
@@ -2195,7 +2197,7 @@ TfLiteStatus EvalInteger8x8_8(
         integer_lstm_param->effective_cell_to_output_scale_a,
         integer_lstm_param->effective_cell_to_output_scale_b,
 
-        proj_weight_ptr, integer_lstm_param->effective_proj_scale_a,
+        projection_weight_ptr, integer_lstm_param->effective_proj_scale_a,
         integer_lstm_param->effective_proj_scale_b,
 
         layer_norm_input_weight_ptr,
@@ -2214,7 +2216,7 @@ TfLiteStatus EvalInteger8x8_8(
         integer_lstm_param->layer_norm_output_scale_b,
 
         input_gate_bias_ptr, forget_gate_bias_ptr, cell_gate_bias_ptr,
-        output_gate_bias_ptr, proj_bias_ptr,
+        output_gate_bias_ptr, projection_bias_ptr,
 
         params, integer_lstm_param->intermediate_scale_a,
         integer_lstm_param->intermediate_scale_b,

From bd98ba765aac3e20c9db89131aa3016964749133 Mon Sep 17 00:00:00 2001
From: Haoyu Zhang 
Date: Fri, 19 Jun 2020 13:00:32 -0700
Subject: [PATCH 327/867] Fix cancellation race condition in
 BaseRendezvousMgr::RegisterCall

PiperOrigin-RevId: 317363743
Change-Id: Ide89dd360a9885b5e8f67b12f362cbce8cb85d80
---
 .../base_rendezvous_mgr.cc                    | 67 +++++++++----------
 .../distributed_runtime/base_rendezvous_mgr.h | 16 ++---
 .../rpc/rpc_rendezvous_mgr.cc                 |  1 +
 3 files changed, 37 insertions(+), 47 deletions(-)

diff --git a/tensorflow/core/distributed_runtime/base_rendezvous_mgr.cc b/tensorflow/core/distributed_runtime/base_rendezvous_mgr.cc
index 7849e094cb9..4b398e4ecef 100644
--- a/tensorflow/core/distributed_runtime/base_rendezvous_mgr.cc
+++ b/tensorflow/core/distributed_runtime/base_rendezvous_mgr.cc
@@ -139,7 +139,7 @@ Status BaseRemoteRendezvous::Initialize(WorkerSession* session) {
   CHECK_NE(session, nullptr) << "session must not be null!";
   std::vector deferred_calls;
   {
-    mutex_lock l(init_mu_);
+    mutex_lock l(mu_);
     if (session_ != nullptr) {
       if (session_->worker_name() == session->worker_name()) {
         VLOG(1) << "Skipping rendezvous re-initialization.";
@@ -161,12 +161,12 @@ Status BaseRemoteRendezvous::Initialize(WorkerSession* session) {
 }
 
 WorkerSession* BaseRemoteRendezvous::session() {
-  tf_shared_lock l(init_mu_);
+  tf_shared_lock l(mu_);
   return session_;
 }
 
 bool BaseRemoteRendezvous::is_initialized() {
-  tf_shared_lock l(init_mu_);
+  tf_shared_lock l(mu_);
   return is_initialized_locked();
 }
 
@@ -176,7 +176,7 @@ Status BaseRemoteRendezvous::Send(const Rendezvous::ParsedKey& parsed,
   VLOG(1) << "BaseRemoteRendezvous Send " << this << " " << parsed.FullKey();
   WorkerSession* sess = nullptr;
   {
-    tf_shared_lock l(init_mu_);
+    tf_shared_lock l(mu_);
     if (!status_.ok()) return status_;
     DCHECK(is_initialized_locked());
     sess = session_;
@@ -198,7 +198,7 @@ Status BaseRemoteRendezvous::ValidateDevices(const ParsedKey& parsed,
   // (e.g. calling session())
   WorkerSession* sess = nullptr;
   {
-    tf_shared_lock l(init_mu_);
+    tf_shared_lock l(mu_);
     if (!status_.ok()) return status_;
     if (!is_initialized_locked()) {
       return errors::Internal("ValidateDevices called before initialization.");
@@ -345,7 +345,7 @@ void BaseRemoteRendezvous::RecvLocalAsync(const ParsedKey& parsed,
   // Test whether the rendezvous is initialized using a shared lock, to avoid
   // the need for exclusive access in the common case.
   if (TF_PREDICT_FALSE(!is_initialized())) {
-    mutex_lock l(init_mu_);
+    mutex_lock l(mu_);
     if (!is_initialized_locked()) {
       // RecvLocalAsync can be called (due to an incoming RecvTensor RPC from a
       // remote worker) before the RunStep (or PartialRunStep) RPC from the
@@ -386,8 +386,7 @@ void BaseRemoteRendezvous::StartAbort(const Status& s) {
   local_->StartAbort(derived_status);
   {
     // Aborts all active RecvTensor calls.
-    mutex_lock l(init_mu_);
-    mutex_lock l2(active_mu_);
+    mutex_lock l(mu_);
     if (status_.ok()) {
       status_ = derived_status;
       for (auto& entry : active_) {
@@ -402,42 +401,36 @@ void BaseRemoteRendezvous::StartAbort(const Status& s) {
 void BaseRemoteRendezvous::RegisterCall(BaseRecvTensorCall* call,
                                         const Rendezvous::Args& args) {
   CancellationManager* cm = args.cancellation_manager;
-  Status captured_status;
-  {
-    tf_shared_lock l(init_mu_);
-    if (!status_.ok()) {
-      captured_status = status_;
-    }
-  }
-  if (!captured_status.ok()) {
-    call->StartAbort(captured_status);
-    return;
-  }
-
   bool already_cancelled = false;
   InactiveCallback callback = [] {};
-  if (cm != nullptr) {
-    auto token = cm->get_cancellation_token();
-    already_cancelled = !cm->RegisterCallback(token, [this, call] {
-      {
-        tf_shared_lock l(active_mu_);
-        if (active_.find(call) == active_.end()) return;
-      }
+  {
+    mutex_lock l(mu_);
+    if (!status_.ok()) {
+      call->StartAbort(status_);
+      return;
+    }
+    if (cm != nullptr) {
+      auto token = cm->get_cancellation_token();
+      already_cancelled = !cm->RegisterCallback(token, [this, call] {
+        {
+          mutex_lock l(mu_);
+          if (active_.find(call) == active_.end()) return;
+          call->StartAbort(
+              errors::Cancelled("RecvFromRemoteAsync is cancelled."));
+        }
+      });
+      callback = [cm, token] { cm->TryDeregisterCallback(token); };
+    }
+    if (already_cancelled) {
       call->StartAbort(errors::Cancelled("RecvFromRemoteAsync is cancelled."));
-    });
-    callback = [cm, token] { cm->TryDeregisterCallback(token); };
-  }
-
-  if (already_cancelled) {
-    call->StartAbort(errors::Cancelled("RecvFromRemoteAsync is cancelled."));
-  } else {
-    mutex_lock l(active_mu_);
-    CHECK(active_.emplace(call, callback).second);
+    } else {
+      CHECK(active_.emplace(call, callback).second);
+    }
   }
 }
 
 void BaseRemoteRendezvous::DeregisterCall(BaseRecvTensorCall* call) {
-  mutex_lock l(active_mu_);
+  mutex_lock l(mu_);
   auto it = active_.find(call);
   if (it != active_.end()) {
     // Deregister the cancellation callback, if one was registered.
diff --git a/tensorflow/core/distributed_runtime/base_rendezvous_mgr.h b/tensorflow/core/distributed_runtime/base_rendezvous_mgr.h
index afa0f74ea2c..63409a31549 100644
--- a/tensorflow/core/distributed_runtime/base_rendezvous_mgr.h
+++ b/tensorflow/core/distributed_runtime/base_rendezvous_mgr.h
@@ -174,14 +174,12 @@ class BaseRemoteRendezvous : public RemoteRendezvous {
  private:
   Rendezvous* local_;  // Owns a Ref on this object.
 
-  // Guards mutable state that is read-mostly after this rendezvous is
-  // initialized.
-  mutable mutex init_mu_;
+  mutable mutex mu_;
 
   // Status given by StartAbort() if any.
-  Status status_ TF_GUARDED_BY(init_mu_);
+  Status status_ TF_GUARDED_BY(mu_);
 
-  WorkerSession* session_ TF_GUARDED_BY(init_mu_);  // Not owned.
+  WorkerSession* session_ TF_GUARDED_BY(mu_);  // Not owned.
 
   // Data structures to handle calls when partially initialized.
   struct DeferredCall {
@@ -190,16 +188,14 @@ class BaseRemoteRendezvous : public RemoteRendezvous {
 
     DeferredCall(const ParsedKey& parsed, DoneCallback done);
   };
-  std::vector deferred_calls_ TF_GUARDED_BY(init_mu_);
+  std::vector deferred_calls_ TF_GUARDED_BY(mu_);
 
   typedef std::function InactiveCallback;
 
-  // Active outstanding RecvTensor calls.
-  mutex active_mu_;
   std::unordered_map active_
-      TF_GUARDED_BY(active_mu_);
+      TF_GUARDED_BY(mu_);
 
-  bool is_initialized_locked() TF_SHARED_LOCKS_REQUIRED(init_mu_) {
+  bool is_initialized_locked() TF_SHARED_LOCKS_REQUIRED(mu_) {
     return session_ != nullptr;
   }
 
diff --git a/tensorflow/core/distributed_runtime/rpc/rpc_rendezvous_mgr.cc b/tensorflow/core/distributed_runtime/rpc/rpc_rendezvous_mgr.cc
index 512c17fcfcf..89fe6ced725 100644
--- a/tensorflow/core/distributed_runtime/rpc/rpc_rendezvous_mgr.cc
+++ b/tensorflow/core/distributed_runtime/rpc/rpc_rendezvous_mgr.cc
@@ -282,6 +282,7 @@ void RpcRemoteRendezvous::RecvFromRemoteAsync(
     // callback.
     call->ReleaseWorker(sess->worker_cache());
     call->done()(call->status(), Args(), Args(), Tensor(), false);
+    DeregisterCall(call);
     get_call_freelist()->Release(call);
     return;
   }

From 5ed030734c63118c3fe09470a1d2da6af204a0eb Mon Sep 17 00:00:00 2001
From: Andrew Audibert 
Date: Fri, 19 Jun 2020 13:03:24 -0700
Subject: [PATCH 328/867] [tf.data service] Avoid holding locks during RPC
 calls.

This CL fixes a deadlock where a worker holds its lock while making an RPC to the master, and the master holds its lock while making an RPC to the worker. The RPCs require locks to serve, so we end up deadlocked.

We can avoid this by never holding a lock while performing RPCs. This CL modifies the master locking to release the lock when making the `ProcessTask` RPC to the worker.

This change shouldn't affect any functionality - it should only reduce the scope of some locking.

PiperOrigin-RevId: 317364346
Change-Id: I21e5ed8cdaced1192a89ffda4f8f93418e5dc4a5
---
 tensorflow/core/data/service/master_impl.cc | 131 ++++++++++++--------
 tensorflow/core/data/service/master_impl.h  |  19 ++-
 2 files changed, 94 insertions(+), 56 deletions(-)

diff --git a/tensorflow/core/data/service/master_impl.cc b/tensorflow/core/data/service/master_impl.cc
index 37a884d540e..5c7917b4154 100644
--- a/tensorflow/core/data/service/master_impl.cc
+++ b/tensorflow/core/data/service/master_impl.cc
@@ -61,7 +61,8 @@ Status DataServiceMasterImpl::RegisterWorker(
   VLOG(3) << "Received register worker request";
   mutex_lock l(mu_);
   int64 worker_id = next_worker_id_++;
-  workers_.emplace_back(worker_id, request->worker_address());
+  workers_.push_back(
+      std::make_shared(worker_id, request->worker_address()));
   response->set_worker_id(worker_id);
 
   // Allocate tasks to the worker.
@@ -70,17 +71,18 @@ Status DataServiceMasterImpl::RegisterWorker(
     if (job->finished()) {
       continue;
     }
-    int64 task_id = CreateTask(job.get(), request->worker_address());
+    const Task& task = CreateTaskLocked(job.get(), request->worker_address());
 
     TaskDef* task_def = response->add_tasks();
     *task_def->mutable_dataset() =
         datasets_by_id_[job->dataset_id()]->dataset_def();
     task_def->set_dataset_id(job->dataset_id());
     task_def->set_job_id(job->job_id());
-    task_def->set_task_id(task_id);
+    task_def->set_task_id(task.task_id());
   }
 
-  VLOG(1) << "Registered worker " << workers_.back().DebugString();
+  VLOG(1) << "Registered worker at address " << request->worker_address()
+          << " with id " << worker_id;
   return Status::OK();
 }
 
@@ -145,7 +147,6 @@ Status DataServiceMasterImpl::CreateJob(const CreateJobRequest* request,
   VLOG(3) << "Received create job request for dataset id "
           << request->dataset_id();
   ProcessingMode processing_mode = ProcessingMode(request->processing_mode());
-  mutex_lock l(mu_);
   int64 job_id;
   TF_RETURN_IF_ERROR(CreateJob(request->dataset_id(), processing_mode,
                                absl::optional(), &job_id));
@@ -161,25 +162,30 @@ Status DataServiceMasterImpl::GetOrCreateJob(
   VLOG(3) << "Received get or create job request for dataset id "
           << request->dataset_id() << " with name " << request->job_name()
           << " and index " << request->job_name_index();
-  mutex_lock l(mu_);
   NamedJobKey key(request->job_name(), request->job_name_index());
   ProcessingMode requested_processing_mode =
       ProcessingMode(request->processing_mode());
-  std::shared_ptr* job = gtl::FindOrNull(named_jobs_, key);
-  if (job != nullptr) {
-    TF_RETURN_IF_ERROR(ValidateMatchingJob(**job, requested_processing_mode,
-                                           request->dataset_id()));
-    int64 job_id = (*job)->job_id();
-    response->set_job_id(job_id);
-    VLOG(3) << "Found existing job for name=" << request->job_name()
-            << ", index=" << request->job_name_index()
-            << ". job_id: " << job_id;
-    return Status::OK();
+  {
+    mutex_lock l(mu_);
+    std::shared_ptr* job = gtl::FindOrNull(named_jobs_, key);
+    if (job != nullptr) {
+      TF_RETURN_IF_ERROR(ValidateMatchingJob(**job, requested_processing_mode,
+                                             request->dataset_id()));
+      int64 job_id = (*job)->job_id();
+      response->set_job_id(job_id);
+      VLOG(3) << "Found existing job for name=" << request->job_name()
+              << ", index=" << request->job_name_index()
+              << ". job_id: " << job_id;
+      return Status::OK();
+    }
   }
   int64 job_id;
   TF_RETURN_IF_ERROR(CreateJob(request->dataset_id(), requested_processing_mode,
                                request->job_name(), &job_id));
-  named_jobs_[key] = jobs_[job_id];
+  {
+    mutex_lock l(mu_);
+    named_jobs_[key] = jobs_[job_id];
+  }
   response->set_job_id(job_id);
   VLOG(3) << "Created job " << job_id << " for dataset "
           << request->dataset_id() << " and name " << request->job_name();
@@ -211,8 +217,7 @@ Status DataServiceMasterImpl::ValidateMatchingJob(
 Status DataServiceMasterImpl::CreateJob(int64 dataset_id,
                                         ProcessingMode processing_mode,
                                         absl::optional job_name,
-                                        int64* out_job_id)
-    EXCLUSIVE_LOCKS_REQUIRED(mu_) {
+                                        int64* out_job_id) LOCKS_EXCLUDED(mu_) {
   switch (processing_mode) {
     case ProcessingMode::PARALLEL_EPOCHS:
       break;
@@ -225,41 +230,64 @@ Status DataServiceMasterImpl::CreateJob(int64 dataset_id,
                                    ProcessingModeToString(processing_mode),
                                    " not recognized");
   }
-  if (!datasets_by_id_.contains(dataset_id)) {
-    return errors::NotFound("Dataset id: <", dataset_id, "> not found.");
+  std::shared_ptr job;
+  std::vector> workers;
+  {
+    mutex_lock l(mu_);
+    if (!datasets_by_id_.contains(dataset_id)) {
+      return errors::NotFound("Dataset id: <", dataset_id, "> not found.");
+    }
+
+    int64 job_id = next_job_id_++;
+    DCHECK(!jobs_.contains(job_id));
+    job = std::make_shared(job_id, dataset_id, processing_mode, job_name);
+    jobs_[job_id] = job;
+
+    // Copy workers_ so that we can iterate through the workers without holding
+    // the lock. When a new worker is added in `RegisterWorker`, we iterate
+    // through the jobs in `jobs_` and give it a task for each job. So even if a
+    // new worker is registered after we release the lock, because this job has
+    // been added to `jobs_`, it will still receive a task for this job.
+    workers = workers_;
+    const Dataset& dataset = *datasets_by_id_[dataset_id];
+    if (VLOG_IS_ON(1)) {
+      VLOG(1) << "Sending tasks to workers for job " << job->job_id()
+              << ". Dataset id: " << dataset_id
+              << ". Dataset fingerprint: " << dataset.fingerprint()
+              << ". Dataset definition size: "
+              << datasets_by_id_[dataset_id]->dataset_def().ByteSizeLong();
+    }
   }
 
-  int64 job_id = next_job_id_++;
-  DCHECK(!jobs_.contains(job_id));
-  auto job =
-      std::make_shared(job_id, dataset_id, processing_mode, job_name);
-  jobs_[job_id] = job;
-
-  for (auto& worker : workers_) {
-    int64 task_id = CreateTask(job.get(), worker.address());
-
-    // TODO(aaudibert): perform these calls asynchronously.
-    // TODO(aaudibert): clean up in case some calls succeed, but later calls
-    // fail
-    TF_RETURN_IF_ERROR(AllocateTaskToWorker(tasks_.at(task_id), &worker));
+  for (auto& worker : workers) {
+    const Task& task = CreateTask(job.get(), worker->address());
+    Status s = AllocateTaskToWorker(task, worker.get());
+    if (!s.ok()) {
+      LOG(WARNING) << "Failed to allocate task with id " << task.task_id()
+                   << " to worker at address " << worker->address() << ": "
+                   << s.error_message();
+    }
   }
+  VLOG(1) << "Done sending tasks to workers for job " << job->job_id();
 
-  *out_job_id = job_id;
+  *out_job_id = job->job_id();
   return Status::OK();
 }
 
-int64 DataServiceMasterImpl::CreateTask(Job* job,
-                                        const std::string& worker_address)
-    EXCLUSIVE_LOCKS_REQUIRED(mu_) {
+const DataServiceMasterImpl::Task& DataServiceMasterImpl::CreateTask(
+    Job* job, const std::string& worker_address) LOCKS_EXCLUDED(mu_) {
+  mutex_lock l(mu_);
+  return CreateTaskLocked(job, worker_address);
+}
+
+const DataServiceMasterImpl::Task& DataServiceMasterImpl::CreateTaskLocked(
+    Job* job, const std::string& worker_address) EXCLUSIVE_LOCKS_REQUIRED(mu_) {
   int64 task_id = next_task_id_++;
   DCHECK(!tasks_.contains(task_id));
-  auto result =
-      tasks_.emplace(std::piecewise_construct, std::forward_as_tuple(task_id),
-                     std::forward_as_tuple(task_id, job->job_id(),
-                                           job->dataset_id(), worker_address));
+  tasks_.insert({task_id, Task(task_id, job->job_id(), job->dataset_id(),
+                               worker_address)});
   job->add_task_id(task_id);
-  DCHECK(result.second);
-  return task_id;
+  return tasks_.at(task_id);
 }
 
 Status DataServiceMasterImpl::EnsureWorkerStubInitialized(Worker* worker) {
@@ -273,14 +301,17 @@ Status DataServiceMasterImpl::EnsureWorkerStubInitialized(Worker* worker) {
 
 Status DataServiceMasterImpl::AllocateTaskToWorker(const Task& task,
                                                    Worker* worker)
-    EXCLUSIVE_LOCKS_REQUIRED(mu_) {
+    LOCKS_EXCLUDED(mu_) {
   TF_RETURN_IF_ERROR(EnsureWorkerStubInitialized(worker));
   grpc::ClientContext client_ctx;
   ProcessTaskRequest req;
   req.mutable_task()->set_dataset_id(task.dataset_id());
-  DCHECK(datasets_by_id_.contains(task.dataset_id()));
-  *req.mutable_task()->mutable_dataset() =
-      datasets_by_id_.at(task.dataset_id())->dataset_def();
+  {
+    mutex_lock l(mu_);
+    DCHECK(datasets_by_id_.contains(task.dataset_id()));
+    *req.mutable_task()->mutable_dataset() =
+        datasets_by_id_.at(task.dataset_id())->dataset_def();
+  }
   req.mutable_task()->set_task_id(task.task_id());
   ProcessTaskResponse resp;
   grpc::Status s = worker->stub()->ProcessTask(&client_ctx, req, &resp);
@@ -321,8 +352,8 @@ Status DataServiceMasterImpl::GetWorkers(const GetWorkersRequest* request,
   VLOG(3) << "Enter GetWorkers";
   for (auto& worker : workers_) {
     WorkerInfo* info = response->add_workers();
-    info->set_address(worker.address());
-    info->set_id(worker.worker_id());
+    info->set_address(worker->address());
+    info->set_id(worker->worker_id());
   }
   VLOG(3) << "Returning list of " << workers_.size()
           << " workers from GetWorkers";
diff --git a/tensorflow/core/data/service/master_impl.h b/tensorflow/core/data/service/master_impl.h
index 0dc049a389c..67df2613118 100644
--- a/tensorflow/core/data/service/master_impl.h
+++ b/tensorflow/core/data/service/master_impl.h
@@ -177,16 +177,23 @@ class DataServiceMasterImpl {
   };
 
   // Registers a dataset with the given fingerprint, returning a new dataset id.
-  int64 RegisterDataset(uint64 fingerprint, const DatasetDef& dataset);
+  int64 RegisterDataset(uint64 fingerprint, const DatasetDef& dataset)
+      EXCLUSIVE_LOCKS_REQUIRED(mu_);
   // Initializes a workers stub, if it hasn't been initialized already.
   Status EnsureWorkerStubInitialized(Worker* worker);
   // Instructs a worker to begin processing a task.
-  Status AllocateTaskToWorker(const Task& task_id, Worker* worker);
+  Status AllocateTaskToWorker(const Task& task_id, Worker* worker)
+      LOCKS_EXCLUDED(mu_);
   // Creates a job and stores its job_id in `*job_id`.
   Status CreateJob(int64 dataset_id, ProcessingMode processing_mode,
-                   absl::optional job_name, int64* out_job_id);
-  // Creates a new task for a job, returning the new task's id.
-  int64 CreateTask(Job* job, const std::string& worker_address);
+                   absl::optional job_name, int64* out_job_id)
+      LOCKS_EXCLUDED(mu_);
+  // Creates a new task for a job, returning a reference to the task.
+  const Task& CreateTask(Job* job, const std::string& worker_address)
+      LOCKS_EXCLUDED(mu_);
+  // Same as `CreateTask`, but expects that the master lock is already held.
+  const Task& CreateTaskLocked(Job* job, const std::string& worker_address)
+      EXCLUSIVE_LOCKS_REQUIRED(mu_);
   // Validates that an existing job matches the given processing_mode and
   // dataset_id, returning an error status describing any difference.
   Status ValidateMatchingJob(const Job& job, ProcessingMode processing_mode,
@@ -202,7 +209,7 @@ class DataServiceMasterImpl {
   int64 next_task_id_ TF_GUARDED_BY(mu_) = 0;
 
   // Registered workers.
-  std::vector workers_ TF_GUARDED_BY(mu_);
+  std::vector> workers_ TF_GUARDED_BY(mu_);
   // Registered datasets, keyed by dataset ids.
   absl::flat_hash_map> datasets_by_id_
       TF_GUARDED_BY(mu_);

From 3ae2cf96105c32a71db7796b901e3c8ae5d8f71a Mon Sep 17 00:00:00 2001
From: "A. Unique TensorFlower" 
Date: Fri, 19 Jun 2020 13:05:16 -0700
Subject: [PATCH 329/867] Suppress Clang error: unused variable 'row_base' when
 building with MSAN instrumentation

PiperOrigin-RevId: 317364738
Change-Id: Ic55fdfa9520f341d254aa42333da67ae5d340680
---
 tensorflow/core/kernels/eigen_contraction_kernel.h | 2 ++
 1 file changed, 2 insertions(+)

diff --git a/tensorflow/core/kernels/eigen_contraction_kernel.h b/tensorflow/core/kernels/eigen_contraction_kernel.h
index 7233020c1c1..ef4b9dbc012 100644
--- a/tensorflow/core/kernels/eigen_contraction_kernel.h
+++ b/tensorflow/core/kernels/eigen_contraction_kernel.h
@@ -171,6 +171,7 @@ struct mkldnn_gemm_kernel
Date: Fri, 19 Jun 2020 13:06:52 -0700
Subject: [PATCH 330/867] Remove `run_deprecated_v1` annotations from
 collective ops tests.

PiperOrigin-RevId: 317365063
Change-Id: Ibf13ad8629947becd40038d41ee213d3466b6292
---
 tensorflow/python/ops/collective_ops_test.py | 380 ++++++++++---------
 1 file changed, 203 insertions(+), 177 deletions(-)

diff --git a/tensorflow/python/ops/collective_ops_test.py b/tensorflow/python/ops/collective_ops_test.py
index 9727593a1c5..8e3a95d7dbf 100644
--- a/tensorflow/python/ops/collective_ops_test.py
+++ b/tensorflow/python/ops/collective_ops_test.py
@@ -104,39 +104,42 @@ class CollectiveOpTest(test.TestCase):
     for i in range(group_size * num_instances):
       self.assertAllClose(results[i], expected, rtol=1e-5, atol=1e-5)
 
-  @test_util.run_deprecated_v1
   def testCollectiveReduce(self):
-    self._testCollectiveReduce(
-        inputs=[[0.1, 1.1, 2.1, 3.1, 4.1, 5.1, 6.1, 7.1],
-                [0.3, 1.3, 2.3, 3.3, 4.3, 5.3, 6.3, 7.3]],
-        expected=[0.2, 1.2, 2.2, 3.2, 4.2, 5.2, 6.2, 7.2],
-        set_graph_key=True)
+    # Tests that execute collectives need to be enclosed in graph or tf.function
+    with ops.Graph().as_default():
+      self._testCollectiveReduce(
+          inputs=[[0.1, 1.1, 2.1, 3.1, 4.1, 5.1, 6.1, 7.1],
+                  [0.3, 1.3, 2.3, 3.3, 4.3, 5.3, 6.3, 7.3]],
+          expected=[0.2, 1.2, 2.2, 3.2, 4.2, 5.2, 6.2, 7.2],
+          set_graph_key=True)
 
-  @test_util.run_deprecated_v1
   def testCollectiveAutoGraphKey(self):
-    self._testCollectiveReduce(
-        inputs=[[0.1, 1.1, 2.1, 3.1, 4.1, 5.1, 6.1, 7.1],
-                [0.3, 1.3, 2.3, 3.3, 4.3, 5.3, 6.3, 7.3]],
-        expected=[0.2, 1.2, 2.2, 3.2, 4.2, 5.2, 6.2, 7.2],
-        set_graph_key=False)
+    # Tests that execute collectives need to be enclosed in graph or tf.function
+    with ops.Graph().as_default():
+      self._testCollectiveReduce(
+          inputs=[[0.1, 1.1, 2.1, 3.1, 4.1, 5.1, 6.1, 7.1],
+                  [0.3, 1.3, 2.3, 3.3, 4.3, 5.3, 6.3, 7.3]],
+          expected=[0.2, 1.2, 2.2, 3.2, 4.2, 5.2, 6.2, 7.2],
+          set_graph_key=False)
 
-  @test_util.run_deprecated_v1
   def testFp16Reduce(self):
-    self._testCollectiveReduce(
-        inputs=[[0.1, 1.1, 2.1, 3.1, 4.1, 5.1, 6.1, 7.1],
-                [0.3, 1.3, 2.3, 3.3, 4.3, 5.3, 6.3, 7.3]],
-        expected=[0.2, 1.2, 2.2, 3.2, 4.2, 5.2, 6.2, 7.2],
-        set_graph_key=True,
-        fp16=True)
+    # Tests that execute collectives need to be enclosed in graph or tf.function
+    with ops.Graph().as_default():
+      self._testCollectiveReduce(
+          inputs=[[0.1, 1.1, 2.1, 3.1, 4.1, 5.1, 6.1, 7.1],
+                  [0.3, 1.3, 2.3, 3.3, 4.3, 5.3, 6.3, 7.3]],
+          expected=[0.2, 1.2, 2.2, 3.2, 4.2, 5.2, 6.2, 7.2],
+          set_graph_key=True,
+          fp16=True)
 
-  @test_util.run_deprecated_v1
   def testCollectiveMultipleConcurrentReduce(self):
-    self._testMultipleConcurrentCollectiveReduce(
-        [0.1, 1.1, 2.1, 3.1, 4.1, 5.1, 6.1, 7.1],
-        [0.3, 1.3, 2.3, 3.3, 4.3, 5.3, 6.3, 7.3],
-        [0.2, 1.2, 2.2, 3.2, 4.2, 5.2, 6.2, 7.2])
+    # Tests that execute collectives need to be enclosed in graph or tf.function
+    with ops.Graph().as_default():
+      self._testMultipleConcurrentCollectiveReduce(
+          [0.1, 1.1, 2.1, 3.1, 4.1, 5.1, 6.1, 7.1],
+          [0.3, 1.3, 2.3, 3.3, 4.3, 5.3, 6.3, 7.3],
+          [0.2, 1.2, 2.2, 3.2, 4.2, 5.2, 6.2, 7.2])
 
-  @test_util.run_deprecated_v1
   def testCollectiveTimeoutV1(self):
     timeout = 4.5
     kwargs = dict(
@@ -145,14 +148,17 @@ class CollectiveOpTest(test.TestCase):
         set_graph_key=True,
         timeout=timeout)
 
-    self._testCollectiveReduce(**kwargs)
+    # Tests that execute collectives need to be enclosed in graph or tf.function
+    with ops.Graph().as_default():
+      self._testCollectiveReduce(**kwargs)
 
     start_time = time.time()
-    with self.assertRaisesRegex(
-        errors.DeadlineExceededError,
-        'Collective has timed out waiting for other workers'):
-      self._testCollectiveReduce(
-          reported_group_size=len(kwargs['inputs']) + 1, **kwargs)
+    with ops.Graph().as_default():
+      with self.assertRaisesRegex(
+          errors.DeadlineExceededError,
+          'Collective has timed out waiting for other workers'):
+        self._testCollectiveReduce(
+            reported_group_size=len(kwargs['inputs']) + 1, **kwargs)
     elapsed = time.time() - start_time
     self.assertAllGreaterEqual(elapsed, timeout)
 
@@ -199,17 +205,18 @@ class CollectiveOpTest(test.TestCase):
     elapsed = time.time() - start_time
     self.assertAllGreaterEqual(elapsed, timeout)
 
-  @test_util.run_deprecated_v1
   def testNcclHintFallbackToRingReduce(self):
     """Tests that setting `communication_hint=nccl` works on non-GPU builds."""
     if kernels.get_registered_kernels_for_op('NcclAllReduce'):
       self.skipTest('Run only on non-GPU environments')
-    self._testCollectiveReduce(
-        inputs=[[0.1, 1.1, 2.1, 3.1, 4.1, 5.1, 6.1, 7.1],
-                [0.3, 1.3, 2.3, 3.3, 4.3, 5.3, 6.3, 7.3]],
-        expected=[0.2, 1.2, 2.2, 3.2, 4.2, 5.2, 6.2, 7.2],
-        set_graph_key=False,
-        communication_hint='nccl')
+    # Tests that execute collectives need to be enclosed in graph or tf.function
+    with ops.Graph().as_default():
+      self._testCollectiveReduce(
+          inputs=[[0.1, 1.1, 2.1, 3.1, 4.1, 5.1, 6.1, 7.1],
+                  [0.3, 1.3, 2.3, 3.3, 4.3, 5.3, 6.3, 7.3]],
+          expected=[0.2, 1.2, 2.2, 3.2, 4.2, 5.2, 6.2, 7.2],
+          set_graph_key=False,
+          communication_hint='nccl')
 
   def _testWhile(self, num_vars, num_iterations, key_base):
     group_size = 2
@@ -262,15 +269,16 @@ class CollectiveOpTest(test.TestCase):
           [((1 << (num_iterations + v)) * 1.) for v in range(num_vars)]
           for _ in range(group_size)])
 
-  @test_util.run_deprecated_v1
   def testSimpleWhile(self):
-    self._testWhile(num_vars=1, num_iterations=4, key_base=20)
+    # Tests that execute collectives need to be enclosed in graph or tf.function
+    with ops.Graph().as_default():
+      self._testWhile(num_vars=1, num_iterations=4, key_base=20)
 
-  @test_util.run_deprecated_v1
   def testWhileMultipleAllReduce(self):
-    self._testWhile(num_vars=2, num_iterations=4, key_base=20)
+    # Tests that execute collectives need to be enclosed in graph or tf.function
+    with ops.Graph().as_default():
+      self._testWhile(num_vars=2, num_iterations=4, key_base=20)
 
-  @test_util.run_deprecated_v1
   def testWhileWithScopedAllocator(self):
     group_size = 2
     group_key = 1
@@ -284,47 +292,52 @@ class CollectiveOpTest(test.TestCase):
     del rewrite_options.scoped_allocator_opts.enable_op[:]
     rewrite_options.scoped_allocator_opts.enable_op.append('CollectiveReduce')
 
-    with self.session(config=config) as sess:
-      run_ops = []
-      for i in range(group_size):
-        with ops.device('CPU:%d' % i):
-          constant = constant_op.constant(0.)
-          cond = lambda i: math_ops.less(i, 10.)
-          body = lambda i: math_ops.add(i, 1.)
-          input0 = control_flow_ops.while_loop(cond, body, [constant])
-          input1 = math_ops.add(constant, 5)
-          colred0 = collective_ops.all_reduce(input0, group_size, group_key,
-                                              instance_key0, 'Add', 'Id')
-          colred1 = collective_ops.all_reduce(input1, group_size, group_key,
-                                              instance_key1, 'Add', 'Id')
-          run_ops.append(math_ops.add_n([colred0, colred1]))
-      results = sess.run(run_ops)
+    # Tests that execute collectives need to be enclosed in graph or tf.function
+    with ops.Graph().as_default():
+      with self.session(config=config) as sess:
+        run_ops = []
+        for i in range(group_size):
+          with ops.device('CPU:%d' % i):
+            constant = constant_op.constant(0.)
+            cond = lambda i: math_ops.less(i, 10.)
+            body = lambda i: math_ops.add(i, 1.)
+            input0 = control_flow_ops.while_loop(cond, body, [constant])
+            input1 = math_ops.add(constant, 5)
+            colred0 = collective_ops.all_reduce(input0, group_size, group_key,
+                                                instance_key0, 'Add', 'Id')
+            colred1 = collective_ops.all_reduce(input1, group_size, group_key,
+                                                instance_key1, 'Add', 'Id')
+            run_ops.append(math_ops.add_n([colred0, colred1]))
+        results = sess.run(run_ops)
       self.assertEqual(results, [30., 30.])
 
-  @test_util.run_deprecated_v1
   def testCollectiveReduceScalar(self):
-    self._testCollectiveReduce(inputs=[0.1, 0.3], expected=0.2,
-                               set_graph_key=True)
+    # Tests that execute collectives need to be enclosed in graph or tf.function
+    with ops.Graph().as_default():
+      self._testCollectiveReduce(inputs=[0.1, 0.3], expected=0.2,
+                                 set_graph_key=True)
 
-  @test_util.run_deprecated_v1
   def testCollectiveReduceMaximum(self):
-    self._testCollectiveReduce(
-        inputs=[[1., 20., 3., 40., 5.], [10., 2., 30., 4., 50.]],
-        expected=[10., 20., 30., 40., 50.],
-        set_graph_key=True,
-        instance_key=30,
-        merge_op='Max',
-        final_op='Id')
+    # Tests that execute collectives need to be enclosed in graph or tf.function
+    with ops.Graph().as_default():
+      self._testCollectiveReduce(
+          inputs=[[1., 20., 3., 40., 5.], [10., 2., 30., 4., 50.]],
+          expected=[10., 20., 30., 40., 50.],
+          set_graph_key=True,
+          instance_key=30,
+          merge_op='Max',
+          final_op='Id')
 
-  @test_util.run_deprecated_v1
   def testCollectiveReduceMinimum(self):
-    self._testCollectiveReduce(
-        inputs=[[1., 20., 3., 40., 5.], [10., 2., 30., 4., 50.]],
-        expected=[1., 2., 3., 4., 5.],
-        set_graph_key=True,
-        instance_key=40,
-        merge_op='Min',
-        final_op='Id')
+    # Tests that execute collectives need to be enclosed in graph or tf.function
+    with ops.Graph().as_default():
+      self._testCollectiveReduce(
+          inputs=[[1., 20., 3., 40., 5.], [10., 2., 30., 4., 50.]],
+          expected=[1., 2., 3., 4., 5.],
+          set_graph_key=True,
+          instance_key=40,
+          merge_op='Min',
+          final_op='Id')
 
   def _testCollectiveBroadcast(self, in_val):
     group_key = 1
@@ -345,13 +358,15 @@ class CollectiveOpTest(test.TestCase):
     self.assertAllClose(results[0], in_val, rtol=1e-5, atol=1e-5)
     self.assertAllClose(results[1], in_val, rtol=1e-5, atol=1e-5)
 
-  @test_util.run_deprecated_v1
   def testCollectiveBroadcast(self):
-    self._testCollectiveBroadcast([0.1, 1.1, 2.1, 3.1, 4.1, 5.1, 6.1, 7.1])
+    # Tests that execute collectives need to be enclosed in graph or tf.function
+    with ops.Graph().as_default():
+      self._testCollectiveBroadcast([0.1, 1.1, 2.1, 3.1, 4.1, 5.1, 6.1, 7.1])
 
-  @test_util.run_deprecated_v1
   def testCollectiveBroadcastBool(self):
-    self._testCollectiveBroadcast([True, False])
+    # Tests that execute collectives need to be enclosed in graph or tf.function
+    with ops.Graph().as_default():
+      self._testCollectiveBroadcast([True, False])
 
   def _testCollectiveGather(self, t0, t1, expected, set_graph_key):
     group_key = 1
@@ -371,94 +386,101 @@ class CollectiveOpTest(test.TestCase):
     self.assertAllClose(results[0], expected, rtol=1e-5, atol=1e-5)
     self.assertAllClose(results[1], expected, rtol=1e-5, atol=1e-5)
 
-  @test_util.run_deprecated_v1
   def testCollectiveGather(self):
-    self._testCollectiveGather([0, 1, 2, 3, 4, 5, 6, 7],
-                               [10, 11, 12, 13, 14, 15, 16, 17],
-                               [0, 1, 2, 3, 4, 5, 6, 7,
-                                10, 11, 12, 13, 14, 15, 16, 17],
-                               True)
-    self._testCollectiveGather([[0, 1, 2, 3], [4, 5, 6, 7]],
-                               [[10, 11, 12, 13], [14, 15, 16, 17]],
-                               [[0, 1, 2, 3], [4, 5, 6, 7],
-                                [10, 11, 12, 13], [14, 15, 16, 17]],
-                               True)
-    self._testCollectiveGather([[[0, 1], [2, 3]], [[4, 5], [6, 7]]],
-                               [[[10, 11], [12, 13]], [[14, 15], [16, 17]]],
-                               [[[0, 1], [2, 3]], [[4, 5], [6, 7]],
-                                [[10, 11], [12, 13]], [[14, 15], [16, 17]]],
-                               True)
+    # Tests that execute collectives need to be enclosed in graph or tf.function
+    with ops.Graph().as_default():
+      self._testCollectiveGather([0, 1, 2, 3, 4, 5, 6, 7],
+                                 [10, 11, 12, 13, 14, 15, 16, 17],
+                                 [0, 1, 2, 3, 4, 5, 6, 7,
+                                  10, 11, 12, 13, 14, 15, 16, 17],
+                                 True)
+      self._testCollectiveGather([[0, 1, 2, 3], [4, 5, 6, 7]],
+                                 [[10, 11, 12, 13], [14, 15, 16, 17]],
+                                 [[0, 1, 2, 3], [4, 5, 6, 7],
+                                  [10, 11, 12, 13], [14, 15, 16, 17]],
+                                 True)
+      self._testCollectiveGather([[[0, 1], [2, 3]], [[4, 5], [6, 7]]],
+                                 [[[10, 11], [12, 13]], [[14, 15], [16, 17]]],
+                                 [[[0, 1], [2, 3]], [[4, 5], [6, 7]],
+                                  [[10, 11], [12, 13]], [[14, 15], [16, 17]]],
+                                 True)
 
-  @test_util.run_deprecated_v1
   def testCollectiveGatherShapeMismatch(self):
     group_key = 1
     instance_key = 1
     t0 = [1, 2, 3, 4]
     t1 = [5, 6, 7, 8]
     t2 = [9, 10]
-    with self.session(
-        config=config_pb2.ConfigProto(device_count={'CPU': 2})) as sess:
-      with ops.device('/CPU:0'):
-        in0 = constant_op.constant(t0)
-        c0 = collective_ops.all_gather(in0, 2, group_key, instance_key)
-      with ops.device('/CPU:1'):
-        in1 = constant_op.constant(t1)
-        in2 = constant_op.constant(t2)
-        c1 = collective_ops.all_gather(in1, 2, group_key, instance_key)
-        c2 = collective_ops.all_gather(in2, 2, group_key, instance_key)
-      run_options = config_pb2.RunOptions()
-      run_options.experimental.collective_graph_key = 1
-      sess.run([c0, c1], options=run_options)
-      with self.assertRaisesRegexp(errors.InvalidArgumentError,
-                                   'Shape mismatch'):
-        sess.run([c0, c2], options=run_options)
+    # Tests that execute collectives need to be enclosed in graph or tf.function
+    with ops.Graph().as_default():
+      with self.session(
+          config=config_pb2.ConfigProto(device_count={'CPU': 2})) as sess:
+        with ops.device('/CPU:0'):
+          in0 = constant_op.constant(t0)
+          c0 = collective_ops.all_gather(in0, 2, group_key, instance_key)
+        with ops.device('/CPU:1'):
+          in1 = constant_op.constant(t1)
+          in2 = constant_op.constant(t2)
+          c1 = collective_ops.all_gather(in1, 2, group_key, instance_key)
+          c2 = collective_ops.all_gather(in2, 2, group_key, instance_key)
+        run_options = config_pb2.RunOptions()
+        run_options.experimental.collective_graph_key = 1
+        sess.run([c0, c1], options=run_options)
+        with self.assertRaisesRegexp(errors.InvalidArgumentError,
+                                     'Shape mismatch'):
+          sess.run([c0, c2], options=run_options)
 
-  @test_util.run_deprecated_v1
   def testCollectiveGatherShapeMismatchAcrossDevices(self):
     group_key = 1
     instance_key = 1
     t0 = [1, 2, 3, 4]
     t1 = [5, 6]
-    with self.session(
-        config=config_pb2.ConfigProto(device_count={'CPU': 2})) as sess:
-      with ops.device('/CPU:0'):
-        in0 = constant_op.constant(t0)
-        c0 = collective_ops.all_gather(in0, 2, group_key, instance_key)
-      with ops.device('/CPU:1'):
-        in1 = constant_op.constant(t1)
-        c1 = collective_ops.all_gather(in1, 2, group_key, instance_key)
-      run_options = config_pb2.RunOptions()
-      run_options.experimental.collective_graph_key = 1
-      with self.assertRaisesRegexp(errors.InvalidArgumentError,
-                                   'Shape mismatch'):
-        sess.run([c0, c1], options=run_options)
+    # Tests that execute collectives need to be enclosed in graph or tf.function
+    with ops.Graph().as_default():
+      with self.session(
+          config=config_pb2.ConfigProto(device_count={'CPU': 2})) as sess:
+        with ops.device('/CPU:0'):
+          in0 = constant_op.constant(t0)
+          c0 = collective_ops.all_gather(in0, 2, group_key, instance_key)
+        with ops.device('/CPU:1'):
+          in1 = constant_op.constant(t1)
+          c1 = collective_ops.all_gather(in1, 2, group_key, instance_key)
+        run_options = config_pb2.RunOptions()
+        run_options.experimental.collective_graph_key = 1
+        with self.assertRaisesRegexp(errors.InvalidArgumentError,
+                                     'Shape mismatch'):
+          sess.run([c0, c1], options=run_options)
 
-  @test_util.run_deprecated_v1
   def testCollectiveGatherPolymorphicShape(self):
     t0 = [0, 1, 2, 3, 4, 5, 6, 7]
     t1 = [10, 11, 12, 13, 14, 15, 16, 17]
     group_size = 2
     group_key = 1
     instance_key = 123
-    with self.session(
-        config=config_pb2.ConfigProto(
-            device_count={'CPU': group_size})) as sess:
-      with ops.device('/CPU:0'):
-        in0 = array_ops.placeholder(dtype=dtypes.int32, shape=[None])
-        c0 = collective_ops.all_gather(in0, group_size, group_key, instance_key)
-      with ops.device('/CPU:1'):
-        in1 = array_ops.placeholder(dtype=dtypes.int32, shape=[None])
-        c1 = collective_ops.all_gather(in1, group_size, group_key, instance_key)
+    # Tests that execute collectives need to be enclosed in graph or tf.function
+    with ops.Graph().as_default():
+      with self.session(
+          config=config_pb2.ConfigProto(
+              device_count={'CPU': group_size})) as sess:
+        with ops.device('/CPU:0'):
+          in0 = array_ops.placeholder(dtype=dtypes.int32, shape=[None])
+          c0 = collective_ops.all_gather(in0, group_size, group_key,
+                                         instance_key)
+        with ops.device('/CPU:1'):
+          in1 = array_ops.placeholder(dtype=dtypes.int32, shape=[None])
+          c1 = collective_ops.all_gather(in1, group_size, group_key,
+                                         instance_key)
 
-      results = sess.run([c0, c1], feed_dict={in0: t0, in1: t1})
-      expected_output = [0, 1, 2, 3, 4, 5, 6, 7, 10, 11, 12, 13, 14, 15, 16, 17]
-      self.assertAllClose(results[0], expected_output, rtol=1e-5, atol=1e-5)
-      self.assertAllClose(results[1], expected_output, rtol=1e-5, atol=1e-5)
+        results = sess.run([c0, c1], feed_dict={in0: t0, in1: t1})
+        results_ = sess.run([c0, c1], feed_dict={in0: t0[1:], in1: t1[1:]})
 
-      results_ = sess.run([c0, c1], feed_dict={in0: t0[1:], in1: t1[1:]})
-      expected_output_ = [1, 2, 3, 4, 5, 6, 7, 11, 12, 13, 14, 15, 16, 17]
-      self.assertAllClose(results_[0], expected_output_, rtol=1e-5, atol=1e-5)
-      self.assertAllClose(results_[1], expected_output_, rtol=1e-5, atol=1e-5)
+    expected_output = [0, 1, 2, 3, 4, 5, 6, 7, 10, 11, 12, 13, 14, 15, 16, 17]
+    self.assertAllClose(results[0], expected_output, rtol=1e-5, atol=1e-5)
+    self.assertAllClose(results[1], expected_output, rtol=1e-5, atol=1e-5)
+
+    expected_output_ = [1, 2, 3, 4, 5, 6, 7, 11, 12, 13, 14, 15, 16, 17]
+    self.assertAllClose(results_[0], expected_output_, rtol=1e-5, atol=1e-5)
+    self.assertAllClose(results_[1], expected_output_, rtol=1e-5, atol=1e-5)
 
   @test_util.run_v2_only
   def testCollectiveGroupSizeMismatch(self):
@@ -492,8 +514,17 @@ class CollectiveOpTest(test.TestCase):
                                  'but that group has size'):
       run_all_reduce()
 
-  @test_util.run_deprecated_v1
+  @test_util.run_v2_only
   def testCollectiveTensorsHaveNoDeviceSpecified(self):
+    context._reset_context()
+    cpus = config.list_physical_devices('CPU')
+    self.assertEqual(len(cpus), 1)
+    config.set_logical_device_configuration(cpus[0], [
+        context.LogicalDeviceConfiguration(),
+        context.LogicalDeviceConfiguration()
+    ])
+    context.ensure_initialized()
+
     group_size = 2
     group_key = 1
     instance_key = 1
@@ -517,20 +548,12 @@ class CollectiveOpTest(test.TestCase):
 
       return results
 
-    with self.session(config=config_pb2.ConfigProto(
-        device_count={'CPU': 2})) as sess:
-      with ops.device('/CPU:0'):
-        in0 = constant_op.constant(1)
-      with ops.device('/CPU:1'):
-        in1 = constant_op.constant(3)
-
-      result_op = fn([in0, in1])
-
-      run_options = config_pb2.RunOptions()
-      run_options.experimental.collective_graph_key = 1
-      result = sess.run(result_op, options=run_options)
-
-      self.assertAllClose(result, [2, 2])
+    with ops.device('/CPU:0'):
+      in0 = constant_op.constant(1)
+    with ops.device('/CPU:1'):
+      in1 = constant_op.constant(3)
+    result = fn([in0, in1])
+    self.assertAllClose(result, [2, 2])
 
   @test_util.run_v2_only
   def testCollectiveGroupSizeOne(self):
@@ -548,7 +571,6 @@ class CollectiveOpTest(test.TestCase):
         in_tensor, group_size, group_key, instance_key)
     self.assertAllEqual(in_value, gathered_tensor.numpy())
 
-  @test_util.run_deprecated_v1
   def testConstantWithScopedAllocator(self):
     group_size = 2
     group_key = 1
@@ -565,21 +587,25 @@ class CollectiveOpTest(test.TestCase):
     del rewrite_options.scoped_allocator_opts.enable_op[:]
     rewrite_options.scoped_allocator_opts.enable_op.append('CollectiveReduce')
 
-    with self.session(config=cfg) as sess:
-      run_ops = []
-      for i in range(group_size):
-        with ops.device('CPU:%d' % i):
-          constant = constant_op.constant(i + 1.)
-          input_tensor1 = array_ops.identity(constant)
-          input_tensor2 = array_ops.identity(constant)
-          reduced_tensor1 = collective_ops.all_reduce(
-              input_tensor1, group_size, group_key, instance_key1, 'Add', 'Id')
-          reduced_tensor2 = collective_ops.all_reduce(
-              input_tensor2, group_size, group_key, instance_key2, 'Add', 'Id')
-          run_ops.append(array_ops.identity(reduced_tensor1))
-          run_ops.append(array_ops.identity(reduced_tensor2))
-      results = sess.run(run_ops)
-      self.assertEqual(results, [3., 3., 3., 3.])
+    # Tests that execute collectives need to be enclosed in graph or tf.function
+    with ops.Graph().as_default():
+      with self.session(config=cfg) as sess:
+        run_ops = []
+        for i in range(group_size):
+          with ops.device('CPU:%d' % i):
+            constant = constant_op.constant(i + 1.)
+            input_tensor1 = array_ops.identity(constant)
+            input_tensor2 = array_ops.identity(constant)
+            reduced_tensor1 = collective_ops.all_reduce(
+                input_tensor1, group_size, group_key, instance_key1, 'Add',
+                'Id')
+            reduced_tensor2 = collective_ops.all_reduce(
+                input_tensor2, group_size, group_key, instance_key2, 'Add',
+                'Id')
+            run_ops.append(array_ops.identity(reduced_tensor1))
+            run_ops.append(array_ops.identity(reduced_tensor2))
+        results = sess.run(run_ops)
+    self.assertEqual(results, [3., 3., 3., 3.])
 
   @test_util.run_v2_only
   def testMultipleGroups(self):

From 72d30dfb8bc58be931604f853bd161a11b7c9fcc Mon Sep 17 00:00:00 2001
From: Haoyu Zhang 
Date: Fri, 19 Jun 2020 13:19:39 -0700
Subject: [PATCH 331/867] Raise error type corresponding to status code instead
 of generic RuntimeError.

PiperOrigin-RevId: 317367352
Change-Id: I35378b88a33269ac225632ae848398b819c694a1
---
 tensorflow/python/eager/pywrap_tensor.cc | 32 ++++++++++++++++--------
 1 file changed, 21 insertions(+), 11 deletions(-)

diff --git a/tensorflow/python/eager/pywrap_tensor.cc b/tensorflow/python/eager/pywrap_tensor.cc
index 031545531f1..0789eab6270 100644
--- a/tensorflow/python/eager/pywrap_tensor.cc
+++ b/tensorflow/python/eager/pywrap_tensor.cc
@@ -180,6 +180,15 @@ int ConvertDeviceName(PyObject* obj, const char** dst) {
   return 1;
 }
 
+void RaiseExceptionTypeFromTFStatus(TF_Status* status) {
+  TF_Code code = TF_GetCode(status);
+  PyObject* exception = tensorflow::PyExceptionRegistry::Lookup(code);
+  PyErr_SetObject(exception,
+                  pybind11::make_tuple(pybind11::none(), pybind11::none(),
+                                       TF_Message(status))
+                      .ptr());
+}
+
 }  // namespace
 
 namespace tensorflow {
@@ -305,13 +314,7 @@ TFE_TensorHandle* ConvertToEagerTensorUncached(TFE_Context* ctx,
                                                     device_name, status.get()));
     const TF_Code code = TF_GetCode(status.get());
     if (code != TF_OK) {
-      // Instead of raising a generic RuntimeError, raise an exception type
-      // based on the status error code.
-      PyObject* exception = PyExceptionRegistry::Lookup(code);
-      PyErr_SetObject(exception,
-                      pybind11::make_tuple(pybind11::none(), pybind11::none(),
-                                           TF_Message(status.get()))
-                          .ptr());
+      RaiseExceptionTypeFromTFStatus(status.get());
       return nullptr;
     }
   }
@@ -512,7 +515,9 @@ static PyObject* EagerTensor_datatype_enum(EagerTensor* self) {
 static PyObject* EagerTensor_shape_tuple(EagerTensor* self) {
   auto handle = self->handle;
   int n = TFE_TensorHandleNumDims(handle, &self->status);
-  if (MaybeRaiseExceptionFromTFStatus(&self->status, nullptr)) {
+  TF_Code code = TF_GetCode(&self->status);
+  if (code != TF_OK) {
+    RaiseExceptionTypeFromTFStatus(&self->status);
     // Cleanup self->status before returning.
     self->status.status = tensorflow::Status::OK();
     return nullptr;
@@ -522,13 +527,18 @@ static PyObject* EagerTensor_shape_tuple(EagerTensor* self) {
   for (int i = 0; i < n; ++i) {
     PyObject* dim =
         PyLong_FromLongLong(TFE_TensorHandleDim(handle, i, &self->status));
-    if (MaybeRaiseExceptionFromTFStatus(&self->status, nullptr) ||
-        dim == nullptr || PyTuple_SetItem(shape, i, dim) != 0) {
+    code = TF_GetCode(&self->status);
+    if (code != TF_OK || dim == nullptr ||
+        PyTuple_SetItem(shape, i, dim) != 0) {
+      if (code != TF_OK) {
+        RaiseExceptionTypeFromTFStatus(&self->status);
+      } else {
+        PyErr_SetString(PyExc_RuntimeError, "Error while creating shape");
+      }
       // Cleanup self->status before returning.
       self->status.status = tensorflow::Status::OK();
       Py_DECREF(shape);
       if (dim != nullptr) Py_DECREF(dim);
-      PyErr_SetString(PyExc_RuntimeError, "Error while creating shape");
       return nullptr;
     }
   }

From 4b3576c08171aee7edc80e929d2babdd899d67b7 Mon Sep 17 00:00:00 2001
From: Frank Chen 
Date: Fri, 19 Jun 2020 14:01:00 -0700
Subject: [PATCH 332/867] Change quick_exit() to exit() to fix Mac OS nightly
 build problems

PiperOrigin-RevId: 317375174
Change-Id: Ia0adc40f0d952cb478118fc07189ccd0f2d9a073
---
 tensorflow/core/tpu/kernels/tpu_compile_op_common.cc | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/tensorflow/core/tpu/kernels/tpu_compile_op_common.cc b/tensorflow/core/tpu/kernels/tpu_compile_op_common.cc
index 92d1fa1337e..79556cfa544 100644
--- a/tensorflow/core/tpu/kernels/tpu_compile_op_common.cc
+++ b/tensorflow/core/tpu/kernels/tpu_compile_op_common.cc
@@ -353,7 +353,7 @@ Status TpuCompileOpKernelCommon::CompileTFFunctionToHlo(
     return;
   }
 
-  std::quick_exit(42);
+  std::exit(42);
 }
 
 /* static */ Status TpuCompileOpKernelCommon::GetDynamicShapes(

From 83fe1bad15e65e8db5e546d683d0ad591f19fad7 Mon Sep 17 00:00:00 2001
From: Andrew Audibert 
Date: Fri, 19 Jun 2020 14:02:55 -0700
Subject: [PATCH 333/867] [tf.data service] Add test that different workers use
 independent shuffle orders.

If shuffle seeds are unspecified, shuffle order should be non-deterministically chosen on each worker.

PiperOrigin-RevId: 317375549
Change-Id: I35e32cfbbfb8558451a079875b495708347a23bf
---
 .../kernel_tests/data_service_ops_test.py     | 23 +++++++++++++++++++
 1 file changed, 23 insertions(+)

diff --git a/tensorflow/python/data/kernel_tests/data_service_ops_test.py b/tensorflow/python/data/kernel_tests/data_service_ops_test.py
index 796ab328980..488bf97f184 100644
--- a/tensorflow/python/data/kernel_tests/data_service_ops_test.py
+++ b/tensorflow/python/data/kernel_tests/data_service_ops_test.py
@@ -31,6 +31,7 @@ from tensorflow.python.eager import def_function
 from tensorflow.python.framework import combinations
 from tensorflow.python.framework import dtypes
 from tensorflow.python.framework import errors
+from tensorflow.python.framework import random_seed
 from tensorflow.python.ops import math_ops
 from tensorflow.python.ops import random_ops
 from tensorflow.python.ops import tensor_array_ops
@@ -78,6 +79,28 @@ class DataServiceOpsTest(test_base.DatasetTestBase, parameterized.TestCase):
     results = [elem.numpy() for elem in ds]
     self.assertEqual(list(range(num_elements)), results)
 
+  @combinations.generate(test_base.eager_only_combinations())
+  def testDifferentShuffleOrders(self):
+    random_seed.set_random_seed(None)
+    num_elements = 100
+    master_address = self.create_cluster(2)
+    ds = dataset_ops.Dataset.range(num_elements)
+    ds = ds.shuffle(num_elements)
+    ds = _make_distributed_dataset(ds, master_address)
+    output = [elem.numpy() for elem in ds]
+
+    # The output will be two sequences of range(num_elements)
+    # non-deterministically interleaved together. If the orders of the elements
+    # were the same, first_order and second_order computed below will be equal.
+    first_order = {}
+    second_order = {}
+    for element in output:
+      if element in first_order:
+        second_order[element] = len(second_order)
+      else:
+        first_order[element] = len(first_order)
+    self.assertNotEqual(first_order, second_order)
+
   @combinations.generate(test_base.eager_only_combinations())
   def testMultipleEpochs(self):
     num_elements = 3

From 94bf57d06c546df46b2af36cfa11dc62dd35aebb Mon Sep 17 00:00:00 2001
From: Sanjoy Das 
Date: Fri, 19 Jun 2020 14:07:16 -0700
Subject: [PATCH 334/867] Do not try to compile trivially dead branches in the
 Case tf2xla lowering

This is important for the upcoming DeviceIndex op which can be used to select
one of many implementations depending on the device, and some of them may not be
compilable by tf2xla.

PiperOrigin-RevId: 317376420
Change-Id: I6428df6f4da238e5d2bc3618d51c579e34454945
---
 tensorflow/compiler/tests/BUILD               | 20 +++++
 tensorflow/compiler/tests/case_test.py        | 87 +++++++++++++++++++
 tensorflow/compiler/tf2xla/kernels/BUILD      |  1 +
 tensorflow/compiler/tf2xla/kernels/case_op.cc | 46 ++++++++--
 tensorflow/compiler/tf2xla/kernels/case_op.h  | 11 ++-
 5 files changed, 156 insertions(+), 9 deletions(-)
 create mode 100644 tensorflow/compiler/tests/case_test.py

diff --git a/tensorflow/compiler/tests/BUILD b/tensorflow/compiler/tests/BUILD
index 034ec82de10..42353451408 100644
--- a/tensorflow/compiler/tests/BUILD
+++ b/tensorflow/compiler/tests/BUILD
@@ -1453,6 +1453,26 @@ tf_xla_py_test(
     ],
 )
 
+tf_xla_py_test(
+    name = "case_test",
+    size = "small",
+    srcs = ["case_test.py"],
+    disabled_backends = ["cpu_ondemand"],
+    python_version = "PY3",
+    tags = [
+        "no_pip",  # TODO(b/149738646): fix pip install so these tests run on kokoro pip
+    ],
+    use_xla_device = False,  # Uses tf.function(experimental_compile=True)
+    deps = [
+        ":xla_test",
+        "//tensorflow/compiler/tf2xla/python:xla",
+        "//tensorflow/python:array_ops",
+        "//tensorflow/python:framework",
+        "//tensorflow/python:platform_test",
+        "//tensorflow/python:training",
+    ],
+)
+
 tf_xla_py_test(
     name = "gather_test",
     size = "medium",
diff --git a/tensorflow/compiler/tests/case_test.py b/tensorflow/compiler/tests/case_test.py
new file mode 100644
index 00000000000..3b2dff537da
--- /dev/null
+++ b/tensorflow/compiler/tests/case_test.py
@@ -0,0 +1,87 @@
+# Copyright 2020 The TensorFlow Authors. All Rights Reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+# ==============================================================================
+"""Tests for while loops in XLA."""
+
+from __future__ import absolute_import
+from __future__ import division
+from __future__ import print_function
+
+from tensorflow.compiler.tests import xla_test
+from tensorflow.python.eager import def_function
+from tensorflow.python.framework import ops
+from tensorflow.python.ops import array_ops
+from tensorflow.python.ops import control_flow_ops
+from tensorflow.python.ops import image_ops
+from tensorflow.python.ops import io_ops
+from tensorflow.python.platform import test
+
+
+class CaseTest(xla_test.XLATestCase):
+
+  def testCaseBasic(self):
+
+    @def_function.function(experimental_compile=True)
+    def switch_case_test(branch_index):
+
+      def f1():
+        return array_ops.constant(17)
+
+      def f2():
+        return array_ops.constant(31)
+
+      def f3():
+        return array_ops.constant(-1)
+
+      return control_flow_ops.switch_case(
+          branch_index, branch_fns={
+              0: f1,
+              1: f2
+          }, default=f3)
+
+    with ops.device(self.device):
+      self.assertEqual(switch_case_test(array_ops.constant(0)).numpy(), 17)
+      self.assertEqual(switch_case_test(array_ops.constant(1)).numpy(), 31)
+      self.assertEqual(switch_case_test(array_ops.constant(2)).numpy(), -1)
+      self.assertEqual(switch_case_test(array_ops.constant(3)).numpy(), -1)
+
+  def testBranchIsPruned(self):
+
+    @def_function.function(experimental_compile=True)
+    def switch_case_test():
+      branch_index = array_ops.constant(0)
+
+      def f1():
+        return array_ops.constant(17)
+
+      def f2():
+        # Some operations that XLA cannot compile.
+        image_ops.decode_image(io_ops.read_file('/tmp/bmp'))
+        return array_ops.constant(31)
+
+      # This tests that we do not try to compile all branches if the branch
+      # index in trivially constant.
+      return control_flow_ops.switch_case(
+          branch_index, branch_fns={
+              0: f1,
+              1: f2
+          }, default=f2)
+
+    with ops.device(self.device):
+      self.assertEqual(switch_case_test().numpy(), 17)
+
+
+if __name__ == '__main__':
+  ops.enable_eager_execution()
+  test.main()
diff --git a/tensorflow/compiler/tf2xla/kernels/BUILD b/tensorflow/compiler/tf2xla/kernels/BUILD
index bfdfe38305b..bdaeeafd295 100644
--- a/tensorflow/compiler/tf2xla/kernels/BUILD
+++ b/tensorflow/compiler/tf2xla/kernels/BUILD
@@ -316,6 +316,7 @@ tf_kernel_library(
         "//tensorflow/compiler/tf2xla/ops:xla_ops",
         "//tensorflow/compiler/xla:literal",
         "//tensorflow/compiler/xla/client:xla_builder",
+        "//tensorflow/compiler/xla/client/lib:constants",
         "//tensorflow/core:framework",
         "//tensorflow/core:lib",
         "//tensorflow/core:protos_all_cc",
diff --git a/tensorflow/compiler/tf2xla/kernels/case_op.cc b/tensorflow/compiler/tf2xla/kernels/case_op.cc
index 1b15c09f7e3..fbd54f1ef39 100644
--- a/tensorflow/compiler/tf2xla/kernels/case_op.cc
+++ b/tensorflow/compiler/tf2xla/kernels/case_op.cc
@@ -21,13 +21,14 @@ limitations under the License.
 #include "tensorflow/compiler/tf2xla/xla_context.h"
 #include "tensorflow/compiler/tf2xla/xla_op_kernel.h"
 #include "tensorflow/compiler/tf2xla/xla_op_registry.h"
+#include "tensorflow/compiler/xla/client/lib/constants.h"
 #include "tensorflow/compiler/xla/client/xla_builder.h"
 #include "tensorflow/core/lib/core/errors.h"
 
 namespace tensorflow {
 
 XlaCaseOp::XlaCaseOp(OpKernelConstruction* ctx) : XlaOpKernel(ctx) {
-  OP_REQUIRES_OK(ctx, ctx->GetAttr("branches", &branches_));
+  OP_REQUIRES_OK(ctx, ctx->GetAttr("branches", &unpruned_branches_));
   OP_REQUIRES_OK(ctx, ctx->GetAttr("Tin", &input_types_));
   OP_REQUIRES_OK(ctx, ctx->GetAttr("Tout", &output_types_));
   if (!ctx->GetAttr(kXlaTokenInputNodesAttrName, &token_input_nodes_).ok()) {
@@ -41,12 +42,29 @@ XlaCaseOp::XlaCaseOp(OpKernelConstruction* ctx) : XlaOpKernel(ctx) {
   }
 }
 
+std::pair, xla::XlaOp>
+XlaCaseOp::GetPrunedBranchesAndIndex(XlaOpKernelContext* ctx) {
+  xla::Literal branch_index_literal;
+  bool branch_index_is_constant =
+      ctx->ConstantInput(0, &branch_index_literal).ok();
+
+  if (!branch_index_is_constant) {
+    return {unpruned_branches_, ctx->Input(0)};
+  }
+
+  int32 branch_index = branch_index_literal.Get({});
+  if (branch_index < 0 || branch_index >= unpruned_branches_.size()) {
+    branch_index = unpruned_branches_.size() - 1;
+  }
+
+  std::vector pruned_branch = {unpruned_branches_[branch_index]};
+  return {pruned_branch, xla::ZerosLike(ctx->Input(0))};
+}
+
 // TODO(b/35949885): There is duplication here with the handling of the
 // while_op/if_op. Refactor the common code out/rework.
 void XlaCaseOp::Compile(XlaOpKernelContext* ctx) {
-  xla::XlaBuilder* b = ctx->builder();
-  int num_branches = branches_.size();
-  OP_REQUIRES(ctx, num_branches >= 1,
+  OP_REQUIRES(ctx, !unpruned_branches_.empty(),
               errors::InvalidArgument("Must provide at least one case branch"));
   OP_REQUIRES(ctx, input_type(0) == DT_INT32,
               errors::InvalidArgument(
@@ -55,6 +73,18 @@ void XlaCaseOp::Compile(XlaOpKernelContext* ctx) {
               errors::InvalidArgument(
                   "branch_index argument must be scalar for XLA compilation"));
 
+  xla::XlaBuilder* b = ctx->builder();
+
+  // We opportunistically prune out branches if the branch index is a
+  // compile-time constant.  This is important in the context of the DeviceIndex
+  // ops (and other such ops that may come later) since we may have a Case with
+  // trivially unselected branches that cannot be compiled into HLO.
+  std::vector branches;
+  xla::XlaOp branch_index;
+  std::tie(branches, branch_index) = GetPrunedBranchesAndIndex(ctx);
+
+  int num_branches = branches.size();
+
   VLOG(1) << "Building Case: " << input_types_.size() << " inputs";
 
   std::vector arguments(input_types_.size());
@@ -94,7 +124,7 @@ void XlaCaseOp::Compile(XlaOpKernelContext* ctx) {
     std::vector case_bodies(num_branches);
     for (int branch_idx = 0; branch_idx < num_branches; branch_idx++) {
       OP_REQUIRES_OK(ctx, FindMustBeConstNodes(
-                              ctx, branches_[branch_idx],
+                              ctx, branches[branch_idx],
                               &case_branch_must_be_const_nodes[branch_idx],
                               &case_bodies[branch_idx]));
     }
@@ -133,7 +163,7 @@ void XlaCaseOp::Compile(XlaOpKernelContext* ctx) {
   std::vector branch_results_p(num_branches);
   for (int j = 0; j < num_branches; ++j) {
     OP_REQUIRES_OK(ctx,
-                   compiler->CompileFunction(options, branches_[j], arguments,
+                   compiler->CompileFunction(options, branches[j], arguments,
                                              &branch_results[j]));
     branch_results_p[j] = &branch_results[j];
   }
@@ -171,7 +201,7 @@ void XlaCaseOp::Compile(XlaOpKernelContext* ctx) {
     for (int j = 0; j < num_branches; ++j) {
       branch_results[j] = {};
       OP_REQUIRES_OK(ctx,
-                     compiler->CompileFunction(options, branches_[j], arguments,
+                     compiler->CompileFunction(options, branches[j], arguments,
                                                &branch_results[j]));
     }
   }
@@ -277,7 +307,7 @@ void XlaCaseOp::Compile(XlaOpKernelContext* ctx) {
   auto input_tuple = xla::Tuple(b, inputs);
 
   xla::XlaOp outputs =
-      xla::Conditional(ctx->Input(0), absl::MakeSpan(result_computations),
+      xla::Conditional(branch_index, absl::MakeSpan(result_computations),
                        std::vector(num_branches, input_tuple));
   // Sets non-variable outputs.
   for (int i = 0; i < output_types_.size(); ++i) {
diff --git a/tensorflow/compiler/tf2xla/kernels/case_op.h b/tensorflow/compiler/tf2xla/kernels/case_op.h
index 4a61707864e..4d22a3db830 100644
--- a/tensorflow/compiler/tf2xla/kernels/case_op.h
+++ b/tensorflow/compiler/tf2xla/kernels/case_op.h
@@ -50,7 +50,16 @@ class XlaCaseOp : public XlaOpKernel {
  private:
   TF_DISALLOW_COPY_AND_ASSIGN(XlaCaseOp);
 
-  std::vector branches_;
+  // If the branch_index input is a constant: prunes out all but the branch
+  // corrresponding to that constant branch index, and returns that branch and
+  // the literal 0 (as the first and second component of the pair).
+  //
+  // If the branch_index input is not a constant: returns unpruned_branches_ and
+  // the branch_index input.
+  std::pair, xla::XlaOp> GetPrunedBranchesAndIndex(
+      XlaOpKernelContext* ctx);
+
+  std::vector unpruned_branches_;
   DataTypeVector input_types_;
   DataTypeVector output_types_;
   bool has_token_input_output_;

From f60f6f0c1f68d729b2e501d5a0a668466acb7cda Mon Sep 17 00:00:00 2001
From: Michael Banfield 
Date: Fri, 19 Jun 2020 14:11:21 -0700
Subject: [PATCH 335/867] Support uploading only new data when a file in GCS
 Filesystem is Flush()'d multiple times.

This uses the GCS compose API to avoid reuploading the entire file.

Also add some Vmodules for profiling GCS write paths.

PiperOrigin-RevId: 317377298
Change-Id: I3f36fd684c44070331ba1d9e6689efd0f74bfc0e
---
 .../core/platform/cloud/gcs_file_system.cc    |  150 +-
 .../core/platform/cloud/gcs_file_system.h     |    6 +-
 .../platform/cloud/gcs_file_system_test.cc    | 1620 ++++++++++-------
 3 files changed, 1065 insertions(+), 711 deletions(-)

diff --git a/tensorflow/core/platform/cloud/gcs_file_system.cc b/tensorflow/core/platform/cloud/gcs_file_system.cc
index 5d395f3d821..1bd4d86eef6 100644
--- a/tensorflow/core/platform/cloud/gcs_file_system.cc
+++ b/tensorflow/core/platform/cloud/gcs_file_system.cc
@@ -14,14 +14,18 @@ limitations under the License.
 ==============================================================================*/
 
 #include "tensorflow/core/platform/cloud/gcs_file_system.h"
+
 #include 
 #include 
+
 #include 
 #include 
 #include 
 #include 
 #include 
 #include 
+
+#include "tensorflow/core/platform/strcat.h"
 #ifdef _WIN32
 #include   // for _mktemp
 #endif
@@ -128,6 +132,15 @@ constexpr char kAllowedBucketLocations[] = "GCS_ALLOWED_BUCKET_LOCATIONS";
 // is running in and restricts to buckets in that region.
 constexpr char kDetectZoneSentinelValue[] = "auto";
 
+// How to upload new data when Flush() is called multiple times.
+// By default the entire file is reuploaded.
+constexpr char kAppendMode[] = "GCS_APPEND_MODE";
+// If GCS_APPEND_MODE=compose then instead the new data is uploaded to a
+// temporary object and composed with the original object. This is disabled by
+// default as the multiple API calls required add a risk of stranding temporary
+// objects.
+constexpr char kComposeAppend[] = "compose";
+
 Status GetTmpFilename(string* filename) {
   *filename = io::GetTempFilename("");
   return Status::OK();
@@ -379,15 +392,18 @@ class GcsWritableFile : public WritableFile {
                   GcsFileSystem* filesystem,
                   GcsFileSystem::TimeoutConfig* timeouts,
                   std::function file_cache_erase,
-                  RetryConfig retry_config)
+                  RetryConfig retry_config, bool compose_append)
       : bucket_(bucket),
         object_(object),
         filesystem_(filesystem),
         timeouts_(timeouts),
         file_cache_erase_(std::move(file_cache_erase)),
         sync_needed_(true),
-        retry_config_(retry_config) {
+        retry_config_(retry_config),
+        compose_append_(compose_append),
+        start_offset_(0) {
     // TODO: to make it safer, outfile_ should be constructed from an FD
+    VLOG(3) << "GcsWritableFile: " << GetGcsPath();
     if (GetTmpFilename(&tmp_content_filename_).ok()) {
       outfile_.open(tmp_content_filename_,
                     std::ofstream::binary | std::ofstream::app);
@@ -403,14 +419,18 @@ class GcsWritableFile : public WritableFile {
                   GcsFileSystem* filesystem, const string& tmp_content_filename,
                   GcsFileSystem::TimeoutConfig* timeouts,
                   std::function file_cache_erase,
-                  RetryConfig retry_config)
+                  RetryConfig retry_config, bool compose_append)
       : bucket_(bucket),
         object_(object),
         filesystem_(filesystem),
         timeouts_(timeouts),
         file_cache_erase_(std::move(file_cache_erase)),
         sync_needed_(true),
-        retry_config_(retry_config) {
+        retry_config_(retry_config),
+        compose_append_(compose_append),
+        start_offset_(0) {
+    VLOG(3) << "GcsWritableFile: " << GetGcsPath() << "with existing file "
+            << tmp_content_filename;
     tmp_content_filename_ = tmp_content_filename;
     outfile_.open(tmp_content_filename_,
                   std::ofstream::binary | std::ofstream::app);
@@ -423,6 +443,7 @@ class GcsWritableFile : public WritableFile {
 
   Status Append(StringPiece data) override {
     TF_RETURN_IF_ERROR(CheckWritable());
+    VLOG(3) << "Append: " << GetGcsPath() << " size " << data.length();
     sync_needed_ = true;
     outfile_ << data;
     if (!outfile_.good()) {
@@ -433,6 +454,7 @@ class GcsWritableFile : public WritableFile {
   }
 
   Status Close() override {
+    VLOG(3) << "Close:" << GetGcsPath();
     if (outfile_.is_open()) {
       Status sync_status = Sync();
       if (sync_status.ok()) {
@@ -443,18 +465,23 @@ class GcsWritableFile : public WritableFile {
     return Status::OK();
   }
 
-  Status Flush() override { return Sync(); }
+  Status Flush() override {
+    VLOG(3) << "Flush:" << GetGcsPath();
+    return Sync();
+  }
 
   Status Name(StringPiece* result) const override {
     return errors::Unimplemented("GCSWritableFile does not support Name()");
   }
 
   Status Sync() override {
+    VLOG(3) << "Sync started:" << GetGcsPath();
     TF_RETURN_IF_ERROR(CheckWritable());
     if (!sync_needed_) {
       return Status::OK();
     }
     Status status = SyncImpl();
+    VLOG(3) << "Sync finished " << GetGcsPath();
     if (status.ok()) {
       sync_needed_ = false;
     }
@@ -483,11 +510,26 @@ class GcsWritableFile : public WritableFile {
           "Could not write to the internal temporary file.");
     }
     string session_uri;
-    TF_RETURN_IF_ERROR(CreateNewUploadSession(&session_uri));
+    uint64 start_offset = 0;
+    string object_to_upload = object_;
+    bool should_compose = false;
+    if (compose_append_) {
+      start_offset = start_offset_;
+      // Only compose if the object has already been uploaded to GCS
+      should_compose = start_offset > 0;
+      if (should_compose) {
+        object_to_upload =
+            strings::StrCat(io::Dirname(object_), "/.tmpcompose/",
+                            io::Basename(object_), ".", start_offset_);
+      }
+    }
+    TF_RETURN_IF_ERROR(
+        CreateNewUploadSession(&session_uri, start_offset, object_to_upload));
     uint64 already_uploaded = 0;
     bool first_attempt = true;
     const Status upload_status = RetryingUtils::CallWithRetries(
-        [&first_attempt, &already_uploaded, &session_uri, this]() {
+        [&first_attempt, &already_uploaded, &session_uri, &start_offset,
+         this]() {
           if (!first_attempt) {
             bool completed;
             TF_RETURN_IF_ERROR(RequestUploadSessionStatus(
@@ -502,7 +544,7 @@ class GcsWritableFile : public WritableFile {
             }
           }
           first_attempt = false;
-          return UploadToSession(session_uri, already_uploaded);
+          return UploadToSession(session_uri, start_offset, already_uploaded);
         },
         retry_config_);
     if (upload_status.code() == errors::Code::NOT_FOUND) {
@@ -512,6 +554,12 @@ class GcsWritableFile : public WritableFile {
           strings::StrCat("Upload to gs://", bucket_, "/", object_,
                           " failed, caused by: ", upload_status.ToString()));
     }
+    if (upload_status.ok()) {
+      if (should_compose) {
+        TF_RETURN_IF_ERROR(AppendObject(object_to_upload));
+      }
+      TF_RETURN_IF_ERROR(GetCurrentFileSize(&start_offset_));
+    }
     return upload_status;
   }
 
@@ -534,7 +582,8 @@ class GcsWritableFile : public WritableFile {
   }
 
   /// Initiates a new resumable upload session.
-  Status CreateNewUploadSession(string* session_uri) {
+  Status CreateNewUploadSession(string* session_uri, uint64 start_offset,
+                                string object_to_upload) {
     uint64 file_size;
     TF_RETURN_IF_ERROR(GetCurrentFileSize(&file_size));
 
@@ -542,10 +591,11 @@ class GcsWritableFile : public WritableFile {
     std::unique_ptr request;
     TF_RETURN_IF_ERROR(filesystem_->CreateHttpRequest(&request));
 
-    request->SetUri(strings::StrCat(
-        kGcsUploadUriBase, "b/", bucket_,
-        "/o?uploadType=resumable&name=", request->EscapeString(object_)));
-    request->AddHeader("X-Upload-Content-Length", std::to_string(file_size));
+    request->SetUri(strings::StrCat(kGcsUploadUriBase, "b/", bucket_,
+                                    "/o?uploadType=resumable&name=",
+                                    request->EscapeString(object_to_upload)));
+    request->AddHeader("X-Upload-Content-Length",
+                       std::to_string(file_size - start_offset));
     request->SetPostEmptyBody();
     request->SetResultBuffer(&output_buffer);
     request->SetTimeouts(timeouts_->connect, timeouts_->idle,
@@ -561,6 +611,37 @@ class GcsWritableFile : public WritableFile {
     return Status::OK();
   }
 
+  /// Appends the data of append_object to the original object and deletes
+  /// append_object.
+  Status AppendObject(string append_object) {
+    VLOG(3) << "AppendObject: " << GetGcsPathWithObject(append_object) << " to "
+            << GetGcsPath();
+    std::unique_ptr request;
+    TF_RETURN_IF_ERROR(filesystem_->CreateHttpRequest(&request));
+
+    request->SetUri(strings::StrCat(kGcsUriBase, "b/", bucket_, "/o/",
+                                    request->EscapeString(object_),
+                                    "/compose"));
+
+    const string request_body =
+        strings::StrCat("{'sourceObjects': [{'name': '", object_,
+                        "'},{'name': '", append_object, "'}]}");
+    request->SetTimeouts(timeouts_->connect, timeouts_->idle,
+                         timeouts_->metadata);
+    request->AddHeader("content-type", "application/json");
+    request->SetPostFromBuffer(request_body.c_str(), request_body.size());
+    return RetryingUtils::CallWithRetries(
+        [&request, &append_object, this]() {
+          TF_RETURN_WITH_CONTEXT_IF_ERROR(request->Send(),
+                                          " when composing to ", GetGcsPath());
+          TF_RETURN_WITH_CONTEXT_IF_ERROR(
+              filesystem_->DeleteFile(GetGcsPathWithObject(append_object)),
+              " when cleaning up.");
+          return Status::OK();
+        },
+        retry_config_);
+  }
+
   /// \brief Requests status of a previously initiated upload session.
   ///
   /// If the upload has already succeeded, sets 'completed' to true.
@@ -628,7 +709,8 @@ class GcsWritableFile : public WritableFile {
     return Status::OK();
   }
 
-  Status UploadToSession(const string& session_uri, uint64 start_offset) {
+  Status UploadToSession(const string& session_uri, uint64 start_offset,
+                         uint64 already_uploaded) {
     uint64 file_size;
     TF_RETURN_IF_ERROR(GetCurrentFileSize(&file_size));
 
@@ -637,13 +719,14 @@ class GcsWritableFile : public WritableFile {
     request->SetUri(session_uri);
     if (file_size > 0) {
       request->AddHeader("Content-Range",
-                         strings::StrCat("bytes ", start_offset, "-",
-                                         file_size - 1, "/", file_size));
+                         strings::StrCat("bytes ", already_uploaded, "-",
+                                         file_size - start_offset - 1, "/",
+                                         file_size - start_offset));
     }
     request->SetTimeouts(timeouts_->connect, timeouts_->idle, timeouts_->write);
 
-    TF_RETURN_IF_ERROR(
-        request->SetPutFromFile(tmp_content_filename_, start_offset));
+    TF_RETURN_IF_ERROR(request->SetPutFromFile(
+        tmp_content_filename_, start_offset + already_uploaded));
     TF_RETURN_WITH_CONTEXT_IF_ERROR(request->Send(), " when uploading ",
                                     GetGcsPath());
     // Erase the file from the file cache on every successful write.
@@ -651,9 +734,10 @@ class GcsWritableFile : public WritableFile {
     return Status::OK();
   }
 
-  string GetGcsPath() const {
-    return strings::StrCat("gs://", bucket_, "/", object_);
+  string GetGcsPathWithObject(string object) const {
+    return strings::StrCat("gs://", bucket_, "/", object);
   }
+  string GetGcsPath() const { return GetGcsPathWithObject(object_); }
 
   string bucket_;
   string object_;
@@ -664,6 +748,8 @@ class GcsWritableFile : public WritableFile {
   std::function file_cache_erase_;
   bool sync_needed_;  // whether there is buffered data that needs to be synced
   RetryConfig retry_config_;
+  bool compose_append_;
+  uint64 start_offset_;
 };
 
 class GcsReadOnlyMemoryRegion : public ReadOnlyMemoryRegion {
@@ -849,6 +935,14 @@ GcsFileSystem::GcsFileSystem(bool make_default_cache) {
 
   GetEnvVar(kAllowedBucketLocations, SplitByCommaToLowercaseSet,
             &allowed_locations_);
+
+  StringPiece append_mode;
+  GetEnvVar(kAppendMode, StringPieceIdentity, &append_mode);
+  if (append_mode == kComposeAppend) {
+    compose_append_ = true;
+  } else {
+    compose_append_ = false;
+  }
 }
 
 GcsFileSystem::GcsFileSystem(
@@ -859,7 +953,8 @@ GcsFileSystem::GcsFileSystem(
     size_t stat_cache_max_entries, uint64 matching_paths_cache_max_age,
     size_t matching_paths_cache_max_entries, RetryConfig retry_config,
     TimeoutConfig timeouts, const std::unordered_set& allowed_locations,
-    std::pair* additional_header)
+    std::pair* additional_header,
+    bool compose_append)
     : auth_provider_(std::move(auth_provider)),
       http_request_factory_(std::move(http_request_factory)),
       zone_provider_(std::move(zone_provider)),
@@ -872,6 +967,7 @@ GcsFileSystem::GcsFileSystem(
       bucket_location_cache_(new BucketLocationCache(
           kCacheNeverExpire, kBucketLocationCacheMaxEntries)),
       allowed_locations_(allowed_locations),
+      compose_append_(compose_append),
       timeouts_(timeouts),
       retry_config_(retry_config),
       additional_header_(additional_header) {}
@@ -1056,9 +1152,10 @@ Status GcsFileSystem::NewWritableFile(const string& fname,
                                       std::unique_ptr* result) {
   string bucket, object;
   TF_RETURN_IF_ERROR(ParseGcsPath(fname, false, &bucket, &object));
-  result->reset(new GcsWritableFile(bucket, object, this, &timeouts_,
-                                    [this, fname]() { ClearFileCaches(fname); },
-                                    retry_config_));
+  result->reset(new GcsWritableFile(
+      bucket, object, this, &timeouts_,
+      [this, fname]() { ClearFileCaches(fname); }, retry_config_,
+      compose_append_));
   return Status::OK();
 }
 
@@ -1098,7 +1195,8 @@ Status GcsFileSystem::NewAppendableFile(const string& fname,
   TF_RETURN_IF_ERROR(ParseGcsPath(fname, false, &bucket, &object));
   result->reset(new GcsWritableFile(
       bucket, object, this, old_content_filename, &timeouts_,
-      [this, fname]() { ClearFileCaches(fname); }, retry_config_));
+      [this, fname]() { ClearFileCaches(fname); }, retry_config_,
+      compose_append_));
   return Status::OK();
 }
 
@@ -1629,6 +1727,7 @@ Status GcsFileSystem::RenameFile(const string& src, const string& target) {
 
 // Uses a GCS API command to copy the object and then deletes the old one.
 Status GcsFileSystem::RenameObject(const string& src, const string& target) {
+  VLOG(3) << "RenameObject: started gs://" << src << " to " << target;
   string src_bucket, src_object, target_bucket, target_object;
   TF_RETURN_IF_ERROR(ParseGcsPath(src, false, &src_bucket, &src_object));
   TF_RETURN_IF_ERROR(
@@ -1664,6 +1763,7 @@ Status GcsFileSystem::RenameObject(const string& src, const string& target) {
         "locations or storage classes is not supported.");
   }
 
+  VLOG(3) << "RenameObject: finished from: gs://" << src << " to " << target;
   // In case the delete API call failed, but the deletion actually happened
   // on the server side, we can't just retry the whole RenameFile operation
   // because the source object is already gone.
diff --git a/tensorflow/core/platform/cloud/gcs_file_system.h b/tensorflow/core/platform/cloud/gcs_file_system.h
index d1d8aed54d4..f066cc31eb4 100644
--- a/tensorflow/core/platform/cloud/gcs_file_system.h
+++ b/tensorflow/core/platform/cloud/gcs_file_system.h
@@ -122,7 +122,8 @@ class GcsFileSystem : public FileSystem {
                 size_t matching_paths_cache_max_entries,
                 RetryConfig retry_config, TimeoutConfig timeouts,
                 const std::unordered_set& allowed_locations,
-                std::pair* additional_header);
+                std::pair* additional_header,
+                bool compose_append);
 
   Status NewRandomAccessFile(
       const string& fname, std::unique_ptr* result) override;
@@ -187,6 +188,8 @@ class GcsFileSystem : public FileSystem {
   std::unordered_set allowed_locations() const {
     return allowed_locations_;
   }
+
+  bool compose_append() const { return compose_append_; }
   string additional_header_name() const {
     return additional_header_ ? additional_header_->first : "";
   }
@@ -373,6 +376,7 @@ class GcsFileSystem : public FileSystem {
   using BucketLocationCache = ExpiringLRUCache;
   std::unique_ptr bucket_location_cache_;
   std::unordered_set allowed_locations_;
+  bool compose_append_;
 
   TimeoutConfig timeouts_;
 
diff --git a/tensorflow/core/platform/cloud/gcs_file_system_test.cc b/tensorflow/core/platform/cloud/gcs_file_system_test.cc
index 544ddc32043..6892bd7cc26 100644
--- a/tensorflow/core/platform/cloud/gcs_file_system_test.cc
+++ b/tensorflow/core/platform/cloud/gcs_file_system_test.cc
@@ -22,6 +22,7 @@ limitations under the License.
 #include "tensorflow/core/platform/cloud/http_request_fake.h"
 #include "tensorflow/core/platform/errors.h"
 #include "tensorflow/core/platform/str_util.h"
+#include "tensorflow/core/platform/strcat.h"
 #include "tensorflow/core/platform/test.h"
 
 // Undef DeleteFile macro defined in wndows.h.
@@ -73,16 +74,16 @@ TEST(GcsFileSystemTest, NewRandomAccessFile_NoBlockCache) {
            "Range: 6-11\n"
            "Timeouts: 5 1 20\n",
            "6789")});
-  GcsFileSystem fs(std::unique_ptr(new FakeAuthProvider),
-                   std::unique_ptr(
-                       new FakeHttpRequestFactory(&requests)),
-                   std::unique_ptr(new FakeZoneProvider),
-                   0 /* block size */, 0 /* max bytes */, 0 /* max staleness */,
-                   0 /* stat cache max age */, 0 /* stat cache max entries */,
-                   0 /* matching paths cache max age */,
-                   0 /* matching paths cache max entries */, kTestRetryConfig,
-                   kTestTimeoutConfig, *kAllowedLocationsDefault,
-                   nullptr /* gcs additional header */);
+  GcsFileSystem fs(
+      std::unique_ptr(new FakeAuthProvider),
+      std::unique_ptr(
+          new FakeHttpRequestFactory(&requests)),
+      std::unique_ptr(new FakeZoneProvider), 0 /* block size */,
+      0 /* max bytes */, 0 /* max staleness */, 0 /* stat cache max age */,
+      0 /* stat cache max entries */, 0 /* matching paths cache max age */,
+      0 /* matching paths cache max entries */, kTestRetryConfig,
+      kTestTimeoutConfig, *kAllowedLocationsDefault,
+      nullptr /* gcs additional header */, false /* compose append */);
 
   std::unique_ptr file;
   TF_EXPECT_OK(fs.NewRandomAccessFile("gs://bucket/random_access.txt", &file));
@@ -129,7 +130,7 @@ TEST(GcsFileSystemTest, NewRandomAccessFile_Buffered) {
       0 /* stat cache max entries */, 0 /* matching paths cache max age */,
       0 /* matching paths cache max entries */, kTestRetryConfig,
       kTestTimeoutConfig, *kAllowedLocationsDefault,
-      nullptr /* gcs additional header */);
+      nullptr /* gcs additional header */, false /* compose append */);
 
   std::unique_ptr file;
   TF_EXPECT_OK(fs.NewRandomAccessFile("gs://bucket/random_access.txt", &file));
@@ -177,7 +178,7 @@ TEST(GcsFileSystemTest, NewRandomAccessFile_Buffered_Errors) {
       0 /* stat cache max entries */, 0 /* matching paths cache max age */,
       0 /* matching paths cache max entries */, kTestRetryConfig,
       kTestTimeoutConfig, *kAllowedLocationsDefault,
-      nullptr /* gcs additional header */);
+      nullptr /* gcs additional header */, false /* compose append */);
 
   std::unique_ptr file;
   TF_EXPECT_OK(fs.NewRandomAccessFile("gs://bucket/random_access.txt", &file));
@@ -224,7 +225,7 @@ TEST(GcsFileSystemTest, NewRandomAccessFile_Buffered_ReadAtEOF) {
       0 /* stat cache max entries */, 0 /* matching paths cache max age */,
       0 /* matching paths cache max entries */, kTestRetryConfig,
       kTestTimeoutConfig, *kAllowedLocationsDefault,
-      nullptr /* gcs additional header */);
+      nullptr /* gcs additional header */, false /* compose append */);
 
   std::unique_ptr file;
   TF_EXPECT_OK(fs.NewRandomAccessFile("gs://bucket/random_access.txt", &file));
@@ -265,7 +266,7 @@ TEST(GcsFileSystemTest, NewRandomAccessFile_Buffered_CachedOutOfRange) {
       0 /* stat cache max entries */, 0 /* matching paths cache max age */,
       0 /* matching paths cache max entries */, kTestRetryConfig,
       kTestTimeoutConfig, *kAllowedLocationsDefault,
-      nullptr /* gcs additional header */);
+      nullptr /* gcs additional header */, false /* compose append */);
 
   std::unique_ptr file;
   TF_EXPECT_OK(fs.NewRandomAccessFile("gs://bucket/random_access.txt", &file));
@@ -316,7 +317,7 @@ TEST(GcsFileSystemTest, NewRandomAccessFile_Buffered_CachedNotSequential) {
       0 /* stat cache max entries */, 0 /* matching paths cache max age */,
       0 /* matching paths cache max entries */, kTestRetryConfig,
       kTestTimeoutConfig, *kAllowedLocationsDefault,
-      nullptr /* gcs additional header */);
+      nullptr /* gcs additional header */, false /* compose append */);
 
   std::unique_ptr file;
   TF_EXPECT_OK(fs.NewRandomAccessFile("gs://bucket/random_access.txt", &file));
@@ -357,7 +358,7 @@ TEST(GcsFileSystemTest, NewRandomAccessFile_Buffered_Growing) {
       0 /* stat cache max entries */, 0 /* matching paths cache max age */,
       0 /* matching paths cache max entries */, kTestRetryConfig,
       kTestTimeoutConfig, *kAllowedLocationsDefault,
-      nullptr /* gcs additional header */);
+      nullptr /* gcs additional header */, false /* compose append */);
 
   std::unique_ptr file;
   TF_EXPECT_OK(fs.NewRandomAccessFile("gs://bucket/random_access.txt", &file));
@@ -404,7 +405,7 @@ TEST(GcsFileSystemTest, NewRandomAccessFile_Buffered_ReadBackwards) {
       0 /* stat cache max entries */, 0 /* matching paths cache max age */,
       0 /* matching paths cache max entries */, kTestRetryConfig,
       kTestTimeoutConfig, *kAllowedLocationsDefault,
-      nullptr /* gcs additional header */);
+      nullptr /* gcs additional header */, false /* compose append */);
 
   std::unique_ptr file;
   TF_EXPECT_OK(fs.NewRandomAccessFile("gs://bucket/random_access.txt", &file));
@@ -437,16 +438,16 @@ TEST(GcsFileSystemTest,
             "location":"US-EAST1"
           })")});
 
-  GcsFileSystem fs(std::unique_ptr(new FakeAuthProvider),
-                   std::unique_ptr(
-                       new FakeHttpRequestFactory(&requests)),
-                   std::unique_ptr(new FakeZoneProvider),
-                   0 /* block size */, 0 /* max bytes */, 0 /* max staleness */,
-                   0 /* stat cache max age */, 0 /* stat cache max entries */,
-                   0 /* matching paths cache max age */,
-                   0 /* matching paths cache max entries */, kTestRetryConfig,
-                   kTestTimeoutConfig, *kAllowedLocationsAuto,
-                   nullptr /* gcs additional header */);
+  GcsFileSystem fs(
+      std::unique_ptr(new FakeAuthProvider),
+      std::unique_ptr(
+          new FakeHttpRequestFactory(&requests)),
+      std::unique_ptr(new FakeZoneProvider), 0 /* block size */,
+      0 /* max bytes */, 0 /* max staleness */, 0 /* stat cache max age */,
+      0 /* stat cache max entries */, 0 /* matching paths cache max age */,
+      0 /* matching paths cache max entries */, kTestRetryConfig,
+      kTestTimeoutConfig, *kAllowedLocationsAuto,
+      nullptr /* gcs additional header */, false /* compose append */);
 
   std::unique_ptr file;
   TF_EXPECT_OK(fs.NewRandomAccessFile("gs://bucket/random_access.txt", &file));
@@ -479,16 +480,16 @@ TEST(GcsFileSystemTest, NewRandomAccessFile_WithLocationConstraintCaching) {
             "location":"US-EAST1"
           })")});
 
-  GcsFileSystem fs(std::unique_ptr(new FakeAuthProvider),
-                   std::unique_ptr(
-                       new FakeHttpRequestFactory(&requests)),
-                   std::unique_ptr(new FakeZoneProvider),
-                   0 /* block size */, 0 /* max bytes */, 0 /* max staleness */,
-                   0 /* stat cache max age */, 0 /* stat cache max entries */,
-                   0 /* matching paths cache max age */,
-                   0 /* matching paths cache max entries */, kTestRetryConfig,
-                   kTestTimeoutConfig, *kAllowedLocationsAuto,
-                   nullptr /* gcs additional header */);
+  GcsFileSystem fs(
+      std::unique_ptr(new FakeAuthProvider),
+      std::unique_ptr(
+          new FakeHttpRequestFactory(&requests)),
+      std::unique_ptr(new FakeZoneProvider), 0 /* block size */,
+      0 /* max bytes */, 0 /* max staleness */, 0 /* stat cache max age */,
+      0 /* stat cache max entries */, 0 /* matching paths cache max age */,
+      0 /* matching paths cache max entries */, kTestRetryConfig,
+      kTestTimeoutConfig, *kAllowedLocationsAuto,
+      nullptr /* gcs additional header */, false /* compose append */);
 
   std::unique_ptr file;
 
@@ -520,16 +521,16 @@ TEST(GcsFileSystemTest,
             "location":"BARFOO"
           })")});
 
-  GcsFileSystem fs(std::unique_ptr(new FakeAuthProvider),
-                   std::unique_ptr(
-                       new FakeHttpRequestFactory(&requests)),
-                   std::unique_ptr(new FakeZoneProvider),
-                   0 /* block size */, 0 /* max bytes */, 0 /* max staleness */,
-                   0 /* stat cache max age */, 0 /* stat cache max entries */,
-                   0 /* matching paths cache max age */,
-                   0 /* matching paths cache max entries */, kTestRetryConfig,
-                   kTestTimeoutConfig, *kAllowedLocationsAuto,
-                   nullptr /* gcs additional header */);
+  GcsFileSystem fs(
+      std::unique_ptr(new FakeAuthProvider),
+      std::unique_ptr(
+          new FakeHttpRequestFactory(&requests)),
+      std::unique_ptr(new FakeZoneProvider), 0 /* block size */,
+      0 /* max bytes */, 0 /* max staleness */, 0 /* stat cache max age */,
+      0 /* stat cache max entries */, 0 /* matching paths cache max age */,
+      0 /* matching paths cache max entries */, kTestRetryConfig,
+      kTestTimeoutConfig, *kAllowedLocationsAuto,
+      nullptr /* gcs additional header */, false /* compose append */);
 
   std::unique_ptr file;
   EXPECT_EQ(tensorflow::errors::FailedPrecondition(
@@ -552,16 +553,16 @@ TEST(GcsFileSystemTest, NewRandomAccessFile_NoBlockCache_DifferentN) {
            "Range: 3-12\n"
            "Timeouts: 5 1 20\n",
            "3456789")});
-  GcsFileSystem fs(std::unique_ptr(new FakeAuthProvider),
-                   std::unique_ptr(
-                       new FakeHttpRequestFactory(&requests)),
-                   std::unique_ptr(new FakeZoneProvider),
-                   0 /* block size */, 0 /* max bytes */, 0 /* max staleness */,
-                   0 /* stat cache max age */, 0 /* stat cache max entries */,
-                   0 /* matching paths cache max age */,
-                   0 /* matching paths cache max entries */, kTestRetryConfig,
-                   kTestTimeoutConfig, *kAllowedLocationsDefault,
-                   nullptr /* gcs additional header */);
+  GcsFileSystem fs(
+      std::unique_ptr(new FakeAuthProvider),
+      std::unique_ptr(
+          new FakeHttpRequestFactory(&requests)),
+      std::unique_ptr(new FakeZoneProvider), 0 /* block size */,
+      0 /* max bytes */, 0 /* max staleness */, 0 /* stat cache max age */,
+      0 /* stat cache max entries */, 0 /* matching paths cache max age */,
+      0 /* matching paths cache max entries */, kTestRetryConfig,
+      kTestTimeoutConfig, *kAllowedLocationsDefault,
+      nullptr /* gcs additional header */, false /* compose append */);
 
   std::unique_ptr file;
   TF_EXPECT_OK(fs.NewRandomAccessFile("gs://bucket/random_access.txt", &file));
@@ -621,7 +622,7 @@ TEST(GcsFileSystemTest, NewRandomAccessFile_WithBlockCache) {
       0 /* stat cache max entries */, 0 /* matching paths cache max age */,
       0 /* matching paths cache max entries */, kTestRetryConfig,
       kTestTimeoutConfig, *kAllowedLocationsDefault,
-      nullptr /* gcs additional header */);
+      nullptr /* gcs additional header */, false /* compose append */);
 
   char scratch[100];
   StringPiece result;
@@ -710,7 +711,7 @@ TEST(GcsFileSystemTest, NewRandomAccessFile_WithBlockCache_Flush) {
       0 /* stat cache max entries */, 0 /* matching paths cache max age */,
       0 /* matching paths cache max entries */, kTestRetryConfig,
       kTestTimeoutConfig, *kAllowedLocationsDefault,
-      nullptr /* gcs additional header */);
+      nullptr /* gcs additional header */, false /* compose append */);
 
   char scratch[100];
   StringPiece result;
@@ -750,17 +751,17 @@ TEST(GcsFileSystemTest, NewRandomAccessFile_WithBlockCache_MaxStaleness) {
                            "Range: 8-15\n"
                            "Timeouts: 5 1 20\n",
                            "89abcdef")});
-  GcsFileSystem fs(std::unique_ptr(new FakeAuthProvider),
-                   std::unique_ptr(
-                       new FakeHttpRequestFactory(&requests)),
-                   std::unique_ptr(new FakeZoneProvider),
-                   8 /* block size */, 16 /* max bytes */,
-                   3600 /* max staleness */, 3600 /* stat cache max age */,
-                   0 /* stat cache max entries */,
-                   0 /* matching paths cache max age */,
-                   0 /* matching paths cache max entries */, kTestRetryConfig,
-                   kTestTimeoutConfig, *kAllowedLocationsDefault,
-                   nullptr /* gcs additional header */);
+  GcsFileSystem fs(
+      std::unique_ptr(new FakeAuthProvider),
+      std::unique_ptr(
+          new FakeHttpRequestFactory(&requests)),
+      std::unique_ptr(new FakeZoneProvider), 8 /* block size */,
+      16 /* max bytes */, 3600 /* max staleness */,
+      3600 /* stat cache max age */, 0 /* stat cache max entries */,
+      0 /* matching paths cache max age */,
+      0 /* matching paths cache max entries */, kTestRetryConfig,
+      kTestTimeoutConfig, *kAllowedLocationsDefault,
+      nullptr /* gcs additional header */, false /* compose append */);
   char scratch[100];
   StringPiece result;
   // There should only be two HTTP requests issued to GCS even though we iterate
@@ -830,7 +831,7 @@ TEST(GcsFileSystemTest,
       0 /* stat cache max entries */, 0 /* matching paths cache max age */,
       0 /* matching paths cache max entries */, kTestRetryConfig,
       kTestTimeoutConfig, *kAllowedLocationsDefault,
-      nullptr /* gcs additional header */);
+      nullptr /* gcs additional header */, false /* compose append */);
 
   std::unique_ptr file;
   TF_EXPECT_OK(fs.NewRandomAccessFile("gs://bucket/random_access.txt", &file));
@@ -849,17 +850,17 @@ TEST(GcsFileSystemTest,
 
 TEST(GcsFileSystemTest, NewRandomAccessFile_NoObjectName) {
   std::vector requests;
-  GcsFileSystem fs(std::unique_ptr(new FakeAuthProvider),
-                   std::unique_ptr(
-                       new FakeHttpRequestFactory(&requests)),
-                   std::unique_ptr(new FakeZoneProvider),
-                   0 /* read ahead bytes */, 0 /* max bytes */,
-                   0 /* max staleness */, 0 /* stat cache max age */,
-                   0 /* stat cache max entries */,
-                   0 /* matching paths cache max age */,
-                   0 /* matching paths cache max entries */, kTestRetryConfig,
-                   kTestTimeoutConfig, *kAllowedLocationsDefault,
-                   nullptr /* gcs additional header */);
+  GcsFileSystem fs(
+      std::unique_ptr(new FakeAuthProvider),
+      std::unique_ptr(
+          new FakeHttpRequestFactory(&requests)),
+      std::unique_ptr(new FakeZoneProvider),
+      0 /* read ahead bytes */, 0 /* max bytes */, 0 /* max staleness */,
+      0 /* stat cache max age */, 0 /* stat cache max entries */,
+      0 /* matching paths cache max age */,
+      0 /* matching paths cache max entries */, kTestRetryConfig,
+      kTestTimeoutConfig, *kAllowedLocationsDefault,
+      nullptr /* gcs additional header */, false /* compose append */);
 
   std::unique_ptr file;
   EXPECT_EQ(errors::Code::INVALID_ARGUMENT,
@@ -883,16 +884,16 @@ TEST(GcsFileSystemTest, NewRandomAccessFile_InconsistentRead) {
            "012")});
 
   // Set stat_cache_max_age to 1000s so that StatCache could work.
-  GcsFileSystem fs(std::unique_ptr(new FakeAuthProvider),
-                   std::unique_ptr(
-                       new FakeHttpRequestFactory(&requests)),
-                   std::unique_ptr(new FakeZoneProvider),
-                   0 /* block size */, 0 /* max bytes */, 0 /* max staleness */,
-                   1e3 /* stat cache max age */, 0 /* stat cache max entries */,
-                   0 /* matching paths cache max age */,
-                   0 /* matching paths cache max entries */, kTestRetryConfig,
-                   kTestTimeoutConfig, *kAllowedLocationsDefault,
-                   nullptr /* gcs additional header */);
+  GcsFileSystem fs(
+      std::unique_ptr(new FakeAuthProvider),
+      std::unique_ptr(
+          new FakeHttpRequestFactory(&requests)),
+      std::unique_ptr(new FakeZoneProvider), 0 /* block size */,
+      0 /* max bytes */, 0 /* max staleness */, 1e3 /* stat cache max age */,
+      0 /* stat cache max entries */, 0 /* matching paths cache max age */,
+      0 /* matching paths cache max entries */, kTestRetryConfig,
+      kTestTimeoutConfig, *kAllowedLocationsDefault,
+      nullptr /* gcs additional header */, false /* compose append */);
 
   // Stat the file first so that the file stats are cached.
   FileStatistics stat;
@@ -959,7 +960,7 @@ TEST(GcsFileSystemTest, NewWritableFile) {
       0 /* stat cache max entries */, 0 /* matching paths cache max age */,
       0 /* matching paths cache max entries */, kTestRetryConfig,
       kTestTimeoutConfig, *kAllowedLocationsDefault,
-      nullptr /* gcs additional header */);
+      nullptr /* gcs additional header */, false /* compose append */);
 
   // Read from the file first, to fill the block cache.
   std::unique_ptr rfile;
@@ -1042,16 +1043,16 @@ TEST(GcsFileSystemTest, NewWritableFile_ResumeUploadSucceeds) {
                            "Timeouts: 5 1 30\n"
                            "Put body: t2\n",
                            "")});
-  GcsFileSystem fs(std::unique_ptr(new FakeAuthProvider),
-                   std::unique_ptr(
-                       new FakeHttpRequestFactory(&requests)),
-                   std::unique_ptr(new FakeZoneProvider),
-                   0 /* block size */, 0 /* max bytes */, 0 /* max staleness */,
-                   0 /* stat cache max age */, 0 /* stat cache max entries */,
-                   0 /* matching paths cache max age */,
-                   0 /* matching paths cache max entries */, kTestRetryConfig,
-                   kTestTimeoutConfig, *kAllowedLocationsDefault,
-                   nullptr /* gcs additional header */);
+  GcsFileSystem fs(
+      std::unique_ptr(new FakeAuthProvider),
+      std::unique_ptr(
+          new FakeHttpRequestFactory(&requests)),
+      std::unique_ptr(new FakeZoneProvider), 0 /* block size */,
+      0 /* max bytes */, 0 /* max staleness */, 0 /* stat cache max age */,
+      0 /* stat cache max entries */, 0 /* matching paths cache max age */,
+      0 /* matching paths cache max entries */, kTestRetryConfig,
+      kTestTimeoutConfig, *kAllowedLocationsDefault,
+      nullptr /* gcs additional header */, false /* compose append */);
 
   std::unique_ptr file;
   TF_EXPECT_OK(fs.NewWritableFile("gs://bucket/path/writeable.txt", &file));
@@ -1112,17 +1113,17 @@ TEST(GcsFileSystemTest, NewWritableFile_ResumeUploadSucceedsOnGetStatus) {
            "Range: 0-7\n"
            "Timeouts: 5 1 20\n",
            "01234567")});
-  GcsFileSystem fs(std::unique_ptr(new FakeAuthProvider),
-                   std::unique_ptr(
-                       new FakeHttpRequestFactory(&requests)),
-                   std::unique_ptr(new FakeZoneProvider),
-                   8 /* block size */, 8 /* max bytes */,
-                   3600 /* max staleness */, 3600 /* stat cache max age */,
-                   0 /* stat cache max entries */,
-                   0 /* matching paths cache max age */,
-                   0 /* matching paths cache max entries */, kTestRetryConfig,
-                   kTestTimeoutConfig, *kAllowedLocationsDefault,
-                   nullptr /* gcs additional header */);
+  GcsFileSystem fs(
+      std::unique_ptr(new FakeAuthProvider),
+      std::unique_ptr(
+          new FakeHttpRequestFactory(&requests)),
+      std::unique_ptr(new FakeZoneProvider), 8 /* block size */,
+      8 /* max bytes */, 3600 /* max staleness */,
+      3600 /* stat cache max age */, 0 /* stat cache max entries */,
+      0 /* matching paths cache max age */,
+      0 /* matching paths cache max entries */, kTestRetryConfig,
+      kTestTimeoutConfig, *kAllowedLocationsDefault,
+      nullptr /* gcs additional header */, false /* compose append */);
   // Pull the file's first block into the cache. This will trigger the first
   // HTTP request to GCS.
   std::unique_ptr rfile;
@@ -1208,7 +1209,8 @@ TEST(GcsFileSystemTest, NewWritableFile_ResumeUploadAllAttemptsFail) {
       0 /* stat cache max entries */, 0 /* matching paths cache max age */,
       0 /* matching paths cache max entries */,
       RetryConfig(2 /* .init_delay_time_us */), kTestTimeoutConfig,
-      *kAllowedLocationsDefault, nullptr /* gcs additional header */);
+      *kAllowedLocationsDefault, nullptr /* gcs additional header */,
+      false /* compose append */);
 
   std::unique_ptr file;
   TF_EXPECT_OK(fs.NewWritableFile("gs://bucket/path/writeable.txt", &file));
@@ -1262,16 +1264,16 @@ TEST(GcsFileSystemTest, NewWritableFile_UploadReturns410) {
                            "Timeouts: 5 1 30\n"
                            "Put body: content1,content2\n",
                            "")});
-  GcsFileSystem fs(std::unique_ptr(new FakeAuthProvider),
-                   std::unique_ptr(
-                       new FakeHttpRequestFactory(&requests)),
-                   std::unique_ptr(new FakeZoneProvider),
-                   0 /* block size */, 0 /* max bytes */, 0 /* max staleness */,
-                   0 /* stat cache max age */, 0 /* stat cache max entries */,
-                   0 /* matching paths cache max age */,
-                   0 /* matching paths cache max entries */, kTestRetryConfig,
-                   kTestTimeoutConfig, *kAllowedLocationsDefault,
-                   nullptr /* gcs additional header */);
+  GcsFileSystem fs(
+      std::unique_ptr(new FakeAuthProvider),
+      std::unique_ptr(
+          new FakeHttpRequestFactory(&requests)),
+      std::unique_ptr(new FakeZoneProvider), 0 /* block size */,
+      0 /* max bytes */, 0 /* max staleness */, 0 /* stat cache max age */,
+      0 /* stat cache max entries */, 0 /* matching paths cache max age */,
+      0 /* matching paths cache max entries */, kTestRetryConfig,
+      kTestTimeoutConfig, *kAllowedLocationsDefault,
+      nullptr /* gcs additional header */, false /* compose append */);
 
   {
     std::unique_ptr file;
@@ -1302,16 +1304,16 @@ TEST(GcsFileSystemTest, NewWritableFile_UploadReturns410) {
 
 TEST(GcsFileSystemTest, NewWritableFile_NoObjectName) {
   std::vector requests;
-  GcsFileSystem fs(std::unique_ptr(new FakeAuthProvider),
-                   std::unique_ptr(
-                       new FakeHttpRequestFactory(&requests)),
-                   std::unique_ptr(new FakeZoneProvider),
-                   0 /* block size */, 0 /* max bytes */, 0 /* max staleness */,
-                   0 /* stat cache max age */, 0 /* stat cache max entries */,
-                   0 /* matching paths cache max age */,
-                   0 /* matching paths cache max entries */, kTestRetryConfig,
-                   kTestTimeoutConfig, *kAllowedLocationsDefault,
-                   nullptr /* gcs additional header */);
+  GcsFileSystem fs(
+      std::unique_ptr(new FakeAuthProvider),
+      std::unique_ptr(
+          new FakeHttpRequestFactory(&requests)),
+      std::unique_ptr(new FakeZoneProvider), 0 /* block size */,
+      0 /* max bytes */, 0 /* max staleness */, 0 /* stat cache max age */,
+      0 /* stat cache max entries */, 0 /* matching paths cache max age */,
+      0 /* matching paths cache max entries */, kTestRetryConfig,
+      kTestTimeoutConfig, *kAllowedLocationsDefault,
+      nullptr /* gcs additional header */, false /* compose append */);
 
   std::unique_ptr file;
   EXPECT_EQ(errors::Code::INVALID_ARGUMENT,
@@ -1375,7 +1377,7 @@ TEST(GcsFileSystemTest, NewAppendableFile) {
       0 /* stat cache max entries */, 0 /* matching paths cache max age */,
       0 /* matching paths cache max entries */, kTestRetryConfig,
       kTestTimeoutConfig, *kAllowedLocationsDefault,
-      nullptr /* gcs additional header */);
+      nullptr /* gcs additional header */, false /* compose append */);
 
   // Create an appendable file. This should read the file from GCS, and pull its
   // contents into the block cache.
@@ -1401,16 +1403,16 @@ TEST(GcsFileSystemTest, NewAppendableFile) {
 
 TEST(GcsFileSystemTest, NewAppendableFile_NoObjectName) {
   std::vector requests;
-  GcsFileSystem fs(std::unique_ptr(new FakeAuthProvider),
-                   std::unique_ptr(
-                       new FakeHttpRequestFactory(&requests)),
-                   std::unique_ptr(new FakeZoneProvider),
-                   0 /* block size */, 0 /* max bytes */, 0 /* max staleness */,
-                   0 /* stat cache max age */, 0 /* stat cache max entries */,
-                   0 /* matching paths cache max age */,
-                   0 /* matching paths cache max entries */, kTestRetryConfig,
-                   kTestTimeoutConfig, *kAllowedLocationsDefault,
-                   nullptr /* gcs additional header */);
+  GcsFileSystem fs(
+      std::unique_ptr(new FakeAuthProvider),
+      std::unique_ptr(
+          new FakeHttpRequestFactory(&requests)),
+      std::unique_ptr(new FakeZoneProvider), 0 /* block size */,
+      0 /* max bytes */, 0 /* max staleness */, 0 /* stat cache max age */,
+      0 /* stat cache max entries */, 0 /* matching paths cache max age */,
+      0 /* matching paths cache max entries */, kTestRetryConfig,
+      kTestTimeoutConfig, *kAllowedLocationsDefault,
+      nullptr /* gcs additional header */, false /* compose append */);
 
   std::unique_ptr file;
   EXPECT_EQ(errors::Code::INVALID_ARGUMENT,
@@ -1435,16 +1437,16 @@ TEST(GcsFileSystemTest, NewReadOnlyMemoryRegionFromFile) {
                            "Range: 0-",
                            content.size() - 1, "\n", "Timeouts: 5 1 20\n"),
            content)});
-  GcsFileSystem fs(std::unique_ptr(new FakeAuthProvider),
-                   std::unique_ptr(
-                       new FakeHttpRequestFactory(&requests)),
-                   std::unique_ptr(new FakeZoneProvider),
-                   0 /* block size */, 0 /* max bytes */, 0 /* max staleness */,
-                   0 /* stat cache max age */, 0 /* stat cache max entries */,
-                   0 /* matching paths cache max age */,
-                   0 /* matching paths cache max entries */, kTestRetryConfig,
-                   kTestTimeoutConfig, *kAllowedLocationsDefault,
-                   nullptr /* gcs additional header */);
+  GcsFileSystem fs(
+      std::unique_ptr(new FakeAuthProvider),
+      std::unique_ptr(
+          new FakeHttpRequestFactory(&requests)),
+      std::unique_ptr(new FakeZoneProvider), 0 /* block size */,
+      0 /* max bytes */, 0 /* max staleness */, 0 /* stat cache max age */,
+      0 /* stat cache max entries */, 0 /* matching paths cache max age */,
+      0 /* matching paths cache max entries */, kTestRetryConfig,
+      kTestTimeoutConfig, *kAllowedLocationsDefault,
+      nullptr /* gcs additional header */, false /* compose append */);
 
   std::unique_ptr region;
   TF_EXPECT_OK(fs.NewReadOnlyMemoryRegionFromFile(
@@ -1456,16 +1458,16 @@ TEST(GcsFileSystemTest, NewReadOnlyMemoryRegionFromFile) {
 
 TEST(GcsFileSystemTest, NewReadOnlyMemoryRegionFromFile_NoObjectName) {
   std::vector requests;
-  GcsFileSystem fs(std::unique_ptr(new FakeAuthProvider),
-                   std::unique_ptr(
-                       new FakeHttpRequestFactory(&requests)),
-                   std::unique_ptr(new FakeZoneProvider),
-                   0 /* block size */, 0 /* max bytes */, 0 /* max staleness */,
-                   0 /* stat cache max age */, 0 /* stat cache max entries */,
-                   0 /* matching paths cache max age */,
-                   0 /* matching paths cache max entries */, kTestRetryConfig,
-                   kTestTimeoutConfig, *kAllowedLocationsDefault,
-                   nullptr /* gcs additional header */);
+  GcsFileSystem fs(
+      std::unique_ptr(new FakeAuthProvider),
+      std::unique_ptr(
+          new FakeHttpRequestFactory(&requests)),
+      std::unique_ptr(new FakeZoneProvider), 0 /* block size */,
+      0 /* max bytes */, 0 /* max staleness */, 0 /* stat cache max age */,
+      0 /* stat cache max entries */, 0 /* matching paths cache max age */,
+      0 /* matching paths cache max entries */, kTestRetryConfig,
+      kTestTimeoutConfig, *kAllowedLocationsDefault,
+      nullptr /* gcs additional header */, false /* compose append */);
 
   std::unique_ptr region;
   EXPECT_EQ(errors::Code::INVALID_ARGUMENT,
@@ -1480,16 +1482,16 @@ TEST(GcsFileSystemTest, FileExists_YesAsObject) {
       "Timeouts: 5 1 10\n",
       strings::StrCat("{\"size\": \"1010\",\"generation\": \"1\","
                       "\"updated\": \"2016-04-29T23:15:24.896Z\"}"))});
-  GcsFileSystem fs(std::unique_ptr(new FakeAuthProvider),
-                   std::unique_ptr(
-                       new FakeHttpRequestFactory(&requests)),
-                   std::unique_ptr(new FakeZoneProvider),
-                   0 /* block size */, 0 /* max bytes */, 0 /* max staleness */,
-                   0 /* stat cache max age */, 0 /* stat cache max entries */,
-                   0 /* matching paths cache max age */,
-                   0 /* matching paths cache max entries */, kTestRetryConfig,
-                   kTestTimeoutConfig, *kAllowedLocationsDefault,
-                   nullptr /* gcs additional header */);
+  GcsFileSystem fs(
+      std::unique_ptr(new FakeAuthProvider),
+      std::unique_ptr(
+          new FakeHttpRequestFactory(&requests)),
+      std::unique_ptr(new FakeZoneProvider), 0 /* block size */,
+      0 /* max bytes */, 0 /* max staleness */, 0 /* stat cache max age */,
+      0 /* stat cache max entries */, 0 /* matching paths cache max age */,
+      0 /* matching paths cache max entries */, kTestRetryConfig,
+      kTestTimeoutConfig, *kAllowedLocationsDefault,
+      nullptr /* gcs additional header */, false /* compose append */);
 
   TF_EXPECT_OK(fs.FileExists("gs://bucket/path/file1.txt"));
 }
@@ -1510,16 +1512,16 @@ TEST(GcsFileSystemTest, FileExists_YesAsFolder) {
            "Timeouts: 5 1 10\n",
            "{\"items\": [ "
            "  { \"name\": \"path/subfolder/\" }]}")});
-  GcsFileSystem fs(std::unique_ptr(new FakeAuthProvider),
-                   std::unique_ptr(
-                       new FakeHttpRequestFactory(&requests)),
-                   std::unique_ptr(new FakeZoneProvider),
-                   0 /* block size */, 0 /* max bytes */, 0 /* max staleness */,
-                   0 /* stat cache max age */, 0 /* stat cache max entries */,
-                   0 /* matching paths cache max age */,
-                   0 /* matching paths cache max entries */, kTestRetryConfig,
-                   kTestTimeoutConfig, *kAllowedLocationsDefault,
-                   nullptr /* gcs additional header */);
+  GcsFileSystem fs(
+      std::unique_ptr(new FakeAuthProvider),
+      std::unique_ptr(
+          new FakeHttpRequestFactory(&requests)),
+      std::unique_ptr(new FakeZoneProvider), 0 /* block size */,
+      0 /* max bytes */, 0 /* max staleness */, 0 /* stat cache max age */,
+      0 /* stat cache max entries */, 0 /* matching paths cache max age */,
+      0 /* matching paths cache max entries */, kTestRetryConfig,
+      kTestTimeoutConfig, *kAllowedLocationsDefault,
+      nullptr /* gcs additional header */, false /* compose append */);
 
   TF_EXPECT_OK(fs.FileExists("gs://bucket/path/subfolder"));
 }
@@ -1536,16 +1538,16 @@ TEST(GcsFileSystemTest, FileExists_YesAsBucket) {
            "Auth Token: fake_token\n"
            "Timeouts: 5 1 10\n",
            "{\"size\": \"100\"}")});
-  GcsFileSystem fs(std::unique_ptr(new FakeAuthProvider),
-                   std::unique_ptr(
-                       new FakeHttpRequestFactory(&requests)),
-                   std::unique_ptr(new FakeZoneProvider),
-                   0 /* block size */, 0 /* max bytes */, 0 /* max staleness */,
-                   0 /* stat cache max age */, 0 /* stat cache max entries */,
-                   0 /* matching paths cache max age */,
-                   0 /* matching paths cache max entries */, kTestRetryConfig,
-                   kTestTimeoutConfig, *kAllowedLocationsDefault,
-                   nullptr /* gcs additional header */);
+  GcsFileSystem fs(
+      std::unique_ptr(new FakeAuthProvider),
+      std::unique_ptr(
+          new FakeHttpRequestFactory(&requests)),
+      std::unique_ptr(new FakeZoneProvider), 0 /* block size */,
+      0 /* max bytes */, 0 /* max staleness */, 0 /* stat cache max age */,
+      0 /* stat cache max entries */, 0 /* matching paths cache max age */,
+      0 /* matching paths cache max entries */, kTestRetryConfig,
+      kTestTimeoutConfig, *kAllowedLocationsDefault,
+      nullptr /* gcs additional header */, false /* compose append */);
 
   TF_EXPECT_OK(fs.FileExists("gs://bucket1"));
   TF_EXPECT_OK(fs.FileExists("gs://bucket1/"));
@@ -1566,16 +1568,16 @@ TEST(GcsFileSystemTest, FileExists_NotAsObjectOrFolder) {
            "Auth Token: fake_token\n"
            "Timeouts: 5 1 10\n",
            "{\"items\": []}")});
-  GcsFileSystem fs(std::unique_ptr(new FakeAuthProvider),
-                   std::unique_ptr(
-                       new FakeHttpRequestFactory(&requests)),
-                   std::unique_ptr(new FakeZoneProvider),
-                   0 /* block size */, 0 /* max bytes */, 0 /* max staleness */,
-                   0 /* stat cache max age */, 0 /* stat cache max entries */,
-                   0 /* matching paths cache max age */,
-                   0 /* matching paths cache max entries */, kTestRetryConfig,
-                   kTestTimeoutConfig, *kAllowedLocationsDefault,
-                   nullptr /* gcs additional header */);
+  GcsFileSystem fs(
+      std::unique_ptr(new FakeAuthProvider),
+      std::unique_ptr(
+          new FakeHttpRequestFactory(&requests)),
+      std::unique_ptr(new FakeZoneProvider), 0 /* block size */,
+      0 /* max bytes */, 0 /* max staleness */, 0 /* stat cache max age */,
+      0 /* stat cache max entries */, 0 /* matching paths cache max age */,
+      0 /* matching paths cache max entries */, kTestRetryConfig,
+      kTestTimeoutConfig, *kAllowedLocationsDefault,
+      nullptr /* gcs additional header */, false /* compose append */);
 
   EXPECT_EQ(errors::Code::NOT_FOUND,
             fs.FileExists("gs://bucket/path/file1.txt").code());
@@ -1593,16 +1595,16 @@ TEST(GcsFileSystemTest, FileExists_NotAsBucket) {
            "Auth Token: fake_token\n"
            "Timeouts: 5 1 10\n",
            "", errors::NotFound("404"), 404)});
-  GcsFileSystem fs(std::unique_ptr(new FakeAuthProvider),
-                   std::unique_ptr(
-                       new FakeHttpRequestFactory(&requests)),
-                   std::unique_ptr(new FakeZoneProvider),
-                   0 /* block size */, 0 /* max bytes */, 0 /* max staleness */,
-                   0 /* stat cache max age */, 0 /* stat cache max entries */,
-                   0 /* matching paths cache max age */,
-                   0 /* matching paths cache max entries */, kTestRetryConfig,
-                   kTestTimeoutConfig, *kAllowedLocationsDefault,
-                   nullptr /* gcs additional header */);
+  GcsFileSystem fs(
+      std::unique_ptr(new FakeAuthProvider),
+      std::unique_ptr(
+          new FakeHttpRequestFactory(&requests)),
+      std::unique_ptr(new FakeZoneProvider), 0 /* block size */,
+      0 /* max bytes */, 0 /* max staleness */, 0 /* stat cache max age */,
+      0 /* stat cache max entries */, 0 /* matching paths cache max age */,
+      0 /* matching paths cache max entries */, kTestRetryConfig,
+      kTestTimeoutConfig, *kAllowedLocationsDefault,
+      nullptr /* gcs additional header */, false /* compose append */);
   EXPECT_EQ(errors::Code::INVALID_ARGUMENT,
             fs.FileExists("gs://bucket2/").code());
   EXPECT_EQ(errors::Code::INVALID_ARGUMENT,
@@ -1641,7 +1643,7 @@ TEST(GcsFileSystemTest, FileExists_StatCache) {
       0 /* stat cache max entries */, 0 /* matching paths cache max age */,
       0 /* matching paths cache max entries */, kTestRetryConfig,
       kTestTimeoutConfig, *kAllowedLocationsDefault,
-      nullptr /* gcs additional header */);
+      nullptr /* gcs additional header */, false /* compose append */);
 
   // The stat cache will ensure that repeated lookups don't trigger additional
   // HTTP requests.
@@ -1668,7 +1670,7 @@ TEST(GcsFileSystemTest, FileExists_DirectoryMark) {
       0 /* stat cache max entries */, 0 /* matching paths cache max age */,
       0 /* matching paths cache max entries */, kTestRetryConfig,
       kTestTimeoutConfig, *kAllowedLocationsDefault,
-      nullptr /* gcs additional header */);
+      nullptr /* gcs additional header */, false /* compose append */);
 
   TF_EXPECT_OK(fs.FileExists("gs://bucket/dir/"));
   TF_EXPECT_OK(fs.IsDirectory("gs://bucket/dir/"));
@@ -1682,16 +1684,16 @@ TEST(GcsFileSystemTest, GetChildren_NoItems) {
       "Auth Token: fake_token\n"
       "Timeouts: 5 1 10\n",
       "{\"prefixes\": [\"path/subpath/\"]}")});
-  GcsFileSystem fs(std::unique_ptr(new FakeAuthProvider),
-                   std::unique_ptr(
-                       new FakeHttpRequestFactory(&requests)),
-                   std::unique_ptr(new FakeZoneProvider),
-                   0 /* block size */, 0 /* max bytes */, 0 /* max staleness */,
-                   0 /* stat cache max age */, 0 /* stat cache max entries */,
-                   0 /* matching paths cache max age */,
-                   0 /* matching paths cache max entries */, kTestRetryConfig,
-                   kTestTimeoutConfig, *kAllowedLocationsDefault,
-                   nullptr /* gcs additional header */);
+  GcsFileSystem fs(
+      std::unique_ptr(new FakeAuthProvider),
+      std::unique_ptr(
+          new FakeHttpRequestFactory(&requests)),
+      std::unique_ptr(new FakeZoneProvider), 0 /* block size */,
+      0 /* max bytes */, 0 /* max staleness */, 0 /* stat cache max age */,
+      0 /* stat cache max entries */, 0 /* matching paths cache max age */,
+      0 /* matching paths cache max entries */, kTestRetryConfig,
+      kTestTimeoutConfig, *kAllowedLocationsDefault,
+      nullptr /* gcs additional header */, false /* compose append */);
 
   std::vector children;
   TF_EXPECT_OK(fs.GetChildren("gs://bucket/path/", &children));
@@ -1710,16 +1712,16 @@ TEST(GcsFileSystemTest, GetChildren_ThreeFiles) {
       "  { \"name\": \"path/file1.txt\" },"
       "  { \"name\": \"path/file3.txt\" }],"
       "\"prefixes\": [\"path/subpath/\"]}")});
-  GcsFileSystem fs(std::unique_ptr(new FakeAuthProvider),
-                   std::unique_ptr(
-                       new FakeHttpRequestFactory(&requests)),
-                   std::unique_ptr(new FakeZoneProvider),
-                   0 /* block size */, 0 /* max bytes */, 0 /* max staleness */,
-                   0 /* stat cache max age */, 0 /* stat cache max entries */,
-                   0 /* matching paths cache max age */,
-                   0 /* matching paths cache max entries */, kTestRetryConfig,
-                   kTestTimeoutConfig, *kAllowedLocationsDefault,
-                   nullptr /* gcs additional header */);
+  GcsFileSystem fs(
+      std::unique_ptr(new FakeAuthProvider),
+      std::unique_ptr(
+          new FakeHttpRequestFactory(&requests)),
+      std::unique_ptr(new FakeZoneProvider), 0 /* block size */,
+      0 /* max bytes */, 0 /* max staleness */, 0 /* stat cache max age */,
+      0 /* stat cache max entries */, 0 /* matching paths cache max age */,
+      0 /* matching paths cache max entries */, kTestRetryConfig,
+      kTestTimeoutConfig, *kAllowedLocationsDefault,
+      nullptr /* gcs additional header */, false /* compose append */);
 
   std::vector children;
   TF_EXPECT_OK(fs.GetChildren("gs://bucket/path/", &children));
@@ -1739,16 +1741,16 @@ TEST(GcsFileSystemTest, GetChildren_SelfDirectoryMarker) {
       "  { \"name\": \"path/\" },"
       "  { \"name\": \"path/file3.txt\" }],"
       "\"prefixes\": [\"path/subpath/\"]}")});
-  GcsFileSystem fs(std::unique_ptr(new FakeAuthProvider),
-                   std::unique_ptr(
-                       new FakeHttpRequestFactory(&requests)),
-                   std::unique_ptr(new FakeZoneProvider),
-                   0 /* block size */, 0 /* max bytes */, 0 /* max staleness */,
-                   0 /* stat cache max age */, 0 /* stat cache max entries */,
-                   0 /* matching paths cache max age */,
-                   0 /* matching paths cache max entries */, kTestRetryConfig,
-                   kTestTimeoutConfig, *kAllowedLocationsDefault,
-                   nullptr /* gcs additional header */);
+  GcsFileSystem fs(
+      std::unique_ptr(new FakeAuthProvider),
+      std::unique_ptr(
+          new FakeHttpRequestFactory(&requests)),
+      std::unique_ptr(new FakeZoneProvider), 0 /* block size */,
+      0 /* max bytes */, 0 /* max staleness */, 0 /* stat cache max age */,
+      0 /* stat cache max entries */, 0 /* matching paths cache max age */,
+      0 /* matching paths cache max entries */, kTestRetryConfig,
+      kTestTimeoutConfig, *kAllowedLocationsDefault,
+      nullptr /* gcs additional header */, false /* compose append */);
 
   std::vector children;
   TF_EXPECT_OK(fs.GetChildren("gs://bucket/path/", &children));
@@ -1767,16 +1769,16 @@ TEST(GcsFileSystemTest, GetChildren_ThreeFiles_NoSlash) {
       "  { \"name\": \"path/file1.txt\" },"
       "  { \"name\": \"path/file3.txt\" }],"
       "\"prefixes\": [\"path/subpath/\"]}")});
-  GcsFileSystem fs(std::unique_ptr(new FakeAuthProvider),
-                   std::unique_ptr(
-                       new FakeHttpRequestFactory(&requests)),
-                   std::unique_ptr(new FakeZoneProvider),
-                   0 /* block size */, 0 /* max bytes */, 0 /* max staleness */,
-                   0 /* stat cache max age */, 0 /* stat cache max entries */,
-                   0 /* matching paths cache max age */,
-                   0 /* matching paths cache max entries */, kTestRetryConfig,
-                   kTestTimeoutConfig, *kAllowedLocationsDefault,
-                   nullptr /* gcs additional header */);
+  GcsFileSystem fs(
+      std::unique_ptr(new FakeAuthProvider),
+      std::unique_ptr(
+          new FakeHttpRequestFactory(&requests)),
+      std::unique_ptr(new FakeZoneProvider), 0 /* block size */,
+      0 /* max bytes */, 0 /* max staleness */, 0 /* stat cache max age */,
+      0 /* stat cache max entries */, 0 /* matching paths cache max age */,
+      0 /* matching paths cache max entries */, kTestRetryConfig,
+      kTestTimeoutConfig, *kAllowedLocationsDefault,
+      nullptr /* gcs additional header */, false /* compose append */);
 
   std::vector children;
   TF_EXPECT_OK(fs.GetChildren("gs://bucket/path", &children));
@@ -1792,16 +1794,16 @@ TEST(GcsFileSystemTest, GetChildren_Root) {
       "Auth Token: fake_token\n"
       "Timeouts: 5 1 10\n",
       "{}")});
-  GcsFileSystem fs(std::unique_ptr(new FakeAuthProvider),
-                   std::unique_ptr(
-                       new FakeHttpRequestFactory(&requests)),
-                   std::unique_ptr(new FakeZoneProvider),
-                   0 /* block size */, 0 /* max bytes */, 0 /* max staleness */,
-                   0 /* stat cache max age */, 0 /* stat cache max entries */,
-                   0 /* matching paths cache max age */,
-                   0 /* matching paths cache max entries */, kTestRetryConfig,
-                   kTestTimeoutConfig, *kAllowedLocationsDefault,
-                   nullptr /* gcs additional header */);
+  GcsFileSystem fs(
+      std::unique_ptr(new FakeAuthProvider),
+      std::unique_ptr(
+          new FakeHttpRequestFactory(&requests)),
+      std::unique_ptr(new FakeZoneProvider), 0 /* block size */,
+      0 /* max bytes */, 0 /* max staleness */, 0 /* stat cache max age */,
+      0 /* stat cache max entries */, 0 /* matching paths cache max age */,
+      0 /* matching paths cache max entries */, kTestRetryConfig,
+      kTestTimeoutConfig, *kAllowedLocationsDefault,
+      nullptr /* gcs additional header */, false /* compose append */);
 
   std::vector children;
   TF_EXPECT_OK(fs.GetChildren("gs://bucket-a-b-c", &children));
@@ -1817,16 +1819,16 @@ TEST(GcsFileSystemTest, GetChildren_Empty) {
       "Auth Token: fake_token\n"
       "Timeouts: 5 1 10\n",
       "{}")});
-  GcsFileSystem fs(std::unique_ptr(new FakeAuthProvider),
-                   std::unique_ptr(
-                       new FakeHttpRequestFactory(&requests)),
-                   std::unique_ptr(new FakeZoneProvider),
-                   0 /* block size */, 0 /* max bytes */, 0 /* max staleness */,
-                   0 /* stat cache max age */, 0 /* stat cache max entries */,
-                   0 /* matching paths cache max age */,
-                   0 /* matching paths cache max entries */, kTestRetryConfig,
-                   kTestTimeoutConfig, *kAllowedLocationsDefault,
-                   nullptr /* gcs additional header */);
+  GcsFileSystem fs(
+      std::unique_ptr(new FakeAuthProvider),
+      std::unique_ptr(
+          new FakeHttpRequestFactory(&requests)),
+      std::unique_ptr(new FakeZoneProvider), 0 /* block size */,
+      0 /* max bytes */, 0 /* max staleness */, 0 /* stat cache max age */,
+      0 /* stat cache max entries */, 0 /* matching paths cache max age */,
+      0 /* matching paths cache max entries */, kTestRetryConfig,
+      kTestTimeoutConfig, *kAllowedLocationsDefault,
+      nullptr /* gcs additional header */, false /* compose append */);
 
   std::vector children;
   TF_EXPECT_OK(fs.GetChildren("gs://bucket/path/", &children));
@@ -1858,16 +1860,16 @@ TEST(GcsFileSystemTest, GetChildren_Pagination) {
            "  { \"name\": \"path/file4.txt\" },"
            "  { \"name\": \"path/file5.txt\" }]}")});
 
-  GcsFileSystem fs(std::unique_ptr(new FakeAuthProvider),
-                   std::unique_ptr(
-                       new FakeHttpRequestFactory(&requests)),
-                   std::unique_ptr(new FakeZoneProvider),
-                   0 /* block size */, 0 /* max bytes */, 0 /* max staleness */,
-                   0 /* stat cache max age */, 0 /* stat cache max entries */,
-                   0 /* matching paths cache max age */,
-                   0 /* matching paths cache max entries */, kTestRetryConfig,
-                   kTestTimeoutConfig, *kAllowedLocationsDefault,
-                   nullptr /* gcs additional header */);
+  GcsFileSystem fs(
+      std::unique_ptr(new FakeAuthProvider),
+      std::unique_ptr(
+          new FakeHttpRequestFactory(&requests)),
+      std::unique_ptr(new FakeZoneProvider), 0 /* block size */,
+      0 /* max bytes */, 0 /* max staleness */, 0 /* stat cache max age */,
+      0 /* stat cache max entries */, 0 /* matching paths cache max age */,
+      0 /* matching paths cache max entries */, kTestRetryConfig,
+      kTestTimeoutConfig, *kAllowedLocationsDefault,
+      nullptr /* gcs additional header */, false /* compose append */);
 
   std::vector children;
   TF_EXPECT_OK(fs.GetChildren("gs://bucket/path", &children));
@@ -1885,16 +1887,16 @@ TEST(GcsFileSystemTest, GetMatchingPaths_NoWildcard) {
       "Timeouts: 5 1 10\n",
       "{\"items\": [ "
       "  { \"name\": \"path/subpath/file2.txt\" }]}")});
-  GcsFileSystem fs(std::unique_ptr(new FakeAuthProvider),
-                   std::unique_ptr(
-                       new FakeHttpRequestFactory(&requests)),
-                   std::unique_ptr(new FakeZoneProvider),
-                   0 /* block size */, 0 /* max bytes */, 0 /* max staleness */,
-                   0 /* stat cache max age */, 0 /* stat cache max entries */,
-                   0 /* matching paths cache max age */,
-                   0 /* matching paths cache max entries */, kTestRetryConfig,
-                   kTestTimeoutConfig, *kAllowedLocationsDefault,
-                   nullptr /* gcs additional header */);
+  GcsFileSystem fs(
+      std::unique_ptr(new FakeAuthProvider),
+      std::unique_ptr(
+          new FakeHttpRequestFactory(&requests)),
+      std::unique_ptr(new FakeZoneProvider), 0 /* block size */,
+      0 /* max bytes */, 0 /* max staleness */, 0 /* stat cache max age */,
+      0 /* stat cache max entries */, 0 /* matching paths cache max age */,
+      0 /* matching paths cache max entries */, kTestRetryConfig,
+      kTestTimeoutConfig, *kAllowedLocationsDefault,
+      nullptr /* gcs additional header */, false /* compose append */);
 
   std::vector result;
   TF_EXPECT_OK(
@@ -1913,16 +1915,16 @@ TEST(GcsFileSystemTest, GetMatchingPaths_BucketAndWildcard) {
       "  { \"name\": \"path/file1.txt\" },"
       "  { \"name\": \"path/subpath/file2.txt\" },"
       "  { \"name\": \"path/file3.txt\" }]}")});
-  GcsFileSystem fs(std::unique_ptr(new FakeAuthProvider),
-                   std::unique_ptr(
-                       new FakeHttpRequestFactory(&requests)),
-                   std::unique_ptr(new FakeZoneProvider),
-                   0 /* block size */, 0 /* max bytes */, 0 /* max staleness */,
-                   0 /* stat cache max age */, 0 /* stat cache max entries */,
-                   0 /* matching paths cache max age */,
-                   0 /* matching paths cache max entries */, kTestRetryConfig,
-                   kTestTimeoutConfig, *kAllowedLocationsDefault,
-                   nullptr /* gcs additional header */);
+  GcsFileSystem fs(
+      std::unique_ptr(new FakeAuthProvider),
+      std::unique_ptr(
+          new FakeHttpRequestFactory(&requests)),
+      std::unique_ptr(new FakeZoneProvider), 0 /* block size */,
+      0 /* max bytes */, 0 /* max staleness */, 0 /* stat cache max age */,
+      0 /* stat cache max entries */, 0 /* matching paths cache max age */,
+      0 /* matching paths cache max entries */, kTestRetryConfig,
+      kTestTimeoutConfig, *kAllowedLocationsDefault,
+      nullptr /* gcs additional header */, false /* compose append */);
 
   std::vector result;
   TF_EXPECT_OK(fs.GetMatchingPaths("gs://bucket/*/*", &result));
@@ -1942,16 +1944,16 @@ TEST(GcsFileSystemTest, GetMatchingPaths_FolderAndWildcard_Matches) {
       "  { \"name\": \"path/file1.txt\" },"
       "  { \"name\": \"path/subpath/file2.txt\" },"
       "  { \"name\": \"path/file3.txt\" }]}")});
-  GcsFileSystem fs(std::unique_ptr(new FakeAuthProvider),
-                   std::unique_ptr(
-                       new FakeHttpRequestFactory(&requests)),
-                   std::unique_ptr(new FakeZoneProvider),
-                   0 /* block size */, 0 /* max bytes */, 0 /* max staleness */,
-                   0 /* stat cache max age */, 0 /* stat cache max entries */,
-                   0 /* matching paths cache max age */,
-                   0 /* matching paths cache max entries */, kTestRetryConfig,
-                   kTestTimeoutConfig, *kAllowedLocationsDefault,
-                   nullptr /* gcs additional header */);
+  GcsFileSystem fs(
+      std::unique_ptr(new FakeAuthProvider),
+      std::unique_ptr(
+          new FakeHttpRequestFactory(&requests)),
+      std::unique_ptr(new FakeZoneProvider), 0 /* block size */,
+      0 /* max bytes */, 0 /* max staleness */, 0 /* stat cache max age */,
+      0 /* stat cache max entries */, 0 /* matching paths cache max age */,
+      0 /* matching paths cache max entries */, kTestRetryConfig,
+      kTestTimeoutConfig, *kAllowedLocationsDefault,
+      nullptr /* gcs additional header */, false /* compose append */);
 
   std::vector result;
   TF_EXPECT_OK(fs.GetMatchingPaths("gs://bucket/path/*/file2.txt", &result));
@@ -1968,16 +1970,16 @@ TEST(GcsFileSystemTest, GetMatchingPaths_SelfDirectoryMarker) {
       "{\"items\": [ "
       "  { \"name\": \"path/\" },"
       "  { \"name\": \"path/file3.txt\" }]}")});
-  GcsFileSystem fs(std::unique_ptr(new FakeAuthProvider),
-                   std::unique_ptr(
-                       new FakeHttpRequestFactory(&requests)),
-                   std::unique_ptr(new FakeZoneProvider),
-                   0 /* block size */, 0 /* max bytes */, 0 /* max staleness */,
-                   0 /* stat cache max age */, 0 /* stat cache max entries */,
-                   0 /* matching paths cache max age */,
-                   0 /* matching paths cache max entries */, kTestRetryConfig,
-                   kTestTimeoutConfig, *kAllowedLocationsDefault,
-                   nullptr /* gcs additional header */);
+  GcsFileSystem fs(
+      std::unique_ptr(new FakeAuthProvider),
+      std::unique_ptr(
+          new FakeHttpRequestFactory(&requests)),
+      std::unique_ptr(new FakeZoneProvider), 0 /* block size */,
+      0 /* max bytes */, 0 /* max staleness */, 0 /* stat cache max age */,
+      0 /* stat cache max entries */, 0 /* matching paths cache max age */,
+      0 /* matching paths cache max entries */, kTestRetryConfig,
+      kTestTimeoutConfig, *kAllowedLocationsDefault,
+      nullptr /* gcs additional header */, false /* compose append */);
 
   std::vector result;
   TF_EXPECT_OK(fs.GetMatchingPaths("gs://bucket/path/*", &result));
@@ -1993,16 +1995,16 @@ TEST(GcsFileSystemTest, GetMatchingPaths_SlashInObjectName) {
       "{\"items\": [ "
       "  { \"name\": \"path/\" },"
       "  { \"name\": \"path//foo.txt\" }]}")});
-  GcsFileSystem fs(std::unique_ptr(new FakeAuthProvider),
-                   std::unique_ptr(
-                       new FakeHttpRequestFactory(&requests)),
-                   std::unique_ptr(new FakeZoneProvider),
-                   0 /* block size */, 0 /* max bytes */, 0 /* max staleness */,
-                   0 /* stat cache max age */, 0 /* stat cache max entries */,
-                   0 /* matching paths cache max age */,
-                   0 /* matching paths cache max entries */, kTestRetryConfig,
-                   kTestTimeoutConfig, *kAllowedLocationsDefault,
-                   nullptr /* gcs additional header */);
+  GcsFileSystem fs(
+      std::unique_ptr(new FakeAuthProvider),
+      std::unique_ptr(
+          new FakeHttpRequestFactory(&requests)),
+      std::unique_ptr(new FakeZoneProvider), 0 /* block size */,
+      0 /* max bytes */, 0 /* max staleness */, 0 /* stat cache max age */,
+      0 /* stat cache max entries */, 0 /* matching paths cache max age */,
+      0 /* matching paths cache max entries */, kTestRetryConfig,
+      kTestTimeoutConfig, *kAllowedLocationsDefault,
+      nullptr /* gcs additional header */, false /* compose append */);
 
   std::vector result;
   TF_EXPECT_OK(fs.GetMatchingPaths("gs://bucket/path/*", &result));
@@ -2018,16 +2020,16 @@ TEST(GcsFileSystemTest, GetMatchingPaths_SlashInObjectNameEscaped) {
       "{\"items\": [ "
       "  { \"name\": \"path/\" },"
       "  { \"name\": \"path//foo.txt\" }]}")});
-  GcsFileSystem fs(std::unique_ptr(new FakeAuthProvider),
-                   std::unique_ptr(
-                       new FakeHttpRequestFactory(&requests)),
-                   std::unique_ptr(new FakeZoneProvider),
-                   0 /* block size */, 0 /* max bytes */, 0 /* max staleness */,
-                   0 /* stat cache max age */, 0 /* stat cache max entries */,
-                   0 /* matching paths cache max age */,
-                   0 /* matching paths cache max entries */, kTestRetryConfig,
-                   kTestTimeoutConfig, *kAllowedLocationsDefault,
-                   nullptr /* gcs additional header */);
+  GcsFileSystem fs(
+      std::unique_ptr(new FakeAuthProvider),
+      std::unique_ptr(
+          new FakeHttpRequestFactory(&requests)),
+      std::unique_ptr(new FakeZoneProvider), 0 /* block size */,
+      0 /* max bytes */, 0 /* max staleness */, 0 /* stat cache max age */,
+      0 /* stat cache max entries */, 0 /* matching paths cache max age */,
+      0 /* matching paths cache max entries */, kTestRetryConfig,
+      kTestTimeoutConfig, *kAllowedLocationsDefault,
+      nullptr /* gcs additional header */, false /* compose append */);
 
   std::vector result;
   TF_EXPECT_OK(fs.GetMatchingPaths("gs://bucket/path/\\/*", &result));
@@ -2044,16 +2046,16 @@ TEST(GcsFileSystemTest, GetMatchingPaths_FolderAndWildcard_NoMatches) {
       "  { \"name\": \"path/file1.txt\" },"
       "  { \"name\": \"path/subpath/file2.txt\" },"
       "  { \"name\": \"path/file3.txt\" }]}")});
-  GcsFileSystem fs(std::unique_ptr(new FakeAuthProvider),
-                   std::unique_ptr(
-                       new FakeHttpRequestFactory(&requests)),
-                   std::unique_ptr(new FakeZoneProvider),
-                   0 /* block size */, 0 /* max bytes */, 0 /* max staleness */,
-                   0 /* stat cache max age */, 0 /* stat cache max entries */,
-                   0 /* matching paths cache max age */,
-                   0 /* matching paths cache max entries */, kTestRetryConfig,
-                   kTestTimeoutConfig, *kAllowedLocationsDefault,
-                   nullptr /* gcs additional header */);
+  GcsFileSystem fs(
+      std::unique_ptr(new FakeAuthProvider),
+      std::unique_ptr(
+          new FakeHttpRequestFactory(&requests)),
+      std::unique_ptr(new FakeZoneProvider), 0 /* block size */,
+      0 /* max bytes */, 0 /* max staleness */, 0 /* stat cache max age */,
+      0 /* stat cache max entries */, 0 /* matching paths cache max age */,
+      0 /* matching paths cache max entries */, kTestRetryConfig,
+      kTestTimeoutConfig, *kAllowedLocationsDefault,
+      nullptr /* gcs additional header */, false /* compose append */);
 
   std::vector result;
   TF_EXPECT_OK(fs.GetMatchingPaths("gs://bucket/path/*/file3.txt", &result));
@@ -2062,16 +2064,16 @@ TEST(GcsFileSystemTest, GetMatchingPaths_FolderAndWildcard_NoMatches) {
 
 TEST(GcsFileSystemTest, GetMatchingPaths_OnlyWildcard) {
   std::vector requests;
-  GcsFileSystem fs(std::unique_ptr(new FakeAuthProvider),
-                   std::unique_ptr(
-                       new FakeHttpRequestFactory(&requests)),
-                   std::unique_ptr(new FakeZoneProvider),
-                   0 /* block size */, 0 /* max bytes */, 0 /* max staleness */,
-                   0 /* stat cache max age */, 0 /* stat cache max entries */,
-                   0 /* matching paths cache max age */,
-                   0 /* matching paths cache max entries */, kTestRetryConfig,
-                   kTestTimeoutConfig, *kAllowedLocationsDefault,
-                   nullptr /* gcs additional header */);
+  GcsFileSystem fs(
+      std::unique_ptr(new FakeAuthProvider),
+      std::unique_ptr(
+          new FakeHttpRequestFactory(&requests)),
+      std::unique_ptr(new FakeZoneProvider), 0 /* block size */,
+      0 /* max bytes */, 0 /* max staleness */, 0 /* stat cache max age */,
+      0 /* stat cache max entries */, 0 /* matching paths cache max age */,
+      0 /* matching paths cache max entries */, kTestRetryConfig,
+      kTestTimeoutConfig, *kAllowedLocationsDefault,
+      nullptr /* gcs additional header */, false /* compose append */);
 
   std::vector result;
   EXPECT_EQ(errors::Code::INVALID_ARGUMENT,
@@ -2096,16 +2098,16 @@ TEST(GcsFileSystemTest, GetMatchingPaths_Cache) {
            "  { \"name\": \"path/file1.txt\" },"
            "  { \"name\": \"path/subpath/file2.txt\" },"
            "  { \"name\": \"path/file3.txt\" }]}")});
-  GcsFileSystem fs(std::unique_ptr(new FakeAuthProvider),
-                   std::unique_ptr(
-                       new FakeHttpRequestFactory(&requests)),
-                   std::unique_ptr(new FakeZoneProvider),
-                   0 /* block size */, 0 /* max bytes */, 0 /* max staleness */,
-                   0 /* stat cache max age */, 0 /* stat cache max entries */,
-                   3600 /* matching paths cache max age */,
-                   0 /* matching paths cache max entries */, kTestRetryConfig,
-                   kTestTimeoutConfig, *kAllowedLocationsDefault,
-                   nullptr /* gcs additional header */);
+  GcsFileSystem fs(
+      std::unique_ptr(new FakeAuthProvider),
+      std::unique_ptr(
+          new FakeHttpRequestFactory(&requests)),
+      std::unique_ptr(new FakeZoneProvider), 0 /* block size */,
+      0 /* max bytes */, 0 /* max staleness */, 0 /* stat cache max age */,
+      0 /* stat cache max entries */, 3600 /* matching paths cache max age */,
+      0 /* matching paths cache max entries */, kTestRetryConfig,
+      kTestTimeoutConfig, *kAllowedLocationsDefault,
+      nullptr /* gcs additional header */, false /* compose append */);
 
   // Repeated calls to fs.GetMatchingPaths on these patterns should not lead to
   // any additional HTTP requests to GCS.
@@ -2139,16 +2141,16 @@ TEST(GcsFileSystemTest, GetMatchingPaths_Cache_Flush) {
            "Timeouts: 5 1 10\n",
            "{\"items\": [ "
            "  { \"name\": \"path/subpath/file2.txt\" }]}")});
-  GcsFileSystem fs(std::unique_ptr(new FakeAuthProvider),
-                   std::unique_ptr(
-                       new FakeHttpRequestFactory(&requests)),
-                   std::unique_ptr(new FakeZoneProvider),
-                   0 /* block size */, 0 /* max bytes */, 0 /* max staleness */,
-                   0 /* stat cache max age */, 0 /* stat cache max entries */,
-                   3600 /* matching paths cache max age */,
-                   0 /* matching paths cache max entries */, kTestRetryConfig,
-                   kTestTimeoutConfig, *kAllowedLocationsDefault,
-                   nullptr /* gcs additional header */);
+  GcsFileSystem fs(
+      std::unique_ptr(new FakeAuthProvider),
+      std::unique_ptr(
+          new FakeHttpRequestFactory(&requests)),
+      std::unique_ptr(new FakeZoneProvider), 0 /* block size */,
+      0 /* max bytes */, 0 /* max staleness */, 0 /* stat cache max age */,
+      0 /* stat cache max entries */, 3600 /* matching paths cache max age */,
+      0 /* matching paths cache max entries */, kTestRetryConfig,
+      kTestTimeoutConfig, *kAllowedLocationsDefault,
+      nullptr /* gcs additional header */, false /* compose append */);
 
   // This loop should trigger the first HTTP request to GCS.
   for (int i = 0; i < 10; i++) {
@@ -2212,7 +2214,7 @@ TEST(GcsFileSystemTest, DeleteFile) {
       0 /* stat cache max entries */, 0 /* matching paths cache max age */,
       0 /* matching paths cache max entries */, kTestRetryConfig,
       kTestTimeoutConfig, *kAllowedLocationsDefault,
-      nullptr /* gcs additional header */);
+      nullptr /* gcs additional header */, false /* compose append */);
 
   // Do an initial read of the file to load its contents into the block cache.
   char scratch[100];
@@ -2231,16 +2233,16 @@ TEST(GcsFileSystemTest, DeleteFile) {
 
 TEST(GcsFileSystemTest, DeleteFile_NoObjectName) {
   std::vector requests;
-  GcsFileSystem fs(std::unique_ptr(new FakeAuthProvider),
-                   std::unique_ptr(
-                       new FakeHttpRequestFactory(&requests)),
-                   std::unique_ptr(new FakeZoneProvider),
-                   0 /* block size */, 0 /* max bytes */, 0 /* max staleness */,
-                   0 /* stat cache max age */, 0 /* stat cache max entries */,
-                   0 /* matching paths cache max age */,
-                   0 /* matching paths cache max entries */, kTestRetryConfig,
-                   kTestTimeoutConfig, *kAllowedLocationsDefault,
-                   nullptr /* gcs additional header */);
+  GcsFileSystem fs(
+      std::unique_ptr(new FakeAuthProvider),
+      std::unique_ptr(
+          new FakeHttpRequestFactory(&requests)),
+      std::unique_ptr(new FakeZoneProvider), 0 /* block size */,
+      0 /* max bytes */, 0 /* max staleness */, 0 /* stat cache max age */,
+      0 /* stat cache max entries */, 0 /* matching paths cache max age */,
+      0 /* matching paths cache max entries */, kTestRetryConfig,
+      kTestTimeoutConfig, *kAllowedLocationsDefault,
+      nullptr /* gcs additional header */, false /* compose append */);
 
   EXPECT_EQ(errors::Code::INVALID_ARGUMENT,
             fs.DeleteFile("gs://bucket/").code());
@@ -2283,7 +2285,7 @@ TEST(GcsFileSystemTest, DeleteFile_StatCacheRemoved) {
       0 /* stat cache max entries */, 0 /* matching paths cache max age */,
       0 /* matching paths cache max entries */, kTestRetryConfig,
       kTestTimeoutConfig, *kAllowedLocationsDefault,
-      nullptr /* gcs additional header */);
+      nullptr /* gcs additional header */, false /* compose append */);
 
   // Stats the file first so the stat is cached.
   FileStatistics stat_before_deletion;
@@ -2304,16 +2306,16 @@ TEST(GcsFileSystemTest, DeleteDir_Empty) {
       "Auth Token: fake_token\n"
       "Timeouts: 5 1 10\n",
       "{}")});
-  GcsFileSystem fs(std::unique_ptr(new FakeAuthProvider),
-                   std::unique_ptr(
-                       new FakeHttpRequestFactory(&requests)),
-                   std::unique_ptr(new FakeZoneProvider),
-                   0 /* block size */, 0 /* max bytes */, 0 /* max staleness */,
-                   0 /* stat cache max age */, 0 /* stat cache max entries */,
-                   0 /* matching paths cache max age */,
-                   0 /* matching paths cache max entries */, kTestRetryConfig,
-                   kTestTimeoutConfig, *kAllowedLocationsDefault,
-                   nullptr /* gcs additional header */);
+  GcsFileSystem fs(
+      std::unique_ptr(new FakeAuthProvider),
+      std::unique_ptr(
+          new FakeHttpRequestFactory(&requests)),
+      std::unique_ptr(new FakeZoneProvider), 0 /* block size */,
+      0 /* max bytes */, 0 /* max staleness */, 0 /* stat cache max age */,
+      0 /* stat cache max entries */, 0 /* matching paths cache max age */,
+      0 /* matching paths cache max entries */, kTestRetryConfig,
+      kTestTimeoutConfig, *kAllowedLocationsDefault,
+      nullptr /* gcs additional header */, false /* compose append */);
 
   TF_EXPECT_OK(fs.DeleteDir("gs://bucket/path/"));
 }
@@ -2333,16 +2335,16 @@ TEST(GcsFileSystemTest, DeleteDir_OnlyDirMarkerLeft) {
                            "Timeouts: 5 1 10\n"
                            "Delete: yes\n",
                            "")});
-  GcsFileSystem fs(std::unique_ptr(new FakeAuthProvider),
-                   std::unique_ptr(
-                       new FakeHttpRequestFactory(&requests)),
-                   std::unique_ptr(new FakeZoneProvider),
-                   0 /* block size */, 0 /* max bytes */, 0 /* max staleness */,
-                   0 /* stat cache max age */, 0 /* stat cache max entries */,
-                   0 /* matching paths cache max age */,
-                   0 /* matching paths cache max entries */, kTestRetryConfig,
-                   kTestTimeoutConfig, *kAllowedLocationsDefault,
-                   nullptr /* gcs additional header */);
+  GcsFileSystem fs(
+      std::unique_ptr(new FakeAuthProvider),
+      std::unique_ptr(
+          new FakeHttpRequestFactory(&requests)),
+      std::unique_ptr(new FakeZoneProvider), 0 /* block size */,
+      0 /* max bytes */, 0 /* max staleness */, 0 /* stat cache max age */,
+      0 /* stat cache max entries */, 0 /* matching paths cache max age */,
+      0 /* matching paths cache max entries */, kTestRetryConfig,
+      kTestTimeoutConfig, *kAllowedLocationsDefault,
+      nullptr /* gcs additional header */, false /* compose append */);
 
   TF_EXPECT_OK(fs.DeleteDir("gs://bucket/path/"));
 }
@@ -2353,16 +2355,16 @@ TEST(GcsFileSystemTest, DeleteDir_BucketOnly) {
       "name%2CnextPageToken&maxResults=2\nAuth Token: fake_token\n"
       "Timeouts: 5 1 10\n",
       "{}")});
-  GcsFileSystem fs(std::unique_ptr(new FakeAuthProvider),
-                   std::unique_ptr(
-                       new FakeHttpRequestFactory(&requests)),
-                   std::unique_ptr(new FakeZoneProvider),
-                   0 /* block size */, 0 /* max bytes */, 0 /* max staleness */,
-                   0 /* stat cache max age */, 0 /* stat cache max entries */,
-                   0 /* matching paths cache max age */,
-                   0 /* matching paths cache max entries */, kTestRetryConfig,
-                   kTestTimeoutConfig, *kAllowedLocationsDefault,
-                   nullptr /* gcs additional header */);
+  GcsFileSystem fs(
+      std::unique_ptr(new FakeAuthProvider),
+      std::unique_ptr(
+          new FakeHttpRequestFactory(&requests)),
+      std::unique_ptr(new FakeZoneProvider), 0 /* block size */,
+      0 /* max bytes */, 0 /* max staleness */, 0 /* stat cache max age */,
+      0 /* stat cache max entries */, 0 /* matching paths cache max age */,
+      0 /* matching paths cache max entries */, kTestRetryConfig,
+      kTestTimeoutConfig, *kAllowedLocationsDefault,
+      nullptr /* gcs additional header */, false /* compose append */);
 
   TF_EXPECT_OK(fs.DeleteDir("gs://bucket"));
 }
@@ -2375,16 +2377,16 @@ TEST(GcsFileSystemTest, DeleteDir_NonEmpty) {
       "Timeouts: 5 1 10\n",
       "{\"items\": [ "
       "  { \"name\": \"path/file1.txt\" }]}")});
-  GcsFileSystem fs(std::unique_ptr(new FakeAuthProvider),
-                   std::unique_ptr(
-                       new FakeHttpRequestFactory(&requests)),
-                   std::unique_ptr(new FakeZoneProvider),
-                   0 /* block size */, 0 /* max bytes */, 0 /* max staleness */,
-                   0 /* stat cache max age */, 0 /* stat cache max entries */,
-                   0 /* matching paths cache max age */,
-                   0 /* matching paths cache max entries */, kTestRetryConfig,
-                   kTestTimeoutConfig, *kAllowedLocationsDefault,
-                   nullptr /* gcs additional header */);
+  GcsFileSystem fs(
+      std::unique_ptr(new FakeAuthProvider),
+      std::unique_ptr(
+          new FakeHttpRequestFactory(&requests)),
+      std::unique_ptr(new FakeZoneProvider), 0 /* block size */,
+      0 /* max bytes */, 0 /* max staleness */, 0 /* stat cache max age */,
+      0 /* stat cache max entries */, 0 /* matching paths cache max age */,
+      0 /* matching paths cache max entries */, kTestRetryConfig,
+      kTestTimeoutConfig, *kAllowedLocationsDefault,
+      nullptr /* gcs additional header */, false /* compose append */);
 
   EXPECT_EQ(error::Code::FAILED_PRECONDITION,
             fs.DeleteDir("gs://bucket/path/").code());
@@ -2398,16 +2400,16 @@ TEST(GcsFileSystemTest, GetFileSize) {
       "Timeouts: 5 1 10\n",
       strings::StrCat("{\"size\": \"1010\",\"generation\": \"1\","
                       "\"updated\": \"2016-04-29T23:15:24.896Z\"}"))});
-  GcsFileSystem fs(std::unique_ptr(new FakeAuthProvider),
-                   std::unique_ptr(
-                       new FakeHttpRequestFactory(&requests)),
-                   std::unique_ptr(new FakeZoneProvider),
-                   0 /* block size */, 0 /* max bytes */, 0 /* max staleness */,
-                   0 /* stat cache max age */, 0 /* stat cache max entries */,
-                   0 /* matching paths cache max age */,
-                   0 /* matching paths cache max entries */, kTestRetryConfig,
-                   kTestTimeoutConfig, *kAllowedLocationsDefault,
-                   nullptr /* gcs additional header */);
+  GcsFileSystem fs(
+      std::unique_ptr(new FakeAuthProvider),
+      std::unique_ptr(
+          new FakeHttpRequestFactory(&requests)),
+      std::unique_ptr(new FakeZoneProvider), 0 /* block size */,
+      0 /* max bytes */, 0 /* max staleness */, 0 /* stat cache max age */,
+      0 /* stat cache max entries */, 0 /* matching paths cache max age */,
+      0 /* matching paths cache max entries */, kTestRetryConfig,
+      kTestTimeoutConfig, *kAllowedLocationsDefault,
+      nullptr /* gcs additional header */, false /* compose append */);
 
   uint64 size;
   TF_EXPECT_OK(fs.GetFileSize("gs://bucket/file.txt", &size));
@@ -2416,16 +2418,16 @@ TEST(GcsFileSystemTest, GetFileSize) {
 
 TEST(GcsFileSystemTest, GetFileSize_NoObjectName) {
   std::vector requests;
-  GcsFileSystem fs(std::unique_ptr(new FakeAuthProvider),
-                   std::unique_ptr(
-                       new FakeHttpRequestFactory(&requests)),
-                   std::unique_ptr(new FakeZoneProvider),
-                   0 /* block size */, 0 /* max bytes */, 0 /* max staleness */,
-                   0 /* stat cache max age */, 0 /* stat cache max entries */,
-                   0 /* matching paths cache max age */,
-                   0 /* matching paths cache max entries */, kTestRetryConfig,
-                   kTestTimeoutConfig, *kAllowedLocationsDefault,
-                   nullptr /* gcs additional header */);
+  GcsFileSystem fs(
+      std::unique_ptr(new FakeAuthProvider),
+      std::unique_ptr(
+          new FakeHttpRequestFactory(&requests)),
+      std::unique_ptr(new FakeZoneProvider), 0 /* block size */,
+      0 /* max bytes */, 0 /* max staleness */, 0 /* stat cache max age */,
+      0 /* stat cache max entries */, 0 /* matching paths cache max age */,
+      0 /* matching paths cache max entries */, kTestRetryConfig,
+      kTestTimeoutConfig, *kAllowedLocationsDefault,
+      nullptr /* gcs additional header */, false /* compose append */);
 
   uint64 size;
   EXPECT_EQ(errors::Code::INVALID_ARGUMENT,
@@ -2502,16 +2504,16 @@ TEST(GcsFileSystemTest, RenameFile_Folder) {
            "Timeouts: 5 1 10\n"
            "Delete: yes\n",
            "")});
-  GcsFileSystem fs(std::unique_ptr(new FakeAuthProvider),
-                   std::unique_ptr(
-                       new FakeHttpRequestFactory(&requests)),
-                   std::unique_ptr(new FakeZoneProvider),
-                   0 /* block size */, 0 /* max bytes */, 0 /* max staleness */,
-                   0 /* stat cache max age */, 0 /* stat cache max entries */,
-                   0 /* matching paths cache max age */,
-                   0 /* matching paths cache max entries */, kTestRetryConfig,
-                   kTestTimeoutConfig, *kAllowedLocationsDefault,
-                   nullptr /* gcs additional header */);
+  GcsFileSystem fs(
+      std::unique_ptr(new FakeAuthProvider),
+      std::unique_ptr(
+          new FakeHttpRequestFactory(&requests)),
+      std::unique_ptr(new FakeZoneProvider), 0 /* block size */,
+      0 /* max bytes */, 0 /* max staleness */, 0 /* stat cache max age */,
+      0 /* stat cache max entries */, 0 /* matching paths cache max age */,
+      0 /* matching paths cache max entries */, kTestRetryConfig,
+      kTestTimeoutConfig, *kAllowedLocationsDefault,
+      nullptr /* gcs additional header */, false /* compose append */);
 
   TF_EXPECT_OK(fs.RenameFile("gs://bucket/path1", "gs://bucket/path2/"));
 }
@@ -2603,7 +2605,7 @@ TEST(GcsFileSystemTest, RenameFile_Object) {
       0 /* stat cache max entries */, 0 /* matching paths cache max age */,
       0 /* matching paths cache max entries */, kTestRetryConfig,
       kTestTimeoutConfig, *kAllowedLocationsDefault,
-      nullptr /* gcs additional header */);
+      nullptr /* gcs additional header */, false /* compose append */);
   // Do an initial read of the source and destination files to load their
   // contents into the block cache.
   char scratch[100];
@@ -2684,7 +2686,7 @@ TEST(GcsFileSystemTest, RenameFile_Object_FlushTargetStatCache) {
       0 /* stat cache max entries */, 0 /* matching paths cache max age */,
       0 /* matching paths cache max entries */, kTestRetryConfig,
       kTestTimeoutConfig, *kAllowedLocationsDefault,
-      nullptr /* gcs additional header */);
+      nullptr /* gcs additional header */, false /* compose append */);
   // Do an initial stat of the destination file to load their contents into the
   // stat cache.
   FileStatistics stat_before_renaming;
@@ -2742,16 +2744,16 @@ TEST(GcsFileSystemTest, RenameFile_Object_DeletionRetried) {
            "Timeouts: 5 1 10\n"
            "Delete: yes\n",
            "", errors::NotFound("404"), 404)});
-  GcsFileSystem fs(std::unique_ptr(new FakeAuthProvider),
-                   std::unique_ptr(
-                       new FakeHttpRequestFactory(&requests)),
-                   std::unique_ptr(new FakeZoneProvider),
-                   0 /* block size */, 0 /* max bytes */, 0 /* max staleness */,
-                   0 /* stat cache max age */, 0 /* stat cache max entries */,
-                   0 /* matching paths cache max age */,
-                   0 /* matching paths cache max entries */, kTestRetryConfig,
-                   kTestTimeoutConfig, *kAllowedLocationsDefault,
-                   nullptr /* gcs additional header */);
+  GcsFileSystem fs(
+      std::unique_ptr(new FakeAuthProvider),
+      std::unique_ptr(
+          new FakeHttpRequestFactory(&requests)),
+      std::unique_ptr(new FakeZoneProvider), 0 /* block size */,
+      0 /* max bytes */, 0 /* max staleness */, 0 /* stat cache max age */,
+      0 /* stat cache max entries */, 0 /* matching paths cache max age */,
+      0 /* matching paths cache max entries */, kTestRetryConfig,
+      kTestTimeoutConfig, *kAllowedLocationsDefault,
+      nullptr /* gcs additional header */, false /* compose append */);
 
   TF_EXPECT_OK(
       fs.RenameFile("gs://bucket/path/src.txt", "gs://bucket/path/dst.txt"));
@@ -2784,16 +2786,16 @@ TEST(GcsFileSystemTest, RenameFile_Object_Incomplete) {
            "Post: yes\n"
            "Timeouts: 5 1 10\n",
            "{\"done\": false}")});
-  GcsFileSystem fs(std::unique_ptr(new FakeAuthProvider),
-                   std::unique_ptr(
-                       new FakeHttpRequestFactory(&requests)),
-                   std::unique_ptr(new FakeZoneProvider),
-                   0 /* block size */, 0 /* max bytes */, 0 /* max staleness */,
-                   0 /* stat cache max age */, 0 /* stat cache max entries */,
-                   0 /* matching paths cache max age */,
-                   0 /* matching paths cache max entries */, kTestRetryConfig,
-                   kTestTimeoutConfig, *kAllowedLocationsDefault,
-                   nullptr /* gcs additional header */);
+  GcsFileSystem fs(
+      std::unique_ptr(new FakeAuthProvider),
+      std::unique_ptr(
+          new FakeHttpRequestFactory(&requests)),
+      std::unique_ptr(new FakeZoneProvider), 0 /* block size */,
+      0 /* max bytes */, 0 /* max staleness */, 0 /* stat cache max age */,
+      0 /* stat cache max entries */, 0 /* matching paths cache max age */,
+      0 /* matching paths cache max entries */, kTestRetryConfig,
+      kTestTimeoutConfig, *kAllowedLocationsDefault,
+      nullptr /* gcs additional header */, false /* compose append */);
 
   EXPECT_EQ(
       errors::Code::UNIMPLEMENTED,
@@ -2809,16 +2811,16 @@ TEST(GcsFileSystemTest, Stat_Object) {
       "Timeouts: 5 1 10\n",
       strings::StrCat("{\"size\": \"1010\",\"generation\": \"1\","
                       "\"updated\": \"2016-04-29T23:15:24.896Z\"}"))});
-  GcsFileSystem fs(std::unique_ptr(new FakeAuthProvider),
-                   std::unique_ptr(
-                       new FakeHttpRequestFactory(&requests)),
-                   std::unique_ptr(new FakeZoneProvider),
-                   0 /* block size */, 0 /* max bytes */, 0 /* max staleness */,
-                   0 /* stat cache max age */, 0 /* stat cache max entries */,
-                   0 /* matching paths cache max age */,
-                   0 /* matching paths cache max entries */, kTestRetryConfig,
-                   kTestTimeoutConfig, *kAllowedLocationsDefault,
-                   nullptr /* gcs additional header */);
+  GcsFileSystem fs(
+      std::unique_ptr(new FakeAuthProvider),
+      std::unique_ptr(
+          new FakeHttpRequestFactory(&requests)),
+      std::unique_ptr(new FakeZoneProvider), 0 /* block size */,
+      0 /* max bytes */, 0 /* max staleness */, 0 /* stat cache max age */,
+      0 /* stat cache max entries */, 0 /* matching paths cache max age */,
+      0 /* matching paths cache max entries */, kTestRetryConfig,
+      kTestTimeoutConfig, *kAllowedLocationsDefault,
+      nullptr /* gcs additional header */, false /* compose append */);
 
   FileStatistics stat;
   TF_EXPECT_OK(fs.Stat("gs://bucket/file.txt", &stat));
@@ -2843,16 +2845,16 @@ TEST(GcsFileSystemTest, Stat_Folder) {
            "Timeouts: 5 1 10\n",
            "{\"items\": [ "
            "  { \"name\": \"subfolder/\" }]}")});
-  GcsFileSystem fs(std::unique_ptr(new FakeAuthProvider),
-                   std::unique_ptr(
-                       new FakeHttpRequestFactory(&requests)),
-                   std::unique_ptr(new FakeZoneProvider),
-                   0 /* block size */, 0 /* max bytes */, 0 /* max staleness */,
-                   0 /* stat cache max age */, 0 /* stat cache max entries */,
-                   0 /* matching paths cache max age */,
-                   0 /* matching paths cache max entries */, kTestRetryConfig,
-                   kTestTimeoutConfig, *kAllowedLocationsDefault,
-                   nullptr /* gcs additional header */);
+  GcsFileSystem fs(
+      std::unique_ptr(new FakeAuthProvider),
+      std::unique_ptr(
+          new FakeHttpRequestFactory(&requests)),
+      std::unique_ptr(new FakeZoneProvider), 0 /* block size */,
+      0 /* max bytes */, 0 /* max staleness */, 0 /* stat cache max age */,
+      0 /* stat cache max entries */, 0 /* matching paths cache max age */,
+      0 /* matching paths cache max entries */, kTestRetryConfig,
+      kTestTimeoutConfig, *kAllowedLocationsDefault,
+      nullptr /* gcs additional header */, false /* compose append */);
 
   FileStatistics stat;
   TF_EXPECT_OK(fs.Stat("gs://bucket/subfolder", &stat));
@@ -2876,16 +2878,16 @@ TEST(GcsFileSystemTest, Stat_ObjectOrFolderNotFound) {
            "Auth Token: fake_token\n"
            "Timeouts: 5 1 10\n",
            "{}")});
-  GcsFileSystem fs(std::unique_ptr(new FakeAuthProvider),
-                   std::unique_ptr(
-                       new FakeHttpRequestFactory(&requests)),
-                   std::unique_ptr(new FakeZoneProvider),
-                   0 /* block size */, 0 /* max bytes */, 0 /* max staleness */,
-                   0 /* stat cache max age */, 0 /* stat cache max entries */,
-                   0 /* matching paths cache max age */,
-                   0 /* matching paths cache max entries */, kTestRetryConfig,
-                   kTestTimeoutConfig, *kAllowedLocationsDefault,
-                   nullptr /* gcs additional header */);
+  GcsFileSystem fs(
+      std::unique_ptr(new FakeAuthProvider),
+      std::unique_ptr(
+          new FakeHttpRequestFactory(&requests)),
+      std::unique_ptr(new FakeZoneProvider), 0 /* block size */,
+      0 /* max bytes */, 0 /* max staleness */, 0 /* stat cache max age */,
+      0 /* stat cache max entries */, 0 /* matching paths cache max age */,
+      0 /* matching paths cache max entries */, kTestRetryConfig,
+      kTestTimeoutConfig, *kAllowedLocationsDefault,
+      nullptr /* gcs additional header */, false /* compose append */);
 
   FileStatistics stat;
   EXPECT_EQ(error::Code::NOT_FOUND, fs.Stat("gs://bucket/path", &stat).code());
@@ -2897,16 +2899,16 @@ TEST(GcsFileSystemTest, Stat_Bucket) {
       "Auth Token: fake_token\n"
       "Timeouts: 5 1 10\n",
       "{}")});
-  GcsFileSystem fs(std::unique_ptr(new FakeAuthProvider),
-                   std::unique_ptr(
-                       new FakeHttpRequestFactory(&requests)),
-                   std::unique_ptr(new FakeZoneProvider),
-                   0 /* block size */, 0 /* max bytes */, 0 /* max staleness */,
-                   0 /* stat cache max age */, 0 /* stat cache max entries */,
-                   0 /* matching paths cache max age */,
-                   0 /* matching paths cache max entries */, kTestRetryConfig,
-                   kTestTimeoutConfig, *kAllowedLocationsDefault,
-                   nullptr /* gcs additional header */);
+  GcsFileSystem fs(
+      std::unique_ptr(new FakeAuthProvider),
+      std::unique_ptr(
+          new FakeHttpRequestFactory(&requests)),
+      std::unique_ptr(new FakeZoneProvider), 0 /* block size */,
+      0 /* max bytes */, 0 /* max staleness */, 0 /* stat cache max age */,
+      0 /* stat cache max entries */, 0 /* matching paths cache max age */,
+      0 /* matching paths cache max entries */, kTestRetryConfig,
+      kTestTimeoutConfig, *kAllowedLocationsDefault,
+      nullptr /* gcs additional header */, false /* compose append */);
 
   FileStatistics stat;
   TF_EXPECT_OK(fs.Stat("gs://bucket/", &stat));
@@ -2921,16 +2923,16 @@ TEST(GcsFileSystemTest, Stat_BucketNotFound) {
       "Auth Token: fake_token\n"
       "Timeouts: 5 1 10\n",
       "", errors::NotFound("404"), 404)});
-  GcsFileSystem fs(std::unique_ptr(new FakeAuthProvider),
-                   std::unique_ptr(
-                       new FakeHttpRequestFactory(&requests)),
-                   std::unique_ptr(new FakeZoneProvider),
-                   0 /* block size */, 0 /* max bytes */, 0 /* max staleness */,
-                   0 /* stat cache max age */, 0 /* stat cache max entries */,
-                   0 /* matching paths cache max age */,
-                   0 /* matching paths cache max entries */, kTestRetryConfig,
-                   kTestTimeoutConfig, *kAllowedLocationsDefault,
-                   nullptr /* gcs additional header */);
+  GcsFileSystem fs(
+      std::unique_ptr(new FakeAuthProvider),
+      std::unique_ptr(
+          new FakeHttpRequestFactory(&requests)),
+      std::unique_ptr(new FakeZoneProvider), 0 /* block size */,
+      0 /* max bytes */, 0 /* max staleness */, 0 /* stat cache max age */,
+      0 /* stat cache max entries */, 0 /* matching paths cache max age */,
+      0 /* matching paths cache max entries */, kTestRetryConfig,
+      kTestTimeoutConfig, *kAllowedLocationsDefault,
+      nullptr /* gcs additional header */, false /* compose append */);
 
   FileStatistics stat;
   EXPECT_EQ(error::Code::NOT_FOUND, fs.Stat("gs://bucket/", &stat).code());
@@ -2968,7 +2970,7 @@ TEST(GcsFileSystemTest, Stat_Cache) {
       0 /* stat cache max entries */, 0 /* matching paths cache max age */,
       0 /* matching paths cache max entries */, kTestRetryConfig,
       kTestTimeoutConfig, *kAllowedLocationsDefault,
-      nullptr /* gcs additional header */);
+      nullptr /* gcs additional header */, false /* compose append */);
 
   // Repeated calls to fs.Stat on these paths should not lead to any additional
   // HTTP requests to GCS.
@@ -3010,7 +3012,7 @@ TEST(GcsFileSystemTest, Stat_Cache_Flush) {
       0 /* stat cache max entries */, 0 /* matching paths cache max age */,
       0 /* matching paths cache max entries */, kTestRetryConfig,
       kTestTimeoutConfig, *kAllowedLocationsDefault,
-      nullptr /* gcs additional header */);
+      nullptr /* gcs additional header */, false /* compose append */);
   // There should be a single HTTP request to GCS for fs.Stat in this loop.
   for (int i = 0; i < 10; i++) {
     FileStatistics stat;
@@ -3038,16 +3040,16 @@ TEST(GcsFileSystemTest, Stat_FilenameEndingWithSlash) {
       "Timeouts: 5 1 10\n",
       strings::StrCat("{\"size\": \"5\",\"generation\": \"1\","
                       "\"updated\": \"2016-04-29T23:15:24.896Z\"}"))});
-  GcsFileSystem fs(std::unique_ptr(new FakeAuthProvider),
-                   std::unique_ptr(
-                       new FakeHttpRequestFactory(&requests)),
-                   std::unique_ptr(new FakeZoneProvider),
-                   0 /* block size */, 0 /* max bytes */, 0 /* max staleness */,
-                   0 /* stat cache max age */, 0 /* stat cache max entries */,
-                   0 /* matching paths cache max age */,
-                   0 /* matching paths cache max entries */, kTestRetryConfig,
-                   kTestTimeoutConfig, *kAllowedLocationsDefault,
-                   nullptr /* gcs additional header */);
+  GcsFileSystem fs(
+      std::unique_ptr(new FakeAuthProvider),
+      std::unique_ptr(
+          new FakeHttpRequestFactory(&requests)),
+      std::unique_ptr(new FakeZoneProvider), 0 /* block size */,
+      0 /* max bytes */, 0 /* max staleness */, 0 /* stat cache max age */,
+      0 /* stat cache max entries */, 0 /* matching paths cache max age */,
+      0 /* matching paths cache max entries */, kTestRetryConfig,
+      kTestTimeoutConfig, *kAllowedLocationsDefault,
+      nullptr /* gcs additional header */, false /* compose append */);
 
   FileStatistics stat;
   TF_EXPECT_OK(fs.Stat("gs://bucket/dir/", &stat));
@@ -3070,16 +3072,16 @@ TEST(GcsFileSystemTest, IsDirectory_NotFound) {
            "Auth Token: fake_token\n"
            "Timeouts: 5 1 10\n",
            "", errors::NotFound("404"), 404)});
-  GcsFileSystem fs(std::unique_ptr(new FakeAuthProvider),
-                   std::unique_ptr(
-                       new FakeHttpRequestFactory(&requests)),
-                   std::unique_ptr(new FakeZoneProvider),
-                   0 /* block size */, 0 /* max bytes */, 0 /* max staleness */,
-                   0 /* stat cache max age */, 0 /* stat cache max entries */,
-                   0 /* matching paths cache max age */,
-                   0 /* matching paths cache max entries */, kTestRetryConfig,
-                   kTestTimeoutConfig, *kAllowedLocationsDefault,
-                   nullptr /* gcs additional header */);
+  GcsFileSystem fs(
+      std::unique_ptr(new FakeAuthProvider),
+      std::unique_ptr(
+          new FakeHttpRequestFactory(&requests)),
+      std::unique_ptr(new FakeZoneProvider), 0 /* block size */,
+      0 /* max bytes */, 0 /* max staleness */, 0 /* stat cache max age */,
+      0 /* stat cache max entries */, 0 /* matching paths cache max age */,
+      0 /* matching paths cache max entries */, kTestRetryConfig,
+      kTestTimeoutConfig, *kAllowedLocationsDefault,
+      nullptr /* gcs additional header */, false /* compose append */);
 
   EXPECT_EQ(error::Code::NOT_FOUND,
             fs.IsDirectory("gs://bucket/file.txt").code());
@@ -3101,16 +3103,16 @@ TEST(GcsFileSystemTest, IsDirectory_NotDirectoryButObject) {
            "Timeouts: 5 1 10\n",
            strings::StrCat("{\"size\": \"1010\",\"generation\": \"1\","
                            "\"updated\": \"2016-04-29T23:15:24.896Z\"}"))});
-  GcsFileSystem fs(std::unique_ptr(new FakeAuthProvider),
-                   std::unique_ptr(
-                       new FakeHttpRequestFactory(&requests)),
-                   std::unique_ptr(new FakeZoneProvider),
-                   0 /* block size */, 0 /* max bytes */, 0 /* max staleness */,
-                   0 /* stat cache max age */, 0 /* stat cache max entries */,
-                   0 /* matching paths cache max age */,
-                   0 /* matching paths cache max entries */, kTestRetryConfig,
-                   kTestTimeoutConfig, *kAllowedLocationsDefault,
-                   nullptr /* gcs additional header */);
+  GcsFileSystem fs(
+      std::unique_ptr(new FakeAuthProvider),
+      std::unique_ptr(
+          new FakeHttpRequestFactory(&requests)),
+      std::unique_ptr(new FakeZoneProvider), 0 /* block size */,
+      0 /* max bytes */, 0 /* max staleness */, 0 /* stat cache max age */,
+      0 /* stat cache max entries */, 0 /* matching paths cache max age */,
+      0 /* matching paths cache max entries */, kTestRetryConfig,
+      kTestTimeoutConfig, *kAllowedLocationsDefault,
+      nullptr /* gcs additional header */, false /* compose append */);
 
   EXPECT_EQ(error::Code::FAILED_PRECONDITION,
             fs.IsDirectory("gs://bucket/file.txt").code());
@@ -3132,16 +3134,16 @@ TEST(GcsFileSystemTest, IsDirectory_Yes) {
            "Auth Token: fake_token\n"
            "Timeouts: 5 1 10\n",
            "{\"items\": [{\"name\": \"subfolder/\"}]}")});
-  GcsFileSystem fs(std::unique_ptr(new FakeAuthProvider),
-                   std::unique_ptr(
-                       new FakeHttpRequestFactory(&requests)),
-                   std::unique_ptr(new FakeZoneProvider),
-                   0 /* block size */, 0 /* max bytes */, 0 /* max staleness */,
-                   0 /* stat cache max age */, 0 /* stat cache max entries */,
-                   0 /* matching paths cache max age */,
-                   0 /* matching paths cache max entries */, kTestRetryConfig,
-                   kTestTimeoutConfig, *kAllowedLocationsDefault,
-                   nullptr /* gcs additional header */);
+  GcsFileSystem fs(
+      std::unique_ptr(new FakeAuthProvider),
+      std::unique_ptr(
+          new FakeHttpRequestFactory(&requests)),
+      std::unique_ptr(new FakeZoneProvider), 0 /* block size */,
+      0 /* max bytes */, 0 /* max staleness */, 0 /* stat cache max age */,
+      0 /* stat cache max entries */, 0 /* matching paths cache max age */,
+      0 /* matching paths cache max entries */, kTestRetryConfig,
+      kTestTimeoutConfig, *kAllowedLocationsDefault,
+      nullptr /* gcs additional header */, false /* compose append */);
 
   TF_EXPECT_OK(fs.IsDirectory("gs://bucket/subfolder"));
   TF_EXPECT_OK(fs.IsDirectory("gs://bucket/subfolder/"));
@@ -3159,16 +3161,16 @@ TEST(GcsFileSystemTest, IsDirectory_Bucket) {
            "Auth Token: fake_token\n"
            "Timeouts: 5 1 10\n",
            "{}")});
-  GcsFileSystem fs(std::unique_ptr(new FakeAuthProvider),
-                   std::unique_ptr(
-                       new FakeHttpRequestFactory(&requests)),
-                   std::unique_ptr(new FakeZoneProvider),
-                   0 /* block size */, 0 /* max bytes */, 0 /* max staleness */,
-                   0 /* stat cache max age */, 0 /* stat cache max entries */,
-                   0 /* matching paths cache max age */,
-                   0 /* matching paths cache max entries */, kTestRetryConfig,
-                   kTestTimeoutConfig, *kAllowedLocationsDefault,
-                   nullptr /* gcs additional header */);
+  GcsFileSystem fs(
+      std::unique_ptr(new FakeAuthProvider),
+      std::unique_ptr(
+          new FakeHttpRequestFactory(&requests)),
+      std::unique_ptr(new FakeZoneProvider), 0 /* block size */,
+      0 /* max bytes */, 0 /* max staleness */, 0 /* stat cache max age */,
+      0 /* stat cache max entries */, 0 /* matching paths cache max age */,
+      0 /* matching paths cache max entries */, kTestRetryConfig,
+      kTestTimeoutConfig, *kAllowedLocationsDefault,
+      nullptr /* gcs additional header */, false /* compose append */);
 
   TF_EXPECT_OK(fs.IsDirectory("gs://bucket"));
   TF_EXPECT_OK(fs.IsDirectory("gs://bucket/"));
@@ -3180,16 +3182,16 @@ TEST(GcsFileSystemTest, IsDirectory_BucketNotFound) {
       "Auth Token: fake_token\n"
       "Timeouts: 5 1 10\n",
       "", errors::NotFound("404"), 404)});
-  GcsFileSystem fs(std::unique_ptr(new FakeAuthProvider),
-                   std::unique_ptr(
-                       new FakeHttpRequestFactory(&requests)),
-                   std::unique_ptr(new FakeZoneProvider),
-                   0 /* block size */, 0 /* max bytes */, 0 /* max staleness */,
-                   0 /* stat cache max age */, 0 /* stat cache max entries */,
-                   0 /* matching paths cache max age */,
-                   0 /* matching paths cache max entries */, kTestRetryConfig,
-                   kTestTimeoutConfig, *kAllowedLocationsDefault,
-                   nullptr /* gcs additional header */);
+  GcsFileSystem fs(
+      std::unique_ptr(new FakeAuthProvider),
+      std::unique_ptr(
+          new FakeHttpRequestFactory(&requests)),
+      std::unique_ptr(new FakeZoneProvider), 0 /* block size */,
+      0 /* max bytes */, 0 /* max staleness */, 0 /* stat cache max age */,
+      0 /* stat cache max entries */, 0 /* matching paths cache max age */,
+      0 /* matching paths cache max entries */, kTestRetryConfig,
+      kTestTimeoutConfig, *kAllowedLocationsDefault,
+      nullptr /* gcs additional header */, false /* compose append */);
 
   EXPECT_EQ(error::Code::NOT_FOUND, fs.IsDirectory("gs://bucket/").code());
 }
@@ -3222,16 +3224,16 @@ TEST(GcsFileSystemTest, CreateDir_Folder) {
            "Timeouts: 5 1 10\n",
            strings::StrCat("{\"size\": \"1010\",\"generation\": \"1\","
                            "\"updated\": \"2016-04-29T23:15:24.896Z\"}"))});
-  GcsFileSystem fs(std::unique_ptr(new FakeAuthProvider),
-                   std::unique_ptr(
-                       new FakeHttpRequestFactory(&requests)),
-                   std::unique_ptr(new FakeZoneProvider),
-                   0 /* block size */, 0 /* max bytes */, 0 /* max staleness */,
-                   0 /* stat cache max age */, 0 /* stat cache max entries */,
-                   0 /* matching paths cache max age */,
-                   0 /* matching paths cache max entries */, kTestRetryConfig,
-                   kTestTimeoutConfig, *kAllowedLocationsDefault,
-                   nullptr /* gcs additional header */);
+  GcsFileSystem fs(
+      std::unique_ptr(new FakeAuthProvider),
+      std::unique_ptr(
+          new FakeHttpRequestFactory(&requests)),
+      std::unique_ptr(new FakeZoneProvider), 0 /* block size */,
+      0 /* max bytes */, 0 /* max staleness */, 0 /* stat cache max age */,
+      0 /* stat cache max entries */, 0 /* matching paths cache max age */,
+      0 /* matching paths cache max entries */, kTestRetryConfig,
+      kTestTimeoutConfig, *kAllowedLocationsDefault,
+      nullptr /* gcs additional header */, false /* compose append */);
 
   TF_EXPECT_OK(fs.CreateDir("gs://bucket/subpath"));
   EXPECT_EQ(errors::AlreadyExists("gs://bucket/subpath/"),
@@ -3250,16 +3252,16 @@ TEST(GcsFileSystemTest, CreateDir_Bucket) {
            "Auth Token: fake_token\n"
            "Timeouts: 5 1 10\n",
            "")});
-  GcsFileSystem fs(std::unique_ptr(new FakeAuthProvider),
-                   std::unique_ptr(
-                       new FakeHttpRequestFactory(&requests)),
-                   std::unique_ptr(new FakeZoneProvider),
-                   0 /* block size */, 0 /* max bytes */, 0 /* max staleness */,
-                   0 /* stat cache max age */, 0 /* stat cache max entries */,
-                   0 /* matching paths cache max age */,
-                   0 /* matching paths cache max entries */, kTestRetryConfig,
-                   kTestTimeoutConfig, *kAllowedLocationsDefault,
-                   nullptr /* gcs additional header */);
+  GcsFileSystem fs(
+      std::unique_ptr(new FakeAuthProvider),
+      std::unique_ptr(
+          new FakeHttpRequestFactory(&requests)),
+      std::unique_ptr(new FakeZoneProvider), 0 /* block size */,
+      0 /* max bytes */, 0 /* max staleness */, 0 /* stat cache max age */,
+      0 /* stat cache max entries */, 0 /* matching paths cache max age */,
+      0 /* matching paths cache max entries */, kTestRetryConfig,
+      kTestTimeoutConfig, *kAllowedLocationsDefault,
+      nullptr /* gcs additional header */, false /* compose append */);
 
   TF_EXPECT_OK(fs.CreateDir("gs://bucket/"));
   TF_EXPECT_OK(fs.CreateDir("gs://bucket"));
@@ -3322,16 +3324,16 @@ TEST(GcsFileSystemTest, DeleteRecursively_Ok) {
                            "Timeouts: 5 1 10\n"
                            "Delete: yes\n",
                            "")});
-  GcsFileSystem fs(std::unique_ptr(new FakeAuthProvider),
-                   std::unique_ptr(
-                       new FakeHttpRequestFactory(&requests)),
-                   std::unique_ptr(new FakeZoneProvider),
-                   0 /* block size */, 0 /* max bytes */, 0 /* max staleness */,
-                   0 /* stat cache max age */, 0 /* stat cache max entries */,
-                   0 /* matching paths cache max age */,
-                   0 /* matching paths cache max entries */, kTestRetryConfig,
-                   kTestTimeoutConfig, *kAllowedLocationsDefault,
-                   nullptr /* gcs additional header */);
+  GcsFileSystem fs(
+      std::unique_ptr(new FakeAuthProvider),
+      std::unique_ptr(
+          new FakeHttpRequestFactory(&requests)),
+      std::unique_ptr(new FakeZoneProvider), 0 /* block size */,
+      0 /* max bytes */, 0 /* max staleness */, 0 /* stat cache max age */,
+      0 /* stat cache max entries */, 0 /* matching paths cache max age */,
+      0 /* matching paths cache max entries */, kTestRetryConfig,
+      kTestTimeoutConfig, *kAllowedLocationsDefault,
+      nullptr /* gcs additional header */, false /* compose append */);
 
   int64 undeleted_files, undeleted_dirs;
   TF_EXPECT_OK(fs.DeleteRecursively("gs://bucket/path", &undeleted_files,
@@ -3415,16 +3417,16 @@ TEST(GcsFileSystemTest, DeleteRecursively_DeletionErrors) {
            "Timeouts: 5 1 10\n",
            "", errors::NotFound("404"), 404)});
 
-  GcsFileSystem fs(std::unique_ptr(new FakeAuthProvider),
-                   std::unique_ptr(
-                       new FakeHttpRequestFactory(&requests)),
-                   std::unique_ptr(new FakeZoneProvider),
-                   0 /* block size */, 0 /* max bytes */, 0 /* max staleness */,
-                   0 /* stat cache max age */, 0 /* stat cache max entries */,
-                   0 /* matching paths cache max age */,
-                   0 /* matching paths cache max entries */, kTestRetryConfig,
-                   kTestTimeoutConfig, *kAllowedLocationsDefault,
-                   nullptr /* gcs additional header */);
+  GcsFileSystem fs(
+      std::unique_ptr(new FakeAuthProvider),
+      std::unique_ptr(
+          new FakeHttpRequestFactory(&requests)),
+      std::unique_ptr(new FakeZoneProvider), 0 /* block size */,
+      0 /* max bytes */, 0 /* max staleness */, 0 /* stat cache max age */,
+      0 /* stat cache max entries */, 0 /* matching paths cache max age */,
+      0 /* matching paths cache max entries */, kTestRetryConfig,
+      kTestTimeoutConfig, *kAllowedLocationsDefault,
+      nullptr /* gcs additional header */, false /* compose append */);
 
   int64 undeleted_files, undeleted_dirs;
   TF_EXPECT_OK(fs.DeleteRecursively("gs://bucket/path", &undeleted_files,
@@ -3450,16 +3452,16 @@ TEST(GcsFileSystemTest, DeleteRecursively_NotAFolder) {
            "Auth Token: fake_token\n"
            "Timeouts: 5 1 10\n",
            "", errors::NotFound("404"), 404)});
-  GcsFileSystem fs(std::unique_ptr(new FakeAuthProvider),
-                   std::unique_ptr(
-                       new FakeHttpRequestFactory(&requests)),
-                   std::unique_ptr(new FakeZoneProvider),
-                   0 /* block size */, 0 /* max bytes */, 0 /* max staleness */,
-                   0 /* stat cache max age */, 0 /* stat cache max entries */,
-                   0 /* matching paths cache max age */,
-                   0 /* matching paths cache max entries */, kTestRetryConfig,
-                   kTestTimeoutConfig, *kAllowedLocationsDefault,
-                   nullptr /* gcs additional header */);
+  GcsFileSystem fs(
+      std::unique_ptr(new FakeAuthProvider),
+      std::unique_ptr(
+          new FakeHttpRequestFactory(&requests)),
+      std::unique_ptr(new FakeZoneProvider), 0 /* block size */,
+      0 /* max bytes */, 0 /* max staleness */, 0 /* stat cache max age */,
+      0 /* stat cache max entries */, 0 /* matching paths cache max age */,
+      0 /* matching paths cache max entries */, kTestRetryConfig,
+      kTestTimeoutConfig, *kAllowedLocationsDefault,
+      nullptr /* gcs additional header */, false /* compose append */);
 
   int64 undeleted_files, undeleted_dirs;
   EXPECT_EQ(error::Code::NOT_FOUND,
@@ -3543,7 +3545,7 @@ TEST(GcsFileSystemTest, AdditionalRequestHeaderTest) {
       0 /* stat cache max entries */, 0 /* matching paths cache max age */,
       0 /* matching paths cache max entries */, kTestRetryConfig,
       kTestTimeoutConfig, *kAllowedLocationsDefault,
-      add_header /* gcs additional header */);
+      add_header /* gcs additional header */, false /* compose append */);
 
   std::unique_ptr request;
   TF_EXPECT_OK(fs7.CreateHttpRequest(&request));
@@ -3613,16 +3615,16 @@ TEST(GcsFileSystemTest, CreateHttpRequest) {
                            "Auth Token: fake_token\n"
                            "Header Hello: world\n",
                            "{}")});
-  GcsFileSystem fs(std::unique_ptr(new FakeAuthProvider),
-                   std::unique_ptr(
-                       new FakeHttpRequestFactory(&requests)),
-                   std::unique_ptr(new FakeZoneProvider),
-                   0 /* block size */, 0 /* max bytes */, 0 /* max staleness */,
-                   0 /* stat cache max age */, 0 /* stat cache max entries */,
-                   0 /* matching paths cache max age */,
-                   0 /* matching paths cache max entries */, kTestRetryConfig,
-                   kTestTimeoutConfig, *kAllowedLocationsDefault,
-                   nullptr /* gcs additional header */);
+  GcsFileSystem fs(
+      std::unique_ptr(new FakeAuthProvider),
+      std::unique_ptr(
+          new FakeHttpRequestFactory(&requests)),
+      std::unique_ptr(new FakeZoneProvider), 0 /* block size */,
+      0 /* max bytes */, 0 /* max staleness */, 0 /* stat cache max age */,
+      0 /* stat cache max entries */, 0 /* matching paths cache max age */,
+      0 /* matching paths cache max entries */, kTestRetryConfig,
+      kTestTimeoutConfig, *kAllowedLocationsDefault,
+      nullptr /* gcs additional header */, false /* compose append */);
 
   std::unique_ptr request;
   TF_EXPECT_OK(fs.CreateHttpRequest(&request));
@@ -3676,16 +3678,16 @@ TEST(GcsFileSystemTest, Stat_StatsRecording) {
       "Timeouts: 5 1 10\n",
       strings::StrCat("{\"size\": \"1010\",\"generation\": \"1\","
                       "\"updated\": \"2016-04-29T23:15:24.896Z\"}"))});
-  GcsFileSystem fs(std::unique_ptr(new FakeAuthProvider),
-                   std::unique_ptr(
-                       new FakeHttpRequestFactory(&requests)),
-                   std::unique_ptr(new FakeZoneProvider),
-                   0 /* block size */, 0 /* max bytes */, 0 /* max staleness */,
-                   0 /* stat cache max age */, 0 /* stat cache max entries */,
-                   0 /* matching paths cache max age */,
-                   0 /* matching paths cache max entries */, kTestRetryConfig,
-                   kTestTimeoutConfig, *kAllowedLocationsDefault,
-                   nullptr /* gcs additional header */);
+  GcsFileSystem fs(
+      std::unique_ptr(new FakeAuthProvider),
+      std::unique_ptr(
+          new FakeHttpRequestFactory(&requests)),
+      std::unique_ptr(new FakeZoneProvider), 0 /* block size */,
+      0 /* max bytes */, 0 /* max staleness */, 0 /* stat cache max age */,
+      0 /* stat cache max entries */, 0 /* matching paths cache max age */,
+      0 /* matching paths cache max entries */, kTestRetryConfig,
+      kTestTimeoutConfig, *kAllowedLocationsDefault,
+      nullptr /* gcs additional header */, false /* compose append */);
 
   TestGcsStats stats;
   fs.SetStats(&stats);
@@ -3703,16 +3705,16 @@ TEST(GcsFileSystemTest, NewRandomAccessFile_StatsRecording) {
       "Range: 0-5\n"
       "Timeouts: 5 1 20\n",
       "012345")});
-  GcsFileSystem fs(std::unique_ptr(new FakeAuthProvider),
-                   std::unique_ptr(
-                       new FakeHttpRequestFactory(&requests)),
-                   std::unique_ptr(new FakeZoneProvider),
-                   0 /* block size */, 0 /* max bytes */, 0 /* max staleness */,
-                   0 /* stat cache max age */, 0 /* stat cache max entries */,
-                   0 /* matching paths cache max age */,
-                   0 /* matching paths cache max entries */, kTestRetryConfig,
-                   kTestTimeoutConfig, *kAllowedLocationsDefault,
-                   nullptr /* gcs additional header */);
+  GcsFileSystem fs(
+      std::unique_ptr(new FakeAuthProvider),
+      std::unique_ptr(
+          new FakeHttpRequestFactory(&requests)),
+      std::unique_ptr(new FakeZoneProvider), 0 /* block size */,
+      0 /* max bytes */, 0 /* max staleness */, 0 /* stat cache max age */,
+      0 /* stat cache max entries */, 0 /* matching paths cache max age */,
+      0 /* matching paths cache max entries */, kTestRetryConfig,
+      kTestTimeoutConfig, *kAllowedLocationsDefault,
+      nullptr /* gcs additional header */, false /* compose append */);
 
   TestGcsStats stats;
   fs.SetStats(&stats);
@@ -3732,5 +3734,253 @@ TEST(GcsFileSystemTest, NewRandomAccessFile_StatsRecording) {
   EXPECT_EQ(6, stats.block_retrieved_bytes_transferred_);
 }
 
+TEST(GcsFileSystemTest, NewAppendableFile_MultipleFlushesWithCompose) {
+  std::vector contents(
+      {"content0,", "content1,", "content2,", "content3,"});
+  std::vector requests({
+      // Fetch the file (stats and then content)
+      new FakeHttpRequest(
+          "Uri: "
+          "https://www.googleapis.com/storage/v1/b/bucket/o/"
+          "some%2Fpath%2Fappendable?fields=size%2Cgeneration%2Cupdated\n"
+          "Auth Token: fake_token\n"
+          "Timeouts: 5 1 10\n",
+          strings::StrCat("{\"size\": \"8\",\"generation\": \"1\","
+                          "\"updated\": \"2016-04-29T23:15:24.896Z\"}")),
+      new FakeHttpRequest(
+          "Uri: "
+          "https://storage.googleapis.com/bucket/some%2Fpath%2Fappendable\n"
+          "Auth Token: fake_token\n"
+          "Range: 0-1048575\n"
+          "Timeouts: 5 1 20\n",
+          contents[0]),
+      // Upload entire file
+      new FakeHttpRequest(
+          "Uri: https://www.googleapis.com/upload/storage/v1/b/bucket/o?"
+          "uploadType=resumable&name=some%2Fpath%2Fappendable\n"
+          "Auth Token: fake_token\n"
+          "Header X-Upload-Content-Length: 18\n"
+          "Post: yes\n"
+          "Timeouts: 5 1 10\n",
+          "", {{"Location", "https://custom/upload/location"}}),
+      new FakeHttpRequest(
+          strings::StrCat("Uri: https://custom/upload/location\n"
+                          "Auth Token: fake_token\n"
+                          "Header Content-Range: bytes 0-17/18\n"
+                          "Timeouts: 5 1 30\n"
+                          "Put body: ",
+                          contents[0], contents[1], "\n"),
+          ""),
+      // Upload new part to a temporary object
+      new FakeHttpRequest(
+          "Uri: "
+          "https://www.googleapis.com/upload/storage/v1/b/bucket/"
+          "o?uploadType=resumable&name=some%2Fpath%2F.tmpcompose%2Fappendable."
+          "18\n"
+          "Auth Token: fake_token\n"
+          "Header X-Upload-Content-Length: 9\n"
+          "Post: yes\n"
+          "Timeouts: 5 1 10\n",
+          "",
+          {{"Location",
+            "https://custom/upload/"
+            "location"}}),
+      new FakeHttpRequest(
+          strings::StrCat("Uri: https://custom/upload/location\n"
+                          "Auth Token: fake_token\n"
+                          "Header Content-Range: bytes 0-8/9\n"
+                          "Timeouts: 5 1 30\n"
+                          "Put body: ",
+                          contents[2], "\n"),
+          ""),
+      // Compose the new part at the end of the original object.
+      new FakeHttpRequest("Uri: "
+                          "https://www.googleapis.com/storage/v1/b/bucket/o/"
+                          "some%2Fpath%2Fappendable/compose\n"
+                          "Auth Token: fake_token\n"
+                          "Timeouts: 5 1 10\n"
+                          "Header content-type: application/json\n"
+                          "Post body: {'sourceObjects': [{'name': "
+                          "'some/path/appendable'},{'name': "
+                          "'some/path/.tmpcompose/appendable.18'}]}\n",
+                          ""),
+      // Delete the temporary object.
+      new FakeHttpRequest("Uri: "
+                          "https://www.googleapis.com/storage/v1/b/bucket/o/"
+                          "some%2Fpath%2F.tmpcompose%2Fappendable.18\n"
+                          "Auth Token: fake_token\n"
+                          "Timeouts: 5 1 10\n"
+                          "Delete: yes\n",
+                          ""),
+      new FakeHttpRequest(
+          "Uri: https://www.googleapis.com/upload/storage/v1/b/bucket/o?"
+          "uploadType=resumable&name=some%2Fpath%2F.tmpcompose%2Fappendable."
+          "27\n"
+          "Auth Token: fake_token\n"
+          "Header X-Upload-Content-Length: 9\n"
+          "Post: yes\n"
+          "Timeouts: 5 1 10\n",
+          "", {{"Location", "https://custom/upload/location"}}),
+      new FakeHttpRequest(
+          strings::StrCat("Uri: https://custom/upload/location\n"
+                          "Auth Token: fake_token\n"
+                          "Header Content-Range: bytes 0-8/9\n"
+                          "Timeouts: 5 1 30\n"
+                          "Put body: ",
+                          contents[3], "\n"),
+          ""),
+      new FakeHttpRequest("Uri: "
+                          "https://www.googleapis.com/storage/v1/b/bucket/o/"
+                          "some%2Fpath%2Fappendable/compose\n"
+                          "Auth Token: fake_token\n"
+                          "Timeouts: 5 1 10\n"
+                          "Header content-type: application/json\n"
+                          "Post body: {'sourceObjects': [{'name': "
+                          "'some/path/appendable'},{'name': "
+                          "'some/path/.tmpcompose/appendable.27'}]}\n",
+                          ""),
+      new FakeHttpRequest("Uri: "
+                          "https://www.googleapis.com/storage/v1/b/bucket/o/"
+                          "some%2Fpath%2F.tmpcompose%2Fappendable."
+                          "27\n"
+                          "Auth Token: fake_token\n"
+                          "Timeouts: 5 1 10\n"
+                          "Delete: yes\n",
+                          ""),
+  });
+  GcsFileSystem fs(
+      std::unique_ptr(new FakeAuthProvider),
+      std::unique_ptr(
+          new FakeHttpRequestFactory(&requests)),
+      std::unique_ptr(new FakeZoneProvider), 32 /* block size */,
+      32 /* max bytes */, 0 /* max staleness */, 3600 /* stat cache max age */,
+      0 /* stat cache max entries */, 0 /* matching paths cache max age */,
+      0 /* matching paths cache max entries */, kTestRetryConfig,
+      kTestTimeoutConfig, *kAllowedLocationsDefault,
+      nullptr /* gcs additional header */, true /* compose append */);
+
+  // Create an appendable file. This should read the file from GCS, and pull its
+  // contents into the block cache.
+  std::unique_ptr wfile;
+  TF_EXPECT_OK(
+      fs.NewAppendableFile("gs://bucket/some/path/appendable", &wfile));
+  TF_EXPECT_OK(wfile->Append(contents[1]));
+  TF_EXPECT_OK(wfile->Flush());
+  TF_EXPECT_OK(wfile->Append(contents[2]));
+  TF_EXPECT_OK(wfile->Flush());
+  TF_EXPECT_OK(wfile->Append(contents[3]));
+  TF_EXPECT_OK(wfile->Close());
+}
+
+TEST(GcsFileSystemTest, NewAppendableFile_MultipleFlushesWithoutCompose) {
+  std::vector contents(
+      {"content0,", "content1,", "content2,", "content3,"});
+  std::vector requests({
+      new FakeHttpRequest(
+          "Uri: https://www.googleapis.com/storage/v1/b/bucket/o/"
+          "path%2Fappendable?fields=size%2Cgeneration%2Cupdated\n"
+          "Auth Token: fake_token\n"
+          "Timeouts: 5 1 10\n",
+          strings::StrCat("{\"size\": \"8\",\"generation\": \"1\","
+                          "\"updated\": \"2016-04-29T23:15:24.896Z\"}")),
+      new FakeHttpRequest(
+          "Uri: https://storage.googleapis.com/bucket/path%2Fappendable\n"
+          "Auth Token: fake_token\n"
+          "Range: 0-1048575\n"
+          "Timeouts: 5 1 20\n",
+          contents[0]),
+      new FakeHttpRequest(
+          "Uri: https://www.googleapis.com/upload/storage/v1/b/bucket/o?"
+          "uploadType=resumable&name=path%2Fappendable\n"
+          "Auth Token: fake_token\n"
+          "Header X-Upload-Content-Length: 18\n"
+          "Post: yes\n"
+          "Timeouts: 5 1 10\n",
+          "", {{"Location", "https://custom/upload/location"}}),
+      // Uploads entire file.
+      new FakeHttpRequest(
+          strings::StrCat("Uri: https://custom/upload/location\n"
+                          "Auth Token: fake_token\n"
+                          "Header Content-Range: bytes 0-17/18\n"
+                          "Timeouts: 5 1 30\n"
+                          "Put body: ",
+                          contents[0], contents[1], "\n"),
+          ""),
+      new FakeHttpRequest("Uri: "
+                          "https://www.googleapis.com/upload/storage/v1/b/"
+                          "bucket/o?"
+                          "uploadType=resumable&name=path%2Fappendable\n"
+                          "Auth Token: fake_token\n"
+                          "Header X-Upload-Content-Length: 27\n"
+                          "Post: yes\n"
+                          "Timeouts: 5 1 10\n",
+                          "",
+                          {{"Location",
+                            "https://custom/upload/"
+                            "location"}}),
+      // Uploads entire file again.
+      new FakeHttpRequest(
+          strings::StrCat("Uri: https://custom/upload/location\n"
+                          "Auth Token: fake_token\n"
+                          "Header Content-Range: bytes 0-26/27\n"
+                          "Timeouts: 5 1 30\n"
+                          "Put body: ",
+                          contents[0], contents[1], contents[2], "\n"),
+          ""),
+      new FakeHttpRequest(
+          "Uri: https://www.googleapis.com/upload/storage/v1/b/bucket/o?"
+          "uploadType=resumable&name=path%2Fappendable\n"
+          "Auth Token: fake_token\n"
+          "Header X-Upload-Content-Length: 36\n"
+          "Post: yes\n"
+          "Timeouts: 5 1 10\n",
+          "", {{"Location", "https://custom/upload/location"}}),
+      // Uploads entire file again.
+      new FakeHttpRequest(
+          strings::StrCat("Uri: https://custom/upload/location\n"
+                          "Auth Token: fake_token\n"
+                          "Header Content-Range: bytes 0-35/36\n"
+                          "Timeouts: 5 1 30\n"
+                          "Put body: ",
+                          contents[0], contents[1], contents[2], contents[3],
+                          "\n"),
+          ""),
+  });
+  GcsFileSystem fs(
+      std::unique_ptr(new FakeAuthProvider),
+      std::unique_ptr(
+          new FakeHttpRequestFactory(&requests)),
+      std::unique_ptr(new FakeZoneProvider), 32 /* block size */,
+      32 /* max bytes */, 0 /* max staleness */, 3600 /* stat cache max age */,
+      0 /* stat cache max entries */, 0 /* matching paths cache max age */,
+      0 /* matching paths cache max entries */, kTestRetryConfig,
+      kTestTimeoutConfig, *kAllowedLocationsDefault,
+      nullptr /* gcs additional header */, false /* compose append */);
+
+  // Create an appendable file. This should read the file from GCS, and pull its
+  // contents into the block cache.
+  std::unique_ptr wfile;
+  TF_EXPECT_OK(fs.NewAppendableFile("gs://bucket/path/appendable", &wfile));
+  TF_EXPECT_OK(wfile->Append(contents[1]));
+  TF_EXPECT_OK(wfile->Flush());
+  TF_EXPECT_OK(wfile->Append(contents[2]));
+  TF_EXPECT_OK(wfile->Flush());
+  TF_EXPECT_OK(wfile->Append(contents[3]));
+  TF_EXPECT_OK(wfile->Close());
+}
+
+TEST(GcsFileSystemTest, AppendModeCompose) {
+  unsetenv("GCS_APPEND_MODE");
+  setenv("GCS_APPEND_MODE", "compose", 1);
+  GcsFileSystem fs1;
+  EXPECT_EQ(true, fs1.compose_append());
+}
+
+TEST(GcsFileSystemTest, AppendModeDefault) {
+  unsetenv("GCS_APPEND_MODE");
+  GcsFileSystem fs1;
+  EXPECT_EQ(false, fs1.compose_append());
+}
+
 }  // namespace
 }  // namespace tensorflow

From 030e65acd56b55084a5c8a9f7e5ed3db3dc63093 Mon Sep 17 00:00:00 2001
From: Robert David 
Date: Fri, 19 Jun 2020 14:17:36 -0700
Subject: [PATCH 336/867] Move GetTensorData calls directly to LstmStep call in
 the int8x8_8 version.

PiperOrigin-RevId: 317378494
Change-Id: I880de60ee5c963468b2eb47d8cca7a3b9765f57c
---
 tensorflow/lite/kernels/lstm_eval.cc | 102 ++++++++-------------------
 1 file changed, 31 insertions(+), 71 deletions(-)

diff --git a/tensorflow/lite/kernels/lstm_eval.cc b/tensorflow/lite/kernels/lstm_eval.cc
index f38fdc95f3e..b2f3d77912b 100644
--- a/tensorflow/lite/kernels/lstm_eval.cc
+++ b/tensorflow/lite/kernels/lstm_eval.cc
@@ -2091,50 +2091,6 @@ TfLiteStatus EvalInteger8x8_8(
   const int n_cell = input_to_output_weights->dims->data[0];
   const int n_output = recurrent_to_output_weights->dims->data[1];
 
-  // Weights and states.
-  const int8_t* input_to_input_weight_ptr =
-      GetTensorData(input_to_input_weights);
-  const int8_t* recurrent_to_input_weight_ptr =
-      GetTensorData(recurrent_to_input_weights);
-  const int8_t* cell_to_input_weight_ptr =
-      GetTensorData(cell_to_input_weights);
-  const int8_t* input_to_forget_weight_ptr =
-      GetTensorData(input_to_forget_weights);
-  const int8_t* recurrent_to_forget_weight_ptr =
-      GetTensorData(recurrent_to_forget_weights);
-  const int8_t* cell_to_forget_weight_ptr =
-      GetTensorData(cell_to_forget_weights);
-  const int8_t* input_to_cell_weight_ptr =
-      GetTensorData(input_to_cell_weights);
-  const int8_t* recurrent_to_cell_weight_ptr =
-      GetTensorData(recurrent_to_cell_weights);
-  const int8_t* input_to_output_weight_ptr =
-      GetTensorData(input_to_output_weights);
-  const int8_t* recurrent_to_output_weight_ptr =
-      GetTensorData(recurrent_to_output_weights);
-  const int8_t* cell_to_output_weight_ptr =
-      GetTensorData(cell_to_output_weights);
-  const int8_t* projection_weight_ptr =
-      GetTensorData(projection_weights);
-  const int16_t* layer_norm_input_weight_ptr =
-      GetTensorData(input_layer_norm_coefficients);
-  const int16_t* layer_norm_forget_weight_ptr =
-      GetTensorData(forget_layer_norm_coefficients);
-  const int16_t* layer_norm_cell_weight_ptr =
-      GetTensorData(cell_layer_norm_coefficients);
-  const int16_t* layer_norm_output_weight_ptr =
-      GetTensorData(output_layer_norm_coefficients);
-  const int32_t* input_gate_bias_ptr = GetTensorData(input_gate_bias);
-  const int32_t* forget_gate_bias_ptr =
-      GetTensorData(forget_gate_bias);
-  const int32_t* cell_gate_bias_ptr = GetTensorData(cell_gate_bias);
-  const int32_t* output_gate_bias_ptr =
-      GetTensorData(output_gate_bias);
-  const int32_t* projection_bias_ptr = GetTensorData(projection_bias);
-  int16_t* cell_ptr = GetTensorData(cell_state);
-  int8_t* output_state_ptr = GetTensorData(output_state);
-  int8_t* output_ptr = nullptr;
-
   const int32_t input_zp = input->params.zero_point;
   const int32_t output_state_zp = output_state->params.zero_point;
 
@@ -2146,89 +2102,93 @@ TfLiteStatus EvalInteger8x8_8(
 
   for (int t = 0; t < max_time; t++) {
     const int t_rel = t;
-    output_ptr = output->data.int8 + t_rel * output_step;
-
+    int8_t* output_ptr = GetTensorData(output) + t_rel * output_step;
     // Input can be int8 asymmetric or int16 symmetric.
-    const int8_t* input_ptr = input->data.int8 + t_rel * input_step;
+    const int8_t* input_ptr = GetTensorData(input) + t_rel * input_step;
     lstm_eval::LstmStepInteger(
         input_ptr, input_zp,
 
-        input_to_input_weight_ptr,
+        GetTensorData(input_to_input_weights),
         integer_lstm_param->effective_input_to_input_scale_a,
         integer_lstm_param->effective_input_to_input_scale_b,
 
-        input_to_forget_weight_ptr,
+        GetTensorData(input_to_forget_weights),
         integer_lstm_param->effective_input_to_forget_scale_a,
         integer_lstm_param->effective_input_to_forget_scale_b,
 
-        input_to_cell_weight_ptr,
+        GetTensorData(input_to_cell_weights),
         integer_lstm_param->effective_input_to_cell_scale_a,
         integer_lstm_param->effective_input_to_cell_scale_b,
 
-        input_to_output_weight_ptr,
+        GetTensorData(input_to_output_weights),
         integer_lstm_param->effective_input_to_output_scale_a,
         integer_lstm_param->effective_input_to_output_scale_b,
 
-        recurrent_to_input_weight_ptr,
+        GetTensorData(recurrent_to_input_weights),
         integer_lstm_param->effective_recurrent_to_input_scale_a,
         integer_lstm_param->effective_recurrent_to_input_scale_b,
 
-        recurrent_to_forget_weight_ptr,
+        GetTensorData(recurrent_to_forget_weights),
         integer_lstm_param->effective_recurrent_to_forget_scale_a,
         integer_lstm_param->effective_recurrent_to_forget_scale_b,
 
-        recurrent_to_cell_weight_ptr,
+        GetTensorData(recurrent_to_cell_weights),
         integer_lstm_param->effective_recurrent_to_cell_scale_a,
         integer_lstm_param->effective_recurrent_to_cell_scale_b,
 
-        recurrent_to_output_weight_ptr,
+        GetTensorData(recurrent_to_output_weights),
         integer_lstm_param->effective_recurrent_to_output_scale_a,
         integer_lstm_param->effective_recurrent_to_output_scale_b,
 
-        cell_to_input_weight_ptr,
+        GetTensorData(cell_to_input_weights),
         integer_lstm_param->effective_cell_to_input_scale_a,
         integer_lstm_param->effective_cell_to_input_scale_b,
 
-        cell_to_forget_weight_ptr,
+        GetTensorData(cell_to_forget_weights),
         integer_lstm_param->effective_cell_to_forget_scale_a,
         integer_lstm_param->effective_cell_to_forget_scale_b,
 
-        cell_to_output_weight_ptr,
+        GetTensorData(cell_to_output_weights),
         integer_lstm_param->effective_cell_to_output_scale_a,
         integer_lstm_param->effective_cell_to_output_scale_b,
 
-        projection_weight_ptr, integer_lstm_param->effective_proj_scale_a,
+        GetTensorData(projection_weights),
+        integer_lstm_param->effective_proj_scale_a,
         integer_lstm_param->effective_proj_scale_b,
 
-        layer_norm_input_weight_ptr,
+        GetTensorData(input_layer_norm_coefficients),
         integer_lstm_param->layer_norm_input_scale_a,
         integer_lstm_param->layer_norm_input_scale_b,
 
-        layer_norm_forget_weight_ptr,
+        GetTensorData(forget_layer_norm_coefficients),
         integer_lstm_param->layer_norm_forget_scale_a,
         integer_lstm_param->layer_norm_forget_scale_b,
 
-        layer_norm_cell_weight_ptr, integer_lstm_param->layer_norm_cell_scale_a,
+        GetTensorData(cell_layer_norm_coefficients),
+        integer_lstm_param->layer_norm_cell_scale_a,
         integer_lstm_param->layer_norm_cell_scale_b,
 
-        layer_norm_output_weight_ptr,
+        GetTensorData(output_layer_norm_coefficients),
         integer_lstm_param->layer_norm_output_scale_a,
         integer_lstm_param->layer_norm_output_scale_b,
 
-        input_gate_bias_ptr, forget_gate_bias_ptr, cell_gate_bias_ptr,
-        output_gate_bias_ptr, projection_bias_ptr,
+        GetTensorData(input_gate_bias),
+        GetTensorData(forget_gate_bias),
+        GetTensorData(cell_gate_bias),
+        GetTensorData(output_gate_bias),
+        GetTensorData(projection_bias),
 
         params, integer_lstm_param->intermediate_scale_a,
         integer_lstm_param->intermediate_scale_b,
         integer_lstm_param->intermediate_zp,
         integer_lstm_param->quantized_cell_clip,
         integer_lstm_param->quantized_proj_clip, n_batch, n_cell, n_input,
-        n_output, output_batch_leading_dim, output_state_ptr, output_state_zp,
-        cell_ptr, output_ptr, GetTensorData(scratch0),
-        GetTensorData(scratch1), GetTensorData(scratch2),
-        GetTensorData(scratch3), GetTensorData(scratch4),
-        GetTensorData(scratch5), GetTensorData(scratch6),
-        GetTensorData(scratch7));
+        n_output, output_batch_leading_dim, GetTensorData(output_state),
+        output_state_zp, GetTensorData(cell_state), output_ptr,
+        GetTensorData(scratch0), GetTensorData(scratch1),
+        GetTensorData(scratch2), GetTensorData(scratch3),
+        GetTensorData(scratch4), GetTensorData(scratch5),
+        GetTensorData(scratch6), GetTensorData(scratch7));
   }
 
   return kTfLiteOk;

From 9e7df609bca2e363aa3217b70c6f95578ba07f10 Mon Sep 17 00:00:00 2001
From: "A. Unique TensorFlower" 
Date: Fri, 19 Jun 2020 14:21:31 -0700
Subject: [PATCH 337/867] Fix tokenization tests and update testing_utils to
 transfer state between layer creation.

PiperOrigin-RevId: 317379253
Change-Id: I786c2eb0506239de0e7f1a5f314a8f1b0bda10d4
---
 tensorflow/python/keras/testing_utils.py | 18 +++++++++++++++---
 1 file changed, 15 insertions(+), 3 deletions(-)

diff --git a/tensorflow/python/keras/testing_utils.py b/tensorflow/python/keras/testing_utils.py
index 1928588fea1..cceaabe37a5 100644
--- a/tensorflow/python/keras/testing_utils.py
+++ b/tensorflow/python/keras/testing_utils.py
@@ -94,7 +94,8 @@ def layer_test(layer_cls,
                expected_output_shape=None,
                validate_training=True,
                adapt_data=None,
-               custom_objects=None):
+               custom_objects=None,
+               test_harness=None):
   """Test routine for a layer with a single input and single output.
 
   Arguments:
@@ -114,6 +115,8 @@ def layer_test(layer_cls,
       be tested for this layer. This is only relevant for PreprocessingLayers.
     custom_objects: Optional dictionary mapping name strings to custom objects
       in the layer class. This is helpful for testing custom layers.
+    test_harness: The Tensorflow test, if any, that this function is being
+      called in.
 
   Returns:
     The output data (Numpy array) returned by the layer, for additional
@@ -143,9 +146,15 @@ def layer_test(layer_cls,
     expected_output_dtype = input_dtype
 
   if dtypes.as_dtype(expected_output_dtype) == dtypes.string:
-    assert_equal = string_test
+    if test_harness:
+      assert_equal = test_harness.assertAllEqual
+    else:
+      assert_equal = string_test
   else:
-    assert_equal = numeric_test
+    if test_harness:
+      assert_equal = test_harness.assertAllClose
+    else:
+      assert_equal = numeric_test
 
   # instantiation
   kwargs = kwargs or {}
@@ -228,6 +237,7 @@ def layer_test(layer_cls,
   # test training mode (e.g. useful for dropout tests)
   # Rebuild the model to avoid the graph being reused between predict() and
   # See b/120160788 for more details. This should be mitigated after 2.0.
+  layer_weights = layer.get_weights()  # Get the layer weights BEFORE training.
   if validate_training:
     model = models.Model(x, layer(x))
     if _thread_local_data.run_eagerly is not None:
@@ -252,6 +262,8 @@ def layer_test(layer_cls,
   model = models.Sequential()
   model.add(layers.Input(shape=input_shape[1:], dtype=input_dtype))
   model.add(layer)
+
+  layer.set_weights(layer_weights)
   actual_output = model.predict(input_data)
   actual_output_shape = actual_output.shape
   for expected_dim, actual_dim in zip(computed_output_shape,

From f129485019052480c99442f774fb7e4a59ae5227 Mon Sep 17 00:00:00 2001
From: Bixia Zheng 
Date: Fri, 19 Jun 2020 14:59:47 -0700
Subject: [PATCH 338/867] [TF:TRT] Cosmetic fix.

Rewrite two lines of #if into #if GOOGLE_CUDA && GOOGLE_TENSORRT.

PiperOrigin-RevId: 317386436
Change-Id: Icc8ae27a17900b6f0a198d32c6d73345084eab50
---
 tensorflow/compiler/tf2tensorrt/common/utils.h       |  6 ++----
 .../compiler/tf2tensorrt/convert/convert_graph.cc    |  6 ++----
 .../compiler/tf2tensorrt/convert/convert_graph.h     |  6 ++----
 .../tf2tensorrt/convert/convert_graph_test.cc        |  6 ++----
 .../compiler/tf2tensorrt/convert/convert_nodes.cc    |  6 ++----
 .../compiler/tf2tensorrt/convert/convert_nodes.h     |  6 ++----
 .../tf2tensorrt/convert/convert_nodes_test.cc        |  6 ++----
 .../compiler/tf2tensorrt/convert/logger_registry.cc  |  6 ++----
 .../compiler/tf2tensorrt/convert/logger_registry.h   |  5 +++--
 .../tf2tensorrt/convert/trt_optimization_pass.cc     |  6 ++----
 .../tf2tensorrt/convert/trt_optimization_pass.h      |  6 ++----
 .../tf2tensorrt/kernels/get_calibration_data_op.cc   |  6 ++----
 .../compiler/tf2tensorrt/kernels/trt_engine_op.cc    |  6 ++----
 .../tf2tensorrt/kernels/trt_engine_op_test.cc        |  6 ++----
 .../tf2tensorrt/kernels/trt_engine_resource_ops.cc   |  6 ++----
 .../kernels/trt_engine_resource_ops_test.cc          |  6 ++----
 .../tf2tensorrt/ops/get_calibration_data_op.cc       |  6 ++----
 tensorflow/compiler/tf2tensorrt/ops/trt_engine_op.cc |  6 ++----
 .../tf2tensorrt/ops/trt_engine_resource_ops.cc       |  6 ++----
 .../compiler/tf2tensorrt/plugin/plugin_cast.cu.cc    |  6 ++----
 tensorflow/compiler/tf2tensorrt/plugin/trt_plugin.cc |  6 ++----
 tensorflow/compiler/tf2tensorrt/plugin/trt_plugin.h  |  6 ++----
 tensorflow/compiler/tf2tensorrt/segment/segment.cc   |  6 ++----
 tensorflow/compiler/tf2tensorrt/segment/segment.h    |  6 ++----
 .../compiler/tf2tensorrt/segment/segment_test.cc     |  6 ++----
 tensorflow/compiler/tf2tensorrt/segment/union_find.h |  6 ++----
 tensorflow/compiler/tf2tensorrt/tensorrt_test.cc     |  6 ++----
 .../compiler/tf2tensorrt/utils/trt_allocator.cc      | 12 ++++--------
 .../compiler/tf2tensorrt/utils/trt_allocator.h       | 12 ++++--------
 .../compiler/tf2tensorrt/utils/trt_engine_utils.cc   |  6 ++----
 .../compiler/tf2tensorrt/utils/trt_engine_utils.h    |  6 ++----
 .../tf2tensorrt/utils/trt_int8_calibrator.cc         |  6 ++----
 .../compiler/tf2tensorrt/utils/trt_int8_calibrator.h |  6 ++----
 tensorflow/compiler/tf2tensorrt/utils/trt_logger.cc  |  6 ++----
 tensorflow/compiler/tf2tensorrt/utils/trt_logger.h   |  6 ++----
 .../compiler/tf2tensorrt/utils/trt_lru_cache.cc      |  6 ++----
 .../compiler/tf2tensorrt/utils/trt_lru_cache.h       |  6 ++----
 .../utils/trt_shape_optimization_profiles.h          |  6 ++----
 .../utils/trt_shape_optimization_profiles_test.cc    |  6 ++----
 39 files changed, 83 insertions(+), 162 deletions(-)

diff --git a/tensorflow/compiler/tf2tensorrt/common/utils.h b/tensorflow/compiler/tf2tensorrt/common/utils.h
index 9ab0145e1ec..b428733ecd4 100644
--- a/tensorflow/compiler/tf2tensorrt/common/utils.h
+++ b/tensorflow/compiler/tf2tensorrt/common/utils.h
@@ -16,8 +16,7 @@ limitations under the License.
 #ifndef TENSORFLOW_COMPILER_TF2TENSORRT_COMMON_UTILS_H_
 #define TENSORFLOW_COMPILER_TF2TENSORRT_COMMON_UTILS_H_
 
-#if GOOGLE_CUDA
-#if GOOGLE_TENSORRT
+#if GOOGLE_CUDA && GOOGLE_TENSORRT
 
 #include "tensorflow/core/platform/logging.h"
 
@@ -29,7 +28,6 @@ namespace tensorrt {
 }  // namespace tensorrt
 }  // namespace tensorflow
 
-#endif
-#endif
+#endif  // GOOGLE_CUDA && GOOGLE_TENSORRT
 
 #endif  // TENSORFLOW_COMPILER_TF2TENSORRT_COMMON_UTILS_H_
diff --git a/tensorflow/compiler/tf2tensorrt/convert/convert_graph.cc b/tensorflow/compiler/tf2tensorrt/convert/convert_graph.cc
index 1c51d51f1c9..5429aaf3362 100644
--- a/tensorflow/compiler/tf2tensorrt/convert/convert_graph.cc
+++ b/tensorflow/compiler/tf2tensorrt/convert/convert_graph.cc
@@ -53,8 +53,7 @@ limitations under the License.
 #include "tensorflow/core/util/device_name_utils.h"
 #include "tensorflow/tools/graph_transforms/transform_utils.h"
 
-#if GOOGLE_CUDA
-#if GOOGLE_TENSORRT
+#if GOOGLE_CUDA && GOOGLE_TENSORRT
 #include "third_party/gpus/cuda/include/cuda_runtime_api.h"
 #include "third_party/tensorrt/NvInfer.h"
 namespace tensorflow {
@@ -884,5 +883,4 @@ Status ConvertAfterShapes(const ConversionParams& params) {
 }  // namespace tensorrt
 }  // namespace tensorflow
 
-#endif  // GOOGLE_TENSORRT
-#endif  // GOOGLE_CUDA
+#endif  // GOOGLE_CUDA && GOOGLE_TENSORRT
diff --git a/tensorflow/compiler/tf2tensorrt/convert/convert_graph.h b/tensorflow/compiler/tf2tensorrt/convert/convert_graph.h
index 53ab84a6fa9..d3897e864fa 100644
--- a/tensorflow/compiler/tf2tensorrt/convert/convert_graph.h
+++ b/tensorflow/compiler/tf2tensorrt/convert/convert_graph.h
@@ -24,8 +24,7 @@ limitations under the License.
 #include "tensorflow/core/lib/core/status.h"
 #include "tensorflow/core/platform/types.h"
 
-#if GOOGLE_CUDA
-#if GOOGLE_TENSORRT
+#if GOOGLE_CUDA && GOOGLE_TENSORRT
 
 namespace tensorflow {
 namespace tensorrt {
@@ -66,7 +65,6 @@ Status RegisterGraphToFunctionLibrary(const GraphDef& segment_graph_def,
 }  // namespace tensorrt
 }  // namespace tensorflow
 
-#endif  // GOOGLE_TENSORRT
-#endif  // GOOGLE_CUDA
+#endif  // GOOGLE_CUDA && GOOGLE_TENSORRT
 
 #endif  // TENSORFLOW_COMPILER_TF2TENSORRT_CONVERT_CONVERT_GRAPH_H_
diff --git a/tensorflow/compiler/tf2tensorrt/convert/convert_graph_test.cc b/tensorflow/compiler/tf2tensorrt/convert/convert_graph_test.cc
index a1f523d6bfa..54fb1d56441 100644
--- a/tensorflow/compiler/tf2tensorrt/convert/convert_graph_test.cc
+++ b/tensorflow/compiler/tf2tensorrt/convert/convert_graph_test.cc
@@ -34,8 +34,7 @@ limitations under the License.
 #include "tensorflow/core/protobuf/config.pb.h"  // NOLINT
 #include "tensorflow/core/public/session.h"
 
-#if GOOGLE_CUDA
-#if GOOGLE_TENSORRT
+#if GOOGLE_CUDA && GOOGLE_TENSORRT
 
 namespace tensorflow {
 namespace tensorrt {
@@ -231,5 +230,4 @@ TEST_F(ConvertAfterShapesTest, DirectlyConnectedEngines) {
 }  // namespace tensorrt
 }  // namespace tensorflow
 
-#endif  // GOOGLE_TENSORRT
-#endif  // GOOGLE_CUDA
+#endif  // GOOGLE_CUDA && GOOGLE_TENSORRT
diff --git a/tensorflow/compiler/tf2tensorrt/convert/convert_nodes.cc b/tensorflow/compiler/tf2tensorrt/convert/convert_nodes.cc
index 96cec556942..2ec616ba621 100644
--- a/tensorflow/compiler/tf2tensorrt/convert/convert_nodes.cc
+++ b/tensorflow/compiler/tf2tensorrt/convert/convert_nodes.cc
@@ -59,8 +59,7 @@ limitations under the License.
 #include "tensorflow/core/util/env_var.h"
 #include "tensorflow/core/util/strided_slice_op.h"
 
-#if GOOGLE_CUDA
-#if GOOGLE_TENSORRT
+#if GOOGLE_CUDA && GOOGLE_TENSORRT
 #include "third_party/tensorrt/NvInfer.h"
 #include "third_party/tensorrt/NvInferPlugin.h"
 
@@ -6258,5 +6257,4 @@ bool OutputEdgeValidator::operator()(const Edge* out_edge) const {
 }  // namespace tensorrt
 }  // namespace tensorflow
 
-#endif  // GOOGLE_TENSORRT
-#endif  // GOOGLE_CUDA
+#endif  // GOOGLE_CUDA && GOOGLE_TENSORRT
diff --git a/tensorflow/compiler/tf2tensorrt/convert/convert_nodes.h b/tensorflow/compiler/tf2tensorrt/convert/convert_nodes.h
index 7a1276c645c..a621735fad1 100644
--- a/tensorflow/compiler/tf2tensorrt/convert/convert_nodes.h
+++ b/tensorflow/compiler/tf2tensorrt/convert/convert_nodes.h
@@ -33,8 +33,7 @@ limitations under the License.
 #include "tensorflow/core/lib/core/status.h"
 #include "tensorflow/stream_executor/lib/statusor.h"
 
-#if GOOGLE_CUDA
-#if GOOGLE_TENSORRT
+#if GOOGLE_CUDA && GOOGLE_TENSORRT
 #include "third_party/tensorrt/NvInfer.h"
 
 namespace tensorflow {
@@ -694,7 +693,6 @@ BinaryOperationMap();
 }  // namespace tensorrt
 }  // namespace tensorflow
 
-#endif  // GOOGLE_TENSORRT
-#endif  // GOOGLE_CUDA
+#endif  // GOOGLE_CUDA && GOOGLE_TENSORRT
 
 #endif  // TENSORFLOW_COMPILER_TF2TENSORRT_CONVERT_CONVERT_NODES_H_
diff --git a/tensorflow/compiler/tf2tensorrt/convert/convert_nodes_test.cc b/tensorflow/compiler/tf2tensorrt/convert/convert_nodes_test.cc
index c24b169f651..53ec9ee7ada 100644
--- a/tensorflow/compiler/tf2tensorrt/convert/convert_nodes_test.cc
+++ b/tensorflow/compiler/tf2tensorrt/convert/convert_nodes_test.cc
@@ -21,8 +21,7 @@ limitations under the License.
 #include 
 #include 
 
-#if GOOGLE_CUDA
-#if GOOGLE_TENSORRT
+#if GOOGLE_CUDA && GOOGLE_TENSORRT
 
 #include 
 #include 
@@ -6636,5 +6635,4 @@ TEST_F(OpConverterTest, ConvertPad) {
 }  // namespace tensorrt
 }  // namespace tensorflow
 
-#endif  // GOOGLE_TENSORRT
-#endif  // GOOGLE_CUDA
+#endif  // GOOGLE_CUDA && GOOGLE_TENSORRT
diff --git a/tensorflow/compiler/tf2tensorrt/convert/logger_registry.cc b/tensorflow/compiler/tf2tensorrt/convert/logger_registry.cc
index 82e68cbb28d..07c9c2f1ea0 100644
--- a/tensorflow/compiler/tf2tensorrt/convert/logger_registry.cc
+++ b/tensorflow/compiler/tf2tensorrt/convert/logger_registry.cc
@@ -12,8 +12,7 @@ WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
 See the License for the specific language governing permissions and
 limitations under the License.
 ==============================================================================*/
-#if GOOGLE_CUDA
-#if GOOGLE_TENSORRT
+#if GOOGLE_CUDA && GOOGLE_TENSORRT
 
 #include "tensorflow/compiler/tf2tensorrt/convert/logger_registry.h"
 
@@ -58,5 +57,4 @@ LoggerRegistry* GetLoggerRegistry() {
 }  // namespace tensorrt
 }  // namespace tensorflow
 
-#endif  // GOOGLE_TENSORRT
-#endif  // GOOGLE_CUDA
+#endif  // GOOGLE_CUDA && GOOGLE_TENSORRT
diff --git a/tensorflow/compiler/tf2tensorrt/convert/logger_registry.h b/tensorflow/compiler/tf2tensorrt/convert/logger_registry.h
index 45b302742d0..2a265cf7caa 100644
--- a/tensorflow/compiler/tf2tensorrt/convert/logger_registry.h
+++ b/tensorflow/compiler/tf2tensorrt/convert/logger_registry.h
@@ -19,7 +19,8 @@ limitations under the License.
 #include "tensorflow/core/platform/macros.h"
 #include "tensorflow/core/platform/types.h"
 
-#if GOOGLE_CUDA
+#if GOOGLE_CUDA && GOOGLE_TENSORRT
+
 #include "third_party/tensorrt/NvInfer.h"
 
 namespace tensorflow {
@@ -53,5 +54,5 @@ class RegisterLogger {
 }  // namespace tensorrt
 }  // namespace tensorflow
 
-#endif  // GOOGLE_CUDA
+#endif  // GOOGLE_CUDA && GOOGLE_TENSORRT
 #endif  // TENSORFLOW_COMPILER_TF2TENSORRT_CONVERT_LOGGER_REGISTRY_H_
diff --git a/tensorflow/compiler/tf2tensorrt/convert/trt_optimization_pass.cc b/tensorflow/compiler/tf2tensorrt/convert/trt_optimization_pass.cc
index 72f4fe5ef9b..1cf98d135cb 100644
--- a/tensorflow/compiler/tf2tensorrt/convert/trt_optimization_pass.cc
+++ b/tensorflow/compiler/tf2tensorrt/convert/trt_optimization_pass.cc
@@ -28,8 +28,7 @@ limitations under the License.
 #include "tensorflow/core/platform/logging.h"
 #include "tensorflow/core/platform/stacktrace.h"
 
-#if GOOGLE_CUDA
-#if GOOGLE_TENSORRT
+#if GOOGLE_CUDA && GOOGLE_TENSORRT
 namespace tensorflow {
 namespace tensorrt {
 namespace convert {
@@ -302,5 +301,4 @@ static VerboseCustomGraphOptimizerRegistrar TRTOptimizationPass_Registrar(
 }  // namespace tensorrt
 }  // namespace tensorflow
 
-#endif
-#endif
+#endif  // GOOGLE_CUDA && GOOGLE_TENSORRT
diff --git a/tensorflow/compiler/tf2tensorrt/convert/trt_optimization_pass.h b/tensorflow/compiler/tf2tensorrt/convert/trt_optimization_pass.h
index f79048bb5f6..e0aaa5500ab 100644
--- a/tensorflow/compiler/tf2tensorrt/convert/trt_optimization_pass.h
+++ b/tensorflow/compiler/tf2tensorrt/convert/trt_optimization_pass.h
@@ -23,8 +23,7 @@ limitations under the License.
 #include "tensorflow/core/grappler/optimizers/custom_graph_optimizer.h"
 #include "tensorflow/core/platform/logging.h"
 
-#if GOOGLE_CUDA
-#if GOOGLE_TENSORRT
+#if GOOGLE_CUDA && GOOGLE_TENSORRT
 
 namespace tensorflow {
 namespace tensorrt {
@@ -83,6 +82,5 @@ class TRTOptimizationPass : public grappler::CustomGraphOptimizer {
 }  // namespace tensorrt
 }  // namespace tensorflow
 
-#endif  // GOOGLE_CUDA
-#endif  // GOOGLE_TENSORRT
+#endif  // GOOGLE_CUDA && GOOGLE_TENSORRT
 #endif  // TENSORFLOW_COMPILER_TF2TENSORRT_CONVERT_TRT_OPTIMIZATION_PASS_H_
diff --git a/tensorflow/compiler/tf2tensorrt/kernels/get_calibration_data_op.cc b/tensorflow/compiler/tf2tensorrt/kernels/get_calibration_data_op.cc
index 3143b06817e..76fb40b9520 100644
--- a/tensorflow/compiler/tf2tensorrt/kernels/get_calibration_data_op.cc
+++ b/tensorflow/compiler/tf2tensorrt/kernels/get_calibration_data_op.cc
@@ -22,8 +22,7 @@ limitations under the License.
 #include "tensorflow/core/framework/resource_mgr.h"
 #include "tensorflow/core/lib/core/refcount.h"
 
-#if GOOGLE_CUDA
-#if GOOGLE_TENSORRT
+#if GOOGLE_CUDA && GOOGLE_TENSORRT
 
 namespace tensorflow {
 namespace tensorrt {
@@ -67,5 +66,4 @@ REGISTER_KERNEL_BUILDER(Name("GetCalibrationDataOp").Device(DEVICE_GPU),
 }  // namespace tensorrt
 }  // namespace tensorflow
 
-#endif  // GOOGLE_TENSORRT
-#endif  // GOOGLE_CUDA
+#endif  // GOOGLE_CUDA && GOOGLE_TENSORRT
diff --git a/tensorflow/compiler/tf2tensorrt/kernels/trt_engine_op.cc b/tensorflow/compiler/tf2tensorrt/kernels/trt_engine_op.cc
index 98d199ca9ab..1094555a622 100644
--- a/tensorflow/compiler/tf2tensorrt/kernels/trt_engine_op.cc
+++ b/tensorflow/compiler/tf2tensorrt/kernels/trt_engine_op.cc
@@ -48,8 +48,7 @@ limitations under the License.
 #include "tensorflow/core/util/env_var.h"
 #include "tensorflow/stream_executor/lib/statusor.h"
 
-#if GOOGLE_CUDA
-#if GOOGLE_TENSORRT
+#if GOOGLE_CUDA && GOOGLE_TENSORRT
 #include "third_party/gpus/cuda/include/cuda_runtime_api.h"
 #include "third_party/tensorrt/NvInfer.h"
 
@@ -1009,5 +1008,4 @@ REGISTER_KERNEL_BUILDER(Name("TRTEngineOp").Device(DEVICE_GPU), TRTEngineOp);
 }  // namespace tensorrt
 }  // namespace tensorflow
 
-#endif  // GOOGLE_TENSORRT
-#endif  // GOOGLE_CUDA
+#endif  // GOOGLE_CUDA && GOOGLE_TENSORRT
diff --git a/tensorflow/compiler/tf2tensorrt/kernels/trt_engine_op_test.cc b/tensorflow/compiler/tf2tensorrt/kernels/trt_engine_op_test.cc
index a06010de1c7..71193dc24cf 100644
--- a/tensorflow/compiler/tf2tensorrt/kernels/trt_engine_op_test.cc
+++ b/tensorflow/compiler/tf2tensorrt/kernels/trt_engine_op_test.cc
@@ -50,8 +50,7 @@ limitations under the License.
 #include "tensorflow/core/platform/status.h"
 #include "tensorflow/core/public/version.h"
 
-#if GOOGLE_CUDA
-#if GOOGLE_TENSORRT
+#if GOOGLE_CUDA && GOOGLE_TENSORRT
 
 namespace tensorflow {
 namespace tensorrt {
@@ -306,5 +305,4 @@ TYPED_TEST(TRTEngineOpTest, Basic) {
 }  // namespace tensorrt
 }  // namespace tensorflow
 
-#endif  // GOOGLE_TENSORRT
-#endif  // GOOGLE_CUDA
+#endif  // GOOGLE_CUDA && GOOGLE_TENSORRT
diff --git a/tensorflow/compiler/tf2tensorrt/kernels/trt_engine_resource_ops.cc b/tensorflow/compiler/tf2tensorrt/kernels/trt_engine_resource_ops.cc
index 2c5821df6ac..3b6e7e91d3b 100644
--- a/tensorflow/compiler/tf2tensorrt/kernels/trt_engine_resource_ops.cc
+++ b/tensorflow/compiler/tf2tensorrt/kernels/trt_engine_resource_ops.cc
@@ -33,8 +33,7 @@ limitations under the License.
 #include "tensorflow/core/platform/mutex.h"
 #include "tensorflow/core/platform/thread_annotations.h"
 
-#if GOOGLE_CUDA
-#if GOOGLE_TENSORRT
+#if GOOGLE_CUDA && GOOGLE_TENSORRT
 #include "third_party/tensorrt/NvInfer.h"
 
 namespace tensorflow {
@@ -251,5 +250,4 @@ REGISTER_KERNEL_BUILDER(Name("SerializeTRTResource").Device(DEVICE_GPU),
 }  // namespace tensorrt
 }  // namespace tensorflow
 
-#endif  // GOOGLE_TENSORRT
-#endif  // GOOGLE_CUDA
+#endif  // GOOGLE_CUDA && GOOGLE_TENSORRT
diff --git a/tensorflow/compiler/tf2tensorrt/kernels/trt_engine_resource_ops_test.cc b/tensorflow/compiler/tf2tensorrt/kernels/trt_engine_resource_ops_test.cc
index 4a24160569d..6a073ee24d0 100644
--- a/tensorflow/compiler/tf2tensorrt/kernels/trt_engine_resource_ops_test.cc
+++ b/tensorflow/compiler/tf2tensorrt/kernels/trt_engine_resource_ops_test.cc
@@ -48,8 +48,7 @@ limitations under the License.
 #include "tensorflow/core/platform/tstring.h"
 #include "tensorflow/core/platform/types.h"
 
-#if GOOGLE_CUDA
-#if GOOGLE_TENSORRT
+#if GOOGLE_CUDA && GOOGLE_TENSORRT
 
 namespace tensorflow {
 namespace tensorrt {
@@ -246,5 +245,4 @@ TEST_F(TRTEngineResourceOpsTest, Basic) {
 }  // namespace tensorrt
 }  // namespace tensorflow
 
-#endif  // GOOGLE_TENSORRT
-#endif  // GOOGLE_CUDA
+#endif  // GOOGLE_CUDA && GOOGLE_TENSORRT
diff --git a/tensorflow/compiler/tf2tensorrt/ops/get_calibration_data_op.cc b/tensorflow/compiler/tf2tensorrt/ops/get_calibration_data_op.cc
index 573172b92e6..2af3164c3e2 100644
--- a/tensorflow/compiler/tf2tensorrt/ops/get_calibration_data_op.cc
+++ b/tensorflow/compiler/tf2tensorrt/ops/get_calibration_data_op.cc
@@ -13,8 +13,7 @@ See the License for the specific language governing permissions and
 limitations under the License.
 ==============================================================================*/
 
-#if GOOGLE_CUDA
-#if GOOGLE_TENSORRT
+#if GOOGLE_CUDA && GOOGLE_TENSORRT
 
 #include "tensorflow/core/framework/common_shape_fns.h"
 #include "tensorflow/core/framework/op.h"
@@ -34,5 +33,4 @@ Returns calibration data for the given resource name
 
 }  // namespace tensorflow
 
-#endif  // GOOGLE_TENSORRT
-#endif  // GOOGLE_CUDA
+#endif  // GOOGLE_CUDA && GOOGLE_TENSORRT
diff --git a/tensorflow/compiler/tf2tensorrt/ops/trt_engine_op.cc b/tensorflow/compiler/tf2tensorrt/ops/trt_engine_op.cc
index bd3c2b299a9..2527fe9b910 100644
--- a/tensorflow/compiler/tf2tensorrt/ops/trt_engine_op.cc
+++ b/tensorflow/compiler/tf2tensorrt/ops/trt_engine_op.cc
@@ -13,8 +13,7 @@ See the License for the specific language governing permissions and
 limitations under the License.
 ==============================================================================*/
 
-#if GOOGLE_CUDA
-#if GOOGLE_TENSORRT
+#if GOOGLE_CUDA && GOOGLE_TENSORRT
 
 #include "tensorflow/core/framework/common_shape_fns.h"
 #include "tensorflow/core/framework/op.h"
@@ -59,5 +58,4 @@ REGISTER_OP("TRTEngineOp")
     .Attr("static_engine: bool = true");
 }  // namespace tensorflow
 
-#endif  // GOOGLE_TENSORRT
-#endif  // GOOGLE_CUDA
+#endif  // GOOGLE_CUDA && GOOGLE_TENSORRT
diff --git a/tensorflow/compiler/tf2tensorrt/ops/trt_engine_resource_ops.cc b/tensorflow/compiler/tf2tensorrt/ops/trt_engine_resource_ops.cc
index 01911de66ec..3141092de03 100644
--- a/tensorflow/compiler/tf2tensorrt/ops/trt_engine_resource_ops.cc
+++ b/tensorflow/compiler/tf2tensorrt/ops/trt_engine_resource_ops.cc
@@ -13,8 +13,7 @@ See the License for the specific language governing permissions and
 limitations under the License.
 ==============================================================================*/
 
-#if GOOGLE_CUDA
-#if GOOGLE_TENSORRT
+#if GOOGLE_CUDA && GOOGLE_TENSORRT
 
 #include "tensorflow/core/framework/common_shape_fns.h"
 #include "tensorflow/core/framework/op.h"
@@ -46,5 +45,4 @@ REGISTER_OP("SerializeTRTResource")
 
 }  // namespace tensorflow
 
-#endif  // GOOGLE_TENSORRT
-#endif  // GOOGLE_CUDA
+#endif  // GOOGLE_CUDA && GOOGLE_TENSORRT
diff --git a/tensorflow/compiler/tf2tensorrt/plugin/plugin_cast.cu.cc b/tensorflow/compiler/tf2tensorrt/plugin/plugin_cast.cu.cc
index 4c0d8b0392a..141a7d1f462 100644
--- a/tensorflow/compiler/tf2tensorrt/plugin/plugin_cast.cu.cc
+++ b/tensorflow/compiler/tf2tensorrt/plugin/plugin_cast.cu.cc
@@ -17,8 +17,7 @@ limitations under the License.
 #include "tensorflow/compiler/tf2tensorrt/plugin/trt_plugin.h"
 #include "tensorflow/core/platform/logging.h"
 
-#if GOOGLE_CUDA
-#if GOOGLE_TENSORRT
+#if GOOGLE_CUDA && GOOGLE_TENSORRT
 #define EIGEN_USE_GPU  // For definition of Eigen::GpuDevice.
 #include "third_party/gpus/cuda/include/cuda_runtime_api.h"
 #include "tensorflow/core/util/gpu_kernel_helper.h"
@@ -234,5 +233,4 @@ REGISTER_TFTRT_PLUGIN(CastPluginCreator);
 }  // namespace tensorrt
 }  // namespace tensorflow
 
-#endif  // GOOGLE_CUDA
-#endif  // GOOGLE_TENSORRT
+#endif  // GOOGLE_CUDA && GOOGLE_TENSORRT
diff --git a/tensorflow/compiler/tf2tensorrt/plugin/trt_plugin.cc b/tensorflow/compiler/tf2tensorrt/plugin/trt_plugin.cc
index 563ce724f43..83d5f9b5965 100644
--- a/tensorflow/compiler/tf2tensorrt/plugin/trt_plugin.cc
+++ b/tensorflow/compiler/tf2tensorrt/plugin/trt_plugin.cc
@@ -17,8 +17,7 @@ limitations under the License.
 
 #include 
 
-#if GOOGLE_CUDA
-#if GOOGLE_TENSORRT
+#if GOOGLE_CUDA && GOOGLE_TENSORRT
 
 namespace tensorflow {
 namespace tensorrt {
@@ -30,5 +29,4 @@ const char* kTfTrtPluginNamespace = "TF";
 }  // namespace tensorrt
 }  // namespace tensorflow
 
-#endif  // GOOGLE_CUDA
-#endif  // GOOGLE_TENSORRT
+#endif  // GOOGLE_CUDA && GOOGLE_TENSORRT
diff --git a/tensorflow/compiler/tf2tensorrt/plugin/trt_plugin.h b/tensorflow/compiler/tf2tensorrt/plugin/trt_plugin.h
index bdb046e6c71..600ac6683da 100644
--- a/tensorflow/compiler/tf2tensorrt/plugin/trt_plugin.h
+++ b/tensorflow/compiler/tf2tensorrt/plugin/trt_plugin.h
@@ -20,8 +20,7 @@ limitations under the License.
 
 #include "tensorflow/core/platform/logging.h"
 
-#if GOOGLE_CUDA
-#if GOOGLE_TENSORRT
+#if GOOGLE_CUDA && GOOGLE_TENSORRT
 #include "third_party/tensorrt/NvInfer.h"
 
 namespace tensorflow {
@@ -90,7 +89,6 @@ class TrtPluginRegistrar {
 }  // namespace tensorrt
 }  // namespace tensorflow
 
-#endif  // GOOGLE_TENSORRT
-#endif  // GOOGLE_CUDA
+#endif  // GOOGLE_CUDA && GOOGLE_TENSORRT
 
 #endif  // TENSORFLOW_COMPILER_TF2TENSORRT_PLUGIN_TRT_PLUGIN_H_
diff --git a/tensorflow/compiler/tf2tensorrt/segment/segment.cc b/tensorflow/compiler/tf2tensorrt/segment/segment.cc
index 32e30006f58..d9080b6f69a 100644
--- a/tensorflow/compiler/tf2tensorrt/segment/segment.cc
+++ b/tensorflow/compiler/tf2tensorrt/segment/segment.cc
@@ -35,8 +35,7 @@ limitations under the License.
 #include "tensorflow/core/platform/types.h"
 #include "tensorflow/core/util/env_var.h"
 
-#if GOOGLE_CUDA
-#if GOOGLE_TENSORRT
+#if GOOGLE_CUDA && GOOGLE_TENSORRT
 
 namespace tensorflow {
 namespace tensorrt {
@@ -1062,5 +1061,4 @@ Status SegmentGraph(const Graph* tf_graph,
 }  // namespace tensorrt
 }  // namespace tensorflow
 
-#endif  // GOOGLE_TENSORRT
-#endif  // GOOGLE_CUDA
+#endif  // GOOGLE_CUDA && GOOGLE_TENSORRT
diff --git a/tensorflow/compiler/tf2tensorrt/segment/segment.h b/tensorflow/compiler/tf2tensorrt/segment/segment.h
index 7295c8f0d9d..3f79983cfd2 100644
--- a/tensorflow/compiler/tf2tensorrt/segment/segment.h
+++ b/tensorflow/compiler/tf2tensorrt/segment/segment.h
@@ -25,8 +25,7 @@ limitations under the License.
 #include "tensorflow/core/lib/core/status.h"
 #include "tensorflow/core/platform/types.h"
 
-#if GOOGLE_CUDA
-#if GOOGLE_TENSORRT
+#if GOOGLE_CUDA && GOOGLE_TENSORRT
 
 namespace tensorflow {
 namespace tensorrt {
@@ -67,7 +66,6 @@ Status SegmentGraph(const Graph* tf_graph,
 }  // namespace tensorrt
 }  // namespace tensorflow
 
-#endif  // GOOGLE_TENSORRT
-#endif  // GOOGLE_CUDA
+#endif  // GOOGLE_CUDA && GOOGLE_TENSORRT
 
 #endif  // TENSORFLOW_COMPILER_TF2TENSORRT_SEGMENT_SEGMENT_H_
diff --git a/tensorflow/compiler/tf2tensorrt/segment/segment_test.cc b/tensorflow/compiler/tf2tensorrt/segment/segment_test.cc
index 2437481a9c4..f3bc5bfbee6 100644
--- a/tensorflow/compiler/tf2tensorrt/segment/segment_test.cc
+++ b/tensorflow/compiler/tf2tensorrt/segment/segment_test.cc
@@ -26,8 +26,7 @@ limitations under the License.
 #include "tensorflow/core/platform/types.h"
 #include "tensorflow/core/public/session.h"
 
-#if GOOGLE_CUDA
-#if GOOGLE_TENSORRT
+#if GOOGLE_CUDA && GOOGLE_TENSORRT
 
 namespace tensorflow {
 namespace tensorrt {
@@ -522,5 +521,4 @@ TEST_F(SegmentTest, IncompatibleBatchSizes) {
 }  // namespace tensorrt
 }  // namespace tensorflow
 
-#endif  // GOOGLE_TENSORRT
-#endif  // GOOGLE_CUDA
+#endif  // GOOGLE_CUDA && GOOGLE_TENSORRT
diff --git a/tensorflow/compiler/tf2tensorrt/segment/union_find.h b/tensorflow/compiler/tf2tensorrt/segment/union_find.h
index 70e83c12fca..b53615ec019 100644
--- a/tensorflow/compiler/tf2tensorrt/segment/union_find.h
+++ b/tensorflow/compiler/tf2tensorrt/segment/union_find.h
@@ -19,8 +19,7 @@ limitations under the License.
 #include "absl/strings/str_format.h"
 #include "absl/types/optional.h"
 
-#if GOOGLE_CUDA
-#if GOOGLE_TENSORRT
+#if GOOGLE_CUDA && GOOGLE_TENSORRT
 
 namespace tensorflow {
 namespace tensorrt {
@@ -217,7 +216,6 @@ UnionFind* UnionFind::FindRoot() {
 }  // namespace tensorrt
 }  // namespace tensorflow
 
-#endif  // GOOGLE_TENSORRT
-#endif  // GOOGLE_CUDA
+#endif  // GOOGLE_CUDA && GOOGLE_TENSORRT
 
 #endif  // TENSORFLOW_COMPILER_TF2TENSORRT_SEGMENT_UNION_FIND_H_
diff --git a/tensorflow/compiler/tf2tensorrt/tensorrt_test.cc b/tensorflow/compiler/tf2tensorrt/tensorrt_test.cc
index 510591bfe00..e994d20df33 100644
--- a/tensorflow/compiler/tf2tensorrt/tensorrt_test.cc
+++ b/tensorflow/compiler/tf2tensorrt/tensorrt_test.cc
@@ -18,8 +18,7 @@ limitations under the License.
 #include "tensorflow/core/platform/stream_executor.h"
 #include "tensorflow/core/platform/test.h"
 
-#if GOOGLE_CUDA
-#if GOOGLE_TENSORRT
+#if GOOGLE_CUDA && GOOGLE_TENSORRT
 #include "third_party/gpus/cuda/include/cuda.h"
 #include "third_party/gpus/cuda/include/cuda_runtime_api.h"
 #include "third_party/tensorrt/NvInfer.h"
@@ -164,5 +163,4 @@ TEST(TensorrtTest, BasicFunctions) {
 }  // namespace
 }  // namespace tensorflow
 
-#endif  // GOOGLE_TENSORRT
-#endif  // GOOGLE_CUDA
+#endif  // GOOGLE_CUDA && GOOGLE_TENSORRT
diff --git a/tensorflow/compiler/tf2tensorrt/utils/trt_allocator.cc b/tensorflow/compiler/tf2tensorrt/utils/trt_allocator.cc
index 617ea7fad5c..d4f3a524577 100644
--- a/tensorflow/compiler/tf2tensorrt/utils/trt_allocator.cc
+++ b/tensorflow/compiler/tf2tensorrt/utils/trt_allocator.cc
@@ -17,11 +17,9 @@ limitations under the License.
 
 #include "tensorflow/core/platform/logging.h"
 
-#if GOOGLE_CUDA
-#if GOOGLE_TENSORRT
+#if GOOGLE_CUDA && GOOGLE_TENSORRT
 #include "third_party/gpus/cuda/include/cuda_runtime_api.h"
-#endif  // GOOGLE_TENSORRT
-#endif  // GOOGLE_CUDA
+#endif  // GOOGLE_CUDA && GOOGLE_TENSORRT
 
 namespace tensorflow {
 namespace tensorrt {
@@ -52,8 +50,7 @@ void* Align(uint64_t alignment, uint64_t size, void*& ptr, uint64_t& space) {
 }  // namespace tensorrt
 }  // namespace tensorflow
 
-#if GOOGLE_CUDA
-#if GOOGLE_TENSORRT
+#if GOOGLE_CUDA && GOOGLE_TENSORRT
 
 namespace tensorflow {
 namespace tensorrt {
@@ -113,5 +110,4 @@ void TRTDeviceAllocator::free(void* memory) {
 }  // namespace tensorrt
 }  // namespace tensorflow
 
-#endif  // GOOGLE_TENSORRT
-#endif  // GOOGLE_CUDA
+#endif  // GOOGLE_CUDA && GOOGLE_TENSORRT
diff --git a/tensorflow/compiler/tf2tensorrt/utils/trt_allocator.h b/tensorflow/compiler/tf2tensorrt/utils/trt_allocator.h
index 4ab8b52f523..d219a8a14e8 100644
--- a/tensorflow/compiler/tf2tensorrt/utils/trt_allocator.h
+++ b/tensorflow/compiler/tf2tensorrt/utils/trt_allocator.h
@@ -20,11 +20,9 @@ limitations under the License.
 
 #include "tensorflow/core/framework/allocator.h"
 
-#if GOOGLE_CUDA
-#if GOOGLE_TENSORRT
+#if GOOGLE_CUDA && GOOGLE_TENSORRT
 #include "third_party/tensorrt/NvInfer.h"
-#endif  // GOOGLE_TENSORRT
-#endif  // GOOGLE_CUDA
+#endif  // GOOGLE_CUDA && GOOGLE_TENSORRT
 
 namespace tensorflow {
 namespace tensorrt {
@@ -33,8 +31,7 @@ void* Align(uint64_t alignment, uint64_t size, void*& ptr, uint64_t& space);
 }  // namespace tensorrt
 }  // namespace tensorflow
 
-#if GOOGLE_CUDA
-#if GOOGLE_TENSORRT
+#if GOOGLE_CUDA && GOOGLE_TENSORRT
 
 namespace tensorflow {
 namespace tensorrt {
@@ -69,6 +66,5 @@ class TRTDeviceAllocator : public TRTBaseAllocator {
 }  // namespace tensorrt
 }  // namespace tensorflow
 
-#endif  // GOOGLE_TENSORRT
-#endif  // GOOGLE_CUDA
+#endif  // GOOGLE_CUDA && GOOGLE_TENSORRT
 #endif  // TENSORFLOW_COMPILER_TF2TENSORRT_UTILS_TRT_ALLOCATOR_H_
diff --git a/tensorflow/compiler/tf2tensorrt/utils/trt_engine_utils.cc b/tensorflow/compiler/tf2tensorrt/utils/trt_engine_utils.cc
index ed997b267b1..8ccfb8b06f0 100644
--- a/tensorflow/compiler/tf2tensorrt/utils/trt_engine_utils.cc
+++ b/tensorflow/compiler/tf2tensorrt/utils/trt_engine_utils.cc
@@ -25,8 +25,7 @@ limitations under the License.
 #include "tensorflow/core/lib/core/status.h"
 #include "tensorflow/core/platform/errors.h"
 
-#if GOOGLE_CUDA
-#if GOOGLE_TENSORRT
+#if GOOGLE_CUDA && GOOGLE_TENSORRT
 #include "third_party/tensorrt/NvInfer.h"
 
 namespace tensorflow {
@@ -257,5 +256,4 @@ Status TrtEnqueue(nvinfer1::IExecutionContext* execution_context,
 }  // namespace tensorrt
 }  // namespace tensorflow
 
-#endif  // GOOGLE_TENSORRT
-#endif  // GOOGLE_CUDA
+#endif  // GOOGLE_CUDA && GOOGLE_TENSORRT
diff --git a/tensorflow/compiler/tf2tensorrt/utils/trt_engine_utils.h b/tensorflow/compiler/tf2tensorrt/utils/trt_engine_utils.h
index a471749877a..1ea4fe28cb4 100644
--- a/tensorflow/compiler/tf2tensorrt/utils/trt_engine_utils.h
+++ b/tensorflow/compiler/tf2tensorrt/utils/trt_engine_utils.h
@@ -24,8 +24,7 @@ limitations under the License.
 #include "tensorflow/core/framework/tensor_shape.h"
 #include "tensorflow/core/lib/core/status.h"
 
-#if GOOGLE_CUDA
-#if GOOGLE_TENSORRT
+#if GOOGLE_CUDA && GOOGLE_TENSORRT
 #include "third_party/tensorrt/NvInfer.h"
 
 namespace tensorflow {
@@ -91,7 +90,6 @@ Status TrtEnqueue(nvinfer1::IExecutionContext* execution_context,
 }  // namespace tensorrt
 }  // namespace tensorflow
 
-#endif  // GOOGLE_TENSORRT
-#endif  // GOOGLE_CUDA
+#endif  // GOOGLE_CUDA && GOOGLE_TENSORRT
 
 #endif  // TENSORFLOW_COMPILER_TF2TENSORRT_UTILS_TRT_ENGINE_UTILS_H_
diff --git a/tensorflow/compiler/tf2tensorrt/utils/trt_int8_calibrator.cc b/tensorflow/compiler/tf2tensorrt/utils/trt_int8_calibrator.cc
index 554c127fa37..24271e352a7 100644
--- a/tensorflow/compiler/tf2tensorrt/utils/trt_int8_calibrator.cc
+++ b/tensorflow/compiler/tf2tensorrt/utils/trt_int8_calibrator.cc
@@ -20,8 +20,7 @@ limitations under the License.
 
 #include "tensorflow/core/platform/logging.h"
 
-#if GOOGLE_CUDA
-#if GOOGLE_TENSORRT
+#if GOOGLE_CUDA && GOOGLE_TENSORRT
 #include "third_party/gpus/cuda/include/cuda_runtime_api.h"
 
 namespace tensorflow {
@@ -147,5 +146,4 @@ TRTInt8Calibrator::~TRTInt8Calibrator() {
 }  // namespace tensorrt
 }  // namespace tensorflow
 
-#endif
-#endif
+#endif  // GOOGLE_CUDA && GOOGLE_TENSORRT
diff --git a/tensorflow/compiler/tf2tensorrt/utils/trt_int8_calibrator.h b/tensorflow/compiler/tf2tensorrt/utils/trt_int8_calibrator.h
index 06b39716490..4c670e85f52 100644
--- a/tensorflow/compiler/tf2tensorrt/utils/trt_int8_calibrator.h
+++ b/tensorflow/compiler/tf2tensorrt/utils/trt_int8_calibrator.h
@@ -22,8 +22,7 @@ limitations under the License.
 #include 
 #include "tensorflow/core/platform/mutex.h"
 
-#if GOOGLE_CUDA
-#if GOOGLE_TENSORRT
+#if GOOGLE_CUDA && GOOGLE_TENSORRT
 
 #include "third_party/gpus/cuda/include/cuda_runtime_api.h"
 #include "third_party/tensorrt/NvInfer.h"
@@ -101,6 +100,5 @@ struct TRTInt8Calibrator : public nvinfer1::IInt8EntropyCalibrator {
 }  // namespace tensorrt
 }  // namespace tensorflow
 
-#endif
-#endif
+#endif  // GOOGLE_CUDA && GOOGLE_TENSORRT
 #endif  // TENSORFLOW_COMPILER_TF2TENSORRT_UTILS_TRT_INT8_CALIBRATOR_H_
diff --git a/tensorflow/compiler/tf2tensorrt/utils/trt_logger.cc b/tensorflow/compiler/tf2tensorrt/utils/trt_logger.cc
index 193687ebc8c..e34bf5e7397 100644
--- a/tensorflow/compiler/tf2tensorrt/utils/trt_logger.cc
+++ b/tensorflow/compiler/tf2tensorrt/utils/trt_logger.cc
@@ -15,8 +15,7 @@ limitations under the License.
 
 #include "tensorflow/compiler/tf2tensorrt/utils/trt_logger.h"
 
-#if GOOGLE_CUDA
-#if GOOGLE_TENSORRT
+#if GOOGLE_CUDA && GOOGLE_TENSORRT
 #include "tensorflow/compiler/tf2tensorrt/common/utils.h"
 #include "tensorflow/compiler/tf2tensorrt/convert/logger_registry.h"
 #include "tensorflow/core/platform/logging.h"
@@ -68,5 +67,4 @@ REGISTER_TENSORRT_LOGGER("DefaultLogger", Logger::GetLogger());
 }  // namespace tensorrt
 }  // namespace tensorflow
 
-#endif  // GOOGLE_CUDA
-#endif  // GOOGLE_TENSORRT
+#endif  // GOOGLE_CUDA && GOOGLE_TENSORRT
diff --git a/tensorflow/compiler/tf2tensorrt/utils/trt_logger.h b/tensorflow/compiler/tf2tensorrt/utils/trt_logger.h
index 2ade1b48f47..ce6552e8fe9 100644
--- a/tensorflow/compiler/tf2tensorrt/utils/trt_logger.h
+++ b/tensorflow/compiler/tf2tensorrt/utils/trt_logger.h
@@ -18,8 +18,7 @@ limitations under the License.
 
 #include "tensorflow/core/platform/types.h"
 
-#if GOOGLE_CUDA
-#if GOOGLE_TENSORRT
+#if GOOGLE_CUDA && GOOGLE_TENSORRT
 #include "third_party/tensorrt/NvInfer.h"
 
 namespace tensorflow {
@@ -40,7 +39,6 @@ class Logger : public nvinfer1::ILogger {
 }  // namespace tensorrt
 }  // namespace tensorflow
 
-#endif  // GOOGLE_TENSORRT
-#endif  // GOOGLE_CUDA
+#endif  // GOOGLE_CUDA && GOOGLE_TENSORRT
 
 #endif  // TENSORFLOW_COMPILER_TF2TENSORRT_UTILS_TRT_LOGGER_H_
diff --git a/tensorflow/compiler/tf2tensorrt/utils/trt_lru_cache.cc b/tensorflow/compiler/tf2tensorrt/utils/trt_lru_cache.cc
index fbcdaad52c0..ee7e6272372 100644
--- a/tensorflow/compiler/tf2tensorrt/utils/trt_lru_cache.cc
+++ b/tensorflow/compiler/tf2tensorrt/utils/trt_lru_cache.cc
@@ -23,8 +23,7 @@ limitations under the License.
 #include "tensorflow/core/framework/tensor_shape.h"
 #include "tensorflow/core/platform/mutex.h"
 
-#if GOOGLE_CUDA
-#if GOOGLE_TENSORRT
+#if GOOGLE_CUDA && GOOGLE_TENSORRT
 #include "third_party/tensorrt/NvInfer.h"
 
 namespace tensorflow {
@@ -141,5 +140,4 @@ EngineContext* TRTEngineCacheResource::GetEngineContext(const int profile_id) {
 }  // namespace tensorrt
 }  // namespace tensorflow
 
-#endif  // GOOGLE_TENSORRT
-#endif  // GOOGLE_CUDA
+#endif  // GOOGLE_CUDA && GOOGLE_TENSORRT
diff --git a/tensorflow/compiler/tf2tensorrt/utils/trt_lru_cache.h b/tensorflow/compiler/tf2tensorrt/utils/trt_lru_cache.h
index 8e345254f75..991b9a949e4 100644
--- a/tensorflow/compiler/tf2tensorrt/utils/trt_lru_cache.h
+++ b/tensorflow/compiler/tf2tensorrt/utils/trt_lru_cache.h
@@ -115,8 +115,7 @@ class LRUCache {
   }
 };
 
-#if GOOGLE_CUDA
-#if GOOGLE_TENSORRT
+#if GOOGLE_CUDA && GOOGLE_TENSORRT
 
 struct EngineContext {
   EngineContext() {}  // Creates an empty context.
@@ -223,8 +222,7 @@ class TRTEngineCacheResource : public ResourceBase {
   TrtShapeOptimizationProfile profiles_;
 };
 
-#endif  // GOOGLE_TENSORRT
-#endif  // GOOGLE_CUDA
+#endif  // GOOGLE_CUDA && GOOGLE_TENSORRT
 
 }  // namespace tensorrt
 }  // namespace tensorflow
diff --git a/tensorflow/compiler/tf2tensorrt/utils/trt_shape_optimization_profiles.h b/tensorflow/compiler/tf2tensorrt/utils/trt_shape_optimization_profiles.h
index 40c7f5dcf31..fc688b14139 100644
--- a/tensorflow/compiler/tf2tensorrt/utils/trt_shape_optimization_profiles.h
+++ b/tensorflow/compiler/tf2tensorrt/utils/trt_shape_optimization_profiles.h
@@ -29,8 +29,7 @@ limitations under the License.
 #include "tensorflow/core/lib/strings/str_util.h"
 #include "tensorflow/core/lib/strings/strcat.h"
 
-#if GOOGLE_CUDA
-#if GOOGLE_TENSORRT
+#if GOOGLE_CUDA && GOOGLE_TENSORRT
 
 #include "third_party/tensorrt/NvInfer.h"
 
@@ -173,6 +172,5 @@ class TrtShapeOptimizationProfile {
 }  // namespace tensorrt
 }  // namespace tensorflow
 
-#endif  // GOOGLE_TENSORRT
-#endif  // GOOGLE_CUDA
+#endif  // GOOGLE_CUDA && GOOGLE_TENSORRT
 #endif  // TENSORFLOW_COMPILER_TF2TENSORRT_UTILS_TRT_SHAPE_OPTIMIZATION_PROFILES_H_
diff --git a/tensorflow/compiler/tf2tensorrt/utils/trt_shape_optimization_profiles_test.cc b/tensorflow/compiler/tf2tensorrt/utils/trt_shape_optimization_profiles_test.cc
index 501810587e0..32c2200fb71 100644
--- a/tensorflow/compiler/tf2tensorrt/utils/trt_shape_optimization_profiles_test.cc
+++ b/tensorflow/compiler/tf2tensorrt/utils/trt_shape_optimization_profiles_test.cc
@@ -13,8 +13,7 @@ See the License for the specific language governing permissions and
 limitations under the License.
 ==============================================================================*/
 
-#if GOOGLE_CUDA
-#if GOOGLE_TENSORRT
+#if GOOGLE_CUDA && GOOGLE_TENSORRT
 
 #include 
 
@@ -214,5 +213,4 @@ TEST_F(TrtShapeOptimizationProfileTest, Dynamic) {
 }  // namespace tensorrt
 }  // namespace tensorflow
 
-#endif  // GOOGLE_TENSORRT
-#endif  // GOOGLE_CUDA
+#endif  // GOOGLE_CUDA && GOOGLE_TENSORRT

From 32d63d0a3efb5e0b65dc6f590e54248054cf9f14 Mon Sep 17 00:00:00 2001
From: Nathan Luehr 
Date: Fri, 15 May 2020 11:40:22 -0500
Subject: [PATCH 339/867] Removed TENSOR_OP disable env vars.

* TF_DISABLE_CUBLAS_TENSOR_OP_MATH
* TF_DISABLE_CUDNN_TENSOR_OP_MATH
* TF_DISABLE_CUDNN_RNN_TENSOR_OP_MATH
---
 tensorflow/stream_executor/cuda/cuda_blas.cc | 21 ++------
 tensorflow/stream_executor/cuda/cuda_dnn.cc  | 55 +++++---------------
 2 files changed, 16 insertions(+), 60 deletions(-)

diff --git a/tensorflow/stream_executor/cuda/cuda_blas.cc b/tensorflow/stream_executor/cuda/cuda_blas.cc
index c9f0fc462c9..65c07e72154 100644
--- a/tensorflow/stream_executor/cuda/cuda_blas.cc
+++ b/tensorflow/stream_executor/cuda/cuda_blas.cc
@@ -101,18 +101,6 @@ static std::string ToString(cublasStatus_t status) {
   }
 }
 
-// Decide whether to enable TENSOR_OP_MATH
-static bool TensorOpMathEnabled() {
-  static bool is_enabled = [] {
-    bool is_disabled;
-    TF_CHECK_OK(
-        tensorflow::ReadBoolFromEnvVar("TF_DISABLE_CUBLAS_TENSOR_OP_MATH",
-                                       /*default_val=*/false, &is_disabled));
-    return !is_disabled;
-  }();
-  return is_enabled;
-}
-
 // cuBLAS has interfaces that permit pointers to be passed from either the host
 // memory space or the device memory space; however, you must instruct it as to
 // which address space those pointers are in with cublasSetPointerMode.
@@ -1640,7 +1628,7 @@ bool CUDABlas::DoBlasGemm(
                                                                    &cc_minor);
 
   // GPUs < sm_70 don't support tensor ops.
-  if (cc_major >= 7 && TensorOpMathEnabled()) {
+  if (cc_major >= 7) {
     use_tensor_ops = true;
   }
 #endif
@@ -1921,8 +1909,7 @@ static bool TensorOpsAvailable(int cc_major) {
   // strictly correct.  We can't simply enable it, though, as that would change
   // clients' behavior significantly: Using tensor ops on fp32 inputs cause them
   // to be rounded to fp16.
-  if (cc_major >= 7 && TensorOpMathEnabled() &&
-      std::is_same::value) {
+  if (cc_major >= 7 && std::is_same::value) {
     return true;
   }
 #endif
@@ -2270,7 +2257,7 @@ port::Status CUDABlas::DoBlasGemmBatchedInternal(
   if (stream->parent()->GetDeviceDescription().cuda_compute_capability(
           &cc_major, &cc_minor) &&
       cc_major >= 5) {
-    bool use_tensor_ops = TensorOpMathEnabled() && data_type == CUDA_R_16F;
+    bool use_tensor_ops = data_type == CUDA_R_16F;
     cublasGemmAlgo_t algo =
         (use_tensor_ops ? CUBLAS_GEMM_DFALT_TENSOR_OP : CUBLAS_GEMM_DFALT);
     cudaDataType_t compute_type =
@@ -2425,7 +2412,7 @@ bool CUDABlas::DoBlasGemmStridedBatched(
   if (stream->parent()->GetDeviceDescription().cuda_compute_capability(
           &cc_major, &cc_minor)) {
     // GPUs < sm_70 don't support tensor ops.
-    if (cc_major >= 7 && TensorOpMathEnabled()) {
+    if (cc_major >= 7) {
       use_tensor_ops = true;
     }
 #if CUDA_VERSION >= 9010
diff --git a/tensorflow/stream_executor/cuda/cuda_dnn.cc b/tensorflow/stream_executor/cuda/cuda_dnn.cc
index be18c989861..e46c271443b 100644
--- a/tensorflow/stream_executor/cuda/cuda_dnn.cc
+++ b/tensorflow/stream_executor/cuda/cuda_dnn.cc
@@ -601,31 +601,6 @@ class CudnnFilterDescriptor {
   SE_DISALLOW_COPY_AND_ASSIGN(CudnnFilterDescriptor);
 };
 
-// A helper function to decide whether to enable the TENSOR_OP_MATH math type
-bool TensorOpMathEnabled() {
-  static bool is_enabled = [] {
-    bool is_disabled = false;
-    TF_CHECK_OK(
-        tensorflow::ReadBoolFromEnvVar("TF_DISABLE_CUDNN_TENSOR_OP_MATH",
-                                       /*default_val=*/false, &is_disabled));
-    return !is_disabled;
-  }();
-  return is_enabled;
-}
-
-// A helper function to decide whether to enable the TENSOR_OP_MATH math type
-// for RNNs.
-bool RnnTensorOpMathEnabled() {
-  static bool is_enabled = [] {
-    bool is_disabled = false;
-    TF_CHECK_OK(
-        tensorflow::ReadBoolFromEnvVar("TF_DISABLE_CUDNN_RNN_TENSOR_OP_MATH",
-                                       /*default_val=*/false, &is_disabled));
-    return !is_disabled;
-  }();
-  return is_enabled;
-}
-
 // A helper function to decide whether to use
 // CUDNN_BATCHNORM_SPATIAL_PERSISTENT in batchnorm. This mode can be faster in
 // some tasks because an optimized path may be selected for CUDNN_DATA_FLOAT
@@ -749,9 +724,7 @@ class CudnnConvolutionDescriptor {
 #if CUDNN_VERSION >= 7000
     cudnnMathType_t math_type =
         (use_tensor_op_math ? CUDNN_TENSOR_OP_MATH : CUDNN_DEFAULT_MATH);
-    if (TensorOpMathEnabled()) {
-      CHECK_CUDNN_OK(cudnnSetConvolutionMathType(handle_.get(), math_type));
-    }
+    CHECK_CUDNN_OK(cudnnSetConvolutionMathType(handle_.get(), math_type));
 #endif
   }
 
@@ -1155,21 +1128,19 @@ class CudnnRnnDescriptor : public dnn::RnnDescriptor {
     // in profile mode, which is run with algorithms returned from
     // GetRnnAlgorithms() (which are non-default and explicitly set whether to
     // use tensor ops). CuDNN 7.2.1 fixed this issue
-    if (RnnTensorOpMathEnabled()) {
-      cudnnMathType_t math_type;
-      if (algorithm_config.algorithm().has_value()) {
-        math_type = algorithm_config.algorithm()->tensor_ops_enabled()
-                        ? CUDNN_TENSOR_OP_MATH
-                        : CUDNN_DEFAULT_MATH;
-      } else {
+    cudnnMathType_t math_type;
+    if (algorithm_config.algorithm().has_value()) {
+      math_type = algorithm_config.algorithm()->tensor_ops_enabled()
+                      ? CUDNN_TENSOR_OP_MATH
+                      : CUDNN_DEFAULT_MATH;
+    } else {
 #if CUDNN_VERSION >= 7201
-        math_type = CUDNN_TENSOR_OP_MATH;
+      math_type = CUDNN_TENSOR_OP_MATH;
 #else
-        math_type = CUDNN_DEFAULT_MATH;
+      math_type = CUDNN_DEFAULT_MATH;
 #endif  // CUDNN_VERSION >= 7201
-      }
-      CHECK_CUDNN_OK(cudnnSetRNNMatrixMathType(rnn_desc.get(), math_type));
     }
+    CHECK_CUDNN_OK(cudnnSetRNNMatrixMathType(rnn_desc.get(), math_type));
 #endif  // CUDNN_VERSION >= 7000
 
     return CudnnRnnDescriptor(cudnn, std::move(rnn_desc), std::move(rnn_plan),
@@ -2686,7 +2657,7 @@ AllocateCudnnConvolutionBackwardFilterWorkspace(
 }
 
 static bool TensorOpMathAvailable(int cc_major) {
-  return cc_major >= 7 && CUDNN_VERSION >= 7000 && TensorOpMathEnabled();
+  return cc_major >= 7 && CUDNN_VERSION >= 7000;
 }
 
 port::StatusOr GetCudnnConvolutionForwardAlgorithm(
@@ -3480,9 +3451,7 @@ bool CudnnSupport::GetRnnAlgorithms(
   for (auto i : algo_types) {
     out_algorithms->push_back({i, /*use_tensor_ops=*/false});
 #if CUDNN_VERSION >= 7100
-    if (RnnTensorOpMathEnabled()) {
-      out_algorithms->push_back({i, /*use_tensor_ops=*/true});
-    }
+    out_algorithms->push_back({i, /*use_tensor_ops=*/true});
 #endif
   }
   return true;

From d2afc9ce83b48170f0821b6b3b5debddd857d320 Mon Sep 17 00:00:00 2001
From: Nathan Luehr 
Date: Fri, 15 May 2020 11:46:41 -0500
Subject: [PATCH 340/867] Add global setting control TF32 execution

---
 tensorflow/core/platform/BUILD         |  7 +++++++
 tensorflow/core/platform/tf32_utils.cc | 27 ++++++++++++++++++++++++++
 tensorflow/core/platform/tf32_utils.h  | 27 ++++++++++++++++++++++++++
 3 files changed, 61 insertions(+)
 create mode 100644 tensorflow/core/platform/tf32_utils.cc
 create mode 100644 tensorflow/core/platform/tf32_utils.h

diff --git a/tensorflow/core/platform/BUILD b/tensorflow/core/platform/BUILD
index 70bb8a89417..33a1e7cfe0a 100644
--- a/tensorflow/core/platform/BUILD
+++ b/tensorflow/core/platform/BUILD
@@ -938,6 +938,13 @@ cc_library(
     alwayslink = 1,
 )
 
+cc_library(
+    name = "tf32_utils",
+    srcs = ["tf32_utils.cc"],
+    hdrs = ["tf32_utils.h"],
+    copts = tf_copts(),
+)
+
 tf_cc_tests(
     name = "low_level_library_tests",
     size = "small",
diff --git a/tensorflow/core/platform/tf32_utils.cc b/tensorflow/core/platform/tf32_utils.cc
new file mode 100644
index 00000000000..715b5996dc3
--- /dev/null
+++ b/tensorflow/core/platform/tf32_utils.cc
@@ -0,0 +1,27 @@
+/* Copyright 2020 The TensorFlow Authors. All Rights Reserved.
+
+Licensed under the Apache License, Version 2.0 (the "License");
+you may not use this file except in compliance with the License.
+You may obtain a copy of the License at
+
+    http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software
+distributed under the License is distributed on an "AS IS" BASIS,
+WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+See the License for the specific language governing permissions and
+limitations under the License.
+==============================================================================*/
+
+#include "tensorflow/core/platform/tf32_utils.h"
+
+namespace tensorflow {
+
+// TODO(nluehr): enable tf32 execution by default after TF32 Ampere testing.
+static bool tf32_enabled = false;
+
+void allow_tf32_execution(bool allow) { tf32_enabled = allow; }
+
+bool tf32_execution_allowed() { return tf32_enabled; }
+
+}  // namespace tensorflow
diff --git a/tensorflow/core/platform/tf32_utils.h b/tensorflow/core/platform/tf32_utils.h
new file mode 100644
index 00000000000..a0ce58f9bbd
--- /dev/null
+++ b/tensorflow/core/platform/tf32_utils.h
@@ -0,0 +1,27 @@
+/* Copyright 2020 The TensorFlow Authors. All Rights Reserved.
+
+Licensed under the Apache License, Version 2.0 (the "License");
+you may not use this file except in compliance with the License.
+You may obtain a copy of the License at
+
+    http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software
+distributed under the License is distributed on an "AS IS" BASIS,
+WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+See the License for the specific language governing permissions and
+limitations under the License.
+==============================================================================*/
+
+#ifndef TENSORFLOW_CORE_PLATFORM_TF32_UTILS_H_
+#define TENSORFLOW_CORE_PLATFORM_TF32_UTILS_H_
+
+namespace tensorflow {
+
+void allow_tf32_execution(bool allow);
+
+bool tf32_execution_allowed();
+
+}  // namespace tensorflow
+
+#endif  // TENSORFLOW_CORE_PLATFORM_TF32_UTILS_H_

From 16033c0b3484409a965acc0dd3054695145311a8 Mon Sep 17 00:00:00 2001
From: Nathan Luehr 
Date: Fri, 15 May 2020 13:33:02 -0500
Subject: [PATCH 341/867] Python tf.config tf32 interface

---
 tensorflow/python/BUILD               | 11 +++++++++++
 tensorflow/python/framework/config.py | 26 ++++++++++++++++++++++++++
 tensorflow/python/util/tf32.cc        | 22 ++++++++++++++++++++++
 3 files changed, 59 insertions(+)
 create mode 100644 tensorflow/python/util/tf32.cc

diff --git a/tensorflow/python/BUILD b/tensorflow/python/BUILD
index de9cf9a24c7..5f9e2dfb1ff 100644
--- a/tensorflow/python/BUILD
+++ b/tensorflow/python/BUILD
@@ -788,6 +788,16 @@ tf_python_pybind_extension(
     ],
 )
 
+tf_python_pybind_extension(
+    name = "_pywrap_tf32_execution",
+    srcs = ["util/tf32.cc"],
+    module_name = "_pywrap_tf32_execution",
+    deps = [
+        "//tensorflow/core/platform:tf32_utils",
+        "@pybind11",
+    ],
+)
+
 tf_python_pybind_extension(
     name = "_pywrap_util_port",
     srcs = ["util/port_wrapper.cc"],
@@ -5678,6 +5688,7 @@ py_library(
         "//tensorflow:composite_tensor_whitelist",
     ],
     deps = [
+        ":_pywrap_tf32_execution",
         ":tf_decorator",
         ":tf_export",
         ":tf_stack",
diff --git a/tensorflow/python/framework/config.py b/tensorflow/python/framework/config.py
index 9ff16f2a327..cb95965dfb2 100644
--- a/tensorflow/python/framework/config.py
+++ b/tensorflow/python/framework/config.py
@@ -18,10 +18,36 @@ from __future__ import absolute_import
 from __future__ import division
 from __future__ import print_function
 
+from tensorflow.python import _pywrap_tf32_execution
 from tensorflow.python.eager import context
 from tensorflow.python.util import deprecation
 from tensorflow.python.util.tf_export import tf_export
 
+def tensor_float32_execution_allowed():
+  """Get if TensorFloat-32 operations are enabled on supported hardware.
+
+  Returns:
+    True if TensorFloat-32 execution is enabled and False otherwise.
+  """
+  return _pywrap_tf32_execution.is_allowed()
+
+def allow_tensor_float_32_execution(allow):
+  """Allow use of TensorFloat-32 with float32 ops on supported hardware.
+
+  TensorFloat-32 is a math mode introduced with the NVIDIA Ampere architecture.
+  TensorFloat-32 kernels take float32 inputs and produce float32 outputs.
+  Internally, the inputs are cast to a custom representation with 10-bit
+  mantissa (similar to float16) and 8-bit exponent (similar to float32) and are
+  executed using TensorCores with float32 accumulation. For more information,
+  see https://blogs.nvidia.com/blog/2020/05/14/tensorfloat-32-precision-format/.
+
+  TensorFloat-32 execution is disabled by default, but this may change in a
+  future version.
+  
+  Args:
+    allow: whether to allow TensorFloat-32 execution
+  """
+  _pywrap_tf32_execution.allow(allow)
 
 @tf_export('config.threading.get_intra_op_parallelism_threads')
 def get_intra_op_parallelism_threads():
diff --git a/tensorflow/python/util/tf32.cc b/tensorflow/python/util/tf32.cc
new file mode 100644
index 00000000000..7dece6ccdae
--- /dev/null
+++ b/tensorflow/python/util/tf32.cc
@@ -0,0 +1,22 @@
+/* Copyright 2020 The TensorFlow Authors. All Rights Reserved.
+
+Licensed under the Apache License, Version 2.0 (the "License");
+you may not use this file except in compliance with the License.
+You may obtain a copy of the License at
+
+    http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software
+distributed under the License is distributed on an "AS IS" BASIS,
+WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+See the License for the specific language governing permissions and
+limitations under the License.
+==============================================================================*/
+
+#include "pybind11/pybind11.h"
+#include "tensorflow/core/platform/tf32_utils.h"
+
+PYBIND11_MODULE(_pywrap_tf32_execution, m) {
+  m.def("allow", &tensorflow::allow_tf32_execution);
+  m.def("is_allowed", &tensorflow::tf32_execution_allowed);
+}

From 376efd71b10eb7d7b900c3f7e7aff99bf15b0196 Mon Sep 17 00:00:00 2001
From: Nathan Luehr 
Date: Tue, 19 May 2020 14:58:30 -0500
Subject: [PATCH 342/867] Convolution TF32 Plumbing

---
 tensorflow/stream_executor/cuda/BUILD       |   1 +
 tensorflow/stream_executor/cuda/cuda_dnn.cc | 200 +++++++++++++-------
 2 files changed, 135 insertions(+), 66 deletions(-)

diff --git a/tensorflow/stream_executor/cuda/BUILD b/tensorflow/stream_executor/cuda/BUILD
index c3cf9f5db15..cdc0de7c72a 100644
--- a/tensorflow/stream_executor/cuda/BUILD
+++ b/tensorflow/stream_executor/cuda/BUILD
@@ -356,6 +356,7 @@ cc_library(
         "@local_config_cuda//cuda:cudnn_header",
         "//tensorflow/core:lib",
         "//tensorflow/core:lib_internal",
+        "//tensorflow/core/platform:tf32_utils",
         "//tensorflow/stream_executor:dnn",
         "//tensorflow/stream_executor:event",
         "//tensorflow/stream_executor:plugin_registry",
diff --git a/tensorflow/stream_executor/cuda/cuda_dnn.cc b/tensorflow/stream_executor/cuda/cuda_dnn.cc
index e46c271443b..53ba31d7d0d 100644
--- a/tensorflow/stream_executor/cuda/cuda_dnn.cc
+++ b/tensorflow/stream_executor/cuda/cuda_dnn.cc
@@ -20,8 +20,8 @@ limitations under the License.
 #include 
 
 #include "absl/strings/str_cat.h"
-#include "third_party/eigen3/Eigen/Core"
 #include "tensorflow/core/lib/core/errors.h"
+#include "tensorflow/core/platform/tf32_utils.h"
 #include "tensorflow/core/util/env_var.h"
 #include "tensorflow/stream_executor/cuda/cuda_activation.h"
 #include "tensorflow/stream_executor/cuda/cuda_diagnostics.h"
@@ -42,6 +42,7 @@ limitations under the License.
 #include "tensorflow/stream_executor/scratch_allocator.h"
 #include "tensorflow/stream_executor/stream.h"
 #include "tensorflow/stream_executor/stream_executor_pimpl.h"
+#include "third_party/eigen3/Eigen/Core"
 // clang-format off
 #include "third_party/gpus/cudnn/cudnn.h"
 #include "absl/strings/string_view.h"
@@ -705,10 +706,6 @@ class CudnnConvolutionDescriptor {
             : CUDNN_CROSS_CORRELATION,
         data_type));
 
-    // NOTE(benbarsdell): This only applies if tensor op math is enabled
-    //                      and algo selection is set to Default.
-    this->set_use_tensor_op_math(true);
-
 #if CUDNN_MAJOR >= 7
     VLOG(2) << "Requesting grouped convolution: "
             << convolution_descriptor.group_count();
@@ -720,10 +717,14 @@ class CudnnConvolutionDescriptor {
 #endif
   }
 
-  void set_use_tensor_op_math(bool use_tensor_op_math) const {
+  void set_use_tensor_op_math(bool use_tensor_op_math) {
 #if CUDNN_VERSION >= 7000
     cudnnMathType_t math_type =
+#if CUDNN_VERSION >= 8000
+        (use_tensor_op_math ? CUDNN_DEFAULT_MATH : CUDNN_FMA_MATH);
+#else
         (use_tensor_op_math ? CUDNN_TENSOR_OP_MATH : CUDNN_DEFAULT_MATH);
+#endif
     CHECK_CUDNN_OK(cudnnSetConvolutionMathType(handle_.get(), math_type));
 #endif
   }
@@ -736,6 +737,38 @@ class CudnnConvolutionDescriptor {
   SE_DISALLOW_COPY_AND_ASSIGN(CudnnConvolutionDescriptor);
 };
 
+// A helper function to query if a CudnnConvolutionDescriptor has tensor_op_math
+// set
+static bool IsTensorMathOpSet(const CudnnConvolutionDescriptor& conv) {
+  cudnnMathType_t math_type;
+  CHECK_CUDNN_OK(cudnnGetConvolutionMathType(conv.handle(), &math_type));
+#if CUDNN_VERSION >= 8000
+  return math_type != CUDNN_FMA_MATH;
+#else
+  return math_type == CUDNN_TENSOR_OP_MATH;
+#endif
+}
+
+static bool TensorOpMathAvailable(int cc_major) {
+  return cc_major >= 7 && CUDNN_VERSION >= 7000;
+}
+
+static bool IsTensorMathAllowed(Stream* stream, dnn::DataType input_type) {
+  int cc_major, cc_minor;
+  std::tie(cc_major, cc_minor) = GetCcMajorMinor(stream);
+  if (!TensorOpMathAvailable(cc_major)) {
+    return false;
+  }
+  if (input_type == dnn::DataType::kFloat) {
+    if (CUDNN_VERSION < 8000) {
+      return false;
+    } else if (!tensorflow::tf32_execution_allowed()) {
+      return false;
+    }
+  }
+  return true;
+}
+
 // Turns a PoolingDescriptor structure into a cudnn pooling descriptor handle
 // within a scope.
 class CudnnPoolingDescriptor {
@@ -2531,10 +2564,11 @@ port::StatusOr> AllocateCudnnConvolutionForwardWorkspace(
     const CudnnTensorDescriptor& output_nd,
     const dnn::AlgorithmDesc& algorithm_desc,
     ScratchAllocator* scratch_allocator) {
-  // TODO(csigg): This has side effects on the convolution descriptor. It is
-  // functionally correct because the convolution is run with the algorithm of
-  // the last call to this function, but should be fixed anyway.
-  conv.set_use_tensor_op_math(algorithm_desc.tensor_ops_enabled());
+  if (IsTensorMathOpSet(conv) != algorithm_desc.tensor_ops_enabled()) {
+    return port::Status(
+        port::error::INTERNAL,
+        "Mismatch between cudnn conv and algorithm descriptors.");
+  }
 
   // Query the size of the workspace and allocate it.
   size_t size_in_bytes;
@@ -2574,10 +2608,11 @@ AllocateCudnnConvolutionBackwardDataWorkspace(
     const CudnnTensorDescriptor& output_nd,
     const dnn::AlgorithmDesc& algorithm_desc,
     ScratchAllocator* scratch_allocator) {
-  // TODO(csigg): This has side effects on the convolution descriptor. It is
-  // functionally correct because the convolution is run with the algorithm of
-  // the last call to this function, but should be fixed anyway.
-  conv.set_use_tensor_op_math(algorithm_desc.tensor_ops_enabled());
+  if (IsTensorMathOpSet(conv) != algorithm_desc.tensor_ops_enabled()) {
+    return port::Status(
+        port::error::INTERNAL,
+        "Mismatch between cudnn conv and algorithm descriptors.");
+  }
 
   // Query the size of the workspace and allocate it.
   size_t size_in_bytes;
@@ -2619,10 +2654,11 @@ AllocateCudnnConvolutionBackwardFilterWorkspace(
     const CudnnTensorDescriptor& output_nd,
     const dnn::AlgorithmDesc& algorithm_desc,
     ScratchAllocator* scratch_allocator) {
-  // TODO(csigg): This has side effects on the convolution descriptor. It is
-  // functionally correct because the convolution is run with the algorithm of
-  // the last call to this function, but should be fixed anyway.
-  conv.set_use_tensor_op_math(algorithm_desc.tensor_ops_enabled());
+  if (IsTensorMathOpSet(conv) != algorithm_desc.tensor_ops_enabled()) {
+    return port::Status(
+        port::error::INTERNAL,
+        "Mismatch between cudnn conv and algorithm descriptors.");
+  }
 
   // Query the size of the workspace and allocate it.
   size_t size_in_bytes;
@@ -2656,18 +2692,39 @@ AllocateCudnnConvolutionBackwardFilterWorkspace(
   return scratch_allocator->AllocateBytes(size_in_bytes);
 }
 
-static bool TensorOpMathAvailable(int cc_major) {
-  return cc_major >= 7 && CUDNN_VERSION >= 7000;
+port::StatusOr UseTensorOps(Stream* stream, dnn::DataType type,
+                                  absl::optional desc) {
+  bool use_tensor_ops;
+  if (desc.has_value()) {
+    use_tensor_ops = desc->tensor_ops_enabled();
+    if (use_tensor_ops && !IsTensorMathAllowed(stream, type)) {
+      return port::Status(port::error::INVALID_ARGUMENT,
+                          "Algo requests disallowed tensor op evaluation.");
+    }
+  } else {
+    use_tensor_ops = IsTensorMathAllowed(stream, type);
+  }
+  return use_tensor_ops;
 }
 
 port::StatusOr GetCudnnConvolutionForwardAlgorithm(
     Stream* stream, const CudnnHandle& cudnn,
     const dnn::AlgorithmConfig& algorithm_config,
     const CudnnTensorDescriptor& input_nd, const CudnnFilterDescriptor& filter,
-    const CudnnConvolutionDescriptor& conv,
+    dnn::DataType element_type,
+    const dnn::ConvolutionDescriptor& convolution_descriptor,
     const CudnnTensorDescriptor& output_nd, ScratchAllocator* scratch_allocator,
     DeviceMemory* scratch) {
   absl::optional algo_desc = algorithm_config.algorithm();
+
+  CudnnConvolutionDescriptor conv(
+      convolution_descriptor,
+      ToCudnnDataType(GetConvAccumulatorType(element_type)));
+  bool use_tensor_ops;
+  SE_ASSIGN_OR_RETURN(use_tensor_ops,
+                      UseTensorOps(stream, element_type, algo_desc));
+  conv.set_use_tensor_op_math(use_tensor_ops);
+
   if (!algo_desc.has_value()) {
     // Pick fastest algorithm within memory limit according to cuDNN's
     // heuristics.
@@ -2680,10 +2737,7 @@ port::StatusOr GetCudnnConvolutionForwardAlgorithm(
                         GetCudnnConvolutionForwardAlgo(
                             cudnn, input_nd, filter, conv, output_nd,
                             specify_workspace_limit, memory_limit_bytes));
-    int cc_major, cc_minor;
-    std::tie(cc_major, cc_minor) = GetCcMajorMinor(stream);
-    algo_desc = dnn::AlgorithmDesc(
-        algo, /*use_tensor_ops=*/TensorOpMathAvailable(cc_major));
+    algo_desc = dnn::AlgorithmDesc(algo, use_tensor_ops);
   }
 
   const auto scratch_or = AllocateCudnnConvolutionForwardWorkspace(
@@ -2707,6 +2761,9 @@ port::StatusOr GetCudnnConvolutionForwardAlgorithm(
                      "Returned status: ", scratch_or.status().ToString()));
   }
 
+  SE_ASSIGN_OR_RETURN(use_tensor_ops,
+                      UseTensorOps(stream, element_type, algo_desc));
+  conv.set_use_tensor_op_math(use_tensor_ops);
   SE_ASSIGN_OR_RETURN(*scratch, AllocateCudnnConvolutionForwardWorkspace(
                                     stream, cudnn, input_nd, filter, conv,
                                     output_nd, *algo_desc, scratch_allocator));
@@ -2717,10 +2774,19 @@ port::StatusOr GetCudnnConvolutionBackwardDataAlgorithm(
     Stream* stream, const CudnnHandle& cudnn,
     const dnn::AlgorithmConfig& algorithm_config,
     const CudnnTensorDescriptor& input_nd, const CudnnFilterDescriptor& filter,
-    const CudnnConvolutionDescriptor& conv,
+    dnn::DataType element_type,
+    const dnn::ConvolutionDescriptor& convolution_descriptor,
     const CudnnTensorDescriptor& output_nd, ScratchAllocator* scratch_allocator,
     DeviceMemory* scratch) {
   absl::optional algo_desc = algorithm_config.algorithm();
+  CudnnConvolutionDescriptor conv(
+      convolution_descriptor,
+      ToCudnnDataType(GetConvAccumulatorType(element_type)));
+  bool use_tensor_ops;
+  SE_ASSIGN_OR_RETURN(use_tensor_ops,
+                      UseTensorOps(stream, element_type, algo_desc));
+  conv.set_use_tensor_op_math(use_tensor_ops);
+
   if (!algo_desc.has_value()) {
     // Pick fastest algorithm within memory limit according to cuDNN's
     // heuristics.
@@ -2733,10 +2799,7 @@ port::StatusOr GetCudnnConvolutionBackwardDataAlgorithm(
                         GetCudnnConvolutionBackwardDataAlgo(
                             cudnn, input_nd, filter, conv, output_nd,
                             specify_workspace_limit, memory_limit_bytes));
-    int cc_major, cc_minor;
-    std::tie(cc_major, cc_minor) = GetCcMajorMinor(stream);
-    algo_desc = dnn::AlgorithmDesc(
-        algo, /*use_tensor_ops=*/TensorOpMathAvailable(cc_major));
+    algo_desc = dnn::AlgorithmDesc(algo, use_tensor_ops);
   }
 
   const auto scratch_or = AllocateCudnnConvolutionBackwardDataWorkspace(
@@ -2759,6 +2822,9 @@ port::StatusOr GetCudnnConvolutionBackwardDataAlgorithm(
         "while a secondary algorithm is not provided.");
   }
 
+  SE_ASSIGN_OR_RETURN(use_tensor_ops,
+                      UseTensorOps(stream, element_type, algo_desc));
+  conv.set_use_tensor_op_math(use_tensor_ops);
   SE_ASSIGN_OR_RETURN(*scratch, AllocateCudnnConvolutionBackwardDataWorkspace(
                                     stream, cudnn, input_nd, filter, conv,
                                     output_nd, *algo_desc, scratch_allocator));
@@ -2769,10 +2835,19 @@ port::StatusOr GetCudnnConvolutionBackwardFilterAlgorithm(
     Stream* stream, const CudnnHandle& cudnn,
     const dnn::AlgorithmConfig& algorithm_config,
     const CudnnTensorDescriptor& input_nd, const CudnnFilterDescriptor& filter,
-    const CudnnConvolutionDescriptor& conv,
+    dnn::DataType element_type,
+    const dnn::ConvolutionDescriptor& convolution_descriptor,
     const CudnnTensorDescriptor& output_nd, ScratchAllocator* scratch_allocator,
     DeviceMemory* scratch) {
   absl::optional algo_desc = algorithm_config.algorithm();
+  CudnnConvolutionDescriptor conv(
+      convolution_descriptor,
+      ToCudnnDataType(GetConvAccumulatorType(element_type)));
+  bool use_tensor_ops;
+  SE_ASSIGN_OR_RETURN(use_tensor_ops,
+                      UseTensorOps(stream, element_type, algo_desc));
+  conv.set_use_tensor_op_math(use_tensor_ops);
+
   if (!algo_desc.has_value()) {
     // Pick fastest algorithm within memory limit according to cuDNN's
     // heuristics.
@@ -2785,10 +2860,7 @@ port::StatusOr GetCudnnConvolutionBackwardFilterAlgorithm(
                         GetCudnnConvolutionBackwardFilterAlgo(
                             cudnn, input_nd, filter, conv, output_nd,
                             specify_workspace_limit, memory_limit_bytes));
-    int cc_major, cc_minor;
-    std::tie(cc_major, cc_minor) = GetCcMajorMinor(stream);
-    algo_desc = dnn::AlgorithmDesc(
-        algo, /*use_tensor_ops=*/TensorOpMathAvailable(cc_major));
+    algo_desc = dnn::AlgorithmDesc(algo, use_tensor_ops);
   }
 
   auto scratch_or = AllocateCudnnConvolutionBackwardFilterWorkspace(
@@ -2811,6 +2883,9 @@ port::StatusOr GetCudnnConvolutionBackwardFilterAlgorithm(
         "while a secondary algorithm is not provided.");
   }
 
+  SE_ASSIGN_OR_RETURN(use_tensor_ops,
+                      UseTensorOps(stream, element_type, algo_desc));
+  conv.set_use_tensor_op_math(use_tensor_ops);
   SE_ASSIGN_OR_RETURN(*scratch, AllocateCudnnConvolutionBackwardFilterWorkspace(
                                     stream, cudnn, input_nd, filter, conv,
                                     output_nd, *algo_desc, scratch_allocator));
@@ -2975,35 +3050,32 @@ port::Status CudnnSupport::DoPrepareForConvolution(
   CudnnTensorDescriptor output_nd(
       output_descriptor,
       ToCudnnDataType(element_type, output_descriptor.layout()));
-  CudnnConvolutionDescriptor conv(
-      convolution_descriptor,
-      ToCudnnDataType(GetConvAccumulatorType(element_type)));
 
   auto cudnn = cudnn_->GetHandle(parent_, stream);
 
   switch (kind) {
     case dnn::ConvolutionKind::FORWARD: {
-      SE_ASSIGN_OR_RETURN(
-          *algorithm_desc,
-          GetCudnnConvolutionForwardAlgorithm(
-              stream, cudnn, algorithm_config, input_nd, filter_nd, conv,
-              output_nd, scratch_allocator, scratch_memory));
+      SE_ASSIGN_OR_RETURN(*algorithm_desc,
+                          GetCudnnConvolutionForwardAlgorithm(
+                              stream, cudnn, algorithm_config, input_nd,
+                              filter_nd, element_type, convolution_descriptor,
+                              output_nd, scratch_allocator, scratch_memory));
       break;
     }
     case dnn::ConvolutionKind::BACKWARD_DATA: {
-      SE_ASSIGN_OR_RETURN(
-          *algorithm_desc,
-          GetCudnnConvolutionBackwardDataAlgorithm(
-              stream, cudnn, algorithm_config, input_nd, filter_nd, conv,
-              output_nd, scratch_allocator, scratch_memory));
+      SE_ASSIGN_OR_RETURN(*algorithm_desc,
+                          GetCudnnConvolutionBackwardDataAlgorithm(
+                              stream, cudnn, algorithm_config, input_nd,
+                              filter_nd, element_type, convolution_descriptor,
+                              output_nd, scratch_allocator, scratch_memory));
       break;
     }
     case dnn::ConvolutionKind::BACKWARD_FILTER: {
-      SE_ASSIGN_OR_RETURN(
-          *algorithm_desc,
-          GetCudnnConvolutionBackwardFilterAlgorithm(
-              stream, cudnn, algorithm_config, input_nd, filter_nd, conv,
-              output_nd, scratch_allocator, scratch_memory));
+      SE_ASSIGN_OR_RETURN(*algorithm_desc,
+                          GetCudnnConvolutionBackwardFilterAlgorithm(
+                              stream, cudnn, algorithm_config, input_nd,
+                              filter_nd, element_type, convolution_descriptor,
+                              output_nd, scratch_allocator, scratch_memory));
       break;
     }
     default:
@@ -3032,8 +3104,9 @@ port::Status CudnnSupport::DoConvolve(
   auto accumulator_type = GetConvAccumulatorType(element_type);
   CudnnConvolutionDescriptor conv(convolution_descriptor,
                                   ToCudnnDataType(accumulator_type));
-  // Set use_tensor_math param to correct value
-  conv.set_use_tensor_op_math(algorithm_desc.tensor_ops_enabled());
+  SE_ASSIGN_OR_RETURN(bool use_tensor_ops,
+                      UseTensorOps(stream, element_type, algorithm_desc));
+  conv.set_use_tensor_op_math(use_tensor_ops);
 
   auto cudnn = cudnn_->GetHandle(parent_, stream);
   // Alpha is the scaling factor for input.
@@ -3266,14 +3339,6 @@ port::Status CudnnSupport::DoConvolve(
   return port::Status::OK();
 }
 
-// A helper function to query if a CudnnConvolutionDescriptor has tensor_op_math
-// set
-static bool IsTensorMathOpSet(const CudnnConvolutionDescriptor& conv) {
-  cudnnMathType_t math_type;
-  CHECK_CUDNN_OK(cudnnGetConvolutionMathType(conv.handle(), &math_type));
-  return math_type == CUDNN_TENSOR_OP_MATH;
-}
-
 template 
 port::Status CudnnSupport::DoFusedConvolveImpl(
@@ -3307,8 +3372,6 @@ port::Status CudnnSupport::DoFusedConvolveImpl(
       filter_descriptor,
       GetCudnnDataType(conv_input_descriptor.layout()));
   CudnnTensorDescriptor bias_nd(bias_descriptor, GetCudnnDataType());
-  CudnnConvolutionDescriptor conv(convolution_descriptor,
-                                  ToCudnnDataType(accumulator_type));
 
   auto cudnn = cudnn_->GetHandle(parent_, stream);
 
@@ -3318,9 +3381,14 @@ port::Status CudnnSupport::DoFusedConvolveImpl(
   SE_ASSIGN_OR_RETURN(
       dnn::AlgorithmDesc algo_desc,
       GetCudnnConvolutionForwardAlgorithm(
-          stream, cudnn, algorithm_config, conv_input_nd, filter, conv,
+          stream, cudnn, algorithm_config, conv_input_nd, filter,
+          dnn::ToDataType::value, convolution_descriptor,
           output_nd, scratch_allocator, &scratch));
 
+  CudnnConvolutionDescriptor conv(convolution_descriptor,
+                                  ToCudnnDataType(accumulator_type));
+  conv.set_use_tensor_op_math(algo_desc.tensor_ops_enabled());
+
   std::unique_ptr timer;
   if (is_profiling) {
     timer.reset(new GpuTimer(parent_));  // NOLINT

From 107e6348236d35a9fcdb4a1375ff07bf4975131c Mon Sep 17 00:00:00 2001
From: Nathan Luehr 
Date: Tue, 19 May 2020 15:54:10 -0500
Subject: [PATCH 343/867] Plumb TF32 for RNN

---
 tensorflow/stream_executor/cuda/cuda_dnn.cc | 30 ++++++++++++++-------
 1 file changed, 21 insertions(+), 9 deletions(-)

diff --git a/tensorflow/stream_executor/cuda/cuda_dnn.cc b/tensorflow/stream_executor/cuda/cuda_dnn.cc
index 53ba31d7d0d..820d39c7201 100644
--- a/tensorflow/stream_executor/cuda/cuda_dnn.cc
+++ b/tensorflow/stream_executor/cuda/cuda_dnn.cc
@@ -1161,17 +1161,26 @@ class CudnnRnnDescriptor : public dnn::RnnDescriptor {
     // in profile mode, which is run with algorithms returned from
     // GetRnnAlgorithms() (which are non-default and explicitly set whether to
     // use tensor ops). CuDNN 7.2.1 fixed this issue
-    cudnnMathType_t math_type;
+    bool allow_tensor_ops =
+        data_type != CUDNN_DATA_FLOAT || tensorflow::tf32_execution_allowed();
+    bool use_tensor_ops;
     if (algorithm_config.algorithm().has_value()) {
-      math_type = algorithm_config.algorithm()->tensor_ops_enabled()
-                      ? CUDNN_TENSOR_OP_MATH
-                      : CUDNN_DEFAULT_MATH;
+      use_tensor_ops = algorithm_config.algorithm()->tensor_ops_enabled();
     } else {
-#if CUDNN_VERSION >= 7201
-      math_type = CUDNN_TENSOR_OP_MATH;
-#else
-      math_type = CUDNN_DEFAULT_MATH;
-#endif  // CUDNN_VERSION >= 7201
+      use_tensor_ops = CUDNN_VERSION >= 7201 && allow_tensor_ops;
+    }
+
+    if (use_tensor_ops && !allow_tensor_ops) {
+      return port::Status(port::error::INVALID_ARGUMENT,
+                          "Algo requests disallowed tensor op evaluation.");
+    }
+
+    cudnnMathType_t math_type;
+    if (use_tensor_ops) {
+      math_type =
+          CUDNN_VERSION >= 8000 ? CUDNN_DEFAULT_MATH : CUDNN_TENSOR_OP_MATH;
+    } else {
+      math_type = CUDNN_VERSION >= 8000 ? CUDNN_FMA_MATH : CUDNN_DEFAULT_MATH;
     }
     CHECK_CUDNN_OK(cudnnSetRNNMatrixMathType(rnn_desc.get(), math_type));
 #endif  // CUDNN_VERSION >= 7000
@@ -2707,6 +2716,9 @@ port::StatusOr UseTensorOps(Stream* stream, dnn::DataType type,
   return use_tensor_ops;
 }
 
+cudnnDataType_t GetRnnComputeType(dnn::DataType data_type);
+dnn::DataType GetConvAccumulatorType(dnn::DataType data_type);
+
 port::StatusOr GetCudnnConvolutionForwardAlgorithm(
     Stream* stream, const CudnnHandle& cudnn,
     const dnn::AlgorithmConfig& algorithm_config,

From 03d836e2616574c657f7702e24e4fc79c661115b Mon Sep 17 00:00:00 2001
From: Nathan Luehr 
Date: Wed, 20 May 2020 10:06:35 -0500
Subject: [PATCH 344/867] Plumb TF32 for cublas gemm

---
 tensorflow/stream_executor/cuda/BUILD        |  1 +
 tensorflow/stream_executor/cuda/cuda_blas.cc | 84 +++++++++-----------
 tensorflow/stream_executor/cuda/cuda_blas.h  |  8 +-
 3 files changed, 43 insertions(+), 50 deletions(-)

diff --git a/tensorflow/stream_executor/cuda/BUILD b/tensorflow/stream_executor/cuda/BUILD
index cdc0de7c72a..3a14be9ad50 100644
--- a/tensorflow/stream_executor/cuda/BUILD
+++ b/tensorflow/stream_executor/cuda/BUILD
@@ -251,6 +251,7 @@ cc_library(
         "@local_config_cuda//cuda:cuda_headers",
         "//tensorflow/core:lib",
         "//tensorflow/core:lib_internal",
+        "//tensorflow/core/platform:tf32_utils",
         "//tensorflow/stream_executor",
         "//tensorflow/stream_executor:event",
         "//tensorflow/stream_executor:host_or_device_scalar",
diff --git a/tensorflow/stream_executor/cuda/cuda_blas.cc b/tensorflow/stream_executor/cuda/cuda_blas.cc
index 65c07e72154..e2cbb0b75df 100644
--- a/tensorflow/stream_executor/cuda/cuda_blas.cc
+++ b/tensorflow/stream_executor/cuda/cuda_blas.cc
@@ -48,7 +48,7 @@ limitations under the License.
 
 #include "absl/strings/str_cat.h"
 #include "absl/strings/str_format.h"
-#include "third_party/eigen3/Eigen/Core"
+#include "tensorflow/core/platform/tf32_utils.h"
 #include "tensorflow/core/util/env_var.h"
 #include "tensorflow/stream_executor/cuda/cuda_activation.h"
 #include "tensorflow/stream_executor/cuda/cuda_gpu_executor.h"
@@ -66,6 +66,7 @@ limitations under the License.
 #include "tensorflow/stream_executor/plugin_registry.h"
 #include "tensorflow/stream_executor/scratch_allocator.h"
 #include "tensorflow/stream_executor/stream_executor.h"
+#include "third_party/eigen3/Eigen/Core"
 
 namespace stream_executor {
 namespace gpu {
@@ -225,6 +226,18 @@ bool CUDABlas::Init() {
     return false;
   }
 
+#if CUDA_VERSION >= 9000
+#if CUBLAS_VER_MAJOR >= 11
+  ret = cublasSetMathMode(blas_, CUBLAS_TF32_TENSOR_OP_MATH);
+#else
+  ret = cublasSetMathMode(blas_, CUBLAS_TENSOR_OP_MATH);
+#endif
+  if (ret != CUBLAS_STATUS_SUCCESS) {
+    LOG(ERROR) << "failed to set cublas default math mode: " << ToString(ret);
+    return false;
+  }
+#endif
+
   return true;
 }
 
@@ -387,7 +400,7 @@ cudaDataType_t CUDAComputationType(blas::ComputationType ty) {
 template 
 bool CUDABlas::DoBlasInternalImpl(FuncT cublas_func, Stream *stream,
                                   bool pointer_mode_host, bool err_on_failure,
-                                  bool use_tensor_op_math, Args... args) {
+                                  Args... args) {
   absl::MutexLock lock(&mu_);
 
   CHECK(blas_ != nullptr);
@@ -401,10 +414,10 @@ bool CUDABlas::DoBlasInternalImpl(FuncT cublas_func, Stream *stream,
                                            : CUBLAS_POINTER_MODE_DEVICE)) {
     return false;
   }
-#if CUDA_VERSION >= 9000
+#if CUBLAS_VER_MAJOR >= 11
   ScopedCublasMathMode math_mode{blas_};
-  if (use_tensor_op_math) {
-    if (!math_mode.Init(CUBLAS_TENSOR_OP_MATH)) {
+  if (!tensorflow::tf32_execution_allowed()) {
+    if (!math_mode.Init(CUBLAS_DEFAULT_MATH)) {
       return false;
     }
   }
@@ -1621,21 +1634,9 @@ bool CUDABlas::DoBlasGemm(
     }
   }
 
-  bool use_tensor_ops = false;
-#if CUDA_VERSION >= 9000
-  int cc_major, cc_minor;
-  stream->parent()->GetDeviceDescription().cuda_compute_capability(&cc_major,
-                                                                   &cc_minor);
-
-  // GPUs < sm_70 don't support tensor ops.
-  if (cc_major >= 7) {
-    use_tensor_ops = true;
-  }
-#endif
-
   return DoBlasInternalImpl(
       cublasSgemmEx, stream, true /* = pointer_mode_host */,
-      true /* = err_on_failure= */, use_tensor_ops, CUDABlasTranspose(transa),
+      true /* = err_on_failure= */, CUDABlasTranspose(transa),
       CUDABlasTranspose(transb), m, n, k, &alpha, GpuMemory(a),
       SE_CUDA_DATA_HALF, lda, GpuMemory(b), SE_CUDA_DATA_HALF, ldb, &beta,
       GpuMemoryMutable(c), SE_CUDA_DATA_HALF, ldc);
@@ -2257,7 +2258,8 @@ port::Status CUDABlas::DoBlasGemmBatchedInternal(
   if (stream->parent()->GetDeviceDescription().cuda_compute_capability(
           &cc_major, &cc_minor) &&
       cc_major >= 5) {
-    bool use_tensor_ops = data_type == CUDA_R_16F;
+    bool use_tensor_ops =
+        data_type == CUDA_R_16F || tensorflow::tf32_execution_allowed();
     cublasGemmAlgo_t algo =
         (use_tensor_ops ? CUBLAS_GEMM_DFALT_TENSOR_OP : CUBLAS_GEMM_DFALT);
     cudaDataType_t compute_type =
@@ -2271,7 +2273,7 @@ port::Status CUDABlas::DoBlasGemmBatchedInternal(
     bool ok;
     ok = DoBlasInternalImpl(
         AS_LAMBDA(cublasGemmBatchedEx), stream, true /* = pointer_mode_host */,
-        true /* = err_on_failure */, use_tensor_ops, CUDABlasTranspose(transa),
+        true /* = err_on_failure */, CUDABlasTranspose(transa),
         CUDABlasTranspose(transb), m, n, k, &alpha, a_void_ptrs, data_type, lda,
         b_void_ptrs, data_type, ldb, &beta, c_void_ptrs, data_type, ldc,
         batch_count, compute_type, algo);
@@ -2406,33 +2408,25 @@ bool CUDABlas::DoBlasGemmStridedBatched(
     int lda, int64 stride_a, const DeviceMemory &b, int ldb,
     int64 stride_b, float beta, DeviceMemory *c, int ldc,
     int64 stride_c, int batch_count) {
-  bool use_tensor_ops = false;
-#if CUDA_VERSION >= 9000
+#if CUDA_VERSION >= 9010
   int cc_major, cc_minor;
   if (stream->parent()->GetDeviceDescription().cuda_compute_capability(
-          &cc_major, &cc_minor)) {
-    // GPUs < sm_70 don't support tensor ops.
-    if (cc_major >= 7) {
-      use_tensor_ops = true;
+          &cc_major, &cc_minor) &&
+      cc_major >= 5) {
+    cublasGemmAlgo_t algo =
+        (cc_major >= 7 ? CUBLAS_GEMM_DFALT_TENSOR_OP : CUBLAS_GEMM_DFALT);
+    bool ok = DoBlasInternalImpl(
+        AS_LAMBDA(cublasGemmStridedBatchedEx), stream,
+        true /* = pointer_mode_host */, true /* = err_on_failure */,
+        CUDABlasTranspose(transa), CUDABlasTranspose(transb), m, n, k, &alpha,
+        GpuMemory(a), CUDA_R_16F, lda, stride_a, GpuMemory(b), CUDA_R_16F, ldb,
+        stride_b, &beta, GpuMemoryMutable(c), CUDA_R_16F, ldc, stride_c,
+        batch_count, CUDA_R_32F, algo);
+    if (ok) {
+      return true;
     }
-#if CUDA_VERSION >= 9010
-    if (cc_major >= 5) {
-      cublasGemmAlgo_t algo =
-          (use_tensor_ops ? CUBLAS_GEMM_DFALT_TENSOR_OP : CUBLAS_GEMM_DFALT);
-      bool ok = DoBlasInternalImpl(
-          AS_LAMBDA(cublasGemmStridedBatchedEx), stream,
-          true /* = pointer_mode_host */, true /* = err_on_failure */,
-          use_tensor_ops, CUDABlasTranspose(transa), CUDABlasTranspose(transb),
-          m, n, k, &alpha, GpuMemory(a), CUDA_R_16F, lda, stride_a,
-          GpuMemory(b), CUDA_R_16F, ldb, stride_b, &beta, GpuMemoryMutable(c),
-          CUDA_R_16F, ldc, stride_c, batch_count, CUDA_R_32F, algo);
-      if (ok) {
-        return true;
-      }
-      LOG(ERROR) << "failed BLAS call, see log for details";
-      return false;
-    }
-#endif
+    LOG(ERROR) << "failed BLAS call, see log for details";
+    return false;
   }
 #endif
   // Either CUDA_VERSION < 9.1 or SM < 5.0. Fall back to a loop.
@@ -2445,7 +2439,7 @@ bool CUDABlas::DoBlasGemmStridedBatched(
         reinterpret_cast<__half *>(GpuMemoryMutable(c) + batch * stride_c);
     bool ok = DoBlasInternalImpl(
         cublasSgemmEx, stream, true /* = pointer_mode_host */,
-        true /* = err_on_failure= */, use_tensor_ops, CUDABlasTranspose(transa),
+        true /* = err_on_failure= */, CUDABlasTranspose(transa),
         CUDABlasTranspose(transb), m, n, k, &alpha, a_matrix, SE_CUDA_DATA_HALF,
         lda, b_matrix, SE_CUDA_DATA_HALF, ldb, &beta, c_matrix,
         SE_CUDA_DATA_HALF, ldc);
diff --git a/tensorflow/stream_executor/cuda/cuda_blas.h b/tensorflow/stream_executor/cuda/cuda_blas.h
index 817bdb72777..556456c83db 100644
--- a/tensorflow/stream_executor/cuda/cuda_blas.h
+++ b/tensorflow/stream_executor/cuda/cuda_blas.h
@@ -83,7 +83,7 @@ class CUDABlas : public blas::BlasSupport {
   template 
   bool DoBlasInternalImpl(FuncT cublas_func, Stream *stream,
                           bool pointer_mode_host, bool err_on_failure,
-                          bool use_tensor_op_math, Args... args);
+                          Args... args);
 
   // Convenience functions that call DoBlasInternalImpl with different values
   // for err_on_failure.
@@ -91,8 +91,7 @@ class CUDABlas : public blas::BlasSupport {
   bool DoBlasInternal(FuncT cublas_func, Stream *stream, bool pointer_mode_host,
                       Args... args) {
     return DoBlasInternalImpl(cublas_func, stream, pointer_mode_host,
-                              /*err_on_failure=*/true, /*use_tensor_ops=*/false,
-                              args...);
+                              /*err_on_failure=*/true, args...);
   }
   template 
   bool DoBlasInternalFailureOK(FuncT cublas_func, Stream *stream,
@@ -100,8 +99,7 @@ class CUDABlas : public blas::BlasSupport {
     // Tensor ops are hard-coded off in this path, but can still be enabled with
     // a specific algorithm choice as in DoBlasGemmWithAlgorithmImpl().
     return DoBlasInternalImpl(cublas_func, stream, pointer_mode_host,
-                              /*err_on_failure=*/false,
-                              /*use_tensor_ops=*/false, args...);
+                              /*err_on_failure=*/false, args...);
   }
 
   // A helper function to implement DoBlasGemmBatched interfaces for generic

From 7791e36a57b3d67e625e298a7e8a1063a26571fd Mon Sep 17 00:00:00 2001
From: Nathan Luehr 
Date: Mon, 8 Jun 2020 11:21:49 -0500
Subject: [PATCH 345/867] Address review comments

---
 tensorflow/core/platform/tf32_utils.cc | 10 ++++++----
 tensorflow/core/platform/tf32_utils.h  |  2 +-
 tensorflow/python/framework/config.py  |  7 +++++--
 3 files changed, 12 insertions(+), 7 deletions(-)

diff --git a/tensorflow/core/platform/tf32_utils.cc b/tensorflow/core/platform/tf32_utils.cc
index 715b5996dc3..4456e768c0a 100644
--- a/tensorflow/core/platform/tf32_utils.cc
+++ b/tensorflow/core/platform/tf32_utils.cc
@@ -14,14 +14,16 @@ limitations under the License.
 ==============================================================================*/
 
 #include "tensorflow/core/platform/tf32_utils.h"
+#include 
 
 namespace tensorflow {
 
-// TODO(nluehr): enable tf32 execution by default after TF32 Ampere testing.
-static bool tf32_enabled = false;
+// Whether TensorFloat-32 should be used where supported.
+// TODO(nluehr): Maybe enable by default after TF32 Ampere testing.
+static std::atomic tf32_allowed{false};
 
-void allow_tf32_execution(bool allow) { tf32_enabled = allow; }
+void allow_tf32_execution(bool allowed) { tf32_allowed = allowed; }
 
-bool tf32_execution_allowed() { return tf32_enabled; }
+bool tf32_execution_allowed() { return tf32_allowed; }
 
 }  // namespace tensorflow
diff --git a/tensorflow/core/platform/tf32_utils.h b/tensorflow/core/platform/tf32_utils.h
index a0ce58f9bbd..7a158d00ad3 100644
--- a/tensorflow/core/platform/tf32_utils.h
+++ b/tensorflow/core/platform/tf32_utils.h
@@ -18,7 +18,7 @@ limitations under the License.
 
 namespace tensorflow {
 
-void allow_tf32_execution(bool allow);
+void allow_tf32_execution(bool allowed);
 
 bool tf32_execution_allowed();
 
diff --git a/tensorflow/python/framework/config.py b/tensorflow/python/framework/config.py
index cb95965dfb2..e80ad1d72c4 100644
--- a/tensorflow/python/framework/config.py
+++ b/tensorflow/python/framework/config.py
@@ -23,6 +23,8 @@ from tensorflow.python.eager import context
 from tensorflow.python.util import deprecation
 from tensorflow.python.util.tf_export import tf_export
 
+
+# No tf_export until TF is built against CUDA11 which is required for TF32.
 def tensor_float32_execution_allowed():
   """Get if TensorFloat-32 operations are enabled on supported hardware.
 
@@ -31,7 +33,8 @@ def tensor_float32_execution_allowed():
   """
   return _pywrap_tf32_execution.is_allowed()
 
-def allow_tensor_float_32_execution(allow):
+# No tf_export until TF is built against CUDA11 which is required for TF32.
+def allow_tensor_float_32_execution(allowed):
   """Allow use of TensorFloat-32 with float32 ops on supported hardware.
 
   TensorFloat-32 is a math mode introduced with the NVIDIA Ampere architecture.
@@ -47,7 +50,7 @@ def allow_tensor_float_32_execution(allow):
   Args:
     allow: whether to allow TensorFloat-32 execution
   """
-  _pywrap_tf32_execution.allow(allow)
+  _pywrap_tf32_execution.allow(allowed)
 
 @tf_export('config.threading.get_intra_op_parallelism_threads')
 def get_intra_op_parallelism_threads():

From 6ff37474dd16d77f65d510a9f3bf0f56d08122b9 Mon Sep 17 00:00:00 2001
From: Nathan Luehr 
Date: Tue, 9 Jun 2020 15:35:47 -0500
Subject: [PATCH 346/867] Use CUDNN_TENSOR_OP_MATH to enable tensor cores.

---
 tensorflow/stream_executor/cuda/cuda_dnn.cc | 5 ++---
 1 file changed, 2 insertions(+), 3 deletions(-)

diff --git a/tensorflow/stream_executor/cuda/cuda_dnn.cc b/tensorflow/stream_executor/cuda/cuda_dnn.cc
index 820d39c7201..166fa0e32d0 100644
--- a/tensorflow/stream_executor/cuda/cuda_dnn.cc
+++ b/tensorflow/stream_executor/cuda/cuda_dnn.cc
@@ -721,7 +721,7 @@ class CudnnConvolutionDescriptor {
 #if CUDNN_VERSION >= 7000
     cudnnMathType_t math_type =
 #if CUDNN_VERSION >= 8000
-        (use_tensor_op_math ? CUDNN_DEFAULT_MATH : CUDNN_FMA_MATH);
+        (use_tensor_op_math ? CUDNN_TENSOR_OP_MATH : CUDNN_FMA_MATH);
 #else
         (use_tensor_op_math ? CUDNN_TENSOR_OP_MATH : CUDNN_DEFAULT_MATH);
 #endif
@@ -1177,8 +1177,7 @@ class CudnnRnnDescriptor : public dnn::RnnDescriptor {
 
     cudnnMathType_t math_type;
     if (use_tensor_ops) {
-      math_type =
-          CUDNN_VERSION >= 8000 ? CUDNN_DEFAULT_MATH : CUDNN_TENSOR_OP_MATH;
+      math_type = CUDNN_TENSOR_OP_MATH;
     } else {
       math_type = CUDNN_VERSION >= 8000 ? CUDNN_FMA_MATH : CUDNN_DEFAULT_MATH;
     }

From 8384080492a80b486fd000e3fbd6626cae0fb2cc Mon Sep 17 00:00:00 2001
From: Nathan Luehr 
Date: Thu, 11 Jun 2020 15:35:04 -0500
Subject: [PATCH 347/867] Make python names consistent

---
 tensorflow/python/framework/config.py | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/tensorflow/python/framework/config.py b/tensorflow/python/framework/config.py
index e80ad1d72c4..c19e514a932 100644
--- a/tensorflow/python/framework/config.py
+++ b/tensorflow/python/framework/config.py
@@ -34,7 +34,7 @@ def tensor_float32_execution_allowed():
   return _pywrap_tf32_execution.is_allowed()
 
 # No tf_export until TF is built against CUDA11 which is required for TF32.
-def allow_tensor_float_32_execution(allowed):
+def allow_tensor_float32_execution(allowed):
   """Allow use of TensorFloat-32 with float32 ops on supported hardware.
 
   TensorFloat-32 is a math mode introduced with the NVIDIA Ampere architecture.

From 5fe636fb4a74b58bd7da6de9ae1f286ad395c272 Mon Sep 17 00:00:00 2001
From: Reed 
Date: Thu, 11 Jun 2020 17:16:28 -0700
Subject: [PATCH 348/867] Use float_32 instead of float32 in function names

---
 tensorflow/python/framework/config.py | 4 ++--
 1 file changed, 2 insertions(+), 2 deletions(-)

diff --git a/tensorflow/python/framework/config.py b/tensorflow/python/framework/config.py
index c19e514a932..767a1a99f4f 100644
--- a/tensorflow/python/framework/config.py
+++ b/tensorflow/python/framework/config.py
@@ -25,7 +25,7 @@ from tensorflow.python.util.tf_export import tf_export
 
 
 # No tf_export until TF is built against CUDA11 which is required for TF32.
-def tensor_float32_execution_allowed():
+def tensor_float_32_execution_allowed():
   """Get if TensorFloat-32 operations are enabled on supported hardware.
 
   Returns:
@@ -34,7 +34,7 @@ def tensor_float32_execution_allowed():
   return _pywrap_tf32_execution.is_allowed()
 
 # No tf_export until TF is built against CUDA11 which is required for TF32.
-def allow_tensor_float32_execution(allowed):
+def allow_tensor_float_32_execution(allowed):
   """Allow use of TensorFloat-32 with float32 ops on supported hardware.
 
   TensorFloat-32 is a math mode introduced with the NVIDIA Ampere architecture.

From 434f49d54fc8165c2f019bd114272ce2200202da Mon Sep 17 00:00:00 2001
From: Nathan Luehr 
Date: Fri, 19 Jun 2020 15:44:04 -0500
Subject: [PATCH 349/867] Rework gemm TF32

---
 tensorflow/stream_executor/cuda/cuda_blas.cc | 142 +++++++++++++++----
 1 file changed, 113 insertions(+), 29 deletions(-)

diff --git a/tensorflow/stream_executor/cuda/cuda_blas.cc b/tensorflow/stream_executor/cuda/cuda_blas.cc
index e2cbb0b75df..e9e3635d8c1 100644
--- a/tensorflow/stream_executor/cuda/cuda_blas.cc
+++ b/tensorflow/stream_executor/cuda/cuda_blas.cc
@@ -226,18 +226,6 @@ bool CUDABlas::Init() {
     return false;
   }
 
-#if CUDA_VERSION >= 9000
-#if CUBLAS_VER_MAJOR >= 11
-  ret = cublasSetMathMode(blas_, CUBLAS_TF32_TENSOR_OP_MATH);
-#else
-  ret = cublasSetMathMode(blas_, CUBLAS_TENSOR_OP_MATH);
-#endif
-  if (ret != CUBLAS_STATUS_SUCCESS) {
-    LOG(ERROR) << "failed to set cublas default math mode: " << ToString(ret);
-    return false;
-  }
-#endif
-
   return true;
 }
 
@@ -1634,6 +1622,13 @@ bool CUDABlas::DoBlasGemm(
     }
   }
 
+#if CUDA_VERSION < 11000
+  ScopedCublasMathMode math_mode{blas_};
+  if (!math_mode.Init(CUBLAS_TENSOR_OP_MATH)) {
+    return false;
+  }
+#endif
+
   return DoBlasInternalImpl(
       cublasSgemmEx, stream, true /* = pointer_mode_host */,
       true /* = err_on_failure= */, CUDABlasTranspose(transa),
@@ -1681,6 +1676,16 @@ bool CUDABlas::DoBlasGemm(Stream *stream, blas::Transpose transa,
                       "precondition violation";
     }
   }
+
+#if CUBLAS_VER_MAJOR >= 11
+  ScopedCublasMathMode math_mode{blas_};
+  if (tensorflow::tf32_execution_allowed()) {
+    if (!math_mode.Init(CUBLAS_TF32_TENSOR_OP_MATH)) {
+      return false;
+    }
+  }
+#endif
+
   return DoBlasInternal(cublasSgemm, stream, true /* = pointer_mode_host */,
                         CUDABlasTranspose(transa), CUDABlasTranspose(transb), m,
                         n, k, &alpha, GpuMemory(a), lda, GpuMemory(b), ldb,
@@ -1707,6 +1712,16 @@ bool CUDABlas::DoBlasGemm(Stream *stream, blas::Transpose transa,
                           DeviceMemory> *c, int ldc) {
   auto cb_alpha = GpuComplexValue(alpha);
   auto cb_beta = GpuComplexValue(beta);
+
+#if CUBLAS_VER_MAJOR >= 11
+  ScopedCublasMathMode math_mode{blas_};
+  if (tensorflow::tf32_execution_allowed()) {
+    if (!math_mode.Init(CUBLAS_TF32_TENSOR_OP_MATH)) {
+      return false;
+    }
+  }
+#endif
+
   return DoBlasInternal(cublasCgemm, stream, true /* = pointer_mode_host */,
                         CUDABlasTranspose(transa), CUDABlasTranspose(transb), m,
                         n, k, GpuComplex(&cb_alpha), GpuComplex(GpuMemory(a)),
@@ -1903,20 +1918,6 @@ static bool UsesTensorOps(blas::AlgorithmType algo) {
 #endif
 }
 
-template 
-static bool TensorOpsAvailable(int cc_major) {
-#if CUDA_VERSION >= 9000
-  // cublas *does* allow tensor ops on inputs that are not fp16, so this is not
-  // strictly correct.  We can't simply enable it, though, as that would change
-  // clients' behavior significantly: Using tensor ops on fp32 inputs cause them
-  // to be rounded to fp16.
-  if (cc_major >= 7 && std::is_same::value) {
-    return true;
-  }
-#endif
-  return false;
-}
-
 template 
 bool CUDABlas::DoBlasGemmWithAlgorithmImpl(
     Stream *stream, blas::Transpose transa, blas::Transpose transb, uint64 m,
@@ -1935,17 +1936,52 @@ bool CUDABlas::DoBlasGemmWithAlgorithmImpl(
     return false;
   }
 
-  if (UsesTensorOps(algorithm) && !TensorOpsAvailable(cc_major)) {
-    if (std::is_same::value) {
+  bool algo_uses_tensor_ops = UsesTensorOps(algorithm);
+  cublasMath_t math_type = CUBLAS_DEFAULT_MATH;
+  if (algo_uses_tensor_ops) {
+    if (cc_major < 7) {
       VLOG(2) << "DoBlasGemmWithAlgorithm returning false because algorithm "
               << algorithm
               << " uses tensor ops, but tensor ops are not available in sm"
               << cc_major << "X devices.";
+      return false;
+    } else if (std::is_same::value) {
+#if CUDA_VERSION < 11000
+      VLOG(2) << "DoBlasGemmWithAlgorithm returning false because algorithm "
+              << algorithm
+              << " uses tensor ops, but tensor ops are not available for fp32"
+              << " inputs.";
+      return false;
+#else
+      if (cc_major < 8) {
+        VLOG(2) << "DoBlasGemmWithAlgorithm returning false because algorithm "
+                << algorithm
+                << " uses tensor ops, but tensor ops are not available in sm"
+                << cc_major << "X devices for float input types.";
+        return false;
+      } else if (!tensorflow::tf32_execution_allowed()) {
+        VLOG(2) << "DoBlasGemmWithAlgorithm returning false because algorithm "
+                << algorithm
+                << " uses tensor ops, but tensor ops are disabled for fp32"
+                << " inputs.";
+        return false;
+      }
+      math_type = CUBLAS_TF32_TENSOR_OP_MATH;
+#endif
+    } else if (std::is_same::value) {
+#if CUDA_VERSION < 11000
+      math_type = CUBLAS_TENSOR_OP_MATH;
+#endif
     } else {
       VLOG(2) << "DoBlasGemmWithAlgorithm returning false because algorithm "
               << algorithm
-              << " uses tensor ops, but the input data type is not fp16.";
+              << " uses tensor ops, which are not supported for InT.";
+      return false;
     }
+  }
+
+  ScopedCublasMathMode math_mode{blas_};
+  if (!math_mode.Init(math_type)) {
     return false;
   }
 
@@ -2325,6 +2361,13 @@ bool CUDABlas::DoBlasGemmBatched(
     int ldc, int batch_count, ScratchAllocator *scratch_allocator) {
   // Note: The func passed here (cublasSgemmBatched) is not actually called,
   // due to special handling of fp16 inside DoBlasGemmBatchedInternal.
+#if CUDA_VERSION < 11000
+  ScopedCublasMathMode math_mode{blas_};
+  if (!math_mode.Init(CUBLAS_TENSOR_OP_MATH)) {
+    return false;
+  }
+#endif
+
   port::Status status = DoBlasGemmBatchedInternal(
       cublasSgemmBatched, stream, transa, transb, m, n, k, alpha, a_array, lda,
       b_array, ldb, beta, c_array, ldc, batch_count, scratch_allocator);
@@ -2341,6 +2384,15 @@ bool CUDABlas::DoBlasGemmBatched(
     const port::ArraySlice *> &b_array, int ldb, float beta,
     const port::ArraySlice *> &c_array, int ldc,
     int batch_count, ScratchAllocator *scratch_allocator) {
+#if CUBLAS_VER_MAJOR >= 11
+  ScopedCublasMathMode math_mode{blas_};
+  if (tensorflow::tf32_execution_allowed()) {
+    if (!math_mode.Init(CUBLAS_TF32_TENSOR_OP_MATH)) {
+      return false;
+    }
+  }
+#endif
+
   port::Status status = DoBlasGemmBatchedInternal(
       cublasSgemmBatched, stream, transa, transb, m, n, k, alpha, a_array, lda,
       b_array, ldb, beta, c_array, ldc, batch_count, scratch_allocator);
@@ -2375,6 +2427,15 @@ bool CUDABlas::DoBlasGemmBatched(
     int ldb, std::complex beta,
     const port::ArraySlice> *> &c_array,
     int ldc, int batch_count, ScratchAllocator *scratch_allocator) {
+#if CUBLAS_VER_MAJOR >= 11
+  ScopedCublasMathMode math_mode{blas_};
+  if (tensorflow::tf32_execution_allowed()) {
+    if (!math_mode.Init(CUBLAS_TF32_TENSOR_OP_MATH)) {
+      return false;
+    }
+  }
+#endif
+
   port::Status status = DoBlasGemmBatchedInternal(
       cublasCgemmBatched, stream, transa, transb, m, n, k, alpha, a_array, lda,
       b_array, ldb, beta, c_array, ldc, batch_count, scratch_allocator);
@@ -2408,6 +2469,13 @@ bool CUDABlas::DoBlasGemmStridedBatched(
     int lda, int64 stride_a, const DeviceMemory &b, int ldb,
     int64 stride_b, float beta, DeviceMemory *c, int ldc,
     int64 stride_c, int batch_count) {
+#if CUDA_VERSION < 11000
+  ScopedCublasMathMode math_mode{blas_};
+  if (!math_mode.Init(CUBLAS_TENSOR_OP_MATH)) {
+    return false;
+  }
+#endif
+
 #if CUDA_VERSION >= 9010
   int cc_major, cc_minor;
   if (stream->parent()->GetDeviceDescription().cuda_compute_capability(
@@ -2457,6 +2525,14 @@ bool CUDABlas::DoBlasGemmStridedBatched(
     int64 stride_a, const DeviceMemory &b, int ldb, int64 stride_b,
     float beta, DeviceMemory *c, int ldc, int64 stride_c,
     int batch_count) {
+#if CUBLAS_VER_MAJOR >= 11
+  ScopedCublasMathMode math_mode{blas_};
+  if (tensorflow::tf32_execution_allowed()) {
+    if (!math_mode.Init(CUBLAS_TF32_TENSOR_OP_MATH)) {
+      return false;
+    }
+  }
+#endif
   return DoBlasInternal(
       cublasSgemmStridedBatched, stream, true /* = pointer_mode_host */,
       CUDABlasTranspose(transa), CUDABlasTranspose(transb), m, n, k, &alpha,
@@ -2484,6 +2560,14 @@ bool CUDABlas::DoBlasGemmStridedBatched(
     const DeviceMemory> &b, int ldb, int64 stride_b,
     std::complex beta, DeviceMemory> *c, int ldc,
     int64 stride_c, int batch_count) {
+#if CUBLAS_VER_MAJOR >= 11
+  ScopedCublasMathMode math_mode{blas_};
+  if (tensorflow::tf32_execution_allowed()) {
+    if (!math_mode.Init(CUBLAS_TF32_TENSOR_OP_MATH)) {
+      return false;
+    }
+  }
+#endif
   auto cb_alpha = GpuComplexValue(alpha);
   auto cb_beta = GpuComplexValue(beta);
   return DoBlasInternal(

From 90443c2b17a321f3589039dd560773b8d47a5cb7 Mon Sep 17 00:00:00 2001
From: Jacques Pienaar 
Date: Fri, 19 Jun 2020 15:15:51 -0700
Subject: [PATCH 350/867] Enable test for generating tf_ops

PiperOrigin-RevId: 317389351
Change-Id: I54f6ac53a974cf603ed5fe1a30b6fbb464c80d28
---
 tensorflow/compiler/mlir/tensorflow/BUILD | 5 +++++
 1 file changed, 5 insertions(+)

diff --git a/tensorflow/compiler/mlir/tensorflow/BUILD b/tensorflow/compiler/mlir/tensorflow/BUILD
index bcabb13d301..11a455eec72 100644
--- a/tensorflow/compiler/mlir/tensorflow/BUILD
+++ b/tensorflow/compiler/mlir/tensorflow/BUILD
@@ -57,6 +57,7 @@ gentbl(
     td_srcs = [
         ":tensorflow_ops_td_files",
     ],
+    test = True,
 )
 
 gentbl(
@@ -88,6 +89,7 @@ gentbl(
     td_srcs = [
         ":tensorflow_ops_td_files",
     ],
+    test = True,
 )
 
 gentbl(
@@ -112,6 +114,7 @@ gentbl(
         "@llvm-project//mlir:include/mlir/IR/OpBase.td",
         "@llvm-project//mlir:include/mlir/Dialect/StandardOps/IR/Ops.td",
     ],
+    test = True,
 )
 
 gentbl(
@@ -137,6 +140,7 @@ gentbl(
         "@llvm-project//mlir:include/mlir/Dialect/StandardOps/IR/Ops.td",
         "@llvm-project//mlir:include/mlir/IR/OpBase.td",
     ],
+    test = True,
 )
 
 gentbl(
@@ -161,6 +165,7 @@ gentbl(
         "@llvm-project//mlir:include/mlir/IR/OpBase.td",
         "@llvm-project//mlir:include/mlir/Dialect/StandardOps/IR/Ops.td",
     ],
+    test = True,
 )
 
 gentbl(

From fb5a3c3c6a4894a15bcc88ba9655c6afff8ff3f5 Mon Sep 17 00:00:00 2001
From: "A. Unique TensorFlower" 
Date: Fri, 19 Jun 2020 15:30:49 -0700
Subject: [PATCH 351/867] Fix issues caught by -Wtautological-constant-compare

This fixes a check that is always evaluates to true.  The check expression is a
conditional operator that returns 1 or 2, both of which convert to true.  This
fix puts in the enum value that causes the check to pass.

PiperOrigin-RevId: 317391929
Change-Id: I12f3f9b06f494b4bcf0e4ff194d4dc4edafd52e2
---
 tensorflow/core/kernels/hexagon/graph_transferer.cc | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/tensorflow/core/kernels/hexagon/graph_transferer.cc b/tensorflow/core/kernels/hexagon/graph_transferer.cc
index 7f15f3ab20d..9d6d0563b5f 100644
--- a/tensorflow/core/kernels/hexagon/graph_transferer.cc
+++ b/tensorflow/core/kernels/hexagon/graph_transferer.cc
@@ -667,7 +667,7 @@ void GraphTransferer::RegisterNodeWithPaddingAndStrides(
       << "Op " << node.type_string() << " not found in map(id = " << op_type_id
       << ")";
   // Safety check of padding id
-  CHECK(padding == Padding::VALID ? 1 : 2);
+  CHECK(padding == Padding::SAME);
   AppendNodeParamsWithIoParams(
       shape_refiner, node, node.name(), id, node.type_string(), op_type_id,
       static_cast(padding), node.num_inputs(), extra_inputs,

From ca7c0657c2e22343810957951cfb27aa5bf4c165 Mon Sep 17 00:00:00 2001
From: Yasir Modak <42785357+ymodak@users.noreply.github.com>
Date: Fri, 19 Jun 2020 15:43:04 -0700
Subject: [PATCH 352/867] update arg doc for sigmoid_cross_entropy

fixes #40593
---
 tensorflow/python/ops/losses/losses_impl.py | 5 +++--
 1 file changed, 3 insertions(+), 2 deletions(-)

diff --git a/tensorflow/python/ops/losses/losses_impl.py b/tensorflow/python/ops/losses/losses_impl.py
index 6a7b4b68420..2887b3d78f9 100644
--- a/tensorflow/python/ops/losses/losses_impl.py
+++ b/tensorflow/python/ops/losses/losses_impl.py
@@ -676,8 +676,9 @@ def sigmoid_cross_entropy(
       `{0, 1}`.
     logits: Float `[batch_size, num_classes]` logits outputs of the network.
     weights: Optional `Tensor` whose rank is either 0, or the same rank as
-      `labels`, and must be broadcastable to `labels` (i.e., all dimensions must
-      be either `1`, or the same as the corresponding `losses` dimension).
+    `multi_class_labels`, and must be broadcastable to `multi_class_labels` 
+    (i.e., all dimensions must be either `1`, or the same as the 
+    corresponding `losses` dimension).
     label_smoothing: If greater than `0` then smooth the labels.
     scope: The scope for the operations performed in computing the loss.
     loss_collection: collection to which the loss will be added.

From 18c43b7bd488f0216ce11127158df7084e26cdb4 Mon Sep 17 00:00:00 2001
From: Haoyu Zhang 
Date: Fri, 19 Jun 2020 15:43:22 -0700
Subject: [PATCH 353/867] Clear cancel callback when gRPC eager call returns
 with state.

PiperOrigin-RevId: 317393892
Change-Id: Ife800821494dd4cc2992eec9a5470d989596a6d7
---
 .../core/distributed_runtime/rpc/eager/grpc_eager_service_impl.h | 1 +
 1 file changed, 1 insertion(+)

diff --git a/tensorflow/core/distributed_runtime/rpc/eager/grpc_eager_service_impl.h b/tensorflow/core/distributed_runtime/rpc/eager/grpc_eager_service_impl.h
index 924112e0d96..1d65f945f27 100644
--- a/tensorflow/core/distributed_runtime/rpc/eager/grpc_eager_service_impl.h
+++ b/tensorflow/core/distributed_runtime/rpc/eager/grpc_eager_service_impl.h
@@ -81,6 +81,7 @@ class GrpcEagerServiceImpl : public AsyncServiceInterface {
       local_impl_.RunComponentFunction(call_opts.get(), &call->request,
                                        &call->response,
                                        [call, call_opts](const Status& s) {
+                                         call->ClearCancelCallback();
                                          call->SendResponse(ToGrpcStatus(s));
                                        });
     });

From e3809400b295e8a66fd8e2b20a348356c78a0435 Mon Sep 17 00:00:00 2001
From: Kuangyuan Chen 
Date: Fri, 19 Jun 2020 15:51:19 -0700
Subject: [PATCH 354/867] Import initialization graph in SignatureDef
 SavedModels as an MLIR function in TF saved model dialect. -Mark the init
 function referenced by SessionInitializerOp as an exported function with the
 reserved name "__tf_saved_model_session_initializer". -Remove variable init
 logic in init function if it is already imported as a global tensor. -Add a
 canonicalizer to SessionInitializerOp to remove empty init functions. -Return
 error on SessionInitializerOp in IREE compiler pipeline.

PiperOrigin-RevId: 317395165
Change-Id: Idb7e54cac08add9fc8f2ccfbaf341135fdf59e3b
---
 tensorflow/compiler/mlir/tensorflow/BUILD     |   4 +-
 .../mlir/tensorflow/ir/tf_saved_model.cc      |  75 +++++-
 .../mlir/tensorflow/ir/tf_saved_model.h       |   4 +
 .../mlir/tensorflow/ir/tf_saved_model_ops.td  |  26 ++
 .../tests/tf_saved_model/common_v1.py         |  12 +-
 .../tests/tf_saved_model/hash_table_v1.py     |  92 +++++++
 .../tf_saved_model/remove_init_variable_v1.py |  74 ++++++
 .../tensorflow/tests/tf_saved_model_ops.mlir  |  16 ++
 .../tests/tf_saved_model_ops_invalid.mlir     |  68 +++++
 .../mlir/tensorflow/translate/import_model.cc | 235 ++++++++++++++----
 10 files changed, 549 insertions(+), 57 deletions(-)
 create mode 100644 tensorflow/compiler/mlir/tensorflow/tests/tf_saved_model/hash_table_v1.py
 create mode 100644 tensorflow/compiler/mlir/tensorflow/tests/tf_saved_model/remove_init_variable_v1.py

diff --git a/tensorflow/compiler/mlir/tensorflow/BUILD b/tensorflow/compiler/mlir/tensorflow/BUILD
index 11a455eec72..b159815d5eb 100644
--- a/tensorflow/compiler/mlir/tensorflow/BUILD
+++ b/tensorflow/compiler/mlir/tensorflow/BUILD
@@ -666,7 +666,9 @@ cc_library(
         ":tensorflow_types",
         ":translate_utils",
         "//tensorflow/cc/saved_model:bundle_v2",
+        "//tensorflow/cc/saved_model:constants",
         "//tensorflow/cc/saved_model:loader_lite",
+        "//tensorflow/cc/saved_model:loader_util",
         "//tensorflow/compiler/jit:shape_inference_helpers",
         "//tensorflow/compiler/mlir:op_or_arg_name_mapper",
         "//tensorflow/compiler/tf2xla:functionalize_control_flow",
@@ -678,6 +680,7 @@ cc_library(
         "//tensorflow/core:lib",
         "//tensorflow/core:protos_all_cc",
         "//tensorflow/core/grappler/utils:transitive_fanin",
+        "//tensorflow/core/platform:protobuf_internal",
         "//tensorflow/core/platform:types",
         "//tensorflow/stream_executor/lib",
         "@com_google_absl//absl/algorithm:container",
@@ -687,7 +690,6 @@ cc_library(
         "@com_google_absl//absl/strings",
         "@com_google_absl//absl/types:optional",
         "@llvm-project//llvm:Support",
-        "@llvm-project//mlir:Analysis",
         "@llvm-project//mlir:IR",
         "@llvm-project//mlir:Pass",
         "@llvm-project//mlir:StandardOps",
diff --git a/tensorflow/compiler/mlir/tensorflow/ir/tf_saved_model.cc b/tensorflow/compiler/mlir/tensorflow/ir/tf_saved_model.cc
index 7db0eed7713..ef55761686e 100644
--- a/tensorflow/compiler/mlir/tensorflow/ir/tf_saved_model.cc
+++ b/tensorflow/compiler/mlir/tensorflow/ir/tf_saved_model.cc
@@ -28,6 +28,7 @@ limitations under the License.
 #include "mlir/IR/Identifier.h"  // from @llvm-project
 #include "mlir/IR/Module.h"  // from @llvm-project
 #include "mlir/IR/OpImplementation.h"  // from @llvm-project
+#include "mlir/IR/PatternMatch.h"  // from @llvm-project
 #include "mlir/IR/StandardTypes.h"  // from @llvm-project
 #include "mlir/IR/SymbolTable.h"  // from @llvm-project
 #include "mlir/IR/TypeUtilities.h"  // from @llvm-project
@@ -76,6 +77,23 @@ static LogicalResult Verify(GlobalTensorOp global_tensor) {
   return success();
 }
 
+static LogicalResult Verify(SessionInitializerOp session_initializer) {
+  mlir::SymbolTable symbol_table(
+      session_initializer.getParentOfType());
+
+  auto init_func_op =
+      symbol_table.lookup(session_initializer.initializer());
+  if (!init_func_op)
+    return session_initializer.emitOpError()
+           << "the initializer function does not exist";
+
+  if (!init_func_op.getType().getResults().empty())
+    return session_initializer.emitOpError()
+           << "the initializer function should have no output";
+
+  return success();
+}
+
 #define GET_OP_CLASSES
 #include "tensorflow/compiler/mlir/tensorflow/ir/tf_saved_model.cc.inc"
 
@@ -214,12 +232,12 @@ static LogicalResult VerifySavedModelModule(
   for (auto func : module.getOps()) {
     const bool is_exported = IsExported(func);
 
-    if (is_exported && !func.isPublic()) {
+    if (is_exported && func.getVisibility() != FuncOp::Visibility::Public) {
       return func.emitError()
              << "exported function @" << func.getName() << " should be public";
     }
 
-    if (!is_exported && func.isPublic()) {
+    if (!is_exported && func.getVisibility() == FuncOp::Visibility::Public) {
       return func.emitError() << "non-exported function @" << func.getName()
                               << " should be private";
     }
@@ -232,6 +250,19 @@ static LogicalResult VerifySavedModelModule(
       }
     }
   }
+
+  auto session_initializers = module.getOps();
+  if (!session_initializers.empty() &&
+      !llvm::hasSingleElement(session_initializers)) {
+    return (*++session_initializers.begin()).emitError()
+           << "there must be no more than one session_initializer op";
+  }
+
+  auto is_init = [&session_initializers](mlir::FuncOp func) {
+    if (session_initializers.empty()) return false;
+    return (*session_initializers.begin()).initializer() == func.getName();
+  };
+
   SymbolTable symbol_table(module);
   auto symbol_uses = SymbolTable::getSymbolUses(&module.getBodyRegion());
   if (!symbol_uses.hasValue()) {
@@ -242,6 +273,12 @@ static LogicalResult VerifySavedModelModule(
     auto func = symbol_table.lookup(
         symbol_use.getSymbolRef().cast().getValue());
     if (func && IsExported(func)) {
+      // If it is an init function, then it can be used by the unique
+      // session_initializer op.
+      if (is_init(func) &&
+          llvm::isa(symbol_use.getUser()))
+        continue;
+
       return symbol_use.getUser()
           ->emitError("exported function cannot be internally referenced")
           .attachNote(func.getLoc())
@@ -361,5 +398,39 @@ GlobalTensorOp LookupBoundInput(FuncOp func, int arg_index,
   return symbol_table.lookup(attr.getValue());
 }
 
+SessionInitializerOp GetSessionInitializerOp(mlir::ModuleOp op) {
+  auto initializers = op.getOps();
+  if (initializers.empty()) return {};
+  return *initializers.begin();
+}
+
+class OptimizeSessionInitializerPattern
+    : public OpRewritePattern {
+ public:
+  using OpRewritePattern::OpRewritePattern;
+
+  LogicalResult matchAndRewrite(SessionInitializerOp op,
+                                PatternRewriter &rewriter) const override {
+    SymbolTable symbol_table(op.getParentOfType());
+    auto init_func_op = symbol_table.lookup(op.initializer());
+
+    // The init function can only be referenced from the SessionInitializerOp.
+    // And there is at most one SessionInitializerOp in the module. So both ops
+    // have no other uses and can be simply erased.
+    if (init_func_op.front().begin()->isKnownTerminator()) {
+      rewriter.eraseOp(init_func_op);
+      rewriter.eraseOp(op);
+      return success();
+    }
+
+    return failure();
+  }
+};
+
+void SessionInitializerOp::getCanonicalizationPatterns(
+    OwningRewritePatternList &results, MLIRContext *context) {
+  results.insert(context);
+}
+
 }  // namespace tf_saved_model
 }  // namespace mlir
diff --git a/tensorflow/compiler/mlir/tensorflow/ir/tf_saved_model.h b/tensorflow/compiler/mlir/tensorflow/ir/tf_saved_model.h
index 47ebb1a1be5..b6f8753cc51 100644
--- a/tensorflow/compiler/mlir/tensorflow/ir/tf_saved_model.h
+++ b/tensorflow/compiler/mlir/tensorflow/ir/tf_saved_model.h
@@ -61,6 +61,10 @@ GlobalTensorOp LookupBoundInput(FuncOp func, int arg_index,
 // should have.
 Type GetBoundInputArgTypeFor(GlobalTensorOp global_tensor);
 
+// Returns the session initializer of this module if it exists. Returns null
+// otherwise.
+SessionInitializerOp GetSessionInitializerOp(mlir::ModuleOp op);
+
 }  // namespace tf_saved_model
 }  // namespace mlir
 
diff --git a/tensorflow/compiler/mlir/tensorflow/ir/tf_saved_model_ops.td b/tensorflow/compiler/mlir/tensorflow/ir/tf_saved_model_ops.td
index 4431a160edf..dc1210a4d2a 100644
--- a/tensorflow/compiler/mlir/tensorflow/ir/tf_saved_model_ops.td
+++ b/tensorflow/compiler/mlir/tensorflow/ir/tf_saved_model_ops.td
@@ -128,4 +128,30 @@ def TfSavedModel_GlobalTensorOp : TfSavedModel_Op<"global_tensor"> {
   let verifier = [{ return Verify(*this); }];
 }
 
+def TfSavedModel_SessionInitializerOp: TfSavedModel_Op<"session_initializer"> {
+  let summary = "Initializes TensorFlow session state.";
+  let description = [{
+    The session initializer op marks a function that must be called by an
+    external agent exactly once to initialize TensorFlow session state, and this
+    must happen before any other exported functions are called. There must be no
+    more than one session initializer in a saved model.
+
+    The `initializer` represents the initialization function. The function have
+    no output and this function should be only called once.
+
+    This is used, for example, to initialize hash tables stored in resources and
+    accessed by resource name (rather than as resource handles or bound inputs
+    which is how `global_tensor`s are referenced)
+  }];
+
+  let arguments = (ins
+    FlatSymbolRefAttr:$initializer
+  );
+
+
+  let verifier = [{ return Verify(*this); }];
+
+  let hasCanonicalizer = 1;
+}
+
 #endif // SAVED_MODEL_DIALECT
diff --git a/tensorflow/compiler/mlir/tensorflow/tests/tf_saved_model/common_v1.py b/tensorflow/compiler/mlir/tensorflow/tests/tf_saved_model/common_v1.py
index 7171f63bb05..5bfcfa5378a 100644
--- a/tensorflow/compiler/mlir/tensorflow/tests/tf_saved_model/common_v1.py
+++ b/tensorflow/compiler/mlir/tensorflow/tests/tf_saved_model/common_v1.py
@@ -46,7 +46,10 @@ def set_tf_options():
 # This function needs to take a "create_module_fn", as opposed to just the
 # module itself, because the creation of the module has to be delayed until
 # after absl and tensorflow have run various initialization steps.
-def do_test(signature_def_map, show_debug_info=False):
+def do_test(signature_def_map,
+            init_op=None,
+            canonicalize=False,
+            show_debug_info=False):
   """Runs test.
 
   1. Performs absl and tf "main"-like initialization that must run before almost
@@ -61,6 +64,9 @@ def do_test(signature_def_map, show_debug_info=False):
   Args:
     signature_def_map: A map from string key to signature_def. The key will be
       used as function name in the resulting MLIR.
+    init_op: The initializer op for the saved model. If set, it will generate a
+      initializer graph in the resulting MLIR.
+    canonicalize: If true, canonicalizer will be run on the resulting MLIR.
     show_debug_info: If true, shows debug locations in the resulting MLIR.
   """
 
@@ -84,6 +90,7 @@ def do_test(signature_def_map, show_debug_info=False):
     builder.add_meta_graph_and_variables(
         sess, [tf.saved_model.tag_constants.SERVING],
         signature_def_map,
+        main_op=init_op,
         strip_default_attrs=True)
     builder.save()
 
@@ -97,6 +104,9 @@ def do_test(signature_def_map, show_debug_info=False):
     mlir = pywrap_mlir.experimental_run_pass_pipeline(mlir,
                                                       'tf-standard-pipeline',
                                                       show_debug_info)
+    if canonicalize:
+      mlir = pywrap_mlir.experimental_run_pass_pipeline(mlir, 'canonicalize',
+                                                        show_debug_info)
     print(mlir)
 
   app.run(app_main)
diff --git a/tensorflow/compiler/mlir/tensorflow/tests/tf_saved_model/hash_table_v1.py b/tensorflow/compiler/mlir/tensorflow/tests/tf_saved_model/hash_table_v1.py
new file mode 100644
index 00000000000..16290455608
--- /dev/null
+++ b/tensorflow/compiler/mlir/tensorflow/tests/tf_saved_model/hash_table_v1.py
@@ -0,0 +1,92 @@
+# Copyright 2019 The TensorFlow Authors. All Rights Reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+# ==============================================================================
+
+# RUN: %p/hash_table_v1 | FileCheck %s
+
+# pylint: disable=missing-docstring,line-too-long
+from __future__ import absolute_import
+from __future__ import division
+from __future__ import print_function
+
+import tensorflow.compat.v1 as tf
+from tensorflow.compiler.mlir.tensorflow.tests.tf_saved_model import common_v1
+
+# Verify that the tf.versions attribute exists. It is difficult to enforce
+# contents, since the version numbers change over time. The conversion logic
+# itself is verified in the common graphdef converter, so here just assert
+# it is being invoked.
+# CHECK: module
+# CHECK-SAME: tf.versions
+# CHECK-SAME: bad_consumers
+# CHECK-SAME: min_consumer
+# CHECK-SAME: producer
+
+# CHECK: "tf_saved_model.session_initializer"() {initializer = [[init:@.*]]} : () -> ()
+# CHECK: "tf_saved_model.global_tensor"()
+
+# CHECK:      func [[init]]
+# CHECK-NEXT: [[R5:%.*]] = "tf.Const"()
+# CHECK-NEXT: [[R6:%.*]] = "tf.Const"()
+# CHECK-NEXT: [[R7:%.*]] = "tf.HashTableV2"()
+# CHECK-SAME: shared_name = "[[hash_table:.*]]"
+# CHECK-NEXT: "tf.LookupTableImportV2"([[R7]], [[R5]], [[R6]])
+
+# CHECK:      func {{@[a-zA-Z_0-9]+}}(
+# CHECK-SAME: [[ARG0:%.*]]: tensor
+# CHECK-SAME: [[ARG1:%.*]]: tensor, value = {{.*}} : tensor<1x3xf32>} : () -> ()
+# CHECK-NOT: session_initializer
+
+# CHECK:      func {{@[a-zA-Z_0-9]+}}(
+# CHECK-SAME:   [[ARG0:%.*]]: tensor<3x1xf32> {tf_saved_model.index_path = ["x"]},
+# CHECK-SAME:   [[ARG1:%.*]]: tensor>> {tf_saved_model.bound_input = @[[VAR]]})
+# CHECK-SAME:             -> (tensor<3x3xf32> {tf_saved_model.index_path = ["r"]})
+# CHECK-SAME: attributes {{.*}} tf_saved_model.exported_names = ["key"]
+
+# CHECK-NEXT: [[R0:%.*]] = "tf.ReadVariableOp"([[ARG1]]) {{{.*}}} : (tensor>>) -> tensor<1x3xf32>
+# CHECK-NEXT: [[R1:%.*]] = "tf.MatMul"([[ARG0]], [[R0]]) {{{.*}}} : (tensor<3x1xf32>, tensor<1x3xf32>) -> tensor<3x3xf32>
+# CHECK-NEXT: return [[R1]] : tensor<3x3xf32>
+
+
+def Test():
+
+  x = tf.constant([[1.0], [1.0], [1.0]])
+  y = tf.compat.v1.get_variable(
+      name='y',
+      shape=(1, 3),
+      initializer=tf.random_normal_initializer(),
+      trainable=True)
+  r = tf.matmul(x, y)
+
+  tensor_info_x = tf.compat.v1.saved_model.utils.build_tensor_info(x)
+  tensor_info_r = tf.compat.v1.saved_model.utils.build_tensor_info(r)
+
+  return {
+      'key': (tf.compat.v1.saved_model.signature_def_utils.build_signature_def(
+          inputs={'x': tensor_info_x},
+          outputs={'r': tensor_info_r},
+          method_name='some_function'))
+  }
+
+
+if __name__ == '__main__':
+  common_v1.set_tf_options()
+  common_v1.do_test(
+      Test(), tf.initializers.global_variables(), canonicalize=True)
diff --git a/tensorflow/compiler/mlir/tensorflow/tests/tf_saved_model_ops.mlir b/tensorflow/compiler/mlir/tensorflow/tests/tf_saved_model_ops.mlir
index e2dc5785cf4..26cdf025a10 100644
--- a/tensorflow/compiler/mlir/tensorflow/tests/tf_saved_model_ops.mlir
+++ b/tensorflow/compiler/mlir/tensorflow/tests/tf_saved_model_ops.mlir
@@ -2,6 +2,11 @@
 
 module attributes {tf_saved_model.semantics} {
 
+  // CHECK: tf_saved_model.session_initializer
+  "tf_saved_model.session_initializer"() {
+    initializer = @init
+  } : () -> ()
+
   // Representation for constants: (immutable) global tensor.
   // CHECK: tf_saved_model.global_tensor
   "tf_saved_model.global_tensor"() {
@@ -39,6 +44,17 @@ module attributes {tf_saved_model.semantics} {
     return
   }
 
+  // Representation for init functions
+  // CHECK: func @init
+  // CHECK-SAME: exported_names = ["__tf_saved_model_session_initializer"]
+  func @init(
+    %arg1: tensor>> {tf_saved_model.bound_input = @some_constant}
+  ) attributes {tf_saved_model.exported_names = ["__tf_saved_model_session_initializer"]}
+  {
+    "tf.some_call"(%arg1) : (tensor>>) -> ()
+    return
+  }
+
 }
 
 // -----
diff --git a/tensorflow/compiler/mlir/tensorflow/tests/tf_saved_model_ops_invalid.mlir b/tensorflow/compiler/mlir/tensorflow/tests/tf_saved_model_ops_invalid.mlir
index 7287fcf66c8..260174b184f 100644
--- a/tensorflow/compiler/mlir/tensorflow/tests/tf_saved_model_ops_invalid.mlir
+++ b/tensorflow/compiler/mlir/tensorflow/tests/tf_saved_model_ops_invalid.mlir
@@ -261,6 +261,39 @@ module attributes {tf_saved_model.semantics} {
 
 // -----
 
+module attributes {tf_saved_model.semantics} {
+
+  // expected-error@+1 {{the initializer function does not exist}}
+  "tf_saved_model.session_initializer"() { initializer = @init } : () -> ()
+}
+
+// -----
+
+module attributes {tf_saved_model.semantics} {
+
+  // expected-error@+1 {{the initializer function should have no output}}
+  "tf_saved_model.session_initializer"() { initializer = @init } : () -> ()
+  func @init() -> tensor<1xf32> attributes {sym_visibility = "private"} {
+    %0 = "tf.Const"() {value = dense<[1.0]> : tensor<1xf32> } : () -> tensor<1xf32>
+    return %0 : tensor<1xf32>
+  }
+}
+
+// -----
+
+module attributes {tf_saved_model.semantics} {
+
+  "tf_saved_model.session_initializer"() { initializer = @init } : () -> ()
+  // expected-error@+1 {{there must be no more than one session_initializer op}}
+  "tf_saved_model.session_initializer"() { initializer = @init } : () -> ()
+  func @init() -> tensor<1xf32> attributes {sym_visibility = "private"} {
+    %0 = "tf.Const"() {value = dense<[1.0]> : tensor<1xf32> } : () -> tensor<1xf32>
+    return %0 : tensor<1xf32>
+  }
+}
+
+// -----
+
 module attributes {tf_saved_model.semantics} {
 
   // expected-error@+1 {{exported function @f should be public}}
@@ -284,3 +317,38 @@ module attributes {tf_saved_model.semantics} {
   }
 
 }
+
+// -----
+
+module attributes {tf_saved_model.semantics} {
+
+  // expected-error@+1 {{the initializer function does not exist}}
+  "tf_saved_model.session_initializer"() { initializer = @init } : () -> ()
+}
+
+// -----
+
+module attributes {tf_saved_model.semantics} {
+
+  // expected-error@+1 {{the initializer function should have no output}}
+  "tf_saved_model.session_initializer"() { initializer = @init } : () -> ()
+  func @init() -> (tensor<1xf32> {tf_saved_model.index_path = ["output"]})
+    attributes { tf_saved_model.exported_names = ["__tf_saved_model_session_initializer"] } {
+    %0 = "tf.Const"() {value = dense<[1.0]> : tensor<1xf32> } : () -> tensor<1xf32>
+    return %0 : tensor<1xf32>
+  }
+}
+
+// -----
+
+module attributes {tf_saved_model.semantics} {
+
+  "tf_saved_model.session_initializer"() { initializer = @init } : () -> ()
+  // expected-error@+1 {{there must be no more than one session_initializer op}}
+  "tf_saved_model.session_initializer"() { initializer = @init } : () -> ()
+  func @init() -> (tensor<1xf32> {tf_saved_model.index_path = ["output"]})
+    attributes { tf_saved_model.exported_names = ["__tf_saved_model_session_initializer"] } {
+    %0 = "tf.Const"() {value = dense<[1.0]> : tensor<1xf32> } : () -> tensor<1xf32>
+    return %0 : tensor<1xf32>
+  }
+}
diff --git a/tensorflow/compiler/mlir/tensorflow/translate/import_model.cc b/tensorflow/compiler/mlir/tensorflow/translate/import_model.cc
index 820d0ce31fb..fea809c0798 100644
--- a/tensorflow/compiler/mlir/tensorflow/translate/import_model.cc
+++ b/tensorflow/compiler/mlir/tensorflow/translate/import_model.cc
@@ -60,6 +60,8 @@ limitations under the License.
 #include "mlir/IR/Types.h"  // from @llvm-project
 #include "mlir/IR/Verifier.h"  // from @llvm-project
 #include "mlir/Pass/PassManager.h"  // from @llvm-project
+#include "tensorflow/cc/saved_model/constants.h"
+#include "tensorflow/cc/saved_model/loader_util.h"
 #include "tensorflow/compiler/jit/shape_inference_helpers.h"
 #include "tensorflow/compiler/mlir/op_or_arg_name_mapper.h"
 #include "tensorflow/compiler/mlir/tensorflow/ir/tf_attributes.h"
@@ -116,6 +118,7 @@ using mlir::NamedAttrList;
 using mlir::TensorType;
 using mlir::TF::VarHandleOp;
 using mlir::tf_saved_model::GlobalTensorOp;
+using mlir::tf_saved_model::SessionInitializerOp;
 using stream_executor::port::StatusOr;
 
 namespace {
@@ -2955,6 +2958,13 @@ void SortSavedModelModule(mlir::ModuleOp module) {
     named_global_tensor.global_tensor.getOperation()->moveBefore(
         &module.getBody()->front());
   }
+
+  auto initializers = module.getOps();
+  if (!initializers.empty()) {
+    (*initializers.begin())
+        .getOperation()
+        ->moveBefore(&module.getBody()->front());
+  }
 }
 
 Status CreateSavedModelIR(
@@ -3241,17 +3251,32 @@ class SavedModelSignatureDefImporter {
                                  absl::Span exported_names,
                                  mlir::MLIRContext* context)
       : bundle_(bundle),
+        flib_def_(OpRegistry::Global(), graph_def().library()),
+        debug_info_(),
         exported_names_(exported_names),
-        module_(mlir::ModuleOp::create(mlir::UnknownLoc::get(context))) {}
+        module_(mlir::ModuleOp::create(mlir::UnknownLoc::get(context))) {
+    // debug_info might not be loaded with loader_lite.
+    if (bundle_.debug_info != nullptr) debug_info_ = *bundle_.debug_info;
+  }
 
   // Converts the SavedModel to the SavedModel dialect. Creates an MLIR function
   // for each signature.
   StatusOr ConvertSignatures();
-  Status ConvertSignature(const GraphDef& graphdef,
-                          const std::string& sig_def_key,
-                          const SignatureDef& signature_def,
-                          const GraphDebugInfo& debug_info,
-                          const FunctionLibraryDefinition& flib_def);
+  Status ConvertSignature(const std::string& sig_def_key,
+                          const SignatureDef& signature_def);
+
+  // Converts the initialization graph in the SavedModel to an MLIR function.
+  Status ConvertInitializer();
+
+  // Converts a graph with feeds and fetches to an MLIR function.
+  StatusOr ConvertGraph(
+      const std::string& name,
+      const std::vector>& inputs,
+      const std::vector>& outputs,
+      const std::vector control_outputs);
+
+  // Coarsens the islands in `module_`.
+  Status CoarsenIslands();
 
   // Creates GlobalTensorOp for each variable and moves each VarHandle op to
   // the enclosing function's arguments.
@@ -3262,6 +3287,10 @@ class SavedModelSignatureDefImporter {
   // tensor's shape is used to provide the most accurate nested shape.
   void LiftVariable(VarHandleOp op, GlobalTensorOp global_tensor);
 
+  // Removes the variable and related ops in the init function if it is already
+  // imported as a global tensor.
+  void RemoveVariable(VarHandleOp op);
+
   using VarGlobalMap = llvm::MapVector<
       llvm::StringRef,
       std::pair>>;
@@ -3273,18 +3302,68 @@ class SavedModelSignatureDefImporter {
   GraphImportConfig::InputArrays ParseInputArrays(
       const std::vector>& inputs);
 
+  const GraphDef& graph_def() const {
+    return bundle_.meta_graph_def.graph_def();
+  }
+  const FunctionLibraryDefinition& flib_def() const { return flib_def_; }
+  const GraphDebugInfo& debug_info() const { return debug_info_; }
+
   const SavedModelBundle& bundle_;
+  FunctionLibraryDefinition flib_def_;
+  GraphDebugInfo debug_info_;
   absl::Span exported_names_;
   mlir::OwningModuleRef module_;
 };
 
+Status SavedModelSignatureDefImporter::ConvertInitializer() {
+  std::vector asset_file_defs;
+  TF_RETURN_IF_ERROR(
+      internal::GetAssetFileDefs(bundle_.meta_graph_def, &asset_file_defs));
+
+  if (!asset_file_defs.empty())
+    return errors::Unimplemented(
+        absl::StrCat("Assets are not supported in signaturedef importer"));
+
+  std::string init_node_name;
+  TF_RETURN_IF_ERROR(
+      internal::GetInitOp("", bundle_.meta_graph_def, &init_node_name));
+
+  if (init_node_name.empty()) return Status::OK();
+
+  TF_ASSIGN_OR_RETURN(auto sub_module,
+                      ConvertGraph(init_node_name, {}, {}, {init_node_name}));
+
+  mlir::SymbolTable symbol_table(*sub_module);
+
+  auto init_func_op = symbol_table.lookup(init_node_name);
+
+  init_func_op.removeAttr("tf.entry_function");
+
+  mlir::OpBuilder builder(module_->getBodyRegion());
+
+  // Set the exported name of init function to an reserved name for
+  // tf_saved_model.
+  init_func_op.setAttr(
+      "tf_saved_model.exported_names",
+      builder.getStrArrayAttr({"__tf_saved_model_session_initializer"}));
+
+  builder.create(
+      module_->getLoc(), builder.getSymbolRefAttr(init_func_op.getName()));
+
+  // Move the converted functions to top level MLIR module.
+  auto* block = module_->getBody();
+  auto* sub_block = sub_module->getBody();
+  block->getOperations().splice(
+      mlir::Block::iterator(block->getTerminator()), sub_block->getOperations(),
+      sub_block->begin(), mlir::Block::iterator(sub_block->getTerminator()));
+
+  return Status::OK();
+}
+
 StatusOr
 SavedModelSignatureDefImporter::ConvertSignatures() {
   const auto& signatures = bundle_.GetSignatures();
-  const auto& graphdef = bundle_.meta_graph_def.graph_def();
-  PopulateTfVersions(module_.get(), graphdef.versions());
-
-  FunctionLibraryDefinition flib_def(OpRegistry::Global(), graphdef.library());
+  PopulateTfVersions(module_.get(), graph_def().versions());
 
   // debug_info might not be loaded with loader_lite.
   GraphDebugInfo debug_info;
@@ -3307,23 +3386,49 @@ SavedModelSignatureDefImporter::ConvertSignatures() {
       continue;
     }
 
-    TF_RETURN_IF_ERROR(ConvertSignature(graphdef, sig_def_key, signature_def,
-                                        debug_info, flib_def));
+    TF_RETURN_IF_ERROR(ConvertSignature(sig_def_key, signature_def));
   }
-  TF_RETURN_IF_ERROR(LiftVariables());
+
+  TF_RETURN_IF_ERROR(ConvertInitializer());
 
   mlir::OpBuilder builder(module_->getBodyRegion());
   module_->setAttr("tf_saved_model.semantics", builder.getUnitAttr());
+
+  TF_RETURN_IF_ERROR(CoarsenIslands());
+  TF_RETURN_IF_ERROR(LiftVariables());
+
   SortSavedModelModule(*module_);
   MarkSavedModelFunctionVisibility(*module_);
 
   return std::move(module_);
 }
 
+StatusOr SavedModelSignatureDefImporter::ConvertGraph(
+    const std::string& name,
+    const std::vector>& inputs,
+    const std::vector>& outputs,
+    const std::vector control_outputs) {
+  GraphImportConfig specs;
+  specs.prune_unused_nodes = true;
+  specs.inputs = ParseInputArrays(inputs);
+  for (auto& output : outputs) specs.outputs.push_back(output.second.name());
+  specs.control_outputs = control_outputs;
+
+  // Convert sub-graphdef to sub-graph.
+  GraphConstructorOptions options;
+  options.allow_internal_ops = true;
+  options.add_default_attributes = true;
+  Graph graph(OpRegistry::Global());
+
+  TF_RETURN_IF_ERROR(ConvertGraphDefToGraph(options, graph_def(), &graph));
+
+  // Convert sub-graph to MLIR module.true
+  return GraphDefImporter::Convert(module_->getContext(), graph, debug_info(),
+                                   flib_def(), specs, name);
+}
+
 Status SavedModelSignatureDefImporter::ConvertSignature(
-    const GraphDef& graphdef, const std::string& sig_def_key,
-    const SignatureDef& signature_def, const GraphDebugInfo& debug_info,
-    const FunctionLibraryDefinition& flib_def) {
+    const std::string& sig_def_key, const SignatureDef& signature_def) {
   // Create local vectors for the input and output and sort them to be
   // deterministic. We don't want anyone to really depend on the order, client
   // should lookup argument/result mapping by attribute name.
@@ -3339,34 +3444,9 @@ Status SavedModelSignatureDefImporter::ConvertSignature(
     return lhs.first.size() < rhs.first.size() || lhs.first > rhs.first;
   });
 
-  GraphImportConfig specs;
-  specs.prune_unused_nodes = true;
-  specs.inputs = ParseInputArrays(inputs);
-  for (auto& output : outputs) specs.outputs.push_back(output.second.name());
-
-  // Remove unused nodes and create sub-graphdef.
-  GraphDef sub_graph_def;
-  TF_RETURN_IF_ERROR(tensorflow::grappler::SetTransitiveFaninGraph(
-      graphdef, &sub_graph_def,
-      /*terminal_nodes=*/{specs.outputs.begin(), specs.outputs.end()}));
-
-  // Set the function library definitions in the pruned graphdef.
-  *sub_graph_def.mutable_library() = flib_def.ToProto();
-
-  // Convert sub-graphdef to sub-graph.
-  GraphConstructorOptions options;
-  options.allow_internal_ops = true;
-  options.add_default_attributes = true;
-  Graph sub_graph(OpRegistry::Global());
-
-  TF_RETURN_IF_ERROR(
-      ConvertGraphDefToGraph(options, sub_graph_def, &sub_graph));
-
   // Convert sub-graph to MLIR module.
-  TF_ASSIGN_OR_RETURN(
-      auto sub_module,
-      GraphDefImporter::Convert(module_->getContext(), sub_graph, debug_info,
-                                flib_def, specs, sig_def_key));
+  TF_ASSIGN_OR_RETURN(auto sub_module,
+                      ConvertGraph(sig_def_key, inputs, outputs, {}));
   mlir::OpBuilder builder(sub_module->getBodyRegion());
 
   // Find the FuncOp which corresponds to current SignatureDef.
@@ -3399,16 +3479,28 @@ Status SavedModelSignatureDefImporter::ConvertSignature(
       sub_block->begin(), mlir::Block::iterator(sub_block->getTerminator()));
 
   return Status::OK();
-}
+}  // namespace
 
 Status SavedModelSignatureDefImporter::LiftVariables() {
   VarGlobalMap var_globals;
+  llvm::SmallVector init_vars;
 
-  auto walker = [&var_globals](mlir::Operation* op) {
-    if (auto var_handle_op = llvm::dyn_cast(op))
-      var_globals[var_handle_op.shared_name()].second.push_back(var_handle_op);
-    else if (op->getName().getStringRef() == "tf.VariableV2")
+  auto session_initializer =
+      mlir::tf_saved_model::GetSessionInitializerOp(*module_);
+
+  auto walker = [&var_globals, &init_vars,
+                 &session_initializer](mlir::Operation* op) {
+    if (auto var_handle_op = llvm::dyn_cast(op)) {
+      if (session_initializer &&
+          session_initializer.initializer() ==
+              var_handle_op.getParentOfType().getName())
+        init_vars.push_back(var_handle_op);
+      else
+        var_globals[var_handle_op.shared_name()].second.push_back(
+            var_handle_op);
+    } else if (op->getName().getStringRef() == "tf.VariableV2") {
       return mlir::WalkResult::interrupt();
+    }
     return mlir::WalkResult::advance();
   };
   bool contains_ref_variable = module_->walk(walker).wasInterrupted();
@@ -3425,9 +3517,51 @@ Status SavedModelSignatureDefImporter::LiftVariables() {
     for (VarHandleOp var_handle : it.second.second)
       LiftVariable(var_handle, it.second.first);
 
+  for (auto op : init_vars) RemoveVariable(op);
+
   return Status::OK();
 }
 
+Status SavedModelSignatureDefImporter::CoarsenIslands() {
+  mlir::StatusScopedDiagnosticHandler diag_handler(module_->getContext());
+
+  mlir::PassManager pm(module_->getContext());
+  pm.addNestedPass(
+      mlir::tf_executor::CreateTFExecutorIslandCoarseningPass());
+  if (mlir::failed(pm.run(*module_)))
+    return diag_handler.Combine(
+        errors::Internal("failed to coarsening islands."));
+
+  return Status::OK();
+}
+
+void SavedModelSignatureDefImporter::RemoveVariable(VarHandleOp op) {
+  llvm::SmallVector work_list;
+  work_list.push_back(op);
+  while (!work_list.empty()) {
+    auto* op = work_list.back();
+    work_list.pop_back();
+
+    for (mlir::Value res : op->getResults()) {
+      for (mlir::Operation* user : res.getUsers()) {
+        work_list.push_back(user);
+      }
+    }
+
+    for (auto& use : op->getOpOperands()) {
+      if (mlir::Value value = use.get()) {
+        mlir::Operation* def = value.getDefiningOp();
+        work_list.push_back(def);
+      }
+    }
+
+    op->dropAllReferences();
+    op->dropAllDefinedValueUses();
+
+    op->erase();
+  }
+}
+
 void SavedModelSignatureDefImporter::LiftVariable(
     VarHandleOp op, GlobalTensorOp global_tensor) {
   mlir::OpBuilder builder(&module_->getBodyRegion());
@@ -3460,12 +3594,7 @@ void SavedModelSignatureDefImporter::LiftVariable(
   // Add the newly added function param to entry block's arguments.
   auto new_value = func_op.front().addArgument(resource_type);
 
-  // Remove the VarHandleOp also updating the containing island's return type.
-  DCHECK(llvm::isa(op.getParentOp()));
-  DCHECK(llvm::cast(op.getParentOp())
-             .WrapsSingleOp());
   op.getOperation()->replaceAllUsesWith(llvm::ArrayRef(new_value));
-  op.getParentOp()->getResult(0).setType(resource_type);
   op.getOperation()->erase();
 }
 

From 1823f877359bb138c57a005c30aba8832dfa79fb Mon Sep 17 00:00:00 2001
From: "A. Unique TensorFlower" 
Date: Fri, 19 Jun 2020 15:55:58 -0700
Subject: [PATCH 355/867] Fix issues with `TypeIndex` on MacOS, i.e. hash on
 the type name where available since this otherwise causes problems when
 loading different shared libraries with `RTLD_LOCAL`.

PiperOrigin-RevId: 317395983
Change-Id: I14b3add5fa19725b2150b68813364d16b8320130
---
 tensorflow/core/framework/type_index.h | 38 ++++++++++++++++++++------
 1 file changed, 30 insertions(+), 8 deletions(-)

diff --git a/tensorflow/core/framework/type_index.h b/tensorflow/core/framework/type_index.h
index fd27d8bcb35..fcf68677a12 100644
--- a/tensorflow/core/framework/type_index.h
+++ b/tensorflow/core/framework/type_index.h
@@ -24,6 +24,10 @@ limitations under the License.
 
 #include "tensorflow/core/platform/types.h"
 
+#if defined(MACOS) || defined(TARGET_OS_MAC)
+#include "tensorflow/core/platform/hash.h"
+#endif  // defined(MACOS) || defined(TARGET_OS_MAC)
+
 namespace tensorflow {
 
 // On some platforms, we would like to avoid using RTTI in order to have smaller
@@ -53,10 +57,33 @@ class TypeIndex {
 
   // Returns a TypeIndex object that corresponds to a typename.
   template 
-  static TypeIndex Make(const char* name) {
+  static TypeIndex Make() {
     static bool hash_bit[1];
+
+#if defined(__GXX_RTTI) || defined(_CPPRTTI)
+
+#if defined(MACOS) || defined(TARGET_OS_MAC)
+    // Use a hash based on the type name to avoid issues due to RTLD_LOCAL on
+    // MacOS (b/156979412).
+    return TypeIndex(Hash64(typeid(T).name()), typeid(T).name());
+#else
+    // Use the real type name if we have RTTI.
     return TypeIndex(static_cast(reinterpret_cast(hash_bit)),
-                     name);
+                     typeid(T).name());
+#endif  // defined(MACOS) || defined(TARGET_OS_MAC)
+
+#else
+#if defined(MACOS) || defined(TARGET_OS_MAC)
+    // Warn MacOS users that not using RTTI can cause problems (b/156979412).
+#warning \
+    "Compiling with RTTI disabled on MacOS can cause problems when comparing " \
+    "types across shared libraries."
+#endif  // defined(MACOS) || defined(TARGET_OS_MAC)
+
+    // No type names available.
+    return TypeIndex(static_cast(reinterpret_cast(hash_bit)),
+                     "[RTTI disabled]");
+#endif  // __GXX_RTTI
   }
 
  private:
@@ -70,12 +97,7 @@ class TypeIndex {
 
 template 
 inline TypeIndex MakeTypeIndex() {
-#if defined(__GXX_RTTI) || defined(_CPPRTTI)
-  // Use the real type name if we have RTTI.
-  return TypeIndex::Make(typeid(T).name());
-#else
-  return TypeIndex::Make("[RTTI disabled]");
-#endif  // __GXX_RTTI
+  return TypeIndex::Make();
 }
 
 }  // namespace tensorflow

From 54a7bd1d738e0f2b8b9b729f48f8e90f2f39b3c6 Mon Sep 17 00:00:00 2001
From: Nathan Luehr 
Date: Fri, 19 Jun 2020 18:02:40 -0500
Subject: [PATCH 356/867] Remove unnecessary tf32_execution_allowed check

---
 tensorflow/stream_executor/cuda/cuda_blas.cc | 8 --------
 1 file changed, 8 deletions(-)

diff --git a/tensorflow/stream_executor/cuda/cuda_blas.cc b/tensorflow/stream_executor/cuda/cuda_blas.cc
index e9e3635d8c1..e387690da26 100644
--- a/tensorflow/stream_executor/cuda/cuda_blas.cc
+++ b/tensorflow/stream_executor/cuda/cuda_blas.cc
@@ -402,14 +402,6 @@ bool CUDABlas::DoBlasInternalImpl(FuncT cublas_func, Stream *stream,
                                            : CUBLAS_POINTER_MODE_DEVICE)) {
     return false;
   }
-#if CUBLAS_VER_MAJOR >= 11
-  ScopedCublasMathMode math_mode{blas_};
-  if (!tensorflow::tf32_execution_allowed()) {
-    if (!math_mode.Init(CUBLAS_DEFAULT_MATH)) {
-      return false;
-    }
-  }
-#endif
   cublasStatus_t ret = cublas_func(blas_, args...);
   if ((err_on_failure || VLOG_IS_ON(3)) && ret != CUBLAS_STATUS_SUCCESS) {
     LOG(ERROR) << "failed to run cuBLAS routine: " << ToString(ret);

From 866eb4828af9aa63e75b51fa05e84cc6b3176e5b Mon Sep 17 00:00:00 2001
From: "A. Unique TensorFlower" 
Date: Fri, 19 Jun 2020 15:58:05 -0700
Subject: [PATCH 357/867] tf.numpy avoid using typing module given Kokoro
 breakage.

PiperOrigin-RevId: 317396321
Change-Id: Ifa05dcbbdd8998708e52ba52cdb03e33e4952f21
---
 tensorflow/python/ops/numpy_ops/np_array_ops.py | 5 ++---
 1 file changed, 2 insertions(+), 3 deletions(-)

diff --git a/tensorflow/python/ops/numpy_ops/np_array_ops.py b/tensorflow/python/ops/numpy_ops/np_array_ops.py
index 47236d45561..74f1bc8cbef 100644
--- a/tensorflow/python/ops/numpy_ops/np_array_ops.py
+++ b/tensorflow/python/ops/numpy_ops/np_array_ops.py
@@ -21,7 +21,6 @@ from __future__ import print_function
 
 import math
 import numbers
-from typing import Sequence
 import numpy as np
 import six
 
@@ -1684,10 +1683,10 @@ def _slice_helper(tensor, slice_spec):
 def _as_spec_tuple(slice_spec):
   """Convert slice_spec to tuple."""
   if isinstance(slice_spec,
-                Sequence) and not isinstance(slice_spec, np.ndarray):
+                (list, tuple)) and not isinstance(slice_spec, np.ndarray):
     is_index = True
     for s in slice_spec:
-      if s is None or s is Ellipsis or isinstance(s, (Sequence, slice)):
+      if s is None or s is Ellipsis or isinstance(s, (list, tuple, slice)):
         is_index = False
         break
       elif isinstance(s, (np_arrays.ndarray, np.ndarray)) and s.ndim != 0:

From fac99746cb7382af83fc9922c345a1bd68caa516 Mon Sep 17 00:00:00 2001
From: Chenkai Kuang 
Date: Fri, 19 Jun 2020 16:08:23 -0700
Subject: [PATCH 358/867] Override "map_resources" in AggregatingVariable.

PiperOrigin-RevId: 317398115
Change-Id: Ic57325aacf0fb45a66a469428f544fb94f3cc031
---
 tensorflow/python/distribute/ps_values.py | 8 ++++++++
 1 file changed, 8 insertions(+)

diff --git a/tensorflow/python/distribute/ps_values.py b/tensorflow/python/distribute/ps_values.py
index 37cd6e12d90..5fb2d42b626 100644
--- a/tensorflow/python/distribute/ps_values.py
+++ b/tensorflow/python/distribute/ps_values.py
@@ -166,6 +166,14 @@ class AggregatingVariable(variables_lib.Variable, core.Tensor):
   def _gather_saveables_for_checkpoint(self):
     return {trackable.VARIABLE_VALUE_KEY: self._v}
 
+  def _map_resources(self):
+    """For implementing `Trackable`."""
+    # By delegating this method to the wrapped variable, SavedModel with
+    # AggregatingVariable are identical to SavedModel with normal variables.
+    obj_map, resource_map = self._v._map_resources()  # pylint:disable=protected-access
+    obj_map[self] = obj_map[self._v]
+    return obj_map, resource_map
+
   # pylint: disable=multiple-statements
   def __add__(self, o):
     return self._v + o

From a70ad66828f198f4f495c3f29445b396045b6c3f Mon Sep 17 00:00:00 2001
From: "A. Unique TensorFlower" 
Date: Fri, 19 Jun 2020 16:23:33 -0700
Subject: [PATCH 359/867] GitHub Issue #40462

PiperOrigin-RevId: 317400375
Change-Id: I13891b5c2f41ac97674ddfda679c4273d53b25ef
---
 tensorflow/python/keras/preprocessing/image_dataset.py | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/tensorflow/python/keras/preprocessing/image_dataset.py b/tensorflow/python/keras/preprocessing/image_dataset.py
index d287c4ef372..6e78ac3c03a 100644
--- a/tensorflow/python/keras/preprocessing/image_dataset.py
+++ b/tensorflow/python/keras/preprocessing/image_dataset.py
@@ -147,7 +147,7 @@ def image_dataset_from_directory(directory,
           'directory. If you wish to infer the labels from the subdirectory '
           'names in the target directory, pass `labels="inferred"`. '
           'If you wish to get a dataset that only contains images '
-          '(no labels), pass `labels=None`.')
+          '(no labels), pass `labels_mode=None`.')
     if class_names:
       raise ValueError('You can only pass `class_names` if the labels are '
                        'inferred from the subdirectory names in the target '

From 62683d061cf31d05588a94cc333b53542cea9568 Mon Sep 17 00:00:00 2001
From: George Karpenkov 
Date: Fri, 19 Jun 2020 16:25:08 -0700
Subject: [PATCH 360/867] [XLA] Rollback of rollback of "Implement
 LocalClient::Run which supports buffer donation"

PiperOrigin-RevId: 317400695
Change-Id: I56f1f8df347d5a3b2bad9526c7315c63ad6ddadb
---
 .../compiler/xla/client/local_client.cc       | 26 ++++++++++++-------
 tensorflow/compiler/xla/client/local_client.h | 21 +++++++++++++++
 .../tests/multiple_devices_on_host_test.cc    |  3 ++-
 tensorflow/compiler/xla/tests/while_test.cc   |  6 +++--
 4 files changed, 44 insertions(+), 12 deletions(-)

diff --git a/tensorflow/compiler/xla/client/local_client.cc b/tensorflow/compiler/xla/client/local_client.cc
index afe115deda8..aa252067e19 100644
--- a/tensorflow/compiler/xla/client/local_client.cc
+++ b/tensorflow/compiler/xla/client/local_client.cc
@@ -176,15 +176,23 @@ StatusOr LocalExecutable::Run(
   for (const ShapedBuffer* const arg : arguments) {
     argument_shapes.push_back(&arg->on_host_shape());
   }
-  TF_ASSIGN_OR_RETURN(auto options_and_stream,
-                      RunHelper(argument_shapes, run_options));
-  ExecutableRunOptions options = options_and_stream.first.run_options();
-  options.set_device_ordinal(-1);
-  auto result = RunAsync(arguments, options);
-  Status block_status = options.stream()->BlockHostUntilDone();
-  TF_RETURN_IF_ERROR(result.status());
-  TF_RETURN_IF_ERROR(block_status);
-  return result;
+  return AsyncCallAndBlockHostUntilDone(
+      argument_shapes, run_options, [&](const ExecutableRunOptions& options) {
+        return RunAsync(arguments, options);
+      });
+}
+
+StatusOr LocalExecutable::Run(
+    std::vector arguments, ExecutableRunOptions run_options) {
+  std::vector argument_shapes;
+  argument_shapes.reserve(arguments.size());
+  for (const ExecutionInput& arg : arguments) {
+    argument_shapes.push_back(&arg.shape());
+  }
+  return AsyncCallAndBlockHostUntilDone(
+      argument_shapes, run_options, [&](const ExecutableRunOptions& options) {
+        return RunAsync(argument_shapes, std::move(arguments), options);
+      });
 }
 
 static std::shared_ptr DumpArguments(
diff --git a/tensorflow/compiler/xla/client/local_client.h b/tensorflow/compiler/xla/client/local_client.h
index 7cdeb9dcbf6..3241ac73d54 100644
--- a/tensorflow/compiler/xla/client/local_client.h
+++ b/tensorflow/compiler/xla/client/local_client.h
@@ -51,6 +51,11 @@ class LocalExecutable {
       const absl::Span arguments,
       ExecutableRunOptions run_options);
 
+  // Similar to Run(), but allows for donating argument buffers to the
+  // executable.
+  StatusOr Run(std::vector arguments,
+                                ExecutableRunOptions run_options);
+
   // Similar to Run(), but need not block the host waiting for the computation
   // to complete before returning.
   StatusOr RunAsync(
@@ -90,6 +95,22 @@ class LocalExecutable {
   // Backend::devices_equivalent).
   int build_device_ordinal() const { return build_options_.device_ordinal(); }
 
+  template 
+  StatusOr AsyncCallAndBlockHostUntilDone(
+      absl::Span argument_shapes,
+      const ExecutableRunOptions& run_options,
+      std::function(const ExecutableRunOptions&)> async_callback) {
+    TF_ASSIGN_OR_RETURN(auto options_and_stream,
+                        RunHelper(argument_shapes, run_options));
+    ExecutableRunOptions options = options_and_stream.first.run_options();
+    options.set_device_ordinal(-1);
+    StatusOr result = async_callback(options);
+    Status block_status = options.stream()->BlockHostUntilDone();
+    TF_RETURN_IF_ERROR(result.status());
+    TF_RETURN_IF_ERROR(block_status);
+    return result;
+  }
+
   // Compiled computation.
   std::unique_ptr executable_;
 
diff --git a/tensorflow/compiler/xla/tests/multiple_devices_on_host_test.cc b/tensorflow/compiler/xla/tests/multiple_devices_on_host_test.cc
index 2b19aaded9c..2231fc6feab 100644
--- a/tensorflow/compiler/xla/tests/multiple_devices_on_host_test.cc
+++ b/tensorflow/compiler/xla/tests/multiple_devices_on_host_test.cc
@@ -45,7 +45,8 @@ void CompileAndExecute(
       xla::ClientLibrary::GetXlaService(client->platform())
           ->backend()
           .memory_allocator());
-  StatusOr result = executable->Run({}, execute_options);
+  StatusOr result =
+      executable->Run(absl::Span(), execute_options);
   {
     absl::MutexLock lock(results_mutex);
     results->emplace_back(device_ordinal, std::move(result));
diff --git a/tensorflow/compiler/xla/tests/while_test.cc b/tensorflow/compiler/xla/tests/while_test.cc
index d575bbb1f3e..8e8c3605cc7 100644
--- a/tensorflow/compiler/xla/tests/while_test.cc
+++ b/tensorflow/compiler/xla/tests/while_test.cc
@@ -1324,14 +1324,16 @@ void BM_WhileLoop(int num_iters) {
   options.set_allocator(&allocator);
   const int kWarmups = 2;
   for (int i = 0; i < kWarmups; ++i) {
-    auto result = executable->Run({}, options);
+    auto result =
+        executable->Run(absl::Span(), options);
     ASSERT_TRUE(result.ok());
   }
 
   // Run benchmark.
   tensorflow::testing::StartTiming();
   for (int i = 0; i < num_iters; ++i) {
-    auto result = executable->Run({}, options);
+    auto result =
+        executable->Run(absl::Span(), options);
     ASSERT_TRUE(result.ok());
   }
 }

From 3c9a2f200e7c439ce5e3345bdd9055b88b9777ca Mon Sep 17 00:00:00 2001
From: Peng Wang 
Date: Fri, 19 Jun 2020 16:34:05 -0700
Subject: [PATCH 361/867] [TF-numpy] Changed all callsites of
 np_doc/np_doc_only to use string names, to avoid blocking imports when some
 numpy symbols are missing (e.g. because of an older version of numpy).

Also moved `np_fun_name` to be the first positional argument of np_doc/np_doc_only.

PiperOrigin-RevId: 317402093
Change-Id: I8b3ba54909e9507c7ab062f5bf5d7f13ad549317
---
 .../python/ops/numpy_ops/np_array_ops.py      | 156 ++++++------
 .../python/ops/numpy_ops/np_math_ops.py       | 228 +++++++++---------
 tensorflow/python/ops/numpy_ops/np_utils.py   |  91 ++++---
 .../python/ops/numpy_ops/np_utils_test.py     |  10 +-
 4 files changed, 254 insertions(+), 231 deletions(-)

diff --git a/tensorflow/python/ops/numpy_ops/np_array_ops.py b/tensorflow/python/ops/numpy_ops/np_array_ops.py
index 74f1bc8cbef..a87c72ed763 100644
--- a/tensorflow/python/ops/numpy_ops/np_array_ops.py
+++ b/tensorflow/python/ops/numpy_ops/np_array_ops.py
@@ -41,17 +41,17 @@ from tensorflow.python.ops.numpy_ops import np_utils
 from tensorflow.python.util import nest
 
 
-@np_utils.np_doc(np.empty)
+@np_utils.np_doc('empty')
 def empty(shape, dtype=float):  # pylint: disable=redefined-outer-name
   return zeros(shape, dtype)
 
 
-@np_utils.np_doc(np.empty_like)
+@np_utils.np_doc('empty_like')
 def empty_like(a, dtype=None):
   return zeros_like(a, dtype)
 
 
-@np_utils.np_doc(np.zeros)
+@np_utils.np_doc('zeros')
 def zeros(shape, dtype=float):  # pylint: disable=redefined-outer-name
   dtype = (
       np_utils.result_type(dtype) if dtype else np_dtypes.default_float_type())
@@ -60,7 +60,7 @@ def zeros(shape, dtype=float):  # pylint: disable=redefined-outer-name
   return np_arrays.tensor_to_ndarray(array_ops.zeros(shape, dtype=dtype))
 
 
-@np_utils.np_doc(np.zeros_like)
+@np_utils.np_doc('zeros_like')
 def zeros_like(a, dtype=None):  # pylint: disable=missing-docstring
   if isinstance(a, np_arrays.ndarray):
     a = a.data
@@ -75,7 +75,7 @@ def zeros_like(a, dtype=None):  # pylint: disable=missing-docstring
   return np_arrays.tensor_to_ndarray(array_ops.zeros_like(a, dtype))
 
 
-@np_utils.np_doc(np.ones)
+@np_utils.np_doc('ones')
 def ones(shape, dtype=float):  # pylint: disable=redefined-outer-name
   if dtype:
     dtype = np_utils.result_type(dtype)
@@ -84,7 +84,7 @@ def ones(shape, dtype=float):  # pylint: disable=redefined-outer-name
   return np_arrays.tensor_to_ndarray(array_ops.ones(shape, dtype=dtype))
 
 
-@np_utils.np_doc(np.ones_like)
+@np_utils.np_doc('ones_like')
 def ones_like(a, dtype=None):
   if isinstance(a, np_arrays.ndarray):
     a = a.data
@@ -95,7 +95,7 @@ def ones_like(a, dtype=None):
   return np_arrays.tensor_to_ndarray(array_ops.ones_like(a, dtype))
 
 
-@np_utils.np_doc(np.eye)
+@np_utils.np_doc('eye')
 def eye(N, M=None, k=0, dtype=float):  # pylint: disable=invalid-name,missing-docstring
   if dtype:
     dtype = np_utils.result_type(dtype)
@@ -127,12 +127,12 @@ def eye(N, M=None, k=0, dtype=float):  # pylint: disable=invalid-name,missing-do
       array_ops.matrix_diag(diagonal=diagonal_, num_rows=N, num_cols=M, k=k))
 
 
-@np_utils.np_doc(np.identity)
+@np_utils.np_doc('identity')
 def identity(n, dtype=float):
   return eye(N=n, M=n, dtype=dtype)
 
 
-@np_utils.np_doc(np.full)
+@np_utils.np_doc('full')
 def full(shape, fill_value, dtype=None):  # pylint: disable=redefined-outer-name
   if not isinstance(shape, np_arrays.ndarray):
     shape = asarray(np_arrays.convert_to_tensor(shape, dtype_hint=np.int32))
@@ -144,7 +144,7 @@ def full(shape, fill_value, dtype=None):  # pylint: disable=redefined-outer-name
 
 # Using doc only here since np full_like signature doesn't seem to have the
 # shape argument (even though it exists in the documentation online).
-@np_utils.np_doc_only(np.full_like)
+@np_utils.np_doc_only('full_like')
 def full_like(a, fill_value, dtype=None, order='K', subok=True, shape=None):  # pylint: disable=missing-docstring,redefined-outer-name
   """order, subok and shape arguments mustn't be changed."""
   if order != 'K':
@@ -163,7 +163,7 @@ def full_like(a, fill_value, dtype=None, order='K', subok=True, shape=None):  #
 
 # TODO(wangpeng): investigate whether we can make `copy` default to False.
 # pylint: disable=g-short-docstring-punctuation,g-no-space-after-docstring-summary,g-doc-return-or-yield,g-doc-args
-@np_utils.np_doc_only(np.array)
+@np_utils.np_doc_only('array')
 def array(val, dtype=None, copy=True, ndmin=0):  # pylint: disable=redefined-outer-name
   """Since Tensors are immutable, a copy is made only if val is placed on a
 
@@ -224,7 +224,7 @@ def array(val, dtype=None, copy=True, ndmin=0):  # pylint: disable=redefined-out
 # pylint: enable=g-short-docstring-punctuation,g-no-space-after-docstring-summary,g-doc-return-or-yield,g-doc-args
 
 
-@np_utils.np_doc(np.asarray)
+@np_utils.np_doc('asarray')
 def asarray(a, dtype=None):
   if dtype:
     dtype = np_utils.result_type(dtype)
@@ -233,18 +233,18 @@ def asarray(a, dtype=None):
   return array(a, dtype, copy=False)
 
 
-@np_utils.np_doc(np.asanyarray)
+@np_utils.np_doc('asanyarray')
 def asanyarray(a, dtype=None):
   return asarray(a, dtype)
 
 
-@np_utils.np_doc(np.ascontiguousarray)
+@np_utils.np_doc('ascontiguousarray')
 def ascontiguousarray(a, dtype=None):
   return array(a, dtype, ndmin=1)
 
 
 # Numerical ranges.
-@np_utils.np_doc(np.arange)
+@np_utils.np_doc('arange')
 def arange(start, stop=None, step=1, dtype=None):
   """Returns `step`-separated values in the range [start, stop).
 
@@ -286,7 +286,7 @@ def arange(start, stop=None, step=1, dtype=None):
 
 
 # Building matrices.
-@np_utils.np_doc(np.diag)
+@np_utils.np_doc('diag')
 def diag(v, k=0):  # pylint: disable=missing-docstring
   """Raises an error if input is not 1- or 2-d."""
   v = asarray(v).data
@@ -321,7 +321,7 @@ def diag(v, k=0):  # pylint: disable=missing-docstring
   return np_utils.tensor_to_ndarray(result)
 
 
-@np_utils.np_doc(np.diagonal)
+@np_utils.np_doc('diagonal')
 def diagonal(a, offset=0, axis1=0, axis2=1):  # pylint: disable=missing-docstring
   a = asarray(a).data
 
@@ -352,7 +352,7 @@ def diagonal(a, offset=0, axis1=0, axis2=1):  # pylint: disable=missing-docstrin
   return a
 
 
-@np_utils.np_doc(np.diagflat)
+@np_utils.np_doc('diagflat')
 def diagflat(v, k=0):
   v = asarray(v)
   return diag(array_ops.reshape(v.data, [-1]), k)
@@ -363,21 +363,21 @@ def _promote_dtype(*arrays):
   return [asarray(a, dtype=dtype) for a in arrays]
 
 
-@np_utils.np_doc(np.all)
+@np_utils.np_doc('all')
 def all(a, axis=None, keepdims=None):  # pylint: disable=redefined-builtin
   a = asarray(a, dtype=bool)
   return np_utils.tensor_to_ndarray(
       math_ops.reduce_all(input_tensor=a.data, axis=axis, keepdims=keepdims))
 
 
-@np_utils.np_doc(np.any)
+@np_utils.np_doc('any')
 def any(a, axis=None, keepdims=None):  # pylint: disable=redefined-builtin
   a = asarray(a, dtype=bool)
   return np_utils.tensor_to_ndarray(
       math_ops.reduce_any(input_tensor=a.data, axis=axis, keepdims=keepdims))
 
 
-@np_utils.np_doc(np.compress)
+@np_utils.np_doc('compress')
 def compress(condition, a, axis=None):  # pylint: disable=redefined-outer-name,missing-function-docstring
   condition = asarray(condition, dtype=bool)
   a = asarray(a)
@@ -408,7 +408,7 @@ def compress(condition, a, axis=None):  # pylint: disable=redefined-outer-name,m
       array_ops.boolean_mask(tensor=a_t, mask=condition_t, axis=axis))
 
 
-@np_utils.np_doc(np.copy)
+@np_utils.np_doc('copy')
 def copy(a):
   return array(a, copy=True)
 
@@ -424,7 +424,7 @@ def _maybe_promote_to_int(a):
   return a
 
 
-@np_utils.np_doc(np.cumprod)
+@np_utils.np_doc('cumprod')
 def cumprod(a, axis=None, dtype=None):  # pylint: disable=missing-docstring
   a = asarray(a, dtype=dtype)
 
@@ -440,7 +440,7 @@ def cumprod(a, axis=None, dtype=None):  # pylint: disable=missing-docstring
   return np_utils.tensor_to_ndarray(math_ops.cumprod(a.data, axis))
 
 
-@np_utils.np_doc(np.cumsum)
+@np_utils.np_doc('cumsum')
 def cumsum(a, axis=None, dtype=None):  # pylint: disable=missing-docstring
   a = asarray(a, dtype=dtype)
 
@@ -456,7 +456,7 @@ def cumsum(a, axis=None, dtype=None):  # pylint: disable=missing-docstring
   return np_utils.tensor_to_ndarray(math_ops.cumsum(a.data, axis))
 
 
-@np_utils.np_doc(np.imag)
+@np_utils.np_doc('imag')
 def imag(a):
   a = asarray(a)
   # TODO(srbs): np.imag returns a scalar if a is a scalar, whereas we always
@@ -536,7 +536,7 @@ def _reduce(tf_fn,
       tf_fn(input_tensor=a.data, axis=axis, keepdims=keepdims))
 
 
-@np_utils.np_doc(np.sum)
+@np_utils.np_doc('sum')
 def sum(a, axis=None, dtype=None, keepdims=None):  # pylint: disable=redefined-builtin
   return _reduce(
       math_ops.reduce_sum,
@@ -547,7 +547,7 @@ def sum(a, axis=None, dtype=None, keepdims=None):  # pylint: disable=redefined-b
       tf_bool_fn=math_ops.reduce_any)
 
 
-@np_utils.np_doc(np.prod)
+@np_utils.np_doc('prod')
 def prod(a, axis=None, dtype=None, keepdims=None):
   return _reduce(
       math_ops.reduce_prod,
@@ -558,7 +558,7 @@ def prod(a, axis=None, dtype=None, keepdims=None):
       tf_bool_fn=math_ops.reduce_all)
 
 
-@np_utils.np_doc(np.mean)
+@np_utils.np_doc('mean')
 def mean(a, axis=None, dtype=None, keepdims=None):
   return _reduce(
       math_ops.reduce_mean,
@@ -569,7 +569,7 @@ def mean(a, axis=None, dtype=None, keepdims=None):
       promote_int=_TO_FLOAT)
 
 
-@np_utils.np_doc(np.amax)
+@np_utils.np_doc('amax')
 def amax(a, axis=None, keepdims=None):
   return _reduce(
       math_ops.reduce_max,
@@ -582,7 +582,7 @@ def amax(a, axis=None, keepdims=None):
       preserve_bool=True)
 
 
-@np_utils.np_doc(np.amin)
+@np_utils.np_doc('amin')
 def amin(a, axis=None, keepdims=None):
   return _reduce(
       math_ops.reduce_min,
@@ -595,7 +595,7 @@ def amin(a, axis=None, keepdims=None):
       preserve_bool=True)
 
 
-@np_utils.np_doc(np.var)
+@np_utils.np_doc('var')
 def var(a, axis=None, dtype=None, out=None, ddof=0, keepdims=None):  # pylint: disable=missing-docstring
   if dtype:
     working_dtype = np_utils.result_type(a, dtype)
@@ -642,7 +642,7 @@ def var(a, axis=None, dtype=None, out=None, ddof=0, keepdims=None):  # pylint: d
   return np_utils.tensor_to_ndarray(result)
 
 
-@np_utils.np_doc(np.std)
+@np_utils.np_doc('std')
 def std(a, axis=None, keepdims=None):  # pylint: disable=missing-function-docstring
   return _reduce(
       math_ops.reduce_std,
@@ -653,7 +653,7 @@ def std(a, axis=None, keepdims=None):  # pylint: disable=missing-function-docstr
       promote_int=_TO_FLOAT)
 
 
-@np_utils.np_doc(np.ravel)
+@np_utils.np_doc('ravel')
 def ravel(a):  # pylint: disable=missing-docstring
   a = asarray(a)
   out = np_utils.cond(
@@ -665,7 +665,7 @@ def ravel(a):  # pylint: disable=missing-docstring
 setattr(np_arrays.ndarray, 'ravel', ravel)
 
 
-@np_utils.np_doc(np.real)
+@np_utils.np_doc('real')
 def real(val):
   val = asarray(val)
   # TODO(srbs): np.real returns a scalar if val is a scalar, whereas we always
@@ -673,7 +673,7 @@ def real(val):
   return np_utils.tensor_to_ndarray(math_ops.real(val.data))
 
 
-@np_utils.np_doc(np.repeat)
+@np_utils.np_doc('repeat')
 def repeat(a, repeats, axis=None):  # pylint: disable=missing-docstring
   a = asarray(a).data
   original_shape = a._shape_as_list()  # pylint: disable=protected-access
@@ -704,7 +704,7 @@ def repeat(a, repeats, axis=None):  # pylint: disable=missing-docstring
   return np_utils.tensor_to_ndarray(result)
 
 
-@np_utils.np_doc(np.around)
+@np_utils.np_doc('around')
 def around(a, decimals=0):  # pylint: disable=missing-docstring
   a = asarray(a)
   dtype = a.dtype
@@ -726,7 +726,7 @@ def around(a, decimals=0):  # pylint: disable=missing-docstring
 setattr(np_arrays.ndarray, '__round__', around)
 
 
-@np_utils.np_doc(np.reshape)
+@np_utils.np_doc('reshape')
 def reshape(a, newshape, order='C'):
   """order argument can only b 'C' or 'F'."""
   if order not in {'C', 'F'}:
@@ -758,19 +758,19 @@ def _reshape_method_wrapper(a, *newshape, **kwargs):
   return reshape(a, newshape, order=order)
 
 
-@np_utils.np_doc(np.expand_dims)
+@np_utils.np_doc('expand_dims')
 def expand_dims(a, axis):
   a = asarray(a)
   return np_utils.tensor_to_ndarray(array_ops.expand_dims(a.data, axis=axis))
 
 
-@np_utils.np_doc(np.squeeze)
+@np_utils.np_doc('squeeze')
 def squeeze(a, axis=None):
   a = asarray(a)
   return np_utils.tensor_to_ndarray(array_ops.squeeze(a, axis))
 
 
-@np_utils.np_doc(np.transpose)
+@np_utils.np_doc('transpose')
 def transpose(a, axes=None):
   a = asarray(a)
   if axes is not None:
@@ -778,7 +778,7 @@ def transpose(a, axes=None):
   return np_utils.tensor_to_ndarray(array_ops.transpose(a=a.data, perm=axes))
 
 
-@np_utils.np_doc(np.swapaxes)
+@np_utils.np_doc('swapaxes')
 def swapaxes(a, axis1, axis2):  # pylint: disable=missing-docstring
   a = asarray(a).data
 
@@ -794,7 +794,7 @@ def swapaxes(a, axis1, axis2):  # pylint: disable=missing-docstring
   return np_utils.tensor_to_ndarray(a)
 
 
-@np_utils.np_doc(np.moveaxis)
+@np_utils.np_doc('moveaxis')
 def moveaxis(a, source, destination):  # pylint: disable=missing-docstring
   """Raises ValueError if source, destination not in (-ndim(a), ndim(a))."""
   if not source and not destination:
@@ -914,7 +914,7 @@ setattr(np_arrays.ndarray, 'reshape', _reshape_method_wrapper)
 setattr(np_arrays.ndarray, '__setitem__', _setitem)
 
 
-@np_utils.np_doc(np.pad)
+@np_utils.np_doc('pad')
 def pad(ary, pad_width, mode, constant_values=0):
   """Only supports modes 'constant', 'reflect' and 'symmetric' currently."""
   if not (mode == 'constant' or mode == 'reflect' or mode == 'symmetric'):
@@ -930,7 +930,7 @@ def pad(ary, pad_width, mode, constant_values=0):
           constant_values=constant_values))
 
 
-@np_utils.np_doc(np.take)
+@np_utils.np_doc('take')
 def take(a, indices, axis=None, out=None, mode='clip'):
   """out argument is not supported, and default mode is clip."""
   if out is not None:
@@ -957,7 +957,7 @@ def take(a, indices, axis=None, out=None, mode='clip'):
   return np_utils.tensor_to_ndarray(array_ops.gather(a, indices, axis=axis))
 
 
-@np_utils.np_doc_only(np.where)
+@np_utils.np_doc_only('where')
 def where(condition, x=None, y=None):
   """Raises ValueError if exactly one of x or y is not None."""
   condition = asarray(condition, dtype=np.bool_)
@@ -970,7 +970,7 @@ def where(condition, x=None, y=None):
   raise ValueError('Both x and y must be ndarrays, or both must be None.')
 
 
-@np_utils.np_doc(np.select)
+@np_utils.np_doc('select')
 def select(condlist, choicelist, default=0):  # pylint: disable=missing-docstring
   if len(condlist) != len(choicelist):
     msg = 'condlist must have length equal to choicelist ({} vs {})'
@@ -987,19 +987,19 @@ def select(condlist, choicelist, default=0):  # pylint: disable=missing-docstrin
   return output
 
 
-@np_utils.np_doc(np.shape)
+@np_utils.np_doc('shape')
 def shape(a):
   a = asarray(a)
   return a.shape
 
 
-@np_utils.np_doc(np.ndim)
+@np_utils.np_doc('ndim')
 def ndim(a):
   a = asarray(a)
   return a.ndim
 
 
-@np_utils.np_doc(np.isscalar)
+@np_utils.np_doc('isscalar')
 def isscalar(a):
   return ndim(a) == 0
 
@@ -1034,7 +1034,7 @@ def _boundaries_to_sizes(a, boundaries, axis):
   return sizes
 
 
-@np_utils.np_doc(np.split)
+@np_utils.np_doc('split')
 def split(ary, indices_or_sections, axis=0):
   ary = asarray(ary)
   if not isinstance(indices_or_sections, six.integer_types):
@@ -1043,26 +1043,26 @@ def split(ary, indices_or_sections, axis=0):
   return [np_utils.tensor_to_ndarray(a) for a in result]
 
 
-def _split_on_axis(np_fun, axis):
+def _split_on_axis(np_fun_name, axis):
 
-  @np_utils.np_doc(np_fun)
+  @np_utils.np_doc(np_fun_name)
   def f(ary, indices_or_sections):
     return split(ary, indices_or_sections, axis=axis)
 
   return f
 
 
-vsplit = _split_on_axis(np.vsplit, axis=0)
-hsplit = _split_on_axis(np.hsplit, axis=1)
-dsplit = _split_on_axis(np.dsplit, axis=2)
+vsplit = _split_on_axis('vsplit', axis=0)
+hsplit = _split_on_axis('hsplit', axis=1)
+dsplit = _split_on_axis('dsplit', axis=2)
 
 
-@np_utils.np_doc(np.broadcast_to)
+@np_utils.np_doc('broadcast_to')
 def broadcast_to(array, shape):  # pylint: disable=redefined-outer-name
   return full(shape, array)
 
 
-@np_utils.np_doc(np.stack)
+@np_utils.np_doc('stack')
 def stack(arrays, axis=0):  # pylint: disable=missing-function-docstring
   if isinstance(arrays, (np_arrays.ndarray, ops.Tensor)):
     arrays = asarray(arrays)
@@ -1077,7 +1077,7 @@ def stack(arrays, axis=0):  # pylint: disable=missing-function-docstring
   return asarray(array_ops.stack(unwrapped_arrays, axis))
 
 
-@np_utils.np_doc(np.hstack)
+@np_utils.np_doc('hstack')
 def hstack(tup):
   arrays = [atleast_1d(a) for a in tup]
   arrays = _promote_dtype(*arrays)  # pylint: disable=protected-access
@@ -1091,7 +1091,7 @@ def hstack(tup):
       lambda: array_ops.concat(unwrapped_arrays, axis=1))
 
 
-@np_utils.np_doc(np.vstack)
+@np_utils.np_doc('vstack')
 def vstack(tup):
   arrays = [atleast_2d(a) for a in tup]
   arrays = _promote_dtype(*arrays)  # pylint: disable=protected-access
@@ -1101,7 +1101,7 @@ def vstack(tup):
   return array_ops.concat(unwrapped_arrays, axis=0)
 
 
-@np_utils.np_doc(np.dstack)
+@np_utils.np_doc('dstack')
 def dstack(tup):
   arrays = [atleast_3d(a) for a in tup]
   arrays = _promote_dtype(*arrays)  # pylint: disable=protected-access
@@ -1148,17 +1148,17 @@ def _atleast_nd(n, new_shape, *arys):
     return arys
 
 
-@np_utils.np_doc(np.atleast_1d)
+@np_utils.np_doc('atleast_1d')
 def atleast_1d(*arys):
   return _atleast_nd(1, _pad_left_to, *arys)
 
 
-@np_utils.np_doc(np.atleast_2d)
+@np_utils.np_doc('atleast_2d')
 def atleast_2d(*arys):
   return _atleast_nd(2, _pad_left_to, *arys)
 
 
-@np_utils.np_doc(np.atleast_3d)
+@np_utils.np_doc('atleast_3d')
 def atleast_3d(*arys):  # pylint: disable=missing-docstring
 
   def new_shape(_, old_shape):
@@ -1175,7 +1175,7 @@ def atleast_3d(*arys):  # pylint: disable=missing-docstring
   return _atleast_nd(3, new_shape, *arys)
 
 
-@np_utils.np_doc(np.nonzero)
+@np_utils.np_doc('nonzero')
 def nonzero(a):
   a = atleast_1d(a).data
   if a.shape.rank is None:
@@ -1189,7 +1189,7 @@ def nonzero(a):
           axis=1))
 
 
-@np_utils.np_doc(np.diag_indices)
+@np_utils.np_doc('diag_indices')
 def diag_indices(n, ndim=2):  # pylint: disable=missing-docstring,redefined-outer-name
   if n < 0:
     raise ValueError(
@@ -1202,7 +1202,7 @@ def diag_indices(n, ndim=2):  # pylint: disable=missing-docstring,redefined-oute
   return (math_ops.range(n),) * ndim
 
 
-@np_utils.np_doc(np.tri)
+@np_utils.np_doc('tri')
 def tri(N, M=None, k=0, dtype=None):  # pylint: disable=invalid-name,missing-docstring
   M = M if M is not None else N
   if dtype is not None:
@@ -1229,7 +1229,7 @@ def tri(N, M=None, k=0, dtype=None):  # pylint: disable=invalid-name,missing-doc
   return np_utils.tensor_to_ndarray(r)
 
 
-@np_utils.np_doc(np.tril)
+@np_utils.np_doc('tril')
 def tril(m, k=0):  # pylint: disable=missing-docstring
   m = asarray(m).data
   if m.shape.ndims is None:
@@ -1251,7 +1251,7 @@ def tril(m, k=0):  # pylint: disable=missing-docstring
           array_ops.broadcast_to(mask, array_ops.shape(m)), m, z))
 
 
-@np_utils.np_doc(np.triu)
+@np_utils.np_doc('triu')
 def triu(m, k=0):  # pylint: disable=missing-docstring
   m = asarray(m).data
   if m.shape.ndims is None:
@@ -1273,7 +1273,7 @@ def triu(m, k=0):  # pylint: disable=missing-docstring
           array_ops.broadcast_to(mask, array_ops.shape(m)), z, m))
 
 
-@np_utils.np_doc(np.flip)
+@np_utils.np_doc('flip')
 def flip(m, axis=None):  # pylint: disable=missing-docstring
   m = asarray(m).data
 
@@ -1286,17 +1286,17 @@ def flip(m, axis=None):  # pylint: disable=missing-docstring
   return np_utils.tensor_to_ndarray(array_ops.reverse(m, [axis]))
 
 
-@np_utils.np_doc(np.flipud)
+@np_utils.np_doc('flipud')
 def flipud(m):  # pylint: disable=missing-docstring
   return flip(m, 0)
 
 
-@np_utils.np_doc(np.fliplr)
+@np_utils.np_doc('fliplr')
 def fliplr(m):  # pylint: disable=missing-docstring
   return flip(m, 1)
 
 
-@np_utils.np_doc(np.roll)
+@np_utils.np_doc('roll')
 def roll(a, shift, axis=None):  # pylint: disable=missing-docstring
   a = asarray(a).data
 
@@ -1309,7 +1309,7 @@ def roll(a, shift, axis=None):  # pylint: disable=missing-docstring
   return np_utils.tensor_to_ndarray(array_ops.reshape(a, original_shape))
 
 
-@np_utils.np_doc(np.rot90)
+@np_utils.np_doc('rot90')
 def rot90(m, k=1, axes=(0, 1)):  # pylint: disable=missing-docstring
   m_rank = array_ops.rank(m)
   ax1, ax2 = np_utils._canonicalize_axes(axes, m_rank)  # pylint: disable=protected-access
@@ -1329,7 +1329,7 @@ def rot90(m, k=1, axes=(0, 1)):  # pylint: disable=missing-docstring
       return flip(transpose(m, perm), ax2)
 
 
-@np_utils.np_doc(np.vander)
+@np_utils.np_doc('vander')
 def vander(x, N=None, increasing=False):  # pylint: disable=missing-docstring,invalid-name
   x = asarray(x).data
 
@@ -1368,7 +1368,7 @@ def vander(x, N=None, increasing=False):  # pylint: disable=missing-docstring,in
           x, math_ops.cast(math_ops.range(start, limit, delta), dtype=x.dtype)))
 
 
-@np_utils.np_doc(np.ix_)
+@np_utils.np_doc('ix_')
 def ix_(*args):  # pylint: disable=missing-docstring
   n = len(args)
   output = []
@@ -1400,7 +1400,7 @@ def ix_(*args):  # pylint: disable=missing-docstring
   return output
 
 
-@np_utils.np_doc(np.broadcast_arrays)
+@np_utils.np_doc('broadcast_arrays')
 def broadcast_arrays(*args, **kwargs):  # pylint: disable=missing-docstring
   subok = kwargs.pop('subok', False)
   if subok:
@@ -1413,7 +1413,7 @@ def broadcast_arrays(*args, **kwargs):  # pylint: disable=missing-docstring
   return [np_utils.tensor_to_ndarray(arg) for arg in args]
 
 
-@np_utils.np_doc_only(np.sign)
+@np_utils.np_doc_only('sign')
 def sign(x, out=None, where=None, **kwargs):  # pylint: disable=missing-docstring,redefined-outer-name
   if out:
     raise ValueError('tf.numpy doesnt support setting out.')
@@ -1434,7 +1434,7 @@ def sign(x, out=None, where=None, **kwargs):  # pylint: disable=missing-docstrin
 
 # Note that np.take_along_axis may not be present in some supported versions of
 # numpy.
-@np_utils.np_doc(None, np_fun_name='take_along_axis')
+@np_utils.np_doc('take_along_axis')
 def take_along_axis(arr, indices, axis):  # pylint: disable=missing-docstring
   arr = asarray(arr)
   indices = asarray(indices)
diff --git a/tensorflow/python/ops/numpy_ops/np_math_ops.py b/tensorflow/python/ops/numpy_ops/np_math_ops.py
index 361bfb50dec..3cf26095dd8 100644
--- a/tensorflow/python/ops/numpy_ops/np_math_ops.py
+++ b/tensorflow/python/ops/numpy_ops/np_math_ops.py
@@ -40,7 +40,7 @@ from tensorflow.python.ops.numpy_ops import np_dtypes
 from tensorflow.python.ops.numpy_ops import np_utils
 
 
-@np_utils.np_doc_only(np.dot)
+@np_utils.np_doc_only('dot')
 def dot(a, b):  # pylint: disable=missing-docstring
 
   def f(a, b):  # pylint: disable=missing-docstring
@@ -67,7 +67,7 @@ def _bin_op(tf_fun, a, b, promote=True):
   return np_utils.tensor_to_ndarray(tf_fun(a.data, b.data))
 
 
-@np_utils.np_doc(np.add)
+@np_utils.np_doc('add')
 def add(x1, x2):
 
   def add_or_or(x1, x2):
@@ -79,12 +79,12 @@ def add(x1, x2):
   return _bin_op(add_or_or, x1, x2)
 
 
-@np_utils.np_doc(np.subtract)
+@np_utils.np_doc('subtract')
 def subtract(x1, x2):
   return _bin_op(math_ops.subtract, x1, x2)
 
 
-@np_utils.np_doc(np.multiply)
+@np_utils.np_doc('multiply')
 def multiply(x1, x2):
 
   def mul_or_and(x1, x2):
@@ -96,7 +96,7 @@ def multiply(x1, x2):
   return _bin_op(mul_or_and, x1, x2)
 
 
-@np_utils.np_doc(np.true_divide)
+@np_utils.np_doc('true_divide')
 def true_divide(x1, x2):  # pylint: disable=missing-function-docstring
 
   def _avoid_float64(x1, x2):
@@ -123,7 +123,7 @@ def true_divide(x1, x2):  # pylint: disable=missing-function-docstring
 divide = true_divide
 
 
-@np_utils.np_doc(np.floor_divide)
+@np_utils.np_doc('floor_divide')
 def floor_divide(x1, x2):  # pylint: disable=missing-function-docstring
 
   def f(x1, x2):
@@ -136,7 +136,7 @@ def floor_divide(x1, x2):  # pylint: disable=missing-function-docstring
   return _bin_op(f, x1, x2)
 
 
-@np_utils.np_doc(np.mod)
+@np_utils.np_doc('mod')
 def mod(x1, x2):  # pylint: disable=missing-function-docstring
 
   def f(x1, x2):
@@ -152,12 +152,12 @@ def mod(x1, x2):  # pylint: disable=missing-function-docstring
 remainder = mod
 
 
-@np_utils.np_doc(np.divmod)
+@np_utils.np_doc('divmod')
 def divmod(x1, x2):  # pylint: disable=redefined-builtin
   return floor_divide(x1, x2), mod(x1, x2)
 
 
-@np_utils.np_doc(np.maximum)
+@np_utils.np_doc('maximum')
 def maximum(x1, x2):
 
   def max_or_or(x1, x2):
@@ -169,7 +169,7 @@ def maximum(x1, x2):
   return _bin_op(max_or_or, x1, x2)
 
 
-@np_utils.np_doc(np.minimum)
+@np_utils.np_doc('minimum')
 def minimum(x1, x2):
 
   def min_or_and(x1, x2):
@@ -181,7 +181,7 @@ def minimum(x1, x2):
   return _bin_op(min_or_and, x1, x2)
 
 
-@np_utils.np_doc(np.clip)
+@np_utils.np_doc('clip')
 def clip(a, a_min, a_max):  # pylint: disable=missing-docstring
   if a_min is None and a_max is None:
     raise ValueError('Not more than one of `a_min` and `a_max` may be `None`.')
@@ -196,7 +196,7 @@ def clip(a, a_min, a_max):  # pylint: disable=missing-docstring
             *np_utils.tf_broadcast(a.data, a_min.data, a_max.data)))
 
 
-@np_utils.np_doc(np.matmul)
+@np_utils.np_doc('matmul')
 def matmul(x1, x2):  # pylint: disable=missing-docstring
 
   def f(x1, x2):
@@ -215,12 +215,12 @@ def matmul(x1, x2):  # pylint: disable=missing-docstring
   return _bin_op(f, x1, x2)
 
 
-@np_utils.np_doc(np.tensordot)
+@np_utils.np_doc('tensordot')
 def tensordot(a, b, axes=2):
   return _bin_op(lambda a, b: math_ops.tensordot(a, b, axes=axes), a, b)
 
 
-@np_utils.np_doc_only(np.inner)
+@np_utils.np_doc_only('inner')
 def inner(a, b):  # pylint: disable=missing-function-docstring
 
   def f(a, b):
@@ -233,7 +233,7 @@ def inner(a, b):  # pylint: disable=missing-function-docstring
   return _bin_op(f, a, b)
 
 
-@np_utils.np_doc(np.cross)
+@np_utils.np_doc('cross')
 def cross(a, b, axisa=-1, axisb=-1, axisc=-1, axis=None):  # pylint: disable=missing-docstring
 
   def f(a, b):  # pylint: disable=missing-docstring
@@ -309,7 +309,7 @@ def cross(a, b, axisa=-1, axisb=-1, axisc=-1, axis=None):  # pylint: disable=mis
   return _bin_op(f, a, b)
 
 
-@np_utils.np_doc_only(np.vdot)
+@np_utils.np_doc_only('vdot')
 def vdot(a, b):  # pylint: disable=missing-docstring
   a, b = np_array_ops._promote_dtype(a, b)
   a = np_array_ops.reshape(a, [-1])
@@ -319,27 +319,27 @@ def vdot(a, b):  # pylint: disable=missing-docstring
   return dot(a, b)
 
 
-@np_utils.np_doc(np.power)
+@np_utils.np_doc('power')
 def power(x1, x2):
   return _bin_op(math_ops.pow, x1, x2)
 
 
-@np_utils.np_doc(np.float_power)
+@np_utils.np_doc('float_power')
 def float_power(x1, x2):
   return power(x1, x2)
 
 
-@np_utils.np_doc(np.arctan2)
+@np_utils.np_doc('arctan2')
 def arctan2(x1, x2):
   return _bin_op(math_ops.atan2, x1, x2)
 
 
-@np_utils.np_doc(np.nextafter)
+@np_utils.np_doc('nextafter')
 def nextafter(x1, x2):
   return _bin_op(math_ops.nextafter, x1, x2)
 
 
-@np_utils.np_doc(np.heaviside)
+@np_utils.np_doc('heaviside')
 def heaviside(x1, x2):  # pylint: disable=missing-function-docstring
 
   def f(x1, x2):
@@ -353,12 +353,12 @@ def heaviside(x1, x2):  # pylint: disable=missing-function-docstring
   return y
 
 
-@np_utils.np_doc(np.hypot)
+@np_utils.np_doc('hypot')
 def hypot(x1, x2):
   return sqrt(square(x1) + square(x2))
 
 
-@np_utils.np_doc(np.kron)
+@np_utils.np_doc('kron')
 def kron(a, b):  # pylint: disable=missing-function-docstring
   # pylint: disable=protected-access,g-complex-comprehension
   a, b = np_array_ops._promote_dtype(a, b)
@@ -389,7 +389,7 @@ def kron(a, b):  # pylint: disable=missing-function-docstring
   return np_array_ops.reshape(a_reshaped * b_reshaped, out_shape)
 
 
-@np_utils.np_doc(np.outer)
+@np_utils.np_doc('outer')
 def outer(a, b):
 
   def f(a, b):
@@ -399,7 +399,7 @@ def outer(a, b):
 
 
 # This can also be implemented via tf.reduce_logsumexp
-@np_utils.np_doc(np.logaddexp)
+@np_utils.np_doc('logaddexp')
 def logaddexp(x1, x2):
   amax = maximum(x1, x2)
   delta = x1 - x2
@@ -409,7 +409,7 @@ def logaddexp(x1, x2):
       amax + log1p(exp(-abs(delta))))
 
 
-@np_utils.np_doc(np.logaddexp2)
+@np_utils.np_doc('logaddexp2')
 def logaddexp2(x1, x2):
   amax = maximum(x1, x2)
   delta = x1 - x2
@@ -419,7 +419,7 @@ def logaddexp2(x1, x2):
       amax + log1p(exp2(-abs(delta))) / np.log(2))
 
 
-@np_utils.np_doc(np.polyval)
+@np_utils.np_doc('polyval')
 def polyval(p, x):  # pylint: disable=missing-function-docstring
 
   def f(p, x):
@@ -437,7 +437,7 @@ def polyval(p, x):  # pylint: disable=missing-function-docstring
   return _bin_op(f, p, x)
 
 
-@np_utils.np_doc(np.isclose)
+@np_utils.np_doc('isclose')
 def isclose(a, b, rtol=1e-05, atol=1e-08, equal_nan=False):  # pylint: disable=missing-docstring
 
   def f(a, b):  # pylint: disable=missing-docstring
@@ -455,7 +455,7 @@ def isclose(a, b, rtol=1e-05, atol=1e-08, equal_nan=False):  # pylint: disable=m
   return _bin_op(f, a, b)
 
 
-@np_utils.np_doc(np.allclose)
+@np_utils.np_doc('allclose')
 def allclose(a, b, rtol=1e-05, atol=1e-08, equal_nan=False):
   return np_array_ops.all(
       isclose(a, b, rtol=rtol, atol=atol, equal_nan=equal_nan))
@@ -490,13 +490,13 @@ def _tf_gcd(x1, x2):  # pylint: disable=missing-function-docstring
 
 
 # Note that np.gcd may not be present in some supported versions of numpy.
-@np_utils.np_doc(None, np_fun_name='gcd')
+@np_utils.np_doc('gcd')
 def gcd(x1, x2):
   return _bin_op(_tf_gcd, x1, x2)
 
 
 # Note that np.lcm may not be present in some supported versions of numpy.
-@np_utils.np_doc(None, np_fun_name='lcm')
+@np_utils.np_doc('lcm')
 def lcm(x1, x2):  # pylint: disable=missing-function-docstring
 
   def f(x1, x2):
@@ -527,22 +527,22 @@ def _bitwise_binary_op(tf_fn, x1, x2):  # pylint: disable=missing-function-docst
   return _bin_op(f, x1, x2)
 
 
-@np_utils.np_doc(np.bitwise_and)
+@np_utils.np_doc('bitwise_and')
 def bitwise_and(x1, x2):
   return _bitwise_binary_op(bitwise_ops.bitwise_and, x1, x2)
 
 
-@np_utils.np_doc(np.bitwise_or)
+@np_utils.np_doc('bitwise_or')
 def bitwise_or(x1, x2):
   return _bitwise_binary_op(bitwise_ops.bitwise_or, x1, x2)
 
 
-@np_utils.np_doc(np.bitwise_xor)
+@np_utils.np_doc('bitwise_xor')
 def bitwise_xor(x1, x2):
   return _bitwise_binary_op(bitwise_ops.bitwise_xor, x1, x2)
 
 
-@np_utils.np_doc(np.bitwise_not)
+@np_utils.np_doc('bitwise_not')
 def bitwise_not(x):
 
   def f(x):
@@ -574,62 +574,62 @@ def _scalar(tf_fn, x, promote_to_float=False):
   return np_utils.tensor_to_ndarray(tf_fn(x.data))
 
 
-@np_utils.np_doc(np.log)
+@np_utils.np_doc('log')
 def log(x):
   return _scalar(math_ops.log, x, True)
 
 
-@np_utils.np_doc(np.exp)
+@np_utils.np_doc('exp')
 def exp(x):
   return _scalar(math_ops.exp, x, True)
 
 
-@np_utils.np_doc(np.sqrt)
+@np_utils.np_doc('sqrt')
 def sqrt(x):
   return _scalar(math_ops.sqrt, x, True)
 
 
-@np_utils.np_doc(np.abs)
+@np_utils.np_doc('abs')
 def abs(x):  # pylint: disable=redefined-builtin
   return _scalar(math_ops.abs, x)
 
 
-@np_utils.np_doc(np.absolute)
+@np_utils.np_doc('absolute')
 def absolute(x):
   return abs(x)
 
 
-@np_utils.np_doc(np.fabs)
+@np_utils.np_doc('fabs')
 def fabs(x):
   return abs(x)
 
 
-@np_utils.np_doc(np.ceil)
+@np_utils.np_doc('ceil')
 def ceil(x):
   return _scalar(math_ops.ceil, x, True)
 
 
-@np_utils.np_doc(np.floor)
+@np_utils.np_doc('floor')
 def floor(x):
   return _scalar(math_ops.floor, x, True)
 
 
-@np_utils.np_doc(np.conj)
+@np_utils.np_doc('conj')
 def conj(x):
   return _scalar(math_ops.conj, x)
 
 
-@np_utils.np_doc(np.negative)
+@np_utils.np_doc('negative')
 def negative(x):
   return _scalar(math_ops.negative, x)
 
 
-@np_utils.np_doc(np.reciprocal)
+@np_utils.np_doc('reciprocal')
 def reciprocal(x):
   return _scalar(math_ops.reciprocal, x)
 
 
-@np_utils.np_doc(np.signbit)
+@np_utils.np_doc('signbit')
 def signbit(x):
 
   def f(x):
@@ -640,67 +640,67 @@ def signbit(x):
   return _scalar(f, x)
 
 
-@np_utils.np_doc(np.sin)
+@np_utils.np_doc('sin')
 def sin(x):
   return _scalar(math_ops.sin, x, True)
 
 
-@np_utils.np_doc(np.cos)
+@np_utils.np_doc('cos')
 def cos(x):
   return _scalar(math_ops.cos, x, True)
 
 
-@np_utils.np_doc(np.tan)
+@np_utils.np_doc('tan')
 def tan(x):
   return _scalar(math_ops.tan, x, True)
 
 
-@np_utils.np_doc(np.sinh)
+@np_utils.np_doc('sinh')
 def sinh(x):
   return _scalar(math_ops.sinh, x, True)
 
 
-@np_utils.np_doc(np.cosh)
+@np_utils.np_doc('cosh')
 def cosh(x):
   return _scalar(math_ops.cosh, x, True)
 
 
-@np_utils.np_doc(np.tanh)
+@np_utils.np_doc('tanh')
 def tanh(x):
   return _scalar(math_ops.tanh, x, True)
 
 
-@np_utils.np_doc(np.arcsin)
+@np_utils.np_doc('arcsin')
 def arcsin(x):
   return _scalar(math_ops.asin, x, True)
 
 
-@np_utils.np_doc(np.arccos)
+@np_utils.np_doc('arccos')
 def arccos(x):
   return _scalar(math_ops.acos, x, True)
 
 
-@np_utils.np_doc(np.arctan)
+@np_utils.np_doc('arctan')
 def arctan(x):
   return _scalar(math_ops.atan, x, True)
 
 
-@np_utils.np_doc(np.arcsinh)
+@np_utils.np_doc('arcsinh')
 def arcsinh(x):
   return _scalar(math_ops.asinh, x, True)
 
 
-@np_utils.np_doc(np.arccosh)
+@np_utils.np_doc('arccosh')
 def arccosh(x):
   return _scalar(math_ops.acosh, x, True)
 
 
-@np_utils.np_doc(np.arctanh)
+@np_utils.np_doc('arctanh')
 def arctanh(x):
   return _scalar(math_ops.atanh, x, True)
 
 
-@np_utils.np_doc(np.deg2rad)
+@np_utils.np_doc('deg2rad')
 def deg2rad(x):
 
   def f(x):
@@ -709,7 +709,7 @@ def deg2rad(x):
   return _scalar(f, x, True)
 
 
-@np_utils.np_doc(np.rad2deg)
+@np_utils.np_doc('rad2deg')
 def rad2deg(x):
   return x * (180.0 / np.pi)
 
@@ -719,7 +719,7 @@ _tf_float_types = [
 ]
 
 
-@np_utils.np_doc(np.angle)
+@np_utils.np_doc('angle')
 def angle(z, deg=False):  # pylint: disable=missing-function-docstring
 
   def f(x):
@@ -735,7 +735,7 @@ def angle(z, deg=False):  # pylint: disable=missing-function-docstring
   return y
 
 
-@np_utils.np_doc(np.cbrt)
+@np_utils.np_doc('cbrt')
 def cbrt(x):
 
   def f(x):
@@ -746,12 +746,12 @@ def cbrt(x):
   return _scalar(f, x, True)
 
 
-@np_utils.np_doc(np.conjugate)
+@np_utils.np_doc('conjugate')
 def conjugate(x):
   return _scalar(math_ops.conj, x)
 
 
-@np_utils.np_doc(np.exp2)
+@np_utils.np_doc('exp2')
 def exp2(x):
 
   def f(x):
@@ -760,12 +760,12 @@ def exp2(x):
   return _scalar(f, x, True)
 
 
-@np_utils.np_doc(np.expm1)
+@np_utils.np_doc('expm1')
 def expm1(x):
   return _scalar(math_ops.expm1, x, True)
 
 
-@np_utils.np_doc(np.fix)
+@np_utils.np_doc('fix')
 def fix(x):
 
   def f(x):
@@ -774,36 +774,36 @@ def fix(x):
   return _scalar(f, x, True)
 
 
-@np_utils.np_doc(np.iscomplex)
+@np_utils.np_doc('iscomplex')
 def iscomplex(x):
   return np_array_ops.imag(x) != 0
 
 
-@np_utils.np_doc(np.isreal)
+@np_utils.np_doc('isreal')
 def isreal(x):
   return np_array_ops.imag(x) == 0
 
 
-@np_utils.np_doc(np.iscomplexobj)
+@np_utils.np_doc('iscomplexobj')
 def iscomplexobj(x):
   x = np_array_ops.array(x)
   return np.issubdtype(x.dtype, np.complexfloating)
 
 
-@np_utils.np_doc(np.isrealobj)
+@np_utils.np_doc('isrealobj')
 def isrealobj(x):
   return not iscomplexobj(x)
 
 
-@np_utils.np_doc(np.isnan)
+@np_utils.np_doc('isnan')
 def isnan(x):
   return _scalar(math_ops.is_nan, x, True)
 
 
-def _make_nan_reduction(onp_reduction, reduction, init_val):
+def _make_nan_reduction(np_fun_name, reduction, init_val):
   """Helper to generate nan* functions."""
 
-  @np_utils.np_doc(onp_reduction)
+  @np_utils.np_doc(np_fun_name)
   def nan_reduction(a, axis=None, dtype=None, keepdims=False):
     a = np_array_ops.array(a)
     v = np_array_ops.array(init_val, dtype=a.dtype)
@@ -816,11 +816,11 @@ def _make_nan_reduction(onp_reduction, reduction, init_val):
   return nan_reduction
 
 
-nansum = _make_nan_reduction(np.nansum, np_array_ops.sum, 0)
-nanprod = _make_nan_reduction(np.nanprod, np_array_ops.prod, 1)
+nansum = _make_nan_reduction('nansum', np_array_ops.sum, 0)
+nanprod = _make_nan_reduction('nanprod', np_array_ops.prod, 1)
 
 
-@np_utils.np_doc(np.nanmean)
+@np_utils.np_doc('nanmean')
 def nanmean(a, axis=None, dtype=None, keepdims=None):  # pylint: disable=missing-docstring
   a = np_array_ops.array(a)
   if np.issubdtype(a.dtype, np.bool_) or np.issubdtype(a.dtype, np.integer):
@@ -833,47 +833,47 @@ def nanmean(a, axis=None, dtype=None, keepdims=None):  # pylint: disable=missing
   return nansum(a, axis=axis, dtype=dtype, keepdims=keepdims) / normalizer
 
 
-@np_utils.np_doc(np.isfinite)
+@np_utils.np_doc('isfinite')
 def isfinite(x):
   return _scalar(math_ops.is_finite, x, True)
 
 
-@np_utils.np_doc(np.isinf)
+@np_utils.np_doc('isinf')
 def isinf(x):
   return _scalar(math_ops.is_inf, x, True)
 
 
-@np_utils.np_doc(np.isneginf)
+@np_utils.np_doc('isneginf')
 def isneginf(x):
   return x == np_array_ops.full_like(x, -np.inf)
 
 
-@np_utils.np_doc(np.isposinf)
+@np_utils.np_doc('isposinf')
 def isposinf(x):
   return x == np_array_ops.full_like(x, np.inf)
 
 
-@np_utils.np_doc(np.log2)
+@np_utils.np_doc('log2')
 def log2(x):
   return log(x) / np.log(2)
 
 
-@np_utils.np_doc(np.log10)
+@np_utils.np_doc('log10')
 def log10(x):
   return log(x) / np.log(10)
 
 
-@np_utils.np_doc(np.log1p)
+@np_utils.np_doc('log1p')
 def log1p(x):
   return _scalar(math_ops.log1p, x, True)
 
 
-@np_utils.np_doc(np.positive)
+@np_utils.np_doc('positive')
 def positive(x):
   return _scalar(lambda x: x, x)
 
 
-@np_utils.np_doc(np.sinc)
+@np_utils.np_doc('sinc')
 def sinc(x):
 
   def f(x):
@@ -884,12 +884,12 @@ def sinc(x):
   return _scalar(f, x, True)
 
 
-@np_utils.np_doc(np.square)
+@np_utils.np_doc('square')
 def square(x):
   return _scalar(math_ops.square, x)
 
 
-@np_utils.np_doc(np.diff)
+@np_utils.np_doc('diff')
 def diff(a, n=1, axis=-1):  # pylint: disable=missing-function-docstring
 
   def f(a):
@@ -964,37 +964,37 @@ def _comparison(tf_fun, x1, x2, cast_bool_to_int=False):
   return np_utils.tensor_to_ndarray(tf_fun(x1, x2))
 
 
-@np_utils.np_doc(np.equal)
+@np_utils.np_doc('equal')
 def equal(x1, x2):
   return _comparison(math_ops.equal, x1, x2)
 
 
-@np_utils.np_doc(np.not_equal)
+@np_utils.np_doc('not_equal')
 def not_equal(x1, x2):
   return _comparison(math_ops.not_equal, x1, x2)
 
 
-@np_utils.np_doc(np.greater)
+@np_utils.np_doc('greater')
 def greater(x1, x2):
   return _comparison(math_ops.greater, x1, x2, True)
 
 
-@np_utils.np_doc(np.greater_equal)
+@np_utils.np_doc('greater_equal')
 def greater_equal(x1, x2):
   return _comparison(math_ops.greater_equal, x1, x2, True)
 
 
-@np_utils.np_doc(np.less)
+@np_utils.np_doc('less')
 def less(x1, x2):
   return _comparison(math_ops.less, x1, x2, True)
 
 
-@np_utils.np_doc(np.less_equal)
+@np_utils.np_doc('less_equal')
 def less_equal(x1, x2):
   return _comparison(math_ops.less_equal, x1, x2, True)
 
 
-@np_utils.np_doc(np.array_equal)
+@np_utils.np_doc('array_equal')
 def array_equal(a1, a2):  # pylint: disable=missing-function-docstring
 
   def f(x1, x2):
@@ -1017,22 +1017,22 @@ def _logical_binary_op(tf_fun, x1, x2):
   return np_utils.tensor_to_ndarray(tf_fun(x1.data, x2.data))
 
 
-@np_utils.np_doc(np.logical_and)
+@np_utils.np_doc('logical_and')
 def logical_and(x1, x2):
   return _logical_binary_op(math_ops.logical_and, x1, x2)
 
 
-@np_utils.np_doc(np.logical_or)
+@np_utils.np_doc('logical_or')
 def logical_or(x1, x2):
   return _logical_binary_op(math_ops.logical_or, x1, x2)
 
 
-@np_utils.np_doc(np.logical_xor)
+@np_utils.np_doc('logical_xor')
 def logical_xor(x1, x2):
   return _logical_binary_op(math_ops.logical_xor, x1, x2)
 
 
-@np_utils.np_doc(np.logical_not)
+@np_utils.np_doc('logical_not')
 def logical_not(x):
   x = np_array_ops.array(x, dtype=np.bool_)
   return np_utils.tensor_to_ndarray(math_ops.logical_not(x.data))
@@ -1047,7 +1047,7 @@ setattr(np_arrays.ndarray, '__eq__', _wrap(equal))
 setattr(np_arrays.ndarray, '__ne__', _wrap(not_equal))
 
 
-@np_utils.np_doc(np.linspace)
+@np_utils.np_doc('linspace')
 def linspace(  # pylint: disable=missing-docstring
     start,
     stop,
@@ -1086,7 +1086,7 @@ def linspace(  # pylint: disable=missing-docstring
     return np_arrays.tensor_to_ndarray(result)
 
 
-@np_utils.np_doc(np.logspace)
+@np_utils.np_doc('logspace')
 def logspace(start, stop, num=50, endpoint=True, base=10.0, dtype=None, axis=0):
   dtype = np_utils.result_type(start, stop, dtype)
   result = linspace(
@@ -1097,7 +1097,7 @@ def logspace(start, stop, num=50, endpoint=True, base=10.0, dtype=None, axis=0):
   return np_arrays.tensor_to_ndarray(result)
 
 
-@np_utils.np_doc(np.geomspace)
+@np_utils.np_doc('geomspace')
 def geomspace(start, stop, num=50, endpoint=True, dtype=None, axis=0):  # pylint: disable=missing-docstring
   dtype = dtype or np_utils.result_type(start, stop, float(num),
                                         np_array_ops.zeros((), dtype))
@@ -1121,13 +1121,13 @@ def geomspace(start, stop, num=50, endpoint=True, dtype=None, axis=0):  # pylint
   return np_utils.tensor_to_ndarray(math_ops.cast(res, dtype))
 
 
-@np_utils.np_doc(np.ptp)
+@np_utils.np_doc('ptp')
 def ptp(a, axis=None, keepdims=None):
   return (np_array_ops.amax(a, axis=axis, keepdims=keepdims) -
           np_array_ops.amin(a, axis=axis, keepdims=keepdims))
 
 
-@np_utils.np_doc_only(np.concatenate)
+@np_utils.np_doc_only('concatenate')
 def concatenate(arys, axis=0):
   if not isinstance(arys, (list, tuple)):
     arys = [arys]
@@ -1138,7 +1138,7 @@ def concatenate(arys, axis=0):
   return np_arrays.tensor_to_ndarray(array_ops.concat(arys, axis))
 
 
-@np_utils.np_doc_only(np.tile)
+@np_utils.np_doc_only('tile')
 def tile(a, reps):  # pylint: disable=missing-function-docstring
   a = np_array_ops.array(a).data
   reps = np_array_ops.array(reps, dtype=dtypes.int32).reshape([-1]).data
@@ -1155,13 +1155,13 @@ def tile(a, reps):  # pylint: disable=missing-function-docstring
   return np_arrays.tensor_to_ndarray(array_ops.tile(a, reps))
 
 
-@np_utils.np_doc(np.count_nonzero)
+@np_utils.np_doc('count_nonzero')
 def count_nonzero(a, axis=None):
   return np_arrays.tensor_to_ndarray(
       math_ops.count_nonzero(np_array_ops.array(a).data, axis))
 
 
-@np_utils.np_doc(np.argsort)
+@np_utils.np_doc('argsort')
 def argsort(a, axis=-1, kind='quicksort', order=None):  # pylint: disable=missing-docstring
   # TODO(nareshmodi): make string tensors also work.
   if kind not in ('quicksort', 'stable'):
@@ -1186,7 +1186,7 @@ def argsort(a, axis=-1, kind='quicksort', order=None):  # pylint: disable=missin
   return np_array_ops.array(tf_ans, dtype=np.intp)
 
 
-@np_utils.np_doc(np.sort)
+@np_utils.np_doc('sort')
 def sort(a, axis=-1, kind='quicksort', order=None):  # pylint: disable=missing-docstring
   if kind != 'quicksort':
     raise ValueError("Only 'quicksort' is supported.")
@@ -1212,17 +1212,17 @@ def _argminmax(fn, a, axis=None):
   return np_utils.tensor_to_ndarray(fn(input=a_t, axis=axis))
 
 
-@np_utils.np_doc(np.argmax)
+@np_utils.np_doc('argmax')
 def argmax(a, axis=None):
   return _argminmax(math_ops.argmax, a, axis)
 
 
-@np_utils.np_doc(np.argmin)
+@np_utils.np_doc('argmin')
 def argmin(a, axis=None):
   return _argminmax(math_ops.argmin, a, axis)
 
 
-@np_utils.np_doc(np.append)
+@np_utils.np_doc('append')
 def append(arr, values, axis=None):
   if axis is None:
     return concatenate([np_array_ops.ravel(arr), np_array_ops.ravel(values)], 0)
@@ -1230,7 +1230,7 @@ def append(arr, values, axis=None):
     return concatenate([arr, values], axis=axis)
 
 
-@np_utils.np_doc(np.average)
+@np_utils.np_doc('average')
 def average(a, axis=None, weights=None, returned=False):  # pylint: disable=missing-docstring
   if axis is not None and not isinstance(axis, six.integer_types):
     # TODO(wangpeng): Support tuple of ints as `axis`
@@ -1293,7 +1293,7 @@ def average(a, axis=None, weights=None, returned=False):  # pylint: disable=miss
   return avg
 
 
-@np_utils.np_doc(np.trace)
+@np_utils.np_doc('trace')
 def trace(a, offset=0, axis1=0, axis2=1, dtype=None):  # pylint: disable=missing-docstring
   if dtype:
     dtype = np_utils.result_type(dtype)
@@ -1311,7 +1311,7 @@ def trace(a, offset=0, axis1=0, axis2=1, dtype=None):  # pylint: disable=missing
   return np_array_ops.sum(a, -1, dtype)
 
 
-@np_utils.np_doc(np.meshgrid)
+@np_utils.np_doc('meshgrid')
 def meshgrid(*xi, **kwargs):
   """This currently requires copy=True and sparse=False."""
   sparse = kwargs.get('sparse', False)
diff --git a/tensorflow/python/ops/numpy_ops/np_utils.py b/tensorflow/python/ops/numpy_ops/np_utils.py
index 186e56816fe..04ec38d611c 100644
--- a/tensorflow/python/ops/numpy_ops/np_utils.py
+++ b/tensorflow/python/ops/numpy_ops/np_utils.py
@@ -233,28 +233,71 @@ def _is_compatible_param_kind(a, b):
   return relax(a) == relax(b)
 
 
-def np_doc(np_fun, np_fun_name=None):
-  """Attachs numpy docstring to a function.
+def _prepare_np_fun_name_and_fun(np_fun_name, np_fun):
+  """Mutually propagates information between `np_fun_name` and `np_fun`.
+
+  If one is None and the other is not, we'll try to make the former not None in
+  a best effort.
 
   Args:
-    np_fun: the numpy function whose docstring will be used.
-    np_fun_name: optional name for the np_fun symbol. At least one of np_fun or
+    np_fun_name: name for the np_fun symbol. At least one of np_fun or
       np_fun_name shoud be set.
+    np_fun: the numpy function whose docstring will be used.
 
   Returns:
-    A function decorator that attaches the docstring from `np_fun` to the
-    decorated function.
+    Processed `np_fun_name` and `np_fun`.
   """
+  if np_fun_name is not None:
+    assert isinstance(np_fun_name, str)
+  if np_fun is not None:
+    assert not isinstance(np_fun, str)
   if np_fun is None:
     assert np_fun_name is not None
     try:
       np_fun = getattr(np, str(np_fun_name))
     except AttributeError:
       np_fun = None
-  np_sig = _np_signature(np_fun)
   if np_fun_name is None:
     assert np_fun is not None
     np_fun_name = np_fun.__name__
+  return np_fun_name, np_fun
+
+
+def _np_doc_helper(f, np_f, np_fun_name=None, unsupported_params=None):
+  """Helper to get docs."""
+  assert np_f or np_fun_name
+  if not np_fun_name:
+    np_fun_name = np_f.__name__
+  doc = 'TensorFlow variant of `numpy.%s`.\n\n' % np_fun_name
+  if unsupported_params:
+    doc += 'Unsupported arguments: ' + ', '.join(
+        '`' + name + '`' for name in unsupported_params) + '.\n\n'
+  if _has_docstring(f):
+    doc += f.__doc__
+    doc = _add_blank_line(doc)
+  if _has_docstring(np_f):
+    doc += 'Documentation for `numpy.%s`:\n\n' % np_f.__name__
+    # TODO(wangpeng): It looks like code snippets in numpy doc don't work
+    # correctly with doctest. Fix that and remove the reformatting of the np_f
+    # comment.
+    doc += np_f.__doc__.replace('>>>', '>')
+  return doc
+
+
+def np_doc(np_fun_name, np_fun=None):
+  """Attachs numpy docstring to a function.
+
+  Args:
+    np_fun_name: name for the np_fun symbol. At least one of np_fun or
+      np_fun_name shoud be set.
+    np_fun: (optional) the numpy function whose docstring will be used.
+
+  Returns:
+    A function decorator that attaches the docstring from `np_fun` to the
+    decorated function.
+  """
+  np_fun_name, np_fun = _prepare_np_fun_name_and_fun(np_fun_name, np_fun)
+  np_sig = _np_signature(np_fun)
 
   def decorator(f):
     """The decorator."""
@@ -294,44 +337,24 @@ def np_doc(np_fun, np_fun_name=None):
   return decorator
 
 
-def _np_doc_helper(f, np_f, np_fun_name=None, unsupported_params=None):
-  """Helper to get docs."""
-  if not unsupported_params and not _has_docstring(f) and _has_docstring(np_f):
-    # TODO(wangpeng): It looks like code snippets in numpy doc don't work
-    # correctly with doctest. Fix that and remove the reformatting of the np_f
-    # comment, here and below.
-    return np_f.__doc__.replace('>>>', '>')
-  assert np_f or np_fun_name
-  if not np_fun_name:
-    np_fun_name = np_f.__name__
-  doc = 'TensorFlow variant of `numpy.%s`.\n\n' % np_fun_name
-  if unsupported_params:
-    doc += 'Unsupported arguments: ' + ', '.join(
-        '`' + name + '`' for name in unsupported_params) + '.\n\n'
-  if _has_docstring(f):
-    doc += f.__doc__
-    doc = _add_blank_line(doc)
-  if _has_docstring(np_f):
-    doc += 'Documentation for `numpy.%s`:\n\n' % np_f.__name__
-    doc += np_f.__doc__.replace('>>>', '>')
-  return doc
-
-
-def np_doc_only(np_f):
+def np_doc_only(np_fun_name, np_fun=None):
   """Attachs numpy docstring to a function.
 
   This differs from np_doc in that it doesn't check for a match in signature.
 
   Args:
-    np_f: the numpy function whose docstring will be used.
+    np_fun_name: name for the np_fun symbol. At least one of np_fun or
+      np_fun_name shoud be set.
+    np_fun: (optional) the numpy function whose docstring will be used.
 
   Returns:
-    A function decorator that attaches the docstring from `np_f` to the
+    A function decorator that attaches the docstring from `np_fun` to the
     decorated function.
   """
+  np_fun_name, np_fun = _prepare_np_fun_name_and_fun(np_fun_name, np_fun)
 
   def decorator(f):
-    f.__doc__ = _np_doc_helper(f, np_f)
+    f.__doc__ = _np_doc_helper(f, np_fun, np_fun_name=np_fun_name)
     return f
 
   return decorator
diff --git a/tensorflow/python/ops/numpy_ops/np_utils_test.py b/tensorflow/python/ops/numpy_ops/np_utils_test.py
index 6d0dfa51185..38b51f05e6e 100644
--- a/tensorflow/python/ops/numpy_ops/np_utils_test.py
+++ b/tensorflow/python/ops/numpy_ops/np_utils_test.py
@@ -31,7 +31,7 @@ class UtilsTest(test.TestCase):
       """np_fun docstring."""
       return
 
-    @np_utils.np_doc(np_fun)
+    @np_utils.np_doc(None, np_fun=np_fun)
     def f():
       """f docstring."""
       return
@@ -47,7 +47,7 @@ np_fun docstring."""
 
   def testNpDocName(self):
 
-    @np_utils.np_doc(None, np_fun_name='foo')
+    @np_utils.np_doc('foo')
     def f():
       """f docstring."""
       return
@@ -70,20 +70,20 @@ f docstring.
     # pylint: disable=unused-variable
     with self.assertRaisesRegexp(TypeError, 'Cannot find parameter'):
 
-      @np_utils.np_doc(np_fun)
+      @np_utils.np_doc(None, np_fun=np_fun)
       def f1(a):
         return
 
     with self.assertRaisesRegexp(TypeError, 'is of kind'):
 
-      @np_utils.np_doc(np_fun)
+      @np_utils.np_doc(None, np_fun=np_fun)
       def f2(x, kwargs):
         return
 
     with self.assertRaisesRegexp(TypeError,
                                  'Parameter "y" should have a default value'):
 
-      @np_utils.np_doc(np_fun)
+      @np_utils.np_doc(None, np_fun=np_fun)
       def f3(x, y):
         return
 

From f5a0fdaa0aeff548623811b887c6da34303ab25f Mon Sep 17 00:00:00 2001
From: Lukas Geiger 
Date: Sat, 20 Jun 2020 01:41:44 +0200
Subject: [PATCH 362/867] Use executing_eagerly_outside_functions

---
 .../keras/mixed_precision/experimental/autocast_variable.py   | 4 ++--
 .../mixed_precision/experimental/autocast_variable_test.py    | 3 +++
 2 files changed, 5 insertions(+), 2 deletions(-)

diff --git a/tensorflow/python/keras/mixed_precision/experimental/autocast_variable.py b/tensorflow/python/keras/mixed_precision/experimental/autocast_variable.py
index b60100c7b48..a717fbb41e2 100644
--- a/tensorflow/python/keras/mixed_precision/experimental/autocast_variable.py
+++ b/tensorflow/python/keras/mixed_precision/experimental/autocast_variable.py
@@ -190,7 +190,7 @@ class AutoCastVariable(variables.Variable, core.Tensor):
 
   def _apply_assign_update(
       self, update_fn, value, use_locking=None, name=None, read_value=True):
-    if context.executing_eagerly() or ops.inside_function():
+    if ops.executing_eagerly_outside_functions():
       assign_op = update_fn(value, use_locking, name, False)
       return self if read_value else assign_op
 
@@ -202,7 +202,7 @@ class AutoCastVariable(variables.Variable, core.Tensor):
 
   def _apply_update(self, update_fn, *args, **kwargs):
     update_var = update_fn(*args, **kwargs)
-    if context.executing_eagerly() or ops.inside_function():
+    if ops.executing_eagerly_outside_functions():
       return self
 
     # Fallback to wrapping the returned variable in graph mode if possible
diff --git a/tensorflow/python/keras/mixed_precision/experimental/autocast_variable_test.py b/tensorflow/python/keras/mixed_precision/experimental/autocast_variable_test.py
index 9036109af96..cb5a5d7cb3f 100644
--- a/tensorflow/python/keras/mixed_precision/experimental/autocast_variable_test.py
+++ b/tensorflow/python/keras/mixed_precision/experimental/autocast_variable_test.py
@@ -345,6 +345,9 @@ class AutoCastVariableTest(test.TestCase, parameterized.TestCase):
 
   @combinations.generate(maybe_distribute)
   def test_assign_tf_function(self, distribution):
+    if not context.executing_eagerly():
+      self.skipTest("Test is not compatible with graph mode")
+
     with distribution.scope():
       x = get_var(0., dtypes.float32)
       x = autocast_variable.create_autocast_variable(x)

From 2c5e5a643cc6ba68da46e4e2058f178434848dd8 Mon Sep 17 00:00:00 2001
From: Peng Wang 
Date: Fri, 19 Jun 2020 16:37:27 -0700
Subject: [PATCH 363/867] [TF-numpy] Adds an accessor class for numpy_ops, in
 order to be tf_exported'ed.

PiperOrigin-RevId: 317402589
Change-Id: I6bb5f4f9d3b42cf8c2653d60d80c20b37d0bd59f
---
 tensorflow/python/ops/numpy_ops/BUILD         |  1 +
 .../python/ops/numpy_ops/np_accessor.py       | 32 +++++++++++++++++++
 2 files changed, 33 insertions(+)
 create mode 100644 tensorflow/python/ops/numpy_ops/np_accessor.py

diff --git a/tensorflow/python/ops/numpy_ops/BUILD b/tensorflow/python/ops/numpy_ops/BUILD
index 3f18a7b3e01..3479a622bc0 100644
--- a/tensorflow/python/ops/numpy_ops/BUILD
+++ b/tensorflow/python/ops/numpy_ops/BUILD
@@ -11,6 +11,7 @@ py_library(
     name = "numpy",
     srcs = [
         "__init__.py",
+        "np_accessor.py",
         "np_array_ops.py",
         "np_arrays.py",
         "np_dtypes.py",
diff --git a/tensorflow/python/ops/numpy_ops/np_accessor.py b/tensorflow/python/ops/numpy_ops/np_accessor.py
new file mode 100644
index 00000000000..64786d2c50a
--- /dev/null
+++ b/tensorflow/python/ops/numpy_ops/np_accessor.py
@@ -0,0 +1,32 @@
+# Copyright 2020 The TensorFlow Authors. All Rights Reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+# ==============================================================================
+"""An accessor class for numpy_ops contents."""
+
+from __future__ import absolute_import
+from __future__ import division
+from __future__ import print_function
+
+from tensorflow.python.ops import numpy_ops
+
+
+class Numpy:
+  """An accessor class that forwards attribute accesses to module `numpy_ops`.
+  """
+
+  def __getattr__(self, attr):
+    return getattr(numpy_ops, attr)
+
+
+numpy = Numpy()

From f51b649394b8b0c2cbf8179ebd4b64b5a915110b Mon Sep 17 00:00:00 2001
From: Austin Anderson 
Date: Fri, 19 Jun 2020 16:39:28 -0700
Subject: [PATCH 364/867] Experimental internal CI changes

PiperOrigin-RevId: 317402932
Change-Id: Ibaef72f01e06f4518b85b21f82000599eed7e4bd
---
 .../per_release/scripts/nonpip_gpu.sh         | 75 +++++++++++++++++++
 1 file changed, 75 insertions(+)
 create mode 100644 tensorflow/tools/ci_build/per_release/scripts/nonpip_gpu.sh

diff --git a/tensorflow/tools/ci_build/per_release/scripts/nonpip_gpu.sh b/tensorflow/tools/ci_build/per_release/scripts/nonpip_gpu.sh
new file mode 100644
index 00000000000..6fd7c3d5854
--- /dev/null
+++ b/tensorflow/tools/ci_build/per_release/scripts/nonpip_gpu.sh
@@ -0,0 +1,75 @@
+#!/bin/bash
+# Copyright 2019 The TensorFlow Authors. All Rights Reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+# ==============================================================================
+set -e
+set -x
+
+if [[ -n "${KOKORO_ARTIFACTS_DIR}" ]]; then
+  cd "${KOKORO_ARTIFACTS_DIR}"
+  ls
+  source "$(find "${KOKORO_ARTIFACTS_DIR}" -name "common_google.sh")"
+  cd git/gob-tensorflow
+
+fi
+
+if [[ -z "${TF_KOKORO_PY_VERSION}" ]]; then
+  echo "You must set TF_KOKORO_PY_VERSION, e.g. '3.7', indicating the "
+  echo "Python version to be used for this build."
+  exit 2
+fi
+
+source tensorflow/tools/ci_build/release/common.sh
+
+install_ubuntu_16_pip_deps "pip${TF_KOKORO_PY_VERSION}"
+# Update bazel
+install_bazelisk
+
+# Run configure.
+export TF_NEED_GCP=1
+export TF_NEED_HDFS=1
+export TF_NEED_S3=1
+export TF_NEED_CUDA=1
+export TF_CUDA_VERSION=10
+export TF_CUDNN_VERSION=7
+export TF_NEED_TENSORRT=1
+export TENSORRT_INSTALL_PATH=/usr/local/tensorrt
+export CC_OPT_FLAGS='-mavx'
+export PYTHON_BIN_PATH=$(which "python${TF_KOKORO_PY_VERSION}")
+export TF2_BEHAVIOR=1
+export PROJECT_NAME="tensorflow_gpu"
+export LD_LIBRARY_PATH="/usr/local/cuda:/usr/local/cuda/lib64:/usr/local/cuda/extras/CUPTI/lib64:$TENSORRT_INSTALL_PATH/lib"
+export TF_CUDA_COMPUTE_CAPABILITIES=sm_35,sm_37,sm_52,sm_60,sm_61,compute_70
+
+yes "" | "$PYTHON_BIN_PATH" configure.py
+
+# Get the default test targets for bazel.
+source tensorflow/tools/ci_build/build_scripts/PRESUBMIT_BUILD_TARGETS.sh
+
+# Exclude -no_oss_py36, for example
+tag_filters="gpu,requires-gpu,-no_gpu,-no_oss,-oss_serial,-no_oss_py${TF_KOKORO_PY_VERSION//.}"
+
+set +e
+bazel test --config=cuda --config=opt \
+  --crosstool_top=//third_party/toolchains/preconfig/ubuntu16.04/gcc7_manylinux2010-nvcc-cuda10.1:toolchain \
+  --linkopt=-lrt \
+  --action_env=TF2_BEHAVIOR="${TF2_BEHAVIOR}" \
+  --test_lang_filters=py \
+  --build_tag_filters=${tag_filters} \
+  --test_tag_filters=${tag_filters} \
+  --test_timeout="300,450,1200,3600" --local_test_jobs=4 \
+  --test_output=errors --verbose_failures=true --keep_going \
+  --run_under=//tensorflow/tools/ci_build/gpu_build:parallel_gpu_execute \
+  -- ${DEFAULT_BAZEL_TARGETS} -//tensorflow/lite/...
+test_xml_summary_exit

From 6116b7f9114f28dcffd685222285a8c5f7db3daa Mon Sep 17 00:00:00 2001
From: George Karpenkov 
Date: Fri, 19 Jun 2020 17:06:02 -0700
Subject: [PATCH 365/867] [XLA] [client] Implement a RunAsync overload which
 does not need a vector of shapes

PiperOrigin-RevId: 317406952
Change-Id: I69d8cc8a68ffdfbf70e2969f5df5e6adba7d2e1d
---
 tensorflow/compiler/xla/client/local_client.cc | 10 ++++++++++
 tensorflow/compiler/xla/client/local_client.h  |  3 +++
 2 files changed, 13 insertions(+)

diff --git a/tensorflow/compiler/xla/client/local_client.cc b/tensorflow/compiler/xla/client/local_client.cc
index aa252067e19..5fc9909fa2a 100644
--- a/tensorflow/compiler/xla/client/local_client.cc
+++ b/tensorflow/compiler/xla/client/local_client.cc
@@ -320,6 +320,16 @@ StatusOr LocalExecutable::RunAsync(
   return std::move(outputs);
 }
 
+StatusOr LocalExecutable::RunAsync(
+    std::vector arguments, ExecutableRunOptions run_options) {
+  std::vector argument_shapes;
+  argument_shapes.reserve(arguments.size());
+  for (const ExecutionInput& arg : arguments) {
+    argument_shapes.push_back(&arg.shape());
+  }
+  return RunAsync(argument_shapes, std::move(arguments), run_options);
+}
+
 se::Platform* LocalClient::platform() const {
   return local_service_->backend().platform();
 }
diff --git a/tensorflow/compiler/xla/client/local_client.h b/tensorflow/compiler/xla/client/local_client.h
index 3241ac73d54..8b91f4a1739 100644
--- a/tensorflow/compiler/xla/client/local_client.h
+++ b/tensorflow/compiler/xla/client/local_client.h
@@ -68,6 +68,9 @@ class LocalExecutable {
       absl::Span argument_host_shapes,
       std::vector arguments, ExecutableRunOptions run_options);
 
+  StatusOr RunAsync(std::vector arguments,
+                                     ExecutableRunOptions run_options);
+
   // Return the options used to build the executable.
   const ExecutableBuildOptions& build_options() const { return build_options_; }
 

From f840a6226841eadee32ada80898279b5d9a7ca3b Mon Sep 17 00:00:00 2001
From: Sanjoy Das 
Date: Fri, 19 Jun 2020 17:08:17 -0700
Subject: [PATCH 366/867] Rollback "[TF:TRT] Cosmetic fix."

PiperOrigin-RevId: 317407274
Change-Id: I73cd486acf9091e6678e553ab9b0545288f73324
---
 tensorflow/compiler/tf2tensorrt/common/utils.h       |  6 ++++--
 .../compiler/tf2tensorrt/convert/convert_graph.cc    |  6 ++++--
 .../compiler/tf2tensorrt/convert/convert_graph.h     |  6 ++++--
 .../tf2tensorrt/convert/convert_graph_test.cc        |  6 ++++--
 .../compiler/tf2tensorrt/convert/convert_nodes.cc    |  6 ++++--
 .../compiler/tf2tensorrt/convert/convert_nodes.h     |  6 ++++--
 .../tf2tensorrt/convert/convert_nodes_test.cc        |  6 ++++--
 .../compiler/tf2tensorrt/convert/logger_registry.cc  |  6 ++++--
 .../compiler/tf2tensorrt/convert/logger_registry.h   |  5 ++---
 .../tf2tensorrt/convert/trt_optimization_pass.cc     |  6 ++++--
 .../tf2tensorrt/convert/trt_optimization_pass.h      |  6 ++++--
 .../tf2tensorrt/kernels/get_calibration_data_op.cc   |  6 ++++--
 .../compiler/tf2tensorrt/kernels/trt_engine_op.cc    |  6 ++++--
 .../tf2tensorrt/kernels/trt_engine_op_test.cc        |  6 ++++--
 .../tf2tensorrt/kernels/trt_engine_resource_ops.cc   |  6 ++++--
 .../kernels/trt_engine_resource_ops_test.cc          |  6 ++++--
 .../tf2tensorrt/ops/get_calibration_data_op.cc       |  6 ++++--
 tensorflow/compiler/tf2tensorrt/ops/trt_engine_op.cc |  6 ++++--
 .../tf2tensorrt/ops/trt_engine_resource_ops.cc       |  6 ++++--
 .../compiler/tf2tensorrt/plugin/plugin_cast.cu.cc    |  6 ++++--
 tensorflow/compiler/tf2tensorrt/plugin/trt_plugin.cc |  6 ++++--
 tensorflow/compiler/tf2tensorrt/plugin/trt_plugin.h  |  6 ++++--
 tensorflow/compiler/tf2tensorrt/segment/segment.cc   |  6 ++++--
 tensorflow/compiler/tf2tensorrt/segment/segment.h    |  6 ++++--
 .../compiler/tf2tensorrt/segment/segment_test.cc     |  6 ++++--
 tensorflow/compiler/tf2tensorrt/segment/union_find.h |  6 ++++--
 tensorflow/compiler/tf2tensorrt/tensorrt_test.cc     |  6 ++++--
 .../compiler/tf2tensorrt/utils/trt_allocator.cc      | 12 ++++++++----
 .../compiler/tf2tensorrt/utils/trt_allocator.h       | 12 ++++++++----
 .../compiler/tf2tensorrt/utils/trt_engine_utils.cc   |  6 ++++--
 .../compiler/tf2tensorrt/utils/trt_engine_utils.h    |  6 ++++--
 .../tf2tensorrt/utils/trt_int8_calibrator.cc         |  6 ++++--
 .../compiler/tf2tensorrt/utils/trt_int8_calibrator.h |  6 ++++--
 tensorflow/compiler/tf2tensorrt/utils/trt_logger.cc  |  6 ++++--
 tensorflow/compiler/tf2tensorrt/utils/trt_logger.h   |  6 ++++--
 .../compiler/tf2tensorrt/utils/trt_lru_cache.cc      |  6 ++++--
 .../compiler/tf2tensorrt/utils/trt_lru_cache.h       |  6 ++++--
 .../utils/trt_shape_optimization_profiles.h          |  6 ++++--
 .../utils/trt_shape_optimization_profiles_test.cc    |  6 ++++--
 39 files changed, 162 insertions(+), 83 deletions(-)

diff --git a/tensorflow/compiler/tf2tensorrt/common/utils.h b/tensorflow/compiler/tf2tensorrt/common/utils.h
index b428733ecd4..9ab0145e1ec 100644
--- a/tensorflow/compiler/tf2tensorrt/common/utils.h
+++ b/tensorflow/compiler/tf2tensorrt/common/utils.h
@@ -16,7 +16,8 @@ limitations under the License.
 #ifndef TENSORFLOW_COMPILER_TF2TENSORRT_COMMON_UTILS_H_
 #define TENSORFLOW_COMPILER_TF2TENSORRT_COMMON_UTILS_H_
 
-#if GOOGLE_CUDA && GOOGLE_TENSORRT
+#if GOOGLE_CUDA
+#if GOOGLE_TENSORRT
 
 #include "tensorflow/core/platform/logging.h"
 
@@ -28,6 +29,7 @@ namespace tensorrt {
 }  // namespace tensorrt
 }  // namespace tensorflow
 
-#endif  // GOOGLE_CUDA && GOOGLE_TENSORRT
+#endif
+#endif
 
 #endif  // TENSORFLOW_COMPILER_TF2TENSORRT_COMMON_UTILS_H_
diff --git a/tensorflow/compiler/tf2tensorrt/convert/convert_graph.cc b/tensorflow/compiler/tf2tensorrt/convert/convert_graph.cc
index 5429aaf3362..1c51d51f1c9 100644
--- a/tensorflow/compiler/tf2tensorrt/convert/convert_graph.cc
+++ b/tensorflow/compiler/tf2tensorrt/convert/convert_graph.cc
@@ -53,7 +53,8 @@ limitations under the License.
 #include "tensorflow/core/util/device_name_utils.h"
 #include "tensorflow/tools/graph_transforms/transform_utils.h"
 
-#if GOOGLE_CUDA && GOOGLE_TENSORRT
+#if GOOGLE_CUDA
+#if GOOGLE_TENSORRT
 #include "third_party/gpus/cuda/include/cuda_runtime_api.h"
 #include "third_party/tensorrt/NvInfer.h"
 namespace tensorflow {
@@ -883,4 +884,5 @@ Status ConvertAfterShapes(const ConversionParams& params) {
 }  // namespace tensorrt
 }  // namespace tensorflow
 
-#endif  // GOOGLE_CUDA && GOOGLE_TENSORRT
+#endif  // GOOGLE_TENSORRT
+#endif  // GOOGLE_CUDA
diff --git a/tensorflow/compiler/tf2tensorrt/convert/convert_graph.h b/tensorflow/compiler/tf2tensorrt/convert/convert_graph.h
index d3897e864fa..53ab84a6fa9 100644
--- a/tensorflow/compiler/tf2tensorrt/convert/convert_graph.h
+++ b/tensorflow/compiler/tf2tensorrt/convert/convert_graph.h
@@ -24,7 +24,8 @@ limitations under the License.
 #include "tensorflow/core/lib/core/status.h"
 #include "tensorflow/core/platform/types.h"
 
-#if GOOGLE_CUDA && GOOGLE_TENSORRT
+#if GOOGLE_CUDA
+#if GOOGLE_TENSORRT
 
 namespace tensorflow {
 namespace tensorrt {
@@ -65,6 +66,7 @@ Status RegisterGraphToFunctionLibrary(const GraphDef& segment_graph_def,
 }  // namespace tensorrt
 }  // namespace tensorflow
 
-#endif  // GOOGLE_CUDA && GOOGLE_TENSORRT
+#endif  // GOOGLE_TENSORRT
+#endif  // GOOGLE_CUDA
 
 #endif  // TENSORFLOW_COMPILER_TF2TENSORRT_CONVERT_CONVERT_GRAPH_H_
diff --git a/tensorflow/compiler/tf2tensorrt/convert/convert_graph_test.cc b/tensorflow/compiler/tf2tensorrt/convert/convert_graph_test.cc
index 54fb1d56441..a1f523d6bfa 100644
--- a/tensorflow/compiler/tf2tensorrt/convert/convert_graph_test.cc
+++ b/tensorflow/compiler/tf2tensorrt/convert/convert_graph_test.cc
@@ -34,7 +34,8 @@ limitations under the License.
 #include "tensorflow/core/protobuf/config.pb.h"  // NOLINT
 #include "tensorflow/core/public/session.h"
 
-#if GOOGLE_CUDA && GOOGLE_TENSORRT
+#if GOOGLE_CUDA
+#if GOOGLE_TENSORRT
 
 namespace tensorflow {
 namespace tensorrt {
@@ -230,4 +231,5 @@ TEST_F(ConvertAfterShapesTest, DirectlyConnectedEngines) {
 }  // namespace tensorrt
 }  // namespace tensorflow
 
-#endif  // GOOGLE_CUDA && GOOGLE_TENSORRT
+#endif  // GOOGLE_TENSORRT
+#endif  // GOOGLE_CUDA
diff --git a/tensorflow/compiler/tf2tensorrt/convert/convert_nodes.cc b/tensorflow/compiler/tf2tensorrt/convert/convert_nodes.cc
index 2ec616ba621..96cec556942 100644
--- a/tensorflow/compiler/tf2tensorrt/convert/convert_nodes.cc
+++ b/tensorflow/compiler/tf2tensorrt/convert/convert_nodes.cc
@@ -59,7 +59,8 @@ limitations under the License.
 #include "tensorflow/core/util/env_var.h"
 #include "tensorflow/core/util/strided_slice_op.h"
 
-#if GOOGLE_CUDA && GOOGLE_TENSORRT
+#if GOOGLE_CUDA
+#if GOOGLE_TENSORRT
 #include "third_party/tensorrt/NvInfer.h"
 #include "third_party/tensorrt/NvInferPlugin.h"
 
@@ -6257,4 +6258,5 @@ bool OutputEdgeValidator::operator()(const Edge* out_edge) const {
 }  // namespace tensorrt
 }  // namespace tensorflow
 
-#endif  // GOOGLE_CUDA && GOOGLE_TENSORRT
+#endif  // GOOGLE_TENSORRT
+#endif  // GOOGLE_CUDA
diff --git a/tensorflow/compiler/tf2tensorrt/convert/convert_nodes.h b/tensorflow/compiler/tf2tensorrt/convert/convert_nodes.h
index a621735fad1..7a1276c645c 100644
--- a/tensorflow/compiler/tf2tensorrt/convert/convert_nodes.h
+++ b/tensorflow/compiler/tf2tensorrt/convert/convert_nodes.h
@@ -33,7 +33,8 @@ limitations under the License.
 #include "tensorflow/core/lib/core/status.h"
 #include "tensorflow/stream_executor/lib/statusor.h"
 
-#if GOOGLE_CUDA && GOOGLE_TENSORRT
+#if GOOGLE_CUDA
+#if GOOGLE_TENSORRT
 #include "third_party/tensorrt/NvInfer.h"
 
 namespace tensorflow {
@@ -693,6 +694,7 @@ BinaryOperationMap();
 }  // namespace tensorrt
 }  // namespace tensorflow
 
-#endif  // GOOGLE_CUDA && GOOGLE_TENSORRT
+#endif  // GOOGLE_TENSORRT
+#endif  // GOOGLE_CUDA
 
 #endif  // TENSORFLOW_COMPILER_TF2TENSORRT_CONVERT_CONVERT_NODES_H_
diff --git a/tensorflow/compiler/tf2tensorrt/convert/convert_nodes_test.cc b/tensorflow/compiler/tf2tensorrt/convert/convert_nodes_test.cc
index 53ec9ee7ada..c24b169f651 100644
--- a/tensorflow/compiler/tf2tensorrt/convert/convert_nodes_test.cc
+++ b/tensorflow/compiler/tf2tensorrt/convert/convert_nodes_test.cc
@@ -21,7 +21,8 @@ limitations under the License.
 #include 
 #include 
 
-#if GOOGLE_CUDA && GOOGLE_TENSORRT
+#if GOOGLE_CUDA
+#if GOOGLE_TENSORRT
 
 #include 
 #include 
@@ -6635,4 +6636,5 @@ TEST_F(OpConverterTest, ConvertPad) {
 }  // namespace tensorrt
 }  // namespace tensorflow
 
-#endif  // GOOGLE_CUDA && GOOGLE_TENSORRT
+#endif  // GOOGLE_TENSORRT
+#endif  // GOOGLE_CUDA
diff --git a/tensorflow/compiler/tf2tensorrt/convert/logger_registry.cc b/tensorflow/compiler/tf2tensorrt/convert/logger_registry.cc
index 07c9c2f1ea0..82e68cbb28d 100644
--- a/tensorflow/compiler/tf2tensorrt/convert/logger_registry.cc
+++ b/tensorflow/compiler/tf2tensorrt/convert/logger_registry.cc
@@ -12,7 +12,8 @@ WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
 See the License for the specific language governing permissions and
 limitations under the License.
 ==============================================================================*/
-#if GOOGLE_CUDA && GOOGLE_TENSORRT
+#if GOOGLE_CUDA
+#if GOOGLE_TENSORRT
 
 #include "tensorflow/compiler/tf2tensorrt/convert/logger_registry.h"
 
@@ -57,4 +58,5 @@ LoggerRegistry* GetLoggerRegistry() {
 }  // namespace tensorrt
 }  // namespace tensorflow
 
-#endif  // GOOGLE_CUDA && GOOGLE_TENSORRT
+#endif  // GOOGLE_TENSORRT
+#endif  // GOOGLE_CUDA
diff --git a/tensorflow/compiler/tf2tensorrt/convert/logger_registry.h b/tensorflow/compiler/tf2tensorrt/convert/logger_registry.h
index 2a265cf7caa..45b302742d0 100644
--- a/tensorflow/compiler/tf2tensorrt/convert/logger_registry.h
+++ b/tensorflow/compiler/tf2tensorrt/convert/logger_registry.h
@@ -19,8 +19,7 @@ limitations under the License.
 #include "tensorflow/core/platform/macros.h"
 #include "tensorflow/core/platform/types.h"
 
-#if GOOGLE_CUDA && GOOGLE_TENSORRT
-
+#if GOOGLE_CUDA
 #include "third_party/tensorrt/NvInfer.h"
 
 namespace tensorflow {
@@ -54,5 +53,5 @@ class RegisterLogger {
 }  // namespace tensorrt
 }  // namespace tensorflow
 
-#endif  // GOOGLE_CUDA && GOOGLE_TENSORRT
+#endif  // GOOGLE_CUDA
 #endif  // TENSORFLOW_COMPILER_TF2TENSORRT_CONVERT_LOGGER_REGISTRY_H_
diff --git a/tensorflow/compiler/tf2tensorrt/convert/trt_optimization_pass.cc b/tensorflow/compiler/tf2tensorrt/convert/trt_optimization_pass.cc
index 1cf98d135cb..72f4fe5ef9b 100644
--- a/tensorflow/compiler/tf2tensorrt/convert/trt_optimization_pass.cc
+++ b/tensorflow/compiler/tf2tensorrt/convert/trt_optimization_pass.cc
@@ -28,7 +28,8 @@ limitations under the License.
 #include "tensorflow/core/platform/logging.h"
 #include "tensorflow/core/platform/stacktrace.h"
 
-#if GOOGLE_CUDA && GOOGLE_TENSORRT
+#if GOOGLE_CUDA
+#if GOOGLE_TENSORRT
 namespace tensorflow {
 namespace tensorrt {
 namespace convert {
@@ -301,4 +302,5 @@ static VerboseCustomGraphOptimizerRegistrar TRTOptimizationPass_Registrar(
 }  // namespace tensorrt
 }  // namespace tensorflow
 
-#endif  // GOOGLE_CUDA && GOOGLE_TENSORRT
+#endif
+#endif
diff --git a/tensorflow/compiler/tf2tensorrt/convert/trt_optimization_pass.h b/tensorflow/compiler/tf2tensorrt/convert/trt_optimization_pass.h
index e0aaa5500ab..f79048bb5f6 100644
--- a/tensorflow/compiler/tf2tensorrt/convert/trt_optimization_pass.h
+++ b/tensorflow/compiler/tf2tensorrt/convert/trt_optimization_pass.h
@@ -23,7 +23,8 @@ limitations under the License.
 #include "tensorflow/core/grappler/optimizers/custom_graph_optimizer.h"
 #include "tensorflow/core/platform/logging.h"
 
-#if GOOGLE_CUDA && GOOGLE_TENSORRT
+#if GOOGLE_CUDA
+#if GOOGLE_TENSORRT
 
 namespace tensorflow {
 namespace tensorrt {
@@ -82,5 +83,6 @@ class TRTOptimizationPass : public grappler::CustomGraphOptimizer {
 }  // namespace tensorrt
 }  // namespace tensorflow
 
-#endif  // GOOGLE_CUDA && GOOGLE_TENSORRT
+#endif  // GOOGLE_CUDA
+#endif  // GOOGLE_TENSORRT
 #endif  // TENSORFLOW_COMPILER_TF2TENSORRT_CONVERT_TRT_OPTIMIZATION_PASS_H_
diff --git a/tensorflow/compiler/tf2tensorrt/kernels/get_calibration_data_op.cc b/tensorflow/compiler/tf2tensorrt/kernels/get_calibration_data_op.cc
index 76fb40b9520..3143b06817e 100644
--- a/tensorflow/compiler/tf2tensorrt/kernels/get_calibration_data_op.cc
+++ b/tensorflow/compiler/tf2tensorrt/kernels/get_calibration_data_op.cc
@@ -22,7 +22,8 @@ limitations under the License.
 #include "tensorflow/core/framework/resource_mgr.h"
 #include "tensorflow/core/lib/core/refcount.h"
 
-#if GOOGLE_CUDA && GOOGLE_TENSORRT
+#if GOOGLE_CUDA
+#if GOOGLE_TENSORRT
 
 namespace tensorflow {
 namespace tensorrt {
@@ -66,4 +67,5 @@ REGISTER_KERNEL_BUILDER(Name("GetCalibrationDataOp").Device(DEVICE_GPU),
 }  // namespace tensorrt
 }  // namespace tensorflow
 
-#endif  // GOOGLE_CUDA && GOOGLE_TENSORRT
+#endif  // GOOGLE_TENSORRT
+#endif  // GOOGLE_CUDA
diff --git a/tensorflow/compiler/tf2tensorrt/kernels/trt_engine_op.cc b/tensorflow/compiler/tf2tensorrt/kernels/trt_engine_op.cc
index 1094555a622..98d199ca9ab 100644
--- a/tensorflow/compiler/tf2tensorrt/kernels/trt_engine_op.cc
+++ b/tensorflow/compiler/tf2tensorrt/kernels/trt_engine_op.cc
@@ -48,7 +48,8 @@ limitations under the License.
 #include "tensorflow/core/util/env_var.h"
 #include "tensorflow/stream_executor/lib/statusor.h"
 
-#if GOOGLE_CUDA && GOOGLE_TENSORRT
+#if GOOGLE_CUDA
+#if GOOGLE_TENSORRT
 #include "third_party/gpus/cuda/include/cuda_runtime_api.h"
 #include "third_party/tensorrt/NvInfer.h"
 
@@ -1008,4 +1009,5 @@ REGISTER_KERNEL_BUILDER(Name("TRTEngineOp").Device(DEVICE_GPU), TRTEngineOp);
 }  // namespace tensorrt
 }  // namespace tensorflow
 
-#endif  // GOOGLE_CUDA && GOOGLE_TENSORRT
+#endif  // GOOGLE_TENSORRT
+#endif  // GOOGLE_CUDA
diff --git a/tensorflow/compiler/tf2tensorrt/kernels/trt_engine_op_test.cc b/tensorflow/compiler/tf2tensorrt/kernels/trt_engine_op_test.cc
index 71193dc24cf..a06010de1c7 100644
--- a/tensorflow/compiler/tf2tensorrt/kernels/trt_engine_op_test.cc
+++ b/tensorflow/compiler/tf2tensorrt/kernels/trt_engine_op_test.cc
@@ -50,7 +50,8 @@ limitations under the License.
 #include "tensorflow/core/platform/status.h"
 #include "tensorflow/core/public/version.h"
 
-#if GOOGLE_CUDA && GOOGLE_TENSORRT
+#if GOOGLE_CUDA
+#if GOOGLE_TENSORRT
 
 namespace tensorflow {
 namespace tensorrt {
@@ -305,4 +306,5 @@ TYPED_TEST(TRTEngineOpTest, Basic) {
 }  // namespace tensorrt
 }  // namespace tensorflow
 
-#endif  // GOOGLE_CUDA && GOOGLE_TENSORRT
+#endif  // GOOGLE_TENSORRT
+#endif  // GOOGLE_CUDA
diff --git a/tensorflow/compiler/tf2tensorrt/kernels/trt_engine_resource_ops.cc b/tensorflow/compiler/tf2tensorrt/kernels/trt_engine_resource_ops.cc
index 3b6e7e91d3b..2c5821df6ac 100644
--- a/tensorflow/compiler/tf2tensorrt/kernels/trt_engine_resource_ops.cc
+++ b/tensorflow/compiler/tf2tensorrt/kernels/trt_engine_resource_ops.cc
@@ -33,7 +33,8 @@ limitations under the License.
 #include "tensorflow/core/platform/mutex.h"
 #include "tensorflow/core/platform/thread_annotations.h"
 
-#if GOOGLE_CUDA && GOOGLE_TENSORRT
+#if GOOGLE_CUDA
+#if GOOGLE_TENSORRT
 #include "third_party/tensorrt/NvInfer.h"
 
 namespace tensorflow {
@@ -250,4 +251,5 @@ REGISTER_KERNEL_BUILDER(Name("SerializeTRTResource").Device(DEVICE_GPU),
 }  // namespace tensorrt
 }  // namespace tensorflow
 
-#endif  // GOOGLE_CUDA && GOOGLE_TENSORRT
+#endif  // GOOGLE_TENSORRT
+#endif  // GOOGLE_CUDA
diff --git a/tensorflow/compiler/tf2tensorrt/kernels/trt_engine_resource_ops_test.cc b/tensorflow/compiler/tf2tensorrt/kernels/trt_engine_resource_ops_test.cc
index 6a073ee24d0..4a24160569d 100644
--- a/tensorflow/compiler/tf2tensorrt/kernels/trt_engine_resource_ops_test.cc
+++ b/tensorflow/compiler/tf2tensorrt/kernels/trt_engine_resource_ops_test.cc
@@ -48,7 +48,8 @@ limitations under the License.
 #include "tensorflow/core/platform/tstring.h"
 #include "tensorflow/core/platform/types.h"
 
-#if GOOGLE_CUDA && GOOGLE_TENSORRT
+#if GOOGLE_CUDA
+#if GOOGLE_TENSORRT
 
 namespace tensorflow {
 namespace tensorrt {
@@ -245,4 +246,5 @@ TEST_F(TRTEngineResourceOpsTest, Basic) {
 }  // namespace tensorrt
 }  // namespace tensorflow
 
-#endif  // GOOGLE_CUDA && GOOGLE_TENSORRT
+#endif  // GOOGLE_TENSORRT
+#endif  // GOOGLE_CUDA
diff --git a/tensorflow/compiler/tf2tensorrt/ops/get_calibration_data_op.cc b/tensorflow/compiler/tf2tensorrt/ops/get_calibration_data_op.cc
index 2af3164c3e2..573172b92e6 100644
--- a/tensorflow/compiler/tf2tensorrt/ops/get_calibration_data_op.cc
+++ b/tensorflow/compiler/tf2tensorrt/ops/get_calibration_data_op.cc
@@ -13,7 +13,8 @@ See the License for the specific language governing permissions and
 limitations under the License.
 ==============================================================================*/
 
-#if GOOGLE_CUDA && GOOGLE_TENSORRT
+#if GOOGLE_CUDA
+#if GOOGLE_TENSORRT
 
 #include "tensorflow/core/framework/common_shape_fns.h"
 #include "tensorflow/core/framework/op.h"
@@ -33,4 +34,5 @@ Returns calibration data for the given resource name
 
 }  // namespace tensorflow
 
-#endif  // GOOGLE_CUDA && GOOGLE_TENSORRT
+#endif  // GOOGLE_TENSORRT
+#endif  // GOOGLE_CUDA
diff --git a/tensorflow/compiler/tf2tensorrt/ops/trt_engine_op.cc b/tensorflow/compiler/tf2tensorrt/ops/trt_engine_op.cc
index 2527fe9b910..bd3c2b299a9 100644
--- a/tensorflow/compiler/tf2tensorrt/ops/trt_engine_op.cc
+++ b/tensorflow/compiler/tf2tensorrt/ops/trt_engine_op.cc
@@ -13,7 +13,8 @@ See the License for the specific language governing permissions and
 limitations under the License.
 ==============================================================================*/
 
-#if GOOGLE_CUDA && GOOGLE_TENSORRT
+#if GOOGLE_CUDA
+#if GOOGLE_TENSORRT
 
 #include "tensorflow/core/framework/common_shape_fns.h"
 #include "tensorflow/core/framework/op.h"
@@ -58,4 +59,5 @@ REGISTER_OP("TRTEngineOp")
     .Attr("static_engine: bool = true");
 }  // namespace tensorflow
 
-#endif  // GOOGLE_CUDA && GOOGLE_TENSORRT
+#endif  // GOOGLE_TENSORRT
+#endif  // GOOGLE_CUDA
diff --git a/tensorflow/compiler/tf2tensorrt/ops/trt_engine_resource_ops.cc b/tensorflow/compiler/tf2tensorrt/ops/trt_engine_resource_ops.cc
index 3141092de03..01911de66ec 100644
--- a/tensorflow/compiler/tf2tensorrt/ops/trt_engine_resource_ops.cc
+++ b/tensorflow/compiler/tf2tensorrt/ops/trt_engine_resource_ops.cc
@@ -13,7 +13,8 @@ See the License for the specific language governing permissions and
 limitations under the License.
 ==============================================================================*/
 
-#if GOOGLE_CUDA && GOOGLE_TENSORRT
+#if GOOGLE_CUDA
+#if GOOGLE_TENSORRT
 
 #include "tensorflow/core/framework/common_shape_fns.h"
 #include "tensorflow/core/framework/op.h"
@@ -45,4 +46,5 @@ REGISTER_OP("SerializeTRTResource")
 
 }  // namespace tensorflow
 
-#endif  // GOOGLE_CUDA && GOOGLE_TENSORRT
+#endif  // GOOGLE_TENSORRT
+#endif  // GOOGLE_CUDA
diff --git a/tensorflow/compiler/tf2tensorrt/plugin/plugin_cast.cu.cc b/tensorflow/compiler/tf2tensorrt/plugin/plugin_cast.cu.cc
index 141a7d1f462..4c0d8b0392a 100644
--- a/tensorflow/compiler/tf2tensorrt/plugin/plugin_cast.cu.cc
+++ b/tensorflow/compiler/tf2tensorrt/plugin/plugin_cast.cu.cc
@@ -17,7 +17,8 @@ limitations under the License.
 #include "tensorflow/compiler/tf2tensorrt/plugin/trt_plugin.h"
 #include "tensorflow/core/platform/logging.h"
 
-#if GOOGLE_CUDA && GOOGLE_TENSORRT
+#if GOOGLE_CUDA
+#if GOOGLE_TENSORRT
 #define EIGEN_USE_GPU  // For definition of Eigen::GpuDevice.
 #include "third_party/gpus/cuda/include/cuda_runtime_api.h"
 #include "tensorflow/core/util/gpu_kernel_helper.h"
@@ -233,4 +234,5 @@ REGISTER_TFTRT_PLUGIN(CastPluginCreator);
 }  // namespace tensorrt
 }  // namespace tensorflow
 
-#endif  // GOOGLE_CUDA && GOOGLE_TENSORRT
+#endif  // GOOGLE_CUDA
+#endif  // GOOGLE_TENSORRT
diff --git a/tensorflow/compiler/tf2tensorrt/plugin/trt_plugin.cc b/tensorflow/compiler/tf2tensorrt/plugin/trt_plugin.cc
index 83d5f9b5965..563ce724f43 100644
--- a/tensorflow/compiler/tf2tensorrt/plugin/trt_plugin.cc
+++ b/tensorflow/compiler/tf2tensorrt/plugin/trt_plugin.cc
@@ -17,7 +17,8 @@ limitations under the License.
 
 #include 
 
-#if GOOGLE_CUDA && GOOGLE_TENSORRT
+#if GOOGLE_CUDA
+#if GOOGLE_TENSORRT
 
 namespace tensorflow {
 namespace tensorrt {
@@ -29,4 +30,5 @@ const char* kTfTrtPluginNamespace = "TF";
 }  // namespace tensorrt
 }  // namespace tensorflow
 
-#endif  // GOOGLE_CUDA && GOOGLE_TENSORRT
+#endif  // GOOGLE_CUDA
+#endif  // GOOGLE_TENSORRT
diff --git a/tensorflow/compiler/tf2tensorrt/plugin/trt_plugin.h b/tensorflow/compiler/tf2tensorrt/plugin/trt_plugin.h
index 600ac6683da..bdb046e6c71 100644
--- a/tensorflow/compiler/tf2tensorrt/plugin/trt_plugin.h
+++ b/tensorflow/compiler/tf2tensorrt/plugin/trt_plugin.h
@@ -20,7 +20,8 @@ limitations under the License.
 
 #include "tensorflow/core/platform/logging.h"
 
-#if GOOGLE_CUDA && GOOGLE_TENSORRT
+#if GOOGLE_CUDA
+#if GOOGLE_TENSORRT
 #include "third_party/tensorrt/NvInfer.h"
 
 namespace tensorflow {
@@ -89,6 +90,7 @@ class TrtPluginRegistrar {
 }  // namespace tensorrt
 }  // namespace tensorflow
 
-#endif  // GOOGLE_CUDA && GOOGLE_TENSORRT
+#endif  // GOOGLE_TENSORRT
+#endif  // GOOGLE_CUDA
 
 #endif  // TENSORFLOW_COMPILER_TF2TENSORRT_PLUGIN_TRT_PLUGIN_H_
diff --git a/tensorflow/compiler/tf2tensorrt/segment/segment.cc b/tensorflow/compiler/tf2tensorrt/segment/segment.cc
index d9080b6f69a..32e30006f58 100644
--- a/tensorflow/compiler/tf2tensorrt/segment/segment.cc
+++ b/tensorflow/compiler/tf2tensorrt/segment/segment.cc
@@ -35,7 +35,8 @@ limitations under the License.
 #include "tensorflow/core/platform/types.h"
 #include "tensorflow/core/util/env_var.h"
 
-#if GOOGLE_CUDA && GOOGLE_TENSORRT
+#if GOOGLE_CUDA
+#if GOOGLE_TENSORRT
 
 namespace tensorflow {
 namespace tensorrt {
@@ -1061,4 +1062,5 @@ Status SegmentGraph(const Graph* tf_graph,
 }  // namespace tensorrt
 }  // namespace tensorflow
 
-#endif  // GOOGLE_CUDA && GOOGLE_TENSORRT
+#endif  // GOOGLE_TENSORRT
+#endif  // GOOGLE_CUDA
diff --git a/tensorflow/compiler/tf2tensorrt/segment/segment.h b/tensorflow/compiler/tf2tensorrt/segment/segment.h
index 3f79983cfd2..7295c8f0d9d 100644
--- a/tensorflow/compiler/tf2tensorrt/segment/segment.h
+++ b/tensorflow/compiler/tf2tensorrt/segment/segment.h
@@ -25,7 +25,8 @@ limitations under the License.
 #include "tensorflow/core/lib/core/status.h"
 #include "tensorflow/core/platform/types.h"
 
-#if GOOGLE_CUDA && GOOGLE_TENSORRT
+#if GOOGLE_CUDA
+#if GOOGLE_TENSORRT
 
 namespace tensorflow {
 namespace tensorrt {
@@ -66,6 +67,7 @@ Status SegmentGraph(const Graph* tf_graph,
 }  // namespace tensorrt
 }  // namespace tensorflow
 
-#endif  // GOOGLE_CUDA && GOOGLE_TENSORRT
+#endif  // GOOGLE_TENSORRT
+#endif  // GOOGLE_CUDA
 
 #endif  // TENSORFLOW_COMPILER_TF2TENSORRT_SEGMENT_SEGMENT_H_
diff --git a/tensorflow/compiler/tf2tensorrt/segment/segment_test.cc b/tensorflow/compiler/tf2tensorrt/segment/segment_test.cc
index f3bc5bfbee6..2437481a9c4 100644
--- a/tensorflow/compiler/tf2tensorrt/segment/segment_test.cc
+++ b/tensorflow/compiler/tf2tensorrt/segment/segment_test.cc
@@ -26,7 +26,8 @@ limitations under the License.
 #include "tensorflow/core/platform/types.h"
 #include "tensorflow/core/public/session.h"
 
-#if GOOGLE_CUDA && GOOGLE_TENSORRT
+#if GOOGLE_CUDA
+#if GOOGLE_TENSORRT
 
 namespace tensorflow {
 namespace tensorrt {
@@ -521,4 +522,5 @@ TEST_F(SegmentTest, IncompatibleBatchSizes) {
 }  // namespace tensorrt
 }  // namespace tensorflow
 
-#endif  // GOOGLE_CUDA && GOOGLE_TENSORRT
+#endif  // GOOGLE_TENSORRT
+#endif  // GOOGLE_CUDA
diff --git a/tensorflow/compiler/tf2tensorrt/segment/union_find.h b/tensorflow/compiler/tf2tensorrt/segment/union_find.h
index b53615ec019..70e83c12fca 100644
--- a/tensorflow/compiler/tf2tensorrt/segment/union_find.h
+++ b/tensorflow/compiler/tf2tensorrt/segment/union_find.h
@@ -19,7 +19,8 @@ limitations under the License.
 #include "absl/strings/str_format.h"
 #include "absl/types/optional.h"
 
-#if GOOGLE_CUDA && GOOGLE_TENSORRT
+#if GOOGLE_CUDA
+#if GOOGLE_TENSORRT
 
 namespace tensorflow {
 namespace tensorrt {
@@ -216,6 +217,7 @@ UnionFind* UnionFind::FindRoot() {
 }  // namespace tensorrt
 }  // namespace tensorflow
 
-#endif  // GOOGLE_CUDA && GOOGLE_TENSORRT
+#endif  // GOOGLE_TENSORRT
+#endif  // GOOGLE_CUDA
 
 #endif  // TENSORFLOW_COMPILER_TF2TENSORRT_SEGMENT_UNION_FIND_H_
diff --git a/tensorflow/compiler/tf2tensorrt/tensorrt_test.cc b/tensorflow/compiler/tf2tensorrt/tensorrt_test.cc
index e994d20df33..510591bfe00 100644
--- a/tensorflow/compiler/tf2tensorrt/tensorrt_test.cc
+++ b/tensorflow/compiler/tf2tensorrt/tensorrt_test.cc
@@ -18,7 +18,8 @@ limitations under the License.
 #include "tensorflow/core/platform/stream_executor.h"
 #include "tensorflow/core/platform/test.h"
 
-#if GOOGLE_CUDA && GOOGLE_TENSORRT
+#if GOOGLE_CUDA
+#if GOOGLE_TENSORRT
 #include "third_party/gpus/cuda/include/cuda.h"
 #include "third_party/gpus/cuda/include/cuda_runtime_api.h"
 #include "third_party/tensorrt/NvInfer.h"
@@ -163,4 +164,5 @@ TEST(TensorrtTest, BasicFunctions) {
 }  // namespace
 }  // namespace tensorflow
 
-#endif  // GOOGLE_CUDA && GOOGLE_TENSORRT
+#endif  // GOOGLE_TENSORRT
+#endif  // GOOGLE_CUDA
diff --git a/tensorflow/compiler/tf2tensorrt/utils/trt_allocator.cc b/tensorflow/compiler/tf2tensorrt/utils/trt_allocator.cc
index d4f3a524577..617ea7fad5c 100644
--- a/tensorflow/compiler/tf2tensorrt/utils/trt_allocator.cc
+++ b/tensorflow/compiler/tf2tensorrt/utils/trt_allocator.cc
@@ -17,9 +17,11 @@ limitations under the License.
 
 #include "tensorflow/core/platform/logging.h"
 
-#if GOOGLE_CUDA && GOOGLE_TENSORRT
+#if GOOGLE_CUDA
+#if GOOGLE_TENSORRT
 #include "third_party/gpus/cuda/include/cuda_runtime_api.h"
-#endif  // GOOGLE_CUDA && GOOGLE_TENSORRT
+#endif  // GOOGLE_TENSORRT
+#endif  // GOOGLE_CUDA
 
 namespace tensorflow {
 namespace tensorrt {
@@ -50,7 +52,8 @@ void* Align(uint64_t alignment, uint64_t size, void*& ptr, uint64_t& space) {
 }  // namespace tensorrt
 }  // namespace tensorflow
 
-#if GOOGLE_CUDA && GOOGLE_TENSORRT
+#if GOOGLE_CUDA
+#if GOOGLE_TENSORRT
 
 namespace tensorflow {
 namespace tensorrt {
@@ -110,4 +113,5 @@ void TRTDeviceAllocator::free(void* memory) {
 }  // namespace tensorrt
 }  // namespace tensorflow
 
-#endif  // GOOGLE_CUDA && GOOGLE_TENSORRT
+#endif  // GOOGLE_TENSORRT
+#endif  // GOOGLE_CUDA
diff --git a/tensorflow/compiler/tf2tensorrt/utils/trt_allocator.h b/tensorflow/compiler/tf2tensorrt/utils/trt_allocator.h
index d219a8a14e8..4ab8b52f523 100644
--- a/tensorflow/compiler/tf2tensorrt/utils/trt_allocator.h
+++ b/tensorflow/compiler/tf2tensorrt/utils/trt_allocator.h
@@ -20,9 +20,11 @@ limitations under the License.
 
 #include "tensorflow/core/framework/allocator.h"
 
-#if GOOGLE_CUDA && GOOGLE_TENSORRT
+#if GOOGLE_CUDA
+#if GOOGLE_TENSORRT
 #include "third_party/tensorrt/NvInfer.h"
-#endif  // GOOGLE_CUDA && GOOGLE_TENSORRT
+#endif  // GOOGLE_TENSORRT
+#endif  // GOOGLE_CUDA
 
 namespace tensorflow {
 namespace tensorrt {
@@ -31,7 +33,8 @@ void* Align(uint64_t alignment, uint64_t size, void*& ptr, uint64_t& space);
 }  // namespace tensorrt
 }  // namespace tensorflow
 
-#if GOOGLE_CUDA && GOOGLE_TENSORRT
+#if GOOGLE_CUDA
+#if GOOGLE_TENSORRT
 
 namespace tensorflow {
 namespace tensorrt {
@@ -66,5 +69,6 @@ class TRTDeviceAllocator : public TRTBaseAllocator {
 }  // namespace tensorrt
 }  // namespace tensorflow
 
-#endif  // GOOGLE_CUDA && GOOGLE_TENSORRT
+#endif  // GOOGLE_TENSORRT
+#endif  // GOOGLE_CUDA
 #endif  // TENSORFLOW_COMPILER_TF2TENSORRT_UTILS_TRT_ALLOCATOR_H_
diff --git a/tensorflow/compiler/tf2tensorrt/utils/trt_engine_utils.cc b/tensorflow/compiler/tf2tensorrt/utils/trt_engine_utils.cc
index 8ccfb8b06f0..ed997b267b1 100644
--- a/tensorflow/compiler/tf2tensorrt/utils/trt_engine_utils.cc
+++ b/tensorflow/compiler/tf2tensorrt/utils/trt_engine_utils.cc
@@ -25,7 +25,8 @@ limitations under the License.
 #include "tensorflow/core/lib/core/status.h"
 #include "tensorflow/core/platform/errors.h"
 
-#if GOOGLE_CUDA && GOOGLE_TENSORRT
+#if GOOGLE_CUDA
+#if GOOGLE_TENSORRT
 #include "third_party/tensorrt/NvInfer.h"
 
 namespace tensorflow {
@@ -256,4 +257,5 @@ Status TrtEnqueue(nvinfer1::IExecutionContext* execution_context,
 }  // namespace tensorrt
 }  // namespace tensorflow
 
-#endif  // GOOGLE_CUDA && GOOGLE_TENSORRT
+#endif  // GOOGLE_TENSORRT
+#endif  // GOOGLE_CUDA
diff --git a/tensorflow/compiler/tf2tensorrt/utils/trt_engine_utils.h b/tensorflow/compiler/tf2tensorrt/utils/trt_engine_utils.h
index 1ea4fe28cb4..a471749877a 100644
--- a/tensorflow/compiler/tf2tensorrt/utils/trt_engine_utils.h
+++ b/tensorflow/compiler/tf2tensorrt/utils/trt_engine_utils.h
@@ -24,7 +24,8 @@ limitations under the License.
 #include "tensorflow/core/framework/tensor_shape.h"
 #include "tensorflow/core/lib/core/status.h"
 
-#if GOOGLE_CUDA && GOOGLE_TENSORRT
+#if GOOGLE_CUDA
+#if GOOGLE_TENSORRT
 #include "third_party/tensorrt/NvInfer.h"
 
 namespace tensorflow {
@@ -90,6 +91,7 @@ Status TrtEnqueue(nvinfer1::IExecutionContext* execution_context,
 }  // namespace tensorrt
 }  // namespace tensorflow
 
-#endif  // GOOGLE_CUDA && GOOGLE_TENSORRT
+#endif  // GOOGLE_TENSORRT
+#endif  // GOOGLE_CUDA
 
 #endif  // TENSORFLOW_COMPILER_TF2TENSORRT_UTILS_TRT_ENGINE_UTILS_H_
diff --git a/tensorflow/compiler/tf2tensorrt/utils/trt_int8_calibrator.cc b/tensorflow/compiler/tf2tensorrt/utils/trt_int8_calibrator.cc
index 24271e352a7..554c127fa37 100644
--- a/tensorflow/compiler/tf2tensorrt/utils/trt_int8_calibrator.cc
+++ b/tensorflow/compiler/tf2tensorrt/utils/trt_int8_calibrator.cc
@@ -20,7 +20,8 @@ limitations under the License.
 
 #include "tensorflow/core/platform/logging.h"
 
-#if GOOGLE_CUDA && GOOGLE_TENSORRT
+#if GOOGLE_CUDA
+#if GOOGLE_TENSORRT
 #include "third_party/gpus/cuda/include/cuda_runtime_api.h"
 
 namespace tensorflow {
@@ -146,4 +147,5 @@ TRTInt8Calibrator::~TRTInt8Calibrator() {
 }  // namespace tensorrt
 }  // namespace tensorflow
 
-#endif  // GOOGLE_CUDA && GOOGLE_TENSORRT
+#endif
+#endif
diff --git a/tensorflow/compiler/tf2tensorrt/utils/trt_int8_calibrator.h b/tensorflow/compiler/tf2tensorrt/utils/trt_int8_calibrator.h
index 4c670e85f52..06b39716490 100644
--- a/tensorflow/compiler/tf2tensorrt/utils/trt_int8_calibrator.h
+++ b/tensorflow/compiler/tf2tensorrt/utils/trt_int8_calibrator.h
@@ -22,7 +22,8 @@ limitations under the License.
 #include 
 #include "tensorflow/core/platform/mutex.h"
 
-#if GOOGLE_CUDA && GOOGLE_TENSORRT
+#if GOOGLE_CUDA
+#if GOOGLE_TENSORRT
 
 #include "third_party/gpus/cuda/include/cuda_runtime_api.h"
 #include "third_party/tensorrt/NvInfer.h"
@@ -100,5 +101,6 @@ struct TRTInt8Calibrator : public nvinfer1::IInt8EntropyCalibrator {
 }  // namespace tensorrt
 }  // namespace tensorflow
 
-#endif  // GOOGLE_CUDA && GOOGLE_TENSORRT
+#endif
+#endif
 #endif  // TENSORFLOW_COMPILER_TF2TENSORRT_UTILS_TRT_INT8_CALIBRATOR_H_
diff --git a/tensorflow/compiler/tf2tensorrt/utils/trt_logger.cc b/tensorflow/compiler/tf2tensorrt/utils/trt_logger.cc
index e34bf5e7397..193687ebc8c 100644
--- a/tensorflow/compiler/tf2tensorrt/utils/trt_logger.cc
+++ b/tensorflow/compiler/tf2tensorrt/utils/trt_logger.cc
@@ -15,7 +15,8 @@ limitations under the License.
 
 #include "tensorflow/compiler/tf2tensorrt/utils/trt_logger.h"
 
-#if GOOGLE_CUDA && GOOGLE_TENSORRT
+#if GOOGLE_CUDA
+#if GOOGLE_TENSORRT
 #include "tensorflow/compiler/tf2tensorrt/common/utils.h"
 #include "tensorflow/compiler/tf2tensorrt/convert/logger_registry.h"
 #include "tensorflow/core/platform/logging.h"
@@ -67,4 +68,5 @@ REGISTER_TENSORRT_LOGGER("DefaultLogger", Logger::GetLogger());
 }  // namespace tensorrt
 }  // namespace tensorflow
 
-#endif  // GOOGLE_CUDA && GOOGLE_TENSORRT
+#endif  // GOOGLE_CUDA
+#endif  // GOOGLE_TENSORRT
diff --git a/tensorflow/compiler/tf2tensorrt/utils/trt_logger.h b/tensorflow/compiler/tf2tensorrt/utils/trt_logger.h
index ce6552e8fe9..2ade1b48f47 100644
--- a/tensorflow/compiler/tf2tensorrt/utils/trt_logger.h
+++ b/tensorflow/compiler/tf2tensorrt/utils/trt_logger.h
@@ -18,7 +18,8 @@ limitations under the License.
 
 #include "tensorflow/core/platform/types.h"
 
-#if GOOGLE_CUDA && GOOGLE_TENSORRT
+#if GOOGLE_CUDA
+#if GOOGLE_TENSORRT
 #include "third_party/tensorrt/NvInfer.h"
 
 namespace tensorflow {
@@ -39,6 +40,7 @@ class Logger : public nvinfer1::ILogger {
 }  // namespace tensorrt
 }  // namespace tensorflow
 
-#endif  // GOOGLE_CUDA && GOOGLE_TENSORRT
+#endif  // GOOGLE_TENSORRT
+#endif  // GOOGLE_CUDA
 
 #endif  // TENSORFLOW_COMPILER_TF2TENSORRT_UTILS_TRT_LOGGER_H_
diff --git a/tensorflow/compiler/tf2tensorrt/utils/trt_lru_cache.cc b/tensorflow/compiler/tf2tensorrt/utils/trt_lru_cache.cc
index ee7e6272372..fbcdaad52c0 100644
--- a/tensorflow/compiler/tf2tensorrt/utils/trt_lru_cache.cc
+++ b/tensorflow/compiler/tf2tensorrt/utils/trt_lru_cache.cc
@@ -23,7 +23,8 @@ limitations under the License.
 #include "tensorflow/core/framework/tensor_shape.h"
 #include "tensorflow/core/platform/mutex.h"
 
-#if GOOGLE_CUDA && GOOGLE_TENSORRT
+#if GOOGLE_CUDA
+#if GOOGLE_TENSORRT
 #include "third_party/tensorrt/NvInfer.h"
 
 namespace tensorflow {
@@ -140,4 +141,5 @@ EngineContext* TRTEngineCacheResource::GetEngineContext(const int profile_id) {
 }  // namespace tensorrt
 }  // namespace tensorflow
 
-#endif  // GOOGLE_CUDA && GOOGLE_TENSORRT
+#endif  // GOOGLE_TENSORRT
+#endif  // GOOGLE_CUDA
diff --git a/tensorflow/compiler/tf2tensorrt/utils/trt_lru_cache.h b/tensorflow/compiler/tf2tensorrt/utils/trt_lru_cache.h
index 991b9a949e4..8e345254f75 100644
--- a/tensorflow/compiler/tf2tensorrt/utils/trt_lru_cache.h
+++ b/tensorflow/compiler/tf2tensorrt/utils/trt_lru_cache.h
@@ -115,7 +115,8 @@ class LRUCache {
   }
 };
 
-#if GOOGLE_CUDA && GOOGLE_TENSORRT
+#if GOOGLE_CUDA
+#if GOOGLE_TENSORRT
 
 struct EngineContext {
   EngineContext() {}  // Creates an empty context.
@@ -222,7 +223,8 @@ class TRTEngineCacheResource : public ResourceBase {
   TrtShapeOptimizationProfile profiles_;
 };
 
-#endif  // GOOGLE_CUDA && GOOGLE_TENSORRT
+#endif  // GOOGLE_TENSORRT
+#endif  // GOOGLE_CUDA
 
 }  // namespace tensorrt
 }  // namespace tensorflow
diff --git a/tensorflow/compiler/tf2tensorrt/utils/trt_shape_optimization_profiles.h b/tensorflow/compiler/tf2tensorrt/utils/trt_shape_optimization_profiles.h
index fc688b14139..40c7f5dcf31 100644
--- a/tensorflow/compiler/tf2tensorrt/utils/trt_shape_optimization_profiles.h
+++ b/tensorflow/compiler/tf2tensorrt/utils/trt_shape_optimization_profiles.h
@@ -29,7 +29,8 @@ limitations under the License.
 #include "tensorflow/core/lib/strings/str_util.h"
 #include "tensorflow/core/lib/strings/strcat.h"
 
-#if GOOGLE_CUDA && GOOGLE_TENSORRT
+#if GOOGLE_CUDA
+#if GOOGLE_TENSORRT
 
 #include "third_party/tensorrt/NvInfer.h"
 
@@ -172,5 +173,6 @@ class TrtShapeOptimizationProfile {
 }  // namespace tensorrt
 }  // namespace tensorflow
 
-#endif  // GOOGLE_CUDA && GOOGLE_TENSORRT
+#endif  // GOOGLE_TENSORRT
+#endif  // GOOGLE_CUDA
 #endif  // TENSORFLOW_COMPILER_TF2TENSORRT_UTILS_TRT_SHAPE_OPTIMIZATION_PROFILES_H_
diff --git a/tensorflow/compiler/tf2tensorrt/utils/trt_shape_optimization_profiles_test.cc b/tensorflow/compiler/tf2tensorrt/utils/trt_shape_optimization_profiles_test.cc
index 32c2200fb71..501810587e0 100644
--- a/tensorflow/compiler/tf2tensorrt/utils/trt_shape_optimization_profiles_test.cc
+++ b/tensorflow/compiler/tf2tensorrt/utils/trt_shape_optimization_profiles_test.cc
@@ -13,7 +13,8 @@ See the License for the specific language governing permissions and
 limitations under the License.
 ==============================================================================*/
 
-#if GOOGLE_CUDA && GOOGLE_TENSORRT
+#if GOOGLE_CUDA
+#if GOOGLE_TENSORRT
 
 #include 
 
@@ -213,4 +214,5 @@ TEST_F(TrtShapeOptimizationProfileTest, Dynamic) {
 }  // namespace tensorrt
 }  // namespace tensorflow
 
-#endif  // GOOGLE_CUDA && GOOGLE_TENSORRT
+#endif  // GOOGLE_TENSORRT
+#endif  // GOOGLE_CUDA

From 715b02167d188d15f0205273b93ea6bbd606f4c5 Mon Sep 17 00:00:00 2001
From: Wenhao Jia 
Date: Fri, 19 Jun 2020 17:27:50 -0700
Subject: [PATCH 367/867] Restore TpuPlatform auto registration code.

PiperOrigin-RevId: 317409587
Change-Id: If44d7a39a45c4c7026f70a4d79d965a54c4db295
---
 tensorflow/stream_executor/tpu/tpu_platform.cc | 8 ++++++++
 1 file changed, 8 insertions(+)

diff --git a/tensorflow/stream_executor/tpu/tpu_platform.cc b/tensorflow/stream_executor/tpu/tpu_platform.cc
index 13a845829c1..4bccd822e91 100644
--- a/tensorflow/stream_executor/tpu/tpu_platform.cc
+++ b/tensorflow/stream_executor/tpu/tpu_platform.cc
@@ -134,4 +134,12 @@ void RegisterTpuPlatform() {
   }
 }
 
+REGISTER_MODULE_INITIALIZER(tpu_platform, RegisterTpuPlatform());
+
+// Note that module initialization sequencing is not supported in the
+// open-source project, so this will be a no-op there.
+REGISTER_MODULE_INITIALIZER_SEQUENCE(tpu_platform, multi_platform_manager);
+REGISTER_MODULE_INITIALIZER_SEQUENCE(multi_platform_manager_listener,
+                                     tpu_platform);
+
 }  // namespace tensorflow

From 3427843d707cf166cbd07755bd11a1c7dea76730 Mon Sep 17 00:00:00 2001
From: Akshay Modi 
Date: Fri, 19 Jun 2020 17:33:29 -0700
Subject: [PATCH 368/867] Dist strat interop

PiperOrigin-RevId: 317410192
Change-Id: Ibfd1e3ac143422ccffa5f240075f5ae93a90ad07
---
 .../python/ops/numpy_ops/np_interop_test.py   | 49 +++++++++++++++++++
 1 file changed, 49 insertions(+)

diff --git a/tensorflow/python/ops/numpy_ops/np_interop_test.py b/tensorflow/python/ops/numpy_ops/np_interop_test.py
index f52d3dae78b..9580b787202 100644
--- a/tensorflow/python/ops/numpy_ops/np_interop_test.py
+++ b/tensorflow/python/ops/numpy_ops/np_interop_test.py
@@ -22,8 +22,13 @@ from __future__ import print_function
 import numpy as onp
 
 
+from tensorflow.python.distribute import distribution_strategy_context
+from tensorflow.python.distribute import mirrored_strategy
+from tensorflow.python.distribute import reduce_util
 from tensorflow.python.eager import backprop
+from tensorflow.python.eager import context
 from tensorflow.python.eager import def_function
+from tensorflow.python.framework import config
 from tensorflow.python.framework import constant_op
 from tensorflow.python.framework import ops
 from tensorflow.python.ops import control_flow_ops
@@ -36,6 +41,17 @@ from tensorflow.python.platform import test
 
 class InteropTest(test.TestCase):
 
+  def setUp(self):
+    super(InteropTest, self).setUp()
+    physical_devices = config.list_physical_devices('CPU')
+    configs = config.get_logical_device_configuration(physical_devices[0])
+    if configs is None:
+      logical_devices = [
+          context.LogicalDeviceConfiguration() for _ in range(3)
+      ]
+      config.set_logical_device_configuration(physical_devices[0],
+                                              logical_devices)
+
   def testGradientTapeInterop(self):
     with backprop.GradientTape() as t:
       x = np_array_ops.asarray(3.0)
@@ -139,6 +155,39 @@ class InteropTest(test.TestCase):
 
 #     self.assertEqual(t.numpy(), [1., 2., 3.])
 
+  def testDistStratInterop(self):
+    strategy = mirrored_strategy.MirroredStrategy(
+        devices=['CPU:0', 'CPU:1', 'CPU:2'])
+
+    multiplier = np_array_ops.asarray(5.)
+
+    with strategy.scope():
+      @def_function.function
+      def run():
+        ctx = distribution_strategy_context.get_replica_context()
+        val = np_array_ops.asarray(ctx.replica_id_in_sync_group)
+        return val * multiplier
+
+      distributed_values = strategy.run(run)
+      reduced = strategy.reduce(reduce_util.ReduceOp.SUM,
+                                distributed_values, axis=None)
+
+    values = distributed_values.values
+
+    # Note that this should match the number of virtual CPUs.
+    self.assertLen(values, 3)
+    self.assertIsInstance(values[0], np_arrays.ndarray)
+    self.assertIsInstance(values[1], np_arrays.ndarray)
+    self.assertIsInstance(values[2], np_arrays.ndarray)
+    self.assertAllClose(values[0], 0)
+    self.assertAllClose(values[1], 5)
+    self.assertAllClose(values[2], 10)
+
+    # "strategy.reduce" doesn't rewrap in ndarray.
+    # self.assertIsInstance(reduced, np_arrays.ndarray)
+    self.assertAllClose(reduced, 15)
+
+
 if __name__ == '__main__':
   ops.enable_eager_execution()
   test.main()

From f0d0485b0de521ff273c8d91acbb2fbabe57baa7 Mon Sep 17 00:00:00 2001
From: Gaurav Jain 
Date: Fri, 19 Jun 2020 18:02:12 -0700
Subject: [PATCH 369/867] Rollforward: Add DT_BOOL support to GPU variable ops

Identity on bool had a HostMemory requirement which was causing excessive copies.

PiperOrigin-RevId: 317413034
Change-Id: Ica75743c9d202f5cc5fb8c12a475eda84507f0be
---
 tensorflow/core/kernels/control_flow_ops.cc            | 10 +++++-----
 tensorflow/core/kernels/identity_op.cc                 |  2 +-
 tensorflow/core/kernels/variable_ops.cc                |  3 +--
 .../debug/lib/debug_graph_reconstruction_test.py       |  6 +++---
 tensorflow/python/ops/control_flow_ops_test.py         |  6 +++---
 5 files changed, 13 insertions(+), 14 deletions(-)

diff --git a/tensorflow/core/kernels/control_flow_ops.cc b/tensorflow/core/kernels/control_flow_ops.cc
index accb2c59540..1a0082c6a3b 100644
--- a/tensorflow/core/kernels/control_flow_ops.cc
+++ b/tensorflow/core/kernels/control_flow_ops.cc
@@ -107,15 +107,17 @@ TF_CALL_QUANTIZED_TYPES(REGISTER_GPU_SWITCH);
 TF_CALL_NUMBER_TYPES_NO_INT32(REGISTER_GPU_REF_SWITCH);
 TF_CALL_QUANTIZED_TYPES(REGISTER_GPU_REF_SWITCH);
 TF_CALL_variant(REGISTER_GPU_SWITCH);
+TF_CALL_bool(REGISTER_GPU_SWITCH);
+TF_CALL_bool(REGISTER_GPU_REF_SWITCH);
 
 #undef REGISTER_CPU_SWITCH
 #undef REGISTER_CPU_REF_SWITCH
 #undef REGISTER_GPU_SWITCH
 #undef REGISTER_GPU_REF_SWITCH
 
-// Special GPU kernels for int32 and string.
-// TODO(b/25387198): Also enable int32 in device memory. This kernel
-// registration requires all int32 inputs and outputs to be in host memory.
+// Special GPU kernels for int32, string & resource handles. Requiring all
+// inputs and outputs to be in host memory.
+// TODO(b/25387198): Also enable int32 in device memory.
 #define REGISTER_GPU_HOST_KERNEL(type)                    \
   REGISTER_KERNEL_BUILDER(Name("Switch")                  \
                               .Device(DEVICE_GPU)         \
@@ -145,8 +147,6 @@ TF_CALL_variant(REGISTER_GPU_SWITCH);
 
 REGISTER_GPU_HOST_KERNEL(int32);
 REGISTER_GPU_HOST_REF_KERNEL(int32);
-REGISTER_GPU_HOST_KERNEL(bool);
-REGISTER_GPU_HOST_REF_KERNEL(bool);
 REGISTER_GPU_HOST_KERNEL(tstring);
 REGISTER_GPU_HOST_REF_KERNEL(tstring);
 REGISTER_GPU_HOST_KERNEL(ResourceHandle);
diff --git a/tensorflow/core/kernels/identity_op.cc b/tensorflow/core/kernels/identity_op.cc
index daa8a1ddb25..4b226dd72d4 100644
--- a/tensorflow/core/kernels/identity_op.cc
+++ b/tensorflow/core/kernels/identity_op.cc
@@ -122,6 +122,7 @@ REGISTER_SYCL_HOST_KERNEL(bool);
 
 TF_CALL_NUMBER_TYPES_NO_INT32(REGISTER_GPU_KERNEL);
 REGISTER_GPU_KERNEL(Variant);
+REGISTER_GPU_KERNEL(bool);
 
 #undef REGISTER_GPU_KERNEL
 
@@ -157,7 +158,6 @@ REGISTER_GPU_KERNEL(Variant);
                           IdentityOp)
 
 REGISTER_GPU_HOST_KERNEL(int32);
-REGISTER_GPU_HOST_KERNEL(bool);
 REGISTER_GPU_HOST_KERNEL(tstring);
 REGISTER_GPU_HOST_KERNEL(ResourceHandle);
 
diff --git a/tensorflow/core/kernels/variable_ops.cc b/tensorflow/core/kernels/variable_ops.cc
index 6f5e0b94eca..ccd33e8c75a 100644
--- a/tensorflow/core/kernels/variable_ops.cc
+++ b/tensorflow/core/kernels/variable_ops.cc
@@ -252,8 +252,7 @@ TF_CALL_GPU_NUMBER_TYPES_NO_HALF(REGISTER_SYCL_KERNEL);
 
 TF_CALL_int64(REGISTER_GPU_KERNELS);
 TF_CALL_uint32(REGISTER_GPU_KERNELS);
-TF_CALL_GPU_NUMBER_TYPES(REGISTER_GPU_KERNELS);
-TF_CALL_COMPLEX_TYPES(REGISTER_GPU_KERNELS);
+TF_CALL_GPU_ALL_TYPES(REGISTER_GPU_KERNELS);
 #undef REGISTER_GPU_KERNELS
 #endif  // GOOGLE_CUDA || TENSORFLOW_USE_ROCM
 
diff --git a/tensorflow/python/debug/lib/debug_graph_reconstruction_test.py b/tensorflow/python/debug/lib/debug_graph_reconstruction_test.py
index fb722efab4e..b3baa6e7bc2 100644
--- a/tensorflow/python/debug/lib/debug_graph_reconstruction_test.py
+++ b/tensorflow/python/debug/lib/debug_graph_reconstruction_test.py
@@ -73,9 +73,9 @@ class ReconstructNonDebugGraphTest(test_util.TensorFlowTestCase):
           for attr_key in new_node.attr:
             if attr_key == "parallel_iterations":
               new_node.attr[attr_key].i = 1
-        elif new_node.op == "Switch":
-          # We don't check the inputs to Switch ops as their inputs may be
-          # Send/Recv nodes.
+        elif new_node.op == "Switch" or new_node.op == "Identity":
+          # We don't check the inputs to Switch or Identity ops as their inputs
+          # may be Send/Recv nodes.
           del new_node.input[:]
 
     return output_graph_def
diff --git a/tensorflow/python/ops/control_flow_ops_test.py b/tensorflow/python/ops/control_flow_ops_test.py
index 9254695d988..3ca9bda82f2 100644
--- a/tensorflow/python/ops/control_flow_ops_test.py
+++ b/tensorflow/python/ops/control_flow_ops_test.py
@@ -396,10 +396,10 @@ class CondTest(test_util.TensorFlowTestCase):
         fn2=lambda: math_ops.add(y, 23))
     self.assertEquals(self.evaluate(z), 24)
 
-  @test_util.run_deprecated_v1
+  @test_util.run_v1_only("Exercises Ref variables")
   def testCondModifyBoolPred(self):
-    # This test in particular used to fail only when running in GPU, hence
-    # use_gpu=True.
+    # We want to use the GPU here because we want to ensure that we can update
+    # a boolean ref variable on the GPU.
     with test_util.use_gpu():
       bool_var = variable_scope.get_variable(
           "bool_var", dtype=dtypes.bool, initializer=True)

From aa7ff6aa28977826e7acae379e82da22482b2bf2 Mon Sep 17 00:00:00 2001
From: George Karpenkov 
Date: Fri, 19 Jun 2020 18:16:58 -0700
Subject: [PATCH 370/867] [TF2XLA] Set up aliasing for resource variables even
 when not returning a tuple

PiperOrigin-RevId: 317414582
Change-Id: I45cd1f314331cb86a0257e7b7cf9d0639be84e99
---
 tensorflow/compiler/tf2xla/xla_compiler.cc | 41 +++++++++++++++-------
 1 file changed, 29 insertions(+), 12 deletions(-)

diff --git a/tensorflow/compiler/tf2xla/xla_compiler.cc b/tensorflow/compiler/tf2xla/xla_compiler.cc
index 1cf3e10b774..c1aef3ff690 100644
--- a/tensorflow/compiler/tf2xla/xla_compiler.cc
+++ b/tensorflow/compiler/tf2xla/xla_compiler.cc
@@ -268,6 +268,7 @@ Status BuildComputation(
               return a->arg_num() < b->arg_num();
             });
 
+  std::vector aliases;
   for (const XlaResource* resource : arg_resources) {
     DCHECK_LT(resource->arg_num(), args.size());
     const XlaCompiler::Argument& arg = args[resource->arg_num()];
@@ -289,20 +290,19 @@ Status BuildComputation(
       update.type = resource->type();
       update.shape = resource->shape();
       update.modified = modified;
-      if (is_entry_computation && always_return_tuple &&
+      if (is_entry_computation &&
           arg.resource_kind != XlaResource::kTensorArray &&
           alias_resource_update) {
         // Assuming tuple arg and results are used.
-        int64 output_index = elems.size();
-        if (use_tuple_arg) {
-          builder->SetUpAlias(/*output_index=*/{output_index},
-                              /*param_number=*/0,
-                              /*param_index=*/{update.input_index});
-        } else {
-          builder->SetUpAlias(/*output_index=*/{output_index},
-                              /*param_number=*/update.input_index,
-                              /*param_index=*/{});
-        }
+        xla::ShapeIndex param_index =
+            use_tuple_arg ? xla::ShapeIndex({update.input_index})
+                          : xla::ShapeIndex{};
+        int param_number = use_tuple_arg ? 0 : update.input_index;
+        int64 output_index_num = elems.size();
+        xla::ShapeIndex output_index = xla::ShapeIndex({output_index_num});
+        VLOG(3) << "Storing alias: " << output_index.ToString() << ": ("
+                << param_number << ", " << param_index.ToString() << ")";
+        aliases.push_back({output_index, param_number, param_index});
       }
       for (const auto& grad : resource->tensor_array_gradients()) {
         update.tensor_array_gradients_accessed.insert(grad.first);
@@ -381,8 +381,25 @@ Status BuildComputation(
     xla::XlaScopedShardingAssignment assign_sharding(builder, op_sharding);
     tuple = xla::Tuple(builder, elems);
   }
-  if (!always_return_tuple && elems.size() == 1) {
+  bool returns_tuple = always_return_tuple || elems.size() != 1;
+  VLOG(3) << "Computation returns a tuple=" << returns_tuple;
+  if (!returns_tuple) {
     xla::GetTupleElement(tuple, 0);
+
+    for (xla::XlaBuilder::InputOutputAlias& alias : aliases) {
+      if (alias.output_index == xla::ShapeIndex({0})) {
+        VLOG(3) << "For aliased parameter " << alias.param_number << ": "
+                << alias.param_index.ToString()
+                << " normalizing output_index from {0} to {}, as a scalar is "
+                   "returned from the cluster";
+        alias.output_index = xla::ShapeIndex({});
+      }
+    }
+  }
+
+  for (xla::XlaBuilder::InputOutputAlias& alias : aliases) {
+    builder->SetUpAlias(alias.output_index, alias.param_number,
+                        alias.param_index);
   }
 
   xla::StatusOr computation_status = builder->Build();

From ce7fef0740ebe0e7f92ba1709e1ce04cb5f59b2c Mon Sep 17 00:00:00 2001
From: Chao Mei 
Date: Fri, 19 Jun 2020 19:09:57 -0700
Subject: [PATCH 371/867] Output unconsumed cmdline flags to help using TFLite
 benchmark tool.

PiperOrigin-RevId: 317419402
Change-Id: I599b487a739808bc8f954b798bd54b75c9b715ad
---
 tensorflow/lite/tools/benchmark/benchmark_model.cc |  7 +++++++
 tensorflow/lite/tools/command_line_flags.cc        | 11 ++++++++++-
 tensorflow/lite/tools/command_line_flags.h         |  3 +++
 tensorflow/lite/tools/command_line_flags_test.cc   |  8 ++++++++
 4 files changed, 28 insertions(+), 1 deletion(-)

diff --git a/tensorflow/lite/tools/benchmark/benchmark_model.cc b/tensorflow/lite/tools/benchmark/benchmark_model.cc
index 2a858e7a326..b67faa8c36b 100644
--- a/tensorflow/lite/tools/benchmark/benchmark_model.cc
+++ b/tensorflow/lite/tools/benchmark/benchmark_model.cc
@@ -229,6 +229,13 @@ TfLiteStatus BenchmarkModel::ParseFlags(int* argc, char** argv) {
     TFLITE_LOG(ERROR) << usage;
     return kTfLiteError;
   }
+
+  std::string unconsumed_args =
+      Flags::ArgsToString(*argc, const_cast(argv));
+  if (!unconsumed_args.empty()) {
+    TFLITE_LOG(WARN) << "Unconsumed cmdline flags: " << unconsumed_args;
+  }
+
   return kTfLiteOk;
 }
 
diff --git a/tensorflow/lite/tools/command_line_flags.cc b/tensorflow/lite/tools/command_line_flags.cc
index 92ddb1622c6..4f646ae27f4 100644
--- a/tensorflow/lite/tools/command_line_flags.cc
+++ b/tensorflow/lite/tools/command_line_flags.cc
@@ -325,6 +325,15 @@ std::string Flag::GetTypeName() const {
     usage_text << "\t" << flag.usage_text_ << "\n";
   }
   return usage_text.str();
-}  // namespace tflite
+}
+
+/*static*/ std::string Flags::ArgsToString(int argc, const char** argv) {
+  std::string args;
+  for (int i = 1; i < argc; ++i) {
+    args.append(argv[i]);
+    if (i != argc - 1) args.append(" ");
+  }
+  return args;
+}
 
 }  // namespace tflite
diff --git a/tensorflow/lite/tools/command_line_flags.h b/tensorflow/lite/tools/command_line_flags.h
index 95e64a19e18..4cc09f4b2c5 100644
--- a/tensorflow/lite/tools/command_line_flags.h
+++ b/tensorflow/lite/tools/command_line_flags.h
@@ -140,6 +140,9 @@ class Flags {
   // usage_text strings in flag_list[].
   static std::string Usage(const std::string& cmdline,
                            const std::vector& flag_list);
+
+  // Return a space separated string containing argv[1, ..., argc-1].
+  static std::string ArgsToString(int argc, const char** argv);
 };
 }  // namespace tflite
 
diff --git a/tensorflow/lite/tools/command_line_flags_test.cc b/tensorflow/lite/tools/command_line_flags_test.cc
index 0216d7a0636..afd1264a0ad 100644
--- a/tensorflow/lite/tools/command_line_flags_test.cc
+++ b/tensorflow/lite/tools/command_line_flags_test.cc
@@ -356,5 +356,13 @@ TEST(CommandLineFlagsTest, DuplicateFlagsAndArgs) {
   EXPECT_EQ(argc, 2);
 }
 
+TEST(CommandLineFlagsTest, ArgsToString) {
+  int argc = 3;
+  const char* argv_strings[] = {"program_name", "--some_int=1", "--some_int=2"};
+  std::string args =
+      Flags::ArgsToString(argc, reinterpret_cast(argv_strings));
+  EXPECT_EQ("--some_int=1 --some_int=2", args);
+}
+
 }  // namespace
 }  // namespace tflite

From f3cb68f0ac458281e5b06cd0066607d112adb0e2 Mon Sep 17 00:00:00 2001
From: Andy Ly 
Date: Fri, 19 Jun 2020 19:18:15 -0700
Subject: [PATCH 372/867] Remove assertions checking device map size in
 TPUVariableRuntimeReformattingPass.

It is possible for more than one replicated device to be populated (e.g. replicated host).

PiperOrigin-RevId: 317419924
Change-Id: I1d8fa95654f324557d78adf4570adc6c3cfdabb4
---
 .../transforms/tpu_variable_runtime_reformatting.cc           | 4 ----
 1 file changed, 4 deletions(-)

diff --git a/tensorflow/compiler/mlir/tensorflow/transforms/tpu_variable_runtime_reformatting.cc b/tensorflow/compiler/mlir/tensorflow/transforms/tpu_variable_runtime_reformatting.cc
index ec4a25c6fdd..d88982d9ee7 100644
--- a/tensorflow/compiler/mlir/tensorflow/transforms/tpu_variable_runtime_reformatting.cc
+++ b/tensorflow/compiler/mlir/tensorflow/transforms/tpu_variable_runtime_reformatting.cc
@@ -261,7 +261,6 @@ tf_device::ReplicateOp AddInputsToReplicateOp(
   // placed in logical core 0.
   // TODO(b/148913020): Remove this constraint once model parallelism is
   // supported.
-  assert(devices.size() == 1);
   assert(devices.find(tensorflow::GetDeviceAliasForLogicalCore(0))
              ->getSecond()
              .size() == num_replicas);
@@ -369,9 +368,6 @@ llvm::SmallVector CreateStateVars(
 
   // TODO(b/148913020): Remove this constraint once model parallelism is
   // supported.
-  assert(devices.size() == 1 &&
-         "As model parallelism is not supported yet, tf_device.replicate "
-         "`devices` attribute should have one dictionary element.");
   const auto& device_list =
       devices.find(tensorflow::GetDeviceAliasForLogicalCore(0))->getSecond();
 

From b1fdd334e7443a4faf54cbca63b8dedd46ee48e1 Mon Sep 17 00:00:00 2001
From: Smit Hinsu 
Date: Fri, 19 Jun 2020 19:18:34 -0700
Subject: [PATCH 373/867] Rewrite masks while applying ellipsis mask to
 tf.StridedSlice op

begin_mask and end_mask needs to be adjusted while simplifying StridedSlice op by removing ellipsis_mask. new_axis_mask is already removed before this and we don't yet support shrink_axis_mask.

PiperOrigin-RevId: 317419960
Change-Id: Ie4a5f404f95f5b909065311a54cbbed64d4ccf4b
---
 .../compiler/mlir/lite/tests/prepare-tf.mlir  | 20 +++++++
 .../mlir/lite/transforms/prepare_tf.cc        | 54 ++++++++++---------
 2 files changed, 50 insertions(+), 24 deletions(-)

diff --git a/tensorflow/compiler/mlir/lite/tests/prepare-tf.mlir b/tensorflow/compiler/mlir/lite/tests/prepare-tf.mlir
index e95f3d011e2..719430959d0 100644
--- a/tensorflow/compiler/mlir/lite/tests/prepare-tf.mlir
+++ b/tensorflow/compiler/mlir/lite/tests/prepare-tf.mlir
@@ -528,6 +528,26 @@ func @PadStridedSliceNewAxisMask2(%arg0: tensor<4x64x64x1xf32>) -> tensor<1x4x64
   return %1 : tensor<1x4x64x64xf32>
 }
 
+// CHECK-LABEL: @StridedSliceRewriteMasks
+func @StridedSliceRewriteMasks(%arg0: tensor<8x4x16x2xf32>) -> tensor<8x4x16x1xf32> {
+  %cst = "tf.Const"() {device = "", value = dense<[1, 0, 1]> : tensor<3xi32>} : () -> tensor<3xi32>
+  %cst_0 = "tf.Const"() {device = "", value = dense<[1, 0, 0]> : tensor<3xi32>} : () -> tensor<3xi32>
+  %cst_1 = "tf.Const"() {device = "", value = dense<1> : tensor<3xi32>} : () -> tensor<3xi32>
+
+  // CHECK: %[[CST:.*]] = constant dense<[1, 0, 0, 1]> : tensor<4xi32>
+  // CHECK: %[[CST0:.*]] = constant dense<[1, 0, 0, 0]> : tensor<4xi32>
+  // CHECK: %[[CST1:.*]] = constant dense<1> : tensor<4xi32>
+  // CHECK: %[[RESULT:.*]] = "tf.StridedSlice"(%arg0, %[[CST]], %[[CST0]], %[[CST1]])
+  // CHECK-SAME: begin_mask = 7 : i64
+  // CHECK-SAME: ellipsis_mask = 0 : i64
+  // CHECK-SAME: end_mask = 14 : i64
+  // CHECK-SAME: new_axis_mask = 0 : i64
+  // CHECK-SAME: shrink_axis_mask = 0 : i64
+
+  %0 = "tf.StridedSlice"(%arg0, %cst, %cst_0, %cst_1) {begin_mask = 1 : i64, device = "", ellipsis_mask = 2 : i64, end_mask = 4 : i64, new_axis_mask = 0 : i64, shrink_axis_mask = 0 : i64} : (tensor<8x4x16x2xf32>, tensor<3xi32>, tensor<3xi32>, tensor<3xi32>) -> tensor<8x4x16x1xf32>
+  return %0 : tensor<8x4x16x1xf32>
+}
+
 // CHECK-LABEL: @MatrixSetDiagV2Conversion
 func @MatrixSetDiagV2Conversion(%arg0: tensor<3x3xi32>, %arg1: tensor<3xi32>) -> tensor<3x3xi32> {
   %cst = constant dense<0> : tensor
diff --git a/tensorflow/compiler/mlir/lite/transforms/prepare_tf.cc b/tensorflow/compiler/mlir/lite/transforms/prepare_tf.cc
index 3310c521a5a..6ee988496fa 100644
--- a/tensorflow/compiler/mlir/lite/transforms/prepare_tf.cc
+++ b/tensorflow/compiler/mlir/lite/transforms/prepare_tf.cc
@@ -584,46 +584,50 @@ struct ConvertTFStridedSlice : public RewritePattern {
 
     const int ellipsis_filled_dim_size = input_size - begin_shape[0] + 1;
 
-    llvm::APInt new_begin_mask = strided_slice_op.begin_mask();
-    llvm::APInt new_end_mask = strided_slice_op.end_mask();
+    int64_t begin_mask = strided_slice_op.begin_mask().getSExtValue();
+    int64_t end_mask = strided_slice_op.end_mask().getSExtValue();
+    int64_t new_begin_mask = 0;
+    int64_t new_end_mask = 0;
 
     SmallVector padded_begin;
     SmallVector padded_end;
     SmallVector padded_stride;
 
     // Before the ellipsis.
-    uint64_t index = 1;
-    int count = 0;
-
-    while (index < ellipsis_mask) {
-      padded_begin.push_back(begin_dense_elem_attr.getValue(count));
-      padded_end.push_back(end_dense_elem_attr.getValue(count));
-      padded_stride.push_back(stride_dense_elem_attr.getValue(count));
-      index <<= 1;
-      count++;
+    int index = 0;
+    int new_index = 0;
+    while (((ellipsis_mask >> index) & 1) == 0) {
+      padded_begin.push_back(begin_dense_elem_attr.getValue(index));
+      padded_end.push_back(end_dense_elem_attr.getValue(index));
+      padded_stride.push_back(stride_dense_elem_attr.getValue(index));
+      if ((begin_mask >> index) & 1) new_begin_mask |= (1 << new_index);
+      if ((end_mask >> index) & 1) new_end_mask |= (1 << new_index);
+      ++index;
+      ++new_index;
     }
 
     // Ellipsis.
-    for (int i = 0; i < ellipsis_filled_dim_size; ++i) {
-      new_begin_mask |= ellipsis_mask;
-      new_end_mask |= ellipsis_mask;
+    for (; new_index < index + ellipsis_filled_dim_size; ++new_index) {
+      new_begin_mask |= (1 << new_index);
+      new_end_mask |= (1 << new_index);
 
       // Mimic the begin/end/strides mask behavior.
       padded_begin.push_back(0);
       padded_end.push_back(0);
       padded_stride.push_back(1);
-
-      ellipsis_mask <<= 1;
     }
 
     // Account for ellipsis mask.
-    count++;
+    ++index;
 
     // After the ellipsis.
-    for (; count < begin_shape[0]; ++count) {
-      padded_begin.push_back(begin_dense_elem_attr.getValue(count));
-      padded_end.push_back(end_dense_elem_attr.getValue(count));
-      padded_stride.push_back(stride_dense_elem_attr.getValue(count));
+    for (; index < begin_shape[0]; ++index) {
+      padded_begin.push_back(begin_dense_elem_attr.getValue(index));
+      padded_end.push_back(end_dense_elem_attr.getValue(index));
+      padded_stride.push_back(stride_dense_elem_attr.getValue(index));
+
+      if ((begin_mask >> index) & 1) new_begin_mask |= (1 << new_index);
+      if ((end_mask >> index) & 1) new_end_mask |= (1 << new_index);
     }
 
     auto attribute_type = rewriter.getIntegerType(64);
@@ -645,7 +649,7 @@ struct ConvertTFStridedSlice : public RewritePattern {
         end_op.getResult(), stride_op.getResult(),
         rewriter.getIntegerAttr(attribute_type, new_begin_mask),
         rewriter.getIntegerAttr(attribute_type, new_end_mask),
-        rewriter.getI64IntegerAttr(0),
+        /*ellipsis_maks=*/rewriter.getI64IntegerAttr(0),
         rewriter.getIntegerAttr(attribute_type,
                                 strided_slice_op.new_axis_mask()),
         rewriter.getIntegerAttr(attribute_type,
@@ -655,10 +659,12 @@ struct ConvertTFStridedSlice : public RewritePattern {
 
   LogicalResult matchAndRewrite(Operation *op,
                                 PatternRewriter &rewriter) const override {
-    // TODO(renjieliu): Consider expand the transformation for shrink
-    // mask as well.
     TF::StridedSliceOp strided_slice_op = llvm::cast(op);
 
+    // TODO(renjieliu): Consider expand the transformation for shrink mask as
+    // well.
+    if (strided_slice_op.shrink_axis_mask().getZExtValue()) return failure();
+
     // Handle new axis mask.
     uint64_t new_axis_mask = strided_slice_op.new_axis_mask().getZExtValue();
     if (new_axis_mask != 0) {

From 429c0b423eb17fc7e782b3188e08637c68177c76 Mon Sep 17 00:00:00 2001
From: Robert David 
Date: Fri, 19 Jun 2020 19:22:08 -0700
Subject: [PATCH 374/867] Integer LSTMs: Name scratch arrays Based on what gate
 they are representing. Make naming consistent with float/hybrid versions.

PiperOrigin-RevId: 317420201
Change-Id: Ia9447e51fce1530e75103c4db3759908592af983
---
 tensorflow/lite/kernels/lstm_eval.cc          | 210 ++++++++++--------
 .../calibration/builtin_logging_ops/lstm.cc   |  13 +-
 2 files changed, 130 insertions(+), 93 deletions(-)

diff --git a/tensorflow/lite/kernels/lstm_eval.cc b/tensorflow/lite/kernels/lstm_eval.cc
index b2f3d77912b..e5bdf3e9a1e 100644
--- a/tensorflow/lite/kernels/lstm_eval.cc
+++ b/tensorflow/lite/kernels/lstm_eval.cc
@@ -216,9 +216,8 @@ inline void LstmStepFloat(
     const float* projection_weights_ptr, const float* projection_bias_ptr,
     const TfLiteLSTMParams* params, int n_batch, int n_cell, int n_input,
     int n_aux_input, int n_output, int output_batch_leading_dim,
-    float* output_state_ptr, float* cell_state_ptr, float* input_gate_scratch,
-    float* forget_gate_scratch, float* cell_gate_scratch,
-    float* output_gate_scratch, float* output_ptr) {
+    float* output_state_ptr, float* cell_state_ptr, float* scratch0,
+    float* scratch1, float* scratch2, float* scratch3, float* output_ptr) {
   ruy::profiler::ScopeLabel label("LstmStepFloat");
   // Since we have already checked that weights are all there or none, we can
   // check the existence of only one to the get the condition.
@@ -226,6 +225,12 @@ inline void LstmStepFloat(
   const bool use_peephole = (cell_to_output_weights_ptr != nullptr);
   const bool use_layer_norm = (forget_layer_norm_coefficients_ptr != nullptr);
 
+  // Make named scratch buffers for the different gates.
+  float* input_gate_scratch = scratch0;
+  float* forget_gate_scratch = scratch1;
+  float* cell_gate_scratch = scratch2;
+  float* output_gate_scratch = scratch3;
+
   // Initialize scratch buffers with bias for regular lstm or initialize with
   // zero for layer norm lstm.
   if (use_layer_norm) {
@@ -531,9 +536,8 @@ inline void LstmStepHybrid(
     const int8_t* projection_weights_ptr, float projection_weights_scale,
     const float* projection_bias_ptr, const TfLiteLSTMParams* params,
     int n_batch, int n_cell, int n_input, int n_aux_input, int n_output,
-    int output_batch_leading_dim, float* input_gate_scratch,
-    float* forget_gate_scratch, float* cell_gate_scratch,
-    float* output_gate_scratch, float* scaling_factors,
+    int output_batch_leading_dim, float* scratch0, float* scratch1,
+    float* scratch2, float* scratch3, float* scaling_factors,
     float* scaling_factors_scratch, float* recovered_cell_weights,
     int8_t* quantized_input_ptr, int8_t* quantized_aux_input_ptr,
     int8_t* quantized_output_state_ptr, int8_t* quantized_cell_state_ptr,
@@ -548,6 +552,12 @@ inline void LstmStepHybrid(
   const bool use_peephole = (cell_to_output_weights_ptr != nullptr);
   const bool use_layer_norm = (forget_layer_norm_coefficients_ptr != nullptr);
 
+  // Make named scratch buffers for the different gates.
+  float* input_gate_scratch = scratch0;
+  float* forget_gate_scratch = scratch1;
+  float* cell_gate_scratch = scratch2;
+  float* output_gate_scratch = scratch3;
+
   // Initialize scratch buffers with bias for regular lstm or initialize with
   // zero for layer norm lstm.
   if (use_layer_norm) {
@@ -974,12 +984,12 @@ inline void LstmStepHybrid(
 //
 // Temporary pre-allocated storage for the calculation. Each is of size n_cell *
 // n_batch.
-//   scratch_0
-//   scratch_1
-//   scratch_2
-//   scratch_3
-//   scratch_4
-//   scratch_5: this scratch buffer is created purely for optimizing the
+//   scratch0
+//   scratch1
+//   scratch2
+//   scratch3
+//   scratch4
+//   scratch5: this scratch buffer is created purely for optimizing the
 //              MatrixBatchVectorMultiplyAccumulate.
 //
 // Outputs:
@@ -1047,10 +1057,15 @@ inline void LstmStepInteger(
     const int32_t* projection_effective_bias, int n_batch, int n_cell,
     int n_input, int n_output, int8_t* output_state_ptr,
     int32_t output_state_zp, int16_t* cell_ptr, int8_t* output_ptr,
-    int16_t* scratch_0_ptr, int16_t* scratch_1_ptr, int16_t* scratch_2_ptr,
-    int16_t* scratch_3_ptr, int8_t* scratch_4_ptr, int32_t* scratch_5_ptr,
-    CpuBackendContext* context) {
+    int16_t* scratch0, int16_t* scratch1, int16_t* scratch2, int16_t* scratch3,
+    int8_t* scratch4, int32_t* scratch5, CpuBackendContext* context) {
   ruy::profiler::ScopeLabel label("LstmStepInteger");
+  // Make named scratch buffers for the different gates.
+  int16_t* input_gate_scratch = scratch0;
+  int16_t* forget_gate_scratch = scratch1;
+  int16_t* cell_gate_scratch = scratch2;
+  int16_t* output_gate_scratch = scratch3;
+
   // Get hyper parameters.
   const bool use_cifg = (input_to_input_weight_ptr == nullptr);
   const bool use_peephole = (cell_to_output_weight_ptr != nullptr);
@@ -1072,99 +1087,103 @@ inline void LstmStepInteger(
 
   // Set scratch to 0.
   if (!use_cifg) {
-    std::fill_n(scratch_0_ptr, n_batch * n_cell, 0);
+    std::fill_n(input_gate_scratch, n_batch * n_cell, 0);
   }
-  std::fill_n(scratch_1_ptr, n_batch * n_cell, 0);
-  std::fill_n(scratch_2_ptr, n_batch * n_cell, 0);
-  std::fill_n(scratch_3_ptr, n_batch * n_cell, 0);
+  std::fill_n(forget_gate_scratch, n_batch * n_cell, 0);
+  std::fill_n(cell_gate_scratch, n_batch * n_cell, 0);
+  std::fill_n(output_gate_scratch, n_batch * n_cell, 0);
 
   // Forget gate.
   tensor_utils::MatrixBatchVectorMultiplyAccumulate(
       input_ptr, input_to_forget_effective_bias, input_to_forget_weight_ptr,
       effective_input_to_forget_scale_a, effective_input_to_forget_scale_b,
-      n_batch, n_input, n_cell, 0, scratch_5_ptr, scratch_1_ptr, context);
+      n_batch, n_input, n_cell, 0, scratch5, forget_gate_scratch, context);
 
   tensor_utils::MatrixBatchVectorMultiplyAccumulate(
       output_state_ptr, recurrent_to_forget_effective_bias,
       recurrent_to_forget_weight_ptr, effective_recurrent_to_forget_scale_a,
       effective_recurrent_to_forget_scale_b, n_batch, n_output, n_cell, 0,
-      scratch_5_ptr, scratch_1_ptr, context);
+      scratch5, forget_gate_scratch, context);
   if (use_peephole) {
     tensor_utils::VectorBatchVectorCwiseProductAccumulate(
         cell_to_forget_weight_ptr, n_output, cell_ptr, n_batch,
         effective_cell_to_forget_scale_a, effective_cell_to_forget_scale_b,
-        scratch_1_ptr);
+        forget_gate_scratch);
   }
 
   if (use_layer_norm) {
     tensor_utils::ApplyLayerNorm(
-        scratch_1_ptr, layer_norm_forget_weight_ptr, forget_gate_bias_ptr,
+        forget_gate_scratch, layer_norm_forget_weight_ptr, forget_gate_bias_ptr,
         layer_norm_forget_scale_a, layer_norm_forget_scale_b,
-        forget_variance_guard, n_batch, n_cell, scratch_1_ptr);
+        forget_variance_guard, n_batch, n_cell, forget_gate_scratch);
   }
 
-  tensor_utils::ApplySigmoid(scratch_1_ptr, n_batch, n_cell, scratch_1_ptr);
+  tensor_utils::ApplySigmoid(forget_gate_scratch, n_batch, n_cell,
+                             forget_gate_scratch);
 
-  // Modulation gate.
+  // Cell gate.
   tensor_utils::MatrixBatchVectorMultiplyAccumulate(
       input_ptr, input_to_cell_effective_bias, input_to_cell_weight_ptr,
       effective_input_to_cell_scale_a, effective_input_to_cell_scale_b, n_batch,
-      n_input, n_cell, 0, scratch_5_ptr, scratch_2_ptr, context);
+      n_input, n_cell, 0, scratch5, cell_gate_scratch, context);
 
   tensor_utils::MatrixBatchVectorMultiplyAccumulate(
       output_state_ptr, recurrent_to_cell_effective_bias,
       recurrent_to_cell_weight_ptr, effective_recurrent_to_cell_scale_a,
       effective_recurrent_to_cell_scale_b, n_batch, n_output, n_cell, 0,
-      scratch_5_ptr, scratch_2_ptr, context);
+      scratch5, cell_gate_scratch, context);
 
   if (use_layer_norm) {
-    tensor_utils::ApplyLayerNorm(scratch_2_ptr, layer_norm_cell_weight_ptr,
+    tensor_utils::ApplyLayerNorm(cell_gate_scratch, layer_norm_cell_weight_ptr,
                                  cell_gate_bias_ptr, layer_norm_cell_scale_a,
                                  layer_norm_cell_scale_b, cell_variance_guard,
-                                 n_batch, n_cell, scratch_2_ptr);
+                                 n_batch, n_cell, cell_gate_scratch);
   }
 
-  tensor_utils::ApplyTanh(3, scratch_2_ptr, n_batch, n_cell, scratch_2_ptr);
+  tensor_utils::ApplyTanh(3, cell_gate_scratch, n_batch, n_cell,
+                          cell_gate_scratch);
 
   // Input gate.
   if (use_cifg) {
-    tensor_utils::Sub1Vector(scratch_1_ptr, n_batch * n_cell, scratch_0_ptr);
+    tensor_utils::Sub1Vector(forget_gate_scratch, n_batch * n_cell,
+                             input_gate_scratch);
   } else {
     tensor_utils::MatrixBatchVectorMultiplyAccumulate(
         input_ptr, input_to_input_effective_bias, input_to_input_weight_ptr,
         effective_input_to_input_scale_a, effective_input_to_input_scale_b,
-        n_batch, n_input, n_cell, 0, scratch_5_ptr, scratch_0_ptr, context);
+        n_batch, n_input, n_cell, 0, scratch5, input_gate_scratch, context);
 
     tensor_utils::MatrixBatchVectorMultiplyAccumulate(
         output_state_ptr, recurrent_to_input_effective_bias,
         recurrent_to_input_weight_ptr, effective_recurrent_to_input_scale_a,
         effective_recurrent_to_input_scale_b, n_batch, n_output, n_cell, 0,
-        scratch_5_ptr, scratch_0_ptr, context);
+        scratch5, input_gate_scratch, context);
     if (use_peephole) {
       tensor_utils::VectorBatchVectorCwiseProductAccumulate(
           cell_to_input_weight_ptr, n_output, cell_ptr, n_batch,
           effective_cell_to_input_scale_a, effective_cell_to_input_scale_b,
-          scratch_0_ptr);
+          input_gate_scratch);
     }
 
     if (use_layer_norm) {
       tensor_utils::ApplyLayerNorm(
-          scratch_0_ptr, layer_norm_input_weight_ptr, input_gate_bias_ptr,
+          input_gate_scratch, layer_norm_input_weight_ptr, input_gate_bias_ptr,
           layer_norm_input_scale_a, layer_norm_input_scale_b,
-          input_variance_guard, n_batch, n_cell, scratch_0_ptr);
+          input_variance_guard, n_batch, n_cell, input_gate_scratch);
     }
-    tensor_utils::ApplySigmoid(scratch_0_ptr, n_batch, n_cell, scratch_0_ptr);
+    tensor_utils::ApplySigmoid(input_gate_scratch, n_batch, n_cell,
+                               input_gate_scratch);
   }
 
   // New cell.
-  tensor_utils::CwiseMul(scratch_1_ptr, cell_ptr, n_batch, n_cell, 15,
-                         scratch_1_ptr);
+  tensor_utils::CwiseMul(forget_gate_scratch, cell_ptr, n_batch, n_cell, 15,
+                         forget_gate_scratch);
 
-  tensor_utils::CwiseMul(scratch_0_ptr, scratch_2_ptr, n_batch, n_cell,
-                         30 + cell_scale, scratch_2_ptr);
+  tensor_utils::CwiseMul(input_gate_scratch, cell_gate_scratch, n_batch, n_cell,
+                         30 + cell_scale, cell_gate_scratch);
 
-  tensor_utils::CwiseAdd(scratch_1_ptr, scratch_2_ptr, n_batch, n_cell,
-                         cell_ptr);
+  tensor_utils::CwiseAdd(forget_gate_scratch, cell_gate_scratch, n_batch,
+                         n_cell, cell_ptr);
 
   if (quantized_cell_clip > 0) {
     tensor_utils::CwiseClipping(cell_ptr, quantized_cell_clip, n_batch, n_cell);
@@ -1174,49 +1193,50 @@ inline void LstmStepInteger(
   tensor_utils::MatrixBatchVectorMultiplyAccumulate(
       input_ptr, input_to_output_effective_bias, input_to_output_weight_ptr,
       effective_input_to_output_scale_a, effective_input_to_output_scale_b,
-      n_batch, n_input, n_cell, 0, scratch_5_ptr, scratch_3_ptr, context);
+      n_batch, n_input, n_cell, 0, scratch5, output_gate_scratch, context);
 
   tensor_utils::MatrixBatchVectorMultiplyAccumulate(
       output_state_ptr, recurrent_to_output_effective_bias,
       recurrent_to_output_weight_ptr, effective_recurrent_to_output_scale_a,
       effective_recurrent_to_output_scale_b, n_batch, n_output, n_cell, 0,
-      scratch_5_ptr, scratch_3_ptr, context);
+      scratch5, output_gate_scratch, context);
   if (use_peephole) {
     tensor_utils::VectorBatchVectorCwiseProductAccumulate(
         cell_to_output_weight_ptr, n_output, cell_ptr, n_batch,
         effective_cell_to_output_scale_a, effective_cell_to_output_scale_b,
-        scratch_3_ptr);
+        output_gate_scratch);
   }
 
   if (use_layer_norm) {
     tensor_utils::ApplyLayerNorm(
-        scratch_3_ptr, layer_norm_output_weight_ptr, output_gate_bias_ptr,
+        output_gate_scratch, layer_norm_output_weight_ptr, output_gate_bias_ptr,
         layer_norm_output_scale_a, layer_norm_output_scale_b,
-        output_variance_guard, n_batch, n_cell, scratch_3_ptr);
+        output_variance_guard, n_batch, n_cell, output_gate_scratch);
   }
 
-  tensor_utils::ApplySigmoid(scratch_3_ptr, n_batch, n_cell, scratch_3_ptr);
+  tensor_utils::ApplySigmoid(output_gate_scratch, n_batch, n_cell,
+                             output_gate_scratch);
 
   // Hidden.
   tensor_utils::ApplyTanh(15 + cell_scale, cell_ptr, n_batch, n_cell,
-                          scratch_0_ptr);
+                          input_gate_scratch);
 
-  tensor_utils::CwiseMul(scratch_3_ptr, scratch_0_ptr, effective_hidden_scale_a,
-                         effective_hidden_scale_b, n_batch, n_cell, hidden_zp,
-                         scratch_4_ptr);
+  tensor_utils::CwiseMul(output_gate_scratch, input_gate_scratch,
+                         effective_hidden_scale_a, effective_hidden_scale_b,
+                         n_batch, n_cell, hidden_zp, scratch4);
   // Projection.
   if (use_projection) {
     std::fill_n(output_ptr, n_batch * n_output, 0);
     tensor_utils::MatrixBatchVectorMultiplyAccumulate(
-        scratch_4_ptr, projection_effective_bias, projection_weight_ptr,
+        scratch4, projection_effective_bias, projection_weight_ptr,
         effective_proj_scale_a, effective_proj_scale_b, n_batch, n_cell,
-        n_output, output_state_zp, scratch_5_ptr, output_ptr, context);
+        n_output, output_state_zp, scratch5, output_ptr, context);
     if (quantized_proj_clip > 0) {
       tensor_utils::CwiseClipping(output_ptr, quantized_proj_clip, n_batch,
                                   n_output);
     }
   } else {
-    std::copy_n(scratch_4_ptr, n_batch * n_output, output_ptr);
+    std::copy_n(scratch4, n_batch * n_output, output_ptr);
   }
   std::copy_n(output_ptr, n_batch * n_output, output_state_ptr);
 }
@@ -1300,14 +1320,14 @@ inline void LstmStepInteger(
 //
 // Temporary pre-allocated storage for the calculation. Each is of size n_cell *
 // n_batch.
-//   scratch_0
-//   scratch_1
-//   scratch_2
-//   scratch_3
-//   scratch_4
-//   scratch_5
-//   scratch_6
-//   scratch_7
+//   scratch0
+//   scratch1
+//   scratch2
+//   scratch3
+//   scratch4
+//   scratch5
+//   scratch6
+//   scratch7
 //
 // Outputs:
 //   output_state_ptr - size 'n_batch * n_output'
@@ -1369,6 +1389,12 @@ void LstmStepInteger(
     int8_t* scratch0, int8_t* scratch1, int16_t* scratch2, int16_t* scratch3,
     int16_t* scratch4, int16_t* scratch5, int16_t* scratch6,
     int16_t* scratch7) {
+  // Make named scratch buffers for the different gates.
+  int16_t* input_gate_scratch = scratch5;
+  int16_t* forget_gate_scratch = scratch2;
+  int16_t* cell_gate_scratch = scratch3;
+  int16_t* output_gate_scratch = scratch4;
+
   // Forget gate.
   std::fill_n(scratch0, n_batch * n_cell, 0);
   std::fill_n(scratch1, n_batch * n_cell, 0);
@@ -1386,16 +1412,17 @@ void LstmStepInteger(
   tensor_utils::TwoGateSaturationgAdd(
       scratch0, intermediate_zp[4], scratch1, intermediate_zp[5],
       intermediate_scale_a[2], intermediate_scale_b[2], intermediate_scale_a[3],
-      intermediate_scale_b[3], n_batch, n_cell, scratch2);
+      intermediate_scale_b[3], n_batch, n_cell, forget_gate_scratch);
 
   // Forget gate layer norm.
   tensor_utils::ApplyLayerNormFloat(
-      scratch2, layer_norm_forget_weight_ptr, layer_norm_forget_scale_a,
-      layer_norm_forget_scale_b, forget_gate_bias_ptr, n_batch, n_cell,
-      scratch2);
+      forget_gate_scratch, layer_norm_forget_weight_ptr,
+      layer_norm_forget_scale_a, layer_norm_forget_scale_b,
+      forget_gate_bias_ptr, n_batch, n_cell, forget_gate_scratch);
 
   // Forget gate sigmoid.
-  tensor_utils::ApplySigmoidFloat(scratch2, n_batch, n_cell, scratch2);
+  tensor_utils::ApplySigmoidFloat(forget_gate_scratch, n_batch, n_cell,
+                                  forget_gate_scratch);
 
   // Update gate.
   std::fill_n(scratch0, n_batch * n_cell, 0);
@@ -1413,15 +1440,17 @@ void LstmStepInteger(
   tensor_utils::TwoGateSaturationgAdd(
       scratch0, intermediate_zp[7], scratch1, intermediate_zp[8],
       intermediate_scale_a[4], intermediate_scale_b[4], intermediate_scale_a[5],
-      intermediate_scale_b[5], n_batch, n_cell, scratch3);
+      intermediate_scale_b[5], n_batch, n_cell, cell_gate_scratch);
 
-  // Update gate with layer norm.
+  // Update gate layer norm.
   tensor_utils::ApplyLayerNormFloat(
-      scratch3, layer_norm_cell_weight_ptr, layer_norm_cell_scale_a,
-      layer_norm_cell_scale_b, cell_gate_bias_ptr, n_batch, n_cell, scratch3);
+      cell_gate_scratch, layer_norm_cell_weight_ptr, layer_norm_cell_scale_a,
+      layer_norm_cell_scale_b, cell_gate_bias_ptr, n_batch, n_cell,
+      cell_gate_scratch);
 
   // Update gate tanh.
-  tensor_utils::ApplyTanhFloat(scratch3, n_batch, n_cell, -12, scratch3);
+  tensor_utils::ApplyTanhFloat(cell_gate_scratch, n_batch, n_cell, -12,
+                               cell_gate_scratch);
 
   // Output gate.
   std::fill_n(scratch0, n_batch * n_cell, 0);
@@ -1440,26 +1469,28 @@ void LstmStepInteger(
   tensor_utils::TwoGateSaturationgAdd(
       scratch0, intermediate_zp[10], scratch1, intermediate_zp[11],
       intermediate_scale_a[6], intermediate_scale_b[6], intermediate_scale_a[7],
-      intermediate_scale_b[7], n_batch, n_cell, scratch4);
+      intermediate_scale_b[7], n_batch, n_cell, output_gate_scratch);
 
   // Output gate with layer norm.
   tensor_utils::ApplyLayerNormFloat(
-      scratch4, layer_norm_output_weight_ptr, layer_norm_output_scale_a,
-      layer_norm_output_scale_b, output_gate_bias_ptr, n_batch, n_cell,
-      scratch4);
+      output_gate_scratch, layer_norm_output_weight_ptr,
+      layer_norm_output_scale_a, layer_norm_output_scale_b,
+      output_gate_bias_ptr, n_batch, n_cell, output_gate_scratch);
 
   // Output gate sigmoid.
-  tensor_utils::ApplySigmoidFloat(scratch4, n_batch, n_cell, scratch4);
+  tensor_utils::ApplySigmoidFloat(output_gate_scratch, n_batch, n_cell,
+                                  output_gate_scratch);
 
   // Input gate with cifg
-  tensor_utils::Sub1Vector(scratch2, n_batch * n_cell, scratch5);
+  tensor_utils::Sub1Vector(forget_gate_scratch, n_batch * n_cell,
+                           input_gate_scratch);
 
   // New cell.
-  tensor_utils::CwiseMul(scratch2, cell_ptr, n_batch, n_cell, 15 + 15 - 15,
-                         scratch6);
+  tensor_utils::CwiseMul(forget_gate_scratch, cell_ptr, n_batch, n_cell,
+                         15 + 15 - 15, scratch6);
 
-  tensor_utils::CwiseMul(scratch5, scratch3, n_batch, n_cell, 15 + 15 - 15,
-                         scratch7);
+  tensor_utils::CwiseMul(input_gate_scratch, cell_gate_scratch, n_batch, n_cell,
+                         15 + 15 - 15, scratch7);
 
   tensor_utils::CwiseAdd(scratch6, scratch7, n_batch, n_cell, cell_ptr);
 
@@ -1468,15 +1499,16 @@ void LstmStepInteger(
   }
 
   // Cell to hidden.
-  tensor_utils::ApplyTanhFloat(cell_ptr, n_batch, n_cell, -15, scratch2);
+  tensor_utils::ApplyTanhFloat(cell_ptr, n_batch, n_cell, -15,
+                               forget_gate_scratch);
 
   std::vector hidden(n_batch * n_cell);
-  tensor_utils::CwiseMul(scratch4, scratch2, n_batch, n_cell, 15 + 15 - 15,
-                         scratch3);
+  tensor_utils::CwiseMul(output_gate_scratch, forget_gate_scratch, n_batch,
+                         n_cell, 15 + 15 - 15, cell_gate_scratch);
 
   // Projection.
   tensor_utils::MatrixBatchVectorMultiply(
-      scratch3, projection_weight_ptr, effective_proj_scale_a,
+      cell_gate_scratch, projection_weight_ptr, effective_proj_scale_a,
       effective_proj_scale_b, projection_bias_ptr, n_batch, n_cell, n_output,
       output_state_zp, output_ptr);
 
diff --git a/tensorflow/lite/tools/optimize/calibration/builtin_logging_ops/lstm.cc b/tensorflow/lite/tools/optimize/calibration/builtin_logging_ops/lstm.cc
index 50138442c25..09ce81c1d97 100644
--- a/tensorflow/lite/tools/optimize/calibration/builtin_logging_ops/lstm.cc
+++ b/tensorflow/lite/tools/optimize/calibration/builtin_logging_ops/lstm.cc
@@ -62,11 +62,16 @@ inline void LstmStepWithAuxInput(
     const float* projection_weights_ptr, const float* projection_bias_ptr,
     const TfLiteLSTMParams* params, int n_batch, int n_cell, int n_input,
     int n_aux_input, int n_output, int output_batch_leading_dim,
-    float* output_state_ptr, float* cell_state_ptr, float* input_gate_scratch,
-    float* forget_gate_scratch, float* cell_gate_scratch,
-    float* output_gate_scratch, float* output_ptr, Logger* logger,
-    const std::vector& intermediate_tensor_indexes,
+    float* output_state_ptr, float* cell_state_ptr, float* scratch0,
+    float* scratch1, float* scratch2, float* scratch3, float* output_ptr,
+    Logger* logger, const std::vector& intermediate_tensor_indexes,
     ErrorReporter* error_reporter) {
+  // Make named scratch buffers for the different gates.
+  float* input_gate_scratch = scratch0;
+  float* forget_gate_scratch = scratch1;
+  float* cell_gate_scratch = scratch2;
+  float* output_gate_scratch = scratch3;
+
   // Since we have already checked that weights are all there or none, we can
   // check the existence of only one to the get the condition.
   const bool use_cifg = (input_to_input_weights_ptr == nullptr);

From 7ee328496f1ea1306f212ff1b9b4242e1cbd286b Mon Sep 17 00:00:00 2001
From: "A. Unique TensorFlower" 
Date: Fri, 19 Jun 2020 19:30:45 -0700
Subject: [PATCH 375/867] Add code examples to README.md

PiperOrigin-RevId: 317420765
Change-Id: I26de720075125387c652fc30d09215055e1dddad
---
 tensorflow/python/ops/numpy_ops/README.md   | 121 +++++++++++++-------
 tensorflow/python/ops/numpy_ops/__init__.py |   3 -
 2 files changed, 78 insertions(+), 46 deletions(-)

diff --git a/tensorflow/python/ops/numpy_ops/README.md b/tensorflow/python/ops/numpy_ops/README.md
index 3dc37423d26..812f6b946c0 100644
--- a/tensorflow/python/ops/numpy_ops/README.md
+++ b/tensorflow/python/ops/numpy_ops/README.md
@@ -8,6 +8,14 @@ change the baseline version of numpy API being supported. A list of some
 systematic differences with numpy are listed later in the "Differences with
 Numpy" section.
 
+## Getting Started
+
+```python
+import tensorflow as tf
+from tf.experimental import numpy as np
+print(np.ones([2,1]) + np.ones([1, 2]))
+```
+
 ## Types
 
 The module provide an `ndarray` class which wraps an immutable `tf.Tensor`.
@@ -15,44 +23,12 @@ Additional functions are provided which accept array-like objects. Here
 array-like objects includes `ndarrays` as defined by this module, as well as
 `tf.Tensor`, in addition to types accepted by `numpy`.
 
-A subset of `numpy` dtypes are supported, along with `tf.bfloat16`.
-Additionally, support is provided for selecting the default float type
-(`np.float32` vs `np.float64`) given that some applications may prefer lower
-precision.
+A subset of `numpy` dtypes are supported. Type promotion follows numpy
+semantics.
 
-## Device Support
-
-Given that `ndarray` and functions wrap TensorFlow constructs, the code will
-have GPU and TPU support on par with TensorFlow. Also the code can be wrapped
-with `tf.function` and XLA compiled. Device placement can be controlled by using
-`with tf.device` scopes.
-
-## Graph and Eager Modes
-
-Eager mode execution should typically match numpy semantics of executing
-op-by-op. However the same code can be executed in graph mode, by putting it
-inside a `tf.function`. This can change behavior of certain operations since
-symbolic execution may not have information that is computed during runtime.
-
-Some differences are:
-
-  * Shapes can be incomplete or unknown. This means that `ndarray.shape`,
-    `ndarray.size` and `ndarray.ndim` can return `ndarray` objects instead of
-    returning integer (or tuple of integer) values.
-  * Python control flow based on `ndarray` values may not work and may have to
-    be rewritten to use `tf.cond` or `tf.while_loop`. Note that autograph
-    conversion as part of `tf.function` should still work.
-  * `__len__`, `__iter__` and `__index__` properties of `ndarray` may similarly
-    not work in graph mode.
-
-## Mutation and Variables
-
-`ndarrays` currently wrap immutable `tf.Tensor`. Also currently mutation
-operations like slice assigns are not supported. This may change in the future.
-
-There is currently no explict construct on par with `tf.Variable`. However one
-can directly construct a `tf.Variable` and use that with the numpy APIs in this
-module. See section on Interop.
+```python
+print(np.ones([1, 2], dtype=np.int16) + np.ones([2, 1], dtype=np.uint8))
+```
 
 ## Interop
 
@@ -67,28 +43,87 @@ be done using `tf.vectorized_map()`.
 In general, the expected behavior should be on par with that of code involving
 `tf.Tensor` and running stateless TensorFlow functions on them.
 
+```python
+np.sum(np.ones([1, 2]) + tf.ones([2, 1]))
+```
+
 ## Array Interface
 
 The `ndarray` class implements the `__array__` interface. This should allow
 these objects to be passed into contexts that expect a `numpy` or array-like
 object (e.g. matplotlib).
 
+```python
+import numpy as onp
+onp.sum(np.ones([1, 2]) + onp.ones([2, 1]))
+```
+
+## Device Support
+
+Given that `ndarray` and functions wrap TensorFlow constructs, the code will
+have GPU and TPU support on par with TensorFlow. Also the code can be wrapped
+with `tf.function` and XLA compiled. Device placement can be controlled by using
+`with tf.device` scopes.
+
+```python
+with tf.device("GPU:0"):
+  x = np.ones([1, 2])
+print(tf.convert_to_tensor(x).device)
+```
+
+## Graph and Eager Modes
+
+Eager mode execution should typically match numpy semantics of executing
+op-by-op. However the same code can be executed in graph mode, by putting it
+inside a `tf.function`. The function body can contain numpy code, and the inputs
+can be ndarray as well.
+
+```python
+@tf.function
+def f(x, y):
+  return np.sum(x + y)
+
+f(np.ones([1, 2]), tf.ones([2, 1]))
+```
+
+Note that this can change behavior of certain operations since symbolic
+execution may not have information that is computed during runtime.
+
+Some differences are:
+
+*   Shapes can be incomplete or unknown. This means that `ndarray.shape`,
+    `ndarray.size` and `ndarray.ndim` can return `ndarray` objects instead of
+    returning integer (or tuple of integer) values.
+*   Python control flow based on `ndarray` values may not work and may have to
+    be rewritten to use `tf.cond` or `tf.while_loop`. Note that autograph
+    conversion as part of `tf.function` should still work.
+*   `__len__`, `__iter__` and `__index__` properties of `ndarray` may similarly
+    not work in graph mode.
+
+## Mutation and Variables
+
+`ndarrays` currently wrap immutable `tf.Tensor`. Also currently mutation
+operations like slice assigns are not supported. This may change in the future.
+
+There is currently no explict construct on par with `tf.Variable`. However one
+can directly construct a `tf.Variable` and use that with the numpy APIs in this
+module. See section on Interop.
 
 ## Differences with Numpy
 
 Here is a non-exhaustive list of differences:
 
-  * Not all dtypes are currently supported. e.g. `np.float96`, `np.float128`.
+*   Not all dtypes are currently supported. e.g. `np.float96`, `np.float128`.
     `np.object`, `np.str`, `np.recarray` types are not supported.
-  * `ndarray` storage is in C order only. Fortran order, views, stride_tricks
+*   `ndarray` storage is in C order only. Fortran order, views, stride_tricks
     are not supported.
-  * Only a subset of functions and modules are supported. This set would be
+*   Only a subset of functions and modules are supported. This set would be
     expanded over time. For supported functions, some arguments or argument
     values may not be supported. This differences are listed in the function
     comments.
-  * Buffer mutation is currently not supported. `ndarrays` wrap immutable
+*   Buffer mutation is currently not supported. `ndarrays` wrap immutable
     tensors. This means that output buffer arguments (e..g `out` in ufuncs) are
     not supported
-  * full `ufunc` support is not provided.
-  * Numpy C API is not supported. Numpy's Cython and Swig integration are not
+*   full `ufunc` support is not provided.
+*   Numpy C API is not supported. Numpy's Cython and Swig integration are not
     supported.
diff --git a/tensorflow/python/ops/numpy_ops/__init__.py b/tensorflow/python/ops/numpy_ops/__init__.py
index 10ace06df9a..c520ece7843 100644
--- a/tensorflow/python/ops/numpy_ops/__init__.py
+++ b/tensorflow/python/ops/numpy_ops/__init__.py
@@ -31,9 +31,6 @@ array-like objects includes `ndarrays` as defined by this module, as well as
 `tf.Tensor`, in addition to types accepted by `numpy`.
 
 A subset of `numpy` dtypes are supported, along with `tf.bfloat16`.
-Additionally, support is provided for selecting the default float type
-(`np.float32` vs `np.float64`) given that some applications may prefer lower
-precision.
 
 Device Support
 -------------

From 09ec15539eece57b257ce9074918282d88523d56 Mon Sep 17 00:00:00 2001
From: Peng Wang 
Date: Fri, 19 Jun 2020 19:42:33 -0700
Subject: [PATCH 376/867] [TF-numpy] Adds some constants.

PiperOrigin-RevId: 317421570
Change-Id: I467e979095fbd2311a64913e6419ca8317f68734
---
 tensorflow/python/ops/numpy_ops/np_math_ops.py | 5 +++++
 1 file changed, 5 insertions(+)

diff --git a/tensorflow/python/ops/numpy_ops/np_math_ops.py b/tensorflow/python/ops/numpy_ops/np_math_ops.py
index 3cf26095dd8..df3d7cf32ab 100644
--- a/tensorflow/python/ops/numpy_ops/np_math_ops.py
+++ b/tensorflow/python/ops/numpy_ops/np_math_ops.py
@@ -40,6 +40,11 @@ from tensorflow.python.ops.numpy_ops import np_dtypes
 from tensorflow.python.ops.numpy_ops import np_utils
 
 
+pi = np.pi
+e = np.e
+inf = np.inf
+
+
 @np_utils.np_doc_only('dot')
 def dot(a, b):  # pylint: disable=missing-docstring
 

From fcba37557a30e707b1368ee270a29e41410ad546 Mon Sep 17 00:00:00 2001
From: Mark Daoust 
Date: Fri, 19 Jun 2020 20:36:23 -0700
Subject: [PATCH 377/867] In `.adapt`, don't freeze any shape values, only the
 number of dimensions.

Currently, this freezes the batch size and other shape elements. This is too strict, leads to downstream failures if you have:

* a different batch_size for `adapt()` and `fit()`.
* variable sequence length or image size.

PiperOrigin-RevId: 317425544
Change-Id: I8cfeceeb6816d2f70ed112a04f51fdd15e6658bf
---
 .../keras/engine/base_preprocessing_layer.py  | 19 +++++++---
 .../engine/base_preprocessing_layer_test.py   | 38 +++++++++++++++++--
 2 files changed, 48 insertions(+), 9 deletions(-)

diff --git a/tensorflow/python/keras/engine/base_preprocessing_layer.py b/tensorflow/python/keras/engine/base_preprocessing_layer.py
index b2ab0880422..c8ba1229ff5 100644
--- a/tensorflow/python/keras/engine/base_preprocessing_layer.py
+++ b/tensorflow/python/keras/engine/base_preprocessing_layer.py
@@ -185,14 +185,21 @@ class CombinerPreprocessingLayer(PreprocessingLayer):
       if not self.built:
         try:
           # If this is a Numpy array or tensor, we can get shape from .shape.
-          # If not, an attribute error will be thrown (and we can assume the
-          # input data is a scalar with shape None.
-          shape = data_element.shape
+          # If not, an attribute error will be thrown.
+          data_shape = data_element.shape
+          data_shape_nones = tuple([None]*len(data_element.shape))
         except AttributeError:
-          shape = None
+          # The input has an unknown number of dimensions.
+          data_shape = None
+          data_shape_nones = None
+
         # TODO (b/159261555): move this to base layer build.
-        self._batch_input_shape = shape
-        self.build(shape)
+        batch_input_shape = getattr(self, '_batch_input_shape', None)
+        if batch_input_shape is None:
+          # Set the number of dimensions.
+          self._batch_input_shape = data_shape_nones
+
+        self.build(data_shape)
 
       # Once we have built the Layer, we can process the input data. We do so
       # until we've gotten an exception indicating that we have no more data.
diff --git a/tensorflow/python/keras/engine/base_preprocessing_layer_test.py b/tensorflow/python/keras/engine/base_preprocessing_layer_test.py
index 70d088cf3d3..a3a36a9bf11 100644
--- a/tensorflow/python/keras/engine/base_preprocessing_layer_test.py
+++ b/tensorflow/python/keras/engine/base_preprocessing_layer_test.py
@@ -122,11 +122,11 @@ class AddingPreprocessingLayerV1(
   pass
 
 
-def get_layer():
+def get_layer(**kwargs):
   if context.executing_eagerly():
-    return AddingPreprocessingLayer()
+    return AddingPreprocessingLayer(**kwargs)
   else:
-    return AddingPreprocessingLayerV1()
+    return AddingPreprocessingLayerV1(**kwargs)
 
 
 @keras_parameterized.run_all_keras_modes
@@ -366,6 +366,38 @@ class PreprocessingLayerTest(keras_parameterized.TestCase):
     with self.assertRaisesRegex(RuntimeError, "Unable to restore a layer of"):
       _ = keras.models.load_model(output_path)
 
+  def test_adapt_sets_input_shape_rank(self):
+    """Check that `.adapt()` sets the `input_shape`'s rank."""
+    # Shape: (3,1,2)
+    adapt_dataset = np.array([[[1., 2.]],
+                              [[3., 4.]],
+                              [[5., 6.]]], dtype=np.float32)
+
+    layer = get_layer()
+    layer.adapt(adapt_dataset)
+
+    input_dataset = np.array([[[1., 2.], [3., 4.]],
+                              [[3., 4.], [5., 6.]]], dtype=np.float32)
+    layer(input_dataset)
+
+    model = keras.Sequential([layer])
+    self.assertTrue(model.built)
+    self.assertEqual(model.input_shape, (None, None, None))
+
+  def test_adapt_doesnt_overwrite_input_shape(self):
+    """Check that `.adapt()` doesn't change the `input_shape`."""
+    # Shape: (3, 1, 2)
+    adapt_dataset = np.array([[[1., 2.]],
+                              [[3., 4.]],
+                              [[5., 6.]]], dtype=np.float32)
+
+    layer = get_layer(input_shape=[1, 2])
+    layer.adapt(adapt_dataset)
+
+    model = keras.Sequential([layer])
+    self.assertTrue(model.built)
+    self.assertEqual(model.input_shape, (None, 1, 2))
+
 
 @keras_parameterized.run_all_keras_modes
 class ConvertToListTest(keras_parameterized.TestCase):

From 728a4a4405d17dc9dc24a0a584cbdb49cd9c001e Mon Sep 17 00:00:00 2001
From: David Majnemer 
Date: Fri, 19 Jun 2020 21:10:52 -0700
Subject: [PATCH 378/867] Unbreak the MSVC build

Don't use inline variables, our MSVC builds do not like them.

Instead, simulate them using inline functions + function-scope static variables.

PiperOrigin-RevId: 317428053
Change-Id: Icf8838c159ab9f132ad32360633046f4c2224a79
---
 .../compiler/xla/tests/manifest_checking_test.cc      |  4 ++--
 tensorflow/compiler/xla/tests/test_macros.cc          |  8 ++++----
 tensorflow/compiler/xla/tests/test_macros.h           | 11 +++++++++--
 3 files changed, 15 insertions(+), 8 deletions(-)

diff --git a/tensorflow/compiler/xla/tests/manifest_checking_test.cc b/tensorflow/compiler/xla/tests/manifest_checking_test.cc
index 8806290472d..ac6204f9df9 100644
--- a/tensorflow/compiler/xla/tests/manifest_checking_test.cc
+++ b/tensorflow/compiler/xla/tests/manifest_checking_test.cc
@@ -37,7 +37,7 @@ using ManifestT = absl::flat_hash_map>;
 ManifestT ReadManifest() {
   ManifestT manifest;
 
-  absl::string_view path = absl::NullSafeStringView(kDisabledManifestPath);
+  absl::string_view path = absl::NullSafeStringView(*DisabledManifestPath());
   if (path.empty()) {
     return manifest;
   }
@@ -115,7 +115,7 @@ void ManifestCheckingTest::SetUp() {
 
   // Expect a full match vs. one of the platform regexps to disable the test.
   const std::vector& disabled_platforms = it->second;
-  auto platform_string = kTestPlatform;
+  auto platform_string = *TestPlatform();
   for (const auto& s : disabled_platforms) {
     if (RE2::FullMatch(/*text=*/platform_string, /*re=*/s)) {
       GTEST_SKIP();
diff --git a/tensorflow/compiler/xla/tests/test_macros.cc b/tensorflow/compiler/xla/tests/test_macros.cc
index 9e85af76e89..eecbb89b877 100644
--- a/tensorflow/compiler/xla/tests/test_macros.cc
+++ b/tensorflow/compiler/xla/tests/test_macros.cc
@@ -20,10 +20,10 @@ limitations under the License.
 namespace xla {
 
 static bool InitModule() {
-  kDisabledManifestPath = XLA_DISABLED_MANIFEST;
-  VLOG(1) << "kDisabledManifestPath: " << kDisabledManifestPath;
-  kTestPlatform = XLA_PLATFORM;
-  VLOG(1) << "kTestPlatform: " << kTestPlatform;
+  *DisabledManifestPath() = XLA_DISABLED_MANIFEST;
+  VLOG(1) << "DisabledManifestPath: " << *DisabledManifestPath();
+  *TestPlatform() = XLA_PLATFORM;
+  VLOG(1) << "TestPlatform: " << *TestPlatform();
   return false;
 }
 
diff --git a/tensorflow/compiler/xla/tests/test_macros.h b/tensorflow/compiler/xla/tests/test_macros.h
index f62bccbe850..16cc9ff6feb 100644
--- a/tensorflow/compiler/xla/tests/test_macros.h
+++ b/tensorflow/compiler/xla/tests/test_macros.h
@@ -73,8 +73,15 @@ limitations under the License.
 
 namespace xla {
 
-inline const char *kDisabledManifestPath = nullptr;
-inline const char *kTestPlatform = nullptr;
+inline const char** DisabledManifestPath() {
+  static const char* disabled_manifest_path = nullptr;
+  return &disabled_manifest_path;
+}
+
+inline const char** TestPlatform() {
+  static const char* test_platform = nullptr;
+  return &test_platform;
+}
 
 }  // namespace xla
 

From 8efcda9caa52c3935e9db3cb9f4a787d0b97dae7 Mon Sep 17 00:00:00 2001
From: Peng Wang 
Date: Fri, 19 Jun 2020 21:56:50 -0700
Subject: [PATCH 379/867] [TF-numpy] Added `einsum` to numpy_ops.

PiperOrigin-RevId: 317431034
Change-Id: I209bc52b5c03369526eb45486c1b6c09352f9f06
---
 .../python/ops/numpy_ops/np_math_ops.py       | 28 +++++++++++++++++++
 1 file changed, 28 insertions(+)

diff --git a/tensorflow/python/ops/numpy_ops/np_math_ops.py b/tensorflow/python/ops/numpy_ops/np_math_ops.py
index df3d7cf32ab..f9d3b34e90d 100644
--- a/tensorflow/python/ops/numpy_ops/np_math_ops.py
+++ b/tensorflow/python/ops/numpy_ops/np_math_ops.py
@@ -34,6 +34,7 @@ from tensorflow.python.ops import clip_ops
 from tensorflow.python.ops import control_flow_ops
 from tensorflow.python.ops import math_ops
 from tensorflow.python.ops import sort_ops
+from tensorflow.python.ops import special_math_ops
 from tensorflow.python.ops.numpy_ops import np_array_ops
 from tensorflow.python.ops.numpy_ops import np_arrays
 from tensorflow.python.ops.numpy_ops import np_dtypes
@@ -1336,3 +1337,30 @@ def meshgrid(*xi, **kwargs):
   outputs = [np_utils.tensor_to_ndarray(output) for output in outputs]
 
   return outputs
+
+
+@np_utils.np_doc('einsum')
+def einsum(subscripts, *operands, casting='safe', optimize=False):  # pylint: disable=missing-docstring
+  if casting == 'safe':
+    operands = np_array_ops._promote_dtype(*operands)  # pylint: disable=protected-access
+  elif casting == 'no':
+    operands = [np_array_ops.asarray(x) for x in operands]
+  else:
+    raise ValueError('casting policy not supported: %s' % casting)
+  if not optimize:
+    # TF doesn't have a "no optimization" option.
+    # TODO(wangpeng): Print a warning that np and tf use different
+    #   optimizations.
+    tf_optimize = 'greedy'
+  elif optimize == True:  # pylint: disable=singleton-comparison,g-explicit-bool-comparison
+    tf_optimize = 'greedy'
+  elif optimize == 'greedy':
+    tf_optimize = 'greedy'
+  elif optimize == 'optimal':
+    tf_optimize = 'optimal'
+  else:
+    raise ValueError('`optimize` method not supported: %s' % optimize)
+  operands = [x.data for x in operands]
+  res = special_math_ops.einsum(subscripts, *operands, optimize=tf_optimize)
+  res = np_utils.tensor_to_ndarray(res)
+  return res

From 7f2bfd570962ab3401cd5744ec638fcf523d26c5 Mon Sep 17 00:00:00 2001
From: Sanjoy Das 
Date: Fri, 19 Jun 2020 22:10:56 -0700
Subject: [PATCH 380/867] Have the HloEvaluator print out intermediate values
 on VLOG(100)

This is useful for tracing through very small HLO programs.

PiperOrigin-RevId: 317432566
Change-Id: I1b70ed0f81ed57915ecccaf8cb85fefc857e2151
---
 tensorflow/compiler/xla/service/hlo_evaluator.cc | 7 +++++++
 1 file changed, 7 insertions(+)

diff --git a/tensorflow/compiler/xla/service/hlo_evaluator.cc b/tensorflow/compiler/xla/service/hlo_evaluator.cc
index 3dd6d82784f..ae8f49df4b4 100644
--- a/tensorflow/compiler/xla/service/hlo_evaluator.cc
+++ b/tensorflow/compiler/xla/service/hlo_evaluator.cc
@@ -274,6 +274,13 @@ StatusOr HloEvaluator::Evaluate(
   engine_.seed(seed_);
 
   TF_RETURN_IF_ERROR(computation.Accept(this));
+
+  if (VLOG_IS_ON(100)) {
+    for (const HloInstruction* instr : computation.instructions()) {
+      VLOG(100) << instr->name() << " = " << GetEvaluatedLiteralFor(instr);
+    }
+  }
+
   return GetEvaluatedLiteralFor(computation.root_instruction()).Clone();
 }
 

From 27cb9aa8342b120f739310a6a6222c37620984a4 Mon Sep 17 00:00:00 2001
From: Frank Chen 
Date: Fri, 19 Jun 2020 22:49:37 -0700
Subject: [PATCH 381/867] Add SnappyInputStream implementation.

This is to support snappy compression/decompression on new version of snapshot dataset, as the dataset uses input streams instead.

PiperOrigin-RevId: 317435680
Change-Id: Ie57b43e73b6b7911398d883c3c5a0de72973288e
---
 tensorflow/core/BUILD                         |   1 +
 tensorflow/core/lib/io/BUILD                  |  17 +-
 .../core/lib/io/snappy/snappy_inputstream.cc  | 153 +++++++++++++++
 .../core/lib/io/snappy/snappy_inputstream.h   |  89 +++++++++
 ...{snappy_buffers_test.cc => snappy_test.cc} | 182 ++++++++++++++++--
 5 files changed, 427 insertions(+), 15 deletions(-)
 create mode 100644 tensorflow/core/lib/io/snappy/snappy_inputstream.cc
 create mode 100644 tensorflow/core/lib/io/snappy/snappy_inputstream.h
 rename tensorflow/core/lib/io/snappy/{snappy_buffers_test.cc => snappy_test.cc} (59%)

diff --git a/tensorflow/core/BUILD b/tensorflow/core/BUILD
index 7f1c1bd549b..0f709750897 100644
--- a/tensorflow/core/BUILD
+++ b/tensorflow/core/BUILD
@@ -1865,6 +1865,7 @@ cc_library(
         "//tensorflow/core/lib/io:record_reader",
         "//tensorflow/core/lib/io:record_writer",
         "//tensorflow/core/lib/io:snappy_inputbuffer",
+        "//tensorflow/core/lib/io:snappy_inputstream",
         "//tensorflow/core/lib/io:snappy_outputbuffer",
         "//tensorflow/core/lib/io:table",
         "//tensorflow/core/lib/io:table_options",
diff --git a/tensorflow/core/lib/io/BUILD b/tensorflow/core/lib/io/BUILD
index d03a895b429..5e1704a50c1 100644
--- a/tensorflow/core/lib/io/BUILD
+++ b/tensorflow/core/lib/io/BUILD
@@ -208,6 +208,19 @@ cc_library(
     alwayslink = True,
 )
 
+cc_library(
+    name = "snappy_inputstream",
+    srcs = ["snappy/snappy_inputstream.cc"],
+    hdrs = ["snappy/snappy_inputstream.h"],
+    deps = [
+        ":inputstream_interface",
+        "//tensorflow/core/platform:errors",
+        "//tensorflow/core/platform:platform_port",
+        "@com_google_absl//absl/memory",
+    ],
+    alwayslink = True,
+)
+
 cc_library(
     name = "cache",
     srcs = [
@@ -354,6 +367,7 @@ filegroup(
         "record_reader.h",
         "record_writer.h",
         "snappy/snappy_inputbuffer.h",
+        "snappy/snappy_inputstream.h",
         "snappy/snappy_outputbuffer.h",
         "table.h",
         "table_builder.h",
@@ -377,7 +391,7 @@ filegroup(
         "random_inputstream_test.cc",
         "record_reader_writer_test.cc",
         "recordio_test.cc",
-        "snappy/snappy_buffers_test.cc",
+        "snappy/snappy_test.cc",
         "table_test.cc",
         "zlib_buffers_test.cc",
     ],
@@ -409,6 +423,7 @@ filegroup(
         "inputbuffer.h",
         "iterator.h",
         "snappy/snappy_inputbuffer.h",
+        "snappy/snappy_inputstream.h",
         "snappy/snappy_outputbuffer.h",
         "zlib_compression_options.h",
         "zlib_inputstream.h",
diff --git a/tensorflow/core/lib/io/snappy/snappy_inputstream.cc b/tensorflow/core/lib/io/snappy/snappy_inputstream.cc
new file mode 100644
index 00000000000..7e77971f4f1
--- /dev/null
+++ b/tensorflow/core/lib/io/snappy/snappy_inputstream.cc
@@ -0,0 +1,153 @@
+/* Copyright 2020 The TensorFlow Authors. All Rights Reserved.
+
+Licensed under the Apache License, Version 2.0 (the "License");
+you may not use this file except in compliance with the License.
+You may obtain a copy of the License at
+
+    http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software
+distributed under the License is distributed on an "AS IS" BASIS,
+WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+See the License for the specific language governing permissions and
+limitations under the License.
+==============================================================================*/
+
+#include "tensorflow/core/lib/io/snappy/snappy_inputstream.h"
+
+#include "absl/memory/memory.h"
+#include "tensorflow/core/platform/errors.h"
+#include "tensorflow/core/platform/snappy.h"
+
+namespace tensorflow {
+namespace io {
+
+SnappyInputStream::SnappyInputStream(InputStreamInterface* input_stream,
+                                     size_t output_buffer_bytes,
+                                     bool owns_input_stream)
+    : input_stream_(input_stream),
+      output_buffer_bytes_(output_buffer_bytes),
+      owns_input_stream_(owns_input_stream),
+      bytes_read_(0),
+      output_buffer_(new char[output_buffer_bytes]),
+      next_out_(nullptr),
+      avail_out_(0) {}
+
+SnappyInputStream::SnappyInputStream(InputStreamInterface* input_stream,
+                                     size_t output_buffer_bytes)
+    : SnappyInputStream(input_stream, output_buffer_bytes, false) {}
+
+SnappyInputStream::~SnappyInputStream() {
+  if (owns_input_stream_) {
+    delete input_stream_;
+  }
+}
+
+Status SnappyInputStream::ReadNBytes(int64 bytes_to_read, tstring* result) {
+  result->clear();
+  result->resize_uninitialized(bytes_to_read);
+
+  char* result_ptr = result->mdata();
+
+  // Read as many bytes as possible from the cache.
+  size_t bytes_read = ReadBytesFromCache(bytes_to_read, result_ptr);
+  bytes_to_read -= bytes_read;
+  result_ptr += bytes_read;
+
+  while (bytes_to_read > 0) {
+    DCHECK_EQ(avail_out_, 0);
+
+    // Fill the cache with more data.
+    TF_RETURN_IF_ERROR(Inflate());
+
+    size_t bytes_read = ReadBytesFromCache(bytes_to_read, result_ptr);
+    bytes_to_read -= bytes_read;
+    result_ptr += bytes_read;
+  }
+
+  return Status::OK();
+}
+
+#if defined(PLATFORM_GOOGLE)
+Status SnappyInputStream::ReadNBytes(int64 bytes_to_read, absl::Cord* result) {
+  // TODO(frankchn): Optimize this instead of bouncing through the buffer.
+  tstring buf;
+  TF_RETURN_IF_ERROR(ReadNBytes(bytes_to_read, &buf));
+  result->Clear();
+  result->Append(buf.data());
+  return Status::OK();
+}
+#endif
+
+Status SnappyInputStream::Inflate() {
+  tstring compressed_block_length_ts;
+  uint32 compressed_block_length;
+
+  TF_RETURN_IF_ERROR(
+      input_stream_->ReadNBytes(sizeof(uint32), &compressed_block_length_ts));
+  for (int i = 0; i < sizeof(uint32); ++i) {
+    compressed_block_length =
+        (compressed_block_length << 8) |
+        static_cast(compressed_block_length_ts.data()[i]);
+  }
+
+  tstring compressed_block;
+  compressed_block.resize_uninitialized(compressed_block_length);
+
+  Status s =
+      input_stream_->ReadNBytes(compressed_block_length, &compressed_block);
+  if (errors::IsOutOfRange(s)) {
+    return errors::DataLoss("Failed to read ", compressed_block_length,
+                            " bytes from file. Possible data corruption.");
+  }
+  TF_RETURN_IF_ERROR(s);
+
+  size_t uncompressed_length;
+  if (!port::Snappy_GetUncompressedLength(compressed_block.data(),
+                                          compressed_block_length,
+                                          &uncompressed_length)) {
+    return errors::DataLoss("Parsing error in Snappy_GetUncompressedLength");
+  }
+
+  DCHECK_EQ(avail_out_, 0);
+  if (output_buffer_bytes_ < uncompressed_length) {
+    return errors::ResourceExhausted(
+        "Output buffer(size: ", output_buffer_bytes_,
+        " bytes"
+        ") too small. Should be larger than ",
+        uncompressed_length, " bytes.");
+  }
+
+  next_out_ = output_buffer_.get();
+  if (!port::Snappy_Uncompress(compressed_block.data(), compressed_block_length,
+                               output_buffer_.get())) {
+    return errors::DataLoss("Snappy_Uncompress failed.");
+  }
+  avail_out_ += uncompressed_length;
+
+  return Status::OK();
+}
+
+size_t SnappyInputStream::ReadBytesFromCache(size_t bytes_to_read,
+                                             char* result) {
+  size_t can_read_bytes = std::min(bytes_to_read, avail_out_);
+  if (can_read_bytes) {
+    memcpy(result, next_out_, can_read_bytes);
+    next_out_ += can_read_bytes;
+    avail_out_ -= can_read_bytes;
+  }
+  bytes_read_ += can_read_bytes;
+  return can_read_bytes;
+}
+
+int64 SnappyInputStream::Tell() const { return bytes_read_; }
+
+Status SnappyInputStream::Reset() {
+  TF_RETURN_IF_ERROR(input_stream_->Reset());
+  avail_out_ = 0;
+  bytes_read_ = 0;
+  return Status::OK();
+}
+
+}  // namespace io
+}  // namespace tensorflow
diff --git a/tensorflow/core/lib/io/snappy/snappy_inputstream.h b/tensorflow/core/lib/io/snappy/snappy_inputstream.h
new file mode 100644
index 00000000000..bbe8eaf0dda
--- /dev/null
+++ b/tensorflow/core/lib/io/snappy/snappy_inputstream.h
@@ -0,0 +1,89 @@
+/* Copyright 2020 The TensorFlow Authors. All Rights Reserved.
+
+Licensed under the Apache License, Version 2.0 (the "License");
+you may not use this file except in compliance with the License.
+You may obtain a copy of the License at
+
+    http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software
+distributed under the License is distributed on an "AS IS" BASIS,
+WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+See the License for the specific language governing permissions and
+limitations under the License.
+==============================================================================*/
+
+#ifndef TENSORFLOW_CORE_LIB_IO_SNAPPY_SNAPPY_INPUTSTREAM_H_
+#define TENSORFLOW_CORE_LIB_IO_SNAPPY_SNAPPY_INPUTSTREAM_H_
+
+#include "tensorflow/core/lib/io/inputstream_interface.h"
+
+namespace tensorflow {
+namespace io {
+
+class SnappyInputStream : public InputStreamInterface {
+ public:
+  // Creates a SnappyInputStream for `input_stream`.
+  //
+  // Takes ownership  of `input_stream` iff `owns_input_stream` is true.
+  SnappyInputStream(InputStreamInterface* input_stream,
+                    size_t output_buffer_bytes, bool owns_input_stream);
+
+  // Equivalent to the previous constructor with owns_input_stream = false.
+  explicit SnappyInputStream(InputStreamInterface* input_stream,
+                             size_t output_buffer_bytes);
+
+  ~SnappyInputStream() override;
+
+  // Reads bytes_to_read bytes into *result, overwriting *result.
+  //
+  // Return Status codes:
+  // OK:           If successful.
+  // OUT_OF_RANGE: If there are not enough bytes to read before
+  //               the end of the stream.
+  // ABORTED:      If inflate() fails, we return the error code with the
+  //               error message in `z_stream_->msg`.
+  // others:       If reading from stream failed.
+  Status ReadNBytes(int64 bytes_to_read, tstring* result) override;
+
+#if defined(PLATFORM_GOOGLE)
+  Status ReadNBytes(int64 bytes_to_read, absl::Cord* result) override;
+#endif
+
+  int64 Tell() const override;
+
+  Status Reset() override;
+
+ private:
+  // Decompress the next chunk of data and place the data into the cache.
+  Status Inflate();
+
+  // Attempt to read `bytes_to_read` from the decompressed data cache. Returns
+  // the actual number of bytes read.
+  size_t ReadBytesFromCache(size_t bytes_to_read, char* result);
+
+  InputStreamInterface* input_stream_;
+  const size_t output_buffer_bytes_;
+  const bool owns_input_stream_;
+
+  // Specifies the number of decompressed bytes currently read.
+  int64 bytes_read_;
+
+  // output_buffer_ contains decompressed data not yet read by the client.
+  std::unique_ptr output_buffer_;
+
+  // next_out_ points to the position in the `output_buffer_` that contains the
+  // next unread byte.
+  char* next_out_;
+
+  // avail_out_ specifies the number of bytes left in the output_buffers_ that
+  // is not yet read.
+  size_t avail_out_;
+
+  TF_DISALLOW_COPY_AND_ASSIGN(SnappyInputStream);
+};
+
+}  // namespace io
+}  // namespace tensorflow
+
+#endif  // TENSORFLOW_CORE_LIB_IO_SNAPPY_SNAPPY_INPUTSTREAM_H_
diff --git a/tensorflow/core/lib/io/snappy/snappy_buffers_test.cc b/tensorflow/core/lib/io/snappy/snappy_test.cc
similarity index 59%
rename from tensorflow/core/lib/io/snappy/snappy_buffers_test.cc
rename to tensorflow/core/lib/io/snappy/snappy_test.cc
index 521b49b73f2..b7d5eae6cc5 100644
--- a/tensorflow/core/lib/io/snappy/snappy_buffers_test.cc
+++ b/tensorflow/core/lib/io/snappy/snappy_test.cc
@@ -15,7 +15,9 @@ limitations under the License.
 
 #include "tensorflow/core/lib/core/status_test_util.h"
 #include "tensorflow/core/lib/io/inputbuffer.h"
+#include "tensorflow/core/lib/io/random_inputstream.h"
 #include "tensorflow/core/lib/io/snappy/snappy_inputbuffer.h"
+#include "tensorflow/core/lib/io/snappy/snappy_inputstream.h"
 #include "tensorflow/core/lib/io/snappy/snappy_outputbuffer.h"
 
 namespace tensorflow {
@@ -50,18 +52,17 @@ static string GenTestString(int copies = 1) {
   return result;
 }
 
-Status TestMultipleWrites(size_t compress_input_buf_size,
-                          size_t compress_output_buf_size,
-                          size_t uncompress_input_buf_size,
-                          size_t uncompress_output_buf_size, int num_writes = 1,
-                          bool with_flush = false, int num_copies = 1,
-                          bool corrupt_compressed_file = false) {
+Status TestMultipleWritesWriteFile(size_t compress_input_buf_size,
+                                   size_t compress_output_buf_size,
+                                   int num_writes, bool with_flush,
+                                   int num_copies, bool corrupt_compressed_file,
+                                   string& fname, string& data,
+                                   string& expected_result) {
   Env* env = Env::Default();
 
-  string fname = testing::TmpDir() + "/snappy_buffers_test";
-  string data = GenTestString(num_copies);
+  fname = testing::TmpDir() + "/snappy_buffers_test";
+  data = GenTestString(num_copies);
   std::unique_ptr file_writer;
-  string expected_result;
 
   TF_RETURN_IF_ERROR(env->NewWritableFile(fname, &file_writer));
   io::SnappyOutputBuffer out(file_writer.get(), compress_input_buf_size,
@@ -112,6 +113,25 @@ Status TestMultipleWrites(size_t compress_input_buf_size,
     fname = corrupt_fname;
   }
 
+  return Status::OK();
+}
+
+Status TestMultipleWrites(size_t compress_input_buf_size,
+                          size_t compress_output_buf_size,
+                          size_t uncompress_input_buf_size,
+                          size_t uncompress_output_buf_size, int num_writes = 1,
+                          bool with_flush = false, int num_copies = 1,
+                          bool corrupt_compressed_file = false) {
+  Env* env = Env::Default();
+
+  string expected_result;
+  string fname;
+  string data;
+
+  TF_RETURN_IF_ERROR(TestMultipleWritesWriteFile(
+      compress_input_buf_size, compress_output_buf_size, num_writes, with_flush,
+      num_copies, corrupt_compressed_file, fname, data, expected_result));
+
   std::unique_ptr file_reader;
   TF_RETURN_IF_ERROR(env->NewRandomAccessFile(fname, &file_reader));
   io::SnappyInputBuffer in(file_reader.get(), uncompress_input_buf_size,
@@ -131,15 +151,56 @@ Status TestMultipleWrites(size_t compress_input_buf_size,
     }
     TF_RETURN_IF_ERROR(in.Reset());
   }
+
   return Status::OK();
 }
 
-void TestTell(size_t compress_input_buf_size, size_t compress_output_buf_size,
-              size_t uncompress_input_buf_size,
-              size_t uncompress_output_buf_size, int num_copies = 1) {
+Status TestMultipleWritesInputStream(
+    size_t compress_input_buf_size, size_t compress_output_buf_size,
+    size_t uncompress_input_buf_size, size_t uncompress_output_buf_size,
+    int num_writes = 1, bool with_flush = false, int num_copies = 1,
+    bool corrupt_compressed_file = false) {
   Env* env = Env::Default();
-  string fname = testing::TmpDir() + "/snappy_buffers_test";
-  string data = GenTestString(num_copies);
+
+  string expected_result;
+  string fname;
+  string data;
+
+  TF_RETURN_IF_ERROR(TestMultipleWritesWriteFile(
+      compress_input_buf_size, compress_output_buf_size, num_writes, with_flush,
+      num_copies, corrupt_compressed_file, fname, data, expected_result));
+
+  std::unique_ptr file_reader;
+  TF_RETURN_IF_ERROR(env->NewRandomAccessFile(fname, &file_reader));
+  io::RandomAccessInputStream random_input_stream(file_reader.get(), false);
+  io::SnappyInputStream snappy_input_stream(&random_input_stream,
+                                            uncompress_output_buf_size);
+
+  for (int attempt = 0; attempt < 2; ++attempt) {
+    string actual_result;
+    for (int i = 0; i < num_writes; ++i) {
+      tstring decompressed_output;
+      TF_RETURN_IF_ERROR(
+          snappy_input_stream.ReadNBytes(data.size(), &decompressed_output));
+      strings::StrAppend(&actual_result, decompressed_output);
+    }
+
+    if (actual_result.compare(expected_result)) {
+      return errors::DataLoss("Actual and expected results don't match.");
+    }
+    TF_RETURN_IF_ERROR(snappy_input_stream.Reset());
+  }
+  return Status::OK();
+}
+
+void TestTellWriteFile(size_t compress_input_buf_size,
+                       size_t compress_output_buf_size,
+                       size_t uncompress_input_buf_size,
+                       size_t uncompress_output_buf_size, int num_copies,
+                       string& fname, string& data) {
+  Env* env = Env::Default();
+  fname = testing::TmpDir() + "/snappy_buffers_test";
+  data = GenTestString(num_copies);
 
   // Write the compressed file.
   std::unique_ptr file_writer;
@@ -150,6 +211,18 @@ void TestTell(size_t compress_input_buf_size, size_t compress_output_buf_size,
   TF_CHECK_OK(out.Flush());
   TF_CHECK_OK(file_writer->Flush());
   TF_CHECK_OK(file_writer->Close());
+}
+
+void TestTell(size_t compress_input_buf_size, size_t compress_output_buf_size,
+              size_t uncompress_input_buf_size,
+              size_t uncompress_output_buf_size, int num_copies = 1) {
+  Env* env = Env::Default();
+  string data;
+  string fname;
+
+  TestTellWriteFile(compress_input_buf_size, compress_output_buf_size,
+                    uncompress_input_buf_size, uncompress_output_buf_size,
+                    num_copies, fname, data);
 
   tstring first_half(string(data, 0, data.size() / 2));
   tstring bytes_read;
@@ -175,6 +248,43 @@ void TestTell(size_t compress_input_buf_size, size_t compress_output_buf_size,
   EXPECT_EQ(bytes_read, data);
 }
 
+void TestTellInputStream(size_t compress_input_buf_size,
+                         size_t compress_output_buf_size,
+                         size_t uncompress_input_buf_size,
+                         size_t uncompress_output_buf_size,
+                         int num_copies = 1) {
+  Env* env = Env::Default();
+  string data;
+  string fname;
+
+  TestTellWriteFile(compress_input_buf_size, compress_output_buf_size,
+                    uncompress_input_buf_size, uncompress_output_buf_size,
+                    num_copies, fname, data);
+
+  tstring first_half(string(data, 0, data.size() / 2));
+  tstring bytes_read;
+  std::unique_ptr file_reader;
+  TF_CHECK_OK(env->NewRandomAccessFile(fname, &file_reader));
+  io::RandomAccessInputStream random_input_stream(file_reader.get(), false);
+  io::SnappyInputStream in(&random_input_stream, uncompress_output_buf_size);
+
+  // Read the first half of the uncompressed file and expect that Tell()
+  // returns half the uncompressed length of the file.
+  TF_CHECK_OK(in.ReadNBytes(first_half.size(), &bytes_read));
+  EXPECT_EQ(in.Tell(), first_half.size());
+  EXPECT_EQ(bytes_read, first_half);
+
+  // Read the remaining half of the uncompressed file and expect that
+  // Tell() points past the end of file.
+  tstring second_half;
+  TF_CHECK_OK(in.ReadNBytes(data.size() - first_half.size(), &second_half));
+  EXPECT_EQ(in.Tell(), data.size());
+  bytes_read.append(second_half);
+
+  // Expect that the file is correctly read.
+  EXPECT_EQ(bytes_read, data);
+}
+
 static bool SnappyCompressionSupported() {
   string out;
   StringPiece in = "aaaaaaaaaaaaaaaaaaaaaaaaaaaaaaa";
@@ -187,6 +297,7 @@ TEST(SnappyBuffers, MultipleWritesWithoutFlush) {
     return;
   }
   TF_CHECK_OK(TestMultipleWrites(10000, 10000, 10000, 10000, 2));
+  TF_CHECK_OK(TestMultipleWritesInputStream(10000, 10000, 10000, 10000, 2));
 }
 
 TEST(SnappyBuffers, MultipleWriteCallsWithFlush) {
@@ -195,6 +306,8 @@ TEST(SnappyBuffers, MultipleWriteCallsWithFlush) {
     return;
   }
   TF_CHECK_OK(TestMultipleWrites(10000, 10000, 10000, 10000, 2, true));
+  TF_CHECK_OK(
+      TestMultipleWritesInputStream(10000, 10000, 10000, 10000, 2, true));
 }
 
 TEST(SnappyBuffers, SmallUncompressInputBuffer) {
@@ -208,6 +321,17 @@ TEST(SnappyBuffers, SmallUncompressInputBuffer) {
                                      COMPRESSED_RECORD_SIZE, " bytes."));
 }
 
+TEST(SnappyBuffers, SmallUncompressInputStream) {
+  if (!SnappyCompressionSupported()) {
+    fprintf(stderr, "skipping compression tests\n");
+    return;
+  }
+  CHECK_EQ(TestMultipleWritesInputStream(10000, 10000, 10000, 10, 2, true),
+           errors::ResourceExhausted(
+               "Output buffer(size: 10 bytes) too small. ",
+               "Should be larger than ", GetRecord().size(), " bytes."));
+}
+
 TEST(SnappyBuffers, CorruptBlock) {
   if (!SnappyCompressionSupported()) {
     fprintf(stderr, "skipping compression tests\n");
@@ -218,6 +342,17 @@ TEST(SnappyBuffers, CorruptBlock) {
                             " bytes from file. ", "Possible data corruption."));
 }
 
+TEST(SnappyBuffers, CorruptBlockInputStream) {
+  if (!SnappyCompressionSupported()) {
+    fprintf(stderr, "skipping compression tests\n");
+    return;
+  }
+  CHECK_EQ(
+      TestMultipleWritesInputStream(10000, 10000, 700, 10000, 2, true, 1, true),
+      errors::DataLoss("Failed to read ", COMPRESSED_RECORD_SIZE,
+                       " bytes from file. ", "Possible data corruption."));
+}
+
 TEST(SnappyBuffers, CorruptBlockLargeInputBuffer) {
   if (!SnappyCompressionSupported()) {
     fprintf(stderr, "skipping compression tests\n");
@@ -227,6 +362,17 @@ TEST(SnappyBuffers, CorruptBlockLargeInputBuffer) {
            errors::OutOfRange("EOF reached"));
 }
 
+TEST(SnappyBuffers, CorruptBlockLargeInputStream) {
+  if (!SnappyCompressionSupported()) {
+    fprintf(stderr, "skipping compression tests\n");
+    return;
+  }
+  CHECK_EQ(TestMultipleWritesInputStream(10000, 10000, 2000, 10000, 2, true, 1,
+                                         true),
+           errors::DataLoss("Failed to read ", COMPRESSED_RECORD_SIZE,
+                            " bytes from file. Possible data corruption."));
+}
+
 TEST(SnappyBuffers, Tell) {
   if (!SnappyCompressionSupported()) {
     fprintf(stderr, "skipping compression tests\n");
@@ -235,4 +381,12 @@ TEST(SnappyBuffers, Tell) {
   TestTell(10000, 10000, 2000, 10000, 2);
 }
 
+TEST(SnappyBuffers, TellInputStream) {
+  if (!SnappyCompressionSupported()) {
+    fprintf(stderr, "skipping compression tests\n");
+    return;
+  }
+  TestTellInputStream(10000, 10000, 2000, 10000, 2);
+}
+
 }  // namespace tensorflow

From ae78d9c3150c448e14c2a624989884722b65a18e Mon Sep 17 00:00:00 2001
From: Ruoxin Sang 
Date: Sat, 20 Jun 2020 00:06:21 -0700
Subject: [PATCH 382/867] Add a `TPUClusterResolver.connect` API to simplify
 TPU initialization.

RELNOTES=Add a `tf.distribute.cluster_resolver.TPUClusterResolver.connect` API to simplify TPU initialization.
PiperOrigin-RevId: 317439811
Change-Id: I2f1a944f3c440356b21da27a72855c969f1c3b3b
---
 .../tpu/tpu_cluster_resolver.py               | 58 ++++++++++++++++---
 tensorflow/python/distribute/tpu_strategy.py  |  5 +-
 tensorflow/python/tpu/tpu_strategy_util.py    | 16 +++--
 ...ter_resolver.-t-p-u-cluster-resolver.pbtxt |  4 ++
 ...ter_resolver.-t-p-u-cluster-resolver.pbtxt |  4 ++
 tensorflow/tools/docs/BUILD                   |  2 +-
 6 files changed, 74 insertions(+), 15 deletions(-)

diff --git a/tensorflow/python/distribute/cluster_resolver/tpu/tpu_cluster_resolver.py b/tensorflow/python/distribute/cluster_resolver/tpu/tpu_cluster_resolver.py
index 943b736fde4..e42420ec644 100644
--- a/tensorflow/python/distribute/cluster_resolver/tpu/tpu_cluster_resolver.py
+++ b/tensorflow/python/distribute/cluster_resolver/tpu/tpu_cluster_resolver.py
@@ -52,16 +52,59 @@ class TPUClusterResolver(cluster_resolver.ClusterResolver):
   """Cluster Resolver for Google Cloud TPUs.
 
   This is an implementation of cluster resolvers for the Google Cloud TPU
-  service. As Cloud TPUs are in alpha, you will need to specify a API definition
-  file for this to consume, in addition to a list of Cloud TPUs in your Google
-  Cloud Platform project.
+  service.
 
   TPUClusterResolver supports the following distinct environments:
   Google Compute Engine
   Google Kubernetes Engine
   Google internal
+
+  It can be passed into `tf.distribute.TPUStrategy` to support TF2 training on
+  Cloud TPUs.
   """
 
+  @staticmethod
+  def connect(tpu=None,
+              zone=None,
+              project=None):
+    """Initializes TPU and returns a TPUClusterResolver.
+
+    This API will connect to remote TPU cluster and initialize the TPU
+    hardwares. Example usage:
+
+    >>> resolver = tf.distribute.cluster_resolver.TPUClusterResolver.connect(
+    ...     tpu='')
+
+    It can be viewed as convenient wrapper of the following code:
+
+    >>> resolver = tf.distribute.cluster_resolver.TPUClusterResolver(tpu='')
+    >>> tf.config.experimental_connect_to_cluster(resolver)
+    >>> tf.tpu.experimental.initialize_tpu_system(resolver)
+
+    Args:
+      tpu: A string corresponding to the TPU to use. It can be the TPU name or
+        TPU worker gRPC address. If not set, it will try automatically resolve
+        the TPU address on Cloud TPUs.
+      zone: Zone where the TPUs are located. If omitted or empty, we will assume
+        that the zone of the TPU is the same as the zone of the GCE VM, which we
+        will try to discover from the GCE metadata service.
+      project: Name of the GCP project containing Cloud TPUs. If omitted or
+        empty, we will try to discover the project name of the GCE VM from the
+        GCE metadata service.
+
+    Returns:
+      An instance of TPUClusterResolver object.
+
+    Raises:
+      NotFoundError: If no TPU devices found in eager mode.
+    """
+    resolver = TPUClusterResolver(tpu, zone, project)
+    from tensorflow.python.eager import remote  # pylint: disable=g-import-not-at-top
+    remote.connect_to_cluster(resolver)
+    from tensorflow.python.tpu import tpu_strategy_util  # pylint: disable=g-import-not-at-top
+    tpu_strategy_util.initialize_tpu_system(resolver)
+    return resolver
+
   @staticmethod
   def _get_device_dict_and_cores(devices):
     """Returns a dict of hosts to cores and total cores given devices names.
@@ -110,12 +153,9 @@ class TPUClusterResolver(cluster_resolver.ClusterResolver):
     for the IP addresses and ports of each Cloud TPU listed.
 
     Args:
-      tpu: A string corresponding to the TPU to use. If the string is an empty
-        string, the string 'local', or a string that begins with 'grpc://', then
-          it is assumed to not correspond with a Cloud TPU and will instead be
-          passed as the session master and no ClusterSpec propagation will be
-          done. In the future, this may also support a list of strings when
-          multiple Cloud TPUs are used.
+      tpu: A string corresponding to the TPU to use. It can be the TPU name or
+        TPU worker gRPC address. If not set, it will try automatically resolve
+        the TPU address on Cloud TPUs.
       zone: Zone where the TPUs are located. If omitted or empty, we will assume
         that the zone of the TPU is the same as the zone of the GCE VM, which we
         will try to discover from the GCE metadata service.
diff --git a/tensorflow/python/distribute/tpu_strategy.py b/tensorflow/python/distribute/tpu_strategy.py
index 7e8f5b97e7e..e2e75260593 100644
--- a/tensorflow/python/distribute/tpu_strategy.py
+++ b/tensorflow/python/distribute/tpu_strategy.py
@@ -338,7 +338,10 @@ class TPUExtended(distribute_lib.StrategyExtendedV1):
     if context.executing_eagerly():
       # In async remote eager, we want to sync the exectors before exiting the
       # program.
-      atexit.register(context.async_wait)
+      def async_wait():
+        if context.context()._context_handle is not None:  # pylint: disable=protected-access
+          context.async_wait()
+      atexit.register(async_wait)
 
   # TODO(bfontain): Remove once a proper dataset API exists for prefetching
   # a dataset to multiple devices exists.
diff --git a/tensorflow/python/tpu/tpu_strategy_util.py b/tensorflow/python/tpu/tpu_strategy_util.py
index 543c91167cd..c315d7c5e1b 100644
--- a/tensorflow/python/tpu/tpu_strategy_util.py
+++ b/tensorflow/python/tpu/tpu_strategy_util.py
@@ -24,6 +24,7 @@ from tensorflow.python.distribute.cluster_resolver.tpu_cluster_resolver import T
 from tensorflow.python.eager import context
 from tensorflow.python.eager import function
 from tensorflow.python.framework import device
+from tensorflow.python.framework import errors
 from tensorflow.python.framework import ops
 from tensorflow.python.platform import tf_logging as logging
 from tensorflow.python.tpu import topology
@@ -47,8 +48,8 @@ def initialize_tpu_system(cluster_resolver=None):
     The tf.tpu.Topology object for the topology of the TPU cluster.
 
   Raises:
-    RuntimeError: If no TPU devices found for eager execution or if run in a
-        tf.function.
+    RuntimeError: If running inside a tf.function.
+    NotFoundError: If no TPU devices found in eager mode.
   """
   job = None
   if cluster_resolver is None:
@@ -93,8 +94,15 @@ def initialize_tpu_system(cluster_resolver=None):
     # The TPU_SYSTEM device must match the device used in tpu.initialize_system
     # exactly, otherwise you can get errors if there are multiple TPU_SYSTEM
     # devices available.
-    with ops.device(tpu._tpu_system_device_name(job)):  # pylint: disable=protected-access
-      output = _tpu_init_fn()
+    try:
+      with ops.device(tpu._tpu_system_device_name(job)):  # pylint: disable=protected-access
+        output = _tpu_init_fn()
+      context.async_wait()
+    except errors.InvalidArgumentError as e:
+      raise errors.NotFoundError(
+          None, None,
+          "TPUs not found in the cluster. Failed in initialization: "
+          + str(e))
 
     # Clear out the eager context caches since the memory is invalid now.
     logging.info("Clearing out eager caches")
diff --git a/tensorflow/tools/api/golden/v1/tensorflow.distribute.cluster_resolver.-t-p-u-cluster-resolver.pbtxt b/tensorflow/tools/api/golden/v1/tensorflow.distribute.cluster_resolver.-t-p-u-cluster-resolver.pbtxt
index 658212aca5e..8eee489df93 100644
--- a/tensorflow/tools/api/golden/v1/tensorflow.distribute.cluster_resolver.-t-p-u-cluster-resolver.pbtxt
+++ b/tensorflow/tools/api/golden/v1/tensorflow.distribute.cluster_resolver.-t-p-u-cluster-resolver.pbtxt
@@ -15,6 +15,10 @@ tf_class {
     name: "cluster_spec"
     argspec: "args=[\'self\'], varargs=None, keywords=None, defaults=None"
   }
+  member_method {
+    name: "connect"
+    argspec: "args=[\'tpu\', \'zone\', \'project\'], varargs=None, keywords=None, defaults=[\'None\', \'None\', \'None\'], "
+  }
   member_method {
     name: "get_job_name"
     argspec: "args=[\'self\'], varargs=None, keywords=None, defaults=None"
diff --git a/tensorflow/tools/api/golden/v2/tensorflow.distribute.cluster_resolver.-t-p-u-cluster-resolver.pbtxt b/tensorflow/tools/api/golden/v2/tensorflow.distribute.cluster_resolver.-t-p-u-cluster-resolver.pbtxt
index 658212aca5e..8eee489df93 100644
--- a/tensorflow/tools/api/golden/v2/tensorflow.distribute.cluster_resolver.-t-p-u-cluster-resolver.pbtxt
+++ b/tensorflow/tools/api/golden/v2/tensorflow.distribute.cluster_resolver.-t-p-u-cluster-resolver.pbtxt
@@ -15,6 +15,10 @@ tf_class {
     name: "cluster_spec"
     argspec: "args=[\'self\'], varargs=None, keywords=None, defaults=None"
   }
+  member_method {
+    name: "connect"
+    argspec: "args=[\'tpu\', \'zone\', \'project\'], varargs=None, keywords=None, defaults=[\'None\', \'None\', \'None\'], "
+  }
   member_method {
     name: "get_job_name"
     argspec: "args=[\'self\'], varargs=None, keywords=None, defaults=None"
diff --git a/tensorflow/tools/docs/BUILD b/tensorflow/tools/docs/BUILD
index f347cab91bb..d2f5469a55c 100644
--- a/tensorflow/tools/docs/BUILD
+++ b/tensorflow/tools/docs/BUILD
@@ -11,7 +11,7 @@ package(
 
 exports_files(["LICENSE"])
 
-tpu_module = "tpu.,distribute.tpu_strategy"
+tpu_module = "tpu.,distribute.tpu_strategy,distribute.cluster_resolver.tpu,distribute.cluster_resolver.tpu_oss"
 
 py_library(
     name = "tf_doctest_lib",

From fd2d4cdb70211c6e3434e72a600b94f4c5398b0e Mon Sep 17 00:00:00 2001
From: Priya Gupta 
Date: Sat, 20 Jun 2020 00:22:08 -0700
Subject: [PATCH 383/867] Expand `Strategy.scope` API docstring to include
 details on what it does and when to use it.

PiperOrigin-RevId: 317440737
Change-Id: Ie9730f32acf19a9848c9fbfaf25720b6af64feda
---
 .../python/distribute/distribute_lib.py       | 87 ++++++++++++++++++-
 1 file changed, 83 insertions(+), 4 deletions(-)

diff --git a/tensorflow/python/distribute/distribute_lib.py b/tensorflow/python/distribute/distribute_lib.py
index ec0b911ebe0..2b4ec361868 100644
--- a/tensorflow/python/distribute/distribute_lib.py
+++ b/tensorflow/python/distribute/distribute_lib.py
@@ -787,17 +787,96 @@ class StrategyBase(object):
     finally:
       self._scale_loss_for_estimator = False
 
+  # pylint: disable=line-too-long
   def scope(self):
-    """Returns a context manager selecting this Strategy as current.
+    """Context manager to make the strategy current and distribute variables.
 
-    Inside a `with strategy.scope():` code block, this thread
-    will use a variable creator set by `strategy`, and will
-    enter its "cross-replica context".
+    This method returns a context manager, and is used as follows:
+
+    >>> strategy = tf.distribute.MirroredStrategy()
+    >>> # Variable created inside scope:
+    >>> with strategy.scope():
+    ...   mirrored_variable = tf.Variable(1.)
+    >>> mirrored_variable
+    MirroredVariable:{
+      0: 
+    }
+    >>> # Variable created outside scope:
+    >>> regular_variable = tf.Variable(1.)
+    >>> regular_variable
+    
+
+    _What happens when Strategy.scope is entered?_
+
+    * `strategy` is installed in the global context as the "current" strategy.
+      Inside this scope, `tf.distribute.get_strategy()` will now return this
+      strategy. Outside this scope, it returns the default no-op strategy.
+    * Entering the scope also enters the "cross-replica context". See
+      `tf.distribute.StrategyExtended` for an explanation on cross-replica and
+      replica contexts.
+    * Variable creation inside `scope` is intercepted by the strategy. Each
+      strategy defines how it wants to affect the variable creation. Sync
+      strategies like `MirroredStrategy`, `TPUStrategy` and
+      `MultiWorkerMiroredStrategy` create variables replicated on each replica,
+      whereas `ParameterServerStrategy` creates variables on the parameter
+      servers. This is done using a custom `tf.variable_creator_scope`.
+    * In some strategies, a default device scope may also be entered: in
+      `MultiWorkerMiroredStrategy`, a default device scope of "/CPU:0" is
+      entered on each worker.
+
+    Note: Entering a scope does not automatically distribute a computation, except
+      in the case of high level training framework like keras `model.fit`. If
+      you're not using `model.fit`, you
+      need to use `strategy.run` API to explicitly distribute that computation.
+      See an example in the [custom training loop tutorial](https://www.tensorflow.org/tutorials/distribute/custom_training).
+
+
+    _What should be in scope and what should be outside?_
+
+    There are a number of requirements on what needs to happen inside the scope.
+    However, in places where we have information about which strategy is in use,
+    we often enter the scope for the user, so they don't have to do it
+    explicitly (i.e. calling those either inside or outside the scope is OK).
+
+    * Anything that creates variables that should be distributed variables
+      must be in `strategy.scope`. This can be either by directly putting it in
+      scope, or relying on another API like `strategy.run` or `model.fit` to
+      enter it for you. Any variable that is created outside scope will not be
+      distributed and may have performance implications. Common things that
+      create variables in TF: models, optimizers, metrics. These should always
+      be created inside the scope. Another source of variable creation can be
+      a checkpoint restore - when variables are created lazily. Note that any
+      variable created inside a strategy captures the strategy information. So
+      reading and writing to these variables outside the `strategy.scope` can
+      also work seamlessly, without the user having to enter the scope.
+    * Some strategy APIs (such as `strategy.run` and `strategy.reduce`) which
+      require to be in a strategy's scope, enter the scope for you
+      automatically, which means when using those APIs you don't need to
+      enter the scope yourself.
+    * When a `tf.keras.Model` is created inside a `strategy.scope`, we capture
+      this information. When high level training frameworks methods such as
+      `model.compile`, `model.fit` etc are then called
+      on this model, we automatically enter the scope, as well as use this
+      strategy to distribute the training etc. See
+      detailed example in [distributed keras tutorial](https://www.tensorflow.org/tutorials/distribute/keras).
+      Note that simply calling the `model(..)` is not impacted - only high
+      level training framework APIs are. `model.compile`, `model.fit`,
+      `model.evaluate`, `model.predict` and `model.save` can all be called
+      inside or outside the scope.
+    * The following can be either inside or outside the scope:
+      ** Creating the input datasets
+      ** Defining `tf.function`s that represent your training step
+      ** Saving APIs such as `tf.saved_model.save`. Loading creates variables,
+         so that should go inside the scope if you want to train the model in a
+         distributed way.
+      ** Checkpoint saving. As mentioned above - `checkpoint.restore` may
+         sometimes need to be inside scope if it creates variables.
 
     Returns:
       A context manager.
     """
     return self._extended._scope(self)  # pylint: disable=protected-access
+  # pylint: enable=line-too-long
 
   @doc_controls.do_not_doc_inheritable  # DEPRECATED, moving to `extended`
   def colocate_vars_with(self, colocate_with_variable):

From cf081ee890a0d0fadaaed6a45f07e59600142722 Mon Sep 17 00:00:00 2001
From: "A. Unique TensorFlower" 
Date: Sat, 20 Jun 2020 02:01:44 -0700
Subject: [PATCH 384/867] Update GraphDef version to 438.

PiperOrigin-RevId: 317446620
Change-Id: I5c667cb1a7edf507d36a10e90cc03b52cf9227b3
---
 tensorflow/core/public/version.h | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/tensorflow/core/public/version.h b/tensorflow/core/public/version.h
index 9a79fc1eddf..64f1ef6b585 100644
--- a/tensorflow/core/public/version.h
+++ b/tensorflow/core/public/version.h
@@ -108,7 +108,7 @@ limitations under the License.
 
 #define TF_GRAPH_DEF_VERSION_MIN_PRODUCER 0
 #define TF_GRAPH_DEF_VERSION_MIN_CONSUMER 0
-#define TF_GRAPH_DEF_VERSION 437  // Updated: 2020/6/19
+#define TF_GRAPH_DEF_VERSION 438  // Updated: 2020/6/20
 
 // Checkpoint compatibility versions (the versions field in SavedSliceMeta).
 //

From 7a0531c5aac0788ee810c8e11fe00840830d8b0d Mon Sep 17 00:00:00 2001
From: "A. Unique TensorFlower" 
Date: Sat, 20 Jun 2020 02:01:45 -0700
Subject: [PATCH 385/867] compat: Update forward compatibility horizon to
 2020-06-20

PiperOrigin-RevId: 317446623
Change-Id: I57979687a0cc094b1caac5caad0bf6641bc3f575
---
 tensorflow/python/compat/compat.py | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/tensorflow/python/compat/compat.py b/tensorflow/python/compat/compat.py
index 22988d26cfc..ba6e375843b 100644
--- a/tensorflow/python/compat/compat.py
+++ b/tensorflow/python/compat/compat.py
@@ -33,7 +33,7 @@ from tensorflow.python.util.tf_export import tf_export
 # This value changes every day with an automatic CL. It can be modified in code
 # via `forward_compatibility_horizon()` or with the environment variable
 # TF_FORWARD_COMPATIBILITY_DELTA_DAYS, which is added to the compatibility date.
-_FORWARD_COMPATIBILITY_HORIZON = datetime.date(2020, 6, 19)
+_FORWARD_COMPATIBILITY_HORIZON = datetime.date(2020, 6, 20)
 _FORWARD_COMPATIBILITY_DELTA_DAYS_VAR_NAME = "TF_FORWARD_COMPATIBILITY_DELTA_DAYS"
 _FORWARD_COMPATIBILITY_DATE_NUMBER = None
 

From 5b46965a7dbcb7d775d9bca1b6bc4ee4f4652101 Mon Sep 17 00:00:00 2001
From: Lukas Geiger 
Date: Sat, 20 Jun 2020 14:34:22 +0200
Subject: [PATCH 386/867] Use ObjectIdentitySet for FuncGraph.control_captures

---
 tensorflow/python/framework/func_graph.py | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/tensorflow/python/framework/func_graph.py b/tensorflow/python/framework/func_graph.py
index b0f8821b17f..94d5913cbd1 100644
--- a/tensorflow/python/framework/func_graph.py
+++ b/tensorflow/python/framework/func_graph.py
@@ -187,7 +187,7 @@ class FuncGraph(ops.Graph):
     self.inputs = []
     self.outputs = []
     self.control_outputs = []
-    self.control_captures = set()
+    self.control_captures = object_identity.ObjectIdentitySet()
     self.structured_input_signature = None
     self.structured_outputs = None
     self._weak_variables = []

From c2009c6bc53f244e59d2788156e3a485f09ce2da Mon Sep 17 00:00:00 2001
From: Vo Van Nghia 
Date: Sat, 20 Jun 2020 21:47:50 +0700
Subject: [PATCH 387/867] Add new appendable file

---
 .../filesystem/plugins/gcs/gcs_filesystem.cc  | 39 ++++++++++++++++++-
 1 file changed, 37 insertions(+), 2 deletions(-)

diff --git a/tensorflow/c/experimental/filesystem/plugins/gcs/gcs_filesystem.cc b/tensorflow/c/experimental/filesystem/plugins/gcs/gcs_filesystem.cc
index 24d85f359ef..954ee4eefe8 100644
--- a/tensorflow/c/experimental/filesystem/plugins/gcs/gcs_filesystem.cc
+++ b/tensorflow/c/experimental/filesystem/plugins/gcs/gcs_filesystem.cc
@@ -135,6 +135,10 @@ static void Init(TF_Filesystem* filesystem, TF_Status* status) {
   TF_SetStatus(status, TF_OK, "");
 }
 
+static void Cleanup(TF_Filesystem* filesystem) {
+  plugin_memory_free(filesystem->plugin_filesystem);
+}
+
 // TODO(vnvo2409): Implement later
 
 static void NewWritableFile(const TF_Filesystem* filesystem, const char* path,
@@ -145,9 +149,37 @@ static void NewWritableFile(const TF_Filesystem* filesystem, const char* path,
   if (TF_GetCode(status) != TF_OK) return;
 
   auto gcs_client = static_cast(filesystem->plugin_filesystem);
-  TempFile outfile(TF_GetTempFileName(""), std::ios::binary | std::ios::out);
+  char* temp_file_name = TF_GetTempFileName("");
   file->plugin_file = new tf_writable_file::GCSFile(
-      {bucket, object, gcs_client, std::move(outfile), true});
+      {bucket, object, gcs_client,
+       TempFile(temp_file_name, std::ios::binary | std::ios::out), true});
+  // We are responsible for freeing the pointer returned by TF_GetTempFileName
+  free(temp_file_name);
+  TF_SetStatus(status, TF_OK, "");
+}
+
+static void NewAppendableFile(const TF_Filesystem* filesystem, const char* path,
+                              TF_WritableFile* file, TF_Status* status) {
+  char* bucket;
+  char* object;
+  ParseGCSPath(path, false, &bucket, &object, status);
+  if (TF_GetCode(status) != TF_OK) return;
+
+  auto gcs_client = static_cast(filesystem->plugin_filesystem);
+  char* temp_file_name = TF_GetTempFileName("");
+
+  auto gcs_status = gcs_client->DownloadToFile(bucket, object, temp_file_name);
+  TF_SetStatusFromGCSStatus(gcs_status, status);
+  auto status_code = TF_GetCode(status);
+  if (status_code != TF_OK && status_code != TF_NOT_FOUND) {
+    return;
+  }
+  // If this file does not exist on server, we will need to sync it.
+  bool sync_need = (status_code == TF_NOT_FOUND);
+  file->plugin_file = new tf_writable_file::GCSFile(
+      {bucket, object, gcs_client,
+       TempFile(temp_file_name, std::ios::binary | std::ios::app), sync_need});
+  free(temp_file_name);
   TF_SetStatus(status, TF_OK, "");
 }
 
@@ -165,7 +197,10 @@ static void ProvideFilesystemSupportFor(TF_FilesystemPluginOps* ops,
   ops->filesystem_ops = static_cast(
       plugin_memory_allocate(TF_FILESYSTEM_OPS_SIZE));
   ops->filesystem_ops->init = tf_gcs_filesystem::Init;
+  ops->filesystem_ops->cleanup = tf_gcs_filesystem::Cleanup;
   ops->filesystem_ops->new_writable_file = tf_gcs_filesystem::NewWritableFile;
+  ops->filesystem_ops->new_appendable_file =
+      tf_gcs_filesystem::NewAppendableFile;
 }
 
 void TF_InitPlugin(TF_FilesystemPluginInfo* info) {

From 5da215da086da25561afb6059ad52ac8c1cc7e98 Mon Sep 17 00:00:00 2001
From: Yong Tang 
Date: Sat, 20 Jun 2020 15:10:30 +0000
Subject: [PATCH 388/867] Fix TypeError when sparse.from_dense is called with
 tf.string

This PR tries to fix the issue raised in 40633 where
sparse.from_dense throws out TypeError with tf.string as input.

The issue was that from_dense uses `tf.constant(0, dtype)` to get
the zero value which fails on tf.string. This PR changes to `zeros_like`
to work with tf.string.

This PR fixes 40633.

Signed-off-by: Yong Tang 
---
 tensorflow/python/ops/sparse_ops.py      | 2 +-
 tensorflow/python/ops/sparse_ops_test.py | 9 +++++++++
 2 files changed, 10 insertions(+), 1 deletion(-)

diff --git a/tensorflow/python/ops/sparse_ops.py b/tensorflow/python/ops/sparse_ops.py
index 07eb35e84ce..cee1dc23aa0 100644
--- a/tensorflow/python/ops/sparse_ops.py
+++ b/tensorflow/python/ops/sparse_ops.py
@@ -120,7 +120,7 @@ def from_dense(tensor, name=None):
   with ops.name_scope(name, "dense_to_sparse"):
     tensor = ops.convert_to_tensor(tensor)
     indices = array_ops.where_v2(
-        math_ops.not_equal(tensor, array_ops.constant(0, tensor.dtype)))
+        math_ops.not_equal(tensor, array_ops.zeros_like(tensor)))
     values = array_ops.gather_nd(tensor, indices)
     shape = array_ops.shape(tensor, out_type=dtypes.int64)
     return sparse_tensor.SparseTensor(indices, values, shape)
diff --git a/tensorflow/python/ops/sparse_ops_test.py b/tensorflow/python/ops/sparse_ops_test.py
index 91151ba8461..cbe611ac647 100644
--- a/tensorflow/python/ops/sparse_ops_test.py
+++ b/tensorflow/python/ops/sparse_ops_test.py
@@ -180,6 +180,15 @@ class SparseOpsTest(test_util.TensorFlowTestCase, parameterized.TestCase):
         array_ops.transpose(dense_of_sparse))
     self.assertAllEqual(expected, result)
 
+  def testConstantStringToSparse(self):
+    # Test case for GitHub issue 40633.
+    tensor = constant_op.constant(list("ababa"))
+    sparse = sparse_ops.from_dense(tensor)
+    result = self.evaluate(sparse)
+    self.assertAllEqual([[0], [1], [2], [3], [4]], result.indices)
+    self.assertAllEqual([b'a', b'b', b'a', b'b', b'a'], result.values)
+    self.assertAllEqual([5], result.dense_shape)
+
 
 if __name__ == '__main__':
   googletest.main()

From 6d2ce43b03cde78a53b5fa3dfddef0cf718ccac8 Mon Sep 17 00:00:00 2001
From: Andy Ly 
Date: Sat, 20 Jun 2020 08:32:38 -0700
Subject: [PATCH 389/867] Remove empty description fields in
 tf_generated_ops.td and tf_ops.td. (NFC)

PiperOrigin-RevId: 317466704
Change-Id: Ic0bbec0c7013c2f2238a4f4e5763632c846f1337
---
 .../mlir/tensorflow/ir/tf_generated_ops.td    | 229 +-----------------
 .../compiler/mlir/tensorflow/ir/tf_ops.td     |  11 +-
 2 files changed, 4 insertions(+), 236 deletions(-)

diff --git a/tensorflow/compiler/mlir/tensorflow/ir/tf_generated_ops.td b/tensorflow/compiler/mlir/tensorflow/ir/tf_generated_ops.td
index dcef99e6971..1e5dad345f8 100644
--- a/tensorflow/compiler/mlir/tensorflow/ir/tf_generated_ops.td
+++ b/tensorflow/compiler/mlir/tensorflow/ir/tf_generated_ops.td
@@ -52,9 +52,6 @@ an output element, this operation computes \\(y = |x|\\).
 def TF_AcosOp : TF_Op<"Acos", [NoSideEffect, SameOperandsAndResultType]> {
   let summary = "Computes acos of x element-wise.";
 
-  let description = [{
-  }];
-
   let arguments = (ins
     TensorOf<[BF16, F16, F32, F64, I32, I64, TF_Complex128, TF_Complex64]>:$x
   );
@@ -371,9 +368,6 @@ retained with length 1.
 def TF_ApproximateEqualOp : TF_Op<"ApproximateEqual", [Commutative, NoSideEffect]> {
   let summary = "Returns the truth value of abs(x-y) < tolerance element-wise.";
 
-  let description = [{
-  }];
-
   let arguments = (ins
     TensorOf<[BF16, F16, F32, F64, I16, I32, I64, I8, TF_Complex128, TF_Complex64, TF_Qint32, TF_Qint8, TF_Quint8, TF_Uint16, TF_Uint32, TF_Uint64, TF_Uint8]>:$x,
     TensorOf<[BF16, F16, F32, F64, I16, I32, I64, I8, TF_Complex128, TF_Complex64, TF_Qint32, TF_Qint8, TF_Quint8, TF_Uint16, TF_Uint32, TF_Uint64, TF_Uint8]>:$y,
@@ -734,9 +728,6 @@ window in `value`.
 def TF_AvgPoolGradOp : TF_Op<"AvgPoolGrad", [NoSideEffect]> {
   let summary = "Computes gradients of the average pooling function.";
 
-  let description = [{
-  }];
-
   let arguments = (ins
     I32Tensor:$orig_input_shape,
     TF_FpTensor:$grad,
@@ -1402,9 +1393,6 @@ An n-way switch statement, implementing the following:
 def TF_CastOp : TF_Op<"Cast", [NoSideEffect, SameOperandsAndResultShape]> {
   let summary = "Cast x of type SrcT to y of DstT.";
 
-  let description = [{
-  }];
-
   let arguments = (ins
     TF_Tensor:$x,
 
@@ -1424,9 +1412,6 @@ def TF_CastOp : TF_Op<"Cast", [NoSideEffect, SameOperandsAndResultShape]> {
 def TF_CeilOp : TF_Op<"Ceil", [NoSideEffect, SameOperandsAndResultType]> {
   let summary = "Returns element-wise smallest integer not less than x.";
 
-  let description = [{
-  }];
-
   let arguments = (ins
     TF_FpTensor:$x
   );
@@ -1485,9 +1470,6 @@ greater than `clip_value_max` are set to `clip_value_max`.
 def TF_CollectiveBcastRecvOp : TF_Op<"CollectiveBcastRecv", []> {
   let summary = "Receives a tensor value broadcast from another device.";
 
-  let description = [{
-  }];
-
   let arguments = (ins
     I64Attr:$group_size,
     I64Attr:$group_key,
@@ -1507,9 +1489,6 @@ def TF_CollectiveBcastRecvOp : TF_Op<"CollectiveBcastRecv", []> {
 def TF_CollectiveBcastSendOp : TF_Op<"CollectiveBcastSend", []> {
   let summary = "Broadcasts a tensor value to one or more other devices.";
 
-  let description = [{
-  }];
-
   let arguments = (ins
     TensorOf<[F16, F32, F64, I1, I32, I64]>:$input,
 
@@ -1533,9 +1512,6 @@ def TF_CollectiveGatherOp : TF_Op<"CollectiveGather", []> {
 Mutually accumulates multiple tensors of identical type and shape.
   }];
 
-  let description = [{
-  }];
-
   let arguments = (ins
     TensorOf<[F16, F32, F64, I32, I64]>:$input,
 
@@ -1559,9 +1535,6 @@ def TF_CollectiveReduceOp : TF_Op<"CollectiveReduce", [SameOperandsAndResultType
 Mutually reduces multiple tensors of identical type and shape.
   }];
 
-  let description = [{
-  }];
-
   let arguments = (ins
     TensorOf<[F16, F32, F64, I32, I64]>:$input,
 
@@ -1641,9 +1614,6 @@ value is computed as \\( \sqrt{a^2 + b^2}\\).
 def TF_ConcatOp : TF_Op<"Concat", [NoSideEffect]> {
   let summary = "Concatenates tensors along one dimension.";
 
-  let description = [{
-  }];
-
   let arguments = (ins
     I32Tensor:$concat_dim,
     Variadic:$values
@@ -1700,9 +1670,6 @@ This is typically used by gradient computations for a concat operation.
 def TF_ConcatV2Op : TF_Op<"ConcatV2", [NoSideEffect]> {
   let summary = "Concatenates tensors along one dimension.";
 
-  let description = [{
-  }];
-
   let arguments = (ins
     Variadic:$values,
     TF_I32OrI64Tensor:$axis
@@ -1842,9 +1809,6 @@ def TF_Conv2DBackpropFilterOp : TF_Op<"Conv2DBackpropFilter", [NoSideEffect, TF_
 Computes the gradients of convolution with respect to the filter.
   }];
 
-  let description = [{
-  }];
-
   let arguments = (ins
     TF_FpTensor:$input,
     I32Tensor:$filter_sizes,
@@ -1878,9 +1842,6 @@ def TF_Conv2DBackpropInputOp : TF_Op<"Conv2DBackpropInput", [NoSideEffect, TF_La
 Computes the gradients of convolution with respect to the input.
   }];
 
-  let description = [{
-  }];
-
   let arguments = (ins
     I32Tensor:$input_sizes,
     TensorOf<[BF16, F16, F32, F64, I32]>:$filter,
@@ -1952,9 +1913,6 @@ def TF_Conv3DBackpropFilterV2Op : TF_Op<"Conv3DBackpropFilterV2", [NoSideEffect]
 Computes the gradients of 3-D convolution with respect to the filter.
   }];
 
-  let description = [{
-  }];
-
   let arguments = (ins
     TF_FpTensor:$input,
     I32Tensor:$filter_sizes,
@@ -1978,9 +1936,6 @@ def TF_Conv3DBackpropInputV2Op : TF_Op<"Conv3DBackpropInputV2", [NoSideEffect]>
 Computes the gradients of 3-D convolution with respect to the input.
   }];
 
-  let description = [{
-  }];
-
   let arguments = (ins
     TF_I32OrI64Tensor:$input_sizes,
     TF_FpTensor:$filter,
@@ -2465,9 +2420,6 @@ horizontal and vertices strides, `strides = [1, stride, stride, 1]`.
 def TF_DeviceIndexOp : TF_Op<"DeviceIndex", [NoSideEffect]> {
   let summary = "Return the index of device the op runs.";
 
-  let description = [{
-  }];
-
   let arguments = (ins
     StrArrayAttr:$device_names
   );
@@ -2792,9 +2744,6 @@ def TF_EluGradOp : TF_Op<"EluGrad", [NoSideEffect, SameOperandsAndResultType]> {
 Computes gradients for the exponential linear (Elu) operation.
   }];
 
-  let description = [{
-  }];
-
   let arguments = (ins
     TF_FpTensor:$gradients,
     TF_FpTensor:$outputs
@@ -2814,9 +2763,6 @@ Creates a tensor with the given shape.
 This operation creates a tensor of `shape` and `dtype`.
   }];
 
-  let description = [{
-  }];
-
   let arguments = (ins
     I32Tensor:$shape,
 
@@ -2946,9 +2892,6 @@ tf.math.equal(x, y) ==> array([True,  True])
 def TF_ErfOp : TF_Op<"Erf", [NoSideEffect, SameOperandsAndResultType]> {
   let summary = "Computes the Gauss error function of `x` element-wise.";
 
-  let description = [{
-  }];
-
   let arguments = (ins
     TF_FpTensor:$x
   );
@@ -2965,9 +2908,6 @@ def TF_ErfcOp : TF_Op<"Erfc", [NoSideEffect, SameOperandsAndResultType]> {
 Computes the complementary error function of `x` element-wise.
   }];
 
-  let description = [{
-  }];
-
   let arguments = (ins
     TF_FpTensor:$x
   );
@@ -2982,9 +2922,6 @@ Computes the complementary error function of `x` element-wise.
 def TF_ErfinvOp : TF_Op<"Erfinv", [NoSideEffect]> {
   let summary = "";
 
-  let description = [{
-  }];
-
   let arguments = (ins
     TF_FpTensor:$x
   );
@@ -3190,9 +3127,6 @@ def TF_FakeParamOp : TF_Op<"FakeParam", [NoSideEffect]> {
   intermediate output needed for the gradient computation of the other branch).
   }];
 
-  let description = [{
-  }];
-
   let arguments = (ins
     TF_ShapeAttr:$shape
   );
@@ -3402,9 +3336,6 @@ fill([2, 3], 9) ==> [[9, 9, 9]
 def TF_FloorOp : TF_Op<"Floor", [NoSideEffect, SameOperandsAndResultType]> {
   let summary = "Returns element-wise largest integer not greater than x.";
 
-  let description = [{
-  }];
-
   let arguments = (ins
     TF_FpTensor:$x
   );
@@ -4212,9 +4143,6 @@ def TF_IgammaGradAOp : TF_Op<"IgammaGradA", [NoSideEffect, ResultsBroadcastableS
                        WithBroadcastableBinOpBuilder {
   let summary = "Computes the gradient of `igamma(a, x)` wrt `a`.";
 
-  let description = [{
-  }];
-
   let arguments = (ins
     TF_F32OrF64Tensor:$a,
     TF_F32OrF64Tensor:$x
@@ -4487,9 +4415,6 @@ tf.math.is_nan(x) ==> [False, True, False, True, False]
 def TF_IteratorGetNextOp : TF_Op<"IteratorGetNext", []> {
   let summary = "Gets the next output from the given iterator .";
 
-  let description = [{
-  }];
-
   let arguments = (ins
     TF_ResourceTensor:$iterator
   );
@@ -4558,9 +4483,6 @@ convolutional neural networks (NIPS 2012)](http://papers.nips.cc/paper/4824-imag
 def TF_LRNGradOp : TF_Op<"LRNGrad", [NoSideEffect]> {
   let summary = "Gradients for Local Response Normalization.";
 
-  let description = [{
-  }];
-
   let arguments = (ins
     TensorOf<[BF16, F16, F32]>:$input_grads,
     TensorOf<[BF16, F16, F32]>:$input_image,
@@ -4582,9 +4504,6 @@ def TF_LRNGradOp : TF_Op<"LRNGrad", [NoSideEffect]> {
 def TF_LeakyReluOp : TF_Op<"LeakyRelu", [NoSideEffect, SameOperandsAndResultType]> {
   let summary = "Computes rectified linear: `max(features, features * alpha)`.";
 
-  let description = [{
-  }];
-
   let arguments = (ins
     TF_FpTensor:$features,
 
@@ -4605,9 +4524,6 @@ def TF_LeakyReluGradOp : TF_Op<"LeakyReluGrad", [NoSideEffect, SameOperandsAndRe
 Computes rectified linear gradients for a LeakyRelu operation.
   }];
 
-  let description = [{
-  }];
-
   let arguments = (ins
     TF_FpTensor:$gradients,
     TF_FpTensor:$features,
@@ -4888,9 +4804,6 @@ def TF_LogicalAndOp : TF_Op<"LogicalAnd", [Commutative, NoSideEffect, ResultsBro
 def TF_LogicalNotOp : TF_Op<"LogicalNot", [NoSideEffect, SameOperandsAndResultType]> {
   let summary = "Returns the truth value of `NOT x` element-wise.";
 
-  let description = [{
-  }];
-
   let arguments = (ins
     I1Tensor:$x
   );
@@ -4971,9 +4884,6 @@ The tensor `values` must be of the type of the table values.
 def TF_LookupTableSizeV2Op : TF_Op<"LookupTableSizeV2", []> {
   let summary = "Computes the number of elements in the given table.";
 
-  let description = [{
-  }];
-
   let arguments = (ins
     TF_ResourceTensor:$table_handle
   );
@@ -5658,9 +5568,6 @@ retained with length 1.
 def TF_MaxPoolOp : TF_Op<"MaxPool", [NoSideEffect, TF_FoldOperandsTransposeInterface]> {
   let summary = "Performs max pooling on the input.";
 
-  let description = [{
-  }];
-
   let arguments = (ins
     TensorOf<[BF16, F16, F32, F64, I16, I32, I64, I8, TF_Qint8, TF_Uint16, TF_Uint8]>:$input,
 
@@ -5687,9 +5594,6 @@ def TF_MaxPoolOp : TF_Op<"MaxPool", [NoSideEffect, TF_FoldOperandsTransposeInter
 def TF_MaxPool3DOp : TF_Op<"MaxPool3D", [NoSideEffect]> {
   let summary = "Performs 3D max pooling on the input.";
 
-  let description = [{
-  }];
-
   let arguments = (ins
     TensorOf<[BF16, F16, F32]>:$input,
 
@@ -5709,9 +5613,6 @@ def TF_MaxPool3DOp : TF_Op<"MaxPool3D", [NoSideEffect]> {
 def TF_MaxPool3DGradOp : TF_Op<"MaxPool3DGrad", [NoSideEffect]> {
   let summary = "Computes gradients of 3D max pooling function.";
 
-  let description = [{
-  }];
-
   let arguments = (ins
     TensorOf<[BF16, F16, F32]>:$orig_input,
     TensorOf<[BF16, F16, F32]>:$orig_output,
@@ -5734,9 +5635,6 @@ def TF_MaxPool3DGradOp : TF_Op<"MaxPool3DGrad", [NoSideEffect]> {
 def TF_MaxPoolGradOp : TF_Op<"MaxPoolGrad", [NoSideEffect]> {
   let summary = "Computes gradients of the maxpooling function.";
 
-  let description = [{
-  }];
-
   let arguments = (ins
     TF_IntOrFpTensor:$orig_input,
     TF_IntOrFpTensor:$orig_output,
@@ -6015,9 +5913,6 @@ Returns x * y element-wise. Returns zero if y is zero, even if x if infinite or
 def TF_MultinomialOp : TF_Op<"Multinomial", []> {
   let summary = "Draws samples from a multinomial distribution.";
 
-  let description = [{
-  }];
-
   let arguments = (ins
     TF_IntOrFpTensor:$logits,
     I32Tensor:$num_samples,
@@ -6037,9 +5932,6 @@ def TF_MultinomialOp : TF_Op<"Multinomial", []> {
 def TF_NdtriOp : TF_Op<"Ndtri", [NoSideEffect]> {
   let summary = "";
 
-  let description = [{
-  }];
-
   let arguments = (ins
     TF_FpTensor:$x
   );
@@ -6074,9 +5966,6 @@ I.e., \\(y = -x\\).
 def TF_NoOp : TF_Op<"NoOp", [NoSideEffect]> {
   let summary = "Does nothing. Only useful as a placeholder for control edges.";
 
-  let description = [{
-  }];
-
   let arguments = (ins);
 
   let results = (outs);
@@ -6330,9 +6219,6 @@ output =
 def TF_OutfeedEnqueueTupleOp : TF_Op<"OutfeedEnqueueTuple", []> {
   let summary = "Enqueue multiple Tensor values on the computation outfeed.";
 
-  let description = [{
-  }];
-
   let arguments = (ins
     Variadic:$inputs
   );
@@ -6617,9 +6503,6 @@ q_full, r_full = qr(a, full_matrices=True)
 def TF_QuantizeAndDequantizeOp : TF_Op<"QuantizeAndDequantize", [NoSideEffect, SameOperandsAndResultType]> {
   let summary = "Use QuantizeAndDequantizeV2 instead.";
 
-  let description = [{
-  }];
-
   let arguments = (ins
     TF_FpTensor:$input,
 
@@ -6871,9 +6754,6 @@ def TF_RandomGammaGradOp : TF_Op<"RandomGammaGrad", [NoSideEffect, ResultsBroadc
 Computes the derivative of a Gamma random sample w.r.t. `alpha`.
   }];
 
-  let description = [{
-  }];
-
   let arguments = (ins
     TF_F32OrF64Tensor:$alpha,
     TF_F32OrF64Tensor:$sample
@@ -7203,9 +7083,6 @@ array([ 0.,  0., -0.,  3.], dtype=float32)
 def TF_Relu6Op : TF_Op<"Relu6", [NoSideEffect, SameOperandsAndResultType]> {
   let summary = "Computes rectified linear 6: `min(max(features, 0), 6)`.";
 
-  let description = [{
-  }];
-
   let arguments = (ins
     TF_IntOrFpTensor:$features
   );
@@ -7220,9 +7097,6 @@ def TF_Relu6Op : TF_Op<"Relu6", [NoSideEffect, SameOperandsAndResultType]> {
 def TF_Relu6GradOp : TF_Op<"Relu6Grad", [NoSideEffect, SameOperandsAndResultType]> {
   let summary = "Computes rectified linear 6 gradients for a Relu6 operation.";
 
-  let description = [{
-  }];
-
   let arguments = (ins
     TF_IntOrFpTensor:$gradients,
     TF_IntOrFpTensor:$features
@@ -7238,9 +7112,6 @@ def TF_Relu6GradOp : TF_Op<"Relu6Grad", [NoSideEffect, SameOperandsAndResultType
 def TF_ReluGradOp : TF_Op<"ReluGrad", [NoSideEffect, SameOperandsAndResultType]> {
   let summary = "Computes rectified linear gradients for a Relu operation.";
 
-  let description = [{
-  }];
-
   let arguments = (ins
     TF_IntOrFpTensor:$gradients,
     TF_IntOrFpTensor:$features
@@ -7365,9 +7236,6 @@ Input images can be of different types but output images are always float.
 def TF_ResizeBilinearGradOp : TF_Op<"ResizeBilinearGrad", [NoSideEffect]> {
   let summary = "Computes the gradient of bilinear interpolation.";
 
-  let description = [{
-  }];
-
   let arguments = (ins
     F32Tensor:$grads,
     TF_FpTensor:$original_image,
@@ -7388,9 +7256,6 @@ def TF_ResizeNearestNeighborOp : TF_Op<"ResizeNearestNeighbor", [NoSideEffect]>
 Resize `images` to `size` using nearest neighbor interpolation.
   }];
 
-  let description = [{
-  }];
-
   let arguments = (ins
     TensorOf<[BF16, F16, F32, F64, I16, I32, I64, I8, TF_Uint16, TF_Uint8]>:$images,
     I32Tensor:$size,
@@ -7507,9 +7372,6 @@ var <- var - mom
 def TF_ResourceApplyGradientDescentOp : TF_Op<"ResourceApplyGradientDescent", []> {
   let summary = "Update '*var' by subtracting 'alpha' * 'delta' from it.";
 
-  let description = [{
-  }];
-
   let arguments = (ins
     TF_ResourceTensor:$var,
     TensorOf<[BF16, F16, F32, F64, I16, I32, I64, I8, TF_Complex128, TF_Complex64, TF_Qint32, TF_Qint8, TF_Quint8, TF_Uint16, TF_Uint32, TF_Uint64, TF_Uint8]>:$alpha,
@@ -8292,9 +8154,6 @@ select(condition, t, e) ==> [[1, 2],
 def TF_SelectV2Op : TF_Op<"SelectV2", [NoSideEffect, ResultsBroadcastableShape]> {
   let summary = "";
 
-  let description = [{
-  }];
-
   let arguments = (ins
     I1Tensor:$condition,
     TF_Tensor:$t,
@@ -8343,9 +8202,6 @@ def TF_SeluGradOp : TF_Op<"SeluGrad", [NoSideEffect, SameOperandsAndResultType]>
 Computes gradients for the scaled exponential linear (Selu) operation.
   }];
 
-  let description = [{
-  }];
-
   let arguments = (ins
     TF_FpTensor:$gradients,
     TF_FpTensor:$outputs
@@ -8596,9 +8452,6 @@ whose values are extracted from 'input' starting at the offsets in
 def TF_SnapshotOp : TF_Op<"Snapshot", [NoSideEffect, SameOperandsAndResultType]> {
   let summary = "Returns a copy of the input tensor.";
 
-  let description = [{
-  }];
-
   let arguments = (ins
     TF_Tensor:$input
   );
@@ -8663,9 +8516,6 @@ Inputs are the logits, not probabilities.
 def TF_SoftplusOp : TF_Op<"Softplus", [NoSideEffect, SameOperandsAndResultType]> {
   let summary = "Computes softplus: `log(exp(features) + 1)`.";
 
-  let description = [{
-  }];
-
   let arguments = (ins
     TF_FpTensor:$features
   );
@@ -8680,9 +8530,6 @@ def TF_SoftplusOp : TF_Op<"Softplus", [NoSideEffect, SameOperandsAndResultType]>
 def TF_SoftplusGradOp : TF_Op<"SoftplusGrad", [NoSideEffect, SameOperandsAndResultType]> {
   let summary = "Computes softplus gradients for a softplus operation.";
 
-  let description = [{
-  }];
-
   let arguments = (ins
     TF_FpTensor:$gradients,
     TF_FpTensor:$features
@@ -8698,9 +8545,6 @@ def TF_SoftplusGradOp : TF_Op<"SoftplusGrad", [NoSideEffect, SameOperandsAndResu
 def TF_SoftsignOp : TF_Op<"Softsign", [NoSideEffect, SameOperandsAndResultType]> {
   let summary = "Computes softsign: `features / (abs(features) + 1)`.";
 
-  let description = [{
-  }];
-
   let arguments = (ins
     TF_FpTensor:$features
   );
@@ -8715,9 +8559,6 @@ def TF_SoftsignOp : TF_Op<"Softsign", [NoSideEffect, SameOperandsAndResultType]>
 def TF_SoftsignGradOp : TF_Op<"SoftsignGrad", [NoSideEffect, SameOperandsAndResultType]> {
   let summary = "Computes softsign gradients for a softsign operation.";
 
-  let description = [{
-  }];
-
   let arguments = (ins
     TF_FpTensor:$gradients,
     TF_FpTensor:$features
@@ -8965,9 +8806,6 @@ are checked during execution.
 def TF_SplitOp : TF_Op<"Split", [NoSideEffect]> {
   let summary = "Splits a tensor into `num_split` tensors along one dimension.";
 
-  let description = [{
-  }];
-
   let arguments = (ins
     I32Tensor:$split_dim,
     TF_Tensor:$value
@@ -8986,9 +8824,6 @@ def TF_SplitOp : TF_Op<"Split", [NoSideEffect]> {
 def TF_SplitVOp : TF_Op<"SplitV", [NoSideEffect]> {
   let summary = "Splits a tensor into `num_split` tensors along one dimension.";
 
-  let description = [{
-  }];
-
   let arguments = (ins
     TF_Tensor:$value,
     TF_I32OrI64Tensor:$size_splits,
@@ -9052,11 +8887,11 @@ I.e., \\(y = x * x = x^2\\).
   }];
 
   let arguments = (ins
-    TensorOf<[BF16, F16, F32, F64, I32, I64, TF_Complex128, TF_Complex64]>:$x
+    TensorOf<[BF16, F16, F32, F64, I16, I32, I64, I8, TF_Complex128, TF_Complex64]>:$x
   );
 
   let results = (outs
-    TensorOf<[BF16, F16, F32, F64, I32, I64, TF_Complex128, TF_Complex64]>:$y
+    TensorOf<[BF16, F16, F32, F64, I16, I32, I64, I8, TF_Complex128, TF_Complex64]>:$y
   );
 
   TF_DerivedOperandTypeAttr T = TF_DerivedOperandTypeAttr<0>;
@@ -9125,9 +8960,6 @@ shape(squeeze(t, [2, 4])) ==> [1, 2, 3, 1]
 def TF_StackCloseV2Op : TF_Op<"StackCloseV2", []> {
   let summary = "Delete the stack from its resource container.";
 
-  let description = [{
-  }];
-
   let arguments = (ins
     TF_ResourceTensor:$handle
   );
@@ -9138,9 +8970,6 @@ def TF_StackCloseV2Op : TF_Op<"StackCloseV2", []> {
 def TF_StackPopV2Op : TF_Op<"StackPopV2", []> {
   let summary = "Pop the element at the top of the stack.";
 
-  let description = [{
-  }];
-
   let arguments = (ins
     TF_ResourceTensor:$handle
   );
@@ -9155,9 +8984,6 @@ def TF_StackPopV2Op : TF_Op<"StackPopV2", []> {
 def TF_StackPushV2Op : TF_Op<"StackPushV2", []> {
   let summary = "Push an element onto the stack.";
 
-  let description = [{
-  }];
-
   let arguments = (ins
     TF_ResourceTensor:$handle,
     TF_Tensor:$elem,
@@ -9175,9 +9001,6 @@ def TF_StackPushV2Op : TF_Op<"StackPushV2", []> {
 def TF_StackV2Op : TF_Op<"StackV2", []> {
   let summary = "A stack that produces elements in first-in last-out order.";
 
-  let description = [{
-  }];
-
   let arguments = (ins
     I32Tensor:$max_size,
 
@@ -9895,9 +9718,6 @@ calculation gets its own TensorArray accumulator.
 def TF_TensorArrayReadV3Op : TF_Op<"TensorArrayReadV3", []> {
   let summary = "Read an element from the TensorArray into output `value`.";
 
-  let description = [{
-  }];
-
   let arguments = (ins
     TF_ResourceTensor:$handle,
     I32Tensor:$index,
@@ -9937,9 +9757,6 @@ Scatter the data from the input value into specific TensorArray elements.
 def TF_TensorArraySizeV3Op : TF_Op<"TensorArraySizeV3", []> {
   let summary = "Get the current size of the TensorArray.";
 
-  let description = [{
-  }];
-
   let arguments = (ins
     TF_ResourceTensor:$handle,
     F32Tensor:$flow_in
@@ -10016,9 +9833,6 @@ Write data via Write and read via Read or Pack.
 def TF_TensorArrayWriteV3Op : TF_Op<"TensorArrayWriteV3", []> {
   let summary = "Push an element onto the tensor_array.";
 
-  let description = [{
-  }];
-
   let arguments = (ins
     TF_ResourceTensor:$handle,
     I32Tensor:$index,
@@ -10139,9 +9953,6 @@ values: The tensor.
 def TF_TensorListGetItemOp : TF_Op<"TensorListGetItem", [NoSideEffect]> {
   let summary = "";
 
-  let description = [{
-  }];
-
   let arguments = (ins
     TF_VariantTensor:$input_handle,
     I32Tensor:$index,
@@ -10271,9 +10082,6 @@ output_handle: The TensorList.
 def TF_TensorListSetItemOp : TF_Op<"TensorListSetItem", [NoSideEffect]> {
   let summary = "";
 
-  let description = [{
-  }];
-
   let arguments = (ins
     TF_VariantTensor:$input_handle,
     I32Tensor:$index,
@@ -11063,9 +10871,6 @@ def TF_XdivyOp : TF_Op<"Xdivy", [NoSideEffect, ResultsBroadcastableShape]>,
                  WithBroadcastableBinOpBuilder {
   let summary = "Returns 0 if x == 0, and x / y otherwise, elementwise.";
 
-  let description = [{
-  }];
-
   let arguments = (ins
     TensorOf<[F16, F32, F64, TF_Complex128, TF_Complex64]>:$x,
     TensorOf<[F16, F32, F64, TF_Complex128, TF_Complex64]>:$y
@@ -11242,9 +11047,6 @@ def TF_XlaHostComputeOp : TF_Op<"XlaHostCompute", []> {
 A pseudo-op to represent host-side computation in an XLA program.
   }];
 
-  let description = [{
-  }];
-
   let arguments = (ins
     Variadic:$inputs,
 
@@ -11315,9 +11117,6 @@ https://www.tensorflow.org/performance/xla/operation_semantics#pad
 def TF_XlaRecvFromHostOp : TF_Op<"XlaRecvFromHost", []> {
   let summary = "An op to receive a tensor from the host.";
 
-  let description = [{
-  }];
-
   let arguments = (ins
     TF_ShapeAttr:$shape,
     StrAttr:$key
@@ -11355,9 +11154,6 @@ https://www.tensorflow.org/performance/xla/operation_semantics#reduce .
 def TF_XlaReplicaIdOp : TF_Op<"XlaReplicaId", [NoSideEffect]> {
   let summary = "Replica ID.";
 
-  let description = [{
-  }];
-
   let arguments = (ins);
 
   let results = (outs
@@ -11397,9 +11193,6 @@ i=0...N-1.
 def TF_XlaSendToHostOp : TF_Op<"XlaSendToHost", []> {
   let summary = "An op to send a tensor to the host.";
 
-  let description = [{
-  }];
-
   let arguments = (ins
     TF_Tensor:$input,
 
@@ -11443,9 +11236,6 @@ tensor such that tensor[...,:,:] = u[..., :, :] * Diag(s[..., :]) * Transpose(v[
 def TF_Xlog1pyOp : TF_Op<"Xlog1py", [NoSideEffect]> {
   let summary = "Returns 0 if x == 0, and x * log1p(y) otherwise, elementwise.";
 
-  let description = [{
-  }];
-
   let arguments = (ins
     TensorOf<[F16, F32, F64, TF_Complex128, TF_Complex64]>:$x,
     TensorOf<[F16, F32, F64, TF_Complex128, TF_Complex64]>:$y
@@ -11462,9 +11252,6 @@ def TF_XlogyOp : TF_Op<"Xlogy", [NoSideEffect, ResultsBroadcastableShape]>,
                  WithBroadcastableBinOpBuilder {
   let summary = "Returns 0 if x == 0, and x * log(y) otherwise, elementwise.";
 
-  let description = [{
-  }];
-
   let arguments = (ins
     TensorOf<[F16, F32, F64, TF_Complex128, TF_Complex64]>:$x,
     TensorOf<[F16, F32, F64, TF_Complex128, TF_Complex64]>:$y
@@ -11480,9 +11267,6 @@ def TF_XlogyOp : TF_Op<"Xlogy", [NoSideEffect, ResultsBroadcastableShape]>,
 def TF_ZerosLikeOp : TF_Op<"ZerosLike", [NoSideEffect, SameOperandsAndResultType]> {
   let summary = "Returns a tensor of zeros with the same shape and type as x.";
 
-  let description = [{
-  }];
-
   let arguments = (ins
     TF_Tensor:$x
   );
@@ -11589,9 +11373,6 @@ expected to create these operators.
 def TF__HostComputeMlirOp : TF_Op<"_HostComputeMlir", []> {
   let summary = "A host-side computation called from a TPU device.";
 
-  let description = [{
-  }];
-
   let arguments = (ins
     Variadic:$inputs,
 
@@ -11671,9 +11452,6 @@ def TF__XlaRecvAtHostOp : TF_Op<"_XlaRecvAtHost", []> {
 A placeholder op to receive values from a running XLA computation.
   }];
 
-  let description = [{
-  }];
-
   let arguments = (ins
     TF_StrTensor:$dynamic_key,
 
@@ -11691,9 +11469,6 @@ A placeholder op to receive values from a running XLA computation.
 def TF__XlaSendFromHostOp : TF_Op<"_XlaSendFromHost", []> {
   let summary = "A placeholder op to send values to a running XLA computation.";
 
-  let description = [{
-  }];
-
   let arguments = (ins
     Variadic:$inputs,
     TF_StrTensor:$dynamic_key,
diff --git a/tensorflow/compiler/mlir/tensorflow/ir/tf_ops.td b/tensorflow/compiler/mlir/tensorflow/ir/tf_ops.td
index d8675bb786f..24e88b0e966 100644
--- a/tensorflow/compiler/mlir/tensorflow/ir/tf_ops.td
+++ b/tensorflow/compiler/mlir/tensorflow/ir/tf_ops.td
@@ -232,6 +232,7 @@ else_branch: A function that takes 'inputs' and returns a list of
 
 def TF_YieldOp : TF_Op<"Yield", [Terminator]> {
   let summary = "Yield operation";
+
   let description = [{
     The "yield" operation represents a return operation within the conditional
     and body of structured control flow (e.g., if and while). The operation
@@ -497,6 +498,7 @@ Inserts a placeholder for a tensor that will be always fed.
 
 def TF_PlaceholderWithDefaultOp : TF_Op<"PlaceholderWithDefault", [NoSideEffect]> {
   let summary = "Placeholder op";
+
   let description = [{
     A placeholder op that passes through input when its output is not fed.
   }];
@@ -839,9 +841,6 @@ def TF_XlaShardingOp : TF_Op<"XlaSharding", [NoSideEffect]> {
 An op which shards the input based on the given sharding attribute.
   }];
 
-  let description = [{
-  }];
-
   let arguments = (ins
     TF_Tensor:$input,
 
@@ -858,9 +857,6 @@ An op which shards the input based on the given sharding attribute.
 def TF_InfeedDequeueTupleOp : TF_Op<"InfeedDequeueTuple", []> {
   let summary = "Fetches multiple values from infeed as an XLA tuple.";
 
-  let description = [{
-  }];
-
   let arguments = (ins
     OptionalAttr:$_XlaSharding
   );
@@ -904,9 +900,6 @@ def TF_BatchDatasetV2Op : TF_Op<"BatchDatasetV2", [NoSideEffect]> {
 Creates a dataset that batches `batch_size` elements from `input_dataset`.
   }];
 
-  let description = [{
-  }];
-
   let arguments = (ins
     TF_VariantTensor:$input_dataset,
     I64Tensor:$batch_size,

From e647a3b425ea63ff5e2e2338815ca4aea188c619 Mon Sep 17 00:00:00 2001
From: "A. Unique TensorFlower" 
Date: Sat, 20 Jun 2020 11:28:17 -0700
Subject: [PATCH 390/867] Add experimental C API to access EagerContext context
 ID.

PiperOrigin-RevId: 317476439
Change-Id: I9e97bce61cf526695f0c903b5f4f837116fef455
---
 tensorflow/c/eager/c_api_experimental.cc | 6 ++++++
 tensorflow/c/eager/c_api_experimental.h  | 8 ++++++++
 tensorflow/python/tfe_wrapper.cc         | 3 +++
 3 files changed, 17 insertions(+)

diff --git a/tensorflow/c/eager/c_api_experimental.cc b/tensorflow/c/eager/c_api_experimental.cc
index 9937fd7551f..7390cf243be 100644
--- a/tensorflow/c/eager/c_api_experimental.cc
+++ b/tensorflow/c/eager/c_api_experimental.cc
@@ -60,6 +60,12 @@ void TFE_ContextDisableGraphCollection(TFE_Context* ctx) {
   context->SetShouldStoreGraphs(false);
 }
 
+uint64_t TFE_GetContextId(TFE_Context* ctx) {
+  tensorflow::EagerContext* context =
+      tensorflow::ContextFromInterface(tensorflow::unwrap(ctx));
+  return context->GetContextId();
+}
+
 void TFE_MonitoringCounterCellIncrementBy(TFE_MonitoringCounterCell* cell,
                                           int64_t value) {
   cell->cell.IncrementBy(value);
diff --git a/tensorflow/c/eager/c_api_experimental.h b/tensorflow/c/eager/c_api_experimental.h
index 1b8efe61ee0..1af76c01154 100644
--- a/tensorflow/c/eager/c_api_experimental.h
+++ b/tensorflow/c/eager/c_api_experimental.h
@@ -300,6 +300,14 @@ TF_CAPI_EXPORT extern void TFE_ContextOptionsSetLazyRemoteInputsCopy(
 TF_CAPI_EXPORT extern void TFE_ContextOptionsSetTfrt(TFE_ContextOptions*,
                                                      bool use_tfrt);
 
+// Returns the context_id from the EagerContext which is used by the
+// EagerService to maintain consistency between client and worker. The
+// context_id is initialized with a dummy value and is later set when the worker
+// is initialized (either locally or remotely). The context_id can change during
+// the process lifetime although this should cause the worker to be
+// reinitialized (e.g. cleared caches) as well.
+TF_CAPI_EXPORT extern uint64_t TFE_GetContextId(TFE_Context* ctx);
+
 // -----------------------------------------------------------------------------
 // Cancellation APIs.
 
diff --git a/tensorflow/python/tfe_wrapper.cc b/tensorflow/python/tfe_wrapper.cc
index 00137f6f492..80cce331353 100644
--- a/tensorflow/python/tfe_wrapper.cc
+++ b/tensorflow/python/tfe_wrapper.cc
@@ -461,6 +461,9 @@ PYBIND11_MODULE(_pywrap_tfe, m) {
   m.def("TFE_ContextClearCaches", [](py::handle& o) {
     TFE_ContextClearCaches(tensorflow::InputTFE_Context(o));
   });
+  m.def("TFE_GetContextId", [](py::handle& ctx) {
+    return TFE_GetContextId(tensorflow::InputTFE_Context(ctx));
+  });
   m.def("TFE_ContextGetDevicePlacementPolicy", [](py::handle& ctx) {
     return TFE_ContextGetDevicePlacementPolicy(
         tensorflow::InputTFE_Context(ctx));

From 39a2286a0e28069e8c1f8100dd807c3d24aacfdc Mon Sep 17 00:00:00 2001
From: Ruoxin Sang 
Date: Sat, 20 Jun 2020 13:06:21 -0700
Subject: [PATCH 391/867] Graduate TPUStrategy from experimental.

RELNOTES=Make TPUStrategy symbol non experimental.
PiperOrigin-RevId: 317482072
Change-Id: I8bf596729699cb02fa275dfb63855c2dc68c1d42
---
 .../debug/lib/check_numerics_callback.py      |   2 +-
 .../python/debug/lib/dumping_callback.py      |   2 +-
 tensorflow/python/distribute/README.md        |   2 +-
 .../python/distribute/distribute_lib.py       |  34 +--
 .../python/distribute/mirrored_strategy.py    |   5 +-
 tensorflow/python/distribute/tpu_strategy.py  | 196 +++++++++++++++++-
 tensorflow/python/tpu/tpu_embedding_v2.py     |  12 +-
 ...ensorflow.distribute.-t-p-u-strategy.pbtxt |  99 +++++++++
 .../api/golden/v2/tensorflow.distribute.pbtxt |   4 +
 .../tools/compatibility/tf_upgrade_v2.py      |   2 +-
 .../tools/compatibility/tf_upgrade_v2_test.py |   2 +-
 11 files changed, 323 insertions(+), 37 deletions(-)
 create mode 100644 tensorflow/tools/api/golden/v2/tensorflow.distribute.-t-p-u-strategy.pbtxt

diff --git a/tensorflow/python/debug/lib/check_numerics_callback.py b/tensorflow/python/debug/lib/check_numerics_callback.py
index 796fabae301..bd88ec5e122 100644
--- a/tensorflow/python/debug/lib/check_numerics_callback.py
+++ b/tensorflow/python/debug/lib/check_numerics_callback.py
@@ -422,7 +422,7 @@ def enable_check_numerics(stack_height_limit=30,
   tf.debugging.enable_check_numerics()
 
   resolver = tf.distribute.cluster_resolver.TPUClusterResolver(tpu='')
-  strategy = tf.distribute.experimental.TPUStrategy(resolver)
+  strategy = tf.distribute.TPUStrategy(resolver)
   with strategy.scope():
     # ...
   ```
diff --git a/tensorflow/python/debug/lib/dumping_callback.py b/tensorflow/python/debug/lib/dumping_callback.py
index 563b52f8f63..56de65d2339 100644
--- a/tensorflow/python/debug/lib/dumping_callback.py
+++ b/tensorflow/python/debug/lib/dumping_callback.py
@@ -737,7 +737,7 @@ def enable_dump_debug_info(dump_root,
       logdir, tensor_debug_mode="FULL_HEALTH")
 
   resolver = tf.distribute.cluster_resolver.TPUClusterResolver(tpu='')
-  strategy = tf.distribute.experimental.TPUStrategy(resolver)
+  strategy = tf.distribute.TPUStrategy(resolver)
   with strategy.scope():
     # ...
   ```
diff --git a/tensorflow/python/distribute/README.md b/tensorflow/python/distribute/README.md
index f44a4ee8531..26a62dacb30 100644
--- a/tensorflow/python/distribute/README.md
+++ b/tensorflow/python/distribute/README.md
@@ -49,7 +49,7 @@ model.evaluate(dataset)
 
 ```python
 # Create the strategy instance.
-tpu_strategy = tf.distribute.experimental.TPUStrategy(resolver)
+tpu_strategy = tf.distribute.TPUStrategy(resolver)
 
 
 # Create the keras model under strategy.scope()
diff --git a/tensorflow/python/distribute/distribute_lib.py b/tensorflow/python/distribute/distribute_lib.py
index 2b4ec361868..5abfb6e1c09 100644
--- a/tensorflow/python/distribute/distribute_lib.py
+++ b/tensorflow/python/distribute/distribute_lib.py
@@ -618,7 +618,7 @@ class InputOptions(
   resolver = tf.distribute.cluster_resolver.TPUClusterResolver(tpu='')
   tf.config.experimental_connect_to_cluster(resolver)
   tf.tpu.experimental.initialize_tpu_system(resolver)
-  strategy = tf.distribute.experimental.TPUStrategy(resolver)
+  strategy = tf.distribute.TPUStrategy(resolver)
 
   dataset = tf.data.Dataset.range(16)
   distributed_dataset_on_host = (
@@ -1462,17 +1462,17 @@ class Strategy(StrategyBase):
     topology = tf.tpu.experimental.initialize_tpu_system(resolver)
     device_assignment = tf.tpu.experimental.DeviceAssignment.build(
         topology,
-        computation_shape=[1, 1, 2],
+        computation_shape=[1, 1, 1, 2],
         num_replicas=4)
-    strategy = tf.distribute.experimental.TPUStrategy(
-        resolver, device_assignment=device_assignment)
+    strategy = tf.distribute.TPUStrategy(
+        resolver, experimental_device_assignment=device_assignment)
     iterator = iter(inputs)
 
     @tf.function()
     def step_fn(inputs):
       output = tf.add(inputs, inputs)
 
-      // Add operation will be executed on logical device 0.
+      # Add operation will be executed on logical device 0.
       output = strategy.experimental_assign_to_logical_device(output, 0)
       return output
 
@@ -1517,10 +1517,10 @@ class Strategy(StrategyBase):
     topology = tf.tpu.experimental.initialize_tpu_system(resolver)
     device_assignment = tf.tpu.experimental.DeviceAssignment.build(
         topology,
-        computation_shape=[2, 2, 2],
+        computation_shape=[1, 2, 2, 2],
         num_replicas=1)
-    strategy = tf.distribute.experimental.TPUStrategy(
-        resolver, device_assignment=device_assignment)
+    strategy = tf.distribute.TPUStrategy(
+        resolver, experimental_device_assignment=device_assignment)
 
     iterator = iter(inputs)
 
@@ -1529,8 +1529,8 @@ class Strategy(StrategyBase):
       inputs = strategy.experimental_split_to_logical_devices(
         inputs, [1, 2, 4, 1])
 
-      // model() function will be executed on 8 logical devices with `inputs`
-      // split 2 * 4  ways.
+      # model() function will be executed on 8 logical devices with `inputs`
+      # split 2 * 4  ways.
       output = model(inputs)
       return output
 
@@ -1571,10 +1571,10 @@ class Strategy(StrategyBase):
     topology = tf.tpu.experimental.initialize_tpu_system(resolver)
     device_assignment = tf.tpu.experimental.DeviceAssignment.build(
         topology,
-        computation_shape=[1, 1, 2],
+        computation_shape=[1, 1, 1, 2],
         num_replicas=4)
-    strategy = tf.distribute.experimental.TPUStrategy(
-        resolver, device_assignment=device_assignment)
+    strategy = tf.distribute.TPUStrategy(
+        resolver, experimental_device_assignment=device_assignment)
 
     iterator = iter(inputs)
 
@@ -1584,12 +1584,12 @@ class Strategy(StrategyBase):
       images = strategy.experimental_split_to_logical_devices(
         inputs, [1, 2, 4, 1])
 
-      // model() function will be executed on 8 logical devices with `inputs`
-      // split 2 * 4  ways.
+      # model() function will be executed on 8 logical devices with `inputs`
+      # split 2 * 4  ways.
       output = model(inputs)
 
-      // For loss calculation, all logical devices share the same logits
-      // and labels.
+      # For loss calculation, all logical devices share the same logits
+      # and labels.
       labels = strategy.experimental_replicate_to_logical_devices(labels)
       output = strategy.experimental_replicate_to_logical_devices(output)
       loss = loss_fn(labels, output)
diff --git a/tensorflow/python/distribute/mirrored_strategy.py b/tensorflow/python/distribute/mirrored_strategy.py
index 36598634fac..e01b5e6792d 100644
--- a/tensorflow/python/distribute/mirrored_strategy.py
+++ b/tensorflow/python/distribute/mirrored_strategy.py
@@ -190,9 +190,8 @@ class MirroredStrategy(distribute_lib.Strategy):
 
   This strategy is typically used for training on one
   machine with multiple GPUs. For TPUs, use
-  `tf.distribute.experimental.TPUStrategy`. To use `MirroredStrategy` with
-  multiple workers, please refer to
-  `tf.distribute.experimental.MultiWorkerMirroredStrategy`.
+  `tf.distribute.TPUStrategy`. To use `MirroredStrategy` with multiple workers,
+  please refer to `tf.distribute.experimental.MultiWorkerMirroredStrategy`.
 
   For example, a variable created under a `MirroredStrategy` is a
   `MirroredVariable`. If no devices are specified in the constructor argument of
diff --git a/tensorflow/python/distribute/tpu_strategy.py b/tensorflow/python/distribute/tpu_strategy.py
index e2e75260593..dcd1671841f 100644
--- a/tensorflow/python/distribute/tpu_strategy.py
+++ b/tensorflow/python/distribute/tpu_strategy.py
@@ -23,6 +23,7 @@ import collections
 import contextlib
 import copy
 import weakref
+from absl import logging
 
 import numpy as np
 
@@ -57,6 +58,7 @@ from tensorflow.python.tpu import tpu
 from tensorflow.python.tpu import tpu_strategy_util
 from tensorflow.python.tpu import training_loop
 from tensorflow.python.tpu.ops import tpu_ops
+from tensorflow.python.util import deprecation
 from tensorflow.python.util import nest
 from tensorflow.python.util.tf_export import tf_export
 
@@ -97,9 +99,188 @@ def validate_run_function(fn):
         "eager behavior is enabled.")
 
 
+@tf_export("distribute.TPUStrategy", v1=[])
+class TPUStrategyV2(distribute_lib.Strategy):
+  """Synchronous training on TPUs and TPU Pods.
+
+  To construct a TPUStrategy object, you need to run the
+  initialization code as below:
+
+  >>> resolver = tf.distribute.cluster_resolver.TPUClusterResolver(tpu='')
+  >>> tf.config.experimental_connect_to_cluster(resolver)
+  >>> tf.tpu.experimental.initialize_tpu_system(resolver)
+  >>> strategy = tf.distribute.TPUStrategy(resolver)
+
+  While using distribution strategies, the variables created within the
+  strategy's scope will be replicated across all the replicas and can be kept in
+  sync using all-reduce algorithms.
+
+  To run TF2 programs on TPUs, you can either use `.compile` and
+  `.fit` APIs in `tf.keras` with TPUStrategy, or write your own customized
+  training loop by calling `strategy.run` directly. Note that
+  TPUStrategy doesn't support pure eager execution, so please make sure the
+  function passed into `strategy.run` is a `tf.function` or
+  `strategy.run` is called inside a `tf.function` if eager
+  behavior is enabled. See more details in https://www.tensorflow.org/guide/tpu.
+
+  `experimental_distribute_datasets_from_function` and
+  `experimental_distribute_dataset` APIs can be used to distribute the dataset
+  across the TPU workers when writing your own training loop. If you are using
+  `fit` and `compile` methods available in `tf.keras.Model`, then Keras will
+  handle the distribution for you.
+
+  An example of writing customized training loop on TPUs:
+
+  >>> with strategy.scope():
+  ...   model = tf.keras.Sequential([
+  ...     tf.keras.layers.Dense(2, input_shape=(5,)),
+  ...   ])
+  ...   optimizer = tf.keras.optimizers.SGD(learning_rate=0.1)
+
+  >>> def dataset_fn(ctx):
+  ...   x = np.random.random((2, 5)).astype(np.float32)
+  ...   y = np.random.randint(2, size=(2, 1))
+  ...   dataset = tf.data.Dataset.from_tensor_slices((x, y))
+  ...   return dataset.repeat().batch(1, drop_remainder=True)
+  >>> dist_dataset = strategy.experimental_distribute_datasets_from_function(
+  ...     dataset_fn)
+  >>> iterator = iter(dist_dataset)
+
+  >>> @tf.function()
+  ... def train_step(iterator):
+  ...
+  ...   def step_fn(inputs):
+  ...     features, labels = inputs
+  ...     with tf.GradientTape() as tape:
+  ...       logits = model(features, training=True)
+  ...       loss = tf.keras.losses.sparse_categorical_crossentropy(
+  ...           labels, logits)
+  ...
+  ...     grads = tape.gradient(loss, model.trainable_variables)
+  ...     optimizer.apply_gradients(zip(grads, model.trainable_variables))
+  ...
+  ...   strategy.run(step_fn, args=(next(iterator),))
+
+  >>> train_step(iterator)
+
+  For the advanced use cases like model parallelism, you can set
+  `experimental_device_assignment` argument when creating TPUStrategy to specify
+  number of replicas and number of logical devices. Below is an example to
+  initialize TPU system with 2 logical devices and 1 replica.
+
+  >>> resolver = tf.distribute.cluster_resolver.TPUClusterResolver(tpu='')
+  >>> tf.config.experimental_connect_to_cluster(resolver)
+  >>> topology = tf.tpu.experimental.initialize_tpu_system(resolver)
+  >>> device_assignment = tf.tpu.experimental.DeviceAssignment.build(
+  ...     topology,
+  ...     computation_shape=[1, 1, 1, 2],
+  ...     num_replicas=1)
+  >>> strategy = tf.distribute.TPUStrategy(
+  ...     resolver, experimental_device_assignment=device_assignment)
+
+  Then you can run a `tf.add` operation only on logical device 0.
+
+  >>> @tf.function()
+  ... def step_fn(inputs):
+  ...   features, _ = inputs
+  ...   output = tf.add(features, features)
+  ...
+  ...   # Add operation will be executed on logical device 0.
+  ...   output = strategy.experimental_assign_to_logical_device(output, 0)
+  ...   return output
+  >>> dist_dataset = strategy.experimental_distribute_datasets_from_function(
+  ...     dataset_fn)
+  >>> iterator = iter(dist_dataset)
+  >>> strategy.run(step_fn, args=(next(iterator),))
+  """
+
+  def __init__(self,
+               tpu_cluster_resolver=None,
+               experimental_device_assignment=None):
+    """Synchronous training in TPU donuts or Pods.
+
+    Args:
+      tpu_cluster_resolver: A tf.distribute.cluster_resolver.TPUClusterResolver,
+        which provides information about the TPU cluster. If None, it will
+        assume running on a local TPU worker.
+      experimental_device_assignment: Optional
+        `tf.tpu.experimental.DeviceAssignment` to specify the placement of
+        replicas on the TPU cluster.
+    """
+    super(TPUStrategyV2, self).__init__(TPUExtended(
+        self, tpu_cluster_resolver,
+        device_assignment=experimental_device_assignment))
+    distribute_lib.distribution_strategy_gauge.get_cell("V2").set("TPUStrategy")
+    distribute_lib.distribution_strategy_replica_gauge.get_cell(
+        "num_workers").set(self.extended.num_hosts)
+    distribute_lib.distribution_strategy_replica_gauge.get_cell(
+        "num_replicas_per_worker").set(self.extended.num_replicas_per_host)
+    # Packed variable is used to reduce the overhead of function execution.
+    # For a DistributedVariable, only one variable handle is captured into a
+    # function graph. It's only supported in eager mode.
+    self._enable_packed_variable_in_eager_mode = False
+
+  def run(self, fn, args=(), kwargs=None, options=None):
+    """Run the computation defined by `fn` on each TPU replica.
+
+    Executes ops specified by `fn` on each replica. If `args` or `kwargs` have
+    `tf.distribute.DistributedValues`, such as those produced by a
+    `tf.distribute.DistributedDataset` from
+    `tf.distribute.Strategy.experimental_distribute_dataset` or
+    `tf.distribute.Strategy.experimental_distribute_datasets_from_function`,
+    when `fn` is executed on a particular replica, it will be executed with the
+    component of `tf.distribute.DistributedValues` that correspond to that
+    replica.
+
+    `fn` may call `tf.distribute.get_replica_context()` to access members such
+    as `all_reduce`.
+
+    All arguments in `args` or `kwargs` should either be nest of tensors or
+    `tf.distribute.DistributedValues` containing tensors or composite tensors.
+
+    Example usage:
+
+    >>> resolver = tf.distribute.cluster_resolver.TPUClusterResolver(tpu='')
+    >>> tf.config.experimental_connect_to_cluster(resolver)
+    >>> tf.tpu.experimental.initialize_tpu_system(resolver)
+    >>> strategy = tf.distribute.TPUStrategy(resolver)
+    >>> @tf.function
+    ... def run():
+    ...   def value_fn(value_context):
+    ...     return value_context.num_replicas_in_sync
+    ...   distributed_values = (
+    ...       strategy.experimental_distribute_values_from_function(value_fn))
+    ...   def replica_fn(input):
+    ...     return input * 2
+    ...   return strategy.run(replica_fn, args=(distributed_values,))
+    >>> result = run()
+
+    Args:
+      fn: The function to run. The output must be a `tf.nest` of `Tensor`s.
+      args: (Optional) Positional arguments to `fn`.
+      kwargs: (Optional) Keyword arguments to `fn`.
+      options: (Optional) An instance of `tf.distribute.RunOptions` specifying
+        the options to run `fn`.
+
+    Returns:
+      Merged return value of `fn` across replicas. The structure of the return
+      value is the same as the return value from `fn`. Each element in the
+      structure can either be `tf.distribute.DistributedValues`, `Tensor`
+      objects, or `Tensor`s (for example, if running on a single replica).
+    """
+    validate_run_function(fn)
+
+    # Note: the target function is converted to graph even when in Eager mode,
+    # so autograph is on by default here.
+    fn = autograph.tf_convert(fn, autograph_ctx.control_status_ctx())
+    options = options or distribute_lib.RunOptions()
+    return self.extended.tpu_run(fn, args, kwargs, options)
+
+
 @tf_export("distribute.experimental.TPUStrategy", v1=[])
+@deprecation.deprecated_endpoints("distribute.experimental.TPUStrategy")
 class TPUStrategy(distribute_lib.Strategy):
-  """TPU distribution strategy implementation.
+  """Synchronous training on TPUs and TPU Pods.
 
   To construct a TPUStrategy object, you need to run the
   initialization code as below:
@@ -109,9 +290,9 @@ class TPUStrategy(distribute_lib.Strategy):
   >>> tf.tpu.experimental.initialize_tpu_system(resolver)
   >>> strategy = tf.distribute.experimental.TPUStrategy(resolver)
 
-  While using distribution strategies, the variables created within strategy's
-  scope will be replicated across all the replicas and can be kept in sync
-  using all-reduce algorithms.
+  While using distribution strategies, the variables created within the
+  strategy's scope will be replicated across all the replicas and can be kept in
+  sync using all-reduce algorithms.
 
   To run TF2 programs on TPUs, you can either use `.compile` and
   `.fit` APIs in `tf.keras` with TPUStrategy, or write your own customized
@@ -131,9 +312,12 @@ class TPUStrategy(distribute_lib.Strategy):
       tpu_cluster_resolver: A tf.distribute.cluster_resolver.TPUClusterResolver,
         which provides information about the TPU cluster.
       device_assignment: Optional `tf.tpu.experimental.DeviceAssignment` to
-        specify the placement of replicas on the TPU cluster. Currently only
-        supports the usecase of using a single core within a TPU cluster.
+        specify the placement of replicas on the TPU cluster.
     """
+    logging.warning(
+        "`tf.distribute.experimental.TPUStrategy` is deprecated, please use "
+        " the non experimental symbol `tf.distribute.TPUStrategy` instead.")
+
     super(TPUStrategy, self).__init__(TPUExtended(
         self, tpu_cluster_resolver, device_assignment=device_assignment))
     distribute_lib.distribution_strategy_gauge.get_cell("V2").set("TPUStrategy")
diff --git a/tensorflow/python/tpu/tpu_embedding_v2.py b/tensorflow/python/tpu/tpu_embedding_v2.py
index e5cfba7c587..6db67fea367 100644
--- a/tensorflow/python/tpu/tpu_embedding_v2.py
+++ b/tensorflow/python/tpu/tpu_embedding_v2.py
@@ -132,7 +132,7 @@ class TPUEmbedding(tracking.AutoTrackable):
   First lets look at the `TPUStrategy` mode. Initial setup looks like:
 
   ```python
-  strategy = tf.distribute.experimental.TPUStrategy(...)
+  strategy = tf.distribute.TPUStrategy(...)
   with strategy.scope():
     embedding = tf.tpu.experimental.embedding.TPUEmbedding(
         feature_config=feature_config,
@@ -234,7 +234,7 @@ class TPUEmbedding(tracking.AutoTrackable):
     """Creates the TPUEmbedding mid level API object.
 
     ```python
-    strategy = tf.distribute.experimental.TPUStrategy(...)
+    strategy = tf.distribute.TPUStrategy(...)
     with strategy.scope():
       embedding = tf.tpu.experimental.embedding.TPUEmbedding(
           feature_config=tf.tpu.experimental.embedding.FeatureConfig(
@@ -512,7 +512,7 @@ class TPUEmbedding(tracking.AutoTrackable):
     ensure you understand the effect of applying a zero gradient.
 
     ```python
-    strategy = tf.distribute.experimental.TPUStrategy(...)
+    strategy = tf.distribute.TPUStrategy(...)
     with strategy.scope():
       embedding = tf.tpu.experimental.embedding.TPUEmbedding(...)
 
@@ -603,7 +603,7 @@ class TPUEmbedding(tracking.AutoTrackable):
     `(batch_size, max_sequence_length, dim)` instead.
 
     ```python
-    strategy = tf.distribute.experimental.TPUStrategy(...)
+    strategy = tf.distribute.TPUStrategy(...)
     with strategy.scope():
       embedding = tf.tpu.experimental.embedding.TPUEmbedding(...)
 
@@ -1054,13 +1054,13 @@ class TPUEmbedding(tracking.AutoTrackable):
     embedding tables. We expect that the batch size of each of the tensors in
     features matches the per core batch size. This will automatically happen if
     your input dataset is batched to the global batch size and you use
-    `tf.distribute.experimental.TPUStrategy`'s `experimental_distribute_dataset`
+    `tf.distribute.TPUStrategy`'s `experimental_distribute_dataset`
     or if you use `experimental_distribute_datasets_from_function` and batch
     to the per core batch size computed by the context passed to your input
     function.
 
     ```python
-    strategy = tf.distribute.experimental.TPUStrategy(...)
+    strategy = tf.distribute.TPUStrategy(...)
     with strategy.scope():
       embedding = tf.tpu.experimental.embedding.TPUEmbedding(...)
 
diff --git a/tensorflow/tools/api/golden/v2/tensorflow.distribute.-t-p-u-strategy.pbtxt b/tensorflow/tools/api/golden/v2/tensorflow.distribute.-t-p-u-strategy.pbtxt
new file mode 100644
index 00000000000..f41a08454e2
--- /dev/null
+++ b/tensorflow/tools/api/golden/v2/tensorflow.distribute.-t-p-u-strategy.pbtxt
@@ -0,0 +1,99 @@
+path: "tensorflow.distribute.TPUStrategy"
+tf_class {
+  is_instance: ""
+  is_instance: ""
+  is_instance: ""
+  is_instance: ""
+  member {
+    name: "extended"
+    mtype: ""
+  }
+  member {
+    name: "num_replicas_in_sync"
+    mtype: ""
+  }
+  member_method {
+    name: "__init__"
+    argspec: "args=[\'self\', \'tpu_cluster_resolver\', \'experimental_device_assignment\'], varargs=None, keywords=None, defaults=[\'None\', \'None\'], "
+  }
+  member_method {
+    name: "colocate_vars_with"
+    argspec: "args=[\'self\', \'colocate_with_variable\'], varargs=None, keywords=None, defaults=None"
+  }
+  member_method {
+    name: "configure"
+    argspec: "args=[\'self\', \'session_config\', \'cluster_spec\', \'task_type\', \'task_id\'], varargs=None, keywords=None, defaults=[\'None\', \'None\', \'None\', \'None\'], "
+  }
+  member_method {
+    name: "experimental_assign_to_logical_device"
+    argspec: "args=[\'self\', \'tensor\', \'logical_device_id\'], varargs=None, keywords=None, defaults=None"
+  }
+  member_method {
+    name: "experimental_distribute_dataset"
+    argspec: "args=[\'self\', \'dataset\', \'options\'], varargs=None, keywords=None, defaults=[\'None\'], "
+  }
+  member_method {
+    name: "experimental_distribute_datasets_from_function"
+    argspec: "args=[\'self\', \'dataset_fn\', \'options\'], varargs=None, keywords=None, defaults=[\'None\'], "
+  }
+  member_method {
+    name: "experimental_distribute_values_from_function"
+    argspec: "args=[\'self\', \'value_fn\'], varargs=None, keywords=None, defaults=None"
+  }
+  member_method {
+    name: "experimental_local_results"
+    argspec: "args=[\'self\', \'value\'], varargs=None, keywords=None, defaults=None"
+  }
+  member_method {
+    name: "experimental_make_numpy_dataset"
+    argspec: "args=[\'self\', \'numpy_input\'], varargs=None, keywords=None, defaults=None"
+  }
+  member_method {
+    name: "experimental_replicate_to_logical_devices"
+    argspec: "args=[\'self\', \'tensor\'], varargs=None, keywords=None, defaults=None"
+  }
+  member_method {
+    name: "experimental_run"
+    argspec: "args=[\'self\', \'fn\', \'input_iterator\'], varargs=None, keywords=None, defaults=[\'None\'], "
+  }
+  member_method {
+    name: "experimental_run_v2"
+    argspec: "args=[\'self\', \'fn\', \'args\', \'kwargs\', \'options\'], varargs=None, keywords=None, defaults=[\'()\', \'None\', \'None\'], "
+  }
+  member_method {
+    name: "experimental_split_to_logical_devices"
+    argspec: "args=[\'self\', \'tensor\', \'partition_dimensions\'], varargs=None, keywords=None, defaults=None"
+  }
+  member_method {
+    name: "group"
+    argspec: "args=[\'self\', \'value\', \'name\'], varargs=None, keywords=None, defaults=[\'None\'], "
+  }
+  member_method {
+    name: "make_dataset_iterator"
+    argspec: "args=[\'self\', \'dataset\'], varargs=None, keywords=None, defaults=None"
+  }
+  member_method {
+    name: "make_input_fn_iterator"
+    argspec: "args=[\'self\', \'input_fn\', \'replication_mode\'], varargs=None, keywords=None, defaults=[\'InputReplicationMode.PER_WORKER\'], "
+  }
+  member_method {
+    name: "reduce"
+    argspec: "args=[\'self\', \'reduce_op\', \'value\', \'axis\'], varargs=None, keywords=None, defaults=None"
+  }
+  member_method {
+    name: "run"
+    argspec: "args=[\'self\', \'fn\', \'args\', \'kwargs\', \'options\'], varargs=None, keywords=None, defaults=[\'()\', \'None\', \'None\'], "
+  }
+  member_method {
+    name: "scope"
+    argspec: "args=[\'self\'], varargs=None, keywords=None, defaults=None"
+  }
+  member_method {
+    name: "unwrap"
+    argspec: "args=[\'self\', \'value\'], varargs=None, keywords=None, defaults=None"
+  }
+  member_method {
+    name: "update_config_proto"
+    argspec: "args=[\'self\', \'config_proto\'], varargs=None, keywords=None, defaults=None"
+  }
+}
diff --git a/tensorflow/tools/api/golden/v2/tensorflow.distribute.pbtxt b/tensorflow/tools/api/golden/v2/tensorflow.distribute.pbtxt
index 009cb7fe400..d3867889a4f 100644
--- a/tensorflow/tools/api/golden/v2/tensorflow.distribute.pbtxt
+++ b/tensorflow/tools/api/golden/v2/tensorflow.distribute.pbtxt
@@ -72,6 +72,10 @@ tf_module {
     name: "StrategyExtended"
     mtype: ""
   }
+  member {
+    name: "TPUStrategy"
+    mtype: ""
+  }
   member {
     name: "cluster_resolver"
     mtype: ""
diff --git a/tensorflow/tools/compatibility/tf_upgrade_v2.py b/tensorflow/tools/compatibility/tf_upgrade_v2.py
index d27c75fb44e..1475a74d62f 100644
--- a/tensorflow/tools/compatibility/tf_upgrade_v2.py
+++ b/tensorflow/tools/compatibility/tf_upgrade_v2.py
@@ -883,7 +883,7 @@ class TFAPIChangeSpec(ast_edits.NoUpdateSpec):
     contrib_tpu_strategy_warning = (
         ast_edits.ERROR,
         "(Manual edit required) tf.contrib.distribute.TPUStrategy has "
-        "been migrated to tf.distribute.experimental.TPUStrategy. Note the "
+        "been migrated to tf.distribute.TPUStrategy. Note the "
         "slight changes in constructor. " + distribute_strategy_api_changes)
 
     contrib_collective_strategy_warning = (
diff --git a/tensorflow/tools/compatibility/tf_upgrade_v2_test.py b/tensorflow/tools/compatibility/tf_upgrade_v2_test.py
index 47b9899a6b7..185a3b07f8d 100644
--- a/tensorflow/tools/compatibility/tf_upgrade_v2_test.py
+++ b/tensorflow/tools/compatibility/tf_upgrade_v2_test.py
@@ -2155,7 +2155,7 @@ def _log_prob(self, x):
     expected = "tf.contrib.distribute.TPUStrategy"
     _, _, errors, new_text = self._upgrade(text)
     self.assertEqual(expected, new_text)
-    self.assertIn("migrated to tf.distribute.experimental.TPUStrategy",
+    self.assertIn("migrated to tf.distribute.TPUStrategy",
                   errors[0])
 
     text = "tf.contrib.distribute.foo"

From 059c05e4da07225faa651c46df9b9be6f83e6914 Mon Sep 17 00:00:00 2001
From: Gaurav Jain 
Date: Sat, 20 Jun 2020 14:12:04 -0700
Subject: [PATCH 392/867] Handle DT_UINT32 & DT_UINT64 in transpose

PiperOrigin-RevId: 317485472
Change-Id: I5c62d7977f44f57f544ebe27e82f141fe07ce73e
---
 tensorflow/core/kernels/transpose_functor.h | 2 ++
 1 file changed, 2 insertions(+)

diff --git a/tensorflow/core/kernels/transpose_functor.h b/tensorflow/core/kernels/transpose_functor.h
index a89fc40d772..0c22b11b7c6 100644
--- a/tensorflow/core/kernels/transpose_functor.h
+++ b/tensorflow/core/kernels/transpose_functor.h
@@ -191,11 +191,13 @@ Status DoTransposeImpl(const Device& d, const Tensor& in,
     case DT_FLOAT:
     case DT_INT32:
     case DT_QINT32:
+    case DT_UINT32:
       Transpose::run(d, in, perm, out);
       break;
 
     case DT_DOUBLE:
     case DT_INT64:
+    case DT_UINT64:
       Transpose::run(d, in, perm, out);
       break;
 

From c654edb1f30c7fb4e4ae60fb20b8a5661fee0ed1 Mon Sep 17 00:00:00 2001
From: "A. Unique TensorFlower" 
Date: Sat, 20 Jun 2020 15:08:56 -0700
Subject: [PATCH 393/867] Update Eigen to
 https://gitlab.com/libeigen/eigen/-/commit/6b9c92fe7eff0dedb031cec38004c9c3667f3057

PiperOrigin-RevId: 317488881
Change-Id: Icf75e932f8568a2135234de236fd0f94869ad05a
---
 tensorflow/workspace.bzl | 8 ++++----
 1 file changed, 4 insertions(+), 4 deletions(-)

diff --git a/tensorflow/workspace.bzl b/tensorflow/workspace.bzl
index 27eca0ee54f..e6a15b422eb 100755
--- a/tensorflow/workspace.bzl
+++ b/tensorflow/workspace.bzl
@@ -237,11 +237,11 @@ def tf_repositories(path_prefix = "", tf_repo_name = ""):
         name = "eigen_archive",
         build_file = clean_dep("//third_party:eigen.BUILD"),
         patch_file = clean_dep("//third_party/eigen3:gpu_packet_math.patch"),
-        sha256 = "a6b790c40e637cddc2554a7bed40622d30311210c8ce94524cdf5496417fd2bc",  # SHARED_EIGEN_SHA
-        strip_prefix = "eigen-6228f27234ca84545e711fb27b7850f6829af3d9",
+        sha256 = "d26ada177ed9b696a9447fc85d209932a032c8ffc51630cf15eea8629b29dad6",  # SHARED_EIGEN_SHA
+        strip_prefix = "eigen-6b9c92fe7eff0dedb031cec38004c9c3667f3057",
         urls = [
-            "https://storage.googleapis.com/mirror.tensorflow.org/gitlab.com/libeigen/eigen/-/archive/6228f27234ca84545e711fb27b7850f6829af3d9/eigen-6228f27234ca84545e711fb27b7850f6829af3d9.tar.gz",
-            "https://gitlab.com/libeigen/eigen/-/archive/6228f27234ca84545e711fb27b7850f6829af3d9/eigen-6228f27234ca84545e711fb27b7850f6829af3d9.tar.gz",
+            "https://storage.googleapis.com/mirror.tensorflow.org/gitlab.com/libeigen/eigen/-/archive/6b9c92fe7eff0dedb031cec38004c9c3667f3057/eigen-6b9c92fe7eff0dedb031cec38004c9c3667f3057.tar.gz",
+            "https://gitlab.com/libeigen/eigen/-/archive/6b9c92fe7eff0dedb031cec38004c9c3667f3057/eigen-6b9c92fe7eff0dedb031cec38004c9c3667f3057.tar.gz",
         ],
     )
 

From ae423bd3bf2eb3d41e37a358429e7436e1f0b56a Mon Sep 17 00:00:00 2001
From: Gaurav Jain 
Date: Sat, 20 Jun 2020 15:09:35 -0700
Subject: [PATCH 394/867] Rollback: Add uint32 & uint64 to
 TF_CALL_INTEGRAL_TYPES

PiperOrigin-RevId: 317488920
Change-Id: I65736e7f4a1004ff634194343dc4ec237a227a19
---
 tensorflow/core/framework/register_types.h    | 21 +++++++++-----
 tensorflow/core/framework/types.cc            |  5 ++++
 tensorflow/core/kernels/BUILD                 |  2 --
 tensorflow/core/kernels/concat_lib_cpu.cc     |  2 ++
 tensorflow/core/kernels/concat_op.cc          |  2 ++
 tensorflow/core/kernels/constant_op.cc        |  1 +
 tensorflow/core/kernels/control_flow_ops.cc   |  5 ++++
 .../core/kernels/data/dataset_test_base.cc    |  2 ++
 tensorflow/core/kernels/dense_update_ops.cc   |  1 +
 .../core/kernels/dynamic_partition_op.cc      |  2 ++
 tensorflow/core/kernels/fill_functor.cc       |  5 +---
 tensorflow/core/kernels/gather_op.cc          |  2 ++
 tensorflow/core/kernels/identity_op.cc        |  1 +
 tensorflow/core/kernels/ragged_gather_op.cc   |  2 ++
 .../kernels/ragged_tensor_from_variant_op.cc  |  2 ++
 .../kernels/ragged_tensor_to_tensor_op.cc     |  2 ++
 .../kernels/ragged_tensor_to_variant_op.cc    |  2 ++
 .../core/kernels/resource_variable_ops.cc     |  1 +
 tensorflow/core/kernels/split_lib_cpu.cc      |  1 +
 tensorflow/core/kernels/split_op.cc           |  1 +
 tensorflow/core/kernels/strided_slice_op.cc   |  2 ++
 .../core/kernels/strided_slice_op_impl.h      |  2 ++
 tensorflow/core/kernels/topk_op.cc            |  2 ++
 .../core/kernels/topk_op_gpu_uint32.cu.cc     | 28 -------------------
 .../core/kernels/topk_op_gpu_uint64.cu.cc     | 28 -------------------
 tensorflow/core/util/batch_util.cc            |  8 ++++++
 .../core/util/saved_tensor_slice_util.h       |  2 --
 27 files changed, 63 insertions(+), 71 deletions(-)
 delete mode 100644 tensorflow/core/kernels/topk_op_gpu_uint32.cu.cc
 delete mode 100644 tensorflow/core/kernels/topk_op_gpu_uint64.cu.cc

diff --git a/tensorflow/core/framework/register_types.h b/tensorflow/core/framework/register_types.h
index 0cf6536e8c2..bc3e5e1743b 100644
--- a/tensorflow/core/framework/register_types.h
+++ b/tensorflow/core/framework/register_types.h
@@ -153,9 +153,16 @@ limitations under the License.
 #endif  // defined(IS_MOBILE_PLATFORM)  - end of TF_CALL_type defines
 
 // Defines for sets of types.
-#define TF_CALL_INTEGRAL_TYPES(m)                                       \
-  TF_CALL_uint64(m) TF_CALL_int64(m) TF_CALL_uint32(m) TF_CALL_int32(m) \
-      TF_CALL_uint16(m) TF_CALL_int16(m) TF_CALL_uint8(m) TF_CALL_int8(m)
+
+// TODO(b/111604096): Add uint32 and uint64 to TF_CALL_INTEGRAL_TYPES.
+//
+// The uint32 and uint64 types were introduced in 10/2017 to be used via XLA and
+// thus were not included in TF_CALL_INTEGRAL_TYPES. Including them in
+// TF_CALL_INTEGRAL_TYPES should only happen after evaluating the effect on the
+// TF binary size and performance.
+#define TF_CALL_INTEGRAL_TYPES(m)                                      \
+  TF_CALL_int64(m) TF_CALL_int32(m) TF_CALL_uint16(m) TF_CALL_int16(m) \
+      TF_CALL_uint8(m) TF_CALL_int8(m)
 
 #define TF_CALL_FLOAT_TYPES(m) \
   TF_CALL_half(m) TF_CALL_bfloat16(m) TF_CALL_float(m) TF_CALL_double(m)
@@ -167,10 +174,10 @@ limitations under the License.
 #define TF_CALL_REAL_NUMBER_TYPES_NO_BFLOAT16(m) \
   TF_CALL_INTEGRAL_TYPES(m) TF_CALL_half(m) TF_CALL_float(m) TF_CALL_double(m)
 
-#define TF_CALL_REAL_NUMBER_TYPES_NO_INT32(m)                                \
-  TF_CALL_half(m) TF_CALL_bfloat16(m) TF_CALL_float(m) TF_CALL_double(m)     \
-      TF_CALL_uint64(m) TF_CALL_int64(m) TF_CALL_uint32(m) TF_CALL_uint16(m) \
-          TF_CALL_int16(m) TF_CALL_uint8(m) TF_CALL_int8(m)
+#define TF_CALL_REAL_NUMBER_TYPES_NO_INT32(m)                              \
+  TF_CALL_half(m) TF_CALL_bfloat16(m) TF_CALL_float(m) TF_CALL_double(m)   \
+      TF_CALL_int64(m) TF_CALL_uint16(m) TF_CALL_int16(m) TF_CALL_uint8(m) \
+          TF_CALL_int8(m)
 
 #define TF_CALL_COMPLEX_TYPES(m) TF_CALL_complex64(m) TF_CALL_complex128(m)
 
diff --git a/tensorflow/core/framework/types.cc b/tensorflow/core/framework/types.cc
index d6455e012d0..97eaec98ffe 100644
--- a/tensorflow/core/framework/types.cc
+++ b/tensorflow/core/framework/types.cc
@@ -238,6 +238,11 @@ int DataTypeSize(DataType dt) {
     TF_CALL_qint16(CASE);
     TF_CALL_quint16(CASE);
 
+    // uint32 and uint64 aren't included in TF_CALL_POD_TYPES because we
+    // don't want to define kernels for them at this stage to avoid binary
+    // bloat.
+    TF_CALL_uint32(CASE);
+    TF_CALL_uint64(CASE);
     default:
       return 0;
   }
diff --git a/tensorflow/core/kernels/BUILD b/tensorflow/core/kernels/BUILD
index e2ff5aed283..7da864a6027 100644
--- a/tensorflow/core/kernels/BUILD
+++ b/tensorflow/core/kernels/BUILD
@@ -4919,9 +4919,7 @@ tf_kernel_library(
         "topk_op_gpu_double.cu.cc",
         "topk_op_gpu_float.cu.cc",
         "topk_op_gpu_half.cu.cc",
-        "topk_op_gpu_uint64.cu.cc",
         "topk_op_gpu_int64.cu.cc",
-        "topk_op_gpu_uint32.cu.cc",
         "topk_op_gpu_int32.cu.cc",
         "topk_op_gpu_int16.cu.cc",
         "topk_op_gpu_uint16.cu.cc",
diff --git a/tensorflow/core/kernels/concat_lib_cpu.cc b/tensorflow/core/kernels/concat_lib_cpu.cc
index 1dec589d3ff..da73d3d2c56 100644
--- a/tensorflow/core/kernels/concat_lib_cpu.cc
+++ b/tensorflow/core/kernels/concat_lib_cpu.cc
@@ -116,6 +116,8 @@ REGISTER(qint8)
 REGISTER(quint16)
 REGISTER(qint16)
 REGISTER(qint32)
+REGISTER(uint32)
+REGISTER(uint64)
 
 #if defined(IS_MOBILE_PLATFORM) && !defined(SUPPORT_SELECTIVE_REGISTRATION) && \
     !defined(__ANDROID_TYPES_FULL__)
diff --git a/tensorflow/core/kernels/concat_op.cc b/tensorflow/core/kernels/concat_op.cc
index d3f3a04f33b..be3e9a67c5f 100644
--- a/tensorflow/core/kernels/concat_op.cc
+++ b/tensorflow/core/kernels/concat_op.cc
@@ -208,6 +208,8 @@ REGISTER_CONCAT(qint8);
 REGISTER_CONCAT(quint16);
 REGISTER_CONCAT(qint16);
 REGISTER_CONCAT(qint32);
+REGISTER_CONCAT(uint32);
+REGISTER_CONCAT(uint64);
 
 #undef REGISTER_CONCAT
 
diff --git a/tensorflow/core/kernels/constant_op.cc b/tensorflow/core/kernels/constant_op.cc
index dc178d17d49..4bcbc076446 100644
--- a/tensorflow/core/kernels/constant_op.cc
+++ b/tensorflow/core/kernels/constant_op.cc
@@ -211,6 +211,7 @@ TF_CALL_ALL_TYPES(REGISTER_CPU_KERNEL);
 // the conversion from uint8 to quint8.
 REGISTER_KERNEL(CPU, quint8);
 REGISTER_KERNEL(CPU, quint16);
+REGISTER_KERNEL(CPU, uint32);
 #undef REGISTER_CPU_KERNEL
 
 #ifdef TENSORFLOW_USE_SYCL
diff --git a/tensorflow/core/kernels/control_flow_ops.cc b/tensorflow/core/kernels/control_flow_ops.cc
index 1a0082c6a3b..435de3c5954 100644
--- a/tensorflow/core/kernels/control_flow_ops.cc
+++ b/tensorflow/core/kernels/control_flow_ops.cc
@@ -101,12 +101,16 @@ TF_CALL_ALL_TYPES(REGISTER_CPU_SWITCH);
 TF_CALL_ALL_TYPES(REGISTER_CPU_REF_SWITCH);
 TF_CALL_QUANTIZED_TYPES(REGISTER_CPU_SWITCH);
 TF_CALL_QUANTIZED_TYPES(REGISTER_CPU_REF_SWITCH);
+REGISTER_CPU_SWITCH(uint64);
 
 TF_CALL_NUMBER_TYPES_NO_INT32(REGISTER_GPU_SWITCH);
 TF_CALL_QUANTIZED_TYPES(REGISTER_GPU_SWITCH);
 TF_CALL_NUMBER_TYPES_NO_INT32(REGISTER_GPU_REF_SWITCH);
 TF_CALL_QUANTIZED_TYPES(REGISTER_GPU_REF_SWITCH);
+REGISTER_GPU_SWITCH(uint64);
 TF_CALL_variant(REGISTER_GPU_SWITCH);
+TF_CALL_uint32(REGISTER_GPU_SWITCH);
+TF_CALL_uint32(REGISTER_GPU_REF_SWITCH);
 TF_CALL_bool(REGISTER_GPU_SWITCH);
 TF_CALL_bool(REGISTER_GPU_REF_SWITCH);
 
@@ -307,6 +311,7 @@ TF_CALL_QUANTIZED_TYPES(REGISTER_GPU_KERNEL);
 TF_CALL_QUANTIZED_TYPES(REGISTER_GPU_REF_KERNEL);
 REGISTER_GPU_KERNEL(bool);
 REGISTER_GPU_REF_KERNEL(bool);
+REGISTER_GPU_KERNEL(uint64);
 TF_CALL_variant(REGISTER_GPU_KERNEL);
 
 #undef REGISTER_GPU_KERNEL
diff --git a/tensorflow/core/kernels/data/dataset_test_base.cc b/tensorflow/core/kernels/data/dataset_test_base.cc
index e41e35be1e9..b91ab9b733c 100644
--- a/tensorflow/core/kernels/data/dataset_test_base.cc
+++ b/tensorflow/core/kernels/data/dataset_test_base.cc
@@ -220,6 +220,8 @@ Status DatasetOpsTestBase::ExpectEqual(const Tensor& a, const Tensor& b) {
     break;
     TF_CALL_NUMBER_TYPES(CASE);
     TF_CALL_tstring(CASE);
+    TF_CALL_uint32(CASE);
+    TF_CALL_uint64(CASE);
     // TODO(feihugis): figure out how to support variant tensors.
 #undef CASE
     default:
diff --git a/tensorflow/core/kernels/dense_update_ops.cc b/tensorflow/core/kernels/dense_update_ops.cc
index 71235fca143..55e4cd7606a 100644
--- a/tensorflow/core/kernels/dense_update_ops.cc
+++ b/tensorflow/core/kernels/dense_update_ops.cc
@@ -98,6 +98,7 @@ typedef Eigen::SyclDevice SYCLDevice;
 
 TF_CALL_ALL_TYPES(REGISTER_KERNELS);
 // uint32 not included in ALL_TYPES
+TF_CALL_uint32(REGISTER_KERNELS);
 TF_CALL_QUANTIZED_TYPES(REGISTER_KERNELS);
 // quint16 not included in QUANTIZIED_TYPES
 TF_CALL_quint16(REGISTER_KERNELS);
diff --git a/tensorflow/core/kernels/dynamic_partition_op.cc b/tensorflow/core/kernels/dynamic_partition_op.cc
index 95af19c4c48..90ed71dccce 100644
--- a/tensorflow/core/kernels/dynamic_partition_op.cc
+++ b/tensorflow/core/kernels/dynamic_partition_op.cc
@@ -164,6 +164,8 @@ class DynamicPartitionOp : public DynamicPartitionOp_Shared {
       DynamicPartitionOp)
 
 TF_CALL_ALL_TYPES(REGISTER_DYNAMIC_PARTITION);
+// For partitioning fingerprints.
+TF_CALL_uint64(REGISTER_DYNAMIC_PARTITION);
 #undef REGISTER_DYNAMIC_PARTITION
 
 }  // namespace tensorflow
diff --git a/tensorflow/core/kernels/fill_functor.cc b/tensorflow/core/kernels/fill_functor.cc
index 174a4e45a79..10dd3df1915 100644
--- a/tensorflow/core/kernels/fill_functor.cc
+++ b/tensorflow/core/kernels/fill_functor.cc
@@ -45,8 +45,6 @@ DEFINE_SETZERO_CPU(Eigen::half);
 DEFINE_SETZERO_CPU(bfloat16);
 DEFINE_SETZERO_CPU(float);
 DEFINE_SETZERO_CPU(double);
-DEFINE_SETZERO_CPU(uint32);
-DEFINE_SETZERO_CPU(uint64);
 DEFINE_SETZERO_CPU(uint8);
 DEFINE_SETZERO_CPU(int8);
 DEFINE_SETZERO_CPU(uint16);
@@ -98,8 +96,6 @@ DEFINE_SETONE_CPU(Eigen::half);
 DEFINE_SETONE_CPU(bfloat16);
 DEFINE_SETONE_CPU(float);
 DEFINE_SETONE_CPU(double);
-DEFINE_SETONE_CPU(uint32);
-DEFINE_SETONE_CPU(uint64);
 DEFINE_SETONE_CPU(uint8);
 DEFINE_SETONE_CPU(int8);
 DEFINE_SETONE_CPU(uint16);
@@ -141,6 +137,7 @@ struct FillFunctor {
 TF_CALL_ALL_TYPES(DEFINE_FILL_CPU);
 DEFINE_FILL_CPU(quint8);
 DEFINE_FILL_CPU(quint16);
+DEFINE_FILL_CPU(uint32);
 #undef DEFINE_FILL_CPU
 
 #ifdef TENSORFLOW_USE_SYCL
diff --git a/tensorflow/core/kernels/gather_op.cc b/tensorflow/core/kernels/gather_op.cc
index 948567e019a..6d493a5f2ea 100644
--- a/tensorflow/core/kernels/gather_op.cc
+++ b/tensorflow/core/kernels/gather_op.cc
@@ -211,6 +211,8 @@ TF_CALL_ALL_TYPES(REGISTER_GATHER_CPU);
 TF_CALL_QUANTIZED_TYPES(REGISTER_GATHER_CPU);
 TF_CALL_quint16(REGISTER_GATHER_CPU);
 TF_CALL_qint16(REGISTER_GATHER_CPU);
+TF_CALL_uint32(REGISTER_GATHER_CPU);
+TF_CALL_uint64(REGISTER_GATHER_CPU);
 
 #undef REGISTER_GATHER_CPU
 
diff --git a/tensorflow/core/kernels/identity_op.cc b/tensorflow/core/kernels/identity_op.cc
index 4b226dd72d4..d15b64597f5 100644
--- a/tensorflow/core/kernels/identity_op.cc
+++ b/tensorflow/core/kernels/identity_op.cc
@@ -122,6 +122,7 @@ REGISTER_SYCL_HOST_KERNEL(bool);
 
 TF_CALL_NUMBER_TYPES_NO_INT32(REGISTER_GPU_KERNEL);
 REGISTER_GPU_KERNEL(Variant);
+TF_CALL_uint32(REGISTER_GPU_KERNEL);
 REGISTER_GPU_KERNEL(bool);
 
 #undef REGISTER_GPU_KERNEL
diff --git a/tensorflow/core/kernels/ragged_gather_op.cc b/tensorflow/core/kernels/ragged_gather_op.cc
index 3bf82cba050..88c0d1ebd69 100644
--- a/tensorflow/core/kernels/ragged_gather_op.cc
+++ b/tensorflow/core/kernels/ragged_gather_op.cc
@@ -296,6 +296,8 @@ TF_CALL_tstring(REGISTER_CPU_KERNEL);
 TF_CALL_QUANTIZED_TYPES(REGISTER_CPU_KERNEL);
 TF_CALL_quint16(REGISTER_CPU_KERNEL);
 TF_CALL_qint16(REGISTER_CPU_KERNEL);
+TF_CALL_uint32(REGISTER_CPU_KERNEL);
+TF_CALL_uint64(REGISTER_CPU_KERNEL);
 #undef REGISTER_CPU_KERNEL
 #undef REGISTER_CPU_KERNEL_WITH_INDEX_TYPE
 
diff --git a/tensorflow/core/kernels/ragged_tensor_from_variant_op.cc b/tensorflow/core/kernels/ragged_tensor_from_variant_op.cc
index ad0712e6fd0..f83bcb38c6c 100644
--- a/tensorflow/core/kernels/ragged_tensor_from_variant_op.cc
+++ b/tensorflow/core/kernels/ragged_tensor_from_variant_op.cc
@@ -308,6 +308,8 @@ TF_CALL_tstring(REGISTER_KERNELS);
 TF_CALL_QUANTIZED_TYPES(REGISTER_KERNELS);
 TF_CALL_quint16(REGISTER_KERNELS);
 TF_CALL_qint16(REGISTER_KERNELS);
+TF_CALL_uint32(REGISTER_KERNELS);
+TF_CALL_uint64(REGISTER_KERNELS);
 #undef REGISTER_KERNELS
 #undef REGISTER_KERNELS_WITH_SPLIT_TYPE
 }  // namespace tensorflow
diff --git a/tensorflow/core/kernels/ragged_tensor_to_tensor_op.cc b/tensorflow/core/kernels/ragged_tensor_to_tensor_op.cc
index 9ae5d7ffbdc..d729c43f25a 100644
--- a/tensorflow/core/kernels/ragged_tensor_to_tensor_op.cc
+++ b/tensorflow/core/kernels/ragged_tensor_to_tensor_op.cc
@@ -561,6 +561,8 @@ TF_CALL_string(REGISTER_CPU_KERNEL);
 TF_CALL_QUANTIZED_TYPES(REGISTER_CPU_KERNEL);
 TF_CALL_quint16(REGISTER_CPU_KERNEL);
 TF_CALL_qint16(REGISTER_CPU_KERNEL);
+TF_CALL_uint32(REGISTER_CPU_KERNEL);
+TF_CALL_uint64(REGISTER_CPU_KERNEL);
 
 #undef REGISTER_CPU_KERNEL
 
diff --git a/tensorflow/core/kernels/ragged_tensor_to_variant_op.cc b/tensorflow/core/kernels/ragged_tensor_to_variant_op.cc
index 64c372b005e..7a5ae1c6240 100644
--- a/tensorflow/core/kernels/ragged_tensor_to_variant_op.cc
+++ b/tensorflow/core/kernels/ragged_tensor_to_variant_op.cc
@@ -213,6 +213,8 @@ TF_CALL_tstring(REGISTER_KERNELS);
 TF_CALL_QUANTIZED_TYPES(REGISTER_KERNELS);
 TF_CALL_quint16(REGISTER_KERNELS);
 TF_CALL_qint16(REGISTER_KERNELS);
+TF_CALL_uint32(REGISTER_KERNELS);
+TF_CALL_uint64(REGISTER_KERNELS);
 #undef REGISTER_KERNELS
 #undef REGISTER_KERNELS_WITH_SPLIT_TYPE
 }  // namespace tensorflow
diff --git a/tensorflow/core/kernels/resource_variable_ops.cc b/tensorflow/core/kernels/resource_variable_ops.cc
index 79a64cb9219..0fc1d53749f 100644
--- a/tensorflow/core/kernels/resource_variable_ops.cc
+++ b/tensorflow/core/kernels/resource_variable_ops.cc
@@ -512,6 +512,7 @@ class AssignVariableOp : public OpKernel {
 
 TF_CALL_ALL_TYPES(REGISTER_KERNELS);
 TF_CALL_QUANTIZED_TYPES(REGISTER_KERNELS);
+TF_CALL_uint32(REGISTER_KERNELS);
 #undef REGISTER_KERNELS
 
 #if GOOGLE_CUDA || TENSORFLOW_USE_ROCM
diff --git a/tensorflow/core/kernels/split_lib_cpu.cc b/tensorflow/core/kernels/split_lib_cpu.cc
index a3060e4e90d..0cb0a94d498 100644
--- a/tensorflow/core/kernels/split_lib_cpu.cc
+++ b/tensorflow/core/kernels/split_lib_cpu.cc
@@ -43,6 +43,7 @@ void Split::operator()(
 
 TF_CALL_ALL_TYPES(DEFINE_CPU_KERNELS)
 DEFINE_CPU_KERNELS(quint8)
+DEFINE_CPU_KERNELS(uint64)
 
 #ifdef TENSORFLOW_USE_SYCL
 template 
diff --git a/tensorflow/core/kernels/split_op.cc b/tensorflow/core/kernels/split_op.cc
index 08575f01f67..f09740c6198 100644
--- a/tensorflow/core/kernels/split_op.cc
+++ b/tensorflow/core/kernels/split_op.cc
@@ -404,6 +404,7 @@ class SplitOpSYCL : public SplitOpBase {
 
 TF_CALL_ALL_TYPES(REGISTER_SPLIT);
 REGISTER_SPLIT(quint8);
+REGISTER_SPLIT(uint64);
 
 #undef REGISTER_SPLIT
 
diff --git a/tensorflow/core/kernels/strided_slice_op.cc b/tensorflow/core/kernels/strided_slice_op.cc
index b4099213303..ccc1984bb98 100644
--- a/tensorflow/core/kernels/strided_slice_op.cc
+++ b/tensorflow/core/kernels/strided_slice_op.cc
@@ -440,6 +440,8 @@ class StridedSliceAssignOp : public OpKernel {
                           StridedSliceAssignOp)
 
 TF_CALL_ALL_TYPES(REGISTER_STRIDED_SLICE);
+TF_CALL_uint32(REGISTER_STRIDED_SLICE);
+TF_CALL_uint64(REGISTER_STRIDED_SLICE);
 
 #undef REGISTER_STRIDED_SLICE
 
diff --git a/tensorflow/core/kernels/strided_slice_op_impl.h b/tensorflow/core/kernels/strided_slice_op_impl.h
index 5ce1d773e33..1ae959b7b3f 100644
--- a/tensorflow/core/kernels/strided_slice_op_impl.h
+++ b/tensorflow/core/kernels/strided_slice_op_impl.h
@@ -287,6 +287,8 @@ TF_CALL_GPU_ALL_TYPES(DECLARE_FOR_N_GPU);
 #endif  // END GOOGLE_CUDA || TENSORFLOW_USE_ROCM
 
 TF_CALL_ALL_TYPES(DECLARE_FOR_N_CPU);
+TF_CALL_uint32(DECLARE_FOR_N_CPU);
+TF_CALL_uint64(DECLARE_FOR_N_CPU);
 
 #ifdef TENSORFLOW_USE_SYCL
 #define PREVENT_FOR_N_SYCL(T) \
diff --git a/tensorflow/core/kernels/topk_op.cc b/tensorflow/core/kernels/topk_op.cc
index 50325b7bcfe..c555b42f005 100644
--- a/tensorflow/core/kernels/topk_op.cc
+++ b/tensorflow/core/kernels/topk_op.cc
@@ -258,6 +258,7 @@ namespace functor {
 
 TF_CALL_GPU_NUMBER_TYPES(DECLARE_GPU_SPEC);
 TF_CALL_INTEGRAL_TYPES(DECLARE_GPU_SPEC);
+TF_CALL_uint32(DECLARE_GPU_SPEC);
 
 #undef DECLARE_GPU_SPEC
 
@@ -275,6 +276,7 @@ TF_CALL_INTEGRAL_TYPES(DECLARE_GPU_SPEC);
 
 TF_CALL_GPU_NUMBER_TYPES(REGISTER_KERNELS);
 TF_CALL_INTEGRAL_TYPES(REGISTER_KERNELS);
+TF_CALL_uint32(REGISTER_KERNELS)
 #undef REGISTER_KERNELS
 
 #endif  // end GOOGLE_CUDA
diff --git a/tensorflow/core/kernels/topk_op_gpu_uint32.cu.cc b/tensorflow/core/kernels/topk_op_gpu_uint32.cu.cc
deleted file mode 100644
index 16e2e0e9420..00000000000
--- a/tensorflow/core/kernels/topk_op_gpu_uint32.cu.cc
+++ /dev/null
@@ -1,28 +0,0 @@
-/* Copyright 2020 The TensorFlow Authors. All Rights Reserved.
-
-Licensed under the Apache License, Version 2.0 (the "License");
-you may not use this file except in compliance with the License.
-You may obtain a copy of the License at
-
-    http://www.apache.org/licenses/LICENSE-2.0
-
-Unless required by applicable law or agreed to in writing, software
-distributed under the License is distributed on an "AS IS" BASIS,
-WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-See the License for the specific language governing permissions and
-limitations under the License.
-==============================================================================*/
-
-#if GOOGLE_CUDA
-#define EIGEN_USE_GPU
-
-#include "tensorflow/core/kernels/topk_op.h"
-#include "tensorflow/core/kernels/topk_op_gpu.h"
-
-namespace tensorflow {
-using Eigen::GpuDevice;
-
-template struct functor::TopKFunctor;
-}  // namespace tensorflow
-
-#endif  // GOOGLE_CUDA
diff --git a/tensorflow/core/kernels/topk_op_gpu_uint64.cu.cc b/tensorflow/core/kernels/topk_op_gpu_uint64.cu.cc
deleted file mode 100644
index 895247a63a2..00000000000
--- a/tensorflow/core/kernels/topk_op_gpu_uint64.cu.cc
+++ /dev/null
@@ -1,28 +0,0 @@
-/* Copyright 2020 The TensorFlow Authors. All Rights Reserved.
-
-Licensed under the Apache License, Version 2.0 (the "License");
-you may not use this file except in compliance with the License.
-You may obtain a copy of the License at
-
-    http://www.apache.org/licenses/LICENSE-2.0
-
-Unless required by applicable law or agreed to in writing, software
-distributed under the License is distributed on an "AS IS" BASIS,
-WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-See the License for the specific language governing permissions and
-limitations under the License.
-==============================================================================*/
-
-#if GOOGLE_CUDA
-#define EIGEN_USE_GPU
-
-#include "tensorflow/core/kernels/topk_op.h"
-#include "tensorflow/core/kernels/topk_op_gpu.h"
-
-namespace tensorflow {
-using Eigen::GpuDevice;
-
-template struct functor::TopKFunctor;
-}  // namespace tensorflow
-
-#endif  // GOOGLE_CUDA
diff --git a/tensorflow/core/util/batch_util.cc b/tensorflow/core/util/batch_util.cc
index e03188b04da..b88c365ced0 100644
--- a/tensorflow/core/util/batch_util.cc
+++ b/tensorflow/core/util/batch_util.cc
@@ -182,6 +182,8 @@ Status CopyElementToSlice(Tensor element, Tensor* parent, int64 index) {
   switch (element.dtype()) {
     TF_CALL_ALL_TYPES(HANDLE_TYPE);
     TF_CALL_QUANTIZED_TYPES(HANDLE_TYPE);
+    TF_CALL_uint32(HANDLE_TYPE);
+    TF_CALL_uint64(HANDLE_TYPE);
 #undef HANDLE_TYPE
     default:
       return errors::Unimplemented("CopyElementToSlice Unhandled data type: ",
@@ -205,6 +207,8 @@ Status CopySliceToElement(const Tensor& parent, Tensor* element, int64 index) {
   switch (parent.dtype()) {
     TF_CALL_ALL_TYPES(HANDLE_TYPE);
     TF_CALL_QUANTIZED_TYPES(HANDLE_TYPE);
+    TF_CALL_uint32(HANDLE_TYPE);
+    TF_CALL_uint64(HANDLE_TYPE);
 #undef HANDLE_TYPE
     default:
       return errors::Unimplemented("CopySliceToElement Unhandled data type: ",
@@ -276,6 +280,8 @@ Status CopyContiguousSlices(const Tensor& src, int64 src_offset,
   switch (src.dtype()) {
     TF_CALL_ALL_TYPES(HANDLE_TYPE);
     TF_CALL_QUANTIZED_TYPES(HANDLE_TYPE);
+    TF_CALL_uint32(HANDLE_TYPE);
+    TF_CALL_uint64(HANDLE_TYPE);
 #undef HANDLE_TYPE
     default:
       return errors::Unimplemented("CopyContiguousSlices unhandled data type: ",
@@ -302,6 +308,8 @@ Status MaybeMoveSliceToElement(Tensor* parent, Tensor* element, int64 index) {
   switch (parent->dtype()) {
     TF_CALL_ALL_TYPES(HANDLE_TYPE);
     TF_CALL_QUANTIZED_TYPES(HANDLE_TYPE);
+    TF_CALL_uint32(HANDLE_TYPE);
+    TF_CALL_uint64(HANDLE_TYPE);
 #undef HANDLE_TYPE
     default:
       return errors::Unimplemented(
diff --git a/tensorflow/core/util/saved_tensor_slice_util.h b/tensorflow/core/util/saved_tensor_slice_util.h
index 1f9768f5163..09b9235b711 100644
--- a/tensorflow/core/util/saved_tensor_slice_util.h
+++ b/tensorflow/core/util/saved_tensor_slice_util.h
@@ -116,9 +116,7 @@ TENSOR_PROTO_EXTRACT_TYPE(double, double, double);
 TENSOR_PROTO_EXTRACT_TYPE_COMPLEX(complex64, scomplex, float);
 TENSOR_PROTO_EXTRACT_TYPE_COMPLEX(complex128, dcomplex, double);
 TENSOR_PROTO_EXTRACT_TYPE(int32, int, int32);
-TENSOR_PROTO_EXTRACT_TYPE(uint32, uint32, uint32);
 TENSOR_PROTO_EXTRACT_TYPE(int64, int64, protobuf_int64);
-TENSOR_PROTO_EXTRACT_TYPE(uint64, uint64, protobuf_uint64);
 TENSOR_PROTO_EXTRACT_TYPE(uint16, int, int32);
 TENSOR_PROTO_EXTRACT_TYPE(uint8, int, int32);
 TENSOR_PROTO_EXTRACT_TYPE(int8, int, int32);

From 793c5ca70b26763c4f85f3c2961d82653d665caf Mon Sep 17 00:00:00 2001
From: Peng Wang 
Date: Sat, 20 Jun 2020 15:46:53 -0700
Subject: [PATCH 395/867] [TF-numpy] Fixed einsum's signature to be Python2
 compatible.

PiperOrigin-RevId: 317490711
Change-Id: Ia33d3958bb3163131d40262cd7247411be4ade52
---
 tensorflow/python/ops/numpy_ops/np_math_ops.py | 4 +++-
 1 file changed, 3 insertions(+), 1 deletion(-)

diff --git a/tensorflow/python/ops/numpy_ops/np_math_ops.py b/tensorflow/python/ops/numpy_ops/np_math_ops.py
index f9d3b34e90d..b242e51a2e5 100644
--- a/tensorflow/python/ops/numpy_ops/np_math_ops.py
+++ b/tensorflow/python/ops/numpy_ops/np_math_ops.py
@@ -1340,7 +1340,9 @@ def meshgrid(*xi, **kwargs):
 
 
 @np_utils.np_doc('einsum')
-def einsum(subscripts, *operands, casting='safe', optimize=False):  # pylint: disable=missing-docstring
+def einsum(subscripts, *operands, **kwargs):  # pylint: disable=missing-docstring
+  casting = kwargs.get('casting', 'safe')
+  optimize = kwargs.get('optimize', False)
   if casting == 'safe':
     operands = np_array_ops._promote_dtype(*operands)  # pylint: disable=protected-access
   elif casting == 'no':

From 443e2ae2847cc75df554507e53c083ed5312039e Mon Sep 17 00:00:00 2001
From: "A. Unique TensorFlower" 
Date: Sat, 20 Jun 2020 17:06:03 -0700
Subject: [PATCH 396/867] Slightly relax test tolerance to fix failure in
 xla_gpu version of image preprocessing test.

PiperOrigin-RevId: 317495267
Change-Id: I47cc7556f207e7f8061d68320b764c2663bcccc1
---
 .../keras/layers/preprocessing/image_preprocessing_test.py      | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/tensorflow/python/keras/layers/preprocessing/image_preprocessing_test.py b/tensorflow/python/keras/layers/preprocessing/image_preprocessing_test.py
index 24011225a29..5cb7cec5b7b 100644
--- a/tensorflow/python/keras/layers/preprocessing/image_preprocessing_test.py
+++ b/tensorflow/python/keras/layers/preprocessing/image_preprocessing_test.py
@@ -976,7 +976,7 @@ class RandomRotationTest(keras_parameterized.TestCase):
         output = strat.run(lambda: layer(input_images, training=True))
       values = output.values
       self.assertAllEqual(2, len(values))
-      self.assertAllClose(values[0], values[1])
+      self.assertAllClose(values[0], values[1], rtol=1e-5)
 
   @tf_test_util.run_v2_only
   def test_config_with_custom_name(self):

From 2e7f0f4f0491fae7ac9c4d0893b87820f8a00fad Mon Sep 17 00:00:00 2001
From: "A. Unique TensorFlower" 
Date: Sat, 20 Jun 2020 20:32:38 -0700
Subject: [PATCH 397/867] GitHub Issue #40462 fix typo

PiperOrigin-RevId: 317504961
Change-Id: I80c98d7f317280588a400eadcd0adf19d728ce66
---
 tensorflow/python/keras/preprocessing/image_dataset.py | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/tensorflow/python/keras/preprocessing/image_dataset.py b/tensorflow/python/keras/preprocessing/image_dataset.py
index 6e78ac3c03a..7c7df083559 100644
--- a/tensorflow/python/keras/preprocessing/image_dataset.py
+++ b/tensorflow/python/keras/preprocessing/image_dataset.py
@@ -147,7 +147,7 @@ def image_dataset_from_directory(directory,
           'directory. If you wish to infer the labels from the subdirectory '
           'names in the target directory, pass `labels="inferred"`. '
           'If you wish to get a dataset that only contains images '
-          '(no labels), pass `labels_mode=None`.')
+          '(no labels), pass `label_mode=None`.')
     if class_names:
       raise ValueError('You can only pass `class_names` if the labels are '
                        'inferred from the subdirectory names in the target '

From 50c07b67b1ae8b2bbbb33282bf5f7506b164dabc Mon Sep 17 00:00:00 2001
From: "A. Unique TensorFlower" 
Date: Sun, 21 Jun 2020 02:02:00 -0700
Subject: [PATCH 398/867] Update GraphDef version to 439.

PiperOrigin-RevId: 317522315
Change-Id: I7af95c686872f428e270b19f3d6ca6e24b77841d
---
 tensorflow/core/public/version.h | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/tensorflow/core/public/version.h b/tensorflow/core/public/version.h
index 64f1ef6b585..bfea50a1932 100644
--- a/tensorflow/core/public/version.h
+++ b/tensorflow/core/public/version.h
@@ -108,7 +108,7 @@ limitations under the License.
 
 #define TF_GRAPH_DEF_VERSION_MIN_PRODUCER 0
 #define TF_GRAPH_DEF_VERSION_MIN_CONSUMER 0
-#define TF_GRAPH_DEF_VERSION 438  // Updated: 2020/6/20
+#define TF_GRAPH_DEF_VERSION 439  // Updated: 2020/6/21
 
 // Checkpoint compatibility versions (the versions field in SavedSliceMeta).
 //

From f45d6083b766183e045552dafa8e46586eb3d4fc Mon Sep 17 00:00:00 2001
From: "A. Unique TensorFlower" 
Date: Sun, 21 Jun 2020 02:02:03 -0700
Subject: [PATCH 399/867] compat: Update forward compatibility horizon to
 2020-06-21

PiperOrigin-RevId: 317522319
Change-Id: Id26ff638f85872a3615fe2fe71e8cf0fced872ed
---
 tensorflow/python/compat/compat.py | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/tensorflow/python/compat/compat.py b/tensorflow/python/compat/compat.py
index ba6e375843b..3f622ef844d 100644
--- a/tensorflow/python/compat/compat.py
+++ b/tensorflow/python/compat/compat.py
@@ -33,7 +33,7 @@ from tensorflow.python.util.tf_export import tf_export
 # This value changes every day with an automatic CL. It can be modified in code
 # via `forward_compatibility_horizon()` or with the environment variable
 # TF_FORWARD_COMPATIBILITY_DELTA_DAYS, which is added to the compatibility date.
-_FORWARD_COMPATIBILITY_HORIZON = datetime.date(2020, 6, 20)
+_FORWARD_COMPATIBILITY_HORIZON = datetime.date(2020, 6, 21)
 _FORWARD_COMPATIBILITY_DELTA_DAYS_VAR_NAME = "TF_FORWARD_COMPATIBILITY_DELTA_DAYS"
 _FORWARD_COMPATIBILITY_DATE_NUMBER = None
 

From a1f9c4d80b68c47deddfbfb4ee5ccc093d347d48 Mon Sep 17 00:00:00 2001
From: Samuel Holt <6444377+samholt@users.noreply.github.com>
Date: Mon, 22 Jun 2020 01:28:32 +0100
Subject: [PATCH 400/867] Update
 tensorflow/python/keras/layers/convolutional.py

Co-authored-by: kyscg 
---
 tensorflow/python/keras/layers/convolutional.py | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/tensorflow/python/keras/layers/convolutional.py b/tensorflow/python/keras/layers/convolutional.py
index 471d94570a5..19d5ea71527 100644
--- a/tensorflow/python/keras/layers/convolutional.py
+++ b/tensorflow/python/keras/layers/convolutional.py
@@ -421,7 +421,7 @@ class Conv1D(Conv):
       specifying the stride length of the convolution.
       Specifying any stride value != 1 is incompatible with specifying
       any `dilation_rate` value != 1.
-    padding: One of `"valid"`, `"causal"` or `"same"` (case-insensitive).
+    padding: One of `"valid"`, `"same"` or `"causal"` (case-insensitive).
       `"valid"` means no padding. `"same"` results in padding evenly to 
       the left/right or up/down of the input such that output has the same 
       height/width dimension as the input.

From ebf57bdfc7f942394d4d6c67d6bbe62276cc75e6 Mon Sep 17 00:00:00 2001
From: Brian Zhao 
Date: Sun, 21 Jun 2020 19:06:43 -0700
Subject: [PATCH 401/867] Moving RAII helpers for TensorHandle, Tensor, and
 Operation to their respective classes.

PiperOrigin-RevId: 317578771
Change-Id: Iaf674696ea7d7dfdf94924f4c60d555a613c5f57
---
 tensorflow/c/eager/abstract_context.h         | 14 +++++
 tensorflow/c/eager/abstract_operation.h       | 15 ++++++
 tensorflow/c/eager/abstract_tensor_handle.h   | 16 ++++++
 .../c/eager/immediate_execution_context.h     | 15 ++++++
 .../c/eager/immediate_execution_operation.h   | 16 ++++++
 .../eager/immediate_execution_tensor_handle.h | 14 +++++
 .../c/experimental/saved_model/core/ops/BUILD | 50 +++--------------
 .../core/ops/owned_eager_context.h            | 54 -------------------
 .../saved_model/core/ops/owned_eager_op.h     | 42 ---------------
 .../saved_model/core/ops/owned_tensor.h       | 42 ---------------
 .../core/ops/owned_tensor_handle.h            | 54 -------------------
 .../saved_model/core/ops/variable_ops.cc      | 29 +++++-----
 .../saved_model/core/ops/variable_ops.h       |  5 +-
 .../saved_model/core/ops/variable_ops_test.cc | 21 ++++----
 tensorflow/c/tensor_interface.h               | 14 +++++
 .../core/common_runtime/eager/context.h       | 13 +++++
 16 files changed, 152 insertions(+), 262 deletions(-)
 delete mode 100644 tensorflow/c/experimental/saved_model/core/ops/owned_eager_context.h
 delete mode 100644 tensorflow/c/experimental/saved_model/core/ops/owned_eager_op.h
 delete mode 100644 tensorflow/c/experimental/saved_model/core/ops/owned_tensor.h
 delete mode 100644 tensorflow/c/experimental/saved_model/core/ops/owned_tensor_handle.h

diff --git a/tensorflow/c/eager/abstract_context.h b/tensorflow/c/eager/abstract_context.h
index 59c726349ac..36d983e1408 100644
--- a/tensorflow/c/eager/abstract_context.h
+++ b/tensorflow/c/eager/abstract_context.h
@@ -15,6 +15,7 @@ limitations under the License.
 #ifndef TENSORFLOW_C_EAGER_ABSTRACT_CONTEXT_H_
 #define TENSORFLOW_C_EAGER_ABSTRACT_CONTEXT_H_
 
+#include 
 #include 
 
 #include "tensorflow/c/eager/abstract_function.h"
@@ -64,6 +65,19 @@ class AbstractContext {
   const AbstractContextKind kind_;
 };
 
+namespace internal {
+struct AbstractContextDeleter {
+  void operator()(AbstractContext* p) const {
+    if (p != nullptr) {
+      p->Release();
+    }
+  }
+};
+}  // namespace internal
+
+using AbstractContextPtr =
+    std::unique_ptr;
+
 }  // namespace tensorflow
 
 #endif  // TENSORFLOW_C_EAGER_ABSTRACT_CONTEXT_H_
diff --git a/tensorflow/c/eager/abstract_operation.h b/tensorflow/c/eager/abstract_operation.h
index da4b6ecb75e..817d7656ec8 100644
--- a/tensorflow/c/eager/abstract_operation.h
+++ b/tensorflow/c/eager/abstract_operation.h
@@ -15,6 +15,8 @@ limitations under the License.
 #ifndef TENSORFLOW_C_EAGER_ABSTRACT_OPERATION_H_
 #define TENSORFLOW_C_EAGER_ABSTRACT_OPERATION_H_
 
+#include 
+
 #include "absl/types/span.h"
 #include "tensorflow/c/eager/abstract_tensor_handle.h"
 #include "tensorflow/c/tensor_interface.h"
@@ -110,6 +112,19 @@ class AbstractOperation {
   const AbstractOperationKind kind_;
 };
 
+namespace internal {
+struct AbstractOperationDeleter {
+  void operator()(AbstractOperation* p) const {
+    if (p != nullptr) {
+      p->Release();
+    }
+  }
+};
+}  // namespace internal
+
+using AbstractOpPtr =
+    std::unique_ptr;
+
 }  // namespace tensorflow
 
 #endif  // TENSORFLOW_C_EAGER_ABSTRACT_OPERATION_H_
diff --git a/tensorflow/c/eager/abstract_tensor_handle.h b/tensorflow/c/eager/abstract_tensor_handle.h
index 14acac29bb9..64b941d0729 100644
--- a/tensorflow/c/eager/abstract_tensor_handle.h
+++ b/tensorflow/c/eager/abstract_tensor_handle.h
@@ -15,6 +15,8 @@ limitations under the License.
 #ifndef TENSORFLOW_C_EAGER_ABSTRACT_TENSOR_HANDLE_H_
 #define TENSORFLOW_C_EAGER_ABSTRACT_TENSOR_HANDLE_H_
 
+#include 
+
 namespace tensorflow {
 
 // Abstract interface to a Tensor handle in either tracing or immediate
@@ -40,6 +42,20 @@ class AbstractTensorHandle {
   const AbstractTensorHandleKind kind_;
 };
 
+namespace internal {
+struct AbstractTensorHandleDeleter {
+  void operator()(AbstractTensorHandle* p) const {
+    if (p != nullptr) {
+      p->Release();
+    }
+  }
+};
+}  // namespace internal
+
+using AbstractTensorHandlePtr =
+    std::unique_ptr;
+
 }  // namespace tensorflow
 
 #endif  // TENSORFLOW_C_EAGER_ABSTRACT_TENSOR_HANDLE_H_
diff --git a/tensorflow/c/eager/immediate_execution_context.h b/tensorflow/c/eager/immediate_execution_context.h
index 0e3fe8cd4e1..77d59dd23e2 100644
--- a/tensorflow/c/eager/immediate_execution_context.h
+++ b/tensorflow/c/eager/immediate_execution_context.h
@@ -15,6 +15,7 @@ limitations under the License.
 #ifndef TENSORFLOW_C_EAGER_IMMEDIATE_EXECUTION_CONTEXT_H_
 #define TENSORFLOW_C_EAGER_IMMEDIATE_EXECUTION_CONTEXT_H_
 
+#include 
 #include 
 
 #include "absl/types/optional.h"
@@ -107,6 +108,20 @@ class ImmediateExecutionContext : public AbstractContext {
   ~ImmediateExecutionContext() override {}
 };
 
+namespace internal {
+struct ImmediateExecutionContextDeleter {
+  void operator()(ImmediateExecutionContext* p) const {
+    if (p != nullptr) {
+      p->Release();
+    }
+  }
+};
+}  // namespace internal
+
+using ImmediateContextPtr =
+    std::unique_ptr;
+
 }  // namespace tensorflow
 
 #endif  // TENSORFLOW_C_EAGER_IMMEDIATE_EXECUTION_CONTEXT_H_
diff --git a/tensorflow/c/eager/immediate_execution_operation.h b/tensorflow/c/eager/immediate_execution_operation.h
index 31413b5b4b9..4e2959ba7af 100644
--- a/tensorflow/c/eager/immediate_execution_operation.h
+++ b/tensorflow/c/eager/immediate_execution_operation.h
@@ -15,6 +15,8 @@ limitations under the License.
 #ifndef TENSORFLOW_C_EAGER_IMMEDIATE_EXECUTION_OPERATION_H_
 #define TENSORFLOW_C_EAGER_IMMEDIATE_EXECUTION_OPERATION_H_
 
+#include 
+
 #include "absl/types/span.h"
 #include "tensorflow/c/eager/abstract_operation.h"
 #include "tensorflow/c/eager/immediate_execution_tensor_handle.h"
@@ -48,6 +50,20 @@ class ImmediateExecutionOperation : public AbstractOperation {
   ~ImmediateExecutionOperation() override {}
 };
 
+namespace internal {
+struct ImmediateExecutionOperationDeleter {
+  void operator()(ImmediateExecutionOperation* p) const {
+    if (p != nullptr) {
+      p->Release();
+    }
+  }
+};
+}  // namespace internal
+
+using ImmediateOpPtr =
+    std::unique_ptr;
+
 }  // namespace tensorflow
 
 #endif  // TENSORFLOW_C_EAGER_IMMEDIATE_EXECUTION_OPERATION_H_
diff --git a/tensorflow/c/eager/immediate_execution_tensor_handle.h b/tensorflow/c/eager/immediate_execution_tensor_handle.h
index 1f5a77e54ee..31aa3aa0f75 100644
--- a/tensorflow/c/eager/immediate_execution_tensor_handle.h
+++ b/tensorflow/c/eager/immediate_execution_tensor_handle.h
@@ -59,6 +59,20 @@ class ImmediateExecutionTensorHandle : public AbstractTensorHandle {
   ~ImmediateExecutionTensorHandle() override {}
 };
 
+namespace internal {
+struct ImmediateExecutionTensorHandleDeleter {
+  void operator()(ImmediateExecutionTensorHandle* p) const {
+    if (p != nullptr) {
+      p->Release();
+    }
+  }
+};
+}  // namespace internal
+
+using ImmediateTensorHandlePtr =
+    std::unique_ptr;
+
 }  // namespace tensorflow
 
 #endif  // TENSORFLOW_C_EAGER_IMMEDIATE_EXECUTION_TENSOR_HANDLE_H_
diff --git a/tensorflow/c/experimental/saved_model/core/ops/BUILD b/tensorflow/c/experimental/saved_model/core/ops/BUILD
index 332b92bec45..1e2496487f9 100644
--- a/tensorflow/c/experimental/saved_model/core/ops/BUILD
+++ b/tensorflow/c/experimental/saved_model/core/ops/BUILD
@@ -14,44 +14,6 @@ package(
     licenses = ["notice"],  # Apache 2.0
 )
 
-cc_library(
-    name = "owned_eager_op",
-    hdrs = [
-        "owned_eager_op.h",
-    ],
-    deps = [
-        "//tensorflow/c/eager:immediate_execution_operation",
-    ],
-)
-
-cc_library(
-    name = "owned_tensor_handle",
-    hdrs = [
-        "owned_tensor_handle.h",
-    ],
-    deps = [
-        "//tensorflow/c/eager:immediate_execution_tensor_handle",
-        "//tensorflow/core/common_runtime/eager:tensor_handle",
-    ],
-)
-
-cc_library(
-    name = "owned_eager_context",
-    hdrs = ["owned_eager_context.h"],
-    deps = [
-        "//tensorflow/c/eager:immediate_execution_context",
-        "//tensorflow/core/common_runtime/eager:context",
-    ],
-)
-
-cc_library(
-    name = "owned_tensor",
-    hdrs = ["owned_tensor.h"],
-    deps = [
-        "//tensorflow/c:tensor_interface",
-    ],
-)
-
 cc_library(
     name = "variable_ops",
     srcs = [
@@ -61,10 +23,10 @@ cc_library(
         "variable_ops.h",
     ],
     deps = [
-        ":owned_eager_op",
-        ":owned_tensor_handle",
+        "//tensorflow/c/eager:abstract_operation",
         "//tensorflow/c/eager:abstract_tensor_handle",
         "//tensorflow/c/eager:immediate_execution_context",
+        "//tensorflow/c/eager:immediate_execution_operation",
         "//tensorflow/c/eager:immediate_execution_tensor_handle",
         "//tensorflow/core:framework",
         "//tensorflow/core:lib",
@@ -79,11 +41,11 @@ tf_cc_test(
         "variable_ops_test.cc",
     ],
     deps = [
-        ":owned_eager_context",
-        ":owned_tensor",
-        ":owned_tensor_handle",
         ":variable_ops",
-        "//tensorflow/core:all_kernels",
+        "//tensorflow/c:tensor_interface",
+        "//tensorflow/c/eager:abstract_tensor_handle",
+        "//tensorflow/c/eager:immediate_execution_context",
+        "//tensorflow/c/eager:immediate_execution_tensor_handle",
         "//tensorflow/core:framework",
         "//tensorflow/core:lib",
         "//tensorflow/core:protos_all_cc",
diff --git a/tensorflow/c/experimental/saved_model/core/ops/owned_eager_context.h b/tensorflow/c/experimental/saved_model/core/ops/owned_eager_context.h
deleted file mode 100644
index d944fcb51a2..00000000000
--- a/tensorflow/c/experimental/saved_model/core/ops/owned_eager_context.h
+++ /dev/null
@@ -1,54 +0,0 @@
-/* Copyright 2020 The TensorFlow Authors. All Rights Reserved.
-
-Licensed under the Apache License, Version 2.0 (the "License");
-you may not use this file except in compliance with the License.
-You may obtain a copy of the License at
-
-    http://www.apache.org/licenses/LICENSE-2.0
-
-Unless required by applicable law or agreed to in writing, software
-distributed under the License is distributed on an "AS IS" BASIS,
-WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-See the License for the specific language governing permissions and
-limitations under the License.
-==============================================================================*/
-
-#ifndef TENSORFLOW_C_EXPERIMENTAL_SAVED_MODEL_CORE_OPS_OWNED_EAGER_CONTEXT_H_
-#define TENSORFLOW_C_EXPERIMENTAL_SAVED_MODEL_CORE_OPS_OWNED_EAGER_CONTEXT_H_
-
-#include 
-
-#include "tensorflow/c/eager/immediate_execution_context.h"
-#include "tensorflow/core/common_runtime/eager/context.h"
-
-namespace tensorflow {
-namespace internal {
-
-struct ImmediateExecutionContextDeleter {
-  void operator()(ImmediateExecutionContext* p) const {
-    if (p != nullptr) {
-      p->Release();
-    }
-  }
-};
-
-struct EagerContextDeleter {
-  void operator()(EagerContext* p) const {
-    if (p != nullptr) {
-      p->Release();
-    }
-  }
-};
-
-}  // namespace internal
-
-using AbstractContextPtr =
-    std::unique_ptr;
-
-using EagerContextPtr =
-    std::unique_ptr;
-
-}  // namespace tensorflow
-
-#endif  // THIRD_PARTY_TENSORFLOW_C_EXPERIMENTAL_SAVED_MODEL_CORE_OPS_OWNED_EAGER_CONTEXT_H_
diff --git a/tensorflow/c/experimental/saved_model/core/ops/owned_eager_op.h b/tensorflow/c/experimental/saved_model/core/ops/owned_eager_op.h
deleted file mode 100644
index b3a08334a97..00000000000
--- a/tensorflow/c/experimental/saved_model/core/ops/owned_eager_op.h
+++ /dev/null
@@ -1,42 +0,0 @@
-/* Copyright 2020 The TensorFlow Authors. All Rights Reserved.
-
-Licensed under the Apache License, Version 2.0 (the "License");
-you may not use this file except in compliance with the License.
-You may obtain a copy of the License at
-
-    http://www.apache.org/licenses/LICENSE-2.0
-
-Unless required by applicable law or agreed to in writing, software
-distributed under the License is distributed on an "AS IS" BASIS,
-WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-See the License for the specific language governing permissions and
-limitations under the License.
-==============================================================================*/
-
-#ifndef TENSORFLOW_C_EXPERIMENTAL_SAVED_MODEL_CORE_OWNED_EAGER_OP_H_
-#define TENSORFLOW_C_EXPERIMENTAL_SAVED_MODEL_CORE_OWNED_EAGER_OP_H_
-
-#include 
-
-#include "tensorflow/c/eager/immediate_execution_operation.h"
-
-namespace tensorflow {
-namespace internal {
-
-struct ImmediateExecutionOperationDeleter {
-  void operator()(ImmediateExecutionOperation* p) const {
-    if (p != nullptr) {
-      p->Release();
-    }
-  }
-};
-
-}  // namespace internal
-
-using AbstractOpPtr =
-    std::unique_ptr;
-
-}  // namespace tensorflow
-
-#endif  // TENSORFLOW_C_EXPERIMENTAL_SAVED_MODEL_CORE_OWNED_EAGER_OP_H_
diff --git a/tensorflow/c/experimental/saved_model/core/ops/owned_tensor.h b/tensorflow/c/experimental/saved_model/core/ops/owned_tensor.h
deleted file mode 100644
index 335d9e46c7a..00000000000
--- a/tensorflow/c/experimental/saved_model/core/ops/owned_tensor.h
+++ /dev/null
@@ -1,42 +0,0 @@
-/* Copyright 2020 The TensorFlow Authors. All Rights Reserved.
-
-Licensed under the Apache License, Version 2.0 (the "License");
-you may not use this file except in compliance with the License.
-You may obtain a copy of the License at
-
-    http://www.apache.org/licenses/LICENSE-2.0
-
-Unless required by applicable law or agreed to in writing, software
-distributed under the License is distributed on an "AS IS" BASIS,
-WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-See the License for the specific language governing permissions and
-limitations under the License.
-==============================================================================*/
-
-#ifndef TENSORFLOW_C_EXPERIMENTAL_SAVED_MODEL_CORE_OPS_OWNED_TENSOR_H_
-#define TENSORFLOW_C_EXPERIMENTAL_SAVED_MODEL_CORE_OPS_OWNED_TENSOR_H_
-
-#include 
-
-#include "tensorflow/c/tensor_interface.h"
-
-namespace tensorflow {
-namespace internal {
-
-struct AbstractTensorInterfaceDeleter {
-  void operator()(AbstractTensorInterface* p) const {
-    if (p != nullptr) {
-      p->Release();
-    }
-  }
-};
-
-}  // namespace internal
-
-using AbstractTensorPtr =
-    std::unique_ptr;
-
-}  // namespace tensorflow
-
-#endif  // TENSORFLOW_C_EXPERIMENTAL_SAVED_MODEL_CORE_OPS_OWNED_TENSOR_H_
diff --git a/tensorflow/c/experimental/saved_model/core/ops/owned_tensor_handle.h b/tensorflow/c/experimental/saved_model/core/ops/owned_tensor_handle.h
deleted file mode 100644
index c52ebaa2479..00000000000
--- a/tensorflow/c/experimental/saved_model/core/ops/owned_tensor_handle.h
+++ /dev/null
@@ -1,54 +0,0 @@
-/* Copyright 2020 The TensorFlow Authors. All Rights Reserved.
-
-Licensed under the Apache License, Version 2.0 (the "License");
-you may not use this file except in compliance with the License.
-You may obtain a copy of the License at
-
-    http://www.apache.org/licenses/LICENSE-2.0
-
-Unless required by applicable law or agreed to in writing, software
-distributed under the License is distributed on an "AS IS" BASIS,
-WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-See the License for the specific language governing permissions and
-limitations under the License.
-==============================================================================*/
-
-#ifndef TENSORFLOW_C_EXPERIMENTAL_SAVED_MODEL_CORE_OWNED_TENSOR_HANDLE_H_
-#define TENSORFLOW_C_EXPERIMENTAL_SAVED_MODEL_CORE_OWNED_TENSOR_HANDLE_H_
-
-#include 
-
-#include "tensorflow/c/eager/immediate_execution_tensor_handle.h"
-#include "tensorflow/core/common_runtime/eager/tensor_handle.h"
-
-namespace tensorflow {
-namespace internal {
-
-struct TensorHandleDeleter {
-  void operator()(TensorHandle* p) const {
-    if (p != nullptr) {
-      p->Release();
-    }
-  }
-};
-
-struct AbstractTensorHandleDeleter {
-  void operator()(ImmediateExecutionTensorHandle* p) const {
-    if (p != nullptr) {
-      p->Release();
-    }
-  }
-};
-
-}  // namespace internal
-
-using TensorHandlePtr =
-    std::unique_ptr;
-
-using AbstractTensorHandlePtr =
-    std::unique_ptr;
-
-}  // namespace tensorflow
-
-#endif  // TENSORFLOW_C_EXPERIMENTAL_SAVED_MODEL_CORE_OWNED_TENSOR_HANDLE_H_
diff --git a/tensorflow/c/experimental/saved_model/core/ops/variable_ops.cc b/tensorflow/c/experimental/saved_model/core/ops/variable_ops.cc
index eb06662722e..67c592fc16b 100644
--- a/tensorflow/c/experimental/saved_model/core/ops/variable_ops.cc
+++ b/tensorflow/c/experimental/saved_model/core/ops/variable_ops.cc
@@ -16,10 +16,11 @@ limitations under the License.
 #include "tensorflow/c/experimental/saved_model/core/ops/variable_ops.h"
 
 #include "absl/types/span.h"
+#include "tensorflow/c/eager/abstract_operation.h"
 #include "tensorflow/c/eager/abstract_tensor_handle.h"
 #include "tensorflow/c/eager/immediate_execution_context.h"
-#include "tensorflow/c/experimental/saved_model/core/ops/owned_eager_op.h"
-#include "tensorflow/c/experimental/saved_model/core/ops/owned_tensor_handle.h"
+#include "tensorflow/c/eager/immediate_execution_operation.h"
+#include "tensorflow/c/eager/immediate_execution_tensor_handle.h"
 #include "tensorflow/core/framework/tensor_shape.h"
 #include "tensorflow/core/framework/types.pb.h"
 #include "tensorflow/core/lib/gtl/inlined_vector.h"
@@ -35,8 +36,8 @@ static const char kNoSharingResourceID[] =
 
 Status CreateUninitializedResourceVariable(ImmediateExecutionContext* ctx,
                                            DataType dtype, TensorShape shape,
-                                           AbstractTensorHandlePtr* handle) {
-  AbstractOpPtr varhandle_op = AbstractOpPtr(ctx->CreateOperation());
+                                           ImmediateTensorHandlePtr* handle) {
+  ImmediateOpPtr varhandle_op(ctx->CreateOperation());
 
   TF_RETURN_IF_ERROR(varhandle_op->Reset("VarHandleOp", nullptr));
   TF_RETURN_IF_ERROR(varhandle_op->SetAttrType("dtype", dtype));
@@ -55,17 +56,19 @@ Status CreateUninitializedResourceVariable(ImmediateExecutionContext* ctx,
   int num_retvals = 1;
   TF_RETURN_IF_ERROR(varhandle_op->Execute(
       absl::MakeSpan(&var_handle, num_retvals), &num_retvals));
-  if (var_handle->getKind() != ImmediateExecutionTensorHandle::kKind) {
+  AbstractTensorHandlePtr owned_var_handle(var_handle);
+  if (owned_var_handle->getKind() != ImmediateExecutionTensorHandle::kKind) {
     return errors::Internal("Unexpected tensor handle kind.");
   }
-  handle->reset(reinterpret_cast(var_handle));
+  handle->reset(reinterpret_cast(
+      owned_var_handle.release()));
   return Status();
 }
 
 Status AssignVariable(ImmediateExecutionContext* ctx,
                       ImmediateExecutionTensorHandle* variable_handle,
                       DataType dtype, ImmediateExecutionTensorHandle* value) {
-  AbstractOpPtr assign_op(ctx->CreateOperation());
+  ImmediateOpPtr assign_op(ctx->CreateOperation());
   TF_RETURN_IF_ERROR(assign_op->Reset("AssignVariableOp", nullptr));
   TF_RETURN_IF_ERROR(assign_op->SetAttrType("dtype", dtype));
   TF_RETURN_IF_ERROR(assign_op->AddInput(variable_handle));
@@ -78,8 +81,8 @@ Status AssignVariable(ImmediateExecutionContext* ctx,
 
 Status ReadVariable(ImmediateExecutionContext* ctx,
                     ImmediateExecutionTensorHandle* variable_handle,
-                    DataType dtype, AbstractTensorHandlePtr* output) {
-  AbstractOpPtr read_op = AbstractOpPtr(ctx->CreateOperation());
+                    DataType dtype, ImmediateTensorHandlePtr* output) {
+  ImmediateOpPtr read_op(ctx->CreateOperation());
   TF_RETURN_IF_ERROR(read_op->Reset("ReadVariableOp", nullptr));
   TF_RETURN_IF_ERROR(read_op->SetAttrType("dtype", dtype));
   TF_RETURN_IF_ERROR(read_op->AddInput(variable_handle));
@@ -88,16 +91,18 @@ Status ReadVariable(ImmediateExecutionContext* ctx,
   int num_retvals = 1;
   TF_RETURN_IF_ERROR(
       read_op->Execute(absl::MakeSpan(&value, num_retvals), &num_retvals));
-  if (value->getKind() != ImmediateExecutionTensorHandle::kKind) {
+  AbstractTensorHandlePtr owned_value(value);
+  if (owned_value->getKind() != ImmediateExecutionTensorHandle::kKind) {
     return errors::Internal("Unexpected tensor handle kind.");
   }
-  output->reset(reinterpret_cast(value));
+  output->reset(
+      reinterpret_cast(owned_value.release()));
   return Status();
 }
 
 Status DestroyResource(ImmediateExecutionContext* ctx,
                        ImmediateExecutionTensorHandle* handle) {
-  AbstractOpPtr destroy_op = AbstractOpPtr(ctx->CreateOperation());
+  ImmediateOpPtr destroy_op(ctx->CreateOperation());
   TF_RETURN_IF_ERROR(destroy_op->Reset("DestroyResourceOp", nullptr));
   TF_RETURN_IF_ERROR(destroy_op->SetAttrBool("ignore_lookup_error", true));
   TF_RETURN_IF_ERROR(destroy_op->AddInput(handle));
diff --git a/tensorflow/c/experimental/saved_model/core/ops/variable_ops.h b/tensorflow/c/experimental/saved_model/core/ops/variable_ops.h
index 038b2c3d62a..13c941a77fe 100644
--- a/tensorflow/c/experimental/saved_model/core/ops/variable_ops.h
+++ b/tensorflow/c/experimental/saved_model/core/ops/variable_ops.h
@@ -18,7 +18,6 @@ limitations under the License.
 
 #include "tensorflow/c/eager/immediate_execution_context.h"
 #include "tensorflow/c/eager/immediate_execution_tensor_handle.h"
-#include "tensorflow/c/experimental/saved_model/core/ops/owned_tensor_handle.h"
 #include "tensorflow/core/framework/tensor_shape.h"
 #include "tensorflow/core/framework/types.pb.h"
 #include "tensorflow/core/platform/status.h"
@@ -32,7 +31,7 @@ namespace internal {
 // https://github.com/tensorflow/tensorflow/blob/516608035f85cec8b126712b0ff8407220206b22/tensorflow/python/ops/resource_variable_ops.py#L1867-L1872
 Status CreateUninitializedResourceVariable(ImmediateExecutionContext* ctx,
                                            DataType dtype, TensorShape shape,
-                                           AbstractTensorHandlePtr* handle);
+                                           ImmediateTensorHandlePtr* handle);
 
 // Executes an AssignVariableOp using `ctx`, assigning the variable associated
 // with `variable_handle` with `value`. `dtype` must be the datatype of the
@@ -48,7 +47,7 @@ Status AssignVariable(ImmediateExecutionContext* ctx,
 // the dtype of the variable associated with `variable_handle`.
 Status ReadVariable(ImmediateExecutionContext* ctx,
                     ImmediateExecutionTensorHandle* variable_handle,
-                    DataType dtype, AbstractTensorHandlePtr* output);
+                    DataType dtype, ImmediateTensorHandlePtr* output);
 
 // Executes DestroyResourceOp on `handle`, using `ctx`. This is equivalent to
 // the cleanup that occurs in a tf.Variable's EagerResourceDeleter:
diff --git a/tensorflow/c/experimental/saved_model/core/ops/variable_ops_test.cc b/tensorflow/c/experimental/saved_model/core/ops/variable_ops_test.cc
index 3c57ed4d38a..09c45332efc 100644
--- a/tensorflow/c/experimental/saved_model/core/ops/variable_ops_test.cc
+++ b/tensorflow/c/experimental/saved_model/core/ops/variable_ops_test.cc
@@ -17,9 +17,8 @@ limitations under the License.
 
 #include 
 
-#include "tensorflow/c/experimental/saved_model/core/ops/owned_eager_context.h"
-#include "tensorflow/c/experimental/saved_model/core/ops/owned_tensor.h"
-#include "tensorflow/c/experimental/saved_model/core/ops/owned_tensor_handle.h"
+#include "tensorflow/c/eager/immediate_execution_tensor_handle.h"
+#include "tensorflow/c/tensor_interface.h"
 #include "tensorflow/core/common_runtime/device_mgr.h"
 #include "tensorflow/core/common_runtime/eager/context.h"
 #include "tensorflow/core/framework/tensor.h"
@@ -30,10 +29,10 @@ limitations under the License.
 namespace tensorflow {
 namespace {
 
-AbstractTensorHandlePtr CreateScalarTensorHandle(EagerContext* context,
-                                                 float value) {
+ImmediateTensorHandlePtr CreateScalarTensorHandle(EagerContext* context,
+                                                  float value) {
   AbstractTensorPtr tensor(context->CreateFloatScalar(value));
-  AbstractTensorHandlePtr handle(context->CreateLocalHandle(tensor.get()));
+  ImmediateTensorHandlePtr handle(context->CreateLocalHandle(tensor.get()));
   return handle;
 }
 
@@ -62,7 +61,7 @@ class VariableOpsTest : public ::testing::Test {
 // Sanity check for variable creation
 TEST_F(VariableOpsTest, CreateVariableSuccessful) {
   // Create a DT_Resource TensorHandle that points to a scalar DT_FLOAT tensor
-  AbstractTensorHandlePtr handle;
+  ImmediateTensorHandlePtr handle;
   TF_EXPECT_OK(internal::CreateUninitializedResourceVariable(
       context(), DT_FLOAT, {}, &handle));
   // The created TensorHandle should be a DT_Resource
@@ -72,7 +71,7 @@ TEST_F(VariableOpsTest, CreateVariableSuccessful) {
 // Sanity check for variable destruction
 TEST_F(VariableOpsTest, DestroyVariableSuccessful) {
   // Create a DT_Resource TensorHandle that points to a scalar DT_FLOAT tensor
-  AbstractTensorHandlePtr handle;
+  ImmediateTensorHandlePtr handle;
   TF_EXPECT_OK(internal::CreateUninitializedResourceVariable(
       context(), DT_FLOAT, {}, &handle));
 
@@ -83,18 +82,18 @@ TEST_F(VariableOpsTest, DestroyVariableSuccessful) {
 // Sanity check for handle assignment and reading
 TEST_F(VariableOpsTest, AssignVariableAndReadSuccessful) {
   // Create a DT_Resource TensorHandle that points to a scalar DT_FLOAT tensor
-  AbstractTensorHandlePtr variable;
+  ImmediateTensorHandlePtr variable;
   TF_EXPECT_OK(internal::CreateUninitializedResourceVariable(
       context(), DT_FLOAT, {}, &variable));
 
   // Create a Scalar float TensorHandle with value 42, and assign it to
   // the variable.
-  AbstractTensorHandlePtr my_value = CreateScalarTensorHandle(context(), 42.0);
+  ImmediateTensorHandlePtr my_value = CreateScalarTensorHandle(context(), 42.0);
   TF_EXPECT_OK(internal::AssignVariable(context(), variable.get(), DT_FLOAT,
                                         my_value.get()));
 
   // Read back the value from the variable, and check that it is 42.
-  AbstractTensorHandlePtr read_value_handle;
+  ImmediateTensorHandlePtr read_value_handle;
   TF_EXPECT_OK(internal::ReadVariable(context(), variable.get(), DT_FLOAT,
                                       &read_value_handle));
   Status status;
diff --git a/tensorflow/c/tensor_interface.h b/tensorflow/c/tensor_interface.h
index eb0d28b0bf9..d165c84980c 100644
--- a/tensorflow/c/tensor_interface.h
+++ b/tensorflow/c/tensor_interface.h
@@ -54,6 +54,20 @@ class AbstractTensorInterface {
   virtual ~AbstractTensorInterface() {}
 };
 
+namespace internal {
+struct AbstractTensorInterfaceDeleter {
+  void operator()(AbstractTensorInterface* p) const {
+    if (p != nullptr) {
+      p->Release();
+    }
+  }
+};
+}  // namespace internal
+
+using AbstractTensorPtr =
+    std::unique_ptr;
+
 }  // namespace tensorflow
 
 #endif  // TENSORFLOW_C_TENSOR_INTERFACE_H_
diff --git a/tensorflow/core/common_runtime/eager/context.h b/tensorflow/core/common_runtime/eager/context.h
index 141327c08cb..3dab7c08d77 100644
--- a/tensorflow/core/common_runtime/eager/context.h
+++ b/tensorflow/core/common_runtime/eager/context.h
@@ -722,6 +722,19 @@ inline EagerContext* ContextFromInterface(ImmediateExecutionContext* context) {
   return down_cast(context);
 }
 
+namespace internal {
+struct EagerContextDeleter {
+  void operator()(EagerContext* p) const {
+    if (p != nullptr) {
+      p->Release();
+    }
+  }
+};
+}  // namespace internal
+
+using EagerContextPtr =
+    std::unique_ptr;
+
 }  // namespace tensorflow
 
 #endif  // TENSORFLOW_CORE_COMMON_RUNTIME_EAGER_CONTEXT_H_

From 8e9a05117bd5c810fa49d56a4cb8a99dc952eac4 Mon Sep 17 00:00:00 2001
From: Shanqing Cai 
Date: Sun, 21 Jun 2020 19:49:53 -0700
Subject: [PATCH 402/867] Improve the documentation of
 tf.debugging.assert_shapes()

- Specify the first argument as a list of (tensor, shape) tuples.
  We no longer want to publicly support hashable Tensors, as in TF2
  Tensors are not hashable.
- Better explain the shape iterables.

PiperOrigin-RevId: 317581098
Change-Id: I6bb2a114c884ee5b282f085865a5e6e30eb6ae15
---
 tensorflow/python/ops/check_ops.py | 29 ++++++++++++++++++++++-------
 1 file changed, 22 insertions(+), 7 deletions(-)

diff --git a/tensorflow/python/ops/check_ops.py b/tensorflow/python/ops/check_ops.py
index 680796df48d..1e4b3ca1bb4 100644
--- a/tensorflow/python/ops/check_ops.py
+++ b/tensorflow/python/ops/check_ops.py
@@ -146,7 +146,7 @@ def _unary_assert_doc(sym, sym_name):
 
     Raises:
       InvalidArgumentError: if the check can be performed immediately and
-        `x {sym}` is False. The check can be performed immediately during 
+        `x {sym}` is False. The check can be performed immediately during
         eager execution or if `x` is statically known.
     """.format(
         sym=sym, sym_name=cap_sym_name, opname=opname)
@@ -209,7 +209,7 @@ def _binary_assert_doc(sym):
 
     Raises:
       InvalidArgumentError: if the check can be performed immediately and
-        `x {sym} y` is False. The check can be performed immediately during 
+        `x {sym} y` is False. The check can be performed immediately during
         eager execution or if `x` and `y` are statically known.
     """.format(
         sym=sym, opname=opname)
@@ -1634,7 +1634,7 @@ def assert_shapes_v2(shapes, data=None, summarize=None, message=None,
   >>> n = 10
   >>> q = 3
   >>> d = 7
-  >>> x = tf.zeros([n,q]) 
+  >>> x = tf.zeros([n,q])
   >>> y = tf.ones([n,d])
   >>> param = tf.Variable([1.0, 2.0, 3.0])
   >>> scalar = 1.0
@@ -1644,9 +1644,9 @@ def assert_shapes_v2(shapes, data=None, summarize=None, message=None,
   ...  (param, ('Q',)),
   ...  (scalar, ()),
   ... ])
-  
+
   >>> tf.debugging.assert_shapes([
-  ...   (x, ('N', 'D')), 
+  ...   (x, ('N', 'D')),
   ...   (y, ('N', 'D'))
   ... ])
   Traceback (most recent call last):
@@ -1745,8 +1745,23 @@ def assert_shapes(shapes, data=None, summarize=None, message=None, name=None):
   prefix) are both treated as having a single dimension of size one.
 
   Args:
-    shapes: dictionary with (`Tensor` to shape) items, or a list of
-      (`Tensor`, shape) tuples. A shape must be an iterable.
+    shapes: A list of (`Tensor`, `shape`) tuples, wherein `shape` is the
+      expected shape of `Tensor`. See the example code above. The `shape` must
+      be an iterable. Each element of the iterable can be either a concrete
+      integer value or a string that abstractly represents the dimension.
+      For example,
+        - `('N', 'Q')` specifies a 2D shape wherein the first and second
+          dimensions of shape may or may not be equal.
+        - `('N', 'N', 'Q')` specifies a 3D shape wherein the first and second
+          dimensions are equal.
+        - `(1, 'N')` specifies a 2D shape wherein the first dimension is
+          exactly 1 and the second dimension can be any value.
+      Note that the abstract dimension letters take effect across different
+      tuple elements of the list. For example,
+      `tf.debugging.assert_shapes([(x, ('N', 'A')), (y, ('N', 'B'))]` asserts
+      that both `x` and `y` are rank-2 tensors and their first dimensions are
+      equal (`N`).
+      `shape` can also be a `tf.TensorShape`.
     data: The tensors to print out if the condition is False.  Defaults to error
       message and first few entries of the violating tensor.
     summarize: Print this many entries of the tensor.

From db121c7f4b9f8d5a993fc9ed7c502bba2b68b135 Mon Sep 17 00:00:00 2001
From: YoungSeok Yoon 
Date: Sun, 21 Jun 2020 21:08:24 -0700
Subject: [PATCH 403/867] Clean up Makefile based build scripts for TFLite iOS

For TFLite iOS, bazel has been the recommended way to build the static framework
and the Makefile based script is not even meeting feature parity with the bazel
build (e.g., no GPU delegate support).

PiperOrigin-RevId: 317587116
Change-Id: Idbdf60687f35e1ef0bdb074668317818eeaa95e5
---
 .../lite/lib_package/create_ios_frameworks.sh | 120 ------------------
 .../tools/make/build_ios_universal_lib.sh     |  12 +-
 2 files changed, 10 insertions(+), 122 deletions(-)
 delete mode 100755 tensorflow/lite/lib_package/create_ios_frameworks.sh

diff --git a/tensorflow/lite/lib_package/create_ios_frameworks.sh b/tensorflow/lite/lib_package/create_ios_frameworks.sh
deleted file mode 100755
index 71aee65848f..00000000000
--- a/tensorflow/lite/lib_package/create_ios_frameworks.sh
+++ /dev/null
@@ -1,120 +0,0 @@
-#!/bin/bash
-# Copyright 2017 The TensorFlow Authors. All Rights Reserved.
-#
-# Licensed under the Apache License, Version 2.0 (the "License");
-# you may not use this file except in compliance with the License.
-# You may obtain a copy of the License at
-#
-#     http://www.apache.org/licenses/LICENSE-2.0
-#
-# Unless required by applicable law or agreed to in writing, software
-# distributed under the License is distributed on an "AS IS" BASIS,
-# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-# See the License for the specific language governing permissions and
-# limitations under the License.
-# ==============================================================================
-
-# TODO(ycling): Refactoring - Move this script into `tools/make`.
-set -e
-
-echo "Starting"
-TFLITE_DIR="$(cd "$(dirname "${BASH_SOURCE[0]}")" && pwd)/.."
-
-usage() {
-  echo "Usage: $(basename "$0") [-a]"
-  echo "-g build with GPU delegate"
-  exit 1
-}
-
-USE_GPU_DELEGATE="false"
-FRAMEWORK_NAME="tensorflow_lite"
-while getopts "g" opt_name; do
-  case "$opt_name" in
-    g)
-        USE_GPU_DELEGATE="true"
-        FRAMEWORK_NAME="tensorflow_lite_gpu"
-        ;;
-    *) usage;;
-  esac
-done
-shift $((OPTIND - 1))
-readonly USE_GPU_DELEGATE
-readonly FRAMEWORK_NAME
-
-if [ $USE_GPU_DELEGATE == "true" ] ; then
-  for filename in metal_delegate.h libmetal_delegate.a ; do
-    if [[ ! -f "${TFLITE_DIR}/delegates/gpu/${filename}" ]] ; then
-      echo "File ${TFLITE_DIR}/delegates/gpu/${filename} doesn't exist."
-      echo "It's required for building TFLite Framework with GPU. Aborting."
-      exit 1
-    fi
-  done
-fi
-
-TMP_DIR=$(mktemp -d)
-echo "Package dir: " $TMP_DIR
-FW_DIR=$TMP_DIR/tensorflow_lite_ios_frameworks
-FW_DIR_TFLITE=$FW_DIR/$FRAMEWORK_NAME.framework
-FW_DIR_TFLITE_HDRS=$FW_DIR_TFLITE/Headers
-
-echo "Creating target Headers directories"
-mkdir -p $FW_DIR_TFLITE_HDRS
-
-echo "Headers, populating: TensorFlow Lite"
-cd $TFLITE_DIR/../..
-
-find tensorflow/lite -name '*.h' \
-    -not -path 'tensorflow/lite/tools/*' \
-    -not -path 'tensorflow/lite/examples/*' \
-    -not -path 'tensorflow/lite/gen/*' \
-    -not -path 'tensorflow/lite/toco/*' \
-    -not -path 'tensorflow/lite/nnapi/*' \
-    -not -path 'tensorflow/lite/java/*' \
-    | tar -cf $FW_DIR_TFLITE_HDRS/tmp.tar -T -
-cd $FW_DIR_TFLITE_HDRS
-tar xf tmp.tar
-rm -f tmp.tar
-
-echo "Headers, populating: Flatbuffer"
-cd $TFLITE_DIR/tools/make/downloads/flatbuffers/include/
-find . -name '*.h' | tar -cf $FW_DIR_TFLITE_HDRS/tmp.tar -T -
-cd $FW_DIR_TFLITE_HDRS
-tar xf tmp.tar
-rm -f tmp.tar
-
-cd $TFLITE_DIR/../..
-echo "Generate LICENSE files and copy to target"
-bazel build //tensorflow/tools/lib_package:clicenses_generate
-cp $TFLITE_DIR/../../LICENSE $FW_DIR_TFLITE
-cp $TFLITE_DIR/../../bazel-bin/tensorflow/tools/lib_package/THIRD_PARTY_TF_C_LICENSES \
-   $FW_DIR_TFLITE
-
-echo "Copying static libraries"
-# Note: There must be a static library with the same name
-# as the framework name.
-cp $TFLITE_DIR/tools/make/gen/lib/libtensorflow-lite.a \
-    $FW_DIR_TFLITE/$FRAMEWORK_NAME
-if [ $USE_GPU_DELEGATE == "true" ] ; then
-  cp "${TFLITE_DIR}/delegates/gpu/libmetal_delegate.a" \
-      $FW_DIR_TFLITE/libmetal_delegate.a
-fi
-
-# This is required, otherwise they interfere with the documentation of the
-# pod at cocoapods.org.
-echo "Remove all README files"
-cd $FW_DIR_TFLITE_HDRS
-find . -type f -name README\* -exec rm -f {} \;
-find . -type f -name readme\* -exec rm -f {} \;
-
-TARGET_GEN_LOCATION="$TFLITE_DIR/gen/ios_frameworks"
-echo "Moving results to target: " $TARGET_GEN_LOCATION
-cd $FW_DIR
-zip -q -r $FRAMEWORK_NAME.framework.zip $FRAMEWORK_NAME.framework -x .DS_Store
-rm -rf $TARGET_GEN_LOCATION
-mkdir -p $TARGET_GEN_LOCATION
-cp -r $FRAMEWORK_NAME.framework.zip $TARGET_GEN_LOCATION
-
-echo "Cleaning up"
-rm -rf $TMP_DIR
-
-echo "Finished"
diff --git a/tensorflow/lite/tools/make/build_ios_universal_lib.sh b/tensorflow/lite/tools/make/build_ios_universal_lib.sh
index 74bf9183541..87ca0649303 100755
--- a/tensorflow/lite/tools/make/build_ios_universal_lib.sh
+++ b/tensorflow/lite/tools/make/build_ios_universal_lib.sh
@@ -1,4 +1,4 @@
-#!/bin/bash -x
+#!/bin/bash +x
 # Copyright 2017 The TensorFlow Authors. All Rights Reserved.
 #
 # Licensed under the Apache License, Version 2.0 (the "License");
@@ -14,7 +14,15 @@
 # limitations under the License.
 # ==============================================================================
 
-set -e
+echo "========================================================================="
+echo "WARNING: This build script is deprecated and no longer maintained. Please"
+echo "         refer to the iOS build guide to learn how to build the latest   "
+echo "         version of TFLite static framework for iOS using bazel.         "
+echo "         https://www.tensorflow.org/lite/guide/build_ios                 "
+echo "========================================================================="
+sleep 5s
+
+set -ex
 
 SCRIPT_DIR="$(cd "$(dirname "${BASH_SOURCE[0]}")" && pwd)"
 cd "$SCRIPT_DIR/../../../.."

From f24d25d8603c3fb7465922ad0e68737dade34db2 Mon Sep 17 00:00:00 2001
From: "Jae H. Yoo" 
Date: Sun, 21 Jun 2020 21:29:52 -0700
Subject: [PATCH 404/867] Rename kTfLiteActRelu1 to kTfLiteActReluN1To1

This CL renames `kTfLiteActRelu1` to `kTfLiteActReluN1To1` because it has minimum clipping at -1, not 0. Also, This CL finishes renaming because TFLite already uses `kTfLiteBuiltinReluN1To1`, `ActivationFunctionType_RELU_N1_TO_1` for this op.

PiperOrigin-RevId: 317589358
Change-Id: I2424104da45234346749b3921d563e9161e809cc
---
 tensorflow/lite/c/builtin_op_data.h           |  5 +++--
 .../lite/core/api/flatbuffer_conversions.cc   |  2 +-
 .../delegates/gpu/common/model_builder.cc     |  6 +++---
 .../hexagon/builders/conv_2d_builder.cc       |  2 +-
 tensorflow/lite/delegates/hexagon/utils.cc    |  2 +-
 .../delegates/xnnpack/xnnpack_delegate.cc     |  4 ++--
 .../builders/activation_layer_builder.cc      |  6 +++---
 .../lite/experimental/writer/enum_mapping.h   |  2 +-
 tensorflow/lite/kernels/fully_connected.cc    |  2 +-
 .../lite/kernels/internal/tensor_utils.h      |  2 +-
 tensorflow/lite/kernels/kernel_util.cc        |  2 +-
 tensorflow/lite/kernels/kernel_util.h         |  2 +-
 .../lite/micro/kernels/activation_utils.h     |  2 +-
 tensorflow/lite/micro/kernels/add_test.cc     |  6 +++---
 tensorflow/lite/micro/kernels/mul_test.cc     |  2 +-
 tensorflow/lite/micro/kernels/pooling_test.cc | 19 ++++++++++---------
 tensorflow/lite/micro/kernels/sub_test.cc     |  6 +++---
 17 files changed, 37 insertions(+), 35 deletions(-)

diff --git a/tensorflow/lite/c/builtin_op_data.h b/tensorflow/lite/c/builtin_op_data.h
index 9e0e82bc906..232f5f95928 100644
--- a/tensorflow/lite/c/builtin_op_data.h
+++ b/tensorflow/lite/c/builtin_op_data.h
@@ -67,8 +67,9 @@ typedef struct {
 typedef enum {
   kTfLiteActNone = 0,
   kTfLiteActRelu,
-  kTfLiteActRelu1,  // min(max(-1, x), 1)
-  kTfLiteActRelu6,  // min(max(0, x), 6)
+  kTfLiteActReluN1To1,                    // min(max(-1, x), 1)
+  kTfLiteActRelu1 = kTfLiteActReluN1To1,  // kTfLiteActRelu1 will be deprecated.
+  kTfLiteActRelu6,                        // min(max(0, x), 6)
   kTfLiteActTanh,
   kTfLiteActSignBit,
   kTfLiteActSigmoid,
diff --git a/tensorflow/lite/core/api/flatbuffer_conversions.cc b/tensorflow/lite/core/api/flatbuffer_conversions.cc
index 2a4dfbb6ff4..73d785bf369 100644
--- a/tensorflow/lite/core/api/flatbuffer_conversions.cc
+++ b/tensorflow/lite/core/api/flatbuffer_conversions.cc
@@ -109,7 +109,7 @@ TfLiteFusedActivation ConvertActivation(ActivationFunctionType activation) {
     case ActivationFunctionType_RELU:
       return kTfLiteActRelu;
     case ActivationFunctionType_RELU_N1_TO_1:
-      return kTfLiteActRelu1;
+      return kTfLiteActReluN1To1;
     case ActivationFunctionType_RELU6:
       return kTfLiteActRelu6;
     case ActivationFunctionType_TANH:
diff --git a/tensorflow/lite/delegates/gpu/common/model_builder.cc b/tensorflow/lite/delegates/gpu/common/model_builder.cc
index 01f94c94888..8b5261cfd98 100644
--- a/tensorflow/lite/delegates/gpu/common/model_builder.cc
+++ b/tensorflow/lite/delegates/gpu/common/model_builder.cc
@@ -109,7 +109,7 @@ absl::Status IsActivationSupported(TfLiteFusedActivation fused_activation) {
   switch (fused_activation) {
     case kTfLiteActNone:
     case kTfLiteActRelu:
-    case kTfLiteActRelu1:
+    case kTfLiteActReluN1To1:
     case kTfLiteActRelu6:
     case kTfLiteActTanh:
       return absl::OkStatus();
@@ -140,12 +140,12 @@ absl::Status MaybeFuseActivation(TfLiteFusedActivation fused_activation,
   }
   switch (fused_activation) {
     case kTfLiteActRelu:
-    case kTfLiteActRelu1:
+    case kTfLiteActReluN1To1:
     case kTfLiteActRelu6: {
       ReLUAttributes attr;
       attr.clip = fused_activation == kTfLiteActRelu
                       ? 0.0f
-                      : (fused_activation == kTfLiteActRelu1 ? 1.0f : 6.0f);
+                      : (fused_activation == kTfLiteActReluN1To1 ? 1.0f : 6.0f);
       for (auto index : output_indices) {
         Node* activation_node;
         RETURN_IF_ERROR(
diff --git a/tensorflow/lite/delegates/hexagon/builders/conv_2d_builder.cc b/tensorflow/lite/delegates/hexagon/builders/conv_2d_builder.cc
index a366522e35c..cfddd2c2b97 100644
--- a/tensorflow/lite/delegates/hexagon/builders/conv_2d_builder.cc
+++ b/tensorflow/lite/delegates/hexagon/builders/conv_2d_builder.cc
@@ -197,7 +197,7 @@ TfLiteStatus Conv2dOpBuilder::PopulateSubGraph(const TfLiteIntArray* inputs,
   if (activation == kTfLiteActRelu6) {
     conv_output_min = 0;
     conv_output_max = 6;
-  } else if (activation == kTfLiteActRelu1) {
+  } else if (activation == kTfLiteActReluN1To1) {
     conv_output_min = -1;
     conv_output_max = 1;
   } else if (activation == kTfLiteActRelu) {
diff --git a/tensorflow/lite/delegates/hexagon/utils.cc b/tensorflow/lite/delegates/hexagon/utils.cc
index 9253836a3b1..223d4a8a826 100644
--- a/tensorflow/lite/delegates/hexagon/utils.cc
+++ b/tensorflow/lite/delegates/hexagon/utils.cc
@@ -26,7 +26,7 @@ namespace {
 
 bool IsActivationReluOrNone(TfLiteFusedActivation activation) {
   return (activation == kTfLiteActRelu || activation == kTfLiteActRelu6 ||
-          activation == kTfLiteActRelu1 || activation == kTfLiteActNone);
+          activation == kTfLiteActReluN1To1 || activation == kTfLiteActNone);
 }
 
 bool TensorTypeMatch(int tensor_id, TfLiteContext* context,
diff --git a/tensorflow/lite/delegates/xnnpack/xnnpack_delegate.cc b/tensorflow/lite/delegates/xnnpack/xnnpack_delegate.cc
index 739e45f62e4..0afc9c32122 100644
--- a/tensorflow/lite/delegates/xnnpack/xnnpack_delegate.cc
+++ b/tensorflow/lite/delegates/xnnpack/xnnpack_delegate.cc
@@ -330,7 +330,7 @@ class Subgraph {
         *output_min = 0.0f;
         *output_max = +std::numeric_limits::infinity();
         return kTfLiteOk;
-      case kTfLiteActRelu1:
+      case kTfLiteActReluN1To1:
         *output_min = -1.0f;
         *output_max = +1.0f;
         return kTfLiteOk;
@@ -497,7 +497,7 @@ class Subgraph {
             context, "unsupported fused activation (Relu) in node #%d",
             node_index);
         return kTfLiteOk;
-      case kTfLiteActRelu1:
+      case kTfLiteActReluN1To1:
         TF_LITE_MAYBE_KERNEL_LOG(
             context, "unsupported fused activation (ReluMinus1To1) in node #%d",
             node_index);
diff --git a/tensorflow/lite/experimental/delegates/coreml/builders/activation_layer_builder.cc b/tensorflow/lite/experimental/delegates/coreml/builders/activation_layer_builder.cc
index ec032d8421e..df853797c8a 100644
--- a/tensorflow/lite/experimental/delegates/coreml/builders/activation_layer_builder.cc
+++ b/tensorflow/lite/experimental/delegates/coreml/builders/activation_layer_builder.cc
@@ -41,7 +41,7 @@ CoreML::Specification::NeuralNetworkLayer* ActivationLayerBuilder::Build() {
       layer_->mutable_activation()->mutable_relu();
       break;
     // Relu1 and Relu6 layers are fully composed in PopulateSubgraph().
-    case kTfLiteActRelu1:  // clip(-1, 1)
+    case kTfLiteActReluN1To1:  // clip(-1, 1)
       layer_->mutable_unary()->set_alpha(-1);
       layer_->mutable_unary()->set_type(
           CoreML::Specification::UnaryFunctionLayerParams::THRESHOLD);
@@ -64,7 +64,7 @@ CoreML::Specification::NeuralNetworkLayer* ActivationLayerBuilder::Build() {
 }
 
 TfLiteStatus ActivationLayerBuilder::PopulateSubgraph(TfLiteContext* context) {
-  if (!(activation_ == kTfLiteActRelu6 || activation_ == kTfLiteActRelu1)) {
+  if (!(activation_ == kTfLiteActRelu6 || activation_ == kTfLiteActReluN1To1)) {
     builder_output_ = AddOutput();
     return kTfLiteOk;
   }
@@ -125,7 +125,7 @@ OpBuilder* CreateReluOpBuilder(GraphBuilder* graph_builder) {
 }
 
 OpBuilder* CreateReluN1To1OpBuilder(GraphBuilder* graph_builder) {
-  return new ActivationLayerBuilder(graph_builder, kTfLiteActRelu1);
+  return new ActivationLayerBuilder(graph_builder, kTfLiteActReluN1To1);
 }
 
 OpBuilder* CreateRelu6OpBuilder(GraphBuilder* graph_builder) {
diff --git a/tensorflow/lite/experimental/writer/enum_mapping.h b/tensorflow/lite/experimental/writer/enum_mapping.h
index b78d610c4c5..5eabbcb2015 100644
--- a/tensorflow/lite/experimental/writer/enum_mapping.h
+++ b/tensorflow/lite/experimental/writer/enum_mapping.h
@@ -29,7 +29,7 @@ inline ActivationFunctionType TfLiteActivationToSchemaActivation(
       return ActivationFunctionType_NONE;
     case kTfLiteActRelu:
       return ActivationFunctionType_RELU;
-    case kTfLiteActRelu1:
+    case kTfLiteActReluN1To1:
       return ActivationFunctionType_RELU_N1_TO_1;
     case kTfLiteActRelu6:
       return ActivationFunctionType_RELU6;
diff --git a/tensorflow/lite/kernels/fully_connected.cc b/tensorflow/lite/kernels/fully_connected.cc
index 8b7a7832dbb..9cbbcae9c51 100644
--- a/tensorflow/lite/kernels/fully_connected.cc
+++ b/tensorflow/lite/kernels/fully_connected.cc
@@ -312,7 +312,7 @@ TfLiteStatus Prepare(TfLiteContext* context, TfLiteNode* node) {
   if (!is_pie && !is_hybrid) {
     TF_LITE_ENSURE(context, params->activation == kTfLiteActNone ||
                                 params->activation == kTfLiteActRelu ||
-                                params->activation == kTfLiteActRelu1 ||
+                                params->activation == kTfLiteActReluN1To1 ||
                                 params->activation == kTfLiteActRelu6);
   }
   return PrepareImpl(context, node);
diff --git a/tensorflow/lite/kernels/internal/tensor_utils.h b/tensorflow/lite/kernels/internal/tensor_utils.h
index e2af88d50e3..8c956c49f5f 100644
--- a/tensorflow/lite/kernels/internal/tensor_utils.h
+++ b/tensorflow/lite/kernels/internal/tensor_utils.h
@@ -587,7 +587,7 @@ inline void ApplyActivationToVector(const float* __restrict__ vector,
       return;
     case kTfLiteActRelu:
       return ApplyReluToVector(vector, v_size, result);
-    case kTfLiteActRelu1:
+    case kTfLiteActReluN1To1:
       return ApplyRelu1ToVector(vector, v_size, result);
     case kTfLiteActRelu6:
       return ApplyRelu6ToVector(vector, v_size, result);
diff --git a/tensorflow/lite/kernels/kernel_util.cc b/tensorflow/lite/kernels/kernel_util.cc
index 032726a7860..164aec3f224 100644
--- a/tensorflow/lite/kernels/kernel_util.cc
+++ b/tensorflow/lite/kernels/kernel_util.cc
@@ -188,7 +188,7 @@ void CalculateActivationRangeQuantizedImpl(TfLiteFusedActivation activation,
   } else if (activation == kTfLiteActRelu6) {
     *act_min = std::max(qmin, quantize(0.0));
     *act_max = std::min(qmax, quantize(6.0));
-  } else if (activation == kTfLiteActRelu1) {
+  } else if (activation == kTfLiteActReluN1To1) {
     *act_min = std::max(qmin, quantize(-1.0));
     *act_max = std::min(qmax, quantize(1.0));
   } else {
diff --git a/tensorflow/lite/kernels/kernel_util.h b/tensorflow/lite/kernels/kernel_util.h
index 6fc69fa1629..6bd6bb1c7ed 100644
--- a/tensorflow/lite/kernels/kernel_util.h
+++ b/tensorflow/lite/kernels/kernel_util.h
@@ -169,7 +169,7 @@ void CalculateActivationRange(TfLiteFusedActivation activation,
   } else if (activation == kTfLiteActRelu6) {
     *activation_min = 0;
     *activation_max = 6;
-  } else if (activation == kTfLiteActRelu1) {
+  } else if (activation == kTfLiteActReluN1To1) {
     *activation_min = -1;
     *activation_max = 1;
   } else {
diff --git a/tensorflow/lite/micro/kernels/activation_utils.h b/tensorflow/lite/micro/kernels/activation_utils.h
index a71826211c0..95ecc26dd52 100644
--- a/tensorflow/lite/micro/kernels/activation_utils.h
+++ b/tensorflow/lite/micro/kernels/activation_utils.h
@@ -35,7 +35,7 @@ inline float ActivationValFloat(TfLiteFusedActivation act, float a) {
       return a;
     case kTfLiteActRelu:
       return TfLiteMax(0.0f, a);
-    case kTfLiteActRelu1:
+    case kTfLiteActReluN1To1:
       return TfLiteMax(-1.0f, TfLiteMin(a, 1.0f));
     case kTfLiteActRelu6:
       return TfLiteMax(0.0f, TfLiteMin(a, 6.0f));
diff --git a/tensorflow/lite/micro/kernels/add_test.cc b/tensorflow/lite/micro/kernels/add_test.cc
index 60164ab4746..6c66e0d4aaf 100644
--- a/tensorflow/lite/micro/kernels/add_test.cc
+++ b/tensorflow/lite/micro/kernels/add_test.cc
@@ -201,7 +201,7 @@ TF_LITE_MICRO_TEST(FloatAddActivationRelu1) {
   float output_data[output_dims_count];
   tflite::testing::TestAddFloat(inout_shape, input1_values, inout_shape,
                                 input2_values, inout_shape, golden_values,
-                                kTfLiteActRelu1, output_data);
+                                kTfLiteActReluN1To1, output_data);
 }
 
 TF_LITE_MICRO_TEST(FloatAddVariousInputShapes) {
@@ -313,7 +313,7 @@ TF_LITE_MICRO_TEST(QuantizedAddActivationRelu1Uint8) {
       inout_shape, input1_values, input1_quantized, scales[0], zero_points[0],
       inout_shape, input2_values, input2_quantized, scales[1], zero_points[1],
       inout_shape, golden_values, golden_quantized, scales[2], zero_points[2],
-      kTfLiteActRelu1, output);
+      kTfLiteActReluN1To1, output);
 }
 
 TF_LITE_MICRO_TEST(QuantizedAddActivationRelu1Int8) {
@@ -334,7 +334,7 @@ TF_LITE_MICRO_TEST(QuantizedAddActivationRelu1Int8) {
       inout_shape, input1_values, input1_quantized, scales[0], zero_points[0],
       inout_shape, input2_values, input2_quantized, scales[1], zero_points[1],
       inout_shape, golden_values, golden_quantized, scales[2], zero_points[2],
-      kTfLiteActRelu1, output);
+      kTfLiteActReluN1To1, output);
 }
 
 TF_LITE_MICRO_TEST(QuantizedAddVariousInputShapesUint8) {
diff --git a/tensorflow/lite/micro/kernels/mul_test.cc b/tensorflow/lite/micro/kernels/mul_test.cc
index 6b4d4f07b64..f69bf2aa17e 100644
--- a/tensorflow/lite/micro/kernels/mul_test.cc
+++ b/tensorflow/lite/micro/kernels/mul_test.cc
@@ -402,7 +402,7 @@ TF_LITE_MICRO_TEST(FloatRelu) {
       {0.1, 0.2, 0.3, 0.5},     // input2 data
       {4, 1, 2, 2, 1},          // output shape
       {-0.2, 0.04, 0.21, 0.4},  // expected output data
-      output_data, kTfLiteActRelu1);
+      output_data, kTfLiteActReluN1To1);
 }
 
 TF_LITE_MICRO_TEST(FloatBroadcast) {
diff --git a/tensorflow/lite/micro/kernels/pooling_test.cc b/tensorflow/lite/micro/kernels/pooling_test.cc
index 9e11e9a4d57..35a77662e07 100644
--- a/tensorflow/lite/micro/kernels/pooling_test.cc
+++ b/tensorflow/lite/micro/kernels/pooling_test.cc
@@ -417,7 +417,8 @@ TF_LITE_MICRO_TEST(SimpleAveragePoolTestInt8PaddingValidStride1Stride2Relu) {
       kTfLitePaddingValid, kTfLiteActRelu, output_data);
 }
 
-TF_LITE_MICRO_TEST(SimpleAveragePoolTestInt8PaddingValidStride2Stride1Relu1) {
+TF_LITE_MICRO_TEST(
+    SimpleAveragePoolTestInt8PaddingValidStride2Stride1ReluN1To1) {
   using tflite::testing::F2QS;
 
   const float input_min = -15.9375;
@@ -439,7 +440,7 @@ TF_LITE_MICRO_TEST(SimpleAveragePoolTestInt8PaddingValidStride2Stride1Relu1) {
        F2QS(-0.25, output_min, output_max), F2QS(0.75, output_min, output_max)},
       {4, 1, 1, 2, 1},         // Output shape
       output_min, output_max,  // output quantization range
-      kTfLitePaddingValid, kTfLiteActRelu1, output_data);
+      kTfLitePaddingValid, kTfLiteActReluN1To1, output_data);
 }
 
 TF_LITE_MICRO_TEST(SimpleAveragePoolTestInt8PaddingValidStride2Relu6) {
@@ -532,7 +533,7 @@ TF_LITE_MICRO_TEST(SimpleMaxPoolTestFloatRelu) {
                                     output_data);
 }
 
-TF_LITE_MICRO_TEST(SimpleMaxPoolTestFloatRelu1) {
+TF_LITE_MICRO_TEST(SimpleMaxPoolTestFloatReluN1To1) {
   float output_data[2];
   tflite::testing::TestMaxPoolFloat({4, 1, 2, 4, 1},  // Input shape
                                     {
@@ -548,7 +549,7 @@ TF_LITE_MICRO_TEST(SimpleMaxPoolTestFloatRelu1) {
                                         0.7,
                                     },
                                     {4, 1, 1, 2, 1},  // Output shape
-                                    kTfLitePaddingValid, kTfLiteActRelu1,
+                                    kTfLitePaddingValid, kTfLiteActReluN1To1,
                                     output_data);
 
   tflite::testing::TestMaxPoolFloat({4, 1, 2, 4, 1},  // Input shape
@@ -565,7 +566,7 @@ TF_LITE_MICRO_TEST(SimpleMaxPoolTestFloatRelu1) {
                                         1.0,
                                     },
                                     {4, 1, 1, 2, 1},  // Output shape
-                                    kTfLitePaddingValid, kTfLiteActRelu1,
+                                    kTfLitePaddingValid, kTfLiteActReluN1To1,
                                     output_data);
 }
 
@@ -713,7 +714,7 @@ TF_LITE_MICRO_TEST(MaxPoolTestUInt8ActRelu) {
       kTfLitePaddingValid, kTfLiteActRelu, output_data);
 }
 
-TF_LITE_MICRO_TEST(MaxPoolTestUInt8ActRelu1) {
+TF_LITE_MICRO_TEST(MaxPoolTestUInt8ActReluN1To1) {
   using tflite::testing::F2Q;
 
   uint8_t output_data[2];
@@ -743,7 +744,7 @@ TF_LITE_MICRO_TEST(MaxPoolTestUInt8ActRelu1) {
       {// Output values
        F2Q(-1.0, output_min, output_max), F2Q(1.0, output_min, output_max)},
       output_min, output_max, {4, 1, 1, 2, 1},  // Output shape
-      kTfLitePaddingValid, kTfLiteActRelu1, output_data);
+      kTfLitePaddingValid, kTfLiteActReluN1To1, output_data);
 }
 
 TF_LITE_MICRO_TEST(MaxPoolTestUInt8ActRelu6) {
@@ -944,7 +945,7 @@ TF_LITE_MICRO_TEST(MaxPoolTestUInt8ActRelu) {
       kTfLitePaddingValid, kTfLiteActRelu, output_data);
 }
 
-TF_LITE_MICRO_TEST(MaxPoolTestUInt8ActRelu1) {
+TF_LITE_MICRO_TEST(MaxPoolTestUInt8ActReluN1To1) {
   using tflite::testing::F2QS;
 
   int8_t output_data[2];
@@ -974,7 +975,7 @@ TF_LITE_MICRO_TEST(MaxPoolTestUInt8ActRelu1) {
       {// Output values
        F2QS(-1.0, output_min, output_max), F2QS(1.0, output_min, output_max)},
       output_min, output_max, {4, 1, 1, 2, 1},  // Output shape
-      kTfLitePaddingValid, kTfLiteActRelu1, output_data);
+      kTfLitePaddingValid, kTfLiteActReluN1To1, output_data);
 }
 
 TF_LITE_MICRO_TEST(MaxPoolTestUInt8ActRelu6) {
diff --git a/tensorflow/lite/micro/kernels/sub_test.cc b/tensorflow/lite/micro/kernels/sub_test.cc
index d6ab48ead36..b8de6eba453 100644
--- a/tensorflow/lite/micro/kernels/sub_test.cc
+++ b/tensorflow/lite/micro/kernels/sub_test.cc
@@ -201,7 +201,7 @@ TF_LITE_MICRO_TEST(FloatSubActivationRelu1) {
   float output_data[output_dims_count];
   tflite::testing::TestSubFloat(inout_shape, input1_values, inout_shape,
                                 input2_values, inout_shape, golden_values,
-                                kTfLiteActRelu1, output_data);
+                                kTfLiteActReluN1To1, output_data);
 }
 
 TF_LITE_MICRO_TEST(FloatSubVariousInputShapes) {
@@ -313,7 +313,7 @@ TF_LITE_MICRO_TEST(QuantizedSubActivationRelu1Uint8) {
       inout_shape, input1_values, input1_quantized, scales[0], zero_points[0],
       inout_shape, input2_values, input2_quantized, scales[1], zero_points[1],
       inout_shape, golden_values, golden_quantized, scales[2], zero_points[2],
-      kTfLiteActRelu1, output);
+      kTfLiteActReluN1To1, output);
 }
 
 TF_LITE_MICRO_TEST(QuantizedSubActivationRelu1Int8) {
@@ -334,7 +334,7 @@ TF_LITE_MICRO_TEST(QuantizedSubActivationRelu1Int8) {
       inout_shape, input1_values, input1_quantized, scales[0], zero_points[0],
       inout_shape, input2_values, input2_quantized, scales[1], zero_points[1],
       inout_shape, golden_values, golden_quantized, scales[2], zero_points[2],
-      kTfLiteActRelu1, output);
+      kTfLiteActReluN1To1, output);
 }
 
 TF_LITE_MICRO_TEST(QuantizedSubVariousInputShapesUint8) {

From 27c8448a544840463c54a749bdefe440b52e4ab7 Mon Sep 17 00:00:00 2001
From: "A. Unique TensorFlower" 
Date: Sun, 21 Jun 2020 21:58:35 -0700
Subject: [PATCH 405/867] tf.numpy: Add support for functions in np.random
 module. Remove np.bfloat16 and add some more dtypes.

PiperOrigin-RevId: 317591184
Change-Id: Iea428ad85119233a66ba04a8c9e7e41908ce23bf
---
 tensorflow/python/ops/numpy_ops/np_dtypes.py  |  23 ++--
 tensorflow/python/ops/numpy_ops/np_random.py  |  75 +++++++++--
 .../python/ops/numpy_ops/np_random_test.py    | 126 +++++++++++++++++-
 3 files changed, 197 insertions(+), 27 deletions(-)

diff --git a/tensorflow/python/ops/numpy_ops/np_dtypes.py b/tensorflow/python/ops/numpy_ops/np_dtypes.py
index cefd09f800d..6f335773ab7 100644
--- a/tensorflow/python/ops/numpy_ops/np_dtypes.py
+++ b/tensorflow/python/ops/numpy_ops/np_dtypes.py
@@ -26,33 +26,32 @@ import numpy as np
 # pylint: disable=unused-import
 # pylint: disable=g-bad-import-order
 from numpy import bool_
+from numpy import complex_
+from numpy import complex128
+from numpy import complex64
+from numpy import float_
+from numpy import float16
+from numpy import float32
+from numpy import float64
+from numpy import inexact
 from numpy import int_
 from numpy import int16
 from numpy import int32
 from numpy import int64
 from numpy import int8
+from numpy import object_
+from numpy import string_
 from numpy import uint16
 from numpy import uint32
 from numpy import uint64
 from numpy import uint8
-from numpy import float_
-from numpy import float16
-from numpy import float32
-from numpy import float64
-from numpy import complex_
-from numpy import complex64
-from numpy import complex128
-
-from numpy import inexact
+from numpy import unicode_
 
 from numpy import iinfo
 from numpy import issubdtype
 
 from numpy import inf
 
-# TODO(wangpeng): Make bfloat16 a numpy dtype instead of using TF's
-from tensorflow.python.framework.dtypes import bfloat16
-# pylint: enable=g-bad-import-order
 # pylint: enable=unused-import
 
 _to_float32 = {
diff --git a/tensorflow/python/ops/numpy_ops/np_random.py b/tensorflow/python/ops/numpy_ops/np_random.py
index 801a7549b97..a7556dbddd4 100644
--- a/tensorflow/python/ops/numpy_ops/np_random.py
+++ b/tensorflow/python/ops/numpy_ops/np_random.py
@@ -13,19 +13,44 @@
 # limitations under the License.
 # ==============================================================================
 """Random functions."""
+
+# pylint: disable=g-direct-tensorflow-import
+
 from __future__ import absolute_import
 from __future__ import division
 from __future__ import print_function
 
-import numpy as np
+import numpy as onp
 
 from tensorflow.python.framework import random_seed
+from tensorflow.python.ops import array_ops
 from tensorflow.python.ops import random_ops
+from tensorflow.python.ops.numpy_ops import np_array_ops
+from tensorflow.python.ops.numpy_ops import np_dtypes
 from tensorflow.python.ops.numpy_ops import np_utils
 
-DEFAULT_RANDN_DTYPE = np.float32
+# TODO(agarwal): deprecate this.
+DEFAULT_RANDN_DTYPE = onp.float32
 
 
+@np_utils.np_doc('random.seed')
+def seed(s):
+  """Sets the seed for the random number generator.
+
+  Uses `tf.set_random_seed`.
+
+  Args:
+    s: an integer.
+  """
+  try:
+    s = int(s)
+  except TypeError:
+    # TODO(wangpeng): support this?
+    raise ValueError('np.seed currently only support integer arguments.')
+  random_seed.set_seed(s)
+
+
+@np_utils.np_doc('random.randn')
 def randn(*args):
   """Returns samples from a normal distribution.
 
@@ -40,17 +65,45 @@ def randn(*args):
   # TODO(wangpeng): Use new stateful RNG
   if np_utils.isscalar(args):
     args = (args,)
+  dtype = np_dtypes.default_float_type()
+  return np_utils.tensor_to_ndarray(random_ops.random_normal(args, dtype=dtype))
+
+
+@np_utils.np_doc('random.uniform')
+def uniform(low=0.0, high=1.0, size=None):
+  dtype = np_dtypes.default_float_type()
+  low = np_array_ops.asarray(low, dtype=dtype)
+  high = np_array_ops.asarray(high, dtype=dtype)
+  if size is None:
+    size = array_ops.broadcast_dynamic_shape(low.shape, high.shape)
   return np_utils.tensor_to_ndarray(
-      random_ops.random_normal(args, dtype=DEFAULT_RANDN_DTYPE))
+      random_ops.random_uniform(
+          shape=size, minval=low, maxval=high, dtype=dtype))
 
 
-def seed(s):
-  """Sets the seed for the random number generator.
+@np_utils.np_doc('random.random')
+def random(size=None):
+  return uniform(0., 1., size)
 
-  Uses `tf.set_random_seed`.
 
-  Args:
-    s: an integer.
-  """
-  # TODO(wangpeng): make the signature the same as numpy
-  random_seed.set_seed(s)
+@np_utils.np_doc('random.rand')
+def rand(*size):
+  return uniform(0., 1., size)
+
+
+@np_utils.np_doc('random.randint')
+def randint(low, high=None, size=None, dtype=onp.int):  # pylint: disable=missing-function-docstring
+  low = int(low)
+  if high is None:
+    high = low
+    low = 0
+  if size is None:
+    size = ()
+  elif isinstance(size, int):
+    size = (size,)
+  dtype = np_utils.result_type(dtype)
+  if dtype not in (onp.int32, onp.int64):
+    raise ValueError('Only np.int32 or np.int64 types are supported')
+  return np_utils.tensor_to_ndarray(
+      random_ops.random_uniform(
+          shape=size, minval=low, maxval=high, dtype=dtype))
diff --git a/tensorflow/python/ops/numpy_ops/np_random_test.py b/tensorflow/python/ops/numpy_ops/np_random_test.py
index 3423b2234e8..95bc7606e1a 100644
--- a/tensorflow/python/ops/numpy_ops/np_random_test.py
+++ b/tensorflow/python/ops/numpy_ops/np_random_test.py
@@ -13,22 +13,138 @@
 # limitations under the License.
 # ==============================================================================
 """Tests for tf numpy random number methods."""
+# pylint: disable=g-direct-tensorflow-import
 
 from __future__ import absolute_import
 from __future__ import division
 from __future__ import print_function
 
-import numpy as np
+from absl.testing import parameterized
+import numpy as onp
 from six.moves import range
 
 from tensorflow.python.framework import ops
+from tensorflow.python.ops import numpy_ops as np
 # Needed for ndarray.reshape.
 from tensorflow.python.ops.numpy_ops import np_array_ops  # pylint: disable=unused-import
+from tensorflow.python.ops.numpy_ops import np_dtypes
 from tensorflow.python.ops.numpy_ops import np_random
 from tensorflow.python.platform import test
 
 
-class RandomTest(test.TestCase):
+class SeedTest(test.TestCase):
+
+  def test(self):
+    np.random.seed(1)
+    np.random.seed(np.int32(1))
+    with self.assertRaises(ValueError):
+      np.random.seed((1, 3))
+
+
+class RandomTestBase(test.TestCase, parameterized.TestCase):
+
+  def _test(self, *args, **kw_args):
+    onp_dtype = kw_args.pop('onp_dtype', None)
+    allow_float64 = kw_args.pop('allow_float64', True)
+    old_allow_float64 = np_dtypes.is_allow_float64()
+    np_dtypes.set_allow_float64(allow_float64)
+    old_func = getattr(self, 'onp_func', None)
+    # TODO(agarwal): Note that onp can return a scalar type while np returns
+    # ndarrays. Currently np does not support scalar types.
+    self.onp_func = lambda *args, **kwargs: onp.asarray(  # pylint: disable=g-long-lambda
+        old_func(*args, **kwargs))
+    np_out = self.np_func(*args, **kw_args)
+    onp_out = onp.asarray(self.onp_func(*args, **kw_args))
+    if onp_dtype is not None:
+      onp_out = onp_out.astype(onp_dtype)
+    self.assertEqual(np_out.shape, onp_out.shape)
+    self.assertEqual(np_out.dtype, onp_out.dtype)
+    np_dtypes.set_allow_float64(old_allow_float64)
+
+
+class RandNTest(RandomTestBase):
+
+  def setUp(self):
+    self.np_func = np.random.randn
+    self.onp_func = onp.random.randn
+    super(RandNTest, self).setUp()
+
+  @parameterized.parameters((), (2), (2, 3))
+  def test_float64(self, *dims):
+    self._test(*dims)
+
+  @parameterized.parameters((), (2), ((2,)), (2, 3))
+  def test_float32(self, *dims):
+    self._test(*dims, allow_float64=False, onp_dtype=np.float32)
+
+
+class UniformTest(RandomTestBase):
+
+  def setUp(self):
+    self.np_func = np.random.uniform
+    self.onp_func = onp.random.uniform
+    super(UniformTest, self).setUp()
+
+  @parameterized.parameters(
+      ((), (), None),
+      (1, (), None),
+      ((), 1, None),
+      (1, 1, None),
+      ((1, 2), (2, 1), None),
+      ((1, 2, 1), (2, 1, 1), (2, 2, 2)),
+      ((), (), (2, 2, 2)),
+  )
+  def test_broadcast(self, low_shape, high_shape, size):
+    low = np.zeros(low_shape).astype(np.float64)
+    high = np.ones(high_shape).astype(np.float64)
+    self._test(low=low, high=high, size=size)
+
+  def test_float32(self):
+    self._test(0, 1, (1, 2), allow_float64=False, onp_dtype=np.float32)
+
+  def test_dtype_cast(self):
+    self._test(np.int8(0), np.uint8(1), (1, 2))
+
+
+class RandomTest(RandomTestBase):
+
+  def setUp(self):
+    self.np_func = np.random.random
+    self.onp_func = onp.random.random
+    super(RandomTest, self).setUp()
+
+  @parameterized.parameters((None,), ((),), ((1,),), ((1, 2),))
+  def test(self, size):
+    self._test(size)
+
+
+class RandTest(RandomTestBase):
+
+  def setUp(self):
+    self.np_func = np.random.rand
+    self.onp_func = onp.random.rand
+    super(RandTest, self).setUp()
+
+  @parameterized.parameters((), (1,), (1, 2))
+  def test(self, *size):
+    self._test(*size)
+
+
+class RandIntTest(RandomTestBase):
+
+  def setUp(self):
+    self.np_func = np.random.randint
+    self.onp_func = onp.random.randint
+    super(RandIntTest, self).setUp()
+
+  @parameterized.parameters((0, 1, None, 'l'), (0, 1, None, np.int64),
+                            (0, 1, 2, np.int32), (0, 1, (), np.int32),
+                            (0, 1, (2), np.int64), (0, 1, (2, 2), 'l'))
+  def test(self, low, high, size, dtype):
+    self._test(low, high, size=size, dtype=dtype)
+
+
+class RandNDistriutionTest(test.TestCase):
 
   def assertNotAllClose(self, a, b, **kwargs):
     try:
@@ -38,7 +154,7 @@ class RandomTest(test.TestCase):
     raise AssertionError('The two values are close at all %d elements' %
                          np.size(a))
 
-  def testRandN(self):
+  def testDistribution(self):
 
     def run_test(*args):
       num_samples = 1000
@@ -50,7 +166,9 @@ class RandomTest(test.TestCase):
       # Test output shape.
       for output in outputs:
         self.assertEqual(output.shape, tuple(args))
-        self.assertEqual(output.dtype.type, np_random.DEFAULT_RANDN_DTYPE)
+        default_dtype = (
+            np.float64 if np_dtypes.is_allow_float64() else np.float32)
+        self.assertEqual(output.dtype.type, default_dtype)
 
       if np.prod(args):  # Don't bother with empty arrays.
         outputs = [output.tolist() for output in outputs]

From 8fc628cf78066539e0ed6e90d93169c231e5a3d7 Mon Sep 17 00:00:00 2001
From: Marat Dukhan 
Date: Sun, 21 Jun 2020 23:01:26 -0700
Subject: [PATCH 406/867] Add unit test for Prelu in XNNPACK delegate

- Add PreluTester class and unit test for XNNPACK-delegated Prelu operator
- Relax restrictions on the number of input/output dimensions in delegated
  Prelu operators

PiperOrigin-RevId: 317596110
Change-Id: I99c97c45212e80ec600c4644b858cd15c3d34220
---
 tensorflow/lite/delegates/xnnpack/BUILD       |  32 +
 .../delegates/xnnpack/leaky_relu_tester.h     |   1 -
 .../lite/delegates/xnnpack/prelu_test.cc      | 565 ++++++++++++++++++
 .../lite/delegates/xnnpack/prelu_tester.cc    | 237 ++++++++
 .../lite/delegates/xnnpack/prelu_tester.h     |  88 +++
 .../delegates/xnnpack/xnnpack_delegate.cc     |   6 +-
 6 files changed, 926 insertions(+), 3 deletions(-)
 create mode 100644 tensorflow/lite/delegates/xnnpack/prelu_test.cc
 create mode 100644 tensorflow/lite/delegates/xnnpack/prelu_tester.cc
 create mode 100644 tensorflow/lite/delegates/xnnpack/prelu_tester.h

diff --git a/tensorflow/lite/delegates/xnnpack/BUILD b/tensorflow/lite/delegates/xnnpack/BUILD
index eaf7d8f6f03..e0d3d39f719 100644
--- a/tensorflow/lite/delegates/xnnpack/BUILD
+++ b/tensorflow/lite/delegates/xnnpack/BUILD
@@ -180,6 +180,23 @@ cc_library(
     ],
 )
 
+cc_library(
+    name = "prelu_tester",
+    testonly = 1,
+    srcs = ["prelu_tester.cc"],
+    hdrs = ["prelu_tester.h"],
+    deps = [
+        "//tensorflow/lite:framework",
+        "//tensorflow/lite:schema_fbs_version",
+        "//tensorflow/lite/c:common",
+        "//tensorflow/lite/kernels:builtin_ops",
+        "//tensorflow/lite/schema:schema_fbs",
+        "@FP16",
+        "@com_google_googletest//:gtest",
+        "@flatbuffers",
+    ],
+)
+
 cc_library(
     name = "reduce_tester",
     testonly = 1,
@@ -527,6 +544,21 @@ cc_test(
     ],
 )
 
+cc_test(
+    name = "prelu_test",
+    srcs = ["prelu_test.cc"],
+    linkopts = select({
+        "//tensorflow:emscripten": EMSCRIPTEN_LINKOPTS,
+        "//conditions:default": [],
+    }),
+    deps = [
+        ":prelu_tester",
+        ":test_main",
+        ":xnnpack_delegate_test_mode",
+        "@com_google_googletest//:gtest",
+    ],
+)
+
 cc_test(
     name = "relu_test",
     srcs = ["relu_test.cc"],
diff --git a/tensorflow/lite/delegates/xnnpack/leaky_relu_tester.h b/tensorflow/lite/delegates/xnnpack/leaky_relu_tester.h
index f1d9efd7209..191dc938e89 100644
--- a/tensorflow/lite/delegates/xnnpack/leaky_relu_tester.h
+++ b/tensorflow/lite/delegates/xnnpack/leaky_relu_tester.h
@@ -21,7 +21,6 @@ limitations under the License.
 
 #include 
 #include "tensorflow/lite/c/common.h"
-#include "tensorflow/lite/schema/schema_generated.h"
 
 namespace tflite {
 namespace xnnpack {
diff --git a/tensorflow/lite/delegates/xnnpack/prelu_test.cc b/tensorflow/lite/delegates/xnnpack/prelu_test.cc
new file mode 100644
index 00000000000..ae40c032f32
--- /dev/null
+++ b/tensorflow/lite/delegates/xnnpack/prelu_test.cc
@@ -0,0 +1,565 @@
+/* Copyright 2020 The TensorFlow Authors. All Rights Reserved.
+
+Licensed under the Apache License, Version 2.0 (the "License");
+you may not use this file except in compliance with the License.
+You may obtain a copy of the License at
+
+    http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software
+distributed under the License is distributed on an "AS IS" BASIS,
+WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+See the License for the specific language governing permissions and
+limitations under the License.
+==============================================================================*/
+
+#include 
+#include 
+#include 
+#include 
+
+#include 
+#include "tensorflow/lite/delegates/xnnpack/prelu_tester.h"
+#include "tensorflow/lite/delegates/xnnpack/xnnpack_delegate.h"
+
+namespace tflite {
+namespace xnnpack {
+
+TEST(Prelu, DISABLED_4DBy4D) {
+  std::unique_ptr
+      xnnpack_delegate(TfLiteXNNPackDelegateCreate(nullptr),
+                       TfLiteXNNPackDelegateDelete);
+
+  std::random_device random_device;
+  auto rng = std::mt19937(random_device());
+  auto shape_rng =
+      std::bind(std::uniform_int_distribution(2, 5), std::ref(rng));
+  const auto batch = shape_rng();
+  const auto height = shape_rng();
+  const auto width = shape_rng();
+  const auto channels = shape_rng();
+
+  PreluTester()
+      .InputShape({batch, height, width, channels})
+      .SlopeShape({batch, height, width, channels})
+      .Test(xnnpack_delegate.get());
+}
+
+TEST(Prelu, 4DBy4DBroadcastChannels) {
+  std::unique_ptr
+      xnnpack_delegate(TfLiteXNNPackDelegateCreate(nullptr),
+                       TfLiteXNNPackDelegateDelete);
+
+  std::random_device random_device;
+  auto rng = std::mt19937(random_device());
+  auto shape_rng =
+      std::bind(std::uniform_int_distribution(2, 5), std::ref(rng));
+  const auto batch = shape_rng();
+  const auto height = shape_rng();
+  const auto width = shape_rng();
+  const auto channels = shape_rng();
+
+  PreluTester()
+      .InputShape({batch, height, width, channels})
+      .SlopeShape({1, 1, 1, channels})
+      .Test(xnnpack_delegate.get());
+}
+
+TEST(Prelu, DISABLED_4DBy4DBroadcastWidth) {
+  std::unique_ptr
+      xnnpack_delegate(TfLiteXNNPackDelegateCreate(nullptr),
+                       TfLiteXNNPackDelegateDelete);
+
+  std::random_device random_device;
+  auto rng = std::mt19937(random_device());
+  auto shape_rng =
+      std::bind(std::uniform_int_distribution(2, 5), std::ref(rng));
+  const auto batch = shape_rng();
+  const auto height = shape_rng();
+  const auto width = shape_rng();
+  const auto channels = shape_rng();
+
+  PreluTester()
+      .InputShape({batch, height, width, channels})
+      .SlopeShape({1, 1, width, 1})
+      .Test(xnnpack_delegate.get());
+}
+
+TEST(Prelu, DISABLED_4DBy4DBroadcastHeight) {
+  std::unique_ptr
+      xnnpack_delegate(TfLiteXNNPackDelegateCreate(nullptr),
+                       TfLiteXNNPackDelegateDelete);
+
+  std::random_device random_device;
+  auto rng = std::mt19937(random_device());
+  auto shape_rng =
+      std::bind(std::uniform_int_distribution(2, 5), std::ref(rng));
+  const auto batch = shape_rng();
+  const auto height = shape_rng();
+  const auto width = shape_rng();
+  const auto channels = shape_rng();
+
+  PreluTester()
+      .InputShape({batch, height, width, channels})
+      .SlopeShape({1, height, 1, 1})
+      .Test(xnnpack_delegate.get());
+}
+
+TEST(Prelu, DISABLED_4DBy4DBroadcastBatch) {
+  std::unique_ptr
+      xnnpack_delegate(TfLiteXNNPackDelegateCreate(nullptr),
+                       TfLiteXNNPackDelegateDelete);
+
+  std::random_device random_device;
+  auto rng = std::mt19937(random_device());
+  auto shape_rng =
+      std::bind(std::uniform_int_distribution(2, 5), std::ref(rng));
+  const auto batch = shape_rng();
+  const auto height = shape_rng();
+  const auto width = shape_rng();
+  const auto channels = shape_rng();
+
+  PreluTester()
+      .InputShape({batch, height, width, channels})
+      .SlopeShape({batch, 1, 1, 1})
+      .Test(xnnpack_delegate.get());
+}
+
+TEST(Prelu, DISABLED_4DBy4DBroadcastHeightWidthChannels) {
+  std::unique_ptr
+      xnnpack_delegate(TfLiteXNNPackDelegateCreate(nullptr),
+                       TfLiteXNNPackDelegateDelete);
+
+  std::random_device random_device;
+  auto rng = std::mt19937(random_device());
+  auto shape_rng =
+      std::bind(std::uniform_int_distribution(2, 5), std::ref(rng));
+  const auto batch = shape_rng();
+  const auto height = shape_rng();
+  const auto width = shape_rng();
+  const auto channels = shape_rng();
+
+  PreluTester()
+      .InputShape({batch, height, width, channels})
+      .SlopeShape({1, height, width, channels})
+      .Test(xnnpack_delegate.get());
+}
+
+TEST(Prelu, DISABLED_4DBy3D) {
+  std::unique_ptr
+      xnnpack_delegate(TfLiteXNNPackDelegateCreate(nullptr),
+                       TfLiteXNNPackDelegateDelete);
+
+  std::random_device random_device;
+  auto rng = std::mt19937(random_device());
+  auto shape_rng =
+      std::bind(std::uniform_int_distribution(2, 5), std::ref(rng));
+  const auto batch = shape_rng();
+  const auto height = shape_rng();
+  const auto width = shape_rng();
+  const auto channels = shape_rng();
+
+  PreluTester()
+      .InputShape({batch, height, width, channels})
+      .SlopeShape({height, width, channels})
+      .Test(xnnpack_delegate.get());
+}
+
+TEST(Prelu, DISABLED_4DBy2D) {
+  std::unique_ptr
+      xnnpack_delegate(TfLiteXNNPackDelegateCreate(nullptr),
+                       TfLiteXNNPackDelegateDelete);
+
+  std::random_device random_device;
+  auto rng = std::mt19937(random_device());
+  auto shape_rng =
+      std::bind(std::uniform_int_distribution(2, 5), std::ref(rng));
+  const auto batch = shape_rng();
+  const auto height = shape_rng();
+  const auto width = shape_rng();
+  const auto channels = shape_rng();
+
+  PreluTester()
+      .InputShape({batch, height, width, channels})
+      .SlopeShape({width, channels})
+      .Test(xnnpack_delegate.get());
+}
+
+TEST(Prelu, 4DBy1D) {
+  std::unique_ptr
+      xnnpack_delegate(TfLiteXNNPackDelegateCreate(nullptr),
+                       TfLiteXNNPackDelegateDelete);
+
+  std::random_device random_device;
+  auto rng = std::mt19937(random_device());
+  auto shape_rng =
+      std::bind(std::uniform_int_distribution(2, 5), std::ref(rng));
+  const auto batch = shape_rng();
+  const auto height = shape_rng();
+  const auto width = shape_rng();
+  const auto channels = shape_rng();
+
+  PreluTester()
+      .InputShape({batch, height, width, channels})
+      .SlopeShape({channels})
+      .Test(xnnpack_delegate.get());
+}
+
+TEST(Prelu, DISABLED_4DBy0D) {
+  std::unique_ptr
+      xnnpack_delegate(TfLiteXNNPackDelegateCreate(nullptr),
+                       TfLiteXNNPackDelegateDelete);
+
+  std::random_device random_device;
+  auto rng = std::mt19937(random_device());
+  auto shape_rng =
+      std::bind(std::uniform_int_distribution(2, 5), std::ref(rng));
+  const auto batch = shape_rng();
+  const auto height = shape_rng();
+  const auto width = shape_rng();
+  const auto channels = shape_rng();
+
+  PreluTester()
+      .InputShape({batch, height, width, channels})
+      .SlopeShape({})
+      .Test(xnnpack_delegate.get());
+}
+
+TEST(Prelu, DISABLED_3DBy3D) {
+  std::unique_ptr
+      xnnpack_delegate(TfLiteXNNPackDelegateCreate(nullptr),
+                       TfLiteXNNPackDelegateDelete);
+
+  std::random_device random_device;
+  auto rng = std::mt19937(random_device());
+  auto shape_rng =
+      std::bind(std::uniform_int_distribution(2, 5), std::ref(rng));
+  const auto batch = shape_rng();
+  const auto width = shape_rng();
+  const auto channels = shape_rng();
+
+  PreluTester()
+      .InputShape({batch, width, channels})
+      .SlopeShape({batch, width, channels})
+      .Test(xnnpack_delegate.get());
+}
+
+TEST(Prelu, 3DBy3DBroadcastChannels) {
+  std::unique_ptr
+      xnnpack_delegate(TfLiteXNNPackDelegateCreate(nullptr),
+                       TfLiteXNNPackDelegateDelete);
+
+  std::random_device random_device;
+  auto rng = std::mt19937(random_device());
+  auto shape_rng =
+      std::bind(std::uniform_int_distribution(2, 5), std::ref(rng));
+  const auto batch = shape_rng();
+  const auto width = shape_rng();
+  const auto channels = shape_rng();
+
+  PreluTester()
+      .InputShape({batch, width, channels})
+      .SlopeShape({1, 1, channels})
+      .Test(xnnpack_delegate.get());
+}
+
+TEST(Prelu, DISABLED_3DBy3DBroadcastWidth) {
+  std::unique_ptr
+      xnnpack_delegate(TfLiteXNNPackDelegateCreate(nullptr),
+                       TfLiteXNNPackDelegateDelete);
+
+  std::random_device random_device;
+  auto rng = std::mt19937(random_device());
+  auto shape_rng =
+      std::bind(std::uniform_int_distribution(2, 5), std::ref(rng));
+  const auto batch = shape_rng();
+  const auto width = shape_rng();
+  const auto channels = shape_rng();
+
+  PreluTester()
+      .InputShape({batch, width, channels})
+      .SlopeShape({1, width, 1})
+      .Test(xnnpack_delegate.get());
+}
+
+TEST(Prelu, DISABLED_3DBy3DBroadcastBatch) {
+  std::unique_ptr
+      xnnpack_delegate(TfLiteXNNPackDelegateCreate(nullptr),
+                       TfLiteXNNPackDelegateDelete);
+
+  std::random_device random_device;
+  auto rng = std::mt19937(random_device());
+  auto shape_rng =
+      std::bind(std::uniform_int_distribution(2, 5), std::ref(rng));
+  const auto batch = shape_rng();
+  const auto width = shape_rng();
+  const auto channels = shape_rng();
+
+  PreluTester()
+      .InputShape({batch, width, channels})
+      .SlopeShape({batch, 1, 1})
+      .Test(xnnpack_delegate.get());
+}
+
+TEST(Prelu, DISABLED_3DBy3DBroadcastWidthChannels) {
+  std::unique_ptr
+      xnnpack_delegate(TfLiteXNNPackDelegateCreate(nullptr),
+                       TfLiteXNNPackDelegateDelete);
+
+  std::random_device random_device;
+  auto rng = std::mt19937(random_device());
+  auto shape_rng =
+      std::bind(std::uniform_int_distribution(2, 5), std::ref(rng));
+  const auto batch = shape_rng();
+  const auto width = shape_rng();
+  const auto channels = shape_rng();
+
+  PreluTester()
+      .InputShape({batch, width, channels})
+      .SlopeShape({1, width, channels})
+      .Test(xnnpack_delegate.get());
+}
+
+TEST(Prelu, DISABLED_3DBy2D) {
+  std::unique_ptr
+      xnnpack_delegate(TfLiteXNNPackDelegateCreate(nullptr),
+                       TfLiteXNNPackDelegateDelete);
+
+  std::random_device random_device;
+  auto rng = std::mt19937(random_device());
+  auto shape_rng =
+      std::bind(std::uniform_int_distribution(2, 5), std::ref(rng));
+  const auto batch = shape_rng();
+  const auto width = shape_rng();
+  const auto channels = shape_rng();
+
+  PreluTester()
+      .InputShape({batch, width, channels})
+      .SlopeShape({width, channels})
+      .Test(xnnpack_delegate.get());
+}
+
+TEST(Prelu, 3DBy1D) {
+  std::unique_ptr
+      xnnpack_delegate(TfLiteXNNPackDelegateCreate(nullptr),
+                       TfLiteXNNPackDelegateDelete);
+
+  std::random_device random_device;
+  auto rng = std::mt19937(random_device());
+  auto shape_rng =
+      std::bind(std::uniform_int_distribution(2, 5), std::ref(rng));
+  const auto batch = shape_rng();
+  const auto width = shape_rng();
+  const auto channels = shape_rng();
+
+  PreluTester()
+      .InputShape({batch, width, channels})
+      .SlopeShape({channels})
+      .Test(xnnpack_delegate.get());
+}
+
+TEST(Prelu, DISABLED_3DBy0D) {
+  std::unique_ptr
+      xnnpack_delegate(TfLiteXNNPackDelegateCreate(nullptr),
+                       TfLiteXNNPackDelegateDelete);
+
+  std::random_device random_device;
+  auto rng = std::mt19937(random_device());
+  auto shape_rng =
+      std::bind(std::uniform_int_distribution(2, 5), std::ref(rng));
+  const auto batch = shape_rng();
+  const auto width = shape_rng();
+  const auto channels = shape_rng();
+
+  PreluTester()
+      .InputShape({batch, width, channels})
+      .SlopeShape({})
+      .Test(xnnpack_delegate.get());
+}
+
+TEST(Prelu, DISABLED_2DBy2D) {
+  std::unique_ptr
+      xnnpack_delegate(TfLiteXNNPackDelegateCreate(nullptr),
+                       TfLiteXNNPackDelegateDelete);
+
+  std::random_device random_device;
+  auto rng = std::mt19937(random_device());
+  auto shape_rng =
+      std::bind(std::uniform_int_distribution(2, 5), std::ref(rng));
+  const auto batch = shape_rng();
+  const auto channels = shape_rng();
+
+  PreluTester()
+      .InputShape({batch, channels})
+      .SlopeShape({batch, channels})
+      .Test(xnnpack_delegate.get());
+}
+
+TEST(Prelu, 2DBy2DBroadcastChannels) {
+  std::unique_ptr
+      xnnpack_delegate(TfLiteXNNPackDelegateCreate(nullptr),
+                       TfLiteXNNPackDelegateDelete);
+
+  std::random_device random_device;
+  auto rng = std::mt19937(random_device());
+  auto shape_rng =
+      std::bind(std::uniform_int_distribution(2, 5), std::ref(rng));
+  const auto batch = shape_rng();
+  const auto channels = shape_rng();
+
+  PreluTester()
+      .InputShape({batch, channels})
+      .SlopeShape({1, channels})
+      .Test(xnnpack_delegate.get());
+}
+
+TEST(Prelu, DISABLED_2DBy2DBroadcastBatch) {
+  std::unique_ptr
+      xnnpack_delegate(TfLiteXNNPackDelegateCreate(nullptr),
+                       TfLiteXNNPackDelegateDelete);
+
+  std::random_device random_device;
+  auto rng = std::mt19937(random_device());
+  auto shape_rng =
+      std::bind(std::uniform_int_distribution(2, 5), std::ref(rng));
+  const auto batch = shape_rng();
+  const auto channels = shape_rng();
+
+  PreluTester()
+      .InputShape({batch, channels})
+      .SlopeShape({batch, 1})
+      .Test(xnnpack_delegate.get());
+}
+
+TEST(Prelu, 2DBy1D) {
+  std::unique_ptr
+      xnnpack_delegate(TfLiteXNNPackDelegateCreate(nullptr),
+                       TfLiteXNNPackDelegateDelete);
+
+  std::random_device random_device;
+  auto rng = std::mt19937(random_device());
+  auto shape_rng =
+      std::bind(std::uniform_int_distribution(2, 5), std::ref(rng));
+  const auto batch = shape_rng();
+  const auto channels = shape_rng();
+
+  PreluTester()
+      .InputShape({batch, channels})
+      .SlopeShape({channels})
+      .Test(xnnpack_delegate.get());
+}
+
+TEST(Prelu, DISABLED_2DBy0D) {
+  std::unique_ptr
+      xnnpack_delegate(TfLiteXNNPackDelegateCreate(nullptr),
+                       TfLiteXNNPackDelegateDelete);
+
+  std::random_device random_device;
+  auto rng = std::mt19937(random_device());
+  auto shape_rng =
+      std::bind(std::uniform_int_distribution(2, 5), std::ref(rng));
+  const auto batch = shape_rng();
+  const auto channels = shape_rng();
+
+  PreluTester()
+      .InputShape({batch, channels})
+      .SlopeShape({})
+      .Test(xnnpack_delegate.get());
+}
+
+TEST(Prelu, 1DBy1D) {
+  std::unique_ptr
+      xnnpack_delegate(TfLiteXNNPackDelegateCreate(nullptr),
+                       TfLiteXNNPackDelegateDelete);
+
+  std::random_device random_device;
+  auto rng = std::mt19937(random_device());
+  auto shape_rng =
+      std::bind(std::uniform_int_distribution(2, 5), std::ref(rng));
+  const auto batch = shape_rng();
+
+  PreluTester().InputShape({batch}).SlopeShape({batch}).Test(
+      xnnpack_delegate.get());
+}
+
+TEST(Prelu, DISABLED_1DBy0D) {
+  std::unique_ptr
+      xnnpack_delegate(TfLiteXNNPackDelegateCreate(nullptr),
+                       TfLiteXNNPackDelegateDelete);
+
+  std::random_device random_device;
+  auto rng = std::mt19937(random_device());
+  auto shape_rng =
+      std::bind(std::uniform_int_distribution(2, 5), std::ref(rng));
+  const auto batch = shape_rng();
+
+  PreluTester().InputShape({batch}).SlopeShape({}).Test(xnnpack_delegate.get());
+}
+
+TEST(Prelu, FP16Weights) {
+  std::unique_ptr
+      xnnpack_delegate(TfLiteXNNPackDelegateCreate(nullptr),
+                       TfLiteXNNPackDelegateDelete);
+
+  std::random_device random_device;
+  auto rng = std::mt19937(random_device());
+  auto shape_rng =
+      std::bind(std::uniform_int_distribution(2, 5), std::ref(rng));
+  const auto batch = shape_rng();
+  const auto height = shape_rng();
+  const auto width = shape_rng();
+  const auto channels = shape_rng();
+
+  PreluTester()
+      .InputShape({batch, height, width, channels})
+      .SlopeShape({channels})
+      .FP16Weights()
+      .Test(xnnpack_delegate.get());
+}
+
+TEST(Prelu, SparseWeights) {
+  std::unique_ptr
+      xnnpack_delegate(TfLiteXNNPackDelegateCreate(nullptr),
+                       TfLiteXNNPackDelegateDelete);
+
+  std::random_device random_device;
+  auto rng = std::mt19937(random_device());
+  auto shape_rng =
+      std::bind(std::uniform_int_distribution(2, 5), std::ref(rng));
+  const auto batch = shape_rng();
+  const auto height = shape_rng();
+  const auto width = shape_rng();
+  const auto channels = shape_rng();
+
+  PreluTester()
+      .InputShape({batch, height, width, channels})
+      .SlopeShape({channels})
+      .SparseWeights()
+      .Test(xnnpack_delegate.get());
+}
+
+TEST(Prelu, MultiThreading) {
+  TfLiteXNNPackDelegateOptions delegate_options =
+      TfLiteXNNPackDelegateOptionsDefault();
+  delegate_options.num_threads = 2;
+  std::unique_ptr
+      xnnpack_delegate(TfLiteXNNPackDelegateCreate(&delegate_options),
+                       TfLiteXNNPackDelegateDelete);
+
+  std::random_device random_device;
+  auto rng = std::mt19937(random_device());
+  auto shape_rng =
+      std::bind(std::uniform_int_distribution(2, 5), std::ref(rng));
+  const auto batch = shape_rng();
+  const auto height = shape_rng();
+  const auto width = shape_rng();
+  const auto channels = shape_rng();
+
+  PreluTester()
+      .InputShape({batch, height, width, channels})
+      .SlopeShape({channels})
+      .Test(xnnpack_delegate.get());
+}
+
+}  // namespace xnnpack
+}  // namespace tflite
diff --git a/tensorflow/lite/delegates/xnnpack/prelu_tester.cc b/tensorflow/lite/delegates/xnnpack/prelu_tester.cc
new file mode 100644
index 00000000000..90fa0fb67c0
--- /dev/null
+++ b/tensorflow/lite/delegates/xnnpack/prelu_tester.cc
@@ -0,0 +1,237 @@
+/* Copyright 2020 The TensorFlow Authors. All Rights Reserved.
+
+Licensed under the Apache License, Version 2.0 (the "License");
+you may not use this file except in compliance with the License.
+You may obtain a copy of the License at
+
+    http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software
+distributed under the License is distributed on an "AS IS" BASIS,
+WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+See the License for the specific language governing permissions and
+limitations under the License.
+==============================================================================*/
+
+#include "tensorflow/lite/delegates/xnnpack/prelu_tester.h"
+
+#include 
+#include 
+#include 
+#include 
+#include 
+#include 
+
+#include 
+#include "third_party/FP16/include/fp16.h"
+#include "flatbuffers/flatbuffers.h"  // from @flatbuffers
+#include "tensorflow/lite/interpreter.h"
+#include "tensorflow/lite/kernels/register.h"
+#include "tensorflow/lite/model.h"
+#include "tensorflow/lite/schema/schema_generated.h"
+#include "tensorflow/lite/version.h"
+
+namespace tflite {
+namespace xnnpack {
+
+void PreluTester::Test(TfLiteDelegate* delegate) const {
+  std::random_device random_device;
+  auto rng = std::mt19937(random_device());
+  auto input_rng = std::bind(std::uniform_real_distribution(-1.0f, 1.0f),
+                             std::ref(rng));
+
+  std::vector buffer = CreateTfLiteModel();
+  const Model* model = GetModel(buffer.data());
+
+  std::unique_ptr delegate_interpreter;
+  ASSERT_EQ(
+      InterpreterBuilder(model, ::tflite::ops::builtin::BuiltinOpResolver())(
+          &delegate_interpreter),
+      kTfLiteOk);
+  std::unique_ptr default_interpreter;
+  ASSERT_EQ(
+      InterpreterBuilder(model, ::tflite::ops::builtin::BuiltinOpResolver())(
+          &default_interpreter),
+      kTfLiteOk);
+
+  ASSERT_TRUE(delegate_interpreter);
+  ASSERT_TRUE(default_interpreter);
+
+  ASSERT_EQ(delegate_interpreter->inputs().size(), 1);
+  ASSERT_EQ(default_interpreter->inputs().size(), 1);
+
+  ASSERT_EQ(delegate_interpreter->outputs().size(), 1);
+  ASSERT_EQ(default_interpreter->outputs().size(), 1);
+
+  ASSERT_EQ(delegate_interpreter->AllocateTensors(), kTfLiteOk);
+  ASSERT_EQ(default_interpreter->AllocateTensors(), kTfLiteOk);
+
+  ASSERT_EQ(delegate_interpreter->ModifyGraphWithDelegate(delegate), kTfLiteOk);
+
+  float* default_input_data = default_interpreter->typed_tensor(
+      default_interpreter->inputs()[0]);
+  std::generate(default_input_data,
+                default_input_data + ComputeSize(InputShape()),
+                std::ref(input_rng));
+
+  float* xnnpack_input_data = delegate_interpreter->typed_tensor(
+      delegate_interpreter->inputs()[0]);
+  std::copy(default_input_data, default_input_data + ComputeSize(InputShape()),
+            xnnpack_input_data);
+
+  ASSERT_EQ(default_interpreter->Invoke(), kTfLiteOk);
+  ASSERT_EQ(delegate_interpreter->Invoke(), kTfLiteOk);
+
+  float* default_output_data = default_interpreter->typed_tensor(
+      default_interpreter->outputs()[0]);
+  float* xnnpack_output_data = delegate_interpreter->typed_tensor(
+      delegate_interpreter->outputs()[0]);
+
+  for (size_t i = 0; i < ComputeSize(OutputShape()); i++) {
+    ASSERT_EQ(default_output_data[i], xnnpack_output_data[i]);
+  }
+}
+
+std::vector PreluTester::CreateTfLiteModel() const {
+  std::random_device random_device;
+  auto rng = std::mt19937(random_device());
+  auto slope_rng = std::bind(std::uniform_real_distribution(0.25f, 0.5f),
+                             std::ref(rng));
+
+  flatbuffers::FlatBufferBuilder builder;
+  std::vector> operator_codes{
+      {CreateOperatorCode(builder, BuiltinOperator_PRELU)}};
+  if (FP16Weights()) {
+    operator_codes.emplace_back(
+        CreateOperatorCode(builder, BuiltinOperator_DEQUANTIZE));
+  } else if (SparseWeights()) {
+    operator_codes.emplace_back(
+        CreateOperatorCode(builder, BuiltinOperator_DENSIFY));
+  }
+
+  std::vector> buffers{{
+      CreateBuffer(builder, builder.CreateVector({})),
+  }};
+
+  if (FP16Weights()) {
+    std::vector slope_data(ComputeSize(SlopeShape()));
+    std::generate(slope_data.begin(), slope_data.end(),
+                  std::bind(fp16_ieee_from_fp32_value, slope_rng));
+
+    buffers.push_back(CreateBuffer(
+        builder, builder.CreateVector(
+                     reinterpret_cast(slope_data.data()),
+                     sizeof(uint16_t) * slope_data.size())));
+  } else {
+    std::vector slope_data(ComputeSize(SlopeShape()));
+    std::generate(slope_data.begin(), slope_data.end(), slope_rng);
+
+    buffers.push_back(CreateBuffer(
+        builder, builder.CreateVector(
+                     reinterpret_cast(slope_data.data()),
+                     sizeof(float) * slope_data.size())));
+  }
+
+  std::vector> tensors;
+  std::vector> operators;
+  if (FP16Weights()) {
+    tensors.emplace_back(CreateTensor(
+        builder,
+        builder.CreateVector(SlopeShape().data(), SlopeShape().size()),
+        TensorType_FLOAT16, /*buffer=*/1));
+  } else if (SparseWeights()) {
+    const int dims_count = SlopeShape().size();
+    std::vector> dim_metadata(
+        dims_count);
+    std::vector traversal_order(dims_count);
+    for (int i = 0; i < dims_count; i++) {
+      traversal_order[i] = i;
+      dim_metadata[i] = CreateDimensionMetadata(builder, DimensionType_DENSE,
+                                                SlopeShape()[i]);
+    }
+    const flatbuffers::Offset sparsity_param =
+        CreateSparsityParameters(builder, builder.CreateVector(traversal_order),
+                                 0, builder.CreateVector(dim_metadata));
+    tensors.emplace_back(CreateTensor(
+        builder,
+        builder.CreateVector(SlopeShape().data(), SlopeShape().size()),
+        TensorType_FLOAT32, /*buffer=*/1, /*name=*/0, /*quantization=*/0,
+        /*is_variable=*/false, /*sparsity=*/sparsity_param));
+  }
+  if (FP16Weights()) {
+    const std::array dequantize_inputs{{0}};
+    const std::array dequantize_outputs{{2}};
+    operators.emplace_back(CreateOperator(
+        builder, /*opcode_index=*/1,
+        builder.CreateVector(dequantize_inputs.data(),
+                                      dequantize_inputs.size()),
+        builder.CreateVector(dequantize_outputs.data(),
+                                      dequantize_outputs.size())));
+  } else if (SparseWeights()) {
+    const std::array densify_inputs{{0}};
+    const std::array densify_outputs{{2}};
+    operators.emplace_back(
+        CreateOperator(builder, /*opcode_index=*/1,
+                       builder.CreateVector(densify_inputs.data(),
+                                                     densify_inputs.size()),
+                       builder.CreateVector(densify_outputs.data(),
+                                                     densify_outputs.size())));
+  }
+  tensors.emplace_back(CreateTensor(
+      builder,
+      builder.CreateVector(InputShape().data(), InputShape().size()),
+      TensorType_FLOAT32));
+  tensors.emplace_back(CreateTensor(
+      builder,
+      builder.CreateVector(SlopeShape().data(), SlopeShape().size()),
+      TensorType_FLOAT32,
+      /*buffer=*/(FP16Weights() || SparseWeights()) ? 0 : 1));
+  tensors.emplace_back(CreateTensor(
+      builder,
+      builder.CreateVector(OutputShape().data(), OutputShape().size()),
+      TensorType_FLOAT32));
+
+  const std::array op_inputs{
+      {static_cast(tensors.size()) - 3,
+       static_cast(tensors.size()) - 2}};
+  const std::array op_outputs{
+      {static_cast(tensors.size()) - 1}};
+  operators.emplace_back(CreateOperator(
+      builder, /*opcode_index=*/0,
+      builder.CreateVector(op_inputs.data(), op_inputs.size()),
+      builder.CreateVector(op_outputs.data(), op_outputs.size())));
+
+  const std::array subgraph_inputs{
+      {static_cast(tensors.size() - 3)}};
+  const std::array subgraph_outputs{
+      {static_cast(tensors.size()) - 1}};
+  flatbuffers::Offset subgraph = CreateSubGraph(
+      builder, builder.CreateVector(tensors.data(), tensors.size()),
+      builder.CreateVector(subgraph_inputs.data(),
+                                    subgraph_inputs.size()),
+      builder.CreateVector(subgraph_outputs.data(),
+                                    subgraph_outputs.size()),
+      builder.CreateVector(operators.data(), operators.size()));
+
+  flatbuffers::Offset description =
+      builder.CreateString("PReLU model");
+
+  flatbuffers::Offset model_buffer = CreateModel(
+      builder, TFLITE_SCHEMA_VERSION,
+      builder.CreateVector(operator_codes.data(), operator_codes.size()),
+      builder.CreateVector(&subgraph, 1), description,
+      builder.CreateVector(buffers.data(), buffers.size()));
+
+  builder.Finish(model_buffer);
+
+  return std::vector(builder.GetBufferPointer(),
+                           builder.GetBufferPointer() + builder.GetSize());
+}
+
+int32_t PreluTester::ComputeSize(const std::vector& shape) {
+  return std::accumulate(shape.cbegin(), shape.cend(), 1,
+                         std::multiplies());
+}
+
+}  // namespace xnnpack
+}  // namespace tflite
diff --git a/tensorflow/lite/delegates/xnnpack/prelu_tester.h b/tensorflow/lite/delegates/xnnpack/prelu_tester.h
new file mode 100644
index 00000000000..e89bae6029b
--- /dev/null
+++ b/tensorflow/lite/delegates/xnnpack/prelu_tester.h
@@ -0,0 +1,88 @@
+/* Copyright 2020 The TensorFlow Authors. All Rights Reserved.
+
+Licensed under the Apache License, Version 2.0 (the "License");
+you may not use this file except in compliance with the License.
+You may obtain a copy of the License at
+
+    http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software
+distributed under the License is distributed on an "AS IS" BASIS,
+WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+See the License for the specific language governing permissions and
+limitations under the License.
+==============================================================================*/
+
+#ifndef TENSORFLOW_LITE_DELEGATES_XNNPACK_PRELU_TESTER_H_
+#define TENSORFLOW_LITE_DELEGATES_XNNPACK_PRELU_TESTER_H_
+
+#include 
+#include 
+
+#include 
+#include "tensorflow/lite/c/common.h"
+
+namespace tflite {
+namespace xnnpack {
+
+class PreluTester {
+ public:
+  PreluTester() = default;
+  PreluTester(const PreluTester&) = delete;
+  PreluTester& operator=(const PreluTester&) = delete;
+
+  inline PreluTester& InputShape(std::initializer_list shape) {
+    for (auto it = shape.begin(); it != shape.end(); ++it) {
+      EXPECT_GT(*it, 0);
+    }
+    input_shape_ = std::vector(shape.begin(), shape.end());
+    return *this;
+  }
+
+  inline const std::vector& InputShape() const { return input_shape_; }
+
+  inline PreluTester& SlopeShape(std::initializer_list shape) {
+    for (auto it = shape.begin(); it != shape.end(); ++it) {
+      EXPECT_GT(*it, 0);
+    }
+    slope_shape_ = std::vector(shape.begin(), shape.end());
+    return *this;
+  }
+
+  inline const std::vector& SlopeShape() const { return slope_shape_; }
+
+  inline const std::vector& OutputShape() const {
+    return InputShape();
+  }
+
+  inline PreluTester& FP16Weights() {
+    fp16_weights_ = true;
+    return *this;
+  }
+
+  inline bool FP16Weights() const { return fp16_weights_; }
+
+  inline PreluTester& SparseWeights() {
+    sparse_weights_ = true;
+    return *this;
+  }
+
+  inline bool SparseWeights() const { return sparse_weights_; }
+
+  void Test(TfLiteDelegate* delegate) const;
+
+ private:
+  std::vector CreateTfLiteModel() const;
+
+  static int32_t ComputeSize(const std::vector& shape);
+
+  std::vector input_shape_;
+  std::vector slope_shape_;
+  bool fp16_weights_ = false;
+  bool sparse_weights_ = false;
+};
+
+}  // namespace xnnpack
+}  // namespace tflite
+
+#endif  // TENSORFLOW_LITE_DELEGATES_XNNPACK_PRELU_TESTER_H_
diff --git a/tensorflow/lite/delegates/xnnpack/xnnpack_delegate.cc b/tensorflow/lite/delegates/xnnpack/xnnpack_delegate.cc
index 0afc9c32122..31468ef7407 100644
--- a/tensorflow/lite/delegates/xnnpack/xnnpack_delegate.cc
+++ b/tensorflow/lite/delegates/xnnpack/xnnpack_delegate.cc
@@ -2266,7 +2266,8 @@ class Subgraph {
     const TfLiteTensor& input_tensor = tensors[node->inputs->data[0]];
     TF_LITE_ENSURE_STATUS(CheckTensorFloatType(
         logging_context, input_tensor, node->inputs->data[0], node_index));
-    TF_LITE_ENSURE_STATUS(CheckTensorShape(logging_context, input_tensor, 4,
+    TF_LITE_ENSURE_STATUS(CheckTensorShape(logging_context, input_tensor, 1,
+                                           XNN_MAX_TENSOR_DIMS,
                                            node->inputs->data[0]));
     TF_LITE_ENSURE_STATUS(CheckTensorNonDynamicAllocation(
         logging_context, input_tensor, node->inputs->data[0], node_index));
@@ -2284,7 +2285,8 @@ class Subgraph {
     const TfLiteTensor& output_tensor = tensors[node->outputs->data[0]];
     TF_LITE_ENSURE_STATUS(CheckTensorFloatType(
         logging_context, output_tensor, node->outputs->data[0], node_index));
-    TF_LITE_ENSURE_STATUS(CheckTensorShape(logging_context, output_tensor, 4,
+    TF_LITE_ENSURE_STATUS(CheckTensorShape(logging_context, output_tensor, 1,
+                                           XNN_MAX_TENSOR_DIMS,
                                            node->outputs->data[0]));
     TF_LITE_ENSURE_STATUS(CheckTensorNonDynamicAllocation(
         logging_context, output_tensor, node->outputs->data[0], node_index));

From 9f43ebd68c35a1f8f0f995352f7108b72fb41f50 Mon Sep 17 00:00:00 2001
From: Bixia Zheng 
Date: Sun, 21 Jun 2020 23:18:37 -0700
Subject: [PATCH 407/867] [TF:TRT] Enforce no native segment execution for
 TfTrtIntegrationTestBase tests.

Report an error if the native segment for an TRTEngineOp is executed. This
ensures that the TRTEngineOp constructed by the bridge is acceptable to
TensorRT to catch bugs.

Modify TrtModeTestBase to not build static engines for a graph that generates
dynamic shaped values. This avoids native segment execution that is not caused
by the inconsistency between the bridge and TensorRT.

Temporarily allow native segment execution for VGGBlockTest and
VGGBlockNCHWTest to workaround b/159459919.

PiperOrigin-RevId: 317597344
Change-Id: I6c268c5c912a1fddcffa5d9763399976ddc0299e
---
 .../compiler/tensorrt/test/tf_trt_integration_test_base.py | 1 +
 tensorflow/python/compiler/tensorrt/test/trt_mode_test.py  | 5 +++++
 .../python/compiler/tensorrt/test/vgg_block_nchw_test.py   | 7 +++++++
 tensorflow/python/compiler/tensorrt/test/vgg_block_test.py | 7 +++++++
 4 files changed, 20 insertions(+)

diff --git a/tensorflow/python/compiler/tensorrt/test/tf_trt_integration_test_base.py b/tensorflow/python/compiler/tensorrt/test/tf_trt_integration_test_base.py
index 8b93750fde4..87fa55a32bd 100644
--- a/tensorflow/python/compiler/tensorrt/test/tf_trt_integration_test_base.py
+++ b/tensorflow/python/compiler/tensorrt/test/tf_trt_integration_test_base.py
@@ -971,4 +971,5 @@ def _AddTests(test_class):
 
 
 if is_tensorrt_enabled():
+  os.environ["TF_TRT_ALLOW_ENGINE_NATIVE_SEGMENT_EXECUTION"] = "False"
   _AddTests(TfTrtIntegrationTestBase)
diff --git a/tensorflow/python/compiler/tensorrt/test/trt_mode_test.py b/tensorflow/python/compiler/tensorrt/test/trt_mode_test.py
index 878ab4cbd8e..c67de7432cd 100644
--- a/tensorflow/python/compiler/tensorrt/test/trt_mode_test.py
+++ b/tensorflow/python/compiler/tensorrt/test/trt_mode_test.py
@@ -40,6 +40,11 @@ class TrtModeTestBase(trt_test.TfTrtIntegrationTestBase):
     q = q + 5.0
     return array_ops.identity(q, name="output_0")
 
+  def ShouldRunTest(self, run_params):
+    # Squeeze op produces dynamic shaped values. Therefore, we don't run the
+    # test with static engine to avoid native segment execution.
+    return (run_params.dynamic_engine, "test dynamic engine only")
+
   def GetParams(self):
     """The input has 1 as a first dimension, which is removed by the squeeze.
 
diff --git a/tensorflow/python/compiler/tensorrt/test/vgg_block_nchw_test.py b/tensorflow/python/compiler/tensorrt/test/vgg_block_nchw_test.py
index 368ffad30a4..8fd9606812d 100644
--- a/tensorflow/python/compiler/tensorrt/test/vgg_block_nchw_test.py
+++ b/tensorflow/python/compiler/tensorrt/test/vgg_block_nchw_test.py
@@ -18,6 +18,8 @@ from __future__ import absolute_import
 from __future__ import division
 from __future__ import print_function
 
+import os
+
 import numpy as np
 
 from tensorflow.python.compiler.tensorrt.test import tf_trt_integration_test_base as trt_test
@@ -69,6 +71,11 @@ class VGGBlockNCHWTest(trt_test.TfTrtIntegrationTestBase):
     """Return the expected engines to build."""
     return ["TRTEngineOp_0"]
 
+  # TODO(b/159459919): remove this routine to disallow native segment execution.
+  def setUp(self):
+    super(trt_test.TfTrtIntegrationTestBase, self).setUp()
+    os.environ["TF_TRT_ALLOW_ENGINE_NATIVE_SEGMENT_EXECUTION"] = "True"
+
 
 if __name__ == "__main__":
   test.main()
diff --git a/tensorflow/python/compiler/tensorrt/test/vgg_block_test.py b/tensorflow/python/compiler/tensorrt/test/vgg_block_test.py
index f1b41327a58..9d81cd6dcc3 100644
--- a/tensorflow/python/compiler/tensorrt/test/vgg_block_test.py
+++ b/tensorflow/python/compiler/tensorrt/test/vgg_block_test.py
@@ -18,6 +18,8 @@ from __future__ import absolute_import
 from __future__ import division
 from __future__ import print_function
 
+import os
+
 import numpy as np
 
 from tensorflow.python.compiler.tensorrt.test import tf_trt_integration_test_base as trt_test
@@ -60,6 +62,11 @@ class VGGBlockTest(trt_test.TfTrtIntegrationTestBase):
     """Return the expected engines to build."""
     return ["TRTEngineOp_0"]
 
+  # TODO(b/159459919): remove this routine to disallow native segment execution.
+  def setUp(self):
+    super(trt_test.TfTrtIntegrationTestBase, self).setUp()
+    os.environ["TF_TRT_ALLOW_ENGINE_NATIVE_SEGMENT_EXECUTION"] = "True"
+
 
 if __name__ == "__main__":
   test.main()

From 5229c77d94bdae111fcba9efb353c67c7c58fa44 Mon Sep 17 00:00:00 2001
From: Adrian Kuegel 
Date: Mon, 22 Jun 2020 00:18:35 -0700
Subject: [PATCH 408/867] Add cuda_root filegroup target to the windows BUILD
 template.

When I added this to the linux BUILD template, I forgot to add it here, too.
Adjust cuda_configure.bzl.oss so that it copies the binaries with .exe
extension on Windows.

PiperOrigin-RevId: 317601952
Change-Id: I0712bcd926372cb9d067ead7f92270d52883bfd9
---
 third_party/gpus/cuda/BUILD.windows.tpl |  8 ++++++++
 third_party/gpus/cuda_configure.bzl     | 13 +++++++------
 2 files changed, 15 insertions(+), 6 deletions(-)

diff --git a/third_party/gpus/cuda/BUILD.windows.tpl b/third_party/gpus/cuda/BUILD.windows.tpl
index e83da1429dd..55a9ec3d1ab 100644
--- a/third_party/gpus/cuda/BUILD.windows.tpl
+++ b/third_party/gpus/cuda/BUILD.windows.tpl
@@ -196,6 +196,14 @@ cc_library(
     data = [":cuda-nvvm"],
 )
 
+filegroup(
+    name = "cuda_root",
+    srcs = [
+        "cuda/bin/fatbinary.exe",
+        "cuda/bin/bin2c.exe",
+    ],
+)
+
 bzl_library(
     name = "build_defs_bzl",
     srcs = ["build_defs.bzl"],
diff --git a/third_party/gpus/cuda_configure.bzl b/third_party/gpus/cuda_configure.bzl
index 0b87ba1ae2a..3374965f415 100644
--- a/third_party/gpus/cuda_configure.bzl
+++ b/third_party/gpus/cuda_configure.bzl
@@ -1062,20 +1062,21 @@ def _create_local_cuda_repository(repository_ctx):
     ))
 
     # copy files mentioned in third_party/nccl/build_defs.bzl.tpl
+    file_ext = ".exe" if is_windows(repository_ctx) else ""
     copy_rules.append(make_copy_files_rule(
         repository_ctx,
         name = "cuda-bin",
         srcs = [
             cuda_config.cuda_toolkit_path + "/bin/" + "crt/link.stub",
-            cuda_config.cuda_toolkit_path + "/bin/" + "nvlink",
-            cuda_config.cuda_toolkit_path + "/bin/" + "fatbinary",
-            cuda_config.cuda_toolkit_path + "/bin/" + "bin2c",
+            cuda_config.cuda_toolkit_path + "/bin/" + "nvlink" + file_ext,
+            cuda_config.cuda_toolkit_path + "/bin/" + "fatbinary" + file_ext,
+            cuda_config.cuda_toolkit_path + "/bin/" + "bin2c" + file_ext,
         ],
         outs = [
             "cuda/bin/" + "crt/link.stub",
-            "cuda/bin/" + "nvlink",
-            "cuda/bin/" + "fatbinary",
-            "cuda/bin/" + "bin2c",
+            "cuda/bin/" + "nvlink" + file_ext,
+            "cuda/bin/" + "fatbinary" + file_ext,
+            "cuda/bin/" + "bin2c" + file_ext,
         ],
     ))
 

From 8717d5c92c17bad370856dfb1debab48b6158f8b Mon Sep 17 00:00:00 2001
From: Jens Elofsson 
Date: Mon, 22 Jun 2020 09:38:10 +0200
Subject: [PATCH 409/867] Fix reviewer comments.

---
 tensorflow/lite/micro/examples/network_tester/network_model.h | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/tensorflow/lite/micro/examples/network_tester/network_model.h b/tensorflow/lite/micro/examples/network_tester/network_model.h
index 0431d7deee7..5b4b4cf3070 100644
--- a/tensorflow/lite/micro/examples/network_tester/network_model.h
+++ b/tensorflow/lite/micro/examples/network_tester/network_model.h
@@ -67,4 +67,4 @@ const unsigned char network_model[] = {
     0x08, 0x00, 0x07, 0x00, 0x06, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x11};
 const unsigned int network_model_len = 576;
 
-#endif
+#endif  // TENSORFLOW_LITE_MICRO_EXAMPLES_NETWORK_TESTER_NETWORK_MODEL_H_

From 072cf7ee4b11e40fdb995e248c82774caf8eb989 Mon Sep 17 00:00:00 2001
From: Akshay Modi 
Date: Mon, 22 Jun 2020 00:42:58 -0700
Subject: [PATCH 410/867] Disable __setitem__ on tf numpy's ndarray for
 experimental release

PiperOrigin-RevId: 317604521
Change-Id: I19c6a78125fcd29109e45a66099e7325f7136fdf
---
 tensorflow/python/ops/numpy_ops/BUILD         | 10 ---
 .../python/ops/numpy_ops/np_array_ops.py      | 56 ---------------
 .../python/ops/numpy_ops/np_array_ops_test.py | 20 ------
 .../python/ops/numpy_ops/np_backprop_test.py  | 69 -------------------
 4 files changed, 155 deletions(-)
 delete mode 100644 tensorflow/python/ops/numpy_ops/np_backprop_test.py

diff --git a/tensorflow/python/ops/numpy_ops/BUILD b/tensorflow/python/ops/numpy_ops/BUILD
index 3479a622bc0..704e5a27b48 100644
--- a/tensorflow/python/ops/numpy_ops/BUILD
+++ b/tensorflow/python/ops/numpy_ops/BUILD
@@ -65,16 +65,6 @@ cuda_py_test(
     ],
 )
 
-cuda_py_test(
-    name = "np_backprop_test",
-    srcs = ["np_backprop_test.py"],
-    deps = [
-        ":numpy",
-        "//tensorflow/python:platform",
-        "@absl_py//absl/testing:parameterized",
-    ],
-)
-
 cuda_py_test(
     name = "np_logic_test",
     srcs = ["np_logic_test.py"],
diff --git a/tensorflow/python/ops/numpy_ops/np_array_ops.py b/tensorflow/python/ops/numpy_ops/np_array_ops.py
index a87c72ed763..25241272699 100644
--- a/tensorflow/python/ops/numpy_ops/np_array_ops.py
+++ b/tensorflow/python/ops/numpy_ops/np_array_ops.py
@@ -852,66 +852,10 @@ def moveaxis(a, source, destination):  # pylint: disable=missing-docstring
   return np_utils.tensor_to_ndarray(a)
 
 
-def _setitem(arr, index, value):
-  """Sets the `value` at `index` in the array `arr`.
-
-  This works by replacing the slice at `index` in the tensor with `value`.
-  Since tensors are immutable, this builds a new tensor using the `tf.concat`
-  op. Currently, only 0-d and 1-d indices are supported.
-
-  Note that this may break gradients e.g.
-
-  a = tf_np.array([1, 2, 3])
-  old_a_t = a.data
-
-  with tf.GradientTape(persistent=True) as g:
-    g.watch(a.data)
-    b = a * 2
-    a[0] = 5
-  g.gradient(b.data, [a.data])  # [None]
-  g.gradient(b.data, [old_a_t])  # [[2., 2., 2.]]
-
-  Here `d_b / d_a` is `[None]` since a.data no longer points to the same
-  tensor.
-
-  Args:
-    arr: array_like.
-    index: scalar or 1-d integer array.
-    value: value to set at index.
-
-  Returns:
-    ndarray
-
-  Raises:
-    ValueError: if `index` is not a scalar or 1-d array.
-  """
-  # TODO(srbs): Figure out a solution to the gradient problem.
-  arr = asarray(arr)
-  index = asarray(index)
-  if index.ndim == 0:
-    index = ravel(index)
-  elif index.ndim > 1:
-    raise ValueError('index must be a scalar or a 1-d array.')
-  value = asarray(value, dtype=arr.dtype)
-  if arr.shape[len(index):] != value.shape:
-    value = full(arr.shape[len(index):], value)
-  prefix_t = arr.data[:index.data[0]]
-  postfix_t = arr.data[index.data[0] + 1:]
-  if len(index) == 1:
-    arr._data = array_ops.concat(  # pylint: disable=protected-access
-        [prefix_t, array_ops.expand_dims(value.data, 0), postfix_t], 0)
-  else:
-    subarray = arr[index.data[0]]
-    _setitem(subarray, index[1:], value)
-    arr._data = array_ops.concat(  # pylint: disable=protected-access
-        [prefix_t, array_ops.expand_dims(subarray.data, 0), postfix_t], 0)
-
-
 # TODO(wangpeng): Make a custom `setattr` that also sets docstring for the
 #   method.
 setattr(np_arrays.ndarray, 'transpose', transpose)
 setattr(np_arrays.ndarray, 'reshape', _reshape_method_wrapper)
-setattr(np_arrays.ndarray, '__setitem__', _setitem)
 
 
 @np_utils.np_doc('pad')
diff --git a/tensorflow/python/ops/numpy_ops/np_array_ops_test.py b/tensorflow/python/ops/numpy_ops/np_array_ops_test.py
index 25a6a53507a..d52e0c4ea83 100644
--- a/tensorflow/python/ops/numpy_ops/np_array_ops_test.py
+++ b/tensorflow/python/ops/numpy_ops/np_array_ops_test.py
@@ -911,26 +911,6 @@ class ArrayMethodsTest(test.TestCase):
     run_test(np.arange(30).reshape(2, 3, 5).tolist(), [2, 0, 1])
     run_test(np.arange(30).reshape(2, 3, 5).tolist(), [2, 1, 0])
 
-  def testSetItem(self):
-
-    def run_test(arr, index, value):
-      for fn in self.array_transforms:
-        value_arg = fn(value)
-        tf_array = np_array_ops.array(arr)
-        np_array = np.array(arr)
-        tf_array[index] = value_arg
-        # TODO(srbs): "setting an array element with a sequence" is thrown
-        # if we do not wrap value_arg in a numpy array. Investigate how this can
-        # be avoided.
-        np_array[index] = np.array(value_arg)
-        self.match(tf_array, np_array)
-
-    run_test([1, 2, 3], 1, 5)
-    run_test([[1, 2], [3, 4]], 0, [6, 7])
-    run_test([[1, 2], [3, 4]], 1, [6, 7])
-    run_test([[1, 2], [3, 4]], (0, 1), 6)
-    run_test([[1, 2], [3, 4]], 0, 6)  # Value needs to broadcast.
-
   def match_shape(self, actual, expected, msg=None):
     if msg:
       msg = 'Shape match failed for: {}. Expected: {} Actual: {}'.format(
diff --git a/tensorflow/python/ops/numpy_ops/np_backprop_test.py b/tensorflow/python/ops/numpy_ops/np_backprop_test.py
deleted file mode 100644
index 65c532153a4..00000000000
--- a/tensorflow/python/ops/numpy_ops/np_backprop_test.py
+++ /dev/null
@@ -1,69 +0,0 @@
-# Copyright 2020 The TensorFlow Authors. All Rights Reserved.
-#
-# Licensed under the Apache License, Version 2.0 (the "License");
-# you may not use this file except in compliance with the License.
-# You may obtain a copy of the License at
-#
-#     http://www.apache.org/licenses/LICENSE-2.0
-#
-# Unless required by applicable law or agreed to in writing, software
-# distributed under the License is distributed on an "AS IS" BASIS,
-# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-# See the License for the specific language governing permissions and
-# limitations under the License.
-# ==============================================================================
-"""Tests for backpropgration on tf-numpy functions."""
-
-from __future__ import absolute_import
-from __future__ import division
-from __future__ import print_function
-
-from tensorflow.python.eager import backprop
-from tensorflow.python.framework import ops
-from tensorflow.python.ops.numpy_ops import np_array_ops
-# Required for operator overloads
-from tensorflow.python.ops.numpy_ops import np_math_ops  # pylint: disable=unused-import
-from tensorflow.python.platform import test
-
-
-class BackpropTest(test.TestCase):
-
-  def test_setitem(self):
-    # Single integer index.
-    a = np_array_ops.array([1., 2., 3.])
-    b = np_array_ops.array(5.)
-    c = np_array_ops.array(10.)
-
-    tensors = [arr.data for arr in [a, b, c]]
-    with backprop.GradientTape() as g:
-      g.watch(tensors)
-      a[1] = b + c
-      loss = np_array_ops.sum(a)
-
-    gradients = g.gradient(loss.data, tensors)
-    self.assertSequenceEqual(
-        np_array_ops.array(gradients[0]).tolist(), [1., 0., 1.])
-    self.assertEqual(np_array_ops.array(gradients[1]).tolist(), 1.)
-    self.assertEqual(np_array_ops.array(gradients[2]).tolist(), 1.)
-
-    # Tuple index.
-    a = np_array_ops.array([[[1., 2.], [3., 4.]], [[5., 6.],
-                                                   [7., 8.]]])  # 2x2x2 array.
-    b = np_array_ops.array([10., 11.])
-
-    tensors = [arr.data for arr in [a, b]]
-    with backprop.GradientTape() as g:
-      g.watch(tensors)
-      a[(1, 0)] = b
-      loss = np_array_ops.sum(a)
-
-    gradients = g.gradient(loss.data, tensors)
-    self.assertSequenceEqual(
-        np_array_ops.array(gradients[0]).tolist(),
-        [[[1., 1.], [1., 1.]], [[0., 0.], [1., 1.]]])
-    self.assertEqual(np_array_ops.array(gradients[1]).tolist(), [1., 1.])
-
-
-if __name__ == '__main__':
-  ops.enable_eager_execution()
-  test.main()

From 149a0a1d5a1b920568b8148a31c57903b116e80a Mon Sep 17 00:00:00 2001
From: Adrian Kuegel 
Date: Mon, 22 Jun 2020 01:38:54 -0700
Subject: [PATCH 411/867] Add support for windows.

The tools from the cuda toolkit have the .exe extension on Windows.

PiperOrigin-RevId: 317610348
Change-Id: I6f4bad4c651c8ffb347f06d1aef8ffe00de90620
---
 tensorflow/core/kernels/cubin_headers/build_defs.bzl | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/tensorflow/core/kernels/cubin_headers/build_defs.bzl b/tensorflow/core/kernels/cubin_headers/build_defs.bzl
index 5880cbe8add..c3e44b7a974 100644
--- a/tensorflow/core/kernels/cubin_headers/build_defs.bzl
+++ b/tensorflow/core/kernels/cubin_headers/build_defs.bzl
@@ -5,7 +5,7 @@ load("@local_config_cuda//cuda:build_defs.bzl", "cuda_gpu_architectures", "if_cu
 def _lookup_file(filegroup, path):
     """Extracts file at (relative) path in filegroup."""
     for file in filegroup.files.to_list():
-        if file.path.endswith(path):
+        if file.path.endswith(path) or file.path.endswith(path + ".exe"):
             return file
     return None
 

From 0868ca7bb26c167918209d5b6885ecc695d7b05a Mon Sep 17 00:00:00 2001
From: Tamara Norman 
Date: Mon, 22 Jun 2020 01:49:32 -0700
Subject: [PATCH 412/867] Allow a shape to be passed to CopyToHostAsync

PiperOrigin-RevId: 317611333
Change-Id: I4526f9dbd1b223eb23fe928326afca0eb133c2f5
---
 tensorflow/compiler/xla/pjrt/BUILD          |  2 +
 tensorflow/compiler/xla/pjrt/pjrt_client.cc | 67 ++++++++++++++-------
 tensorflow/compiler/xla/pjrt/pjrt_client.h  | 24 ++++++--
 3 files changed, 67 insertions(+), 26 deletions(-)

diff --git a/tensorflow/compiler/xla/pjrt/BUILD b/tensorflow/compiler/xla/pjrt/BUILD
index e401a798d68..695ba9dee93 100644
--- a/tensorflow/compiler/xla/pjrt/BUILD
+++ b/tensorflow/compiler/xla/pjrt/BUILD
@@ -149,6 +149,7 @@ cc_library(
         "//tensorflow/stream_executor/host:host_platform_id",
         "//tensorflow/stream_executor/lib",
         "@com_google_absl//absl/base",
+        "@com_google_absl//absl/container:flat_hash_map",
         "@com_google_absl//absl/container:flat_hash_set",
         "@com_google_absl//absl/container:inlined_vector",
         "@com_google_absl//absl/memory",
@@ -156,6 +157,7 @@ cc_library(
         "@com_google_absl//absl/strings:str_format",
         "@com_google_absl//absl/synchronization",
         "@com_google_absl//absl/time",
+        "@com_google_absl//absl/types:optional",
         "@com_google_absl//absl/types:span",
     ],
 )
diff --git a/tensorflow/compiler/xla/pjrt/pjrt_client.cc b/tensorflow/compiler/xla/pjrt/pjrt_client.cc
index b4f0363e69a..e341a11d64f 100644
--- a/tensorflow/compiler/xla/pjrt/pjrt_client.cc
+++ b/tensorflow/compiler/xla/pjrt/pjrt_client.cc
@@ -76,11 +76,13 @@ limitations under the License.
 #include "absl/strings/str_format.h"
 #include "absl/synchronization/mutex.h"
 #include "absl/time/time.h"
+#include "absl/types/optional.h"
 #include "absl/types/span.h"
 #include "tensorflow/compiler/xla/client/local_client.h"
 #include "tensorflow/compiler/xla/client/xla_computation.h"
 #include "tensorflow/compiler/xla/cpu_function_runtime.h"
 #include "tensorflow/compiler/xla/executable_run_options.h"
+#include "tensorflow/compiler/xla/layout.h"
 #include "tensorflow/compiler/xla/literal.h"
 #include "tensorflow/compiler/xla/literal_util.h"
 #include "tensorflow/compiler/xla/pjrt/distributed/protocol.pb.h"
@@ -861,10 +863,10 @@ StatusOr> PjRtBuffer::Release(
     if (device_buffer_ == nullptr) {
       return std::shared_ptr();
     }
-    // Set host_value_ and device_buffer_ to null now so that no other thread
-    // can add a hold while we are in WaitForOutstandingUsageHolds()
+    // Clear host_values_ and set device_buffer_ to null now so that no other
+    // thread can add a hold while we are in WaitForOutstandingUsageHolds()
     // below.
-    host_value_ = nullptr;
+    host_values_.clear();
     std::swap(device_buffer_, device_buffer);
     WaitForOutstandingUsageHolds();
     // Now that all holds have completed and no more can be added, we can get
@@ -999,7 +1001,7 @@ void PjRtBuffer::ConfirmDonation(TrackedDeviceBuffer* device_buffer) {
     device_buffer->ReleaseDeviceMemory();
     // Make *this invalid so it can't be used again. Any threads blocking in
     // Release or GetBufferWithHold will see an invalid buffer and return.
-    host_value_ = nullptr;
+    host_values_.clear();
     device_buffer_.reset();
   }
   // Unblock another thread, if any, trying to get a donation hold.
@@ -1019,7 +1021,14 @@ void PjRtBuffer::DropHold(ScopedHold::Type type, TrackedDeviceBuffer* buffer) {
   }
 }
 
-Status PjRtBuffer::CopyToHostAsync() {
+Status PjRtBuffer::CopyToHostAsync(absl::optional layout) {
+  return CopyToHostAsyncInternal(/*discard_cached_copy=*/false, layout)
+      .status();
+}
+
+StatusOr>
+PjRtBuffer::CopyToHostAsyncInternal(bool discard_cached_copy,
+                                    absl::optional layout) {
   if (IsEmptyTuple()) {
     return InvalidArgument("CopyToHostAsync called on empty tuple");
   }
@@ -1027,6 +1036,8 @@ Status PjRtBuffer::CopyToHostAsync() {
   std::shared_ptr host_value;
   LocalDeviceState* local_device = device_->local_device_state();
   se::Stream* stream = local_device->GetDeviceToHostStream();
+  const xla::Layout& host_layout =
+      layout.has_value() ? layout.value() : on_host_shape_.layout();
   {
     absl::MutexLock lock(&mu_);
     // We can't perform any other action while a donation hold is in progress.
@@ -1034,17 +1045,36 @@ Status PjRtBuffer::CopyToHostAsync() {
     if (device_buffer_ == nullptr) {
       return InvalidArgument("CopyToHostAsync() called on invalid buffer.");
     }
-    if (host_value_) {
-      // The host value has already been requested or is available.
-      return Status::OK();
+    if (discard_cached_copy) {
+      auto it = host_values_.find(host_layout);
+      if (it != host_values_.end()) {
+        host_value = it->second;
+        host_values_.erase(it);
+        return host_value;
+      } else {
+        host_value = std::make_shared();
+      }
+    } else {
+      std::shared_ptr& host_value_ref = host_values_[host_layout];
+      if (host_value_ref) {
+        return host_value_ref;
+      }
+      host_value = host_value_ref = std::make_shared();
     }
-    host_value = host_value_ = std::make_shared();
     AcquireHoldLocked(&device_buffer);
   }
   WaitForBufferDefinitionEventsOnStream(*device_buffer, stream);
-  host_value->value = std::make_shared(on_host_shape_);
+  Shape host_shape;
+  if (layout.has_value()) {
+    host_shape = ShapeUtil::MakeShape(on_host_shape_.element_type(),
+                                      on_host_shape_.dimensions());
+    *host_shape.mutable_layout() = host_layout;
+  } else {
+    host_shape = on_host_shape_;
+  }
+  host_value->value = std::make_shared(host_shape);
   ShapedBuffer shaped_buffer = device_buffer->AsShapedBuffer(
-      on_host_shape_, on_device_shape_, client_->client()->platform());
+      host_shape, on_device_shape_, client_->client()->platform());
   client_->client()->backend().transfer_manager()->TransferLiteralFromDevice(
       stream, shaped_buffer, host_value->value.get(),
       [host_value](Status done_status) {
@@ -1074,21 +1104,14 @@ Status PjRtBuffer::CopyToHostAsync() {
   RecordUsage(std::move(device_buffer), local_device, local_device, usage_event,
               stream,
               /*prefer_to_retain_reference=*/true);
-  return Status::OK();
+  return host_value;
 }
 
 StatusOr> PjRtBuffer::ToLiteral(
-    const bool discard_cached_copy) {
+    const bool discard_cached_copy, absl::optional layout) {
   tensorflow::profiler::TraceMe traceme("PjRtBuffer::ToLiteral");
-  TF_RETURN_IF_ERROR(CopyToHostAsync());
-  std::shared_ptr host_value;
-  {
-    absl::MutexLock lock(&mu_);
-    host_value = host_value_;
-    if (discard_cached_copy) {
-      host_value_ = nullptr;
-    }
-  }
+  TF_ASSIGN_OR_RETURN(std::shared_ptr host_value,
+                      CopyToHostAsyncInternal(discard_cached_copy, layout));
   if (host_value == nullptr) {
     return InvalidArgument("ToLiteral called on invalid buffer");
   }
diff --git a/tensorflow/compiler/xla/pjrt/pjrt_client.h b/tensorflow/compiler/xla/pjrt/pjrt_client.h
index 8f74e6244d6..c50d09f631c 100644
--- a/tensorflow/compiler/xla/pjrt/pjrt_client.h
+++ b/tensorflow/compiler/xla/pjrt/pjrt_client.h
@@ -20,15 +20,18 @@ limitations under the License.
 #include 
 #include 
 
+#include "absl/container/flat_hash_map.h"
 #include "absl/container/flat_hash_set.h"
 #include "absl/container/inlined_vector.h"
 #include "absl/strings/string_view.h"
 #include "absl/synchronization/mutex.h"
 #include "absl/synchronization/notification.h"
+#include "absl/types/optional.h"
 #include "absl/types/span.h"
 #include "tensorflow/compiler/xla/client/executable_build_options.h"
 #include "tensorflow/compiler/xla/client/local_client.h"
 #include "tensorflow/compiler/xla/client/xla_computation.h"
+#include "tensorflow/compiler/xla/layout.h"
 #include "tensorflow/compiler/xla/pjrt/local_device_state.h"
 #include "tensorflow/compiler/xla/pjrt/tracked_device_buffer.h"
 #include "tensorflow/compiler/xla/service/computation_placer.h"
@@ -481,14 +484,17 @@ class PjRtBuffer {
   // copies the buffer to the host. Blocks until the value is ready. If
   // `discard_cached_copy` is true then buffer will no longer keep hold of a
   // cached copy of the literal (i.e. The reference to the host value will be
-  // removed.)
+  // removed.) If a layout is passed than a literal with this layout will be
+  // returned.
   StatusOr> ToLiteral(
-      bool discard_cached_copy = false);
+      bool discard_cached_copy = false,
+      absl::optional layout = {});
 
   // Initiates a copy of the buffer to the host. Does not block waiting for
   // the transfer to complete. The value can be retrieved by a later call to
-  // ToLiteral().
-  Status CopyToHostAsync();
+  // ToLiteral(). If a layout is passed then a cached copy with this layout will
+  // be created.
+  Status CopyToHostAsync(absl::optional layout = {});
 
   // Drops the buffer's reference to its associated device memory, leaving the
   // buffer in an invalid state. The memory will be freed lazily when all async
@@ -596,6 +602,14 @@ class PjRtBuffer {
   // successfully donated to an execution.
   void ConfirmDonation(TrackedDeviceBuffer* device_buffer);
 
+  // Initiates a copy of the buffer to the host. Does not block waiting for
+  // the transfer to complete. A host value is returned and if
+  // `discard_cached_copy` is false stored in an internal buffer so that future
+  // transfers don't have to transfer the data from host again. If a layout is
+  // passed then a literal of this layout will be returned and possibly cached.
+  StatusOr> CopyToHostAsyncInternal(
+      bool discard_cached_copy, absl::optional layout);
+
   // Drops a hold without taking any other action. Does a sanity check that
   // buffer==device_buffer_ or device_buffer_==nullptr.
   void DropHold(ScopedHold::Type type, TrackedDeviceBuffer* buffer);
@@ -614,6 +628,8 @@ class PjRtBuffer {
 
   mutable absl::Mutex mu_;
   std::shared_ptr device_buffer_ TF_GUARDED_BY(mu_);
+  absl::flat_hash_map> host_values_
+      TF_GUARDED_BY(mu_);
   std::shared_ptr host_value_ TF_GUARDED_BY(mu_);
   // Count of holds on the buffer.
   std::array holds_ TF_GUARDED_BY(mu_);

From f3a8d13fa908f13e1fdfe83a9140fb6b9d815dc1 Mon Sep 17 00:00:00 2001
From: "A. Unique TensorFlower" 
Date: Mon, 22 Jun 2020 02:02:14 -0700
Subject: [PATCH 413/867] Update GraphDef version to 440.

PiperOrigin-RevId: 317612594
Change-Id: I43b88a42cc0f461da2356f3d71ce710d6da2df0b
---
 tensorflow/core/public/version.h | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/tensorflow/core/public/version.h b/tensorflow/core/public/version.h
index bfea50a1932..e63ca48863f 100644
--- a/tensorflow/core/public/version.h
+++ b/tensorflow/core/public/version.h
@@ -108,7 +108,7 @@ limitations under the License.
 
 #define TF_GRAPH_DEF_VERSION_MIN_PRODUCER 0
 #define TF_GRAPH_DEF_VERSION_MIN_CONSUMER 0
-#define TF_GRAPH_DEF_VERSION 439  // Updated: 2020/6/21
+#define TF_GRAPH_DEF_VERSION 440  // Updated: 2020/6/22
 
 // Checkpoint compatibility versions (the versions field in SavedSliceMeta).
 //

From 1bb6b68f211a3005f3d42757cb81da9463361e55 Mon Sep 17 00:00:00 2001
From: "A. Unique TensorFlower" 
Date: Mon, 22 Jun 2020 02:02:23 -0700
Subject: [PATCH 414/867] compat: Update forward compatibility horizon to
 2020-06-22

PiperOrigin-RevId: 317612617
Change-Id: I0656a5a1850d1bba744f5638dc2bb55eaf81702a
---
 tensorflow/python/compat/compat.py | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/tensorflow/python/compat/compat.py b/tensorflow/python/compat/compat.py
index 3f622ef844d..301f4608aba 100644
--- a/tensorflow/python/compat/compat.py
+++ b/tensorflow/python/compat/compat.py
@@ -33,7 +33,7 @@ from tensorflow.python.util.tf_export import tf_export
 # This value changes every day with an automatic CL. It can be modified in code
 # via `forward_compatibility_horizon()` or with the environment variable
 # TF_FORWARD_COMPATIBILITY_DELTA_DAYS, which is added to the compatibility date.
-_FORWARD_COMPATIBILITY_HORIZON = datetime.date(2020, 6, 21)
+_FORWARD_COMPATIBILITY_HORIZON = datetime.date(2020, 6, 22)
 _FORWARD_COMPATIBILITY_DELTA_DAYS_VAR_NAME = "TF_FORWARD_COMPATIBILITY_DELTA_DAYS"
 _FORWARD_COMPATIBILITY_DATE_NUMBER = None
 

From 9b1e89b77cf7dc9948ab4684e43867ac2485c2f6 Mon Sep 17 00:00:00 2001
From: Christian Sigg 
Date: Mon, 22 Jun 2020 03:58:04 -0700
Subject: [PATCH 415/867] Add RBE Dockerfile for CUDA11+cuDNN8.

PiperOrigin-RevId: 317626204
Change-Id: I75d4dc8dbde5dd4f3c59e047afe38043cbfb5a1c
---
 ...dnn8-ubuntu18.04-manylinux2010-multipython | 97 +++++++++++++++++++
 1 file changed, 97 insertions(+)
 create mode 100644 tensorflow/tools/ci_build/Dockerfile.rbe.cuda11.0-cudnn8-ubuntu18.04-manylinux2010-multipython

diff --git a/tensorflow/tools/ci_build/Dockerfile.rbe.cuda11.0-cudnn8-ubuntu18.04-manylinux2010-multipython b/tensorflow/tools/ci_build/Dockerfile.rbe.cuda11.0-cudnn8-ubuntu18.04-manylinux2010-multipython
new file mode 100644
index 00000000000..cd841a77aba
--- /dev/null
+++ b/tensorflow/tools/ci_build/Dockerfile.rbe.cuda11.0-cudnn8-ubuntu18.04-manylinux2010-multipython
@@ -0,0 +1,97 @@
+# Dockerfile to build a manylinux 2010 compliant cross-compiler.
+#
+# Builds a devtoolset gcc/libstdc++ that targets manylinux 2010 compatible
+# glibc (2.12) and system libstdc++ (4.4).
+#
+# To push a new version, run:
+# $ docker build -f Dockerfile.rbe.cuda11.0-cudnn8-ubuntu18.04-manylinux2010-multipython \
+#  --tag "gcr.io/tensorflow-testing/nosla-cuda11.0-cudnn8-ubuntu18.04-manylinux2010-multipython" .
+# $ docker push gcr.io/tensorflow-testing/nosla-cuda11.0-cudnn8-ubuntu18.04-manylinux2010-multipython
+
+FROM nvidia/cuda:11.0-cudnn8-devel-ubuntu18.04-rc as devtoolset
+
+ENV DEBIAN_FRONTEND=noninteractive
+RUN apt-get update && apt-get install -y \
+      cpio \
+      file \
+      flex \
+      g++ \
+      make \
+      patch \
+      rpm2cpio \
+      unar \
+      wget \
+      xz-utils \
+      && \
+    rm -rf /var/lib/apt/lists/*
+
+ADD devtoolset/fixlinks.sh fixlinks.sh
+ADD devtoolset/build_devtoolset.sh build_devtoolset.sh
+ADD devtoolset/rpm-patch.sh rpm-patch.sh
+
+# Set up a sysroot for glibc 2.12 / libstdc++ 4.4 / devtoolset-7 in /dt7.
+RUN /build_devtoolset.sh devtoolset-7 /dt7
+# Set up a sysroot for glibc 2.12 / libstdc++ 4.4 / devtoolset-8 in /dt8.
+RUN /build_devtoolset.sh devtoolset-8 /dt8
+
+# TODO(klimek): Split up into two different docker images.
+FROM nvidia/cuda:11.0-cudnn8-devel-ubuntu18.04-rc
+COPY --from=devtoolset /dt7 /dt7
+COPY --from=devtoolset /dt8 /dt8
+
+# Install TensorRT.
+RUN echo \
+    deb https://developer.download.nvidia.com/compute/machine-learning/repos/ubuntu1604/x86_64 / \
+    > /etc/apt/sources.list.d/nvidia-ml.list \
+      && \
+    apt-get update && apt-get install -y \
+    libnvinfer-dev=7.1.3-1+cuda11.0 \
+    libnvinfer7=7.1.3-1+cuda11.0 \
+    libnvinfer-plugin-dev=7.1.3-1+cuda11.0 \
+    libnvinfer-plugin7=7.1.3-1+cuda11.0 \
+      && \
+    rm -rf /var/lib/apt/lists/*
+
+# Copy and run the install scripts.
+ARG DEBIAN_FRONTEND=noninteractive
+
+COPY install/install_bootstrap_deb_packages.sh /install/
+RUN /install/install_bootstrap_deb_packages.sh
+
+COPY install/install_deb_packages.sh /install/
+RUN /install/install_deb_packages.sh
+
+# Install additional packages needed for this image:
+# - dependencies to build Python from source
+# - patchelf, as it is required by auditwheel
+RUN apt-get update && apt-get install -y \
+    libbz2-dev \
+    libffi-dev \
+    libgdbm-dev \
+    libncurses5-dev \
+    libnss3-dev \
+    libreadline-dev \
+    patchelf \
+      && \
+    rm -rf /var/lib/apt/lists/*
+
+COPY install/install_bazel.sh /install/
+RUN /install/install_bazel.sh
+
+COPY install/build_and_install_python.sh /install/
+RUN /install/build_and_install_python.sh "2.7.17" "--enable-unicode=ucs4"
+RUN /install/build_and_install_python.sh "3.5.9"
+RUN /install/build_and_install_python.sh "3.6.9"
+RUN /install/build_and_install_python.sh "3.7.7"
+RUN /install/build_and_install_python.sh "3.8.2"
+
+COPY install/install_pip_packages_by_version.sh /install/
+RUN /install/install_pip_packages_by_version.sh "/usr/local/bin/pip2.7"
+RUN /install/install_pip_packages_by_version.sh "/usr/local/bin/pip3.8"
+RUN /install/install_pip_packages_by_version.sh "/usr/local/bin/pip3.5"
+RUN /install/install_pip_packages_by_version.sh "/usr/local/bin/pip3.6"
+RUN /install/install_pip_packages_by_version.sh "/usr/local/bin/pip3.7"
+
+ENV CLANG_VERSION="r42cab985fd95ba4f3f290e7bb26b93805edb447d"
+COPY install/install_latest_clang.sh /install/
+RUN /install/install_latest_clang.sh

From b00a7808a7b29a78762b54e29aac87a77254b4b6 Mon Sep 17 00:00:00 2001
From: Christian Sigg 
Date: Mon, 22 Jun 2020 03:58:58 -0700
Subject: [PATCH 416/867] Add extra header file for cuDNN 8.

PiperOrigin-RevId: 317626279
Change-Id: I99b969a73555932b25081f37b64f71ac6de662d6
---
 third_party/gpus/cuda_configure.bzl | 9 ++++-----
 1 file changed, 4 insertions(+), 5 deletions(-)

diff --git a/third_party/gpus/cuda_configure.bzl b/third_party/gpus/cuda_configure.bzl
index 3374965f415..a192c022a47 100644
--- a/third_party/gpus/cuda_configure.bzl
+++ b/third_party/gpus/cuda_configure.bzl
@@ -1081,17 +1081,16 @@ def _create_local_cuda_repository(repository_ctx):
     ))
 
     # Select the headers based on the cuDNN version (strip '64_' for Windows).
-    if cuda_config.cudnn_version.rsplit("_", 1)[0] < "8":
-        cudnn_headers = ["cudnn.h"]
-    else:
-        cudnn_headers = [
+    cudnn_headers = ["cudnn.h"]
+    if cuda_config.cudnn_version.rsplit("_", 1)[0] >= "8":
+        cudnn_headers += [
+            "cudnn_backend.h",
             "cudnn_adv_infer.h",
             "cudnn_adv_train.h",
             "cudnn_cnn_infer.h",
             "cudnn_cnn_train.h",
             "cudnn_ops_infer.h",
             "cudnn_ops_train.h",
-            "cudnn.h",
             "cudnn_version.h",
         ]
 

From 3f28ac1b635a4b33c67a0676e4029741530d96d4 Mon Sep 17 00:00:00 2001
From: Dan Moldovan 
Date: Mon, 22 Jun 2020 06:37:01 -0700
Subject: [PATCH 417/867] Always return the interned version from as_dtype, so
 that dtypes created dynamically using `DType(value)` retain the exact value
 and type of their original counterparts. This allows more brittle code
 performing exact type checks to continue to work as expected once #40132
 lands.

PiperOrigin-RevId: 317643144
Change-Id: Ia17008a65f9300b28a0f2c7bf18a2213b2f407af
---
 tensorflow/python/framework/dtypes.py      |  5 ++++-
 tensorflow/python/framework/dtypes_test.py | 10 +++++++---
 2 files changed, 11 insertions(+), 4 deletions(-)

diff --git a/tensorflow/python/framework/dtypes.py b/tensorflow/python/framework/dtypes.py
index 994a7eea494..9eeae83c68a 100644
--- a/tensorflow/python/framework/dtypes.py
+++ b/tensorflow/python/framework/dtypes.py
@@ -607,6 +607,9 @@ assert len(_ANY_TO_TF) == sum(
 def as_dtype(type_value):
   """Converts the given `type_value` to a `DType`.
 
+  Note: `DType` values are interned. When passed a new `DType` object,
+  `as_dtype` always returns the interned value.
+
   Args:
     type_value: A value that can be converted to a `tf.DType` object. This may
       currently be a `tf.DType` object, a [`DataType`
@@ -620,7 +623,7 @@ def as_dtype(type_value):
     TypeError: If `type_value` cannot be converted to a `DType`.
   """
   if isinstance(type_value, DType):
-    return type_value
+    return _INTERN_TABLE[type_value.as_datatype_enum]
 
   if isinstance(type_value, np.dtype):
     try:
diff --git a/tensorflow/python/framework/dtypes_test.py b/tensorflow/python/framework/dtypes_test.py
index 041cc5280cd..1b7e02b6179 100644
--- a/tensorflow/python/framework/dtypes_test.py
+++ b/tensorflow/python/framework/dtypes_test.py
@@ -325,15 +325,19 @@ class TypesTest(test_util.TensorFlowTestCase):
     for enum in dtypes._TYPE_TO_STRING:
       dtype = dtypes.DType(enum)
       ctor, args = dtype.__reduce__()
-      self.assertEquals(ctor, dtypes.as_dtype)
-      self.assertEquals(args, (dtype.name,))
+      self.assertEqual(ctor, dtypes.as_dtype)
+      self.assertEqual(args, (dtype.name,))
       reconstructed = ctor(*args)
-      self.assertEquals(reconstructed, dtype)
+      self.assertEqual(reconstructed, dtype)
 
   def testAsDtypeInvalidArgument(self):
     with self.assertRaises(TypeError):
       dtypes.as_dtype((dtypes.int32, dtypes.float32))
 
+  def testAsDtypeReturnsInternedVersion(self):
+    dt = dtypes.DType(types_pb2.DT_VARIANT)
+    self.assertIs(dtypes.as_dtype(dt), dtypes.variant)
+
 
 if __name__ == "__main__":
   googletest.main()

From 2624f6c405e12012c8cdd90594fbc4ea573a2d91 Mon Sep 17 00:00:00 2001
From: Lukas Geiger 
Date: Mon, 22 Jun 2020 15:44:17 +0200
Subject: [PATCH 418/867] Replace deprecated numpy.ndarray.tostring() with
 tobytes()

---
 tensorflow/python/framework/tensor_util.py           | 2 +-
 tensorflow/python/kernel_tests/decode_raw_op_test.py | 8 ++++----
 2 files changed, 5 insertions(+), 5 deletions(-)

diff --git a/tensorflow/python/framework/tensor_util.py b/tensorflow/python/framework/tensor_util.py
index 968b635250a..3d5a16dd0f6 100644
--- a/tensorflow/python/framework/tensor_util.py
+++ b/tensorflow/python/framework/tensor_util.py
@@ -525,7 +525,7 @@ def make_tensor_proto(values, dtype=None, shape=None, verify_shape=False,
     if nparray.size * nparray.itemsize >= (1 << 31):
       raise ValueError(
           "Cannot create a tensor proto whose content is larger than 2GB.")
-    tensor_proto.tensor_content = nparray.tostring()
+    tensor_proto.tensor_content = nparray.tobytes()
     return tensor_proto
 
   # If we were not given values as a numpy array, compute the proto_values
diff --git a/tensorflow/python/kernel_tests/decode_raw_op_test.py b/tensorflow/python/kernel_tests/decode_raw_op_test.py
index ae8d40dbaea..00bdea2b43d 100644
--- a/tensorflow/python/kernel_tests/decode_raw_op_test.py
+++ b/tensorflow/python/kernel_tests/decode_raw_op_test.py
@@ -84,22 +84,22 @@ class DecodeRawOpTest(test.TestCase):
   def testToFloat16(self):
     result = np.matrix([[1, -2, -3, 4]], dtype="
Date: Mon, 22 Jun 2020 15:46:12 +0200
Subject: [PATCH 419/867] Replace deprecated np.asscalar with .item()

---
 tensorflow/python/ops/numpy_ops/np_arrays.py | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/tensorflow/python/ops/numpy_ops/np_arrays.py b/tensorflow/python/ops/numpy_ops/np_arrays.py
index 65e8273375f..97602806aac 100644
--- a/tensorflow/python/ops/numpy_ops/np_arrays.py
+++ b/tensorflow/python/ops/numpy_ops/np_arrays.py
@@ -282,7 +282,7 @@ class ndarray(composite_tensor.CompositeTensor):  # pylint: disable=invalid-name
     # TODO(wangpeng): Handle graph mode
     if not isinstance(self.data, ops.EagerTensor):
       raise TypeError('Indexing using symbolic tensor is not allowed')
-    return np.asscalar(self.data.numpy())
+    return self.data.numpy().item()
 
   def tolist(self):
     return self.data.numpy().tolist()

From a44416d0fc2ef89a6119e75ad35ef3824edb8cbd Mon Sep 17 00:00:00 2001
From: zilinzhu 
Date: Mon, 22 Jun 2020 21:49:59 +0800
Subject: [PATCH 420/867] fix broken figures

---
 tensorflow/compiler/xla/g3doc/tiled_layout.md | 14 ++++++++------
 tensorflow/compiler/xla/xla_data.proto        |  2 +-
 2 files changed, 9 insertions(+), 7 deletions(-)

diff --git a/tensorflow/compiler/xla/g3doc/tiled_layout.md b/tensorflow/compiler/xla/g3doc/tiled_layout.md
index 21e88ceab62..b40f0a95a3d 100644
--- a/tensorflow/compiler/xla/g3doc/tiled_layout.md
+++ b/tensorflow/compiler/xla/g3doc/tiled_layout.md
@@ -3,9 +3,10 @@
 Caution: Tiled layout is *pre-release* and this describes how it's intended to
 work. Errors may be silently ignored.
 
-
 ![](images/xla_array_layout_figure1.png)
-
-Figure 1 

+

+  
+  Figure 1
+

 
 Figure 1 shows how an array F32[3,5] is laid out in memory with 2x2 tiling. A
 shape with this layout is written as F32[3,5]{1,0:(2,2)}, where 1,0 relates to
@@ -120,9 +121,10 @@ element follows the formula above as expected.
 
 XLA's tiling becomes even more flexible by applying it repeatedly.
 
-
 ![](images/xla_array_layout_figure2.png)
-
-Figure 2 

+

+  
+  Figure 2
+

 
 Figure 2 shows how an array of size 4x8 is tiled by two levels of tiling (first
 2x4 then 2x1). We represent this repeated tiling as (2,4)(2,1). Each color
diff --git a/tensorflow/compiler/xla/xla_data.proto b/tensorflow/compiler/xla/xla_data.proto
index 5c21121b98e..e8b6105d3fe 100644
--- a/tensorflow/compiler/xla/xla_data.proto
+++ b/tensorflow/compiler/xla/xla_data.proto
@@ -120,7 +120,7 @@ enum Format {
 }
 
 // Describes a tile used in tiling-based layout. Refer to
-// g3doc/third_party/tensorflow/compiler/xla/g3doc/layout_with_tiling.md for
+// g3doc/third_party/tensorflow/compiler/xla/g3doc/tiled_layout.md for
 // details about tiling-based layout.
 message TileProto {
   // Number of elements in each dimension of the tile. It's ordered from the

From 12f5cd7dde06b2e2e76d39fd114d8ce37231f33b Mon Sep 17 00:00:00 2001
From: Stephan Herhut 
Date: Mon, 22 Jun 2020 06:59:05 -0700
Subject: [PATCH 421/867] Propagate noalias and alignment properties of
 TensorFlow ABI into kernels.

PiperOrigin-RevId: 317646198
Change-Id: I20a649c3127586106c250e213e6b6700fb302495
---
 .../mlir/tools/kernel_gen/cubin_creator.cc    | 23 +++++++++++++------
 1 file changed, 16 insertions(+), 7 deletions(-)

diff --git a/tensorflow/compiler/mlir/tools/kernel_gen/cubin_creator.cc b/tensorflow/compiler/mlir/tools/kernel_gen/cubin_creator.cc
index b534b5a5604..79969a22572 100644
--- a/tensorflow/compiler/mlir/tools/kernel_gen/cubin_creator.cc
+++ b/tensorflow/compiler/mlir/tools/kernel_gen/cubin_creator.cc
@@ -135,11 +135,11 @@ Status LowerTfOpToLhloWithDynamicShapes(mlir::ModuleOp module) {
   return Status::OK();
 }
 
-struct PropagateStaticKnowledge
-    : public mlir::PassWrapper> {
-  explicit PropagateStaticKnowledge(mlir::FunctionType type,
-                                    llvm::ArrayRef same_shape_)
+  explicit PropagateTensorFlowABIKnowledge(mlir::FunctionType type,
+                                           llvm::ArrayRef same_shape_)
       : func_type(type), same_shape(same_shape_) {}
 
   void runOnOperation() override {
@@ -148,6 +148,11 @@ struct PropagateStaticKnowledge
     // we insert constants into the code and replace usages accordingly.
     // We do not change the signature so that we keep a somewhat stable ABI
     // that is easy to undertand by tools.
+    // We also know that tensorflow aligns all allocated pointers by 16, so
+    // we pass this on. Furthermore, we know that arguments never alias. More
+    // precicely, they may only alias (due to reuse) if the kernel does not
+    // read from a position it previously has written to. We express this with
+    // the noalias attribute.
     mlir::LLVM::LLVMFuncOp func = getOperation();
 
     // This only works if the function is local and we can rewrite it.
@@ -172,6 +177,9 @@ struct PropagateStaticKnowledge
         return;
       }
       positions.push_back(arg_pos);
+      // Set alignment and aliasing on the pointers.
+      func.setArgAttr(arg_pos + 1, "llvm.noalias", b.getBoolAttr(true));
+      func.setArgAttr(arg_pos + 1, "llvm.align", b.getIndexAttr(16));
       // Replace the offset with zero. Offset is argument number 3.
       func.getArgument(arg_pos + 2).replaceAllUsesWith(zero);
       // Forward over base_ptr, aligned_ptr, offset, size and stride arguments.
@@ -213,7 +221,7 @@ struct PropagateStaticKnowledge
   llvm::ArrayRef same_shape;
 };
 
-Status PropagateStaticShapeKnowledgeToKernel(
+Status PropagateTensorFlowABIKnowledgeToKernel(
     mlir::ModuleOp module, llvm::ArrayRef same_shape) {
   // Grab the original signature from the single function.
   auto func = *module.getBody()->op_begin();
@@ -228,7 +236,8 @@ Status PropagateStaticShapeKnowledgeToKernel(
                       /*printAfterOnlyOnChange=*/false, llvm::dbgs());
   auto& kernel_pm = pm.nest<::mlir::gpu::GPUModuleOp>();
   kernel_pm.addNestedPass(
-      absl::make_unique(func.getType(), same_shape));
+      absl::make_unique(func.getType(),
+                                                         same_shape));
 
   if (failed(pm.run(module))) {
     return InternalError("Static knowledge propagation failed.");
@@ -263,7 +272,7 @@ StatusOr> tensorflow::kernel_gen::GenerateCubinForTfCode(
   }
   TF_RETURN_IF_ERROR(xla::mlir_gpu::LowerKernelBodiesToNVVM(module.get()));
   TF_RETURN_IF_ERROR(
-      PropagateStaticShapeKnowledgeToKernel(module.get(), same_shape));
+      PropagateTensorFlowABIKnowledgeToKernel(module.get(), same_shape));
 
   mlir::OwningModuleRef kernel_module =
       xla::mlir_gpu::ExtractKernelModule(*module).ValueOrDie();

From bb73be3e363c26d94d44f54edb176006b5403735 Mon Sep 17 00:00:00 2001
From: Alex Zinenko 
Date: Mon, 22 Jun 2020 07:24:04 -0700
Subject: [PATCH 422/867] [mlir][xla] LHLO-to-Affine: simplify DotOp conversion
 with nested builders

MLIR recently introduced a new idiom for constructing loop nests. Use it to
make the legalization of LHLO to affine loops more concise and readable.

PiperOrigin-RevId: 317649515
Change-Id: Idfab27b4655d6df90d940fb7b064ea9941d8a700
---
 .../xla/transforms/lhlo_legalize_to_affine.cc | 89 ++++++++++---------
 1 file changed, 46 insertions(+), 43 deletions(-)

diff --git a/tensorflow/compiler/mlir/xla/transforms/lhlo_legalize_to_affine.cc b/tensorflow/compiler/mlir/xla/transforms/lhlo_legalize_to_affine.cc
index 56b9f5879f6..904a30e847a 100644
--- a/tensorflow/compiler/mlir/xla/transforms/lhlo_legalize_to_affine.cc
+++ b/tensorflow/compiler/mlir/xla/transforms/lhlo_legalize_to_affine.cc
@@ -31,6 +31,17 @@ namespace mlir {
 namespace xla_lhlo {
 namespace {
 
+// Builds an affine loop nest iterating from zeros to "upper_bounds" with unit
+// steps, and populates the body of the innermost loop using "body_builder".
+static void BuildBoundedAffineLoopNest(
+    OpBuilder& builder, Location location, ArrayRef upper_bounds,
+    function_ref body_builder) {
+  SmallVector lower_bounds(upper_bounds.size(), /*Value=*/0);
+  SmallVector steps(upper_bounds.size(), /*Value=*/1);
+  buildAffineLoopNest(builder, location, lower_bounds, upper_bounds, steps,
+                      body_builder);
+}
+
 struct DotOpConverter : public OpRewritePattern {
   using OpRewritePattern::OpRewritePattern;
 
@@ -48,37 +59,29 @@ struct DotOpConverter : public OpRewritePattern {
     if ((lhs_type.getRank() != 2) || (rhs_type.getRank() != 2)) {
       return failure();
     }
-    SmallVector lhs_indices, rhs_indices, result_indices;
-    const auto& loc = op.getLoc();
 
-    // Create the canonical ijk form of matmul.
-    auto forOp = rewriter.create(loc, 0, shape_lhs[0]);
-    lhs_indices.push_back(forOp.getInductionVar());
-    result_indices.push_back(forOp.getInductionVar());
+    LogicalResult map_status = success();
+    auto body_builder = [&](OpBuilder& builder, Location loc, ValueRange ivs) {
+      SmallVector lhs_indices{ivs[0], ivs[2]},
+          rhs_indices{ivs[2], ivs[1]}, result_indices{ivs[0], ivs[1]};
 
-    rewriter.setInsertionPointToStart(forOp.getBody());
-    forOp = rewriter.create(loc, 0, shape_rhs.back());
-    result_indices.push_back(forOp.getInductionVar());
-    rhs_indices.resize(2);
-    rhs_indices[1] = forOp.getInductionVar();
+      auto l = builder.create(loc, lhs, lhs_indices);
+      auto r = builder.create(loc, rhs, rhs_indices);
+      auto result =
+          rewriter.create(loc, op.output(), result_indices);
+      Value op_result = xla_lhlo::XlaOpToStdScalarOp::map(
+          op, element_type, {l, r, result}, &builder);
+      map_status = success(op_result != nullptr);
+      if (failed(map_status)) return;
+      builder.create(loc, op_result, op.output(),
+                                    result_indices);
+    };
 
-    rewriter.setInsertionPointToStart(forOp.getBody());
-    forOp = rewriter.create(loc, 0, shape_rhs.front());
-    lhs_indices.push_back(forOp.getInductionVar());
-    rhs_indices[0] = forOp.getInductionVar();
+    BuildBoundedAffineLoopNest(rewriter, op.getLoc(),
+                               {shape_lhs[0], shape_rhs[1], shape_rhs[0]},
+                               body_builder);
+    if (failed(map_status)) return failure();
 
-    // Construct the innermost loop body.
-    rewriter.setInsertionPointToStart(forOp.getBody());
-    auto l = rewriter.create(loc, lhs, lhs_indices);
-    auto r = rewriter.create(loc, rhs, rhs_indices);
-    auto result =
-        rewriter.create(loc, op.output(), result_indices);
-    Value op_result = xla_lhlo::XlaOpToStdScalarOp::map(
-        op, element_type, {l, r, result}, &rewriter);
-    if (op_result == nullptr) {
-      return failure();
-    }
-    rewriter.create(loc, op_result, op.output(), result_indices);
     rewriter.eraseOp(op);
     return success();
   }
@@ -99,22 +102,22 @@ struct BinaryOpConverter : public OpRewritePattern {
     if (lhs_type.getShape() != rhs_type.getShape()) {
       return failure();
     }
-    const auto& shape = lhs_type.getShape();
-    SmallVector induction_vars;
-    const auto loc = op.getLoc();
-    for (int i = 0; i < shape.size(); ++i) {
-      auto forOp = rewriter.create(loc, 0, shape[i]);
-      induction_vars.push_back(forOp.getInductionVar());
-      rewriter.setInsertionPointToStart(forOp.getBody());
-    }
-    auto l = rewriter.create(loc, lhs, induction_vars);
-    auto r = rewriter.create(loc, rhs, induction_vars);
-    Value opResult = xla_lhlo::XlaOpToStdScalarOp::map(
-        op, element_type, {l, r}, &rewriter);
-    if (opResult == nullptr) {
-      return failure();
-    }
-    rewriter.create(loc, opResult, op.out(), induction_vars);
+
+    LogicalResult map_status = success();
+    auto body_builder = [&](OpBuilder& builder, Location loc,
+                            ValueRange induction_vars) {
+      auto l = builder.create(loc, lhs, induction_vars);
+      auto r = builder.create(loc, rhs, induction_vars);
+      Value op_result = xla_lhlo::XlaOpToStdScalarOp::map(
+          op, element_type, {l, r}, &builder);
+      map_status = success(op_result != nullptr);
+      if (failed(map_status)) return;
+      rewriter.create(loc, op_result, op.out(), induction_vars);
+    };
+
+    BuildBoundedAffineLoopNest(rewriter, op.getLoc(), lhs_type.getShape(),
+                               body_builder);
+    if (failed(map_status)) return failure();
     rewriter.eraseOp(op);
     return success();
   }

From e25fcc8393ccf903ccfe75ca26b3e7dffdd69c5f Mon Sep 17 00:00:00 2001
From: Raman Sarokin 
Date: Mon, 22 Jun 2020 09:41:48 -0700
Subject: [PATCH 423/867] Removed virtual method GPUObjectDescriptor*
 GetGPUDescriptor() for GPUObject.

PiperOrigin-RevId: 317672008
Change-Id: Iae9189569290c5d50dd1d2645dceac5385b2641c
---
 tensorflow/lite/delegates/gpu/cl/arguments.cc | 13 ++++----
 tensorflow/lite/delegates/gpu/cl/arguments.h  |  4 ++-
 tensorflow/lite/delegates/gpu/cl/gpu_object.h |  1 -
 .../lite/delegates/gpu/cl/kernels/prelu.h     | 14 ++++----
 .../lite/delegates/gpu/cl/kernels/winograd.cc | 33 ++++++++++---------
 .../lite/delegates/gpu/cl/linear_storage.cc   |  9 ++---
 .../lite/delegates/gpu/cl/linear_storage.h    | 29 +++++++++++++---
 tensorflow/lite/delegates/gpu/cl/tensor.h     |  3 --
 8 files changed, 62 insertions(+), 44 deletions(-)

diff --git a/tensorflow/lite/delegates/gpu/cl/arguments.cc b/tensorflow/lite/delegates/gpu/cl/arguments.cc
index 53303eab079..6955681a366 100644
--- a/tensorflow/lite/delegates/gpu/cl/arguments.cc
+++ b/tensorflow/lite/delegates/gpu/cl/arguments.cc
@@ -221,8 +221,9 @@ void Arguments::AddObjectRef(const std::string& name, AccessType access_type,
 }
 
 void Arguments::AddObject(const std::string& name, AccessType access_type,
-                          GPUObjectPtr&& object) {
-  objects_[name] = {access_type, std::move(object)};
+                          GPUObjectPtr&& object,
+                          GPUObjectDescriptorPtr&& descriptor_ptr) {
+  objects_[name] = {access_type, std::move(object), std::move(descriptor_ptr)};
 }
 
 void Arguments::AddGPUResources(const std::string& name,
@@ -411,7 +412,8 @@ absl::Status Arguments::Merge(Arguments&& args, const std::string& postfix) {
       return absl::InvalidArgumentError(
           absl::StrCat("Object name collision. Name - ", name));
     }
-    objects_[name] = {v.second.access_type, std::move(v.second.obj_ptr)};
+    objects_[name] = {v.second.access_type, std::move(v.second.obj_ptr),
+                      std::move(v.second.descriptor)};
   }
   for (const auto& v : args.int_values_) {
     AddInt(RenameArg(object_names, postfix, v.first), v.second.value);
@@ -677,7 +679,7 @@ absl::Status Arguments::ResolveSelector(
     desc_ptr = it->second.descriptor.get();
     access_type = it->second.access_type;
   } else if (auto it = objects_.find(object_name); it != objects_.end()) {
-    desc_ptr = it->second.obj_ptr->GetGPUDescriptor();
+    desc_ptr = it->second.descriptor.get();
     access_type = it->second.access_type;
   } else {
     return absl::NotFoundError(
@@ -760,8 +762,7 @@ absl::Status Arguments::ResolveSelectorsPass(
 absl::Status Arguments::AddObjectArgs() {
   for (auto& t : objects_) {
     AddGPUResources(t.first,
-                    t.second.obj_ptr->GetGPUDescriptor()->GetGPUResources(
-                        t.second.access_type));
+                    t.second.descriptor->GetGPUResources(t.second.access_type));
     RETURN_IF_ERROR(SetGPUResources(
         t.first, t.second.obj_ptr->GetGPUResources(t.second.access_type)));
   }
diff --git a/tensorflow/lite/delegates/gpu/cl/arguments.h b/tensorflow/lite/delegates/gpu/cl/arguments.h
index edeab4a603b..4bebb0b2628 100644
--- a/tensorflow/lite/delegates/gpu/cl/arguments.h
+++ b/tensorflow/lite/delegates/gpu/cl/arguments.h
@@ -50,7 +50,8 @@ class Arguments {
   void AddObjectRef(const std::string& name, AccessType access_type,
                     GPUObjectDescriptorPtr&& descriptor_ptr);
   void AddObject(const std::string& name, AccessType access_type,
-                 GPUObjectPtr&& object);
+                 GPUObjectPtr&& object,
+                 GPUObjectDescriptorPtr&& descriptor_ptr);
 
   absl::Status SetInt(const std::string& name, int value);
   absl::Status SetFloat(const std::string& name, float value);
@@ -162,6 +163,7 @@ class Arguments {
   struct ObjectArg {
     AccessType access_type;
     GPUObjectPtr obj_ptr;
+    GPUObjectDescriptorPtr descriptor;
   };
   std::map objects_;
 };
diff --git a/tensorflow/lite/delegates/gpu/cl/gpu_object.h b/tensorflow/lite/delegates/gpu/cl/gpu_object.h
index fec8999e2bc..a6d28436872 100644
--- a/tensorflow/lite/delegates/gpu/cl/gpu_object.h
+++ b/tensorflow/lite/delegates/gpu/cl/gpu_object.h
@@ -149,7 +149,6 @@ class GPUObject {
   GPUObject(const GPUObject&) = delete;
   GPUObject& operator=(const GPUObject&) = delete;
   virtual ~GPUObject() = default;
-  virtual const GPUObjectDescriptor* GetGPUDescriptor() const = 0;
   virtual GPUResourcesWithValue GetGPUResources(
       AccessType access_type) const = 0;
 };
diff --git a/tensorflow/lite/delegates/gpu/cl/kernels/prelu.h b/tensorflow/lite/delegates/gpu/cl/kernels/prelu.h
index 68e4c7b7626..80ee49f77ca 100644
--- a/tensorflow/lite/delegates/gpu/cl/kernels/prelu.h
+++ b/tensorflow/lite/delegates/gpu/cl/kernels/prelu.h
@@ -69,17 +69,17 @@ absl::Status CreatePReLU(const CreationContext& creation_context,
 template 
 absl::Status PReLU::UploadParameters(
     const tflite::gpu::Tensor& parameters, CLContext* context) {
-  LinearStorageCreateInfo create_info;
-  create_info.storage_type =
+  TensorLinearDescriptor desc;
+  desc.storage_type =
       DeduceLinearStorageType(definition_.GetPrimaryStorageType());
-  create_info.data_type = definition_.GetPrimaryDataType();
-  RETURN_IF_ERROR(
-      CreateLinearStorage(create_info, parameters, context, &alpha_));
+  desc.element_type = definition_.GetPrimaryDataType();
+  RETURN_IF_ERROR(CreateLinearStorage(desc, parameters, context, &alpha_));
 
   LinearStorage lt;
-  RETURN_IF_ERROR(CreateLinearStorage(create_info, parameters, context, <));
+  RETURN_IF_ERROR(CreateLinearStorage(desc, parameters, context, <));
   args_.AddObject("alpha", AccessType::READ,
-                  absl::make_unique(std::move(lt)));
+                  absl::make_unique(std::move(lt)),
+                  absl::make_unique(desc));
 
   return absl::OkStatus();
 }
diff --git a/tensorflow/lite/delegates/gpu/cl/kernels/winograd.cc b/tensorflow/lite/delegates/gpu/cl/kernels/winograd.cc
index 2dcb72637ec..d38b72e61a6 100644
--- a/tensorflow/lite/delegates/gpu/cl/kernels/winograd.cc
+++ b/tensorflow/lite/delegates/gpu/cl/kernels/winograd.cc
@@ -378,14 +378,15 @@ absl::Status Winograd4x4To36::UploadBt(CLContext* context) {
     bt_aligned.data[y * 8 + 7] = 0.0f;
   }
 
-  LinearStorageCreateInfo create_info;
-  create_info.storage_type = LinearStorageType::TEXTURE_2D;
-  create_info.data_type = definition_.GetDataType();
+  TensorLinearDescriptor desc;
+  desc.storage_type = LinearStorageType::TEXTURE_2D;
+  desc.element_type = definition_.GetDataType();
 
   LinearStorage lt;
-  RETURN_IF_ERROR(CreateLinearStorage(create_info, bt_aligned, context, <));
+  RETURN_IF_ERROR(CreateLinearStorage(desc, bt_aligned, context, <));
   args_.AddObject("bt", AccessType::READ,
-                  absl::make_unique(std::move(lt)));
+                  absl::make_unique(std::move(lt)),
+                  absl::make_unique(desc));
   return absl::OkStatus();
 }
 
@@ -492,13 +493,14 @@ absl::Status Winograd36To4x4::UploadAt(CLContext* context) {
     at_aligned.data[y * 8 + 7] = 0.0f;
   }
 
-  LinearStorageCreateInfo create_info;
-  create_info.storage_type = LinearStorageType::TEXTURE_2D;
-  create_info.data_type = definition_.GetDataType();
+  TensorLinearDescriptor desc;
+  desc.storage_type = LinearStorageType::TEXTURE_2D;
+  desc.element_type = definition_.GetDataType();
   LinearStorage lt;
-  RETURN_IF_ERROR(CreateLinearStorage(create_info, at_aligned, context, <));
+  RETURN_IF_ERROR(CreateLinearStorage(desc, at_aligned, context, <));
   args_.AddObject("at", AccessType::READ,
-                  absl::make_unique(std::move(lt)));
+                  absl::make_unique(std::move(lt)),
+                  absl::make_unique(desc));
   return absl::OkStatus();
 }
 
@@ -550,14 +552,15 @@ absl::Status CreateWinograd36To4x4(
     const tflite::gpu::Tensor& biases,
     Winograd36To4x4* result) {
   *result = Winograd36To4x4(definition);
-  LinearStorageCreateInfo create_info;
-  create_info.storage_type = LinearStorageType::TEXTURE_2D;
-  create_info.data_type = definition.GetDataType();
+  TensorLinearDescriptor desc;
+  desc.storage_type = LinearStorageType::TEXTURE_2D;
+  desc.element_type = definition.GetDataType();
   LinearStorage lt;
   RETURN_IF_ERROR(
-      CreateLinearStorage(create_info, biases, creation_context.context, <));
+      CreateLinearStorage(desc, biases, creation_context.context, <));
   result->args_.AddObject("biases", AccessType::READ,
-                          absl::make_unique(std::move(lt)));
+                          absl::make_unique(std::move(lt)),
+                          absl::make_unique(desc));
   return result->UploadAt(creation_context.context);
 }
 
diff --git a/tensorflow/lite/delegates/gpu/cl/linear_storage.cc b/tensorflow/lite/delegates/gpu/cl/linear_storage.cc
index 84d91b9136e..47504a34c2b 100644
--- a/tensorflow/lite/delegates/gpu/cl/linear_storage.cc
+++ b/tensorflow/lite/delegates/gpu/cl/linear_storage.cc
@@ -76,10 +76,7 @@ absl::Status TensorLinearDescriptor::PerformReadSelector(
 
 LinearStorage::LinearStorage(int depth, LinearStorageType storage_type,
                              DataType data_type)
-    : depth_(depth), storage_type_(storage_type), data_type_(data_type) {
-  desc_.storage_type = storage_type;
-  desc_.element_type = data_type;
-}
+    : depth_(depth), storage_type_(storage_type), data_type_(data_type) {}
 
 LinearStorage::LinearStorage(LinearStorage&& storage)
     : GPUObject(std::move(storage)),
@@ -89,8 +86,7 @@ LinearStorage::LinearStorage(LinearStorage&& storage)
       depth_(storage.depth_),
       name_(std::move(storage.name_)),
       storage_type_(storage.storage_type_),
-      data_type_(storage.data_type_),
-      desc_(storage.desc_) {
+      data_type_(storage.data_type_) {
   storage.memory_ = nullptr;
 }
 
@@ -103,7 +99,6 @@ LinearStorage& LinearStorage::operator=(LinearStorage&& storage) {
     name_ = std::move(storage.name_);
     std::swap(storage_type_, storage.storage_type_);
     std::swap(data_type_, storage.data_type_);
-    desc_ = storage.desc_;
     GPUObject::operator=(std::move(storage));
   }
   return *this;
diff --git a/tensorflow/lite/delegates/gpu/cl/linear_storage.h b/tensorflow/lite/delegates/gpu/cl/linear_storage.h
index 474c5652db2..c7f55e1d91d 100644
--- a/tensorflow/lite/delegates/gpu/cl/linear_storage.h
+++ b/tensorflow/lite/delegates/gpu/cl/linear_storage.h
@@ -92,9 +92,6 @@ class LinearStorage : public GPUObject {
   std::string ReadLinearFLT4(const std::string& z_coord) const;
   std::string GetDeclaration() const;
 
-  const GPUObjectDescriptor* GetGPUDescriptor() const override {
-    return &desc_;
-  }
   GPUResourcesWithValue GetGPUResources(AccessType access_type) const override;
 
  private:
@@ -115,7 +112,6 @@ class LinearStorage : public GPUObject {
   std::string name_;
   LinearStorageType storage_type_;
   DataType data_type_;
-  TensorLinearDescriptor desc_;
 };
 
 absl::Status CreateBufferLinearStorage(int size, DataType data_type, void* data,
@@ -152,6 +148,31 @@ absl::Status CreateLinearStorage(const LinearStorageCreateInfo& creation_info,
   return absl::OkStatus();
 }
 
+template 
+absl::Status CreateLinearStorage(const TensorLinearDescriptor& descriptor,
+                                 const tflite::gpu::Tensor& tensor,
+                                 CLContext* context, LinearStorage* result) {
+  LinearStorageCreateInfo creation_info;
+  creation_info.storage_type = descriptor.storage_type;
+  creation_info.data_type = descriptor.element_type;
+  int size = creation_info.aligned_size != 0 ? creation_info.aligned_size
+                                             : tensor.shape.v;
+  const int depth = DivideRoundUp(size, 4);
+  if (creation_info.data_type == DataType::FLOAT32) {
+    std::vector gpu_data(depth);
+    CopyLinearFLT4(tensor, absl::MakeSpan(gpu_data));
+    RETURN_IF_ERROR(CreateLinearStorage(creation_info, depth, gpu_data.data(),
+                                        context, result));
+  } else {
+    std::vector gpu_data(depth);
+    CopyLinearFLT4(tensor, absl::MakeSpan(gpu_data));
+    RETURN_IF_ERROR(CreateLinearStorage(creation_info, depth, gpu_data.data(),
+                                        context, result));
+  }
+  result->SetName(creation_info.name);
+  return absl::OkStatus();
+}
+
 }  // namespace cl
 }  // namespace gpu
 }  // namespace tflite
diff --git a/tensorflow/lite/delegates/gpu/cl/tensor.h b/tensorflow/lite/delegates/gpu/cl/tensor.h
index c1b0b14709f..7de42a810ec 100644
--- a/tensorflow/lite/delegates/gpu/cl/tensor.h
+++ b/tensorflow/lite/delegates/gpu/cl/tensor.h
@@ -58,9 +58,6 @@ class Tensor : public GPUObject {
 
   virtual ~Tensor() { Release(); }
 
-  const GPUObjectDescriptor* GetGPUDescriptor() const override {
-    return &descriptor_;
-  }
   GPUResourcesWithValue GetGPUResources(AccessType access_type) const override;
 
   int Width() const { return shape_.w; }

From ea00842c91376ffee148f54ddb1ec7f09bfc95db Mon Sep 17 00:00:00 2001
From: Vo Van Nghia 
Date: Tue, 23 Jun 2020 00:00:36 +0700
Subject: [PATCH 424/867] Add the licenses

---
 .../filesystem/plugins/gcs/gcs_filesystem_test.cc  | 14 ++++++++++++++
 1 file changed, 14 insertions(+)

diff --git a/tensorflow/c/experimental/filesystem/plugins/gcs/gcs_filesystem_test.cc b/tensorflow/c/experimental/filesystem/plugins/gcs/gcs_filesystem_test.cc
index eb0fbfc33f7..6dc5e164a1c 100644
--- a/tensorflow/c/experimental/filesystem/plugins/gcs/gcs_filesystem_test.cc
+++ b/tensorflow/c/experimental/filesystem/plugins/gcs/gcs_filesystem_test.cc
@@ -1,3 +1,17 @@
+/* Copyright 2020 The TensorFlow Authors. All Rights Reserved.
+
+Licensed under the Apache License, Version 2.0 (the "License");
+you may not use this file except in compliance with the License.
+You may obtain a copy of the License at
+
+    http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software
+distributed under the License is distributed on an "AS IS" BASIS,
+WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+See the License for the specific language governing permissions and
+limitations under the License.
+==============================================================================*/
 #include "gtest/gtest.h"
 #include "tensorflow/c/experimental/filesystem/filesystem_interface.h"
 #include "tensorflow/c/tf_status_helper.h"

From 22f6939be9cc23d14610c3e21fd30f7345c1b125 Mon Sep 17 00:00:00 2001
From: Feng Liu 
Date: Mon, 22 Jun 2020 09:43:32 -0700
Subject: [PATCH 425/867] Apply default quantization parameter before
 quantization pass

By this order, the default quantization parameter is only applied on the
activations and the weight quantization parameter will use the parameters from
the weight content.

PiperOrigin-RevId: 317672365
Change-Id: Ib7b02ae19105db124721242ea51a5ccc1d5aa68e
---
 tensorflow/compiler/mlir/lite/tf_tfl_passes.cc | 14 +++++---------
 1 file changed, 5 insertions(+), 9 deletions(-)

diff --git a/tensorflow/compiler/mlir/lite/tf_tfl_passes.cc b/tensorflow/compiler/mlir/lite/tf_tfl_passes.cc
index fed2896035b..3fa2eae42f2 100644
--- a/tensorflow/compiler/mlir/lite/tf_tfl_passes.cc
+++ b/tensorflow/compiler/mlir/lite/tf_tfl_passes.cc
@@ -39,22 +39,18 @@ namespace tensorflow {
 void AddQuantizationPasses(const mlir::TFL::QuantizationSpecs& quant_specs,
                            mlir::OpPassManager* pass_manager) {
   pass_manager->addPass(mlir::TFL::CreatePrepareQuantizePass(quant_specs));
-  pass_manager->addPass(mlir::TFL::CreateQuantizePass());
-  bool emit_quant_adaptor_ops =
-      quant_specs.inference_type != quant_specs.inference_input_type;
-  pass_manager->addPass(
-      mlir::TFL::CreatePostQuantizePass(emit_quant_adaptor_ops));
-
   if (quant_specs.default_ranges.first.hasValue() ||
       quant_specs.default_ranges.second.hasValue()) {
     pass_manager->addPass(mlir::TFL::CreateDefaultQuantParamsPass(
         quant_specs.default_ranges.first.getValueOr(0.0),
         quant_specs.default_ranges.second.getValueOr(0.0),
         quant_specs.IsSignedInferenceType()));
-    pass_manager->addPass(mlir::TFL::CreateQuantizePass());
-    pass_manager->addPass(
-        mlir::TFL::CreatePostQuantizePass(emit_quant_adaptor_ops));
   }
+  pass_manager->addPass(mlir::TFL::CreateQuantizePass());
+  bool emit_quant_adaptor_ops =
+      quant_specs.inference_type != quant_specs.inference_input_type;
+  pass_manager->addPass(
+      mlir::TFL::CreatePostQuantizePass(emit_quant_adaptor_ops));
 }
 
 void AddTFToTFLConversionPasses(const mlir::TFL::PassConfig& pass_config,

From eb32bf1ae46571375b00c9e3e146dad7706286ce Mon Sep 17 00:00:00 2001
From: Bruce Fontaine 
Date: Mon, 22 Jun 2020 09:43:37 -0700
Subject: [PATCH 426/867] Clean up _set_prefetch_on_host method from
 TPUStrategy.

PiperOrigin-RevId: 317672388
Change-Id: Icb03401a792df4d071659f101ba06edf59010577
---
 .../custom_training_loop_input_test.py        | 10 +--
 tensorflow/python/distribute/tpu_strategy.py  | 43 +++++++-----
 .../python/distribute/tpu_strategy_test.py    | 69 +++++++++++++++++--
 3 files changed, 98 insertions(+), 24 deletions(-)

diff --git a/tensorflow/python/distribute/custom_training_loop_input_test.py b/tensorflow/python/distribute/custom_training_loop_input_test.py
index 5660b5839ce..748cb7834fc 100644
--- a/tensorflow/python/distribute/custom_training_loop_input_test.py
+++ b/tensorflow/python/distribute/custom_training_loop_input_test.py
@@ -454,10 +454,11 @@ class InputIterationTest(test.TestCase, parameterized.TestCase,
       ))
   def testDistributeDatasetHostPrefetch(self, distribution):
     data = [5., 6., 7., 8.]
-    distribution.extended._set_prefetch_on_host(True)  # pylint: disable=protected-access
     input_iterator = iter(
         distribution.experimental_distribute_dataset(
-            get_dataset_from_tensor_slices(data).batch(2)))
+            get_dataset_from_tensor_slices(data).batch(2),
+            distribute_lib.InputOptions(
+                experimental_prefetch_to_device=False)))
 
     local_results = distribution.experimental_local_results(
         input_iterator.get_next())
@@ -473,10 +474,11 @@ class InputIterationTest(test.TestCase, parameterized.TestCase,
       ))
   def testDistributeDatasetFunctionHostPrefetch(self, distribution):
     data = [5., 6., 7., 8.]
-    distribution.extended._set_prefetch_on_host(True)  # pylint: disable=protected-access
     input_iterator = iter(
         distribution.experimental_distribute_datasets_from_function(
-            lambda _: get_dataset_from_tensor_slices(data)))
+            lambda _: get_dataset_from_tensor_slices(data),
+            distribute_lib.InputOptions(
+                experimental_prefetch_to_device=False)))
 
     local_results = distribution.experimental_local_results(
         input_iterator.get_next())
diff --git a/tensorflow/python/distribute/tpu_strategy.py b/tensorflow/python/distribute/tpu_strategy.py
index dcd1671841f..4b3c4be0ccd 100644
--- a/tensorflow/python/distribute/tpu_strategy.py
+++ b/tensorflow/python/distribute/tpu_strategy.py
@@ -47,12 +47,14 @@ from tensorflow.python.framework import constant_op
 from tensorflow.python.framework import device_spec
 from tensorflow.python.framework import dtypes
 from tensorflow.python.framework import ops
+from tensorflow.python.framework import sparse_tensor
 from tensorflow.python.framework import tensor_shape
 from tensorflow.python.framework import tensor_util
 from tensorflow.python.ops import array_ops
 from tensorflow.python.ops import control_flow_ops
 from tensorflow.python.ops import math_ops
 from tensorflow.python.ops import resource_variable_ops
+from tensorflow.python.ops.ragged import ragged_tensor
 from tensorflow.python.tpu import device_assignment as device_assignment_lib  # pylint: disable=unused-import
 from tensorflow.python.tpu import tpu
 from tensorflow.python.tpu import tpu_strategy_util
@@ -515,7 +517,6 @@ class TPUExtended(distribute_lib.StrategyExtendedV1):
     self._require_static_shapes = True
 
     self.experimental_enable_get_next_as_optional = True
-    self._prefetch_to_device = True
 
     self._logical_device_stack = [0]
 
@@ -527,16 +528,6 @@ class TPUExtended(distribute_lib.StrategyExtendedV1):
           context.async_wait()
       atexit.register(async_wait)
 
-  # TODO(bfontain): Remove once a proper dataset API exists for prefetching
-  # a dataset to multiple devices exists.
-  # If value is true, this forces prefetch of data to the host's memeory rather
-  # than the individual TPU device's memory. This is needed when using for TPU
-  # Embeddings as a) sparse tensors cannot be prefetched to the TPU device
-  # memory and b) TPU Embedding enqueue operation are CPU ops and this avoids
-  # a copy back to the host for dense tensors
-  def _set_prefetch_on_host(self, value):
-    self._prefetch_to_device = not value
-
   def _validate_colocate_with_variable(self, colocate_with_variable):
     distribute_utils. validate_colocate(colocate_with_variable, self)
 
@@ -575,17 +566,32 @@ class TPUExtended(distribute_lib.StrategyExtendedV1):
         session)
 
   def _get_input_workers(self, options):
-    prefetch_to_device = self._prefetch_to_device
-    if options:
-      prefetch_to_device = options.experimental_prefetch_to_device
-    if prefetch_to_device:
+    if not options or options.experimental_prefetch_to_device:
       return input_lib.InputWorkers(
           tuple(self._device_input_worker_devices.items()))
     else:
       return input_lib.InputWorkers(
           tuple(self._host_input_worker_devices.items()))
 
+  def _check_spec(self, element_spec):
+    if isinstance(element_spec, values.PerReplicaSpec):
+      element_spec = element_spec._component_specs  # pylint: disable=protected-access
+    specs = nest.flatten_with_joined_string_paths(element_spec)
+    for path, spec in specs:
+      if isinstance(spec, (sparse_tensor.SparseTensorSpec,
+                           ragged_tensor.RaggedTensorSpec)):
+        raise ValueError(
+            "Found tensor {} with spec {}. TPUStrategy does not support "
+            "distributed datasets with device prefetch when using sparse or "
+            "ragged tensors. If you indend to use sparse or ragged tensors, "
+            "please pass a tf.distribute.InputOptions object with "
+            "experimental_prefetch_to_device set to False to your dataset "
+            "distribution function.".format(path, type(spec)))
+
   def _experimental_distribute_dataset(self, dataset, options):
+    if options is None or options.experimental_prefetch_to_device:
+      self._check_spec(dataset.element_spec)
+
     return input_lib.get_distributed_dataset(
         dataset,
         self._get_input_workers(options),
@@ -603,12 +609,17 @@ class TPUExtended(distribute_lib.StrategyExtendedV1):
           input_pipeline_id=i,
           num_replicas_in_sync=self._num_replicas_in_sync))
 
-    return input_lib.get_distributed_datasets_from_function(
+    distributed_dataset = input_lib.get_distributed_datasets_from_function(
         dataset_fn,
         input_workers,
         input_contexts,
         self._container_strategy())
 
+    # We can only check after the dataset_fn is called.
+    if options is None or options.experimental_prefetch_to_device:
+      self._check_spec(distributed_dataset.element_spec)
+    return distributed_dataset
+
   def _experimental_distribute_values_from_function(self, value_fn):
     per_replica_values = []
     for replica_id in range(self._num_replicas_in_sync):
diff --git a/tensorflow/python/distribute/tpu_strategy_test.py b/tensorflow/python/distribute/tpu_strategy_test.py
index 4070336aae8..5e47e750d87 100644
--- a/tensorflow/python/distribute/tpu_strategy_test.py
+++ b/tensorflow/python/distribute/tpu_strategy_test.py
@@ -44,6 +44,7 @@ from tensorflow.python.ops import embedding_ops
 from tensorflow.python.ops import math_ops
 from tensorflow.python.ops import random_ops
 from tensorflow.python.ops import variables
+from tensorflow.python.ops.ragged import ragged_tensor
 from tensorflow.python.platform import flags
 from tensorflow.python.platform import tf_logging as logging
 from tensorflow.python.tpu import device_assignment as device_assignment_lib
@@ -475,9 +476,11 @@ class TPUStrategyTest(test.TestCase, parameterized.TestCase):
 
       return dataset.map(make_sparse)
 
-    strategy.extended._set_prefetch_on_host(True)  # pylint: disable=protected-access
     dataset = iter(
-        strategy.experimental_distribute_datasets_from_function(dataset_fn))
+        strategy.experimental_distribute_datasets_from_function(
+            dataset_fn,
+            distribute_lib.InputOptions(
+                experimental_prefetch_to_device=False)))
 
     result = sparse_lookup(dataset)
     self.assertAllEqual(result,
@@ -520,9 +523,11 @@ class TPUStrategyTest(test.TestCase, parameterized.TestCase):
 
       return dataset.map(make_sparse)
 
-    strategy.extended._set_prefetch_on_host(True)  # pylint: disable=protected-access
     dataset = iter(
-        strategy.experimental_distribute_datasets_from_function(dataset_fn))
+        strategy.experimental_distribute_datasets_from_function(
+            dataset_fn,
+            options=distribute_lib.InputOptions(
+                experimental_prefetch_to_device=False)))
 
     result = sparse_lookup(dataset)
     self.assertAllEqual(result, [[0.0, 2.0], [1.5, 5.0]])
@@ -594,5 +599,61 @@ class TPUStrategyDataPrefetchTest(test.TestCase):
         dataset_item.values[0].device)
     self.assertEqual(dataset_location.device_type, "CPU")
 
+  def test_prefetch_to_device_sparse_dataset(self):
+    strategy = get_tpu_strategy()
+    # Values here aren't important.
+    dataset = dataset_ops.Dataset.from_tensors(
+        sparse_tensor.SparseTensor(indices=[[0, 0], [0, 1], [1, 0]],
+                                   values=[1, 2, 3],
+                                   dense_shape=[2, 2]))
+    dataset = dataset.repeat()
+    dataset = dataset.batch(strategy.num_replicas_in_sync)
+
+    with self.assertRaisesRegex(ValueError, "TPUStrategy does not support"):
+      iter(strategy.experimental_distribute_dataset(dataset))
+
+  def test_prefetch_to_device_ragged_dataset(self):
+    strategy = get_tpu_strategy()
+    # Values here aren't important.
+    dataset = dataset_ops.Dataset.from_tensors(
+        ragged_tensor.RaggedTensor.from_row_splits(
+            values=[1, 2, 3],
+            row_splits=[0, 2, 3]))
+    dataset = dataset.repeat()
+    dataset = dataset.batch(strategy.num_replicas_in_sync)
+
+    with self.assertRaisesRegex(ValueError, "TPUStrategy does not support"):
+      iter(strategy.experimental_distribute_dataset(dataset))
+
+  def test_prefetch_to_device_sparse_dataset_fn(self):
+    strategy = get_tpu_strategy()
+    def dataset_fn(ctx):
+      del ctx
+      # Values here aren't important.
+      dataset = dataset_ops.Dataset.from_tensors(
+          sparse_tensor.SparseTensor(indices=[[0, 0], [0, 1], [1, 0]],
+                                     values=[1, 2, 3],
+                                     dense_shape=[2, 2]))
+      dataset = dataset.repeat()
+      return dataset.batch(strategy.num_replicas_in_sync)
+
+    with self.assertRaisesRegex(ValueError, "TPUStrategy does not support"):
+      iter(strategy.experimental_distribute_datasets_from_function(dataset_fn))
+
+  def test_prefetch_to_device_ragged_dataset_fn(self):
+    strategy = get_tpu_strategy()
+    def dataset_fn(ctx):
+      del ctx
+      # Values here aren't important.
+      dataset = dataset_ops.Dataset.from_tensors(
+          ragged_tensor.RaggedTensor.from_row_splits(
+              values=[1, 2, 3],
+              row_splits=[0, 2, 3]))
+      dataset = dataset.repeat()
+      return dataset.batch(strategy.num_replicas_in_sync)
+
+    with self.assertRaisesRegex(ValueError, "TPUStrategy does not support"):
+      iter(strategy.experimental_distribute_datasets_from_function(dataset_fn))
+
 if __name__ == "__main__":
   test.main()

From 7f12fa50f1615e4a2e78482a39e1a5d2db809734 Mon Sep 17 00:00:00 2001
From: Christian Sigg 
Date: Mon, 22 Jun 2020 09:46:09 -0700
Subject: [PATCH 427/867] Update NCCL to v2.7.3.

PiperOrigin-RevId: 317672859
Change-Id: Ice6b6ac0875f1b6f8daa6ad9d9539621b22e6666
---
 tensorflow/workspace.bzl       |   8 +-
 third_party/nccl/archive.BUILD |   1 +
 third_party/nccl/archive.patch | 181 ++++++++++++++++++++++++++++++++-
 3 files changed, 183 insertions(+), 7 deletions(-)

diff --git a/tensorflow/workspace.bzl b/tensorflow/workspace.bzl
index e6a15b422eb..73698987a08 100755
--- a/tensorflow/workspace.bzl
+++ b/tensorflow/workspace.bzl
@@ -802,11 +802,11 @@ def tf_repositories(path_prefix = "", tf_repo_name = ""):
         name = "nccl_archive",
         build_file = clean_dep("//third_party:nccl/archive.BUILD"),
         patch_file = clean_dep("//third_party/nccl:archive.patch"),
-        sha256 = "7ff66aca18392b162829612e02c00b123a58ec35869334f72d7e5afaf5ea4a13",
-        strip_prefix = "nccl-3701130b3c1bcdb01c14b3cb70fe52498c1e82b7",
+        sha256 = "67e15ce3d12ba9ea1e0cb239599202b0f61c146149699341043c072de388e90a",
+        strip_prefix = "nccl-5949d96f36d050e59d05872f8bbffd2549318e95",
         urls = [
-            "https://storage.googleapis.com/mirror.tensorflow.org/github.com/nvidia/nccl/archive/3701130b3c1bcdb01c14b3cb70fe52498c1e82b7.tar.gz",
-            "https://github.com/nvidia/nccl/archive/3701130b3c1bcdb01c14b3cb70fe52498c1e82b7.tar.gz",
+            "https://storage.googleapis.com/mirror.tensorflow.org/github.com/nvidia/nccl/archive/5949d96f36d050e59d05872f8bbffd2549318e95.tar.gz",
+            "https://github.com/nvidia/nccl/archive/5949d96f36d050e59d05872f8bbffd2549318e95.tar.gz",
         ],
     )
 
diff --git a/third_party/nccl/archive.BUILD b/third_party/nccl/archive.BUILD
index 9b0c9bdda1d..028b348caff 100644
--- a/third_party/nccl/archive.BUILD
+++ b/third_party/nccl/archive.BUILD
@@ -46,6 +46,7 @@ gen_device_srcs(
         "src/collectives/device/broadcast.cu.cc",
         "src/collectives/device/reduce.cu.cc",
         "src/collectives/device/reduce_scatter.cu.cc",
+        "src/collectives/device/sendrecv.cu.cc",
     ],
 )
 
diff --git a/third_party/nccl/archive.patch b/third_party/nccl/archive.patch
index ea0ead71fb2..94ef48d00e8 100644
--- a/third_party/nccl/archive.patch
+++ b/third_party/nccl/archive.patch
@@ -22,6 +22,10 @@ diff --git a/src/collectives/device/reduce_scatter.cu b/src/collectives/device/r
 similarity index 100%
 rename from src/collectives/device/reduce_scatter.cu
 rename to src/collectives/device/reduce_scatter.cu.cc
+diff --git a/src/collectives/device/sendrecv.cu b/src/collectives/device/sendrecv.cu.cc
+similarity index 100%
+rename from src/collectives/device/sendrecv.cu
+rename to src/collectives/device/sendrecv.cu.cc
 diff --git a/src/nccl.h.in b/src/nccl.h
 similarity index 98%
 rename from src/nccl.h.in
@@ -38,12 +42,183 @@ index 985274e..7ebb1e1 100644
 -#define NCCL_PATCH ${nccl:Patch}
 -#define NCCL_SUFFIX "${nccl:Suffix}"
 +#define NCCL_MAJOR 2
-+#define NCCL_MINOR 5
-+#define NCCL_PATCH 7
++#define NCCL_MINOR 7
++#define NCCL_PATCH 3
 +#define NCCL_SUFFIX ""
  
 -#define NCCL_VERSION_CODE ${nccl:Version}
-+#define NCCL_VERSION_CODE 2507
++#define NCCL_VERSION_CODE 2703
  #define NCCL_VERSION(X,Y,Z) ((X) * 1000 + (Y) * 100 + (Z))
  
  #ifdef __cplusplus
+See https://github.com/NVIDIA/nccl/pull/322.patch
+From 410d341bd4569f60282576daa5c991717dbd560e Mon Sep 17 00:00:00 2001
+From: Danilo 
+Date: Tue, 14 Apr 2020 14:52:42 +0200
+Subject: [PATCH 1/2] Fix memory leak in xml.cc.
+
+This patch fixes the memory leak documented in
+https://github.com/NVIDIA/nccl/issues/321, where one of the buffers
+allocated by realpath(), inside getPciPath() is not freed.
+
+The memory management aspect of this function also seemed odd and
+unecessary, as the realpath() function is documented to only write up to
+PATH_MAX bytes to the buffer passed to it, meaning we don't need dynamic
+memory allocation at all. I also changed the function signature of
+getPciPath to enforce the use of a fixed-size buffer.
+---
+ src/graph/xml.cc | 23 ++++++++++++-----------
+ 1 file changed, 12 insertions(+), 11 deletions(-)
+
+diff --git a/src/graph/xml.cc b/src/graph/xml.cc
+index 550cfcd0c..8fea91950 100644
+--- a/src/graph/xml.cc
++++ b/src/graph/xml.cc
+@@ -323,12 +323,14 @@ ncclResult_t ncclTopoGetXmlFromFile(const char* xmlTopoFile, struct ncclXml* xml
+ static void memcpylower(char* dst, const char* src, const size_t size) {
+   for (int i=0; iparent;
+   if (parent == NULL) {
+-    if (path == NULL) NCCLCHECK(getPciPath(busId, &path));
++    NCCLCHECK(getPciPath(busId, path));
+ 
+     // Save that for later in case next step is a CPU
+     char numaIdStr[MAX_STR_LEN];
+@@ -544,7 +546,6 @@ ncclResult_t ncclTopoGetXmlFromSys(struct ncclXmlNode* pciNode, struct ncclXml*
+   } else if (strcmp(parent->name, "cpu") == 0) {
+     NCCLCHECK(ncclTopoGetXmlFromCpu(parent, xml));
+   }
+-  free(path);
+   return ncclSuccess;
+ }
+ 
+@@ -640,8 +641,8 @@ ncclResult_t ncclTopoGetXmlFromGpu(struct ncclXmlNode* pciNode, nvmlDevice_t nvm
+     if (index == -1) {
+       const char* busId;
+       NCCLCHECK(xmlGetAttr(sub, "target", &busId));
+-      char* path;
+-      NCCLCHECK(getPciPath(busId, &path));
++      char path[PATH_MAX+1];
++      NCCLCHECK(getPciPath(busId, path));
+       NCCLCHECK(ncclTopoSetAttrFromSys(sub, path, "class", "tclass"));
+     }
+   }
+
+From f02d51952ac587237ea5f7c607a5b379381d09d7 Mon Sep 17 00:00:00 2001
+From: Danilo 
+Date: Tue, 14 Apr 2020 22:17:49 +0200
+Subject: [PATCH 2/2] Performance tweaks in ncclTopoGetXmlFromSys.
+
+Reduce the number of getPciPath calls to a single one per invocation
+and split the function in two so that the large `path` buffer does
+not linger the in the stack during recursive calls.
+---
+ src/graph/xml.cc | 17 +++++++++++------
+ 1 file changed, 11 insertions(+), 6 deletions(-)
+
+diff --git a/src/graph/xml.cc b/src/graph/xml.cc
+index 8fea91950..42eb68a4b 100644
+--- a/src/graph/xml.cc
++++ b/src/graph/xml.cc
+@@ -460,20 +460,21 @@ int checkBDFFormat(char* bdf) {
+   return 1;
+ }
+ 
+-ncclResult_t ncclTopoGetXmlFromSys(struct ncclXmlNode* pciNode, struct ncclXml* xml) {
++ncclResult_t ncclTopoGetXmlNodeFromSys(struct ncclXmlNode* pciNode,
++                                       struct ncclXml* xml,
++                                       struct ncclXmlNode** return_parent) {
+   // Fill info, then parent
+   const char* busId;
+   NCCLCHECK(xmlGetAttr(pciNode, "busid", &busId));
+   char path[PATH_MAX+1];
++  NCCLCHECK(getPciPath(busId, path));
+   int index;
+   NCCLCHECK(xmlGetAttrIndex(pciNode, "class", &index));
+   if (index == -1) {
+-    NCCLCHECK(getPciPath(busId, path));
+     NCCLCHECK(ncclTopoSetAttrFromSys(pciNode, path, "class", "class"));
+   }
+   NCCLCHECK(xmlGetAttrIndex(pciNode, "link_speed", &index));
+   if (index == -1) {
+-    NCCLCHECK(getPciPath(busId, path));
+     char deviceSpeedStr[MAX_STR_LEN];
+     float deviceSpeed;
+     NCCLCHECK(ncclTopoGetStrFromSys(path, "max_link_speed", deviceSpeedStr));
+@@ -486,7 +487,6 @@ ncclResult_t ncclTopoGetXmlFromSys(struct ncclXmlNode* pciNode, struct ncclXml*
+   }
+   NCCLCHECK(xmlGetAttrIndex(pciNode, "link_width", &index));
+   if (index == -1) {
+-    NCCLCHECK(getPciPath(busId, path));
+     char strValue[MAX_STR_LEN];
+     NCCLCHECK(ncclTopoGetStrFromSys(path, "max_link_width", strValue));
+     int deviceWidth = strtol(strValue, NULL, 0);
+@@ -496,8 +496,6 @@ ncclResult_t ncclTopoGetXmlFromSys(struct ncclXmlNode* pciNode, struct ncclXml*
+   }
+   struct ncclXmlNode* parent = pciNode->parent;
+   if (parent == NULL) {
+-    NCCLCHECK(getPciPath(busId, path));
+-
+     // Save that for later in case next step is a CPU
+     char numaIdStr[MAX_STR_LEN];
+     NCCLCHECK(ncclTopoGetStrFromSys(path, "numa_node", numaIdStr));
+@@ -541,6 +539,13 @@ ncclResult_t ncclTopoGetXmlFromSys(struct ncclXmlNode* pciNode, struct ncclXml*
+     pciNode->parent = parent;
+     parent->subs[parent->nSubs++] = pciNode;
+   }
++  *return_parent = parent;
++  return ncclSuccess;
++}
++
++ncclResult_t ncclTopoGetXmlFromSys(struct ncclXmlNode* pciNode, struct ncclXml* xml) {
++  struct ncclXmlNode* parent;
++  ncclTopoGetXmlNodeFromSys(pciNode, xml, &parent);
+   if (strcmp(parent->name, "pci") == 0) {
+     NCCLCHECK(ncclTopoGetXmlFromSys(parent, xml));
+   } else if (strcmp(parent->name, "cpu") == 0) {

From 604e993cd0a7d542be7e22ce24c116e270642fc9 Mon Sep 17 00:00:00 2001
From: Akshay Modi 
Date: Mon, 22 Jun 2020 09:46:27 -0700
Subject: [PATCH 428/867] Special case tfnp ndarrays in some places, and minor
 improvements to tfnp array functions

PiperOrigin-RevId: 317672896
Change-Id: I3495884c4367775b387ad34cf35e937ad3b195c8
---
 tensorflow/python/eager/BUILD                  |  1 +
 tensorflow/python/eager/function.py            | 18 ++++++++++++++++--
 tensorflow/python/ops/numpy_ops/np_arrays.py   | 17 +++++++++++++++--
 .../python/ops/numpy_ops/np_interop_test.py    |  3 ++-
 4 files changed, 34 insertions(+), 5 deletions(-)

diff --git a/tensorflow/python/eager/BUILD b/tensorflow/python/eager/BUILD
index af5f3d16408..f51bd97e488 100644
--- a/tensorflow/python/eager/BUILD
+++ b/tensorflow/python/eager/BUILD
@@ -560,6 +560,7 @@ py_library(
         "//tensorflow/python:graph_to_function_def",
         "//tensorflow/python:pywrap_tf_session",
         "//tensorflow/python:util",
+        "//tensorflow/python/ops/numpy_ops:numpy",
         "//third_party/py/numpy",
         "@six_archive//:six",
     ],
diff --git a/tensorflow/python/eager/function.py b/tensorflow/python/eager/function.py
index ca1e60c1b7b..845abe43134 100644
--- a/tensorflow/python/eager/function.py
+++ b/tensorflow/python/eager/function.py
@@ -81,6 +81,9 @@ from tensorflow.python.util import tf_inspect
 ag_ctx = lazy_loader.LazyLoader(
     "ag_ctx", globals(),
     "tensorflow.python.autograph.core.ag_ctx")
+np_arrays = lazy_loader.LazyLoader(
+    "np_arrays", globals(),
+    "tensorflow.python.ops.numpy_ops.np_arrays")
 
 
 FORWARD_FUNCTION_ATTRIBUTE_NAME = "forward_function_name"
@@ -1487,6 +1490,11 @@ class ConcreteFunction(object):
     self._func_graph = func_graph
     self._captured_inputs = self._func_graph.external_captures
     self._captured_closures = self._func_graph.deferred_external_captures
+    structured_outputs = self._func_graph.structured_outputs
+    self._ndarrays_list = (
+        isinstance(structured_outputs, (list, tuple)) and
+        all([isinstance(o, np_arrays.ndarray) for o in structured_outputs]))
+    self._ndarray_singleton = isinstance(structured_outputs, np_arrays.ndarray)
 
     # function_spec defines the structured signature.
     self._set_function_spec(function_spec)
@@ -2153,9 +2161,15 @@ class ConcreteFunction(object):
     if self._func_graph.structured_outputs is None:
       return result
 
+    if result:
+      if self._ndarrays_list:
+        return [np_arrays.tensor_to_ndarray(o) for o in result]
+      elif self._ndarray_singleton:
+        return np_arrays.tensor_to_ndarray(result[0])
+
     # Replace outputs with results, skipping over any 'None' values.
-    outputs_list = nest.flatten(self._func_graph.structured_outputs,
-                                expand_composites=True)
+    outputs_list = nest.flatten(
+        self._func_graph.structured_outputs, expand_composites=True)
     j = 0
     for i, o in enumerate(outputs_list):
       if o is not None:
diff --git a/tensorflow/python/ops/numpy_ops/np_arrays.py b/tensorflow/python/ops/numpy_ops/np_arrays.py
index 65e8273375f..d5ad87a887e 100644
--- a/tensorflow/python/ops/numpy_ops/np_arrays.py
+++ b/tensorflow/python/ops/numpy_ops/np_arrays.py
@@ -149,10 +149,23 @@ class ndarray(composite_tensor.CompositeTensor):  # pylint: disable=invalid-name
     if dtype and dtype != buffer.dtype:
       buffer = array_ops.bitcast(buffer, dtype)
     self._data = buffer
+    self._type_spec_internal = None
+
+  @classmethod
+  def from_tensor(cls, tensor):
+    o = cls.__new__(cls, None)
+    # pylint: disable=protected-access
+    o._data = tensor
+    o._type_spec_internal = None
+    # pylint: enable=protected-access
+    return o
 
   @property
   def _type_spec(self):
-    return NdarraySpec(type_spec.type_spec_from_value(self._data))
+    if self._type_spec_internal is None:
+      self._type_spec_internal = NdarraySpec(
+          type_spec.type_spec_from_value(self._data))
+    return self._type_spec_internal
 
   @property
   def data(self):
@@ -295,7 +308,7 @@ class ndarray(composite_tensor.CompositeTensor):  # pylint: disable=invalid-name
 
 
 def tensor_to_ndarray(tensor):
-  return ndarray(tensor._shape_tuple(), dtype=tensor.dtype, buffer=tensor)  # pylint: disable=protected-access
+  return ndarray.from_tensor(tensor)
 
 
 def ndarray_to_tensor(arr, dtype=None, name=None, as_ref=False):
diff --git a/tensorflow/python/ops/numpy_ops/np_interop_test.py b/tensorflow/python/ops/numpy_ops/np_interop_test.py
index 9580b787202..f6882e968b5 100644
--- a/tensorflow/python/ops/numpy_ops/np_interop_test.py
+++ b/tensorflow/python/ops/numpy_ops/np_interop_test.py
@@ -64,7 +64,8 @@ class InteropTest(test.TestCase):
 
     dx, dy = t.gradient([xx, yy], [x, y])
 
-    # TODO(nareshmodi): Gradient tape returns tensors. Is it possible to rewrap?
+    self.assertIsInstance(dx, np_arrays.ndarray)
+    self.assertIsInstance(dy, np_arrays.ndarray)
     self.assertAllClose(dx, 2.0)
     self.assertAllClose(dy, 3.0)
 

From 9afaf559d9e9553bc651faa35fc0e6d6f9a04a84 Mon Sep 17 00:00:00 2001
From: Nathan Luehr 
Date: Mon, 22 Jun 2020 12:28:02 -0500
Subject: [PATCH 429/867] Locate CUDNN_FMA_MATH inside CUDNN 8 macros

---
 tensorflow/stream_executor/cuda/cuda_dnn.cc | 10 +++++-----
 1 file changed, 5 insertions(+), 5 deletions(-)

diff --git a/tensorflow/stream_executor/cuda/cuda_dnn.cc b/tensorflow/stream_executor/cuda/cuda_dnn.cc
index 166fa0e32d0..6f0f94c7da2 100644
--- a/tensorflow/stream_executor/cuda/cuda_dnn.cc
+++ b/tensorflow/stream_executor/cuda/cuda_dnn.cc
@@ -1176,11 +1176,11 @@ class CudnnRnnDescriptor : public dnn::RnnDescriptor {
     }
 
     cudnnMathType_t math_type;
-    if (use_tensor_ops) {
-      math_type = CUDNN_TENSOR_OP_MATH;
-    } else {
-      math_type = CUDNN_VERSION >= 8000 ? CUDNN_FMA_MATH : CUDNN_DEFAULT_MATH;
-    }
+#if CUDNN_VERSION >= 8000
+    math_type = use_tensor_ops ? CUDNN_TENSOR_OP_MATH : CUDNN_FMA_MATH;
+#else
+    math_type = use_tensor_ops ? CUDNN_TENSOR_OP_MATH : CUDNN_DEFAULT_MATH;
+#endif
     CHECK_CUDNN_OK(cudnnSetRNNMatrixMathType(rnn_desc.get(), math_type));
 #endif  // CUDNN_VERSION >= 7000
 

From 2afae99d7fd51dc9c3143718cbd19051a0b2140b Mon Sep 17 00:00:00 2001
From: "A. Unique TensorFlower" 
Date: Mon, 22 Jun 2020 09:51:54 -0700
Subject: [PATCH 430/867] Migrating deprecated aliases for absl::StrSplit, and
 the corresponding predicates

PiperOrigin-RevId: 317673991
Change-Id: I77569b9f0c18095c9ade9bcf3c306d2db0fb2048
---
 tensorflow/lite/toco/args.cc | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/tensorflow/lite/toco/args.cc b/tensorflow/lite/toco/args.cc
index c30b98ce516..94fd850fdc5 100644
--- a/tensorflow/lite/toco/args.cc
+++ b/tensorflow/lite/toco/args.cc
@@ -114,7 +114,7 @@ inline bool TryStripSuffixString(absl::string_view str,
 bool Arg::Parse(std::string text) {
   parsed_value_.elements.clear();
   specified_ = true;
-  // strings::Split("") produces {""}, but we need {} on empty input.
+  // absl::StrSplit("") produces {""}, but we need {} on empty input.
   // TODO(aselle): Moved this from elsewhere, but ahentz recommends we could
   // use absl::SplitLeadingDec32Values(text.c_str(), &parsed_values_.elements)
   if (!text.empty()) {

From 251923169ddde0b69c58fb41b23443423ababd2d Mon Sep 17 00:00:00 2001
From: Rahul Joshi 
Date: Mon, 22 Jun 2020 09:58:59 -0700
Subject: [PATCH 431/867] [NFC] Eliminate use of .getBlocks() when not needed  
     Also use llvm::hasSingleElement() instead of .size() == 1

PiperOrigin-RevId: 317675565
Change-Id: I4f0e8892957c2b20e115584fe7424da68d53b67a
---
 tensorflow/compiler/mlir/lite/flatbuffer_export.cc    |  6 +++---
 .../compiler/mlir/lite/transforms/post_quantize.cc    |  2 +-
 tensorflow/compiler/mlir/tensorflow/ir/tf_device.cc   |  6 +++---
 tensorflow/compiler/mlir/tensorflow/ir/tf_executor.cc |  6 +++---
 .../mlir/tensorflow/transforms/fold_switch.cc         |  2 +-
 .../transforms/materialize_mlir_passthrough_op.cc     |  2 +-
 .../transforms/promote_resources_to_args.cc           |  4 ++--
 .../mlir/tensorflow/transforms/resource_op_lifting.cc |  2 +-
 .../transforms/tpu_sharding_identification_pass.cc    | 11 ++++-------
 .../mlir/tensorflow/translate/export_graphdef.cc      |  2 +-
 .../tensorflow/translate/tf_functional_to_executor.cc |  4 ++--
 .../tensorflow/translate/translate_tf_dialect_op.cc   |  5 +++--
 tensorflow/compiler/mlir/xla/mlir_hlo_to_hlo.cc       |  4 ++--
 .../mlir/xla/transforms/hlo_legalize_to_lhlo.cc       |  8 ++++----
 .../mlir/xla/transforms/legalize_tf_with_tf2xla.cc    |  8 +++++---
 .../compiler/mlir/xla/transforms/lhlo_fuse_linalg.cc  |  5 +++--
 .../compiler/mlir/xla/transforms/xla_hlo_fusion.cc    |  2 +-
 .../compiler/xla/service/mlir_gpu/kernel_lowering.cc  |  2 +-
 18 files changed, 41 insertions(+), 40 deletions(-)

diff --git a/tensorflow/compiler/mlir/lite/flatbuffer_export.cc b/tensorflow/compiler/mlir/lite/flatbuffer_export.cc
index a260670015a..e34e7ae7ca6 100644
--- a/tensorflow/compiler/mlir/lite/flatbuffer_export.cc
+++ b/tensorflow/compiler/mlir/lite/flatbuffer_export.cc
@@ -240,10 +240,10 @@ static bool IsValidTFLiteMlirModule(ModuleOp module) {
   }
 
   for (auto fn : module.getOps()) {
-    if (fn.getBlocks().size() != 1) {
+    if (!llvm::hasSingleElement(fn)) {
       return fn.emitError("should have exactly one basic block"), false;
     }
-    auto& bb = fn.getBlocks().front();
+    auto& bb = fn.front();
 
     for (auto arg : bb.getArguments()) {
       if (!HasValidTFLiteType(arg, fn))
@@ -1089,7 +1089,7 @@ void Translator::InitializeNamesFromAttribute(FuncOp fn, bool* has_input_attr) {
           dict_attr.get("outputs").dyn_cast_or_null()) {
     str.getValue().split(output_names, ',', /*MaxSplit=*/-1,
                          /*KeepEmpty=*/false);
-    auto term = fn.getBlocks().back().getTerminator();
+    auto term = fn.back().getTerminator();
     if (output_names.size() != term->getNumOperands()) {
       fn.emitWarning() << "output names (" << output_names.size()
                        << ") != terminator operands (" << term->getNumOperands()
diff --git a/tensorflow/compiler/mlir/lite/transforms/post_quantize.cc b/tensorflow/compiler/mlir/lite/transforms/post_quantize.cc
index 9a1da0ad03d..33380e00543 100644
--- a/tensorflow/compiler/mlir/lite/transforms/post_quantize.cc
+++ b/tensorflow/compiler/mlir/lite/transforms/post_quantize.cc
@@ -52,7 +52,7 @@ class PostQuantizePass : public PassWrapper {
 
 void RemoveQuantizationAdaptorOps(FuncOp func) {
   mlir::OpBuilder builder(func.getBody());
-  auto& bb = func.getBlocks().front();
+  auto& bb = func.front();
   auto* terminator = bb.getTerminator();
 
   int num_args = bb.getNumArguments();
diff --git a/tensorflow/compiler/mlir/tensorflow/ir/tf_device.cc b/tensorflow/compiler/mlir/tensorflow/ir/tf_device.cc
index b8f0585040c..7dd74282487 100644
--- a/tensorflow/compiler/mlir/tensorflow/ir/tf_device.cc
+++ b/tensorflow/compiler/mlir/tensorflow/ir/tf_device.cc
@@ -299,13 +299,13 @@ ParseResult ParseReplicateOp(OpAsmParser* parser, OperationState* state) {
       parser->parseRegion(body, region_args, region_arg_types))
     return failure();
 
-  if (body.getBlocks().size() > 1)
-    return parser->emitError(loc) << "expects a single block region";
-
   // Ensure that the region is well formed: it contains at least a block with
   // a ReturnOp terminator.
   ReplicateOp::ensureTerminator(body, parser->getBuilder(), state->location);
 
+  if (!llvm::hasSingleElement(body))
+    return parser->emitError(loc) << "expects a single block region";
+
   Operation& terminator = body.front().back();
   if (!isa(terminator))
     return parser->emitError(loc) << "expects a tf_device.return terminator";
diff --git a/tensorflow/compiler/mlir/tensorflow/ir/tf_executor.cc b/tensorflow/compiler/mlir/tensorflow/ir/tf_executor.cc
index 3403651eef8..1e66eee06bb 100644
--- a/tensorflow/compiler/mlir/tensorflow/ir/tf_executor.cc
+++ b/tensorflow/compiler/mlir/tensorflow/ir/tf_executor.cc
@@ -220,13 +220,13 @@ ParseResult ParseGraphOp(OpAsmParser &parser, OperationState &result) {
   Region &body = *result.addRegion();
   if (parser.parseRegion(body, llvm::None, llvm::None)) return failure();
 
-  if (body.getBlocks().size() > 1)
-    return parser.emitError(loc) << "expects a single block region";
-
   // Ensure that the region is well formed: it contains at least a block with
   // a FetchOp terminator.
   GraphOp::ensureTerminator(body, parser.getBuilder(), result.location);
 
+  if (!llvm::hasSingleElement(body))
+    return parser.emitError(loc) << "expects a single block region";
+
   // Get the results type from the terminator type inside the graph.
   Operation &fetch = body.back().back();
   if (!isa(fetch))
diff --git a/tensorflow/compiler/mlir/tensorflow/transforms/fold_switch.cc b/tensorflow/compiler/mlir/tensorflow/transforms/fold_switch.cc
index 4d26747ebdc..b47378762a9 100644
--- a/tensorflow/compiler/mlir/tensorflow/transforms/fold_switch.cc
+++ b/tensorflow/compiler/mlir/tensorflow/transforms/fold_switch.cc
@@ -199,7 +199,7 @@ static void MatchSwitchFoldOps(tf_executor::SwitchOp switch_op,
 // Folds merge nodes with only a single non-dead input.
 static LogicalResult FoldMergeNodes(FuncOp function, const DeadQueue& queue) {
   // Create builder for val_index of MergeOp.
-  auto* block = &function.getBlocks().front();
+  auto* block = &function.front();
   OpBuilder builder = OpBuilder::atBlockEnd(block);
   auto type = builder.getIntegerType(32);
   auto build_index = [&](Location loc, int value) {
diff --git a/tensorflow/compiler/mlir/tensorflow/transforms/materialize_mlir_passthrough_op.cc b/tensorflow/compiler/mlir/tensorflow/transforms/materialize_mlir_passthrough_op.cc
index 94fdfb310ac..3ed27d7ce30 100644
--- a/tensorflow/compiler/mlir/tensorflow/transforms/materialize_mlir_passthrough_op.cc
+++ b/tensorflow/compiler/mlir/tensorflow/transforms/materialize_mlir_passthrough_op.cc
@@ -71,7 +71,7 @@ void MaterializePassthroughOpPass::runOnFunction() {
       return;
     }
     Region &body = main.getBody();
-    if (body.getBlocks().size() != 1) {
+    if (!llvm::hasSingleElement(body)) {
       op->emitError() << "MLIR Opaque Op expects a main() entry point with a "
                          "single block\n";
       return;
diff --git a/tensorflow/compiler/mlir/tensorflow/transforms/promote_resources_to_args.cc b/tensorflow/compiler/mlir/tensorflow/transforms/promote_resources_to_args.cc
index cece23b4750..af36770f496 100644
--- a/tensorflow/compiler/mlir/tensorflow/transforms/promote_resources_to_args.cc
+++ b/tensorflow/compiler/mlir/tensorflow/transforms/promote_resources_to_args.cc
@@ -80,11 +80,11 @@ constexpr char kResourceNameArgAttr[] = "tf.resource_name";
 
 // Checks if a function has only one block.
 mlir::LogicalResult CheckSingleBlockFunction(FuncOp function) {
-  if (!hasSingleElement(function.getBlocks()))
+  if (!llvm::hasSingleElement(function)) {
     return function.emitError()
            << "expects function '" << function.getName()
            << "' to have 1 block, got " << function.getBlocks().size();
-
+  }
   return success();
 }
 
diff --git a/tensorflow/compiler/mlir/tensorflow/transforms/resource_op_lifting.cc b/tensorflow/compiler/mlir/tensorflow/transforms/resource_op_lifting.cc
index ed7ebc25c9f..799ab3a0f0d 100644
--- a/tensorflow/compiler/mlir/tensorflow/transforms/resource_op_lifting.cc
+++ b/tensorflow/compiler/mlir/tensorflow/transforms/resource_op_lifting.cc
@@ -1113,7 +1113,7 @@ LogicalResult ResourceLiftingForFunctionalControlFlow(FuncOp function) {
   // This routine should only be called when control flow operations are still
   // represented with TF IfOp and WhileOp operations. In this case, there should
   // be only one basic blocks in the MLIR representation.
-  if (!hasSingleElement(function.getBlocks())) {
+  if (!llvm::hasSingleElement(function)) {
     return function.emitError()
            << "expect the function to have 1 block while it has "
            << function.getBlocks().size();
diff --git a/tensorflow/compiler/mlir/tensorflow/transforms/tpu_sharding_identification_pass.cc b/tensorflow/compiler/mlir/tensorflow/transforms/tpu_sharding_identification_pass.cc
index f8b6e364f55..b05e87c6485 100644
--- a/tensorflow/compiler/mlir/tensorflow/transforms/tpu_sharding_identification_pass.cc
+++ b/tensorflow/compiler/mlir/tensorflow/transforms/tpu_sharding_identification_pass.cc
@@ -159,8 +159,7 @@ llvm::SmallVector ExtractFunctionsConnectedToArg(
   while (!functions_to_parse.empty()) {
     llvm::SmallVector newly_discovered_functions;
     for (auto function_info : functions_to_parse) {
-      Block& func_entry_block =
-          function_info.func.getBody().getBlocks().front();
+      Block& func_entry_block = function_info.func.front();
       auto argument =
           func_entry_block.getArgument(function_info.argument_index);
 
@@ -186,8 +185,7 @@ void IdentifyXlaShardingForComputationInputs(
     StringRef logical_core_0_sharding, tf_device::ClusterFuncOp cluster_func_op,
     FuncOp cluster_function, Builder* builder) {
   // Look up function definition from module.
-  Block& cluster_function_block =
-      cluster_function.getBody().getBlocks().front();
+  Block& cluster_function_block = cluster_function.front();
   ModuleOp module = cluster_func_op.getParentOfType();
 
   llvm::SmallVector sharding_for_args(
@@ -215,8 +213,7 @@ void IdentifyXlaShardingForComputationInputs(
 
         const int function_argument_index = function_arg_info.argument_index;
         auto& parsed_function = function_arg_info.func;
-        Block& parsed_function_block =
-            parsed_function.getBody().getBlocks().front();
+        Block& parsed_function_block = parsed_function.front();
         arg_sharding = ParseInputSharding(
             parsed_function_block.getArgument(function_argument_index));
       }
@@ -245,7 +242,7 @@ void IdentifyXlaShardingForComputationOutputs(
     tf_device::ClusterFuncOp cluster_func, Builder* builder) {
   // By default return values from logical core 0 is used if no sharding
   // configuration is defined.
-  Block& function_block = func.getBody().getBlocks().front();
+  Block& function_block = func.front();
   Operation* terminator = function_block.getTerminator();
   llvm::SmallVector sharding_for_rets(
       terminator->getNumOperands(), logical_core_0_sharding);
diff --git a/tensorflow/compiler/mlir/tensorflow/translate/export_graphdef.cc b/tensorflow/compiler/mlir/tensorflow/translate/export_graphdef.cc
index 262f6f4e50c..8cd14894f8f 100644
--- a/tensorflow/compiler/mlir/tensorflow/translate/export_graphdef.cc
+++ b/tensorflow/compiler/mlir/tensorflow/translate/export_graphdef.cc
@@ -128,7 +128,7 @@ class LegalizedOpOrValLocNameMapper : public OpOrArgLocNameMapper {
 Status HasSingleGraphSingleOpIslandsFunctions(mlir::ModuleOp module) {
   Status status = Status::OK();
   module.walk([&](mlir::FuncOp function) {
-    if (function.getBlocks().size() != 1) {
+    if (!llvm::hasSingleElement(function)) {
       status = errors::FailedPrecondition(
           kInvalidExecutorGraphMsg,
           "only single block functions are supported.");
diff --git a/tensorflow/compiler/mlir/tensorflow/translate/tf_functional_to_executor.cc b/tensorflow/compiler/mlir/tensorflow/translate/tf_functional_to_executor.cc
index 29f98de6448..78019119d9d 100644
--- a/tensorflow/compiler/mlir/tensorflow/translate/tf_functional_to_executor.cc
+++ b/tensorflow/compiler/mlir/tensorflow/translate/tf_functional_to_executor.cc
@@ -46,13 +46,13 @@ struct FunctionalToExecutorDialectConversion
 }  // end anonymous namespace
 
 void FunctionalToExecutorDialectConversion::runOnFunction() {
-  if (getFunction().getBlocks().size() != 1) {
+  if (!llvm::hasSingleElement(getFunction())) {
     LLVM_DEBUG(llvm::dbgs() << "Expect single block function, skip conversion "
                                "to tf_executor dialect\n");
     return;
   }
   auto loc = getFunction().getLoc();
-  mlir::Block& body = getFunction().getBody().front();
+  mlir::Block& body = getFunction().front();
   // Find region of interest and ReturnOp.
   auto copy_range = body.without_terminator();
   if (copy_range.begin() != copy_range.end() &&
diff --git a/tensorflow/compiler/mlir/tensorflow/translate/translate_tf_dialect_op.cc b/tensorflow/compiler/mlir/tensorflow/translate/translate_tf_dialect_op.cc
index bd3fe9876ff..5236bdeffbf 100644
--- a/tensorflow/compiler/mlir/tensorflow/translate/translate_tf_dialect_op.cc
+++ b/tensorflow/compiler/mlir/tensorflow/translate/translate_tf_dialect_op.cc
@@ -13,6 +13,7 @@ See the License for the specific language governing permissions and
 limitations under the License.
 ==============================================================================*/
 
+#include "llvm/ADT/STLExtras.h"
 #include "llvm/Support/ToolOutputFile.h"
 #include "mlir/IR/Function.h"  // from @llvm-project
 #include "mlir/IR/Location.h"  // from @llvm-project
@@ -26,12 +27,12 @@ static mlir::Operation* ExtractOnlyOp(mlir::ModuleOp module) {
   mlir::FuncOp fn = module.lookupSymbol("main");
   if (!fn) return nullptr;
 
-  if (fn.getBlocks().size() != 1) return nullptr;
+  if (!llvm::hasSingleElement(fn)) return nullptr;
 
   // Here, modules with exactly two operations in the only basic block are
   // supported. The last operation should be a terminator operation and the
   // other operation is the operation of interest.
-  auto& block = fn.getBlocks().front();
+  auto& block = fn.front();
   if (block.getOperations().size() != 2) return nullptr;
   if (!block.back().isKnownTerminator()) return nullptr;
 
diff --git a/tensorflow/compiler/mlir/xla/mlir_hlo_to_hlo.cc b/tensorflow/compiler/mlir/xla/mlir_hlo_to_hlo.cc
index 60d9a698731..7a576780c61 100644
--- a/tensorflow/compiler/mlir/xla/mlir_hlo_to_hlo.cc
+++ b/tensorflow/compiler/mlir/xla/mlir_hlo_to_hlo.cc
@@ -1148,13 +1148,13 @@ LogicalResult ConvertToHloModule::LowerFunctionCall(
 
 LogicalResult ConvertToHloModule::RunOnFunction(mlir::FuncOp f) {
   if (lowered_computation_.count(f)) return success();
-  if (f.getBlocks().size() != 1) {
+  if (!llvm::hasSingleElement(f)) {
     return f.emitError("only single block Function supported");
   }
 
   // Create a sub-builder if this is not the main function.
   std::unique_ptr builder_up;
-  bool entry_function = f.getName().str() == "main";
+  bool entry_function = f.getName() == "main";
   if (!entry_function)
     builder_up = module_builder_.CreateSubBuilder(f.getName().str());
   auto& builder = entry_function ? module_builder_ : *builder_up;
diff --git a/tensorflow/compiler/mlir/xla/transforms/hlo_legalize_to_lhlo.cc b/tensorflow/compiler/mlir/xla/transforms/hlo_legalize_to_lhlo.cc
index 446f2aae833..7cdc0d92207 100644
--- a/tensorflow/compiler/mlir/xla/transforms/hlo_legalize_to_lhlo.cc
+++ b/tensorflow/compiler/mlir/xla/transforms/hlo_legalize_to_lhlo.cc
@@ -230,10 +230,10 @@ struct HloToLhloReduceOpConverter : public BaseOpConversion {
     auto loc = op.getLoc();
     // TODO(b/137624192) Implement variadic reduce.
     if (op.getNumResults() != 1) return failure();
-    if (op.getParentRegion()->getBlocks().size() != 1) {
-      op.emitOpError() << "tensor to buffer conversion expects a single block "
-                          "in the region containing the operation";
-      return failure();
+    if (!llvm::hasSingleElement(op.body())) {
+      return op.emitOpError()
+             << "tensor to buffer conversion expects a single block "
+                "in the region containing the operation";
     }
     const auto& original_results = op.getResults();
     SmallVector buffer_args(operands.begin(), operands.end());
diff --git a/tensorflow/compiler/mlir/xla/transforms/legalize_tf_with_tf2xla.cc b/tensorflow/compiler/mlir/xla/transforms/legalize_tf_with_tf2xla.cc
index 8f96f4d1305..54453406ef7 100644
--- a/tensorflow/compiler/mlir/xla/transforms/legalize_tf_with_tf2xla.cc
+++ b/tensorflow/compiler/mlir/xla/transforms/legalize_tf_with_tf2xla.cc
@@ -23,6 +23,7 @@ limitations under the License.
 #include "absl/strings/string_view.h"
 #include "llvm/ADT/DenseSet.h"
 #include "llvm/ADT/Optional.h"
+#include "llvm/ADT/STLExtras.h"
 #include "mlir/Dialect/StandardOps/IR/Ops.h"  // from @llvm-project
 #include "mlir/IR/Diagnostics.h"  // from @llvm-project
 #include "mlir/IR/Function.h"  // from @llvm-project
@@ -320,13 +321,14 @@ LogicalResult FuncLegalizer::PrepareParams() {
 }
 
 LogicalResult FuncLegalizer::Legalize() {
+  if (func_.empty()) return success();
+
   // TensorFlow functions don't use CFGs.
-  if (func_.getBlocks().size() > 1) {
+  if (!llvm::hasSingleElement(func_)) {
     emitError(func_.getLoc()) << "requires at most one block in a TF function";
     return failure();
   }
-  if (func_.getBlocks().empty()) return success();
-  Block& block = func_.getBlocks().front();
+  Block& block = func_.front();
 
   std::vector ops;
   ops.reserve(block.getOperations().size());
diff --git a/tensorflow/compiler/mlir/xla/transforms/lhlo_fuse_linalg.cc b/tensorflow/compiler/mlir/xla/transforms/lhlo_fuse_linalg.cc
index e16ab571b4d..f0971fdf76e 100644
--- a/tensorflow/compiler/mlir/xla/transforms/lhlo_fuse_linalg.cc
+++ b/tensorflow/compiler/mlir/xla/transforms/lhlo_fuse_linalg.cc
@@ -19,6 +19,7 @@ limitations under the License.
 #include "mlir/Dialect/Linalg/Analysis/DependenceAnalysis.h"
 #include "absl/memory/memory.h"
 #include "llvm/ADT/ArrayRef.h"
+#include "llvm/ADT/STLExtras.h"
 #include "mlir/Dialect/Linalg/Transforms/Transforms.h"  // from @llvm-project
 #include "mlir/Dialect/StandardOps/IR/Ops.h"  // from @llvm-project
 #include "mlir/Pass/Pass.h"  // from @llvm-project
@@ -44,7 +45,7 @@ class LhloFuseLinalg : public PassWrapper {
     auto func = getFunction();
 
     // TODO(pifon): Remove assumption that the function has a single block.
-    if (func.getBlocks().size() != 1) {
+    if (!llvm::hasSingleElement(func)) {
       emitError(func.getLoc(), "The function needs to have a single block.");
       signalPassFailure();
       return;
@@ -58,7 +59,7 @@ class LhloFuseLinalg : public PassWrapper {
     for (auto func_arg : func.getArguments()) {
       result_buffers.insert(func_arg);
     }
-    for (auto& block : func.getBlocks()) {
+    for (auto& block : func) {
       auto returnOp = mlir::dyn_cast(block.getTerminator());
       if (!returnOp) continue;
       for (auto operand : returnOp.getOperands()) {
diff --git a/tensorflow/compiler/mlir/xla/transforms/xla_hlo_fusion.cc b/tensorflow/compiler/mlir/xla/transforms/xla_hlo_fusion.cc
index c4eb0e143d2..5d3eda0bea5 100644
--- a/tensorflow/compiler/mlir/xla/transforms/xla_hlo_fusion.cc
+++ b/tensorflow/compiler/mlir/xla/transforms/xla_hlo_fusion.cc
@@ -487,7 +487,7 @@ struct XlaHloFusion : public mlir::PassWrapper {
     }
 
     // process each block and do fusion within a block.
-    for (Block& block : func.getBlocks()) {
+    for (Block& block : func) {
       SmallVector op_list;
       for (Operation& op : block) {
         op_list.push_back(&op);
diff --git a/tensorflow/compiler/xla/service/mlir_gpu/kernel_lowering.cc b/tensorflow/compiler/xla/service/mlir_gpu/kernel_lowering.cc
index ecd1308be4b..3f99d40c717 100644
--- a/tensorflow/compiler/xla/service/mlir_gpu/kernel_lowering.cc
+++ b/tensorflow/compiler/xla/service/mlir_gpu/kernel_lowering.cc
@@ -301,7 +301,7 @@ struct RewriteKernelSignature
         signalPassFailure();
         return;
       }
-      if (func.getBlocks().size() != 1) {
+      if (!llvm::hasSingleElement(func)) {
         func.emitError() << "surrounding function has more than one block";
         signalPassFailure();
         return;

From 834f2bd72627ee8499637ec3c1a526b1dcd19a79 Mon Sep 17 00:00:00 2001
From: "A. Unique TensorFlower" 
Date: Mon, 22 Jun 2020 10:01:26 -0700
Subject: [PATCH 432/867] Enable tf.image.ssim to run on tensor-like inputs and
 add SSIM name space.

PiperOrigin-RevId: 317676190
Change-Id: I08408eaf397ace235b5f50513096cbd9ba46d5a8
---
 tensorflow/python/ops/image_ops_impl.py | 31 ++++++++++++++-----------
 tensorflow/python/ops/image_ops_test.py | 23 ++++++++++++++++++
 2 files changed, 41 insertions(+), 13 deletions(-)

diff --git a/tensorflow/python/ops/image_ops_impl.py b/tensorflow/python/ops/image_ops_impl.py
index a934639d524..683681b5c98 100644
--- a/tensorflow/python/ops/image_ops_impl.py
+++ b/tensorflow/python/ops/image_ops_impl.py
@@ -3642,20 +3642,25 @@ def ssim(img1,
     values are in range (-1, 1], when pixel values are non-negative. Returns
     a tensor with shape: broadcast(img1.shape[:-3], img2.shape[:-3]).
   """
-  _, _, checks = _verify_compatible_image_shapes(img1, img2)
-  with ops.control_dependencies(checks):
-    img1 = array_ops.identity(img1)
+  with ops.name_scope(None, 'SSIM', [img1, img2]):
+    # Convert to tensor if needed.
+    img1 = ops.convert_to_tensor(img1, name='img1')
+    img2 = ops.convert_to_tensor(img2, name='img2')
+    # Shape checking.
+    _, _, checks = _verify_compatible_image_shapes(img1, img2)
+    with ops.control_dependencies(checks):
+      img1 = array_ops.identity(img1)
 
-  # Need to convert the images to float32.  Scale max_val accordingly so that
-  # SSIM is computed correctly.
-  max_val = math_ops.cast(max_val, img1.dtype)
-  max_val = convert_image_dtype(max_val, dtypes.float32)
-  img1 = convert_image_dtype(img1, dtypes.float32)
-  img2 = convert_image_dtype(img2, dtypes.float32)
-  ssim_per_channel, _ = _ssim_per_channel(img1, img2, max_val, filter_size,
-                                          filter_sigma, k1, k2)
-  # Compute average over color channels.
-  return math_ops.reduce_mean(ssim_per_channel, [-1])
+    # Need to convert the images to float32.  Scale max_val accordingly so that
+    # SSIM is computed correctly.
+    max_val = math_ops.cast(max_val, img1.dtype)
+    max_val = convert_image_dtype(max_val, dtypes.float32)
+    img1 = convert_image_dtype(img1, dtypes.float32)
+    img2 = convert_image_dtype(img2, dtypes.float32)
+    ssim_per_channel, _ = _ssim_per_channel(img1, img2, max_val, filter_size,
+                                            filter_sigma, k1, k2)
+    # Compute average over color channels.
+    return math_ops.reduce_mean(ssim_per_channel, [-1])
 
 
 # Default values obtained by Wang et al.
diff --git a/tensorflow/python/ops/image_ops_test.py b/tensorflow/python/ops/image_ops_test.py
index 3530885fe07..0206ccf9b33 100644
--- a/tensorflow/python/ops/image_ops_test.py
+++ b/tensorflow/python/ops/image_ops_test.py
@@ -4865,6 +4865,29 @@ class SSIMTest(test_util.TensorFlowTestCase):
     with self.cached_session(use_gpu=True):
       self.assertAllClose(expected, self.evaluate(ssim), atol=1e-4)
 
+  def testBatchNumpyInputs(self):
+    img = self._LoadTestImages()
+    expected = self._ssim[np.triu_indices(3, k=1)]
+
+    img1, img2 = zip(*itertools.combinations(img, 2))
+    img1 = np.concatenate(img1)
+    img2 = np.concatenate(img2)
+
+    with self.cached_session(use_gpu=True):
+      img1 = self.evaluate(constant_op.constant(img1))
+      img2 = self.evaluate(constant_op.constant(img2))
+
+    ssim = image_ops.ssim(
+        img1,
+        img2,
+        1.0,
+        filter_size=11,
+        filter_sigma=1.5,
+        k1=0.01,
+        k2=0.03)
+    with self.cached_session(use_gpu=True):
+      self.assertAllClose(expected, self.evaluate(ssim), atol=1e-4)
+
   def testBroadcast(self):
     img = self._LoadTestImages()[:2]
     expected = self._ssim[:2, :2]

From 3f973428767cfbbbeeb412b3d7e0be8e566591fb Mon Sep 17 00:00:00 2001
From: "A. Unique TensorFlower" 
Date: Mon, 22 Jun 2020 10:01:29 -0700
Subject: [PATCH 433/867] Fix the api docstrings for on_*_batch_begin methods.

PiperOrigin-RevId: 317676199
Change-Id: I15f8eb706c5841f058d5ac237c2939a7a4d31809
---
 tensorflow/python/keras/callbacks.py | 15 +++++++++------
 1 file changed, 9 insertions(+), 6 deletions(-)

diff --git a/tensorflow/python/keras/callbacks.py b/tensorflow/python/keras/callbacks.py
index 1fae5abd84b..138a682c739 100644
--- a/tensorflow/python/keras/callbacks.py
+++ b/tensorflow/python/keras/callbacks.py
@@ -420,8 +420,9 @@ class CallbackList(object):
 
     Arguments:
         batch: Integer, index of batch within the current epoch.
-        logs: Dict. Has keys `batch` and `size` representing the current batch
-          number and the size of the batch.
+        logs: Dict, contains the return value of `model.train_step`. Typically,
+          the values of the `Model`'s metrics are returned.  Example:
+          `{'loss': 0.2, 'accuracy': 0.7}`.
     """
     # TODO(b/150629188): Make ProgBarLogger callback not use batch hooks
     # when verbose != 1
@@ -443,8 +444,9 @@ class CallbackList(object):
 
     Arguments:
         batch: Integer, index of batch within the current epoch.
-        logs: Dict. Has keys `batch` and `size` representing the current batch
-          number and the size of the batch.
+        logs: Dict, contains the return value of `model.test_step`. Typically,
+          the values of the `Model`'s metrics are returned.  Example:
+          `{'loss': 0.2, 'accuracy': 0.7}`.
     """
     if self._should_call_test_batch_hooks:
       self._call_batch_hook(ModeKeys.TEST, 'begin', batch, logs=logs)
@@ -464,8 +466,9 @@ class CallbackList(object):
 
     Arguments:
         batch: Integer, index of batch within the current epoch.
-        logs: Dict. Has keys `batch` and `size` representing the current batch
-          number and the size of the batch.
+        logs: Dict, contains the return value of `model.predict_step`,
+          it typically returns a dict with a key 'outputs' containing
+          the model's outputs.
     """
     if self._should_call_predict_batch_hooks:
       self._call_batch_hook(ModeKeys.PREDICT, 'begin', batch, logs=logs)

From 60be0c3c0ea9e8ffd9189f1cc755199384d8faf0 Mon Sep 17 00:00:00 2001
From: "T.J. Alumbaugh" 
Date: Mon, 22 Jun 2020 10:06:04 -0700
Subject: [PATCH 434/867] Restore functionality of
 --define=tflite_with_ruy=true

PiperOrigin-RevId: 317677225
Change-Id: If6533fdfeb21f676dd4b77a536b1aca894a03003
---
 tensorflow/lite/kernels/BUILD                 | 58 +++++++++----------
 tensorflow/lite/kernels/conv.cc               | 14 ++---
 tensorflow/lite/kernels/conv_test.cc          |  2 +-
 tensorflow/lite/kernels/cpu_backend_context.h |  2 +-
 tensorflow/lite/kernels/cpu_backend_gemm.h    |  6 +-
 .../kernels/cpu_backend_gemm_custom_gemv.h    |  6 +-
 .../lite/kernels/cpu_backend_gemm_eigen.cc    |  4 +-
 .../lite/kernels/cpu_backend_gemm_eigen.h     |  4 +-
 .../lite/kernels/cpu_backend_gemm_gemmlowp.h  |  4 +-
 .../lite/kernels/cpu_backend_threadpool.h     |  6 +-
 .../optimized/depthwiseconv_multithread.h     |  2 +-
 tensorflow/lite/tools/make/Makefile           |  2 +-
 12 files changed, 55 insertions(+), 55 deletions(-)

diff --git a/tensorflow/lite/kernels/BUILD b/tensorflow/lite/kernels/BUILD
index b16a85c65d8..9d3e5929d82 100644
--- a/tensorflow/lite/kernels/BUILD
+++ b/tensorflow/lite/kernels/BUILD
@@ -14,8 +14,8 @@ package(
 # This will cause TFLite to build with ruy only, providing a smaller binary.
 # WARNING: This build flag is experimental and subject to change.
 config_setting(
-    name = "tflite_with_ruy_only_explicit_true",
-    define_values = {"TFLITE_WITH_RUY_ONLY": "true"},
+    name = "tflite_with_ruy_explicit_true",
+    define_values = {"tflite_with_ruy": "true"},
 )
 
 # Disables usage of ruy as the exclusive GEMM backend in TFLite kernels.
@@ -23,14 +23,14 @@ config_setting(
 # the default GEMM option at runtime.
 # WARNING: This build flag is experimental and subject to change.
 config_setting(
-    name = "tflite_with_ruy_only_explicit_false",
-    define_values = {"TFLITE_WITH_RUY_ONLY": "false"},
+    name = "tflite_with_ruy_explicit_false",
+    define_values = {"tflite_with_ruy": "false"},
 )
 
 ###### Beginning of config_setting's to match aarch64 ######
 #
 # We need to identify the aarch64 instruction set to decide whether to enable
-# TFLITE_WITH_RUY_ONLY by default. This is surprisingly hard to do because select()
+# TFLITE_WITH_RUY by default. This is surprisingly hard to do because select()
 # can only consume config_setting's, these config_settings are not centralized,
 # and the "cpu" value which they define are free-form strings and there is no
 # standardization of the strings that we need to match for the aarch64 architecture.
@@ -239,45 +239,45 @@ cc_test(
 )
 
 cc_library(
-    name = "tflite_with_ruy_only_enabled",
+    name = "tflite_with_ruy_enabled",
     build_for_embedded = True,
-    defines = ["TFLITE_WITH_RUY_ONLY"],
+    defines = ["TFLITE_WITH_RUY"],
     visibility = ["//visibility:private"],
 )
 
 cc_library(
-    name = "tflite_with_ruy_only_and_caching_enabled",
+    name = "tflite_with_ruy_and_caching_enabled",
     defines = [
-        "TFLITE_WITH_RUY_ONLY",
+        "TFLITE_WITH_RUY",
         "TFLITE_WITH_RUY_GEMV",
     ],
     visibility = ["//visibility:private"],
 )
 
 cc_library(
-    name = "tflite_with_ruy_only_default",
+    name = "tflite_with_ruy_default",
     build_for_embedded = True,
     select_deps = {
-        ":chromiumos_arm64": [":tflite_with_ruy_only_enabled"],
-        ":cpu_aarch64": [":tflite_with_ruy_only_enabled"],
-        ":cpu_arm64": [":tflite_with_ruy_only_enabled"],
-        ":cpu_arm64e": [":tflite_with_ruy_only_enabled"],
-        ":cpu_ios_arm64": [":tflite_with_ruy_only_enabled"],
-        ":cpu_ios_arm64e": [":tflite_with_ruy_only_enabled"],
-        ":cpu_arm64_v8a": [":tflite_with_ruy_only_enabled"],
-        "//tensorflow:android_arm": ["tflite_with_ruy_only_enabled"],
+        ":chromiumos_arm64": [":tflite_with_ruy_enabled"],
+        ":cpu_aarch64": [":tflite_with_ruy_enabled"],
+        ":cpu_arm64": [":tflite_with_ruy_enabled"],
+        ":cpu_arm64e": [":tflite_with_ruy_enabled"],
+        ":cpu_ios_arm64": [":tflite_with_ruy_enabled"],
+        ":cpu_ios_arm64e": [":tflite_with_ruy_enabled"],
+        ":cpu_arm64_v8a": [":tflite_with_ruy_enabled"],
+        "//tensorflow:android_arm": ["tflite_with_ruy_enabled"],
         "//conditions:default": [],
     },
     visibility = ["//visibility:private"],
 )
 
 cc_library(
-    name = "tflite_with_ruy_only",
+    name = "tflite_with_ruy",
     build_for_embedded = True,
     select_deps = {
-        ":tflite_with_ruy_only_explicit_true": [":tflite_with_ruy_only_enabled"],
-        ":tflite_with_ruy_only_explicit_false": [],
-        "//conditions:default": [":tflite_with_ruy_only_default"],
+        ":tflite_with_ruy_explicit_true": [":tflite_with_ruy_enabled"],
+        ":tflite_with_ruy_explicit_false": [],
+        "//conditions:default": [":tflite_with_ruy_default"],
     },
 )
 
@@ -291,7 +291,7 @@ cc_library(
     ],
     copts = tflite_copts(),
     deps = [
-        ":tflite_with_ruy_only",
+        ":tflite_with_ruy",
         ":op_macros",
         # For now this unconditionally depends on both ruy and gemmlowp.
         # See the comment inside class CpuBackendContext on the
@@ -311,11 +311,11 @@ cc_library(
     copts = tflite_copts(),
     deps = [
         ":cpu_backend_context",
-        ":tflite_with_ruy_only",
+        ":tflite_with_ruy",
         "//tensorflow/lite/kernels/internal:compatibility",
         "//tensorflow/lite/kernels/internal:types",
         # For now this unconditionally depends on both ruy and gemmlowp.
-        # We only need to depend on gemmlowp when tflite_with_ruy_only
+        # We only need to depend on gemmlowp when tflite_with_ruy
         # is false, but putting these dependencies in a select() seems to
         # defeat copybara's rewriting rules.
         "@ruy//ruy:context",
@@ -349,20 +349,20 @@ cc_library(
     ],
     copts = tflite_copts(),
     deps = [
-        ":tflite_with_ruy_only",
+        ":tflite_with_ruy",
         "//tensorflow/lite/kernels/internal:common",
         "//tensorflow/lite/kernels/internal:compatibility",
         "//tensorflow/lite/kernels/internal:cpu_check",
         "//tensorflow/lite/kernels/internal:types",
         ":cpu_backend_context",
         ":cpu_backend_threadpool",
-        # Depend on ruy regardless of `tflite_with_ruy_only`. See the comment in
+        # Depend on ruy regardless of `tflite_with_ruy`. See the comment in
         # cpu_backend_gemm.h about why ruy is the generic path.
         "@ruy//ruy",
         "@ruy//ruy:matrix",
         "@ruy//ruy:path",
         "@ruy//ruy/profiler:instrumentation",
-        # We only need to depend on gemmlowp and Eigen when tflite_with_ruy_only
+        # We only need to depend on gemmlowp and Eigen when tflite_with_ruy
         # is false, but putting these dependencies in a select() seems to
         # defeat copybara's rewriting rules.
         "@gemmlowp",
@@ -605,7 +605,7 @@ cc_library(
         "//tensorflow/lite/kernels/internal:cppmath",
         "//tensorflow/lite:string",
         "@farmhash_archive//:farmhash",
-    ] + [":tflite_with_ruy_only_and_caching_enabled"],
+    ] + [":tflite_with_ruy_and_caching_enabled"],
 )
 
 cc_library(
diff --git a/tensorflow/lite/kernels/conv.cc b/tensorflow/lite/kernels/conv.cc
index 88765b2f9c4..81069de1abe 100644
--- a/tensorflow/lite/kernels/conv.cc
+++ b/tensorflow/lite/kernels/conv.cc
@@ -26,7 +26,7 @@ limitations under the License.
 #include "tensorflow/lite/kernels/internal/compatibility.h"
 #include "tensorflow/lite/kernels/internal/types.h"
 // b/131835803 forces us to include multithreaded_conv.h before optimized_ops.h
-#ifndef TFLITE_WITH_RUY_ONLY
+#ifndef TFLITE_WITH_RUY
 #include "tensorflow/lite/kernels/internal/optimized/multithreaded_conv.h"
 #endif
 #include "tensorflow/lite/kernels/internal/optimized/optimized_ops.h"
@@ -765,8 +765,8 @@ void EvalFloat(TfLiteContext* context, TfLiteNode* node,
       break;
     }
     case kMultithreadOptimized: {
-#ifdef TFLITE_WITH_RUY_ONLY
-      // See Register_CONV_2D: we should never be here when TFLITE_WITH_RUY_ONLY
+#ifdef TFLITE_WITH_RUY
+      // See Register_CONV_2D: we should never be here when TFLITE_WITH_RUY
       // was enabled. We #if out this code in order to get the corresponding
       // binary size benefits.
       TFLITE_DCHECK(false);
@@ -1051,8 +1051,8 @@ TfLiteRegistration* Register_CONVOLUTION_CBLAS_OPT() {
 TfLiteRegistration* Register_CONV_2D() {
 #if defined TFLITE_USE_APPLE_ACCELERATE_FOR_CONV
   return Register_CONVOLUTION_CBLAS_OPT();
-#elif defined TFLITE_WITH_RUY_ONLY
-  // TFLITE_WITH_RUY_ONLY optimizes the generic kernel type.
+#elif defined TFLITE_WITH_RUY
+  // TFLITE_WITH_RUY optimizes the generic kernel type.
   return Register_CONVOLUTION_GENERIC_OPT();
 #else
   return Register_CONVOLUTION_MULTITHREADED_OPT();
@@ -1063,8 +1063,8 @@ TfLiteRegistration* Register_CONV_2D() {
 // models only need the UINT8 type. TFLite's op registration mechanism doesn't
 // yet allow for more nuanced registration mechanisms.
 TfLiteRegistration* Register_CONV_2D_UINT8() {
-#if defined TFLITE_WITH_RUY_ONLY
-  // TFLITE_WITH_RUY_ONLY optimizes the generic kernel type.
+#if defined TFLITE_WITH_RUY
+  // TFLITE_WITH_RUY optimizes the generic kernel type.
   return Register_CONVOLUTION_GENERIC_OPT_UINT8();
 #else
   return Register_CONV_2D();
diff --git a/tensorflow/lite/kernels/conv_test.cc b/tensorflow/lite/kernels/conv_test.cc
index a1fd34eb1cb..ac78bc6b353 100644
--- a/tensorflow/lite/kernels/conv_test.cc
+++ b/tensorflow/lite/kernels/conv_test.cc
@@ -148,7 +148,7 @@ class ConvolutionOpModel : public BaseConvolutionOpModel {
 const auto kKernelMap = new std::map({
     {"Reference", ops::builtin::Register_CONVOLUTION_REF()},
     {"GenericOptimized", ops::builtin::Register_CONVOLUTION_GENERIC_OPT()},
-#ifndef TFLITE_WITH_RUY_ONLY
+#ifndef TFLITE_WITH_RUY
     {"MultithreadedOptimized",
      ops::builtin::Register_CONVOLUTION_MULTITHREADED_OPT()},
 #endif
diff --git a/tensorflow/lite/kernels/cpu_backend_context.h b/tensorflow/lite/kernels/cpu_backend_context.h
index 19ef88bf8e3..124b9b849a2 100644
--- a/tensorflow/lite/kernels/cpu_backend_context.h
+++ b/tensorflow/lite/kernels/cpu_backend_context.h
@@ -56,7 +56,7 @@ class CpuBackendContext final : public TfLiteInternalBackendContext {
   // (see :cpu_backend_gemm), for now a CpuBackendContext always
   // stores both a gemmlowp context and a ruy context.
   // TODO(b/131416458): Once call sites all go through abstractions,
-  // elide what can be elided based on TFLITE_WITH_RUY_ONLY.
+  // elide what can be elided based on TFLITE_WITH_RUY.
   const std::unique_ptr ruy_context_;
   const std::unique_ptr gemmlowp_context_;
 
diff --git a/tensorflow/lite/kernels/cpu_backend_gemm.h b/tensorflow/lite/kernels/cpu_backend_gemm.h
index f4d20d8970a..a95c4d15a82 100644
--- a/tensorflow/lite/kernels/cpu_backend_gemm.h
+++ b/tensorflow/lite/kernels/cpu_backend_gemm.h
@@ -24,7 +24,7 @@ limitations under the License.
 #include "tensorflow/lite/kernels/cpu_backend_gemm_params.h"
 #include "tensorflow/lite/kernels/cpu_backend_gemm_ruy.h"
 
-#ifndef TFLITE_WITH_RUY_ONLY
+#ifndef TFLITE_WITH_RUY
 #include "tensorflow/lite/kernels/cpu_backend_gemm_eigen.h"
 #include "tensorflow/lite/kernels/cpu_backend_gemm_gemmlowp.h"
 #endif
@@ -42,7 +42,7 @@ template  {};
 
-#ifndef TFLITE_WITH_RUY_ONLY
+#ifndef TFLITE_WITH_RUY
 
 /* Specializations using gemmlowp */
 
@@ -82,7 +82,7 @@ template <>
 struct GemmImpl
     : detail::GemmImplUsingEigen {};
 
-#endif  // not TFLITE_WITH_RUY_ONLY
+#endif  // not TFLITE_WITH_RUY
 
 /* Public entry point */
 
diff --git a/tensorflow/lite/kernels/cpu_backend_gemm_custom_gemv.h b/tensorflow/lite/kernels/cpu_backend_gemm_custom_gemv.h
index 224f8ecea41..2712d7d2cd1 100644
--- a/tensorflow/lite/kernels/cpu_backend_gemm_custom_gemv.h
+++ b/tensorflow/lite/kernels/cpu_backend_gemm_custom_gemv.h
@@ -591,10 +591,10 @@ struct CustomGemvImpl
 
 #include "tensorflow/lite/kernels/internal/compatibility.h"
-#ifndef TFLITE_WITH_RUY_ONLY
+#ifndef TFLITE_WITH_RUY
 
 #include 
 #include 
@@ -190,6 +190,6 @@ struct GemmImplUsingGemmlowpmax_num_threads();
   thread_count = std::max(1, std::min(thread_count, max_threads));
-#ifndef TFLITE_WITH_RUY_ONLY
+#ifndef TFLITE_WITH_RUY
   // Cap the number of threads to 2 for float path to avoid regression in
   // performance (b/132294857).
   if (std::is_floating_point::value) {
diff --git a/tensorflow/lite/tools/make/Makefile b/tensorflow/lite/tools/make/Makefile
index 266aa94ad93..f8b67fbbe7d 100644
--- a/tensorflow/lite/tools/make/Makefile
+++ b/tensorflow/lite/tools/make/Makefile
@@ -187,7 +187,7 @@ ifeq ($(TARGET_ARCH),aarch64)
 	BUILD_WITH_RUY=true
 endif
 ifeq ($(BUILD_WITH_RUY),true)
-  CXXFLAGS += -DTFLITE_WITH_RUY_ONLY
+  CXXFLAGS += -DTFLITE_WITH_RUY
 endif
 
 BUILD_WITH_RUY_PROFILER ?= false

From 3f2c98610e06fa5fbf296f6c25e5b4020dbbefa5 Mon Sep 17 00:00:00 2001
From: Nathan Luehr 
Date: Mon, 22 Jun 2020 12:29:33 -0500
Subject: [PATCH 435/867] Move ScopedCublasMathMode inside DoBlasInternalImpl

---
 tensorflow/stream_executor/cuda/cuda_blas.cc | 127 ++++++++++---------
 tensorflow/stream_executor/cuda/cuda_blas.h  |  10 +-
 2 files changed, 76 insertions(+), 61 deletions(-)

diff --git a/tensorflow/stream_executor/cuda/cuda_blas.cc b/tensorflow/stream_executor/cuda/cuda_blas.cc
index e387690da26..e86e3089038 100644
--- a/tensorflow/stream_executor/cuda/cuda_blas.cc
+++ b/tensorflow/stream_executor/cuda/cuda_blas.cc
@@ -386,9 +386,9 @@ cudaDataType_t CUDAComputationType(blas::ComputationType ty) {
 }  // namespace
 
 template 
-bool CUDABlas::DoBlasInternalImpl(FuncT cublas_func, Stream *stream,
+bool CUDABlas::DoBlasInternalImpl(FuncT cublas_func, Stream* stream,
                                   bool pointer_mode_host, bool err_on_failure,
-                                  Args... args) {
+                                  cublasMath_t math_type, Args... args) {
   absl::MutexLock lock(&mu_);
 
   CHECK(blas_ != nullptr);
@@ -396,6 +396,20 @@ bool CUDABlas::DoBlasInternalImpl(FuncT cublas_func, Stream *stream,
     return false;
   }
 
+#if CUDA_VERSION >= 9000
+  ScopedCublasMathMode math_mode{blas_};
+#if CUBLAS_VER_MAJOR >= 11
+  if (math_type == CUBLAS_TF32_TENSOR_OP_MATH &&
+      tensorflow::tf32_execution_allowed()) {
+#else
+  if (math_type == CUBLAS_TENSOR_OP_MATH) {
+#endif
+    if (!math_mode.Init(math_type)) {
+      return false;
+    }
+  }
+#endif
+
   gpu::ScopedActivateExecutorContext sac{parent_};
   ScopedCublasPointerMode pointer_mode{blas_};
   if (!pointer_mode.Init(pointer_mode_host ? CUBLAS_POINTER_MODE_HOST
@@ -1615,15 +1629,14 @@ bool CUDABlas::DoBlasGemm(
   }
 
 #if CUDA_VERSION < 11000
-  ScopedCublasMathMode math_mode{blas_};
-  if (!math_mode.Init(CUBLAS_TENSOR_OP_MATH)) {
-    return false;
-  }
+  cublasMath_t math_type = CUBLAS_TENSOR_OP_MATH;
+#else
+  cublasMath_t math_type = CUBLAS_DEFAULT_MATH;
 #endif
 
   return DoBlasInternalImpl(
       cublasSgemmEx, stream, true /* = pointer_mode_host */,
-      true /* = err_on_failure= */, CUDABlasTranspose(transa),
+      true /* = err_on_failure= */, math_type, CUDABlasTranspose(transa),
       CUDABlasTranspose(transb), m, n, k, &alpha, GpuMemory(a),
       SE_CUDA_DATA_HALF, lda, GpuMemory(b), SE_CUDA_DATA_HALF, ldb, &beta,
       GpuMemoryMutable(c), SE_CUDA_DATA_HALF, ldc);
@@ -1669,19 +1682,17 @@ bool CUDABlas::DoBlasGemm(Stream *stream, blas::Transpose transa,
     }
   }
 
-#if CUBLAS_VER_MAJOR >= 11
-  ScopedCublasMathMode math_mode{blas_};
-  if (tensorflow::tf32_execution_allowed()) {
-    if (!math_mode.Init(CUBLAS_TF32_TENSOR_OP_MATH)) {
-      return false;
-    }
-  }
+#if CUDA_VERSION < 11000
+  cublasMath_t math_type = CUBLAS_DEFAULT_MATH;
+#else
+  cublasMath_t math_type = CUBLAS_TF32_TENSOR_OP_MATH;
 #endif
 
-  return DoBlasInternal(cublasSgemm, stream, true /* = pointer_mode_host */,
-                        CUDABlasTranspose(transa), CUDABlasTranspose(transb), m,
-                        n, k, &alpha, GpuMemory(a), lda, GpuMemory(b), ldb,
-                        &beta, GpuMemoryMutable(c), ldc);
+  return DoBlasInternalImpl(
+      cublasSgemm, stream, true /* = pointer_mode_host */,
+      true /* = err_on_failure */, math_type, CUDABlasTranspose(transa),
+      CUDABlasTranspose(transb), m, n, k, &alpha, GpuMemory(a), lda,
+      GpuMemory(b), ldb, &beta, GpuMemoryMutable(c), ldc);
 }
 
 bool CUDABlas::DoBlasGemm(Stream *stream, blas::Transpose transa,
@@ -1704,16 +1715,6 @@ bool CUDABlas::DoBlasGemm(Stream *stream, blas::Transpose transa,
                           DeviceMemory> *c, int ldc) {
   auto cb_alpha = GpuComplexValue(alpha);
   auto cb_beta = GpuComplexValue(beta);
-
-#if CUBLAS_VER_MAJOR >= 11
-  ScopedCublasMathMode math_mode{blas_};
-  if (tensorflow::tf32_execution_allowed()) {
-    if (!math_mode.Init(CUBLAS_TF32_TENSOR_OP_MATH)) {
-      return false;
-    }
-  }
-#endif
-
   return DoBlasInternal(cublasCgemm, stream, true /* = pointer_mode_host */,
                         CUDABlasTranspose(transa), CUDABlasTranspose(transb), m,
                         n, k, GpuComplex(&cb_alpha), GpuComplex(GpuMemory(a)),
@@ -2286,10 +2287,27 @@ port::Status CUDABlas::DoBlasGemmBatchedInternal(
   if (stream->parent()->GetDeviceDescription().cuda_compute_capability(
           &cc_major, &cc_minor) &&
       cc_major >= 5) {
-    bool use_tensor_ops =
-        data_type == CUDA_R_16F || tensorflow::tf32_execution_allowed();
-    cublasGemmAlgo_t algo =
-        (use_tensor_ops ? CUBLAS_GEMM_DFALT_TENSOR_OP : CUBLAS_GEMM_DFALT);
+    cublasMath_t math_type;
+    cublasGemmAlgo_t algo;
+    if (data_type == CUDA_R_16F) {
+#if CUDA_VERSION < 11000
+      math_type = CUBLAS_TENSOR_OP_MATH;
+#else
+      math_type = CUBLAS_DEFAULT_MATH;
+#endif
+      algo = CUBLAS_GEMM_DFALT_TENSOR_OP;
+#if CUBLAS_VER_MAJOR >= 11
+    } else if (data_type == CUDA_R_32F) {
+      // DoBlassInternalImpl will switch math_type back to CUBLAS_DEFAULT_MATH
+      // if TF32 is disabled.
+      math_type = CUBLAS_TF32_TENSOR_OP_MATH;
+      algo = tensorflow::tf32_execution_allowed() ? CUBLAS_GEMM_DFALT_TENSOR_OP
+                                                  : CUBLAS_GEMM_DFALT;
+#endif
+    } else {
+      math_type = CUBLAS_DEFAULT_MATH;
+      algo = CUBLAS_GEMM_DFALT;
+    }
     cudaDataType_t compute_type =
         (data_type == CUDA_R_16F ? CUDA_R_32F : data_type);
     const void **a_void_ptrs = reinterpret_cast(
@@ -2301,7 +2319,7 @@ port::Status CUDABlas::DoBlasGemmBatchedInternal(
     bool ok;
     ok = DoBlasInternalImpl(
         AS_LAMBDA(cublasGemmBatchedEx), stream, true /* = pointer_mode_host */,
-        true /* = err_on_failure */, CUDABlasTranspose(transa),
+        true /* = err_on_failure */, math_type, CUDABlasTranspose(transa),
         CUDABlasTranspose(transb), m, n, k, &alpha, a_void_ptrs, data_type, lda,
         b_void_ptrs, data_type, ldb, &beta, c_void_ptrs, data_type, ldc,
         batch_count, compute_type, algo);
@@ -2475,9 +2493,14 @@ bool CUDABlas::DoBlasGemmStridedBatched(
       cc_major >= 5) {
     cublasGemmAlgo_t algo =
         (cc_major >= 7 ? CUBLAS_GEMM_DFALT_TENSOR_OP : CUBLAS_GEMM_DFALT);
+#if CUDA_VERSION < 11000
+    cublasMath_t math_type = CUBLAS_TENSOR_OP_MATH;
+#else
+    cublasMath_t math_type = CUBLAS_DEFAULT_MATH;
+#endif
     bool ok = DoBlasInternalImpl(
         AS_LAMBDA(cublasGemmStridedBatchedEx), stream,
-        true /* = pointer_mode_host */, true /* = err_on_failure */,
+        true /* = pointer_mode_host */, true /* = err_on_failure */, math_type,
         CUDABlasTranspose(transa), CUDABlasTranspose(transb), m, n, k, &alpha,
         GpuMemory(a), CUDA_R_16F, lda, stride_a, GpuMemory(b), CUDA_R_16F, ldb,
         stride_b, &beta, GpuMemoryMutable(c), CUDA_R_16F, ldc, stride_c,
@@ -2499,10 +2522,10 @@ bool CUDABlas::DoBlasGemmStridedBatched(
         reinterpret_cast<__half *>(GpuMemoryMutable(c) + batch * stride_c);
     bool ok = DoBlasInternalImpl(
         cublasSgemmEx, stream, true /* = pointer_mode_host */,
-        true /* = err_on_failure= */, CUDABlasTranspose(transa),
-        CUDABlasTranspose(transb), m, n, k, &alpha, a_matrix, SE_CUDA_DATA_HALF,
-        lda, b_matrix, SE_CUDA_DATA_HALF, ldb, &beta, c_matrix,
-        SE_CUDA_DATA_HALF, ldc);
+        true /* = err_on_failure= */, CUBLAS_DEFAULT_MATH,
+        CUDABlasTranspose(transa), CUDABlasTranspose(transb), m, n, k, &alpha,
+        a_matrix, SE_CUDA_DATA_HALF, lda, b_matrix, SE_CUDA_DATA_HALF, ldb,
+        &beta, c_matrix, SE_CUDA_DATA_HALF, ldc);
     if (!ok) {
       LOG(ERROR) << "failed BLAS call, see log for details";
       return false;
@@ -2517,19 +2540,17 @@ bool CUDABlas::DoBlasGemmStridedBatched(
     int64 stride_a, const DeviceMemory &b, int ldb, int64 stride_b,
     float beta, DeviceMemory *c, int ldc, int64 stride_c,
     int batch_count) {
-#if CUBLAS_VER_MAJOR >= 11
-  ScopedCublasMathMode math_mode{blas_};
-  if (tensorflow::tf32_execution_allowed()) {
-    if (!math_mode.Init(CUBLAS_TF32_TENSOR_OP_MATH)) {
-      return false;
-    }
-  }
+#if CUDA_VERSION < 11000
+  cublasMath_t math_type = CUBLAS_DEFAULT_MATH;
+#else
+  cublasMath_t math_type = CUBLAS_TF32_TENSOR_OP_MATH;
 #endif
-  return DoBlasInternal(
+  return DoBlasInternalImpl(
       cublasSgemmStridedBatched, stream, true /* = pointer_mode_host */,
-      CUDABlasTranspose(transa), CUDABlasTranspose(transb), m, n, k, &alpha,
-      GpuMemory(a), lda, stride_a, GpuMemory(b), ldb, stride_b, &beta,
-      GpuMemoryMutable(c), ldc, stride_c, batch_count);
+      true /* = err_on_failure */, math_type, CUDABlasTranspose(transa),
+      CUDABlasTranspose(transb), m, n, k, &alpha, GpuMemory(a), lda, stride_a,
+      GpuMemory(b), ldb, stride_b, &beta, GpuMemoryMutable(c), ldc, stride_c,
+      batch_count);
 }
 
 bool CUDABlas::DoBlasGemmStridedBatched(
@@ -2552,14 +2573,6 @@ bool CUDABlas::DoBlasGemmStridedBatched(
     const DeviceMemory> &b, int ldb, int64 stride_b,
     std::complex beta, DeviceMemory> *c, int ldc,
     int64 stride_c, int batch_count) {
-#if CUBLAS_VER_MAJOR >= 11
-  ScopedCublasMathMode math_mode{blas_};
-  if (tensorflow::tf32_execution_allowed()) {
-    if (!math_mode.Init(CUBLAS_TF32_TENSOR_OP_MATH)) {
-      return false;
-    }
-  }
-#endif
   auto cb_alpha = GpuComplexValue(alpha);
   auto cb_beta = GpuComplexValue(beta);
   return DoBlasInternal(
diff --git a/tensorflow/stream_executor/cuda/cuda_blas.h b/tensorflow/stream_executor/cuda/cuda_blas.h
index 556456c83db..ca03ce96465 100644
--- a/tensorflow/stream_executor/cuda/cuda_blas.h
+++ b/tensorflow/stream_executor/cuda/cuda_blas.h
@@ -81,9 +81,9 @@ class CUDABlas : public blas::BlasSupport {
   // err_on_failure:     Whether to print an error if the cublas function fails.
   // args:               Arguments of cuBLAS function.
   template 
-  bool DoBlasInternalImpl(FuncT cublas_func, Stream *stream,
+  bool DoBlasInternalImpl(FuncT cublas_func, Stream* stream,
                           bool pointer_mode_host, bool err_on_failure,
-                          Args... args);
+                          cublasMath_t math_type, Args... args);
 
   // Convenience functions that call DoBlasInternalImpl with different values
   // for err_on_failure.
@@ -91,7 +91,8 @@ class CUDABlas : public blas::BlasSupport {
   bool DoBlasInternal(FuncT cublas_func, Stream *stream, bool pointer_mode_host,
                       Args... args) {
     return DoBlasInternalImpl(cublas_func, stream, pointer_mode_host,
-                              /*err_on_failure=*/true, args...);
+                              /*err_on_failure=*/true, CUBLAS_DEFAULT_MATH,
+                              args...);
   }
   template 
   bool DoBlasInternalFailureOK(FuncT cublas_func, Stream *stream,
@@ -99,7 +100,8 @@ class CUDABlas : public blas::BlasSupport {
     // Tensor ops are hard-coded off in this path, but can still be enabled with
     // a specific algorithm choice as in DoBlasGemmWithAlgorithmImpl().
     return DoBlasInternalImpl(cublas_func, stream, pointer_mode_host,
-                              /*err_on_failure=*/false, args...);
+                              /*err_on_failure=*/false, CUBLAS_DEFAULT_MATH,
+                              args...);
   }
 
   // A helper function to implement DoBlasGemmBatched interfaces for generic

From 01c3f45f0fc90dc8e2d0b6d697beb8b48940dc15 Mon Sep 17 00:00:00 2001
From: Geoffrey Martin-Noble 
Date: Mon, 22 Jun 2020 10:07:49 -0700
Subject: [PATCH 436/867] Update LLVM commit to match upstream

PiperOrigin-RevId: 317677703
Change-Id: I7ae6332a7811402bbcce17f205b1c8a5fd854133
---
 tensorflow/workspace.bzl | 4 ++--
 1 file changed, 2 insertions(+), 2 deletions(-)

diff --git a/tensorflow/workspace.bzl b/tensorflow/workspace.bzl
index 73698987a08..5af8c8272b8 100755
--- a/tensorflow/workspace.bzl
+++ b/tensorflow/workspace.bzl
@@ -710,8 +710,8 @@ def tf_repositories(path_prefix = "", tf_repo_name = ""):
     )
 
     # Check out LLVM and MLIR from llvm-project.
-    LLVM_COMMIT = "7f0d7f32631648acf48bc23047635ab5e2058a1a"
-    LLVM_SHA256 = "2f1dbae231b3b8f9c67d6a4f578c8ce29f3aa2831313b34c40ff2edb4014476a"
+    LLVM_COMMIT = "7e825abd5704ce28b166f9463d4bd304348fd2a9"
+    LLVM_SHA256 = "a21b752ee1866e195f3f72c7931c79f8c4ecc0f14861488284bdc2bdf14d6fe9"
     LLVM_URLS = [
         "https://storage.googleapis.com/mirror.tensorflow.org/github.com/llvm/llvm-project/archive/{commit}.tar.gz".format(commit = LLVM_COMMIT),
         "https://github.com/llvm/llvm-project/archive/{commit}.tar.gz".format(commit = LLVM_COMMIT),

From 36c56b0faac9cae5e17738796a77a0d1e5108bea Mon Sep 17 00:00:00 2001
From: Raman Sarokin 
Date: Mon, 22 Jun 2020 10:10:18 -0700
Subject: [PATCH 437/867] Removed trivial copy constructor, copy assignment
 operator.

PiperOrigin-RevId: 317678239
Change-Id: I02723cbe192c225f18231719201e350a138d3818
---
 tensorflow/lite/delegates/gpu/cl/gpu_object.h     |  8 --------
 tensorflow/lite/delegates/gpu/cl/linear_storage.h | 14 --------------
 tensorflow/lite/delegates/gpu/cl/tensor_type.h    | 14 --------------
 3 files changed, 36 deletions(-)

diff --git a/tensorflow/lite/delegates/gpu/cl/gpu_object.h b/tensorflow/lite/delegates/gpu/cl/gpu_object.h
index a6d28436872..faf18b539e2 100644
--- a/tensorflow/lite/delegates/gpu/cl/gpu_object.h
+++ b/tensorflow/lite/delegates/gpu/cl/gpu_object.h
@@ -105,14 +105,6 @@ struct GPUResourcesWithValue {
 class GPUObjectDescriptor {
  public:
   GPUObjectDescriptor() = default;
-  GPUObjectDescriptor(const GPUObjectDescriptor& obj_desc)
-      : state_vars_(obj_desc.state_vars_) {}
-  GPUObjectDescriptor& operator=(const GPUObjectDescriptor& obj_desc) {
-    if (this != &obj_desc) {
-      state_vars_ = obj_desc.state_vars_;
-    }
-    return *this;
-  }
   virtual ~GPUObjectDescriptor() = default;
 
   void SetStateVar(const std::string& key, const std::string& value) const {
diff --git a/tensorflow/lite/delegates/gpu/cl/linear_storage.h b/tensorflow/lite/delegates/gpu/cl/linear_storage.h
index c7f55e1d91d..2c0770ef3dc 100644
--- a/tensorflow/lite/delegates/gpu/cl/linear_storage.h
+++ b/tensorflow/lite/delegates/gpu/cl/linear_storage.h
@@ -41,20 +41,6 @@ struct TensorLinearDescriptor : public GPUObjectDescriptor {
   LinearStorageType storage_type;
   DataType element_type;  // FLOAT32 or FLOAT16
 
-  TensorLinearDescriptor() = default;
-  TensorLinearDescriptor(const TensorLinearDescriptor& desc)
-      : GPUObjectDescriptor(desc),
-        storage_type(desc.storage_type),
-        element_type(desc.element_type) {}
-  TensorLinearDescriptor& operator=(const TensorLinearDescriptor& desc) {
-    if (this != &desc) {
-      storage_type = desc.storage_type;
-      element_type = desc.element_type;
-      GPUObjectDescriptor::operator=(desc);
-    }
-    return *this;
-  }
-
   absl::Status PerformSelector(const std::string& selector,
                                const std::vector& args,
                                const std::vector& template_args,
diff --git a/tensorflow/lite/delegates/gpu/cl/tensor_type.h b/tensorflow/lite/delegates/gpu/cl/tensor_type.h
index 2d4ae0c7335..7e173753217 100644
--- a/tensorflow/lite/delegates/gpu/cl/tensor_type.h
+++ b/tensorflow/lite/delegates/gpu/cl/tensor_type.h
@@ -48,20 +48,6 @@ struct TensorDescriptor : public GPUObjectDescriptor {
   TensorDescriptor() = default;
   TensorDescriptor(DataType dt, TensorStorageType st, Layout l)
       : data_type(dt), storage_type(st), layout(l) {}
-  TensorDescriptor(const TensorDescriptor& desc)
-      : GPUObjectDescriptor(desc),
-        data_type(desc.data_type),
-        storage_type(desc.storage_type),
-        layout(desc.layout) {}
-  TensorDescriptor& operator=(const TensorDescriptor& desc) {
-    if (this != &desc) {
-      data_type = desc.data_type;
-      storage_type = desc.storage_type;
-      layout = desc.layout;
-      GPUObjectDescriptor::operator=(desc);
-    }
-    return *this;
-  }
 
   bool operator==(const TensorDescriptor& d) const {
     return data_type == d.data_type && storage_type == d.storage_type &&

From a6c009467a89875eb78339c7556755bbbbf4c84e Mon Sep 17 00:00:00 2001
From: Gaurav Jain 
Date: Mon, 22 Jun 2020 10:20:00 -0700
Subject: [PATCH 438/867] Retrace tf.function when using different variables

Since variable ids can be used to python data structures it can be
dangerous to reuse the previous function trace if the shape and dtype is
the same. We thus remove the optimization and ensure we always retrace.

PiperOrigin-RevId: 317680370
Change-Id: I1dfa3a626074e623b735869aa724138072e2c274
---
 tensorflow/python/eager/def_function_test.py |  4 +-
 tensorflow/python/eager/function.py          | 29 +++++------
 tensorflow/python/eager/function_test.py     | 52 +++++++++++++++-----
 3 files changed, 55 insertions(+), 30 deletions(-)

diff --git a/tensorflow/python/eager/def_function_test.py b/tensorflow/python/eager/def_function_test.py
index 0549da2c256..6dc4e322bbd 100644
--- a/tensorflow/python/eager/def_function_test.py
+++ b/tensorflow/python/eager/def_function_test.py
@@ -568,10 +568,10 @@ class DefFunctionTest(test.TestCase, parameterized.TestCase):
     self.assertEqual(trace_count[0], 1)
     self.assertEqual(self.evaluate(v1), 2.0)
     double_variable(v2)
-    self.assertEqual(trace_count[0], 1 if ops.Tensor._USE_EQUALITY else 2)
+    self.assertEqual(trace_count[0], 2)
     self.assertEqual(self.evaluate(v2), 4.0)
     double_variable(v3)
-    self.assertEqual(trace_count[0], 2 if ops.Tensor._USE_EQUALITY else 3)
+    self.assertEqual(trace_count[0], 3)
     self.assertEqual(self.evaluate(v3), 8)
 
   def testShapeCache(self):
diff --git a/tensorflow/python/eager/function.py b/tensorflow/python/eager/function.py
index 845abe43134..7b235749533 100644
--- a/tensorflow/python/eager/function.py
+++ b/tensorflow/python/eager/function.py
@@ -92,7 +92,7 @@ IMPLEMENTS_ATTRIBUTE_NAME = "_implements"
 SHARED_RENDEZVOUS_ATTRIBUTE_NAME = "shared_rendezvous"
 
 
-def _make_input_signature_hashable(elem, variable_map=None):
+def _make_input_signature_hashable(elem):
   """Rewrite input signature to be hashable.
 
   We replace nested variables in the input signature with TensorSpec in order to
@@ -100,18 +100,13 @@ def _make_input_signature_hashable(elem, variable_map=None):
 
   Args:
     elem: Input signature element
-    variable_map: Internal argument used for tracking variable aliases
 
   Returns:
     A hashable object for the requested input signature
   """
-  if variable_map is None:
-    variable_map = {}
-
   # TODO(slebedev): consider using nest.
   if isinstance(elem, tuple):
-    return tuple(map(lambda e: _make_input_signature_hashable(e, variable_map),
-                     elem))
+    return tuple(map(_make_input_signature_hashable, elem))
 
   try:
     hash(elem)
@@ -122,15 +117,17 @@ def _make_input_signature_hashable(elem, variable_map=None):
     v = elem()
 
     if resource_variable_ops.is_resource_variable(v):
-      idx = variable_map.get(id(v))
-      if idx is None:
-        idx = len(variable_map)
-        variable_map[id(v)] = idx
-
-      # We include the class name to avoid having different types of variables
-      # having the same hash. We Also include the variable index which allows
-      # us to return a different hash if variables have been aliased in a call.
-      return v.__class__, tensor_spec.TensorSpec(v.shape, v.dtype), idx
+      # We special case variables here to use unique_id as the cache key. This
+      # ensures we have to retrace whenever a different variable is passed in.
+      # This is needed to support cases where the user may use the id of a
+      # variable in the function perhaps as a lookup in a dictionary.
+      #
+      # This choice leads to more retracing when we could have possibly used the
+      # shape and dtype instead. However, we expect the number of variables in a
+      # program to be bounded, and correspondingly the number of retraces.
+      #
+      # Note we also include the class name to avoid collisions with strings.
+      return v.__class__, v._unique_id  # pylint: disable=protected-access
 
     if _is_ndarray(v):
       # Numpy arrays are not hashable, but when calling functions we treat them
diff --git a/tensorflow/python/eager/function_test.py b/tensorflow/python/eager/function_test.py
index f6a36701deb..2c49795ba8a 100644
--- a/tensorflow/python/eager/function_test.py
+++ b/tensorflow/python/eager/function_test.py
@@ -18,6 +18,7 @@ from __future__ import division
 from __future__ import print_function
 
 import collections
+import copy
 import functools
 import itertools
 import multiprocessing.pool
@@ -2930,30 +2931,57 @@ class FunctionTest(test.TestCase, parameterized.TestCase):
     # should only get a miss if the aliasing changed.
     defined(x, y, z)
     self.assertLen(total_function_cache(defined), 1)
-
-    # Calling again is a cache hit
     defined(x, y, z)
     self.assertLen(total_function_cache(defined), 1)
 
-    # Re-arranging arguments doesn't change signature
+    # Re-arranging arguments causes cache miss
     defined(z, y, x)
-    self.assertLen(total_function_cache(defined),
-                   1 if ops.Tensor._USE_EQUALITY else 2)
+    self.assertLen(total_function_cache(defined), 2)
+    defined(z, y, x)
+    self.assertLen(total_function_cache(defined), 2)
 
     # Aliasing causes cache miss
     defined(x, x, z)
-    self.assertLen(total_function_cache(defined),
-                   2 if ops.Tensor._USE_EQUALITY else 3)
+    self.assertLen(total_function_cache(defined), 3)
+    defined(x, x, z)
+    self.assertLen(total_function_cache(defined), 3)
 
-    # Re-arranging arguments doesn't change signature
+    # Re-arranging arguments causes cache miss
     defined(y, y, z)
-    self.assertLen(total_function_cache(defined),
-                   2 if ops.Tensor._USE_EQUALITY else 4)
+    self.assertLen(total_function_cache(defined), 4)
+    defined(y, y, z)
+    self.assertLen(total_function_cache(defined), 4)
 
     # Different alias positions causes cache miss
     defined(z, y, y)
-    self.assertLen(total_function_cache(defined),
-                   3 if ops.Tensor._USE_EQUALITY else 5)
+    self.assertLen(total_function_cache(defined), 5)
+    defined(z, y, y)
+    self.assertLen(total_function_cache(defined), 5)
+
+    x_copy = copy.deepcopy(x)
+
+    # Deep copy causes cache miss
+    defined(x_copy, y, z)
+    self.assertLen(total_function_cache(defined), 6)
+    defined(x_copy, y, z)
+    self.assertLen(total_function_cache(defined), 6)
+
+  def testVariableRetracing(self):
+    v1 = variables.Variable(1.)
+    v2 = variables.Variable(1.)
+    v3 = copy.deepcopy(variables.Variable(1.))
+
+    var_dict = {id(v1): constant_op.constant(1),
+                id(v2): constant_op.constant(2),
+                id(v3): constant_op.constant(3)}
+
+    @function.defun
+    def lookup_tensor(v):
+      return var_dict[id(v)]
+
+    self.assertEqual(1, lookup_tensor(v1).numpy())
+    self.assertEqual(2, lookup_tensor(v2).numpy())
+    self.assertEqual(3, lookup_tensor(v3).numpy())
 
   def testDecoratedMethodInspect(self):
 

From a2100ebf3e049f3a3ff3f7e797e9aef5a5f8a1d3 Mon Sep 17 00:00:00 2001
From: Jonah Kohn <51345541+jonah-kohn@users.noreply.github.com>
Date: Mon, 22 Jun 2020 11:02:25 -0700
Subject: [PATCH 439/867] Respond to linter syntax error

---
 tensorflow/python/keras/utils/vis_utils_test.py | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/tensorflow/python/keras/utils/vis_utils_test.py b/tensorflow/python/keras/utils/vis_utils_test.py
index 0eb38a1d48f..b1dc4c21ab8 100644
--- a/tensorflow/python/keras/utils/vis_utils_test.py
+++ b/tensorflow/python/keras/utils/vis_utils_test.py
@@ -64,7 +64,7 @@ class ModelToDotFormatTest(test.TestCase):
     try:
       vis_utils.plot_model(
           model, to_file=dot_img_file, show_shapes=True,
-          show_dtype=True expand_nested=True)
+          show_dtype=True, expand_nested=True)
       self.assertTrue(file_io.file_exists(dot_img_file))
       file_io.delete_file(dot_img_file)
     except ImportError:

From 9001cc1d4583a413b7a8d636ff7735886cca0ffd Mon Sep 17 00:00:00 2001
From: Jonah Kohn <51345541+jonah-kohn@users.noreply.github.com>
Date: Mon, 22 Jun 2020 11:02:46 -0700
Subject: [PATCH 440/867] Fit linter indentation requirements.

---
 tensorflow/python/keras/utils/vis_utils.py | 8 ++++----
 1 file changed, 4 insertions(+), 4 deletions(-)

diff --git a/tensorflow/python/keras/utils/vis_utils.py b/tensorflow/python/keras/utils/vis_utils.py
index aa2a4c978cb..acfb589f51b 100644
--- a/tensorflow/python/keras/utils/vis_utils.py
+++ b/tensorflow/python/keras/utils/vis_utils.py
@@ -192,10 +192,10 @@ def model_to_dot(model,
     # Rebuild the label as a table including the layer's dtype.
     if show_dtype: 
       def format_dtype(dtype):
-          if dtype is None:
-            return '?'
-          else:
-            return str(dtype)
+        if dtype is None:
+          return '?'
+        else:
+          return str(dtype)
           
       label = '%s|%s' % (label, format_dtype(layer.dtype))
 

From 75a3975ab83fef46213e8d3a84a42f308cb93f6f Mon Sep 17 00:00:00 2001
From: Nick Kreeger 
Date: Mon, 22 Jun 2020 10:22:18 -0700
Subject: [PATCH 441/867] Fix c-style casts and const usage in MicroAllocator.

This change was introduced in cl/316533499 (PR: https://github.com/tensorflow/tensorflow/pull/38121). Lint was complaining of c-style casts, upon fixing it also was hiding const usage.

PiperOrigin-RevId: 317680917
Change-Id: I4d874564875e58eb5f6905c7b75562f90588bb22
---
 tensorflow/lite/micro/micro_allocator.cc | 21 ++++++++++++---------
 1 file changed, 12 insertions(+), 9 deletions(-)

diff --git a/tensorflow/lite/micro/micro_allocator.cc b/tensorflow/lite/micro/micro_allocator.cc
index f3b64bc9f39..d6c192c3ed3 100644
--- a/tensorflow/lite/micro/micro_allocator.cc
+++ b/tensorflow/lite/micro/micro_allocator.cc
@@ -166,11 +166,12 @@ class AllocationInfoBuilder {
   //  - If there's no metadata available, offline_planner_offsets is not set
   //  - If there's metadata available, offline_planner_offsets will point to the
   //    first offset in the metadata buffer list.
-  TfLiteStatus GetOfflinePlannedOffsets(const Model* model,
-                                        int32_t** offline_planner_offsets);
+  TfLiteStatus GetOfflinePlannedOffsets(
+      const Model* model, const int32_t** offline_planner_offsets);
 
   // Add allocaiton information for the tensors.
-  TfLiteStatus AddTensors(const SubGraph* subgraph, int32_t* offline_offsets,
+  TfLiteStatus AddTensors(const SubGraph* subgraph,
+                          const int32_t* offline_offsets,
                           TfLiteTensor* runtime_tensors);
 
   // Add allocation information for the scratch buffers.
@@ -206,7 +207,7 @@ TfLiteStatus AllocationInfoBuilder::Allocate() {
 }
 
 TfLiteStatus AllocationInfoBuilder::AddTensors(const SubGraph* subgraph,
-                                               int32_t* offline_offsets,
+                                               const int32_t* offline_offsets,
                                                TfLiteTensor* runtime_tensors) {
   // Set up allocation info for all tensors.
   for (size_t i = 0; i < tensor_count_; ++i) {
@@ -299,7 +300,7 @@ TfLiteStatus AllocationInfoBuilder::AddTensors(const SubGraph* subgraph,
 // |    4    | Arena byte offset of tensor #1 or -1 to allocate at runtime     |
 // | 3+(n-1) | Arena byte offset of tensor #(n-1) or -1 to allocate at runtime |
 TfLiteStatus AllocationInfoBuilder::GetOfflinePlannedOffsets(
-    const Model* model, int32_t** offline_planner_offsets) {
+    const Model* model, const int32_t** offline_planner_offsets) {
   if (model->metadata()) {
     for (size_t i = 0; i < model->metadata()->size(); ++i) {
       auto metadata = model->metadata()->Get(i);
@@ -309,9 +310,11 @@ TfLiteStatus AllocationInfoBuilder::GetOfflinePlannedOffsets(
             model->buffers();
         auto* buffer = (*buffers)[metadata->buffer()];
         auto* array = buffer->data();
-        const uint32_t* metadata_buffer = (uint32_t*)array->data();
-        const size_t nbr_tensors = (size_t)metadata_buffer[2];
-        *offline_planner_offsets = (int32_t*)&metadata_buffer[3];
+        const uint32_t* metadata_buffer =
+            reinterpret_cast(array->data());
+        const size_t nbr_tensors = static_cast(metadata_buffer[2]);
+        *offline_planner_offsets =
+            reinterpret_cast(&metadata_buffer[3]);
 
         if (tensor_count_ != nbr_tensors) {
           TF_LITE_REPORT_ERROR(reporter_,
@@ -893,7 +896,7 @@ TfLiteStatus MicroAllocator::CommitStaticMemoryPlan(const Model* model,
     TF_LITE_ENSURE_STATUS(
         builder.Init(subgraph->tensors()->size(), scratch_buffer_count_));
 
-    int32_t* offline_planner_offsets = nullptr;
+    const int32_t* offline_planner_offsets = nullptr;
     TF_LITE_ENSURE_STATUS(
         builder.GetOfflinePlannedOffsets(model, &offline_planner_offsets));
     TF_LITE_ENSURE_STATUS(builder.AddTensors(subgraph, offline_planner_offsets,

From 3a22b091cca63a3ae4c680d21a2e5e2c72712968 Mon Sep 17 00:00:00 2001
From: Robert David 
Date: Mon, 22 Jun 2020 10:29:31 -0700
Subject: [PATCH 442/867] Rename a few remaining cell_ptr and cell_scale to
 cell_state_ptr and cell_state_scale.

PiperOrigin-RevId: 317682665
Change-Id: I89ddd7893fd1478bb7e4b9ce8873d3c5e084deb1
---
 tensorflow/lite/kernels/lstm_eval.cc | 47 +++++++++++++++-------------
 1 file changed, 25 insertions(+), 22 deletions(-)

diff --git a/tensorflow/lite/kernels/lstm_eval.cc b/tensorflow/lite/kernels/lstm_eval.cc
index e5bdf3e9a1e..5b4e8a8d479 100644
--- a/tensorflow/lite/kernels/lstm_eval.cc
+++ b/tensorflow/lite/kernels/lstm_eval.cc
@@ -976,7 +976,7 @@ inline void LstmStepHybrid(
 // Scalar values:
 //   quantized_cell_clip: quantized clip value for cell.
 //   quantized_proj_clip: quantized clip value for projection.
-//   cell_scale: the power of two scale for cell state.
+//   cell_state_scale: the power of two scale for cell state.
 //
 // Zero points:
 //   output_state_zp: zero point of output state
@@ -1043,9 +1043,10 @@ inline void LstmStepInteger(
     int32_t layer_norm_output_scale_a, int32_t layer_norm_output_scale_b,
     const int32_t* input_gate_bias_ptr, const int32_t* forget_gate_bias_ptr,
     const int32_t* cell_gate_bias_ptr, const int32_t* output_gate_bias_ptr,
-    int16_t quantized_cell_clip, int8_t quantized_proj_clip, int32_t cell_scale,
-    int32_t input_variance_guard, int32_t forget_variance_guard,
-    int32_t cell_variance_guard, int32_t output_variance_guard,
+    int16_t quantized_cell_clip, int8_t quantized_proj_clip,
+    int32_t cell_state_scale, int32_t input_variance_guard,
+    int32_t forget_variance_guard, int32_t cell_variance_guard,
+    int32_t output_variance_guard,
     const int32_t* input_to_forget_effective_bias,
     const int32_t* recurrent_to_forget_effective_bias,
     const int32_t* input_to_cell_effective_bias,
@@ -1056,7 +1057,7 @@ inline void LstmStepInteger(
     const int32_t* recurrent_to_input_effective_bias,
     const int32_t* projection_effective_bias, int n_batch, int n_cell,
     int n_input, int n_output, int8_t* output_state_ptr,
-    int32_t output_state_zp, int16_t* cell_ptr, int8_t* output_ptr,
+    int32_t output_state_zp, int16_t* cell_state_ptr, int8_t* output_ptr,
     int16_t* scratch0, int16_t* scratch1, int16_t* scratch2, int16_t* scratch3,
     int8_t* scratch4, int32_t* scratch5, CpuBackendContext* context) {
   ruy::profiler::ScopeLabel label("LstmStepInteger");
@@ -1106,7 +1107,7 @@ inline void LstmStepInteger(
       scratch5, forget_gate_scratch, context);
   if (use_peephole) {
     tensor_utils::VectorBatchVectorCwiseProductAccumulate(
-        cell_to_forget_weight_ptr, n_output, cell_ptr, n_batch,
+        cell_to_forget_weight_ptr, n_output, cell_state_ptr, n_batch,
         effective_cell_to_forget_scale_a, effective_cell_to_forget_scale_b,
         forget_gate_scratch);
   }
@@ -1160,7 +1161,7 @@ inline void LstmStepInteger(
         scratch5, input_gate_scratch, context);
     if (use_peephole) {
       tensor_utils::VectorBatchVectorCwiseProductAccumulate(
-          cell_to_input_weight_ptr, n_output, cell_ptr, n_batch,
+          cell_to_input_weight_ptr, n_output, cell_state_ptr, n_batch,
           effective_cell_to_input_scale_a, effective_cell_to_input_scale_b,
           input_gate_scratch);
     }
@@ -1175,18 +1176,19 @@ inline void LstmStepInteger(
                                input_gate_scratch);
   }
 
-  // New cell.
-  tensor_utils::CwiseMul(forget_gate_scratch, cell_ptr, n_batch, n_cell, 15,
-                         forget_gate_scratch);
+  // New cell state.
+  tensor_utils::CwiseMul(forget_gate_scratch, cell_state_ptr, n_batch, n_cell,
+                         15, forget_gate_scratch);
 
   tensor_utils::CwiseMul(input_gate_scratch, cell_gate_scratch, n_batch, n_cell,
-                         30 + cell_scale, cell_gate_scratch);
+                         30 + cell_state_scale, cell_gate_scratch);
 
   tensor_utils::CwiseAdd(forget_gate_scratch, cell_gate_scratch, n_batch,
-                         n_cell, cell_ptr);
+                         n_cell, cell_state_ptr);
 
   if (quantized_cell_clip > 0) {
-    tensor_utils::CwiseClipping(cell_ptr, quantized_cell_clip, n_batch, n_cell);
+    tensor_utils::CwiseClipping(cell_state_ptr, quantized_cell_clip, n_batch,
+                                n_cell);
   }
 
   // Ouptut gate.
@@ -1202,7 +1204,7 @@ inline void LstmStepInteger(
       scratch5, output_gate_scratch, context);
   if (use_peephole) {
     tensor_utils::VectorBatchVectorCwiseProductAccumulate(
-        cell_to_output_weight_ptr, n_output, cell_ptr, n_batch,
+        cell_to_output_weight_ptr, n_output, cell_state_ptr, n_batch,
         effective_cell_to_output_scale_a, effective_cell_to_output_scale_b,
         output_gate_scratch);
   }
@@ -1218,8 +1220,8 @@ inline void LstmStepInteger(
                              output_gate_scratch);
 
   // Hidden.
-  tensor_utils::ApplyTanh(15 + cell_scale, cell_ptr, n_batch, n_cell,
-                          input_gate_scratch);
+  tensor_utils::ApplyTanh(15 + cell_state_scale, cell_state_ptr, n_batch,
+                          n_cell, input_gate_scratch);
 
   tensor_utils::CwiseMul(output_gate_scratch, input_gate_scratch,
                          effective_hidden_scale_a, effective_hidden_scale_b,
@@ -1312,7 +1314,7 @@ inline void LstmStepInteger(
 // Scalar values:
 //   quantized_cell_clip: quantized clip value for cell.
 //   quantized_proj_clip: quantized clip value for projection.
-//   cell_scale: the power of two scale for cell state.
+//   cell_state_scale: the power of two scale for cell state.
 //
 // Zero points:
 //   output_state_zp: zero point of output state.
@@ -1385,7 +1387,7 @@ void LstmStepInteger(
     const int32_t* intermediate_zp, int16_t quantized_cell_clip,
     int8_t quantized_proj_clip, int n_batch, int n_cell, int n_input,
     int n_output, int output_batch_leading_dim, int8_t* output_state_ptr,
-    int32_t output_state_zp, int16_t* cell_ptr, int8_t* output_ptr,
+    int32_t output_state_zp, int16_t* cell_state_ptr, int8_t* output_ptr,
     int8_t* scratch0, int8_t* scratch1, int16_t* scratch2, int16_t* scratch3,
     int16_t* scratch4, int16_t* scratch5, int16_t* scratch6,
     int16_t* scratch7) {
@@ -1486,20 +1488,21 @@ void LstmStepInteger(
                            input_gate_scratch);
 
   // New cell.
-  tensor_utils::CwiseMul(forget_gate_scratch, cell_ptr, n_batch, n_cell,
+  tensor_utils::CwiseMul(forget_gate_scratch, cell_state_ptr, n_batch, n_cell,
                          15 + 15 - 15, scratch6);
 
   tensor_utils::CwiseMul(input_gate_scratch, cell_gate_scratch, n_batch, n_cell,
                          15 + 15 - 15, scratch7);
 
-  tensor_utils::CwiseAdd(scratch6, scratch7, n_batch, n_cell, cell_ptr);
+  tensor_utils::CwiseAdd(scratch6, scratch7, n_batch, n_cell, cell_state_ptr);
 
   if (quantized_cell_clip > 0) {
-    tensor_utils::CwiseClipping(cell_ptr, quantized_cell_clip, n_batch, n_cell);
+    tensor_utils::CwiseClipping(cell_state_ptr, quantized_cell_clip, n_batch,
+                                n_cell);
   }
 
   // Cell to hidden.
-  tensor_utils::ApplyTanhFloat(cell_ptr, n_batch, n_cell, -15,
+  tensor_utils::ApplyTanhFloat(cell_state_ptr, n_batch, n_cell, -15,
                                forget_gate_scratch);
 
   std::vector hidden(n_batch * n_cell);

From 1bd014826e1a6113d8b6af47912c8e9c5c06a560 Mon Sep 17 00:00:00 2001
From: Jonah Kohn <51345541+jonah-kohn@users.noreply.github.com>
Date: Mon, 22 Jun 2020 11:29:18 -0700
Subject: [PATCH 443/867] Cast optimizer params as floats during serialization.

---
 tensorflow/python/keras/optimizer_v2/optimizer_v2.py | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/tensorflow/python/keras/optimizer_v2/optimizer_v2.py b/tensorflow/python/keras/optimizer_v2/optimizer_v2.py
index c55b332bfc0..d8992bbe3e0 100644
--- a/tensorflow/python/keras/optimizer_v2/optimizer_v2.py
+++ b/tensorflow/python/keras/optimizer_v2/optimizer_v2.py
@@ -910,7 +910,7 @@ class OptimizerV2(trackable.Trackable):
       return value()
     if tensor_util.is_tensor(value):
       return backend.get_value(value)
-    return value
+    return float(value)
 
   def variables(self):
     """Returns variables of this Optimizer based on the order created."""

From 7e59a2ab28df8f9cc50de9f064b61d23622af38a Mon Sep 17 00:00:00 2001
From: "A. Unique TensorFlower" 
Date: Mon, 22 Jun 2020 10:33:03 -0700
Subject: [PATCH 444/867] Qualify uses of std::string

PiperOrigin-RevId: 317683481
Change-Id: I3f6bb93d67e18623bf5d049ae0cc03bd2f937e5f
---
 .../convert_expanddims_to_reshape.cc          |  5 ++-
 .../convert_reorder_axes.cc                   |  7 ++--
 .../convert_trivial_pack_to_reshape.cc        |  2 +-
 .../convert_trivial_transpose_to_reshape.cc   |  5 ++-
 .../create_im2col_arrays.cc                   |  4 +-
 .../toco/graph_transformations/dequantize.cc  | 22 +++++-----
 .../ensure_bias_vectors.cc                    |  9 ++--
 ...int8_weights_safe_for_fast_int8_kernels.cc |  2 +-
 .../fuse_binary_into_preceding_affine.cc      |  2 +-
 .../graph_transformations.cc                  | 20 ++++-----
 .../graph_transformations.h                   | 14 +++----
 .../group_bidirectional_sequence_ops.cc       | 34 +++++++--------
 .../graph_transformations/hardcode_min_max.cc |  8 ++--
 .../identify_dilated_conv.cc                  |  2 +-
 .../identify_hardswish.cc                     |  2 +-
 .../graph_transformations/identify_lstm.cc    | 10 ++---
 .../identify_lstm_merge_inputs.cc             | 13 +++---
 .../identify_lstm_split_inputs.cc             |  2 +-
 .../toco/graph_transformations/lstm_utils.cc  | 15 +++----
 .../toco/graph_transformations/lstm_utils.h   | 15 +++----
 .../make_initial_dequantize_operator.cc       |  5 ++-
 .../merge_reshape_into_preceding_transpose.cc |  4 +-
 .../move_binary_operator_before_reshape.cc    |  2 +-
 ...gate_activation_function_into_constants.cc |  2 +-
 .../propagate_array_data_types.cc             |  4 +-
 .../propagate_default_min_max.cc              |  2 +-
 .../propagate_fixed_sizes.cc                  | 42 +++++++++----------
 .../quantization_util.cc                      |  4 +-
 .../graph_transformations/quantization_util.h |  2 +-
 .../toco/graph_transformations/quantize.cc    |  2 +-
 ...minmax_and_narrow_range_from_fake_quant.cc |  2 +-
 .../remove_successive_transpose.cc            |  4 +-
 .../remove_trivial_concatenation_input.cc     |  8 ++--
 .../remove_trivial_passthrough.cc             |  7 ++--
 ...emove_trivial_quantized_activation_func.cc |  4 +-
 .../remove_trivial_quantized_min_max.cc       |  4 +-
 .../graph_transformations/remove_unused_op.cc |  2 +-
 .../reorder_elementwise_unary.cc              |  6 +--
 .../reorder_reshape_transpose.cc              |  6 +--
 .../resolve_batch_normalization.cc            | 10 +++--
 .../resolve_constant_concatenation.cc         |  6 +--
 .../resolve_reorder_axes.cc                   | 14 +++----
 .../resolve_tensorflow_concat.cc              |  4 +-
 .../resolve_tensorflow_matmul.cc              |  6 +--
 .../resolve_tensorflow_switch.cc              |  2 +-
 .../shuffle_fc_weights.cc                     |  4 +-
 .../unfuse_activation_functions.cc            |  2 +-
 .../unpartition_embedding_lookup.cc           |  9 ++--
 .../unroll_batch_matmul.cc                    | 25 +++++------
 49 files changed, 200 insertions(+), 187 deletions(-)

diff --git a/tensorflow/lite/toco/graph_transformations/convert_expanddims_to_reshape.cc b/tensorflow/lite/toco/graph_transformations/convert_expanddims_to_reshape.cc
index 6b632fb5f46..6695f5a844b 100644
--- a/tensorflow/lite/toco/graph_transformations/convert_expanddims_to_reshape.cc
+++ b/tensorflow/lite/toco/graph_transformations/convert_expanddims_to_reshape.cc
@@ -76,8 +76,9 @@ namespace toco {
   reshape_op->outputs = expand_op->outputs;
 
   // Create a new input array
-  string axis_array_name = expand_op->inputs[1];
-  string shape_array_name = toco::AvailableArrayName(*model, axis_array_name);
+  std::string axis_array_name = expand_op->inputs[1];
+  std::string shape_array_name =
+      toco::AvailableArrayName(*model, axis_array_name);
   Array& shape_array = model->GetOrCreateArray(shape_array_name);
   *(shape_array.mutable_shape()->mutable_dims()) = {
       1, static_cast(reshape_dims.size())};
diff --git a/tensorflow/lite/toco/graph_transformations/convert_reorder_axes.cc b/tensorflow/lite/toco/graph_transformations/convert_reorder_axes.cc
index 1440cd1c1a7..c2214d6f97c 100644
--- a/tensorflow/lite/toco/graph_transformations/convert_reorder_axes.cc
+++ b/tensorflow/lite/toco/graph_transformations/convert_reorder_axes.cc
@@ -41,7 +41,7 @@ TensorFlowReshapeOperator* CreateReshapeFromReorderAxes(
                                    input_shape.dims(3) * input_shape.dims(2)};
 
   // Create a new input array for Reshape.
-  string reshape_array_name =
+  std::string reshape_array_name =
       AvailableArrayName(*model, reshape_op->outputs[0]);
   reshape_op->inputs.push_back(reshape_array_name);
 
@@ -71,7 +71,8 @@ TransposeOperator* CreateTransposeFromReorderAxes(
   GetShuffleShape(input_axes_order, output_axes_order, &permutations_data);
 
   // Create a new input permutations array for Transpose.
-  string perm_array_name = AvailableArrayName(*model, transpose_op->outputs[0]);
+  std::string perm_array_name =
+      AvailableArrayName(*model, transpose_op->outputs[0]);
   transpose_op->inputs.push_back(perm_array_name);
 
   Array& perm_array = model->GetOrCreateArray(perm_array_name);
@@ -104,7 +105,7 @@ TransposeOperator* CreateTransposeFromReorderAxes(
 
   // Get input array. If kFakeQuant is the input into ReorderAxes, get the input
   // array passed into kFakeQuant. kFakeQuant op is dropped when possible.
-  string constant_input_array_name = input_array_name;
+  std::string constant_input_array_name = input_array_name;
   if (!input_array.buffer) {
     const auto* op_producing_input = GetOpWithOutput(*model, input_array_name);
     if (op_producing_input &&
diff --git a/tensorflow/lite/toco/graph_transformations/convert_trivial_pack_to_reshape.cc b/tensorflow/lite/toco/graph_transformations/convert_trivial_pack_to_reshape.cc
index 31f9ced8cf5..2ad6280b955 100644
--- a/tensorflow/lite/toco/graph_transformations/convert_trivial_pack_to_reshape.cc
+++ b/tensorflow/lite/toco/graph_transformations/convert_trivial_pack_to_reshape.cc
@@ -59,7 +59,7 @@ namespace toco {
   reshape_op->outputs = pack_op->outputs;
 
   // Create shape param.
-  string shape_array_name =
+  std::string shape_array_name =
       AvailableArrayName(*model, pack_op->outputs[0] + "_shape");
   Array& shape_array = model->GetOrCreateArray(shape_array_name);
   const int shape_array_dims = 1 + input_array.shape().dimensions_count();
diff --git a/tensorflow/lite/toco/graph_transformations/convert_trivial_transpose_to_reshape.cc b/tensorflow/lite/toco/graph_transformations/convert_trivial_transpose_to_reshape.cc
index 2b5aaea2b23..2d1e5090f4a 100644
--- a/tensorflow/lite/toco/graph_transformations/convert_trivial_transpose_to_reshape.cc
+++ b/tensorflow/lite/toco/graph_transformations/convert_trivial_transpose_to_reshape.cc
@@ -90,8 +90,9 @@ bool TransposeAffectsMemoryOrder(std::vector perm,
   reshape_op->outputs = transpose_op->outputs;
 
   // Create a new input array for the shape input
-  string perm_array_name = transpose_op->inputs[1];
-  string shape_array_name = toco::AvailableArrayName(*model, perm_array_name);
+  std::string perm_array_name = transpose_op->inputs[1];
+  std::string shape_array_name =
+      toco::AvailableArrayName(*model, perm_array_name);
   Array& shape_array = model->GetOrCreateArray(shape_array_name);
   *(shape_array.mutable_shape()->mutable_dims()) = {
       1, static_cast(output_dims.size())};
diff --git a/tensorflow/lite/toco/graph_transformations/create_im2col_arrays.cc b/tensorflow/lite/toco/graph_transformations/create_im2col_arrays.cc
index 8e93bc23789..a788b81672d 100644
--- a/tensorflow/lite/toco/graph_transformations/create_im2col_arrays.cc
+++ b/tensorflow/lite/toco/graph_transformations/create_im2col_arrays.cc
@@ -49,7 +49,7 @@ bool ProcessConvOperator(Model* model, ConvOperator* op) {
 
   // Create the im2col array.
   CHECK_EQ(op->outputs.size(), 1);
-  const string& im2col_array_name =
+  const std::string& im2col_array_name =
       AvailableArrayName(*model, op->inputs[0] + "_im2col");
   model->GetOrCreateArray(im2col_array_name);
   op->outputs.push_back(im2col_array_name);
@@ -65,7 +65,7 @@ bool ProcessTransposeConvOperator(Model* model, TransposeConvOperator* op) {
 
   // Always create an im2col array for transpose_conv.
   CHECK_EQ(op->outputs.size(), 1);
-  const string& im2col_array_name = AvailableArrayName(
+  const std::string& im2col_array_name = AvailableArrayName(
       *model, op->inputs[TransposeConvOperator::DATA_INPUT] + "_im2col");
   model->GetOrCreateArray(im2col_array_name);
   op->outputs.push_back(im2col_array_name);
diff --git a/tensorflow/lite/toco/graph_transformations/dequantize.cc b/tensorflow/lite/toco/graph_transformations/dequantize.cc
index cc5dddbb40e..6872fc47344 100644
--- a/tensorflow/lite/toco/graph_transformations/dequantize.cc
+++ b/tensorflow/lite/toco/graph_transformations/dequantize.cc
@@ -41,7 +41,7 @@ void DequantizeBuffer(Array* array) {
 }
 
 std::vector>::iterator FindFirstOpWithInput(
-    Model* model, const string& array_name) {
+    Model* model, const std::string& array_name) {
   for (auto it = model->operators.begin(); it != model->operators.end(); ++it) {
     for (const auto& input : it->get()->inputs) {
       if (input == array_name) {
@@ -52,7 +52,7 @@ std::vector>::iterator FindFirstOpWithInput(
   return model->operators.end();
 }
 
-void ClearArrayQuantizationParams(const string& array_name, Model* model) {
+void ClearArrayQuantizationParams(const std::string& array_name, Model* model) {
   auto* array = &model->GetArray(array_name);
   CHECK(array->quantization_params);
   for (auto& input_array : *model->flags.mutable_input_arrays()) {
@@ -75,7 +75,7 @@ void ClearArrayQuantizationParams(const string& array_name, Model* model) {
   array->quantization_params = nullptr;
 }
 
-bool DequantizeArray(const string& array_name,
+bool DequantizeArray(const std::string& array_name,
                      GraphTransformation* transformation, Model* model) {
   auto* array = &model->GetArray(array_name);
   if (!array->quantization_params) {
@@ -133,7 +133,7 @@ bool DequantizeArray(const string& array_name,
   if (IsInputArray(*model, array_name)) {
     must_insert_fakequant_after = true;
   }
-  for (const string& output_array : model->flags.output_arrays()) {
+  for (const std::string& output_array : model->flags.output_arrays()) {
     if (array_name == output_array) {
       must_insert_fakequant_before = true;
     }
@@ -152,7 +152,7 @@ bool DequantizeArray(const string& array_name,
   auto* fakequant_op = new FakeQuantOperator;
   model->operators.emplace(FindFirstOpWithInput(model, array_name),
                            fakequant_op);
-  const string& new_array_name = AvailableArrayName(*model, array_name);
+  const std::string& new_array_name = AvailableArrayName(*model, array_name);
   auto& new_array = model->GetOrCreateArray(new_array_name);
   new_array.data_type = ArrayDataType::kFloat;
   new_array.copy_shape(array->shape());
@@ -162,7 +162,7 @@ bool DequantizeArray(const string& array_name,
   fakequant_op->narrow_range = array->narrow_range;
   if (must_insert_fakequant_before) {
     for (const auto& op : model->operators) {
-      for (string& output : op->outputs) {
+      for (std::string& output : op->outputs) {
         if (output == array_name) {
           output = new_array_name;
         }
@@ -172,7 +172,7 @@ bool DequantizeArray(const string& array_name,
     fakequant_op->outputs = {array_name};
   } else {
     for (const auto& op : model->operators) {
-      for (string& input : op->inputs) {
+      for (std::string& input : op->inputs) {
         if (input == array_name) {
           input = new_array_name;
         }
@@ -209,15 +209,15 @@ bool DequantizeArray(const string& array_name,
     return ::tensorflow::Status::OK();
   }
 
-  std::vector arrays;
-  for (const string& input : op->inputs) {
+  std::vector arrays;
+  for (const std::string& input : op->inputs) {
     arrays.push_back(input);
   }
-  for (const string& output : op->outputs) {
+  for (const std::string& output : op->outputs) {
     arrays.push_back(output);
   }
   bool changed = false;
-  for (const string& array : arrays) {
+  for (const std::string& array : arrays) {
     if (!model->IsOptionalArray(array)) {
       changed |= DequantizeArray(array, this, model);
     }
diff --git a/tensorflow/lite/toco/graph_transformations/ensure_bias_vectors.cc b/tensorflow/lite/toco/graph_transformations/ensure_bias_vectors.cc
index fcad8bc0086..09bd9aedcf0 100644
--- a/tensorflow/lite/toco/graph_transformations/ensure_bias_vectors.cc
+++ b/tensorflow/lite/toco/graph_transformations/ensure_bias_vectors.cc
@@ -27,7 +27,7 @@ namespace toco {
 namespace {
 
 int GetOutputDepthFromWeights(const Model& model, const Operator& op) {
-  const string& weights_name = op.inputs[1];
+  const std::string& weights_name = op.inputs[1];
   const auto& weights_shape = model.GetArray(weights_name).shape();
   if (op.type == OperatorType::kConv ||
       op.type == OperatorType::kFullyConnected ||
@@ -56,13 +56,14 @@ bool ProcessLinearOperator(Model* model, Operator* op) {
   if (CheckOpInputSize(*op)) {
     return false;
   }
-  const string& output_name = op->outputs[0];
-  const string& weights_name = op->inputs[1];
+  const std::string& output_name = op->outputs[0];
+  const std::string& weights_name = op->inputs[1];
   if (!model->GetArray(weights_name).has_shape()) {
     return false;
   }
   const int depth = GetOutputDepthFromWeights(*model, *op);
-  const string& bias_name = AvailableArrayName(*model, output_name + "_bias");
+  const std::string& bias_name =
+      AvailableArrayName(*model, output_name + "_bias");
   op->inputs.push_back(bias_name);
   auto& bias_array = model->GetOrCreateArray(bias_name);
   bias_array.data_type = ArrayDataType::kFloat;
diff --git a/tensorflow/lite/toco/graph_transformations/ensure_uint8_weights_safe_for_fast_int8_kernels.cc b/tensorflow/lite/toco/graph_transformations/ensure_uint8_weights_safe_for_fast_int8_kernels.cc
index 918bb489995..5854e74b507 100644
--- a/tensorflow/lite/toco/graph_transformations/ensure_uint8_weights_safe_for_fast_int8_kernels.cc
+++ b/tensorflow/lite/toco/graph_transformations/ensure_uint8_weights_safe_for_fast_int8_kernels.cc
@@ -152,7 +152,7 @@ namespace toco {
       return ::tensorflow::Status::OK();
   }
 
-  const string& name = op.inputs[weights_index];
+  const std::string& name = op.inputs[weights_index];
   auto& array = model->GetArray(name);
   if (!array.buffer) {
     return ::tensorflow::Status::OK();
diff --git a/tensorflow/lite/toco/graph_transformations/fuse_binary_into_preceding_affine.cc b/tensorflow/lite/toco/graph_transformations/fuse_binary_into_preceding_affine.cc
index 05a2fecf31d..1de12b4f959 100644
--- a/tensorflow/lite/toco/graph_transformations/fuse_binary_into_preceding_affine.cc
+++ b/tensorflow/lite/toco/graph_transformations/fuse_binary_into_preceding_affine.cc
@@ -260,7 +260,7 @@ void FuseMulOrDivParamsIntoPrecedingAffine(Model* model, Operator* preceding_op,
     return ::tensorflow::Status::OK();
   }
 
-  for (const string& output_array : model->flags.output_arrays()) {
+  for (const std::string& output_array : model->flags.output_arrays()) {
     if (preceding_op->outputs[0] == output_array) {
       return ::tensorflow::Status::OK();
     }
diff --git a/tensorflow/lite/toco/graph_transformations/graph_transformations.cc b/tensorflow/lite/toco/graph_transformations/graph_transformations.cc
index e4eb7698597..bee666531a7 100644
--- a/tensorflow/lite/toco/graph_transformations/graph_transformations.cc
+++ b/tensorflow/lite/toco/graph_transformations/graph_transformations.cc
@@ -29,7 +29,7 @@ namespace toco {
 
 namespace {
 
-void PrintModelStats(const string& label, const Model& model) {
+void PrintModelStats(const std::string& label, const Model& model) {
   int quantized_arrays = 0;
   for (const auto& array : model.GetArrayMap()) {
     if (array.second->quantization_params) {
@@ -57,8 +57,8 @@ void PrintModelStats(const string& label, const Model& model) {
 void DiscardUselessConnectedComponentsAndRNNBackEdges(Model* model) {
   // Identify the set of arrays that are in 'useful' connected components
   // of the graph, which means connected to output arrays.
-  std::unordered_set useful_arrays;
-  for (const string& output_array : model->flags.output_arrays()) {
+  std::unordered_set useful_arrays;
+  for (const std::string& output_array : model->flags.output_arrays()) {
     useful_arrays.insert(output_array);
   }
   bool found_new_useful_arrays;
@@ -66,15 +66,15 @@ void DiscardUselessConnectedComponentsAndRNNBackEdges(Model* model) {
     found_new_useful_arrays = false;
     for (const auto& op : model->operators) {
       bool op_touches_useful_arrays = false;
-      for (const string& output : op->outputs) {
+      for (const std::string& output : op->outputs) {
         op_touches_useful_arrays |= useful_arrays.count(output);
       }
       if (op_touches_useful_arrays) {
-        for (const string& input : op->inputs) {
+        for (const std::string& input : op->inputs) {
           found_new_useful_arrays |= !useful_arrays.count(input);
           useful_arrays.insert(input);
         }
-        for (const string& output : op->outputs) {
+        for (const std::string& output : op->outputs) {
           found_new_useful_arrays |= !useful_arrays.count(output);
           useful_arrays.insert(output);
         }
@@ -91,7 +91,7 @@ void DiscardUselessConnectedComponentsAndRNNBackEdges(Model* model) {
     }
   } while (found_new_useful_arrays);
   // Erase arrays that aren't useful, and that are discardable.
-  model->EraseArrays([&](const string& name) {
+  model->EraseArrays([&](const std::string& name) {
     return (!useful_arrays.count(name) && IsDiscardableArray(*model, name));
   });
   // Erase operators that do not produce a useful output array.
@@ -101,7 +101,7 @@ void DiscardUselessConnectedComponentsAndRNNBackEdges(Model* model) {
     if (useful_arrays.count((*it)->outputs[0])) {
       ++it;
     } else {
-      for (const string& output : (*it)->outputs) {
+      for (const std::string& output : (*it)->outputs) {
         CHECK(!useful_arrays.count(output));
       }
       it = model->operators.erase(it);
@@ -156,7 +156,7 @@ bool GraphTransformationsPass(int increment, Model* model,
                         << " at op_index=" << op_index << "/"
                         << model->operators.size() - 1;
       }
-      for (const string& message : transformation->Messages()) {
+      for (const std::string& message : transformation->Messages()) {
         VLOG(log_level) << transformation->Name() << " " << made_a_change_msg
                         << " at op_index=" << op_index << "/"
                         << model->operators.size() - 1 << ": " << message;
@@ -191,7 +191,7 @@ bool GraphTransformationsPass(int increment, Model* model,
 }  // namespace
 
 tensorflow::Status RunGraphTransformationsWithStatus(
-    Model* model, const string& msg,
+    Model* model, const std::string& msg,
     const GraphTransformationsSet& transformations) {
   PrintModelStats(toco::port::StringF("Before %s", msg), *model);
   int pass_index = 0;
diff --git a/tensorflow/lite/toco/graph_transformations/graph_transformations.h b/tensorflow/lite/toco/graph_transformations/graph_transformations.h
index 07b9fd4c5cf..4d7278fcaf9 100644
--- a/tensorflow/lite/toco/graph_transformations/graph_transformations.h
+++ b/tensorflow/lite/toco/graph_transformations/graph_transformations.h
@@ -33,7 +33,7 @@ class GraphTransformation {
   virtual ~GraphTransformation() {}
   // Returns the list of messages that this graph transformation
   // generated since ClearMessages() was called.
-  const std::vector& Messages() const { return messages_; }
+  const std::vector& Messages() const { return messages_; }
   // Clears the list of messages; should be called after every
   // run of this graph transformation.
   void ClearMessages() { return messages_.clear(); }
@@ -48,7 +48,7 @@ class GraphTransformation {
   GraphTransformation() {}
 
   // List of messages generated by this graph transformation.
-  std::vector messages_;
+  std::vector messages_;
 
  private:
   GraphTransformation(const GraphTransformation& other) = delete;
@@ -74,7 +74,7 @@ class GraphTransformationsSet {
     }
   }
   void Add(GraphTransformation* transformation) {
-    const string& name = transformation->Name();
+    const std::string& name = transformation->Name();
     CHECK(!names_.count(name));
     names_.insert(name);
     transformations_.emplace_back(transformation);
@@ -92,7 +92,7 @@ class GraphTransformationsSet {
   GraphTransformationsSet(const GraphTransformationsSet&& other) = delete;
   std::vector> transformations_;
   // Names of transformations in the set. Only used to guard against dupes.
-  std::unordered_set names_;
+  std::unordered_set names_;
 };
 
 // Run the given list of graph transformations on the model.
@@ -103,11 +103,11 @@ class GraphTransformationsSet {
 // the resulting raw pointers, and this RunGraphTransformations
 // takes care of delete'ing these pointers.
 tensorflow::Status RunGraphTransformationsWithStatus(
-    Model* model, const string& msg,
+    Model* model, const std::string& msg,
     const GraphTransformationsSet& transformations);
 
 inline void RunGraphTransformations(
-    Model* model, const string& msg,
+    Model* model, const std::string& msg,
     const GraphTransformationsSet& transformations) {
   auto s = RunGraphTransformationsWithStatus(model, msg, transformations);
   CHECK(s.ok()) << s.error_message();
@@ -232,7 +232,7 @@ class PropagateDefaultMinMax : public GraphTransformation {
   }
 
  private:
-  bool SetArrayMinMax(const string& array_name, Array* array);
+  bool SetArrayMinMax(const std::string& array_name, Array* array);
   std::vector> type_ranges_;
 };
 
diff --git a/tensorflow/lite/toco/graph_transformations/group_bidirectional_sequence_ops.cc b/tensorflow/lite/toco/graph_transformations/group_bidirectional_sequence_ops.cc
index fa252b1a61b..4f9caeb77b0 100644
--- a/tensorflow/lite/toco/graph_transformations/group_bidirectional_sequence_ops.cc
+++ b/tensorflow/lite/toco/graph_transformations/group_bidirectional_sequence_ops.cc
@@ -197,7 +197,7 @@ void ConstructBidirectionalSequenceOp(
   constexpr int kBwInputActivationStartIndex = 37;
   constexpr int kAuxInputStartIndex = 39;
   (*bi_op)->inputs.reserve(kBidirectionalSequenceLstmInputsCount);
-  const string& input_array_name =
+  const std::string& input_array_name =
       AvailableArrayName(*model, "bidirectional_sequence_lstm_input_0");
   model->GetOrCreateArray(input_array_name);
   // The input will be changed later.
@@ -232,7 +232,7 @@ void ConstructBidirectionalSequenceOp(
 
   // TODO(renjieliu): Deal with Auxiliary input and weights for 39 - 47.
   for (; i <= kBidirectionalSequenceLstmInputsCount; ++i) {
-    const string& temp_array_name = AvailableArrayName(
+    const std::string& temp_array_name = AvailableArrayName(
         *model, "bidirectional_sequence_lstm_temp_" + std::to_string(i));
     model->CreateOptionalArray(temp_array_name);
     (*bi_op)->inputs.push_back(temp_array_name);
@@ -240,9 +240,9 @@ void ConstructBidirectionalSequenceOp(
 
   // Deal with outputs.
   (*bi_op)->outputs.reserve(2);
-  const string& fw_output_array_name =
+  const std::string& fw_output_array_name =
       AvailableArrayName(*model, "bidirectional_sequence_lstm_fw_output_0");
-  const string& bw_output_array_name =
+  const std::string& bw_output_array_name =
       AvailableArrayName(*model, "bidirectional_sequence_lstm_bw_output_0");
   model->GetOrCreateArray(fw_output_array_name);
   model->GetOrCreateArray(bw_output_array_name);
@@ -260,7 +260,7 @@ void ConstructBidirectionalSequenceOp(
   constexpr int kBwInputsStartIndex = 5;
   constexpr int kAuxInputsStartIndex = 9;
   (*bi_op)->inputs.reserve(kBidirectionalSequenceRnnInputsCount);
-  const string& input_array_name =
+  const std::string& input_array_name =
       AvailableArrayName(*model, "bidirectional_sequence_rnn_input_0");
   model->GetOrCreateArray(input_array_name);
   // The input will be changed later.
@@ -280,7 +280,7 @@ void ConstructBidirectionalSequenceOp(
 
   // TODO(renjieliu): Deal with optional weights.
   for (; i < kBidirectionalSequenceRnnInputsCount; ++i) {
-    const string& temp_array_name = AvailableArrayName(
+    const std::string& temp_array_name = AvailableArrayName(
         *model, "bidirectional_sequence_rnn_temp_" + std::to_string(i));
     model->CreateOptionalArray(temp_array_name);
     (*bi_op)->inputs.push_back(temp_array_name);
@@ -288,9 +288,9 @@ void ConstructBidirectionalSequenceOp(
 
   // Deal with outputs.
   (*bi_op)->outputs.reserve(2);
-  const string& fw_output_array_name =
+  const std::string& fw_output_array_name =
       AvailableArrayName(*model, "bidirectional_sequence_rnn_fw_output_0");
-  const string& bw_output_array_name =
+  const std::string& bw_output_array_name =
       AvailableArrayName(*model, "bidirectional_sequence_rnn_bw_output_0");
   model->GetOrCreateArray(fw_output_array_name);
   model->GetOrCreateArray(bw_output_array_name);
@@ -318,7 +318,7 @@ void GroupFwBwSequenceOps(Model* model, std::stack fw_sequence_ops,
 
 template 
 void RewireBidirectionalSequenceSequenceOpsConnections(
-    OperatorType operator_type, const string& input_array_name,
+    OperatorType operator_type, const std::string& input_array_name,
     const std::vector& bidirectional_sequence_ops,
     std::vector>::iterator* op_it, Model* model) {
   int aux_input_index = -1;
@@ -333,8 +333,8 @@ void RewireBidirectionalSequenceSequenceOpsConnections(
       // Should not reach here.
       DCHECK(false);
   }
-  string cur_fw_input = input_array_name;
-  string cur_bw_input = input_array_name;
+  std::string cur_fw_input = input_array_name;
+  std::string cur_bw_input = input_array_name;
   for (size_t i = 0; i < bidirectional_sequence_ops.size(); ++i) {
     DeleteArrayIfUnusedOutsideOfOp(bidirectional_sequence_ops[i]->inputs[0],
                                    bidirectional_sequence_ops[i], model);
@@ -371,8 +371,8 @@ void RewireFinalUnpackOutputs(const UnpackOperator& original_unpack_operator,
   (*final_unpack_operator)->num = original_unpack_operator.num;
 
   for (size_t i = 0; i < original_unpack_operator.outputs.size(); ++i) {
-    const string& output_array_name = original_unpack_operator.outputs[i];
-    const string& final_unpack_output_array_name = AvailableArrayName(
+    const std::string& output_array_name = original_unpack_operator.outputs[i];
+    const std::string& final_unpack_output_array_name = AvailableArrayName(
         *model, "bidirectional_sequence_unpack_" + std::to_string(i));
     model->GetOrCreateArray(final_unpack_output_array_name);
     (*final_unpack_operator)->outputs.push_back(final_unpack_output_array_name);
@@ -381,7 +381,7 @@ void RewireFinalUnpackOutputs(const UnpackOperator& original_unpack_operator,
       // If there's a following op after the unpack, it must be a concat op.
       DCHECK(unpack_following_op->type == OperatorType::kConcatenation);
       // For every output of the concat, rewire the outputs.
-      for (const string& concat_output : unpack_following_op->outputs) {
+      for (const std::string& concat_output : unpack_following_op->outputs) {
         (*final_unpack_operator)->outputs[i] = concat_output;
       }
       // Remove the concat op.
@@ -454,7 +454,7 @@ template 
                        &bidirectional_sequence_ops);
 
   // Rewire the inputs & outputs.
-  string current_input = first_fw_sequence_input->outputs[0];
+  std::string current_input = first_fw_sequence_input->outputs[0];
   RewireBidirectionalSequenceSequenceOpsConnections(
       operator_type, current_input, bidirectional_sequence_ops, &op_it, model);
 
@@ -525,7 +525,7 @@ template 
                        &bidirectional_sequence_lstm_ops);
 
   // Rewire the inputs & outputs.
-  string current_input = first_fw_lstm_input->outputs[0];
+  std::string current_input = first_fw_lstm_input->outputs[0];
   RewireBidirectionalSequenceSequenceOpsConnections(
       OperatorType::kBidirectionalSequenceLstm, current_input,
       bidirectional_sequence_lstm_ops, &op_it, model);
@@ -601,7 +601,7 @@ template 
                        &bidirectional_sequence_rnn_ops);
 
   // Rewire the inputs & outputs.
-  string current_input = first_fw_rnn_input->outputs[0];
+  std::string current_input = first_fw_rnn_input->outputs[0];
   RewireBidirectionalSequenceSequenceOpsConnections(
       OperatorType::kBidirectionalSequenceRnn, current_input,
       bidirectional_sequence_rnn_ops, &op_it, model);
diff --git a/tensorflow/lite/toco/graph_transformations/hardcode_min_max.cc b/tensorflow/lite/toco/graph_transformations/hardcode_min_max.cc
index 171d522daa7..c065e32c4df 100644
--- a/tensorflow/lite/toco/graph_transformations/hardcode_min_max.cc
+++ b/tensorflow/lite/toco/graph_transformations/hardcode_min_max.cc
@@ -279,10 +279,10 @@ bool MinMaxApproximatelyEqual(const MinMax& minmax1, const MinMax& minmax2) {
 // If multiple of these arrays have MinMax, then these are required
 // to agree with each other.
 bool PropagateMinMaxAmongArrays(Model* model,
-                                const std::vector& array_names) {
-  string reference_array_name;
+                                const std::vector& array_names) {
+  std::string reference_array_name;
   MinMax* reference_minmax = nullptr;
-  for (const string& array_name : array_names) {
+  for (const std::string& array_name : array_names) {
     if (model->GetArray(array_name).minmax) {
       reference_array_name = array_name;
       reference_minmax = model->GetArray(array_name).minmax.get();
@@ -294,7 +294,7 @@ bool PropagateMinMaxAmongArrays(Model* model,
     return false;
   }
   bool changed = false;
-  for (const string& array_name : array_names) {
+  for (const std::string& array_name : array_names) {
     auto& array = model->GetArray(array_name);
     if (array.minmax) {
       CHECK(MinMaxApproximatelyEqual(*array.minmax, *reference_minmax))
diff --git a/tensorflow/lite/toco/graph_transformations/identify_dilated_conv.cc b/tensorflow/lite/toco/graph_transformations/identify_dilated_conv.cc
index 1940068d32a..2ba39e74d3a 100644
--- a/tensorflow/lite/toco/graph_transformations/identify_dilated_conv.cc
+++ b/tensorflow/lite/toco/graph_transformations/identify_dilated_conv.cc
@@ -206,7 +206,7 @@ bool ResolveDilatedConv(Model* model, Operator* conv_base_op, Operator* stb_op,
   }
 
   // Conv Op
-  const string& input_of_conv_op =
+  const std::string& input_of_conv_op =
       has_expand_op ? post_stb_op->outputs[0] : stb_op->outputs[0];
   auto* conv_base_op = GetOpWithInput(*model, input_of_conv_op);
   bool changed = false;
diff --git a/tensorflow/lite/toco/graph_transformations/identify_hardswish.cc b/tensorflow/lite/toco/graph_transformations/identify_hardswish.cc
index 00758a22177..5bc49899e0b 100644
--- a/tensorflow/lite/toco/graph_transformations/identify_hardswish.cc
+++ b/tensorflow/lite/toco/graph_transformations/identify_hardswish.cc
@@ -78,7 +78,7 @@ using util::IsBinaryOp;
   // 1. non-constant input of add_with_relu6_op
   // 2. 1/6
   // 3. (and add_with_relu6_op[0].outputs[0] - which we already know!)
-  std::vector mul_inputs = mul_op->inputs;
+  std::vector mul_inputs = mul_op->inputs;
   mul_inputs.insert(mul_inputs.end(), output_op->inputs.begin(),
                     output_op->inputs.end());
 
diff --git a/tensorflow/lite/toco/graph_transformations/identify_lstm.cc b/tensorflow/lite/toco/graph_transformations/identify_lstm.cc
index 43ce90a0444..14f81779147 100644
--- a/tensorflow/lite/toco/graph_transformations/identify_lstm.cc
+++ b/tensorflow/lite/toco/graph_transformations/identify_lstm.cc
@@ -35,7 +35,7 @@ std::vector>::iterator FindOperator(
   return it;
 }
 
-bool ValidateSourceOp(const Model& model, const string& array_name,
+bool ValidateSourceOp(const Model& model, const std::string& array_name,
                       OperatorType op_type, Operator** source_op) {
   if (op_type == OperatorType::kNone) {
     CHECK(!source_op);
@@ -184,7 +184,7 @@ bool MatchOperatorInputs(const Operator& op, const Model& model,
                            &state_remember_mul)) {
     return ::tensorflow::Status::OK();
   }
-  const string prev_state = state_forget_mul->inputs[0];
+  const std::string prev_state = state_forget_mul->inputs[0];
 
   // State forget gate
   Operator* state_forget_sig;
@@ -271,16 +271,16 @@ bool MatchOperatorInputs(const Operator& op, const Model& model,
               LogName(*lstm_cell_op));
 
   // Create temp arrays used internally during runtime.
-  const string base_name(FindLongestCommonPrefix(
+  const std::string base_name(FindLongestCommonPrefix(
       lstm_cell_op->outputs[LstmCellOperator::STATE_OUTPUT],
       lstm_cell_op->outputs[LstmCellOperator::ACTIV_OUTPUT]));
-  const string& concat_temp_array_name =
+  const std::string& concat_temp_array_name =
       AvailableArrayName(*model, base_name + "concat_temp");
   auto& concat_temp_array = model->GetOrCreateArray(concat_temp_array_name);
   concat_temp_array.data_type =
       model->GetArray(concat_inputs->outputs[0]).data_type;
   lstm_cell_op->outputs[LstmCellOperator::CONCAT_TEMP] = concat_temp_array_name;
-  const string& activ_temp_array_name =
+  const std::string& activ_temp_array_name =
       AvailableArrayName(*model, base_name + "activ_temp");
   auto& activ_temp_array = model->GetOrCreateArray(activ_temp_array_name);
   activ_temp_array.data_type =
diff --git a/tensorflow/lite/toco/graph_transformations/identify_lstm_merge_inputs.cc b/tensorflow/lite/toco/graph_transformations/identify_lstm_merge_inputs.cc
index 2ac1d380813..cfa5f879f44 100644
--- a/tensorflow/lite/toco/graph_transformations/identify_lstm_merge_inputs.cc
+++ b/tensorflow/lite/toco/graph_transformations/identify_lstm_merge_inputs.cc
@@ -45,12 +45,12 @@ namespace toco {
 
   // Identify prev_activ_input, prev_state_input as required Op inputs,
   // using the rnn_states in the model flag.
-  string prev_activ_input;
+  std::string prev_activ_input;
   if (!GetMatchingRnnArray(model, src_op->outputs[kOutputTensor],
                            &prev_activ_input)) {
     return ::tensorflow::Status::OK();
   }
-  string prev_state_input;
+  std::string prev_state_input;
   if (!GetMatchingRnnArray(model, src_op->outputs[kCellStateTensor],
                            &prev_state_input)) {
     return ::tensorflow::Status::OK();
@@ -72,9 +72,10 @@ namespace toco {
   CHECK_EQ(num_cell, num_output);
 
   // Create tensorflow_graphdef style's one big weight tensor.
-  const string base_name(FindLongestCommonPrefix(
+  const std::string base_name(FindLongestCommonPrefix(
       src_op->outputs[kOutputTensor], src_op->outputs[kCellStateTensor]));
-  string merged_weights = AvailableArrayName(*model, base_name + "weights");
+  std::string merged_weights =
+      AvailableArrayName(*model, base_name + "weights");
   auto& array = model->GetOrCreateArray(merged_weights);
   array.data_type = ArrayDataType::kFloat;
   int weights_dim1 = 4 * num_cell;
@@ -117,7 +118,7 @@ namespace toco {
       num_cell * 3, num_input);
 
   // Create tensorflow_graphdef style's one big bias tensor.
-  string merged_biases = AvailableArrayName(*model, base_name + "biases");
+  std::string merged_biases = AvailableArrayName(*model, base_name + "biases");
   auto& bias_array = model->GetOrCreateArray(merged_biases);
   bias_array.data_type = ArrayDataType::kFloat;
   bias_array.copy_shape(Shape({weights_dim1}));
@@ -160,7 +161,7 @@ namespace toco {
   lstm_cell_op->outputs[LstmCellOperator::ACTIV_TEMP] =
       src_op->outputs[kOutputStateTensor];
   // Create a new temp array for the fourth output.
-  const string& concat_temp_array_name =
+  const std::string& concat_temp_array_name =
       AvailableArrayName(*model, base_name + "concat_temp");
   model->GetOrCreateArray(concat_temp_array_name);
   lstm_cell_op->outputs[LstmCellOperator::CONCAT_TEMP] = concat_temp_array_name;
diff --git a/tensorflow/lite/toco/graph_transformations/identify_lstm_split_inputs.cc b/tensorflow/lite/toco/graph_transformations/identify_lstm_split_inputs.cc
index 62f4124fb4e..8359534435a 100644
--- a/tensorflow/lite/toco/graph_transformations/identify_lstm_split_inputs.cc
+++ b/tensorflow/lite/toco/graph_transformations/identify_lstm_split_inputs.cc
@@ -86,7 +86,7 @@ namespace toco {
   // Get original weight tensor and decompose 1 tensor to 8 sub tensors.
   Array& kernel =
       model->GetArray(curr_op->inputs[LstmCellOperator::WEIGHTS_INPUT]);
-  const string base_name(FindLongestCommonPrefix(
+  const std::string base_name(FindLongestCommonPrefix(
       curr_op->outputs[LstmCellOperator::ACTIV_OUTPUT],
       curr_op->outputs[LstmCellOperator::STATE_OUTPUT]));
 
diff --git a/tensorflow/lite/toco/graph_transformations/lstm_utils.cc b/tensorflow/lite/toco/graph_transformations/lstm_utils.cc
index 3414a7fd7fe..a8b69205c0d 100644
--- a/tensorflow/lite/toco/graph_transformations/lstm_utils.cc
+++ b/tensorflow/lite/toco/graph_transformations/lstm_utils.cc
@@ -16,8 +16,8 @@ limitations under the License.
 
 namespace toco {
 
-void CreateOptionalArray(Model* model, string* input_array_buffer,
-                         const string& array_name) {
+void CreateOptionalArray(Model* model, std::string* input_array_buffer,
+                         const std::string& array_name) {
   *input_array_buffer = array_name;
   model->CreateOptionalArray(array_name);
 }
@@ -39,7 +39,7 @@ void CopyArrayData(const Buffer& src_buffer,
 }
 
 Buffer* CreateFloatArrayBuffer(Model* model,
-                                                      string* array_name,
+                                                      std::string* array_name,
                                                       const Shape& shape) {
   *array_name = AvailableArrayName(*model, *array_name);
   auto& array = model->GetOrCreateArray(*array_name);
@@ -51,8 +51,8 @@ Buffer* CreateFloatArrayBuffer(Model* model,
   return buffer;
 }
 
-void CopySubArrayToArray(Model* model, string* array_name,
-                         const string& tensor_name, int dim1_size,
+void CopySubArrayToArray(Model* model, std::string* array_name,
+                         const std::string& tensor_name, int dim1_size,
                          int dim2_size, const Array& original_array,
                          int start_idx1, int start_idx2) {
   // Determine whether it's bias or not, create shape, buffer.
@@ -83,8 +83,9 @@ void CopyArrayToSubArray(Buffer& tensor_buffer,
                 dim1_copy_size, dim2_copy_size);
 }
 
-bool GetMatchingRnnArray(Model* model, const string& back_edge_source_array,
-                         string* rnn_array) {
+bool GetMatchingRnnArray(Model* model,
+                         const std::string& back_edge_source_array,
+                         std::string* rnn_array) {
   for (const auto& rnn_state : model->flags.rnn_states()) {
     if (rnn_state.back_edge_source_array() == back_edge_source_array) {
       *rnn_array = rnn_state.state_array();
diff --git a/tensorflow/lite/toco/graph_transformations/lstm_utils.h b/tensorflow/lite/toco/graph_transformations/lstm_utils.h
index 949292ee84b..102fe7d6cfc 100644
--- a/tensorflow/lite/toco/graph_transformations/lstm_utils.h
+++ b/tensorflow/lite/toco/graph_transformations/lstm_utils.h
@@ -62,12 +62,12 @@ enum ExtendedLstmCellOutputs {
 };
 
 // Create optional array used for optional tensor in ExtendedLstmCell inputs.
-void CreateOptionalArray(Model* model, string* input_array_buffer,
-                         const string& array_name);
+void CreateOptionalArray(Model* model, std::string* input_array_buffer,
+                         const std::string& array_name);
 
 // Create float array and get its buffer.
 Buffer* CreateFloatArrayBuffer(Model* model,
-                                                      string* array_name,
+                                                      std::string* array_name,
                                                       const Shape& shape);
 
 // Copy data from one array to the other one (supports 1D and 2D array),
@@ -91,8 +91,8 @@ void CopyArrayData(const Buffer& src_buffer,
 
 // Copy a subset of array data and create a smaller array,
 // mostly used for spliting weights and bias for Lstm cell.
-void CopySubArrayToArray(Model* model, string* array_name,
-                         const string& tensor_name, int dim1_size,
+void CopySubArrayToArray(Model* model, std::string* array_name,
+                         const std::string& tensor_name, int dim1_size,
                          int dim2_size, const Array& original_array,
                          int start_idx1, int start_idx2);
 
@@ -103,8 +103,9 @@ void CopyArrayToSubArray(Buffer& tensor_buffer,
                          int start_idx1, int start_idx2);
 
 // Get mating rnn array inputs using rnn_states flag.
-bool GetMatchingRnnArray(Model* model, const string& back_edge_source_array,
-                         string* rnn_array);
+bool GetMatchingRnnArray(Model* model,
+                         const std::string& back_edge_source_array,
+                         std::string* rnn_array);
 
 }  // namespace toco
 
diff --git a/tensorflow/lite/toco/graph_transformations/make_initial_dequantize_operator.cc b/tensorflow/lite/toco/graph_transformations/make_initial_dequantize_operator.cc
index b914838b91c..7783b41767c 100644
--- a/tensorflow/lite/toco/graph_transformations/make_initial_dequantize_operator.cc
+++ b/tensorflow/lite/toco/graph_transformations/make_initial_dequantize_operator.cc
@@ -31,7 +31,8 @@ namespace toco {
 // generate this output to be removed by graph transformations.  Note that there
 // may be more than one operator that takes the input_array as their input, and
 // that some of these may be removed by graph transformations.
-bool AddDequantizeOperatorToInput(const string& input_name, const Operator* op,
+bool AddDequantizeOperatorToInput(const std::string& input_name,
+                                  const Operator* op,
                                   GraphTransformation* transformation,
                                   Model* model) {
   // An operator with the required output may be a dequantize operator already
@@ -65,7 +66,7 @@ bool AddDequantizeOperatorToInput(const string& input_name, const Operator* op,
   const auto& dequantized_input_name =
       AvailableArrayName(*model, input_name + "_dequantized");
   for (auto& other_op : model->operators) {
-    for (string& other_op_input : other_op->inputs) {
+    for (std::string& other_op_input : other_op->inputs) {
       if (other_op_input == input_name) {
         other_op_input = dequantized_input_name;
       }
diff --git a/tensorflow/lite/toco/graph_transformations/merge_reshape_into_preceding_transpose.cc b/tensorflow/lite/toco/graph_transformations/merge_reshape_into_preceding_transpose.cc
index 80170fe8bcb..96ccc22d9e9 100644
--- a/tensorflow/lite/toco/graph_transformations/merge_reshape_into_preceding_transpose.cc
+++ b/tensorflow/lite/toco/graph_transformations/merge_reshape_into_preceding_transpose.cc
@@ -117,8 +117,8 @@ std::vector ReshapeToTranspose(const Model& model,
     return ::tensorflow::Status::OK();
   }
 
-  const string intermediate_name = reshape_op->inputs[0];
-  const string output_name = reshape_op->outputs[0];
+  const std::string intermediate_name = reshape_op->inputs[0];
+  const std::string output_name = reshape_op->outputs[0];
 
   // Guarantee the input is only consume by the reshape.
   if (CountOpsWithInput(*model, intermediate_name) != 1) {
diff --git a/tensorflow/lite/toco/graph_transformations/move_binary_operator_before_reshape.cc b/tensorflow/lite/toco/graph_transformations/move_binary_operator_before_reshape.cc
index 0f3c4d34d66..222d1fd1e08 100644
--- a/tensorflow/lite/toco/graph_transformations/move_binary_operator_before_reshape.cc
+++ b/tensorflow/lite/toco/graph_transformations/move_binary_operator_before_reshape.cc
@@ -141,7 +141,7 @@ bool IsTailOfShape(const Shape& tail, const Shape& shape) {
   }
 
   // EXTRA CHECKS ON CONNECTING ARRAY
-  for (const string& output_array : model->flags.output_arrays()) {
+  for (const std::string& output_array : model->flags.output_arrays()) {
     if (binary_op->inputs[variable_input_idx] == output_array) {
       AddMessageF(
           "Not moving %s because the output of reshape op %s is an output op.",
diff --git a/tensorflow/lite/toco/graph_transformations/propagate_activation_function_into_constants.cc b/tensorflow/lite/toco/graph_transformations/propagate_activation_function_into_constants.cc
index 95de60262e7..a66a4cd0124 100644
--- a/tensorflow/lite/toco/graph_transformations/propagate_activation_function_into_constants.cc
+++ b/tensorflow/lite/toco/graph_transformations/propagate_activation_function_into_constants.cc
@@ -52,7 +52,7 @@ namespace toco {
   }
 
   // Filter to the list of supported ops.
-  string src_op_input;
+  std::string src_op_input;
   switch (src_op->type) {
     case OperatorType::kGather:
       src_op_input = src_op->inputs[0];
diff --git a/tensorflow/lite/toco/graph_transformations/propagate_array_data_types.cc b/tensorflow/lite/toco/graph_transformations/propagate_array_data_types.cc
index 49d59de860b..5eda1950745 100644
--- a/tensorflow/lite/toco/graph_transformations/propagate_array_data_types.cc
+++ b/tensorflow/lite/toco/graph_transformations/propagate_array_data_types.cc
@@ -48,7 +48,7 @@ void SetDataTypeForAllOutputs(Model* model, Operator* op,
   }
   // Record data types of output before processing, so we can see at the
   // end if we changed anything, and return the correct boolean value.
-  std::unordered_map old_output_data_types;
+  std::unordered_map old_output_data_types;
   for (const auto& output : op->outputs) {
     old_output_data_types[output] = model->GetArray(output).data_type;
   }
@@ -171,7 +171,7 @@ void SetDataTypeForAllOutputs(Model* model, Operator* op,
         return ::tensorflow::Status::OK();
       }
       for (int i = 0; i < op->outputs.size(); ++i) {
-        const string& output = op->outputs[i];
+        const std::string& output = op->outputs[i];
         const ArrayDataType data_type = unsupported_op->output_data_types[i];
         model->GetArray(output).data_type = data_type;
       }
diff --git a/tensorflow/lite/toco/graph_transformations/propagate_default_min_max.cc b/tensorflow/lite/toco/graph_transformations/propagate_default_min_max.cc
index d31ba956afd..bf1109ddba5 100644
--- a/tensorflow/lite/toco/graph_transformations/propagate_default_min_max.cc
+++ b/tensorflow/lite/toco/graph_transformations/propagate_default_min_max.cc
@@ -70,7 +70,7 @@ bool SupportsMinMax(const Array& array) {
 
 // Sets the min/max on the given array, adjusting the reference_minmax for the
 // final data type of the array if it is already specified.
-bool PropagateDefaultMinMax::SetArrayMinMax(const string& array_name,
+bool PropagateDefaultMinMax::SetArrayMinMax(const std::string& array_name,
                                             Array* array) {
   CHECK(!array->minmax);
 
diff --git a/tensorflow/lite/toco/graph_transformations/propagate_fixed_sizes.cc b/tensorflow/lite/toco/graph_transformations/propagate_fixed_sizes.cc
index 006e624eb7a..1524cfe7f35 100644
--- a/tensorflow/lite/toco/graph_transformations/propagate_fixed_sizes.cc
+++ b/tensorflow/lite/toco/graph_transformations/propagate_fixed_sizes.cc
@@ -268,7 +268,7 @@ void ProcessDepthwiseConvOperator(Model* model, DepthwiseConvOperator* op) {
   const auto& weights_shape = weights_array.shape();
   CHECK_EQ(weights_shape.dimensions_count(), 4);
 
-  const string& output_name = op->outputs[0];
+  const std::string& output_name = op->outputs[0];
   const int input_depth = input_shape.dims(3);
   const int output_depth = weights_shape.dims(3);
   // TensorFlow doesn't define the depth_multiplier value on DepthwiseConv ops,
@@ -302,7 +302,7 @@ void ProcessDepthToSpaceOperator(Model* model, DepthToSpaceOperator* op) {
   const auto& input_shape = input_array.shape();
   CHECK_EQ(input_shape.dimensions_count(), 4);
 
-  const string& output_name = op->outputs[0];
+  const std::string& output_name = op->outputs[0];
   const int block_size = op->block_size;
   CHECK_NE(block_size, 0) << "Invalid block_size in " << output_name;
   const int batch = input_shape.dims(0);
@@ -325,7 +325,7 @@ void ProcessSpaceToDepthOperator(Model* model, SpaceToDepthOperator* op) {
   const auto& input_shape = input_array.shape();
   CHECK_EQ(input_shape.dimensions_count(), 4);
 
-  const string& output_name = op->outputs[0];
+  const std::string& output_name = op->outputs[0];
   const int block_size = op->block_size;
   CHECK_NE(block_size, 0) << "Invalid block_size in " << output_name;
   const int batch = input_shape.dims(0);
@@ -470,7 +470,7 @@ void ProcessSimpleOperator(Model* model, Operator* op, int input_index) {
     return;
   }
 
-  const string& output_name = op->outputs[0];
+  const std::string& output_name = op->outputs[0];
   auto& output_array = model->GetArray(output_name);
   if (output_array.has_shape()) {
     return;
@@ -487,7 +487,7 @@ void ProcessSimpleBinaryOperator(Model* model, Operator* op) {
   if (!input0_array.has_shape() || !input1_array.has_shape()) {
     return;
   }
-  const string& output_name = op->outputs[0];
+  const std::string& output_name = op->outputs[0];
   auto& output_array = model->GetArray(output_name);
   ComputeBinaryOperatorOutputSize(input0_array.shape(), input1_array.shape(),
                                   &output_array);
@@ -639,14 +639,14 @@ void ProcessSliceOperator(Model* model, SliceOperator* op) {
 }
 
 void ProcessReorderAxesOperator(Model* model, ReorderAxesOperator* op) {
-  const string& input_name = op->inputs[0];
+  const std::string& input_name = op->inputs[0];
   const auto& input_array = model->GetArray(input_name);
   // Yield until input dims have been resolved.
   if (!input_array.has_shape()) {
     return;
   }
   const auto& input_shape = input_array.shape();
-  const string& output_name = op->outputs[0];
+  const std::string& output_name = op->outputs[0];
   Shape* output_shape = model->GetArray(output_name).mutable_shape();
   ShuffleDims(input_shape, op->input_axes_order, op->output_axes_order,
               output_shape);
@@ -757,7 +757,7 @@ void ProcessRangeOperator(Model* model, RangeOperator* op) {
 
 void ProcessTensorFlowSplitOperator(Model* model, TensorFlowSplitOperator* op) {
   CHECK_EQ(op->inputs.size(), 2);
-  const string& input_name = op->inputs[1];
+  const std::string& input_name = op->inputs[1];
   const auto& input_array = model->GetArray(input_name);
   // Yield until input dims have been resolved.
   if (!input_array.has_shape()) {
@@ -892,7 +892,7 @@ void ProcessTensorFlowSplitVOperator(Model* model,
 }
 
 void ProcessAveragePoolOperator(Model* model, AveragePoolOperator* op) {
-  const string& input_name = op->inputs[0];
+  const std::string& input_name = op->inputs[0];
   const auto& input_array = model->GetArray(input_name);
   // Yield until input dims have been resolved.
   if (!input_array.has_shape()) {
@@ -900,7 +900,7 @@ void ProcessAveragePoolOperator(Model* model, AveragePoolOperator* op) {
   }
   const auto& input_shape = input_array.shape();
   CHECK_EQ(input_shape.dimensions_count(), 4);
-  const string& output_name = op->outputs[0];
+  const std::string& output_name = op->outputs[0];
   const int output_depth = input_shape.dims(3);
   ComputeConvSizes(input_shape, output_depth, op->kwidth, op->kheight,
                    op->stride_width, op->stride_height, 1, 1, op->padding.type,
@@ -909,7 +909,7 @@ void ProcessAveragePoolOperator(Model* model, AveragePoolOperator* op) {
 }
 
 void ProcessMaxPoolOperator(Model* model, MaxPoolOperator* op) {
-  const string& input_name = op->inputs[0];
+  const std::string& input_name = op->inputs[0];
   const auto& input_array = model->GetArray(input_name);
   // Yield until input dims have been resolved.
   if (!input_array.has_shape()) {
@@ -917,7 +917,7 @@ void ProcessMaxPoolOperator(Model* model, MaxPoolOperator* op) {
   }
   const auto& input_shape = input_array.shape();
   CHECK_EQ(input_shape.dimensions_count(), 4);
-  const string& output_name = op->outputs[0];
+  const std::string& output_name = op->outputs[0];
   const int output_depth = input_shape.dims(3);
   ComputeConvSizes(input_shape, output_depth, op->kwidth, op->kheight,
                    op->stride_width, op->stride_height, 1, 1, op->padding.type,
@@ -926,7 +926,7 @@ void ProcessMaxPoolOperator(Model* model, MaxPoolOperator* op) {
 }
 
 void ProcessL2PoolOperator(Model* model, L2PoolOperator* op) {
-  const string& input_name = op->inputs[0];
+  const std::string& input_name = op->inputs[0];
   const auto& input_array = model->GetArray(input_name);
   // Yield until input dims have been resolved.
   if (!input_array.has_shape()) {
@@ -936,7 +936,7 @@ void ProcessL2PoolOperator(Model* model, L2PoolOperator* op) {
   if (input_shape.dimensions_count() < 4) {
     LOG(FATAL) << "missing dimensions for " << input_name;
   }
-  const string& output_name = op->outputs[0];
+  const std::string& output_name = op->outputs[0];
   const int output_depth = input_shape.dims(3);
   ComputeConvSizes(input_shape, output_depth, op->kwidth, op->kheight,
                    op->stride_width, op->stride_height, 1, 1, op->padding.type,
@@ -954,7 +954,7 @@ void ProcessResizeBilinearOperator(Model* model, ResizeBilinearOperator* op) {
   }
   const auto& input_data_shape = model->GetArray(op->inputs[0]).shape();
 
-  const string& output_size_name = op->inputs[1];
+  const std::string& output_size_name = op->inputs[1];
   const auto& output_size_array = model->GetArray(output_size_name);
   CHECK(output_size_array.data_type == ArrayDataType::kInt32);
   CHECK(output_size_array.has_shape());
@@ -982,7 +982,7 @@ void ProcessResizeNearestNeighborOperator(Model* model,
   }
   const auto& input_data_shape = model->GetArray(op->inputs[0]).shape();
 
-  const string& output_size_name = op->inputs[1];
+  const std::string& output_size_name = op->inputs[1];
   const auto& output_size_array = model->GetArray(output_size_name);
   CHECK(output_size_array.data_type == ArrayDataType::kInt32);
   CHECK(output_size_array.has_shape());
@@ -1862,7 +1862,7 @@ void ProcessArgMinMaxOperator(Model* model, Op* op) {
     }
   }
 
-  const string& output_name = op->outputs[0];
+  const std::string& output_name = op->outputs[0];
   auto& output_array = model->GetArray(output_name);
   if (output_array.has_shape()) {
     return;
@@ -1880,7 +1880,7 @@ void ProcessSparseToDenseOperator(Model* model, SparseToDenseOperator* op) {
   // Output should not go over four dimensions.
   CHECK_LE(output_shape_array.shape().dims(0), 4);
 
-  const string& output_name = op->outputs[0];
+  const std::string& output_name = op->outputs[0];
   Array& output_array = model->GetArray(output_name);
   if (output_array.has_shape()) return;
 
@@ -2015,7 +2015,7 @@ void ProcessUnpackOperator(Model* model, UnpackOperator* op) {
       output_dims.push_back(input_dims[i]);
     }
   }
-  for (const string& output_name : op->outputs) {
+  for (const std::string& output_name : op->outputs) {
     auto& output_array = model->GetArray(output_name);
     if (output_array.has_shape()) {
       return;
@@ -2149,7 +2149,7 @@ void ProcessScatterNdOperator(Model* model, ScatterNdOperator* op) {
   *modified = false;
   auto it = model->operators.begin() + op_index;
   auto* op = it->get();
-  std::unordered_map> old_output_dims;
+  std::unordered_map> old_output_dims;
   for (const auto& output : op->outputs) {
     if (model->GetArray(output).has_shape()) {
       old_output_dims[output] = model->GetArray(output).shape().dims();
@@ -2400,7 +2400,7 @@ void ProcessScatterNdOperator(Model* model, ScatterNdOperator* op) {
         return ::tensorflow::Status::OK();
       }
       for (int i = 0; i < op->outputs.size(); ++i) {
-        const string& output = op->outputs[i];
+        const std::string& output = op->outputs[i];
         model->GetArray(output).copy_shape(unsupported_op->output_shapes.at(i));
       }
       break;
diff --git a/tensorflow/lite/toco/graph_transformations/quantization_util.cc b/tensorflow/lite/toco/graph_transformations/quantization_util.cc
index 23749abf0b1..76ead658107 100644
--- a/tensorflow/lite/toco/graph_transformations/quantization_util.cc
+++ b/tensorflow/lite/toco/graph_transformations/quantization_util.cc
@@ -164,7 +164,7 @@ std::unique_ptr QuantizeBuffer(
 
 template 
 void QuantizeArray(GraphTransformation* transformation, Model* model,
-                   const string& name,
+                   const std::string& name,
                    const QuantizationParams& quantization_params) {
   auto& array = model->GetArray(name);
   CHECK(array.data_type == ArrayDataType::kFloat);
@@ -184,7 +184,7 @@ void QuantizeArray(GraphTransformation* transformation, Model* model,
 }  // namespace
 
 void QuantizeArray(GraphTransformation* transformation, Model* model,
-                   const string& name, ArrayDataType quantized_data_type,
+                   const std::string& name, ArrayDataType quantized_data_type,
                    const QuantizationParams& quantization_params) {
   ArrayDataType adjusted_data_type = quantized_data_type;
   auto& array = model->GetArray(name);
diff --git a/tensorflow/lite/toco/graph_transformations/quantization_util.h b/tensorflow/lite/toco/graph_transformations/quantization_util.h
index d226aeab8b7..d1d72b98f9e 100644
--- a/tensorflow/lite/toco/graph_transformations/quantization_util.h
+++ b/tensorflow/lite/toco/graph_transformations/quantization_util.h
@@ -47,7 +47,7 @@ void ChooseQuantizationParamsForArrayAndQuantizedDataType(
 // Quantizes an array by setting its data type and (if constant) quantizing
 // all values in the array.
 void QuantizeArray(GraphTransformation* transformation, Model* model,
-                   const string& name, ArrayDataType quantized_data_type,
+                   const std::string& name, ArrayDataType quantized_data_type,
                    const QuantizationParams& quantization_params);
 
 // Returns true if the given array, when quantized, contains only values between
diff --git a/tensorflow/lite/toco/graph_transformations/quantize.cc b/tensorflow/lite/toco/graph_transformations/quantize.cc
index e6fd88c9787..c5848f83dd3 100644
--- a/tensorflow/lite/toco/graph_transformations/quantize.cc
+++ b/tensorflow/lite/toco/graph_transformations/quantize.cc
@@ -121,7 +121,7 @@ bool SupportOutputTypeFloatInQuantizedOp(const Operator& op) {
   }
   return false;
 }
-const MinMax& GetOrComputeMinMax(Model* model, const string& array_name) {
+const MinMax& GetOrComputeMinMax(Model* model, const std::string& array_name) {
   auto& array = model->GetArray(array_name);
   // Normally we should have a MinMax recorded on this Array,
   // so we just use it.
diff --git a/tensorflow/lite/toco/graph_transformations/read_array_minmax_and_narrow_range_from_fake_quant.cc b/tensorflow/lite/toco/graph_transformations/read_array_minmax_and_narrow_range_from_fake_quant.cc
index 4d621018dc3..30875c7e59e 100644
--- a/tensorflow/lite/toco/graph_transformations/read_array_minmax_and_narrow_range_from_fake_quant.cc
+++ b/tensorflow/lite/toco/graph_transformations/read_array_minmax_and_narrow_range_from_fake_quant.cc
@@ -29,7 +29,7 @@ namespace {
 
 bool ApplyAttrsToArray(GraphTransformation* transformation, Model* model,
                        const FakeQuantOperator& fq_op,
-                       const string& array_name) {
+                       const std::string& array_name) {
   bool changed = false;
   auto& annotated_array = model->GetArray(array_name);
   if (!annotated_array.minmax) {
diff --git a/tensorflow/lite/toco/graph_transformations/remove_successive_transpose.cc b/tensorflow/lite/toco/graph_transformations/remove_successive_transpose.cc
index 6eccda04c18..5e3b6f7b615 100644
--- a/tensorflow/lite/toco/graph_transformations/remove_successive_transpose.cc
+++ b/tensorflow/lite/toco/graph_transformations/remove_successive_transpose.cc
@@ -43,8 +43,8 @@ bool TransformsToIdentity(std::vector const& perm1,
   return true;
 }
 
-void ReplaceOpInputsWith(Model* model, const string& lookfor,
-                         const string& replacewith) {
+void ReplaceOpInputsWith(Model* model, const std::string& lookfor,
+                         const std::string& replacewith) {
   for (const auto& op : model->operators) {
     for (int i = 0; i < op->inputs.size(); ++i) {
       if (op->inputs[i] == lookfor) {
diff --git a/tensorflow/lite/toco/graph_transformations/remove_trivial_concatenation_input.cc b/tensorflow/lite/toco/graph_transformations/remove_trivial_concatenation_input.cc
index df4a4ea51c4..dc210fc03d8 100644
--- a/tensorflow/lite/toco/graph_transformations/remove_trivial_concatenation_input.cc
+++ b/tensorflow/lite/toco/graph_transformations/remove_trivial_concatenation_input.cc
@@ -41,9 +41,9 @@ namespace toco {
   if (concat_op->type != OperatorType::kConcatenation) {
     return ::tensorflow::Status::OK();
   }
-  std::vector trivial_inputs;
-  std::vector nontrivial_inputs;
-  for (const string& input : concat_op->inputs) {
+  std::vector trivial_inputs;
+  std::vector nontrivial_inputs;
+  for (const std::string& input : concat_op->inputs) {
     const auto& input_array = model->GetArray(input);
     const bool is_trivial =
         input_array.has_shape() && input_array.shape().dimensions_count() == 0;
@@ -60,7 +60,7 @@ namespace toco {
 
   // Drop trivial inputs.
   concat_op->inputs = nontrivial_inputs;
-  for (const string& input : trivial_inputs) {
+  for (const std::string& input : trivial_inputs) {
     DeleteArrayIfUnusedOutsideOfOp(input, concat_op, model);
   }
   *modified = true;
diff --git a/tensorflow/lite/toco/graph_transformations/remove_trivial_passthrough.cc b/tensorflow/lite/toco/graph_transformations/remove_trivial_passthrough.cc
index bd529bd9ecd..45dbec83471 100644
--- a/tensorflow/lite/toco/graph_transformations/remove_trivial_passthrough.cc
+++ b/tensorflow/lite/toco/graph_transformations/remove_trivial_passthrough.cc
@@ -29,7 +29,7 @@ namespace {
 // array instead. from_array is assumed to be discardable, and consequently
 // this only updates operator edges (since discardable arrays only
 // appear there, and not e.g. in model flags).
-void Reroute(const string& from, const string& to, Model* model) {
+void Reroute(const std::string& from, const std::string& to, Model* model) {
   for (const auto& op : model->operators) {
     for (auto& output : op->outputs) {
       if (output == from) {
@@ -92,8 +92,9 @@ bool RemoveTrivialPassthroughOp(GraphTransformation* transformation,
     }
   }
 
-  const string main_input_name = passthru_op->inputs[main_input_array_index];
-  const string output_name = passthru_op->outputs[0];
+  const std::string main_input_name =
+      passthru_op->inputs[main_input_array_index];
+  const std::string output_name = passthru_op->outputs[0];
 
   if (IsDiscardableArray(*model, output_name)) {
     transformation->AddMessageF(
diff --git a/tensorflow/lite/toco/graph_transformations/remove_trivial_quantized_activation_func.cc b/tensorflow/lite/toco/graph_transformations/remove_trivial_quantized_activation_func.cc
index 56acf22f7f1..80d28e0fc6d 100644
--- a/tensorflow/lite/toco/graph_transformations/remove_trivial_quantized_activation_func.cc
+++ b/tensorflow/lite/toco/graph_transformations/remove_trivial_quantized_activation_func.cc
@@ -32,7 +32,7 @@ namespace {
 
 bool IsTrivialUnfusedActivationFunc(GraphTransformation* transformation,
                                     const Model& model, OperatorType op_type,
-                                    const string& input_array_name) {
+                                    const std::string& input_array_name) {
   double clamp_min;
   double clamp_max;
   switch (op_type) {
@@ -60,7 +60,7 @@ bool IsTrivialUnfusedActivationFunc(GraphTransformation* transformation,
 bool IsTrivialFusedActivationFunc(
     GraphTransformation* transformation, const Model& model,
     FusedActivationFunctionType activation_function,
-    const string& output_array_name) {
+    const std::string& output_array_name) {
   double clamp_min;
   double clamp_max;
   switch (activation_function) {
diff --git a/tensorflow/lite/toco/graph_transformations/remove_trivial_quantized_min_max.cc b/tensorflow/lite/toco/graph_transformations/remove_trivial_quantized_min_max.cc
index f1037994c97..bd9281fe34e 100644
--- a/tensorflow/lite/toco/graph_transformations/remove_trivial_quantized_min_max.cc
+++ b/tensorflow/lite/toco/graph_transformations/remove_trivial_quantized_min_max.cc
@@ -31,8 +31,8 @@ namespace toco {
 namespace {
 
 bool IsTrivialMinMax(GraphTransformation* transformation, const Model& model,
-                     OperatorType op_type, const string& input_array_name,
-                     const string& clamp_value_array_name) {
+                     OperatorType op_type, const std::string& input_array_name,
+                     const std::string& clamp_value_array_name) {
   const auto& clamp_value_array = model.GetArray(clamp_value_array_name);
   if (!IsConstantParameterArray(model, clamp_value_array_name)) {
     transformation->AddMessageF("Clip value array %s is non-constant",
diff --git a/tensorflow/lite/toco/graph_transformations/remove_unused_op.cc b/tensorflow/lite/toco/graph_transformations/remove_unused_op.cc
index 384b5f22911..bcdb4cbe77e 100644
--- a/tensorflow/lite/toco/graph_transformations/remove_unused_op.cc
+++ b/tensorflow/lite/toco/graph_transformations/remove_unused_op.cc
@@ -58,7 +58,7 @@ namespace toco {
     if (found_output_as_rnn_state_array) {
       continue;
     }
-    for (const string& output_array : model->flags.output_arrays()) {
+    for (const std::string& output_array : model->flags.output_arrays()) {
       if (output == output_array) {
         return ::tensorflow::Status::OK();
       }
diff --git a/tensorflow/lite/toco/graph_transformations/reorder_elementwise_unary.cc b/tensorflow/lite/toco/graph_transformations/reorder_elementwise_unary.cc
index 17a5e9a1d6a..158c7f95085 100644
--- a/tensorflow/lite/toco/graph_transformations/reorder_elementwise_unary.cc
+++ b/tensorflow/lite/toco/graph_transformations/reorder_elementwise_unary.cc
@@ -75,7 +75,7 @@ bool IsMoveOperator(OperatorType optype) {
     return ::tensorflow::Status::OK();
   }
 
-  const string intermediate_name = element_op->inputs[0];
+  const std::string intermediate_name = element_op->inputs[0];
   auto it = FindOpWithOutput(*model, intermediate_name);
   if (it == model->operators.end()) {
     AddMessageF("No preceding operator");
@@ -103,8 +103,8 @@ bool IsMoveOperator(OperatorType optype) {
   }
 
   // op->inputs may change so we need to keep a value by copy.
-  const string input_name = move_op->inputs[0];
-  const string output_name = element_op->outputs[0];
+  const std::string input_name = move_op->inputs[0];
+  const std::string output_name = element_op->outputs[0];
 
   AddMessageF("Swapping around operators with %s and %s", LogName(*element_op),
               LogName(*move_op));
diff --git a/tensorflow/lite/toco/graph_transformations/reorder_reshape_transpose.cc b/tensorflow/lite/toco/graph_transformations/reorder_reshape_transpose.cc
index 0fbcf9f73b1..9852b3382cd 100644
--- a/tensorflow/lite/toco/graph_transformations/reorder_reshape_transpose.cc
+++ b/tensorflow/lite/toco/graph_transformations/reorder_reshape_transpose.cc
@@ -138,9 +138,9 @@ std::vector ComputeNewPerm(std::vector input_dims,
   }
 
   // Need to copy to keep static if permutated.
-  const string input_name = reshape_op->inputs[0];
-  const string intermediate_name = reshape_op->outputs[0];
-  const string output_name = transpose_op->outputs[0];
+  const std::string input_name = reshape_op->inputs[0];
+  const std::string intermediate_name = reshape_op->outputs[0];
+  const std::string output_name = transpose_op->outputs[0];
 
   // Intermediate should not be consumed by any other operators.
   if (CountOpsWithInput(*model, intermediate_name) != 1) {
diff --git a/tensorflow/lite/toco/graph_transformations/resolve_batch_normalization.cc b/tensorflow/lite/toco/graph_transformations/resolve_batch_normalization.cc
index 6e5815ee94d..aee511e2beb 100644
--- a/tensorflow/lite/toco/graph_transformations/resolve_batch_normalization.cc
+++ b/tensorflow/lite/toco/graph_transformations/resolve_batch_normalization.cc
@@ -62,12 +62,14 @@ namespace toco {
   // Create the new Mul, Add operators
   auto* mul_op = new MulOperator;
   auto* add_op = new AddOperator;
-  const string mul_name =
+  const std::string mul_name =
       AvailableArrayName(*model, bn_op->outputs[0] + "_mul");
-  const string add_name =
+  const std::string add_name =
       AvailableArrayName(*model, bn_op->outputs[0] + "_add");
-  const string mul_param_name = AvailableArrayName(*model, mul_name + "_param");
-  const string add_param_name = AvailableArrayName(*model, add_name + "_param");
+  const std::string mul_param_name =
+      AvailableArrayName(*model, mul_name + "_param");
+  const std::string add_param_name =
+      AvailableArrayName(*model, add_name + "_param");
   mul_op->inputs = {bn_op->inputs[0], mul_param_name};
   mul_op->outputs = {mul_name};
   add_op->inputs = {mul_name, add_param_name};
diff --git a/tensorflow/lite/toco/graph_transformations/resolve_constant_concatenation.cc b/tensorflow/lite/toco/graph_transformations/resolve_constant_concatenation.cc
index 7c9aa025f64..208c345639b 100644
--- a/tensorflow/lite/toco/graph_transformations/resolve_constant_concatenation.cc
+++ b/tensorflow/lite/toco/graph_transformations/resolve_constant_concatenation.cc
@@ -147,7 +147,7 @@ void SetMinMaxForConcatenedArray(GraphTransformation* transformation,
   const auto* concat_op =
       static_cast(concat_base_op);
 
-  for (const string& input_name : concat_op->inputs) {
+  for (const std::string& input_name : concat_op->inputs) {
     // We only expect constant unquantized arrays as input, otherwise we return.
     // We  also make sure the shapes of the input arrays are known and they are
     // all discardable.
@@ -166,10 +166,10 @@ void SetMinMaxForConcatenedArray(GraphTransformation* transformation,
   const int concatenation_axis = concat_op->axis;
 
   CHECK_EQ(concat_op->outputs.size(), 1);
-  string concatenated_array_name = concat_op->outputs[0];
+  std::string concatenated_array_name = concat_op->outputs[0];
   Array& concatenated_array = model->GetOrCreateArray(concatenated_array_name);
   std::vector input_arrays;
-  for (const string& input_name : concat_op->inputs) {
+  for (const std::string& input_name : concat_op->inputs) {
     input_arrays.push_back(&model->GetArray(input_name));
   }
 
diff --git a/tensorflow/lite/toco/graph_transformations/resolve_reorder_axes.cc b/tensorflow/lite/toco/graph_transformations/resolve_reorder_axes.cc
index a3d3e863757..a685f67745b 100644
--- a/tensorflow/lite/toco/graph_transformations/resolve_reorder_axes.cc
+++ b/tensorflow/lite/toco/graph_transformations/resolve_reorder_axes.cc
@@ -27,19 +27,19 @@ namespace toco {
 
 namespace {
 
-void RenameArray(Model* model, const string& oldname,
-                 const string& desired_newname) {
-  const string& newname = AvailableArrayName(*model, desired_newname);
+void RenameArray(Model* model, const std::string& oldname,
+                 const std::string& desired_newname) {
+  const std::string& newname = AvailableArrayName(*model, desired_newname);
   auto& arrays = model->GetMutableArrayMap();
   arrays[newname] = std::move(arrays[oldname]);
   arrays.erase(oldname);
   for (const auto& op : model->operators) {
-    for (string& input : op->inputs) {
+    for (std::string& input : op->inputs) {
       if (input == oldname) {
         input = newname;
       }
     }
-    for (string& output : op->outputs) {
+    for (std::string& output : op->outputs) {
       if (output == oldname) {
         output = newname;
       }
@@ -89,8 +89,8 @@ void ReorderAxes(AxesOrder input_axes_order, AxesOrder output_axes_order,
   auto* reorder_op = static_cast(op);
 
   // Intentionally copies, not references.
-  const string input_array_name = reorder_op->inputs[0];
-  const string output_array_name = reorder_op->outputs[0];
+  const std::string input_array_name = reorder_op->inputs[0];
+  const std::string output_array_name = reorder_op->outputs[0];
 
   auto& input_array = model->GetArray(input_array_name);
   auto& output_array = model->GetArray(output_array_name);
diff --git a/tensorflow/lite/toco/graph_transformations/resolve_tensorflow_concat.cc b/tensorflow/lite/toco/graph_transformations/resolve_tensorflow_concat.cc
index b5a11529764..1ce05336be9 100644
--- a/tensorflow/lite/toco/graph_transformations/resolve_tensorflow_concat.cc
+++ b/tensorflow/lite/toco/graph_transformations/resolve_tensorflow_concat.cc
@@ -44,8 +44,8 @@ namespace toco {
   if (tf_concat_op->type == OperatorType::kConcatV2) {
     axis_pos = tf_concat_op->inputs.size() - 1;
   }
-  const string axis_name = tf_concat_op->inputs[axis_pos];
-  std::vector concat_input_names;
+  const std::string axis_name = tf_concat_op->inputs[axis_pos];
+  std::vector concat_input_names;
   for (std::size_t i = 0; i < tf_concat_op->inputs.size(); i++) {
     if (i != axis_pos) {
       concat_input_names.push_back(tf_concat_op->inputs[i]);
diff --git a/tensorflow/lite/toco/graph_transformations/resolve_tensorflow_matmul.cc b/tensorflow/lite/toco/graph_transformations/resolve_tensorflow_matmul.cc
index ac95d609e91..a6d653d055d 100644
--- a/tensorflow/lite/toco/graph_transformations/resolve_tensorflow_matmul.cc
+++ b/tensorflow/lite/toco/graph_transformations/resolve_tensorflow_matmul.cc
@@ -27,7 +27,7 @@ namespace toco {
 namespace {
 
 TransposeOperator* FindTransposeOpWithInput(const Model& model,
-                                            const string& array_name) {
+                                            const std::string& array_name) {
   for (auto it = model.operators.begin(); it != model.operators.end(); ++it) {
     Operator* op = it->get();
     if (op->type != OperatorType::kTranspose) {
@@ -74,8 +74,8 @@ TransposeOperator* FindTransposeOpWithInput(const Model& model,
     DCHECK_EQ(matmul_it->get(), matmul_op);
   };
 
-  string input_lhs = matmul_op->inputs[0];
-  string input_rhs = matmul_op->inputs[1];
+  std::string input_lhs = matmul_op->inputs[0];
+  std::string input_rhs = matmul_op->inputs[1];
 
   // Handle `transpose_a` with best effort: If the dimension of lhs is known,
   // insert a `Transpose` op.
diff --git a/tensorflow/lite/toco/graph_transformations/resolve_tensorflow_switch.cc b/tensorflow/lite/toco/graph_transformations/resolve_tensorflow_switch.cc
index 5c3176ced34..854dce39a27 100644
--- a/tensorflow/lite/toco/graph_transformations/resolve_tensorflow_switch.cc
+++ b/tensorflow/lite/toco/graph_transformations/resolve_tensorflow_switch.cc
@@ -37,7 +37,7 @@ namespace toco {
 
   CHECK_EQ(switch_op->inputs.size(), 2);
   CHECK_EQ(switch_op->outputs.size(), 2);
-  const string& predicate_name = switch_op->inputs[1];
+  const std::string& predicate_name = switch_op->inputs[1];
   // If the predicate array hasn't been resolved to a constant yet,
   // we need to yield.
   if (!IsConstantParameterArray(*model, predicate_name)) {
diff --git a/tensorflow/lite/toco/graph_transformations/shuffle_fc_weights.cc b/tensorflow/lite/toco/graph_transformations/shuffle_fc_weights.cc
index 195ea70e34b..7eadd01c949 100644
--- a/tensorflow/lite/toco/graph_transformations/shuffle_fc_weights.cc
+++ b/tensorflow/lite/toco/graph_transformations/shuffle_fc_weights.cc
@@ -37,7 +37,7 @@ namespace toco {
     return ::tensorflow::Status::OK();
   }
   const Array& input_array = model->GetArray(fc_op->inputs[0]);
-  const string& weights_name = fc_op->inputs[1];
+  const std::string& weights_name = fc_op->inputs[1];
   Array& weights_array = model->GetArray(weights_name);
   const Array& output_array = model->GetArray(fc_op->outputs[0]);
   // Exit if this FC op isn't quantized with uint8 inputs and int16 outputs,
@@ -143,7 +143,7 @@ namespace toco {
   // Add a second output array to this FC op, serving as a workspace to perform
   // runtime shuffling/xoring of its input activations.
   CHECK_EQ(fc_op->outputs.size(), 1);
-  const string& shuffled_input_workspace_array_name =
+  const std::string& shuffled_input_workspace_array_name =
       AvailableArrayName(*model, fc_op->inputs[0] + "_shuffled");
   fc_op->outputs.push_back(shuffled_input_workspace_array_name);
   auto& shuffled_input_workspace_array =
diff --git a/tensorflow/lite/toco/graph_transformations/unfuse_activation_functions.cc b/tensorflow/lite/toco/graph_transformations/unfuse_activation_functions.cc
index 3e36dd5a45c..2a3cd91551b 100644
--- a/tensorflow/lite/toco/graph_transformations/unfuse_activation_functions.cc
+++ b/tensorflow/lite/toco/graph_transformations/unfuse_activation_functions.cc
@@ -64,7 +64,7 @@ namespace toco {
   // Wire up arrays, constructing a new intermediate array to connect the
   // op to its new unfused activation function.
   ac_op->outputs = op->outputs;
-  const string& tmp_array_name =
+  const std::string& tmp_array_name =
       AvailableArrayName(*model, op->outputs[0] + "_unfused");
   CHECK(!model->HasArray(tmp_array_name));
 
diff --git a/tensorflow/lite/toco/graph_transformations/unpartition_embedding_lookup.cc b/tensorflow/lite/toco/graph_transformations/unpartition_embedding_lookup.cc
index 1f7035c21e2..294c39069f7 100644
--- a/tensorflow/lite/toco/graph_transformations/unpartition_embedding_lookup.cc
+++ b/tensorflow/lite/toco/graph_transformations/unpartition_embedding_lookup.cc
@@ -55,8 +55,8 @@ namespace toco {
   auto* stitch_op = static_cast(op_it->get());
 
   // Split up the DynamicStitch inputs into the indices and data.
-  std::vector stitch_indices_inputs;
-  std::vector stitch_data_inputs;
+  std::vector stitch_indices_inputs;
+  std::vector stitch_data_inputs;
   for (size_t i = 0; i < stitch_op->num_partitions; ++i) {
     stitch_indices_inputs.push_back(stitch_op->inputs[i]);
   }
@@ -67,7 +67,8 @@ namespace toco {
 
   // Validate all indices come from the same DynamicPartition.
   DynamicPartitionOperator* indices_partition_op = nullptr;
-  for (const string& indices_partition_output_name : stitch_indices_inputs) {
+  for (const std::string& indices_partition_output_name :
+       stitch_indices_inputs) {
     auto* op = GetOpWithOutput(*model, indices_partition_output_name);
     CHECK(op) << "Source of " << indices_partition_output_name << " not found";
     if (op->type != OperatorType::kDynamicPartition) {
@@ -112,7 +113,7 @@ namespace toco {
 
   // Find all of the gathers used for the data inputs.
   std::vector gather_ops;
-  for (const string& gather_output_name : stitch_data_inputs) {
+  for (const std::string& gather_output_name : stitch_data_inputs) {
     auto* op = GetOpWithOutput(*model, gather_output_name);
     CHECK(op) << "Source of " << gather_output_name << " not found";
     if (op->type != OperatorType::kGather) {
diff --git a/tensorflow/lite/toco/graph_transformations/unroll_batch_matmul.cc b/tensorflow/lite/toco/graph_transformations/unroll_batch_matmul.cc
index 16dfaf7fc80..5bf000e2784 100644
--- a/tensorflow/lite/toco/graph_transformations/unroll_batch_matmul.cc
+++ b/tensorflow/lite/toco/graph_transformations/unroll_batch_matmul.cc
@@ -34,9 +34,10 @@ absl::InlinedVector ToInlinedVector(const std::vector& vec) {
   return absl::InlinedVector(vec.begin(), vec.end());
 }
 
-std::vector SliceInput(
-    const string& input, const string& base_name, const string& input_name,
-    const int batch_size, const Array& input_array, Model* model,
+std::vector SliceInput(
+    const std::string& input, const std::string& base_name,
+    const std::string& input_name, const int batch_size,
+    const Array& input_array, Model* model,
     std::vector>::iterator* tail_it) {
   int rank = input_array.shape().dimensions_count();
   int num_rows = input_array.shape().dims(rank - 2);
@@ -54,7 +55,7 @@ std::vector SliceInput(
   *tail_it = model->operators.emplace(*tail_it, reshape_op) + 1;
 
   // Slice along each batch index and remember the slice output for future use.
-  std::vector slice_outputs;
+  std::vector slice_outputs;
   for (int batch_idx = 0; batch_idx < batch_size; ++batch_idx) {
     std::string batch_name =
         absl::StrCat(base_name, "_b", batch_idx, "/slice_", input_name);
@@ -110,10 +111,10 @@ std::vector GetTransposeShape(const Shape& input_shape,
   return output_shape;
 }
 
-TransposeOperator* TransposeInput(const string& input, Model* model) {
+TransposeOperator* TransposeInput(const std::string& input, Model* model) {
   const auto& input_array = model->GetArray(input);
   const auto perm_array = GetTransposePerm(input_array);
-  const string perm_array_name = CreateInt32Array(
+  const std::string perm_array_name = CreateInt32Array(
       model, AvailableArrayName(*model, input + "/transpose/perm"), perm_array);
   auto* transpose_op = new TransposeOperator;
   transpose_op->inputs = {input, perm_array_name};
@@ -141,8 +142,8 @@ TransposeOperator* TransposeInput(const string& input, Model* model) {
       static_cast(batch_op_it->get());
   auto& tail_it = batch_op_it;
 
-  string input_lhs = batch_op->inputs[0];
-  string input_rhs = batch_op->inputs[1];
+  std::string input_lhs = batch_op->inputs[0];
+  std::string input_rhs = batch_op->inputs[1];
   const auto& input_lhs_array = model->GetArray(input_lhs);
   const auto& input_rhs_array = model->GetArray(input_rhs);
   if (!input_lhs_array.has_shape() || !input_rhs_array.has_shape())
@@ -195,19 +196,19 @@ TransposeOperator* TransposeInput(const string& input, Model* model) {
   }
   AddMessageF("Unrolling BatchMatMul %s %d times", LogName(*batch_op),
               bcast.output_batch_size());
-  string base_name = std::string(batch_op->outputs[0]);
+  std::string base_name = std::string(batch_op->outputs[0]);
 
   // Compute slices for each batch in the LHS and RHS.
-  std::vector slice_a_outputs =
+  std::vector slice_a_outputs =
       SliceInput(input_lhs, base_name, "a", bcast.x_batch_size(), input_array_a,
                  model, &tail_it);
-  std::vector slice_b_outputs =
+  std::vector slice_b_outputs =
       SliceInput(input_rhs, base_name, "b", bcast.y_batch_size(), input_array_b,
                  model, &tail_it);
 
   // Compute (single batch) MatMul for each output batch. The MatMul outputs are
   // then packed together into one output Tensor.
-  std::vector pack_inputs;
+  std::vector pack_inputs;
   for (int batch_idx = 0; batch_idx < bcast.output_batch_size(); ++batch_idx) {
     std::string batch_name =
         absl::StrCat(batch_op->outputs[0], "_b", batch_idx);

From f5eb8a40e98224d98be8ca4881c7261d2d04ad17 Mon Sep 17 00:00:00 2001
From: Christian Sigg 
Date: Mon, 22 Jun 2020 10:38:48 -0700
Subject: [PATCH 445/867] Change gpu_device_test to cope with CUDA stream
 priority supporting [-2,0] in CUDA 8 instead of [-1,0] before.

PiperOrigin-RevId: 317684869
Change-Id: Ib741cd1015f64358d323dda761250c990ce4ca41
---
 .../core/common_runtime/gpu/gpu_device_test.cc | 18 +++++++-----------
 1 file changed, 7 insertions(+), 11 deletions(-)

diff --git a/tensorflow/core/common_runtime/gpu/gpu_device_test.cc b/tensorflow/core/common_runtime/gpu/gpu_device_test.cc
index dae744380e9..6448fc56af7 100644
--- a/tensorflow/core/common_runtime/gpu/gpu_device_test.cc
+++ b/tensorflow/core/common_runtime/gpu/gpu_device_test.cc
@@ -234,9 +234,9 @@ TEST_F(GPUDeviceTest, SingleVirtualDeviceWithInvalidPriority) {
     SessionOptions opts =
         MakeSessionOptions("0", 0, 1, {{123, 456}}, {{-1, 2}});
 #else
-    // Priority outside the range (-1, 0) for NVidia GPUs
+    // Priority outside the range (-2, 0) for NVidia GPUs
     SessionOptions opts =
-        MakeSessionOptions("0", 0, 1, {{123, 456}}, {{-2, 0}});
+        MakeSessionOptions("0", 0, 1, {{123, 456}}, {{-3, 0}});
 #endif
     std::vector> devices;
     Status status = DeviceFactory::GetFactory("GPU")->CreateDevices(
@@ -249,9 +249,7 @@ TEST_F(GPUDeviceTest, SingleVirtualDeviceWithInvalidPriority) {
         " virtual device 0 on GPU# 0");
 #else
     ExpectErrorMessageSubstr(
-        status,
-        "Priority -2 is outside the range of supported priorities [-1,0] for"
-        " virtual device 0 on GPU# 0");
+        status, "Priority -3 is outside the range of supported priorities");
 #endif
   }
   {
@@ -259,7 +257,7 @@ TEST_F(GPUDeviceTest, SingleVirtualDeviceWithInvalidPriority) {
     // Priority outside the range (0, 2) for AMD GPUs
     SessionOptions opts = MakeSessionOptions("0", 0, 1, {{123, 456}}, {{0, 3}});
 #else
-    // Priority outside the range (-1, 0) for NVidia GPUs
+    // Priority outside the range (-2, 0) for NVidia GPUs
     SessionOptions opts = MakeSessionOptions("0", 0, 1, {{123, 456}}, {{0, 1}});
 #endif
     std::vector> devices;
@@ -273,9 +271,7 @@ TEST_F(GPUDeviceTest, SingleVirtualDeviceWithInvalidPriority) {
         " virtual device 0 on GPU# 0");
 #else
     ExpectErrorMessageSubstr(
-        status,
-        "Priority 1 is outside the range of supported priorities [-1,0] for"
-        " virtual device 0 on GPU# 0");
+        status, "Priority 1 is outside the range of supported priorities");
 #endif
   }
 }
@@ -296,7 +292,7 @@ TEST_F(GPUDeviceTest, MultipleVirtualDevices) {
   // Valid range for priority values on AMD GPUs in (0,2)
   SessionOptions opts = MakeSessionOptions("0", 0, 1, {{123, 456}}, {{0, 1}});
 #else
-  // Valid range for priority values on NVidia GPUs in (-1, 0)
+  // Valid range for priority values on NVidia GPUs in (-2, 0)
   SessionOptions opts = MakeSessionOptions("0", 0, 1, {{123, 456}}, {{0, -1}});
 #endif
   std::vector> devices;
@@ -347,7 +343,7 @@ TEST_F(GPUDeviceTest, MultipleVirtualDevicesWithPriority) {
     // Valid range for priority values on AMD GPUs in (0,2)
     SessionOptions opts = MakeSessionOptions("0", 0, 1, {{123, 456}}, {{2, 1}});
 #else
-    // Valid range for priority values on NVidia GPUs in (-1, 0)
+    // Valid range for priority values on NVidia GPUs in (-2, 0)
     SessionOptions opts =
         MakeSessionOptions("0", 0, 1, {{123, 456}}, {{-1, 0}});
 #endif

From 6010d444ad3b9a45cb9247616830e207bb17f820 Mon Sep 17 00:00:00 2001
From: Vo Van Nghia 
Date: Tue, 23 Jun 2020 02:10:52 +0700
Subject: [PATCH 446/867] Fix internal build

---
 tensorflow/c/experimental/filesystem/plugins/gcs/BUILD        | 3 ++-
 .../filesystem/plugins/gcs/gcs_filesystem_test.cc             | 4 +++-
 2 files changed, 5 insertions(+), 2 deletions(-)

diff --git a/tensorflow/c/experimental/filesystem/plugins/gcs/BUILD b/tensorflow/c/experimental/filesystem/plugins/gcs/BUILD
index 3a65824cd7c..395ce633370 100644
--- a/tensorflow/c/experimental/filesystem/plugins/gcs/BUILD
+++ b/tensorflow/c/experimental/filesystem/plugins/gcs/BUILD
@@ -61,6 +61,7 @@ tf_cc_test(
         "@com_google_absl//absl/strings",
     ] + [
         "//tensorflow/c:tf_status_helper",
-        "@com_google_googletest//:gtest",
+        "//tensorflow/core/platform:stacktrace_handler",
+        "//tensorflow/core/platform:test",
     ],
 )
diff --git a/tensorflow/c/experimental/filesystem/plugins/gcs/gcs_filesystem_test.cc b/tensorflow/c/experimental/filesystem/plugins/gcs/gcs_filesystem_test.cc
index 6dc5e164a1c..a4cb62a673b 100644
--- a/tensorflow/c/experimental/filesystem/plugins/gcs/gcs_filesystem_test.cc
+++ b/tensorflow/c/experimental/filesystem/plugins/gcs/gcs_filesystem_test.cc
@@ -12,9 +12,10 @@ WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
 See the License for the specific language governing permissions and
 limitations under the License.
 ==============================================================================*/
-#include "gtest/gtest.h"
 #include "tensorflow/c/experimental/filesystem/filesystem_interface.h"
 #include "tensorflow/c/tf_status_helper.h"
+#include "tensorflow/core/platform/stacktrace_handler.h"
+#include "tensorflow/core/platform/test.h"
 
 #define ASSERT_TF_OK(x) ASSERT_EQ(TF_OK, TF_GetCode(x))
 
@@ -54,6 +55,7 @@ TEST_F(GCSFilesystemTest, TestInit) { ASSERT_TF_OK(status); }
 }  // namespace tensorflow
 
 GTEST_API_ int main(int argc, char** argv) {
+  tensorflow::testing::InstallStacktraceHandler();
   ::testing::InitGoogleTest(&argc, argv);
   return RUN_ALL_TESTS();
 }

From c041b5de75463f76bf8d9461e0f79ea9ecec498f Mon Sep 17 00:00:00 2001
From: Chris Kennelly 
Date: Mon, 22 Jun 2020 10:55:18 -0700
Subject: [PATCH 447/867] Replace ARCH_K8 with __x86_64__.

PiperOrigin-RevId: 317689006
Change-Id: I7e47b17ef53b3cc223b64ff179fcdc3777c61eb7
---
 tensorflow/core/lib/gtl/manual_constructor_test.cc | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/tensorflow/core/lib/gtl/manual_constructor_test.cc b/tensorflow/core/lib/gtl/manual_constructor_test.cc
index 35cbc78b664..77824326a7d 100644
--- a/tensorflow/core/lib/gtl/manual_constructor_test.cc
+++ b/tensorflow/core/lib/gtl/manual_constructor_test.cc
@@ -92,7 +92,7 @@ TEST(ManualConstructorTest, Alignment) {
 
   EXPECT_EQ(reinterpret_cast(test2.b.get()) - &test2.a,
             reinterpret_cast(&control2.b) - &control2.a);
-#ifdef ARCH_K8
+#ifdef __x86_64__
   EXPECT_EQ(reinterpret_cast(test2.b.get()) % 16, 0);
 #endif
 }

From 18569dde74a866758b2008042a6ed2e3c8bdb027 Mon Sep 17 00:00:00 2001
From: Christian Sigg 
Date: Mon, 22 Jun 2020 11:08:36 -0700
Subject: [PATCH 448/867] Set up RBE builds with CUDA11, cuDNN8, TensorRT7.

PiperOrigin-RevId: 317692556
Change-Id: Ifd1e78cec6a6083e507e98baf6702401f18f67aa
---
 .bazelrc                                      | 80 +++++++++++--------
 .../preconfig/generate/containers.bzl         |  1 +
 .../toolchains/remote_config/configs.bzl      | 13 +++
 .../toolchains/remote_config/containers.bzl   |  7 ++
 4 files changed, 67 insertions(+), 34 deletions(-)

diff --git a/.bazelrc b/.bazelrc
index e67c3eecc3b..f11c376df65 100644
--- a/.bazelrc
+++ b/.bazelrc
@@ -58,13 +58,12 @@
 #
 #
 # Remote build execution options (only configured to work with TF team projects for now.)
-#     rbe:        General RBE options shared by all flavors.
-#     rbe_linux:  General RBE options used on all linux builds.
-#     rbe_win:    General RBE options used on all windows builds.
+#     rbe:       General RBE options shared by all flavors.
+#     rbe_linux: General RBE options used on all linux builds.
+#     rbe_win:   General RBE options used on all windows builds.
 #
-#     rbe_cpu_linux:        RBE options to build with only CPU support.
-#     rbe_linux_cuda_nvcc:  RBE options to build with GPU support using nvcc.
-#     rbe_gpu_linux:        An alias for rbe_linux_cuda_nvcc
+#     rbe_cpu_linux:           RBE options to build with only CPU support.
+#     rbe_linux_cuda_nvcc_py*: RBE options to build with GPU support using nvcc.
 #
 #     rbe_linux_py2: Linux Python 2 RBE config.
 #     rbe_linux_py3: Linux Python 3 RBE config
@@ -400,33 +399,48 @@ build:rbe_linux_cuda_base --repo_env=REMOTE_GPU_TESTING=1
 build:rbe_linux_cuda_base --repo_env=TF_NEED_CUDA=1
 test:rbe_linux_cuda_base --test_env=LD_LIBRARY_PATH="/usr/local/cuda/lib64:/usr/local/cuda/extras/CUPTI/lib64"
 
-build:rbe_linux_cuda_nvcc --config=rbe_linux_cuda_base
-build:rbe_linux_cuda_nvcc --crosstool_top="@ubuntu18.04-py3-gcc7_manylinux2010-cuda10.1-cudnn7-tensorrt6.0_config_cuda//crosstool:toolchain"
-build:rbe_linux_cuda_nvcc --extra_toolchains="@ubuntu18.04-py3-gcc7_manylinux2010-cuda10.1-cudnn7-tensorrt6.0_config_cuda//crosstool:toolchain-linux-x86_64"
-build:rbe_linux_cuda_nvcc --extra_execution_platforms="@ubuntu18.04-py3-gcc7_manylinux2010-cuda10.1-cudnn7-tensorrt6.0_config_platform//:platform"
-build:rbe_linux_cuda_nvcc --host_platform="@ubuntu18.04-py3-gcc7_manylinux2010-cuda10.1-cudnn7-tensorrt6.0_config_platform//:platform"
-build:rbe_linux_cuda_nvcc --platforms="@ubuntu18.04-py3-gcc7_manylinux2010-cuda10.1-cudnn7-tensorrt6.0_config_platform//:platform"
-build:rbe_linux_cuda_nvcc --repo_env=TF_CUDA_CONFIG_REPO="@ubuntu18.04-py3-gcc7_manylinux2010-cuda10.1-cudnn7-tensorrt6.0_config_cuda"
-build:rbe_linux_cuda_nvcc --repo_env=TF_TENSORRT_CONFIG_REPO="@ubuntu18.04-py3-gcc7_manylinux2010-cuda10.1-cudnn7-tensorrt6.0_config_tensorrt"
-build:rbe_linux_cuda_nvcc --repo_env=TF_NCCL_CONFIG_REPO="@ubuntu18.04-py3-gcc7_manylinux2010-cuda10.1-cudnn7-tensorrt6.0_config_nccl"
-build:rbe_linux_cuda_nvcc --define=using_cuda_nvcc=true
-test:rbe_linux_cuda_nvcc --config=rbe_linux_cuda_base
+build:rbe_linux_cuda10.1_nvcc_base --config=rbe_linux_cuda_base
+build:rbe_linux_cuda10.1_nvcc_base --define=using_cuda_nvcc=true
+build:rbe_linux_cuda10.1_nvcc_base --crosstool_top="@ubuntu18.04-gcc7_manylinux2010-cuda10.1-cudnn7-tensorrt6.0_config_cuda//crosstool:toolchain"
+build:rbe_linux_cuda10.1_nvcc_base --extra_toolchains="@ubuntu18.04-gcc7_manylinux2010-cuda10.1-cudnn7-tensorrt6.0_config_cuda//crosstool:toolchain-linux-x86_64"
+build:rbe_linux_cuda10.1_nvcc_base --extra_execution_platforms="@ubuntu18.04-gcc7_manylinux2010-cuda10.1-cudnn7-tensorrt6.0_config_platform//:platform"
+build:rbe_linux_cuda10.1_nvcc_base --host_platform="@ubuntu18.04-gcc7_manylinux2010-cuda10.1-cudnn7-tensorrt6.0_config_platform//:platform"
+build:rbe_linux_cuda10.1_nvcc_base --platforms="@ubuntu18.04-gcc7_manylinux2010-cuda10.1-cudnn7-tensorrt6.0_config_platform//:platform"
+build:rbe_linux_cuda10.1_nvcc_base --repo_env=TF_CUDA_CONFIG_REPO="@ubuntu18.04-gcc7_manylinux2010-cuda10.1-cudnn7-tensorrt6.0_config_cuda"
+build:rbe_linux_cuda10.1_nvcc_base --repo_env=TF_TENSORRT_CONFIG_REPO="@ubuntu18.04-gcc7_manylinux2010-cuda10.1-cudnn7-tensorrt6.0_config_tensorrt"
+build:rbe_linux_cuda10.1_nvcc_base --repo_env=TF_NCCL_CONFIG_REPO="@ubuntu18.04-gcc7_manylinux2010-cuda10.1-cudnn7-tensorrt6.0_config_nccl"
+build:rbe_linux_cuda10.1_nvcc_py2.7 --config=rbe_linux_cuda10.1_nvcc_base --repo_env=TF_PYTHON_CONFIG_REPO="@ubuntu18.04-gcc7_manylinux2010-cuda10.1-cudnn7-tensorrt6.0_config_python2.7"
+build:rbe_linux_cuda10.1_nvcc_py3.5 --config=rbe_linux_cuda10.1_nvcc_base --repo_env=TF_PYTHON_CONFIG_REPO="@ubuntu18.04-gcc7_manylinux2010-cuda10.1-cudnn7-tensorrt6.0_config_python3.5"
+build:rbe_linux_cuda10.1_nvcc_py3.6 --config=rbe_linux_cuda10.1_nvcc_base --repo_env=TF_PYTHON_CONFIG_REPO="@ubuntu18.04-gcc7_manylinux2010-cuda10.1-cudnn7-tensorrt6.0_config_python3.6"
+build:rbe_linux_cuda10.1_nvcc_py3.7 --config=rbe_linux_cuda10.1_nvcc_base --repo_env=TF_PYTHON_CONFIG_REPO="@ubuntu18.04-gcc7_manylinux2010-cuda10.1-cudnn7-tensorrt6.0_config_python3.7"
+build:rbe_linux_cuda10.1_nvcc_py3.8 --config=rbe_linux_cuda10.1_nvcc_base --repo_env=TF_PYTHON_CONFIG_REPO="@ubuntu18.04-gcc7_manylinux2010-cuda10.1-cudnn7-tensorrt6.0_config_python3.8"
 
-build:rbe_linux_cuda_nvcc_base --config=rbe_linux_cuda_base
-build:rbe_linux_cuda_nvcc_base --crosstool_top="@ubuntu18.04-gcc7_manylinux2010-cuda10.1-cudnn7-tensorrt6.0_config_cuda//crosstool:toolchain"
-build:rbe_linux_cuda_nvcc_base --extra_toolchains="@ubuntu18.04-gcc7_manylinux2010-cuda10.1-cudnn7-tensorrt6.0_config_cuda//crosstool:toolchain-linux-x86_64"
-build:rbe_linux_cuda_nvcc_base --extra_execution_platforms="@ubuntu18.04-gcc7_manylinux2010-cuda10.1-cudnn7-tensorrt6.0_config_platform//:platform"
-build:rbe_linux_cuda_nvcc_base --host_platform="@ubuntu18.04-gcc7_manylinux2010-cuda10.1-cudnn7-tensorrt6.0_config_platform//:platform"
-build:rbe_linux_cuda_nvcc_base --platforms="@ubuntu18.04-gcc7_manylinux2010-cuda10.1-cudnn7-tensorrt6.0_config_platform//:platform"
-build:rbe_linux_cuda_nvcc_base --repo_env=TF_CUDA_CONFIG_REPO="@ubuntu18.04-gcc7_manylinux2010-cuda10.1-cudnn7-tensorrt6.0_config_cuda"
-build:rbe_linux_cuda_nvcc_base --repo_env=TF_TENSORRT_CONFIG_REPO="@ubuntu18.04-gcc7_manylinux2010-cuda10.1-cudnn7-tensorrt6.0_config_tensorrt"
-build:rbe_linux_cuda_nvcc_base --repo_env=TF_NCCL_CONFIG_REPO="@ubuntu18.04-gcc7_manylinux2010-cuda10.1-cudnn7-tensorrt6.0_config_nccl"
-build:rbe_linux_cuda_nvcc_base --define=using_cuda_nvcc=true
-build:rbe_linux_cuda_nvcc_py27 --config=rbe_linux_cuda_nvcc_base --repo_env=TF_PYTHON_CONFIG_REPO="@ubuntu18.04-gcc7_manylinux2010-cuda10.1-cudnn7-tensorrt6.0_config_python2.7"
-build:rbe_linux_cuda_nvcc_py35 --config=rbe_linux_cuda_nvcc_base --repo_env=TF_PYTHON_CONFIG_REPO="@ubuntu18.04-gcc7_manylinux2010-cuda10.1-cudnn7-tensorrt6.0_config_python3.5"
-build:rbe_linux_cuda_nvcc_py36 --config=rbe_linux_cuda_nvcc_base --repo_env=TF_PYTHON_CONFIG_REPO="@ubuntu18.04-gcc7_manylinux2010-cuda10.1-cudnn7-tensorrt6.0_config_python3.6"
-build:rbe_linux_cuda_nvcc_py37 --config=rbe_linux_cuda_nvcc_base --repo_env=TF_PYTHON_CONFIG_REPO="@ubuntu18.04-gcc7_manylinux2010-cuda10.1-cudnn7-tensorrt6.0_config_python3.7"
-build:rbe_linux_cuda_nvcc_py38 --config=rbe_linux_cuda_nvcc_base --repo_env=TF_PYTHON_CONFIG_REPO="@ubuntu18.04-gcc7_manylinux2010-cuda10.1-cudnn7-tensorrt6.0_config_python3.8"
+build:rbe_linux_cuda11.0_nvcc_base --config=rbe_linux_cuda_base
+build:rbe_linux_cuda11.0_nvcc_base --define=using_cuda_nvcc=true
+build:rbe_linux_cuda11.0_nvcc_base --crosstool_top="@ubuntu18.04-gcc7_manylinux2010-cuda11.0-cudnn8-tensorrt7.1_config_cuda//crosstool:toolchain"
+build:rbe_linux_cuda11.0_nvcc_base --extra_toolchains="@ubuntu18.04-gcc7_manylinux2010-cuda11.0-cudnn8-tensorrt7.1_config_cuda//crosstool:toolchain-linux-x86_64"
+build:rbe_linux_cuda11.0_nvcc_base --extra_execution_platforms="@ubuntu18.04-gcc7_manylinux2010-cuda11.0-cudnn8-tensorrt7.1_config_platform//:platform"
+build:rbe_linux_cuda11.0_nvcc_base --host_platform="@ubuntu18.04-gcc7_manylinux2010-cuda11.0-cudnn8-tensorrt7.1_config_platform//:platform"
+build:rbe_linux_cuda11.0_nvcc_base --platforms="@ubuntu18.04-gcc7_manylinux2010-cuda11.0-cudnn8-tensorrt7.1_config_platform//:platform"
+build:rbe_linux_cuda11.0_nvcc_base --repo_env=TF_CUDA_CONFIG_REPO="@ubuntu18.04-gcc7_manylinux2010-cuda11.0-cudnn8-tensorrt7.1_config_cuda"
+build:rbe_linux_cuda11.0_nvcc_base --repo_env=TF_TENSORRT_CONFIG_REPO="@ubuntu18.04-gcc7_manylinux2010-cuda11.0-cudnn8-tensorrt7.1_config_tensorrt"
+build:rbe_linux_cuda11.0_nvcc_base --repo_env=TF_NCCL_CONFIG_REPO="@ubuntu18.04-gcc7_manylinux2010-cuda11.0-cudnn8-tensorrt7.1_config_nccl"
+build:rbe_linux_cuda11.0_nvcc_py2.7 --config=rbe_linux_cuda11.0_nvcc_base --repo_env=TF_PYTHON_CONFIG_REPO="@ubuntu18.04-gcc7_manylinux2010-cuda11.0-cudnn8-tensorrt7.1_config_python2.7"
+build:rbe_linux_cuda11.0_nvcc_py3.5 --config=rbe_linux_cuda11.0_nvcc_base --repo_env=TF_PYTHON_CONFIG_REPO="@ubuntu18.04-gcc7_manylinux2010-cuda11.0-cudnn8-tensorrt7.1_config_python3.5"
+build:rbe_linux_cuda11.0_nvcc_py3.6 --config=rbe_linux_cuda11.0_nvcc_base --repo_env=TF_PYTHON_CONFIG_REPO="@ubuntu18.04-gcc7_manylinux2010-cuda11.0-cudnn8-tensorrt7.1_config_python3.6"
+build:rbe_linux_cuda11.0_nvcc_py3.7 --config=rbe_linux_cuda11.0_nvcc_base --repo_env=TF_PYTHON_CONFIG_REPO="@ubuntu18.04-gcc7_manylinux2010-cuda11.0-cudnn8-tensorrt7.1_config_python3.7"
+build:rbe_linux_cuda11.0_nvcc_py3.8 --config=rbe_linux_cuda11.0_nvcc_base --repo_env=TF_PYTHON_CONFIG_REPO="@ubuntu18.04-gcc7_manylinux2010-cuda11.0-cudnn8-tensorrt7.1_config_python3.8"
+
+# Map default to CUDA 10.1.
+build:rbe_linux_cuda_nvcc_py27 --config=rbe_linux_cuda10.1_nvcc_py2.7
+build:rbe_linux_cuda_nvcc_py35 --config=rbe_linux_cuda10.1_nvcc_py3.5
+build:rbe_linux_cuda_nvcc_py36 --config=rbe_linux_cuda10.1_nvcc_py3.6
+build:rbe_linux_cuda_nvcc_py37 --config=rbe_linux_cuda10.1_nvcc_py3.7
+build:rbe_linux_cuda_nvcc_py38 --config=rbe_linux_cuda10.1_nvcc_py3.8
+
+# Deprecated configs that people might still use.
+build:rbe_linux_cuda_nvcc --config=rbe_linux_cuda_nvcc_py36
+build:rbe_gpu_linux       --config=rbe_linux_cuda_nvcc
 
 build:rbe_linux_cuda_clang_base --config=rbe_linux_cuda_base
 build:rbe_linux_cuda_clang_base --crosstool_top="@ubuntu16.04-clang_manylinux2010-cuda10.1-cudnn7-tensorrt6.0_config_cuda//crosstool:toolchain"
@@ -444,8 +458,6 @@ build:rbe_linux_cuda_clang_py36 --config=rbe_linux_cuda_clang_base --repo_env=TF
 build:rbe_linux_cuda_clang_py37 --config=rbe_linux_cuda_clang_base --repo_env=TF_PYTHON_CONFIG_REPO="@ubuntu16.04-clang_manylinux2010-cuda10.1-cudnn7-tensorrt6.0_config_python3.7"
 build:rbe_linux_cuda_clang_py38 --config=rbe_linux_cuda_clang_base --repo_env=TF_PYTHON_CONFIG_REPO="@ubuntu16.04-clang_manylinux2010-cuda10.1-cudnn7-tensorrt6.0_config_python3.8"
 
-common:rbe_gpu_linux --config=rbe_linux_cuda_nvcc
-
 build:rbe_linux_py2 --config=rbe_linux
 build:rbe_linux_py2 --repo_env=PYTHON_BIN_PATH="/usr/bin/python2"
 build:rbe_linux_py2 --python_path="/usr/bin/python2"
diff --git a/third_party/toolchains/preconfig/generate/containers.bzl b/third_party/toolchains/preconfig/generate/containers.bzl
index 4a3ed0b5225..260b7c31717 100644
--- a/third_party/toolchains/preconfig/generate/containers.bzl
+++ b/third_party/toolchains/preconfig/generate/containers.bzl
@@ -11,6 +11,7 @@ container_digests = {
     "cuda10.1-cudnn7-ubuntu16.04-manylinux2010": "sha256:5e6d21c8ef226316eb6df5e2e6015244c16a8e5d936b52a09820442d2f8a919f",
     "cuda10.1-cudnn7-ubuntu16.04-manylinux2010-multipython": "sha256:3f890a951c81a201d60d0161a56ce628a90323be0c7f795550caa37f6f41a85c",
     "cuda10.1-cudnn7-ubuntu18.04-manylinux2010-multipython": "sha256:bd7666d1ef49b2b2e2a64981f1c9234deeccdb0d5198b30ff4289c3dfcffedbf",
+    "cuda11.0-cudnn8-ubuntu18.04-manylinux2010-multipython": "sha256:b52edb4e35c780334ba417b008927722ae668847715a1624e9b2984e99c05338",
     "rocm-ubuntu16.04": "sha256:e645447dd6127325f3e97b8bf23424f637a8579d963b34fcc6772cf7cfaa0ebe",
     "windows-1803": "sha256:f109576c7c0c8a1783ff22b666e8923b52dbbe7933f69a1c7a7275202c304a12",
 }
diff --git a/third_party/toolchains/remote_config/configs.bzl b/third_party/toolchains/remote_config/configs.bzl
index 3bbf99e0e36..501759dcd50 100644
--- a/third_party/toolchains/remote_config/configs.bzl
+++ b/third_party/toolchains/remote_config/configs.bzl
@@ -52,6 +52,19 @@ def initialize_rbe_configs():
         python_install_path = "/usr/local",
     )
 
+    tensorflow_rbe_config(
+        name = "ubuntu18.04-gcc7_manylinux2010-cuda11.0-cudnn8-tensorrt7.1",
+        compiler = "/dt7/usr/bin/gcc",
+        compiler_prefix = "/usr/bin",
+        cuda_version = "11.0",
+        cudnn_version = "8",
+        os = "ubuntu18.04-manylinux2010-multipython",
+        python_versions = ["2.7", "3.5", "3.6", "3.7", "3.8"],
+        tensorrt_install_path = "/usr",
+        tensorrt_version = "7.1",
+        python_install_path = "/usr/local",
+    )
+
     # TODO(klimek): Delete this once all users are migrated to a python-version
     # independent configuration. In the future, use
     # "ubuntu16.04-gcc7_manylinux2010-cuda10.1-cudnn7-tensorrt6.0" instead.
diff --git a/third_party/toolchains/remote_config/containers.bzl b/third_party/toolchains/remote_config/containers.bzl
index 8f6dae7f311..d94bce91675 100644
--- a/third_party/toolchains/remote_config/containers.bzl
+++ b/third_party/toolchains/remote_config/containers.bzl
@@ -38,6 +38,13 @@ containers = {
         "digest": container_digests["cuda10.1-cudnn7-ubuntu18.04-manylinux2010-multipython"],
     },
 
+    # Built with //tensorflow/tools/ci_build/Dockerfile.rbe.cuda11.0-cudnn8-ubuntu18.04-manylinux2010-multipython.
+    "cuda11.0-cudnn8-ubuntu18.04-manylinux2010-multipython": {
+        "registry": "gcr.io",
+        "repository": "tensorflow-testing/nosla-cuda11.0-cudnn8-ubuntu18.04-manylinux2010-multipython",
+        "digest": container_digests["cuda11.0-cudnn8-ubuntu18.04-manylinux2010-multipython"],
+    },
+
     # Built with //tensorflow/tools/ci_build/Dockerfile.rbe.rocm-ubuntu16.04
     "rocm-ubuntu16.04": {
         "registry": "gcr.io",

From 8785b4f5b8701c3aab07695e47b8ff036f41e672 Mon Sep 17 00:00:00 2001
From: Christian Sigg 
Date: Mon, 22 Jun 2020 11:32:54 -0700
Subject: [PATCH 449/867] Make gpu_ftz_test ready for CUDA 11.

CUDA 11 changed the libdevice implementation of expf from previously two to one ex2.approx.ftz. Change the CHECK directives to handle both cases.

PiperOrigin-RevId: 317698875
Change-Id: Idd7b28c77427f299b80fa1b7f4b9be8c7881f963
---
 .../xla/service/gpu/tests/gpu_ftz_test.cc       | 17 +++++++----------
 1 file changed, 7 insertions(+), 10 deletions(-)

diff --git a/tensorflow/compiler/xla/service/gpu/tests/gpu_ftz_test.cc b/tensorflow/compiler/xla/service/gpu/tests/gpu_ftz_test.cc
index 282f7b24a31..9b58457d129 100644
--- a/tensorflow/compiler/xla/service/gpu/tests/gpu_ftz_test.cc
+++ b/tensorflow/compiler/xla/service/gpu/tests/gpu_ftz_test.cc
@@ -92,26 +92,23 @@ TEST_F(GpuFtzDisabledTest, MultiplyFtz) {
 }
 
 // In NVPTX, exp(float) is implemented in libdevice, and consults __nvvm_reflect
-// to determine whether or not ftz is enabled.  The implementation uses two
-// calls to ex2.approx.  When ftz is on, we get two calls to the ftz version;
-// when ftz is off, we get one call to the ftz version and one call to the
-// regular version.
+// to determine whether or not ftz is enabled.
+// The implementation in CUDA 11 uses one ex2.approx.ftz, irrespective of ftz
+// being enabled or not. In previous CUDA versions, there is a leading
+// ex2.approx that does obey the ftz setting.
+// Instead of pattern matching implementation details, it might be better to
+// value-test the actual result instead. TODO(csigg): change to value-test.
 TEST_F(GpuFtzEnabledTest, ExpFtz) {
   CompileAndOptionallyVerifyPtx(CreateUnaryOpModule(HloOpcode::kExp), R"(
     CHECK-NOT: ex2.approx.f32
     CHECK:     ex2.approx.ftz.f32
     CHECK-NOT: ex2.approx.f32
-    CHECK:     ex2.approx.ftz.f32
-    CHECK-NOT: ex2.approx.f32
-    CHECK-NOT: ex2.approx.ftz.f32
   )");
 }
 
 TEST_F(GpuFtzDisabledTest, ExpFtz) {
   CompileAndOptionallyVerifyPtx(CreateUnaryOpModule(HloOpcode::kExp), R"(
-    CHECK-NOT: ex2.approx.f32
-    CHECK-DAG: ex2.approx.ftz.f32
-    CHECK-DAG: ex2.approx.f32
+    CHECK:     ex2.approx.ftz.f32
     CHECK-NOT: ex2.approx.f32
     CHECK-NOT: ex2.approx.ftz.f32
   )");

From 79518facb4b857af9d9d5df2da463fdbf7eb0e3e Mon Sep 17 00:00:00 2001
From: Mihai Maruseac 
Date: Mon, 22 Jun 2020 11:37:07 -0700
Subject: [PATCH 450/867] Upper bound `numpy` dependency to `1.19.0`.

Since `numpy==1.19.0` contains at least one breaking ABI change (numpy/numpy#15355), we need to either upper bound the dependency's range or fix our code to support both ABIs. Since the second requires more changes, we prefer the first one for now.

PiperOrigin-RevId: 317699730
Change-Id: Ia62a779f9ec42d63d3fac1b69cd75e6084358d2f
---
 tensorflow/tools/pip_package/setup.py | 4 +++-
 1 file changed, 3 insertions(+), 1 deletion(-)

diff --git a/tensorflow/tools/pip_package/setup.py b/tensorflow/tools/pip_package/setup.py
index 92061b396ce..42233f80c1c 100644
--- a/tensorflow/tools/pip_package/setup.py
+++ b/tensorflow/tools/pip_package/setup.py
@@ -58,7 +58,9 @@ REQUIRED_PACKAGES = [
     'google_pasta >= 0.1.8',
     'h5py >= 2.10.0, < 2.11.0',
     'keras_preprocessing >= 1.1.1, < 1.2',
-    'numpy >= 1.16.0, < 2.0',
+    # TODO(mihaimaruseac): numpy 1.19.0 has ABI breakage
+    # https://github.com/numpy/numpy/pull/15355
+    'numpy >= 1.16.0, < 1.19.0',
     'opt_einsum >= 2.3.2',
     'protobuf >= 3.9.2',
     'tensorboard >= 2.2.0, < 2.3.0',

From 44067f0783c56ad092f6ef5ea1034e6926559d86 Mon Sep 17 00:00:00 2001
From: Davide Libenzi 
Date: Mon, 22 Jun 2020 11:41:43 -0700
Subject: [PATCH 451/867] Make XRT CPU/GPU use MaybeOwning buffer interface, so
 the new copy protection CL won't break aliasing.

PiperOrigin-RevId: 317700747
Change-Id: Ie7b5bb1989cd4359b30ad86a450de5bff0962c31
---
 .../xla/service/cpu/cpu_executable.cc         |  11 +-
 tensorflow/compiler/xla/service/executable.cc |  47 ++++
 tensorflow/compiler/xla/service/executable.h  |  38 +--
 .../xla/service/maybe_owning_device_memory.cc |   8 +
 .../xla/service/maybe_owning_device_memory.h  |   4 +
 tensorflow/compiler/xla/shape_util.cc         |  32 ++-
 tensorflow/compiler/xla/shape_util.h          |   6 +-
 tensorflow/compiler/xrt/BUILD                 |   1 +
 .../compiler/xrt/kernels/xrt_execute_op.cc    | 255 ++++++++----------
 tensorflow/compiler/xrt/xrt_util.cc           | 134 +++++++++
 tensorflow/compiler/xrt/xrt_util.h            |  21 ++
 11 files changed, 385 insertions(+), 172 deletions(-)

diff --git a/tensorflow/compiler/xla/service/cpu/cpu_executable.cc b/tensorflow/compiler/xla/service/cpu/cpu_executable.cc
index 4552d7b5ba9..d095d220b97 100644
--- a/tensorflow/compiler/xla/service/cpu/cpu_executable.cc
+++ b/tensorflow/compiler/xla/service/cpu/cpu_executable.cc
@@ -299,12 +299,11 @@ StatusOr CpuExecutable::ExecuteAsyncOnStream(
       const Shape& expected_shape =
           entry_comp->parameter_instruction(i)->shape();
       const Shape& actual_shape = arguments[i].Buffers().shape();
-      CHECK(
-          Shape::Equal().IgnoreDynamicDimension()(expected_shape, actual_shape))
-          << absl::StreamFormat(
-                 "Shape mismatch on argument %d.  Expected %s, but was %s.", i,
-                 expected_shape.ToString(/*print_layout=*/true),
-                 actual_shape.ToString(/*print_layout=*/true));
+      TF_RET_CHECK(
+          ShapeUtil::DynamicShapeIsCompatible(actual_shape, expected_shape))
+          << "Shape mismatch on argument " << i << ", "
+          << expected_shape.ToString(/*print_layout=*/true) << " vs. "
+          << actual_shape.ToString(/*print_layout=*/true);
     }
   }
 
diff --git a/tensorflow/compiler/xla/service/executable.cc b/tensorflow/compiler/xla/service/executable.cc
index ebf7cc440dd..61ce6200a28 100644
--- a/tensorflow/compiler/xla/service/executable.cc
+++ b/tensorflow/compiler/xla/service/executable.cc
@@ -28,10 +28,57 @@ limitations under the License.
 #include "tensorflow/core/lib/io/path.h"
 #include "tensorflow/core/lib/strings/proto_serialization.h"
 #include "tensorflow/core/platform/env.h"
+#include "tensorflow/core/platform/errors.h"
 #include "tensorflow/stream_executor/device_description.h"
 
 namespace xla {
 
+ExecutionInput::~ExecutionInput() {
+  for (auto& index : unowned_indices_) {
+    auto buffer = buffers_.mutable_element(index)->Release();
+    if (buffer) {
+      buffer->Release();
+    }
+  }
+}
+
+Status ExecutionInput::SetDynamicShape(Shape dynamic_shape) {
+  const Shape& input_shape = shape();
+  if (!ShapeUtil::DynamicShapeIsCompatible(input_shape, dynamic_shape)) {
+    return tensorflow::errors::InvalidArgument(
+        "Cannot set dynamic shape: ", input_shape.DebugString(), " vs. ",
+        dynamic_shape.DebugString());
+  }
+  dynamic_shape_ = absl::make_unique(std::move(dynamic_shape));
+  return Status::OK();
+}
+
+void ExecutionInput::SetUnownedBuffer(const ShapeIndex& index,
+                                      MaybeOwningDeviceMemory buffer) {
+  *buffers_.mutable_element(index) = std::move(buffer);
+  unowned_indices_.insert(index);
+}
+
+xla::StatusOr ExecutionInput::ToShapedBuffer(
+    se::DeviceMemoryAllocator* allocator, int device_ordinal) const {
+  const Shape& input_shape = shape();
+  xla::ShapedBuffer shaped_buffer(input_shape, input_shape,
+                                  allocator->platform(), device_ordinal);
+  for (const auto& index_buffer : Buffers()) {
+    const tensorflow::se::OwningDeviceMemory* mem =
+        index_buffer.second.AsOwningDeviceMemory();
+    if (mem != nullptr && (mem->allocator() != allocator ||
+                           mem->device_ordinal() != device_ordinal)) {
+      return tensorflow::errors::InvalidArgument(
+          "Device buffer at index ", index_buffer.first.ToString(),
+          " has mismatching allocator/device");
+    }
+    shaped_buffer.set_buffer(index_buffer.second.AsDeviceMemoryBase(),
+                             index_buffer.first);
+  }
+  return std::move(shaped_buffer);
+}
+
 StatusOr Executable::ExecuteOnStream(
     const ServiceExecutableRunOptions* run_options,
     absl::Span arguments,
diff --git a/tensorflow/compiler/xla/service/executable.h b/tensorflow/compiler/xla/service/executable.h
index 2c979662d24..6881f6dd68a 100644
--- a/tensorflow/compiler/xla/service/executable.h
+++ b/tensorflow/compiler/xla/service/executable.h
@@ -17,6 +17,7 @@ limitations under the License.
 #define TENSORFLOW_COMPILER_XLA_SERVICE_EXECUTABLE_H_
 
 #include 
+#include 
 #include 
 #include 
 
@@ -65,31 +66,32 @@ class ExecutionInput {
       : buffers_(std::move(buffers)) {}
   ExecutionInput(ExecutionInput&&) = default;
 
-  ~ExecutionInput() {
-    for (auto& index : unowned_indices_) {
-      auto buffer = buffers_.mutable_element(index)->Release();
-      if (buffer) {
-        buffer->Release();
-      }
-    }
-  }
+  ~ExecutionInput();
 
   ExecutionInput& operator=(ExecutionInput&&) = default;
 
-  const Shape& shape() const { return buffers_.shape(); }
+  const Shape& shape() const {
+    return dynamic_shape_ != nullptr ? *dynamic_shape_ : buffers_.shape();
+  }
+
+  Status SetDynamicShape(Shape dynamic_shape);
+
+  xla::StatusOr ToShapedBuffer(
+      se::DeviceMemoryAllocator* allocator, int device_ordinal) const;
 
   void SetBuffer(const ShapeIndex& index, MaybeOwningDeviceMemory buffer) {
     *buffers_.mutable_element(index) = std::move(buffer);
   }
 
   void SetUnownedBuffer(const ShapeIndex& index,
-                        MaybeOwningDeviceMemory buffer) {
-    *buffers_.mutable_element(index) = std::move(buffer);
-    unowned_indices_.push_back(index);
-  }
+                        MaybeOwningDeviceMemory buffer);
 
   void SetUnownedIndex(const ShapeIndex& index) {
-    unowned_indices_.push_back(index);
+    unowned_indices_.insert(index);
+  }
+
+  void ClearUnownedIndex(const ShapeIndex& index) {
+    unowned_indices_.erase(index);
   }
 
   const ShapeTree& Buffers() const { return buffers_; }
@@ -106,9 +108,10 @@ class ExecutionInput {
 
  private:
   ShapeTree buffers_;
-  // (Unordered) set of indices of buffers that should be returned to the
-  // caller if an error occurs when enqueuing the computation.
-  std::vector unowned_indices_;
+  // Set of indices of buffers that should be returned to the caller if an error
+  // occurs when enqueuing the computation.
+  std::set unowned_indices_;
+  std::unique_ptr dynamic_shape_;
 };
 
 // ExecutionOutput encapsulates the output buffers of a execution and the
@@ -145,7 +148,6 @@ class ExecutionOutput {
     to_be_released_.push_back(std::move(mem));
   }
 
-
   // Should be called once it is known that the execute operation succeeded,
   // before returning the ExecutionOutput to the caller.
   ExecutionOutput& Commit() {
diff --git a/tensorflow/compiler/xla/service/maybe_owning_device_memory.cc b/tensorflow/compiler/xla/service/maybe_owning_device_memory.cc
index c4bf48bcc00..c7505f5fa4a 100644
--- a/tensorflow/compiler/xla/service/maybe_owning_device_memory.cc
+++ b/tensorflow/compiler/xla/service/maybe_owning_device_memory.cc
@@ -14,7 +14,9 @@ limitations under the License.
 ==============================================================================*/
 
 #include "tensorflow/compiler/xla/service/maybe_owning_device_memory.h"
+
 #include "absl/types/variant.h"
+
 namespace xla {
 
 tensorflow::se::DeviceMemoryBase MaybeOwningDeviceMemory::AsDeviceMemoryBase()
@@ -38,4 +40,10 @@ MaybeOwningDeviceMemory::Release() {
   return std::move(absl::get(mem_));
 }
 
+const tensorflow::se::OwningDeviceMemory*
+MaybeOwningDeviceMemory::AsOwningDeviceMemory() const {
+  return HasOwnership() ? &absl::get(mem_)
+                        : nullptr;
+}
+
 }  // namespace xla
diff --git a/tensorflow/compiler/xla/service/maybe_owning_device_memory.h b/tensorflow/compiler/xla/service/maybe_owning_device_memory.h
index 7d23d178130..0b56fed0a72 100644
--- a/tensorflow/compiler/xla/service/maybe_owning_device_memory.h
+++ b/tensorflow/compiler/xla/service/maybe_owning_device_memory.h
@@ -57,6 +57,10 @@ class MaybeOwningDeviceMemory {
   // A nullopt is returned if the HasOwnership() == false;
   absl::optional Release();
 
+  // If the device memory is owned, returns a pointer to the internal
+  // OwningDeviceMemory, otherwise nullptr is returned.
+  const tensorflow::se::OwningDeviceMemory* AsOwningDeviceMemory() const;
+
   // Returns true if the device_memory has ownership over underlying memory.
   bool HasOwnership() const;
 
diff --git a/tensorflow/compiler/xla/shape_util.cc b/tensorflow/compiler/xla/shape_util.cc
index ab46e49b181..bce40578132 100644
--- a/tensorflow/compiler/xla/shape_util.cc
+++ b/tensorflow/compiler/xla/shape_util.cc
@@ -1461,7 +1461,7 @@ ShapeUtil::ReshapeLeavesDimensionsUnmodified(
   return shape;
 }
 
-/* static */ bool ShapeUtil::DynamicShapeIsCompatible(
+/* static */ bool ShapeUtil::DynamicArrayShapeIsCompatible(
     const xla::Shape& dynamic_shape, const xla::Shape& bounded_shape) {
   if (dynamic_shape.rank() != bounded_shape.rank()) {
     return false;
@@ -1474,6 +1474,36 @@ ShapeUtil::ReshapeLeavesDimensionsUnmodified(
   return true;
 }
 
+/* static */ bool ShapeUtil::DynamicShapeIsCompatible(
+    const xla::Shape& dynamic_shape, const xla::Shape& bounded_shape) {
+  bool compatible = true;
+  xla::ShapeUtil::ForEachSubshape(dynamic_shape, [&](const Shape& sub_shape,
+                                                     const ShapeIndex& index) {
+    if (compatible) {
+      auto subshape_result = TryGetSubshape(bounded_shape, index);
+      if (subshape_result.ok()) {
+        const Shape* bounded_sub_shape = subshape_result.ConsumeValueOrDie();
+        if (sub_shape.IsTuple()) {
+          if (!bounded_sub_shape->IsTuple()) {
+            compatible = false;
+          }
+        } else {
+          if (bounded_sub_shape->IsTuple()) {
+            compatible = false;
+          } else if (!sub_shape.is_static() &&
+                     !DynamicArrayShapeIsCompatible(sub_shape,
+                                                    *bounded_sub_shape)) {
+            compatible = false;
+          }
+        }
+      } else {
+        compatible = false;
+      }
+    }
+  });
+  return compatible;
+}
+
 /* static */ Shape ShapeUtil::FilterDimensions(
     const std::function& p, Shape shape) {
   CHECK(shape.IsArray());
diff --git a/tensorflow/compiler/xla/shape_util.h b/tensorflow/compiler/xla/shape_util.h
index dde56587482..fe1a8acf6e4 100644
--- a/tensorflow/compiler/xla/shape_util.h
+++ b/tensorflow/compiler/xla/shape_util.h
@@ -657,7 +657,11 @@ class ShapeUtil {
                                 Shape shape);
 
   // Returns true if `dynamic_shape` has dimensions that are less-equal to the
-  // "bounded_shape".
+  // "bounded_shape". Shapes must be arrays.
+  static bool DynamicArrayShapeIsCompatible(const xla::Shape& dynamic_shape,
+                                            const xla::Shape& bounded_shape);
+
+  // Same as DynamicArrayShapeIsCompatible() but supports tuples.
   static bool DynamicShapeIsCompatible(const xla::Shape& dynamic_shape,
                                        const xla::Shape& bounded_shape);
 
diff --git a/tensorflow/compiler/xrt/BUILD b/tensorflow/compiler/xrt/BUILD
index 332c8ff9a14..6a704be4adb 100644
--- a/tensorflow/compiler/xrt/BUILD
+++ b/tensorflow/compiler/xrt/BUILD
@@ -74,6 +74,7 @@ cc_library(
         "//tensorflow/compiler/xla/client:local_client",
         "//tensorflow/compiler/xla/service:backend",
         "//tensorflow/compiler/xla/service:executable",
+        "//tensorflow/compiler/xla/service:hlo",
         "//tensorflow/compiler/xla/service:shaped_buffer",
         "//tensorflow/core:core_cpu_internal",
         "//tensorflow/core:framework",
diff --git a/tensorflow/compiler/xrt/kernels/xrt_execute_op.cc b/tensorflow/compiler/xrt/kernels/xrt_execute_op.cc
index 2fc599e42df..3bd8af577c8 100644
--- a/tensorflow/compiler/xrt/kernels/xrt_execute_op.cc
+++ b/tensorflow/compiler/xrt/kernels/xrt_execute_op.cc
@@ -51,12 +51,6 @@ namespace tensorflow {
 
 namespace {
 
-struct InputBuffers {
-  std::vector> input_tuples;
-  std::vector input_allocations;
-  std::vector input_pointers;
-};
-
 uint32 InitialRandomSeed() {
   // Support plumbing the TF seed through to XLA is being worked on.
   // If a user wants deterministic behavior, their best option
@@ -80,75 +74,51 @@ uint32 GetXLARandomSeed() {
   return counter.fetch_add(2);
 }
 
-xla::StatusOr GetInputBuffers(
-    XRTMemoryManager::WorkingSet* working_set, xla::Backend* backend,
-    const std::vector& input_coords, bool release_inputs) {
-  InputBuffers input_buffers;
-  input_buffers.input_tuples.reserve(input_coords.size());
-  input_buffers.input_allocations.reserve(input_coords.size());
-  input_buffers.input_pointers.reserve(input_coords.size());
-  for (size_t i = 0; i < input_coords.size(); ++i) {
-    TF_RETURN_IF_ERROR(
-        working_set->LookupAndPin(backend, input_coords[i].handle));
-    auto tuple = working_set->PinnedTuples().back();
-    input_buffers.input_tuples.emplace_back(tuple);
-    if (release_inputs) {
-      // We are holding a reference to the tuple, so we can safely delete it
-      // from the resource manager here.
-      TF_RETURN_IF_ERROR(
-          working_set->MemoryManager()->Release(input_coords[i].handle));
-      VLOG(2) << "Released allocation handle " << input_coords[i].handle;
-    }
-    if (input_coords[i].index.empty()) {
-      TF_ASSIGN_OR_RETURN(xla::ShapedBuffer shaped_buffer,
-                          tuple->ToShapedBuffer());
-      input_buffers.input_allocations.emplace_back(std::move(shaped_buffer));
-    } else {
-      TF_ASSIGN_OR_RETURN(xla::ShapedBuffer shaped_buffer,
-                          tuple->ToShapedBuffer());
-      TF_ASSIGN_OR_RETURN(xla::ShapedBuffer sub_shaped_buffer,
-                          shaped_buffer.SubShapedBuffer(input_coords[i].index));
-      input_buffers.input_allocations.emplace_back(
-          std::move(sub_shaped_buffer));
-    }
+std::vector GetDynamicInputInfo(
+    const xla::ComputationLayout& computation_layout) {
+  std::vector input_is_dynamic;
+  input_is_dynamic.reserve(computation_layout.parameter_count());
+  for (int64 i = 0; i < computation_layout.parameter_count(); ++i) {
+    input_is_dynamic.push_back(
+        !computation_layout.parameter_shape(i).is_static());
   }
-  for (size_t i = 0; i < input_buffers.input_allocations.size(); ++i) {
-    input_buffers.input_pointers.push_back(&input_buffers.input_allocations[i]);
-  }
-  return std::move(input_buffers);
+  return input_is_dynamic;
 }
 
-xla::StatusOr GetChainedOpInputs(
+xla::StatusOr>> GetInputTuples(
+    xla::LocalExecutable* executable, XRTMemoryManager::WorkingSet* working_set,
+    xla::Backend* backend, const std::vector& input_coords,
+    bool release_inputs) {
+  const xla::ComputationLayout& computation_layout =
+      executable->executable()->module_config().entry_computation_layout();
+
+  return GetInputTupleAllocations(
+      input_coords, working_set, backend, computation_layout.parameter_count(),
+      [&](int64 i) { return computation_layout.parameter_shape(i); },
+      release_inputs);
+}
+
+xla::StatusOr>> GetChainedOpInputTuples(
     const xrt::XRTChainedExecuteOp& op,
     absl::Span> op_inputs) {
-  InputBuffers input_buffers;
-  input_buffers.input_tuples.reserve(op.inputs_size());
-  input_buffers.input_allocations.reserve(op.inputs_size());
-  input_buffers.input_pointers.reserve(op.inputs_size());
+  std::vector> input_tuples;
+  input_tuples.reserve(op.inputs_size());
   for (int i = 0; i < op.inputs_size(); ++i) {
     auto& input = op.inputs(i);
-    input_buffers.input_tuples.emplace_back(op_inputs[i]);
     // Thanks to the greatness of proto3, there is no way to query for
     // explicitly set fields, so the default for output_index (zero) means no
     // sub-index. As consequence, the real index is output_index - 1.
     if (input.output_index() == 0) {
-      TF_ASSIGN_OR_RETURN(xla::ShapedBuffer shaped_buffer,
-                          input_buffers.input_tuples.back()->ToShapedBuffer());
-      input_buffers.input_allocations.emplace_back(std::move(shaped_buffer));
+      input_tuples.emplace_back(op_inputs[i]);
     } else {
-      TF_ASSIGN_OR_RETURN(xla::ShapedBuffer shaped_buffer,
-                          input_buffers.input_tuples.back()->ToShapedBuffer());
-      TF_ASSIGN_OR_RETURN(
-          xla::ShapedBuffer sub_shaped_buffer,
-          shaped_buffer.SubShapedBuffer({input.output_index() - 1}));
-      input_buffers.input_allocations.emplace_back(
-          std::move(sub_shaped_buffer));
+      XRTTupleAllocation* sub_tuple;
+      TF_RETURN_IF_ERROR(XRTTupleAllocation::MakeSubBuffer(
+          op_inputs[i].get(), {input.output_index() - 1}, &sub_tuple,
+          /*alias_parent_allocation=*/true));
+      input_tuples.emplace_back(sub_tuple);
     }
   }
-  for (size_t i = 0; i < input_buffers.input_allocations.size(); ++i) {
-    input_buffers.input_pointers.push_back(&input_buffers.input_allocations[i]);
-  }
-  return std::move(input_buffers);
+  return input_tuples;
 }
 
 // Given a shape, returns a byte array representing the shape metadata of the
@@ -228,12 +198,11 @@ Status UpdateMetadata(se::Stream* stream, se::DeviceMemory* buffer,
 // As we can't expand the size of an existing memory allocation, a reallocation
 // is required. A list of new allocations are returned after this function. The
 // caller is reponsible for maintaining those allocations.
-xla::StatusOr> UpdateDynamicInputs(
+Status UpdateDynamicInputs(
     se::Stream* stream, se::DeviceMemoryAllocator* allocator,
-    std::vector runtime_inputs,
+    std::vector* execution_inputs,
     const std::vector& compile_time_shapes) {
-  std::vector new_allocations;
-  TF_RET_CHECK(runtime_inputs.size() == compile_time_shapes.size());
+  TF_RET_CHECK(execution_inputs->size() == compile_time_shapes.size());
   TF_ASSIGN_OR_RETURN(auto compiler, xla::Compiler::GetForPlatform(
                                          stream->parent()->platform()));
   auto shape_size_fn = compiler->ShapeSizeBytesFunction();
@@ -242,57 +211,61 @@ xla::StatusOr> UpdateDynamicInputs(
     if (compile_time_shape.is_static()) {
       continue;
     }
-    auto* runtime_input = runtime_inputs[i];
-
+    xla::ExecutionInput* execution_input = &(*execution_inputs)[i];
     bool element_modified = false;
     TF_RETURN_IF_ERROR(xla::ShapeUtil::ForEachSubshapeWithStatus(
         compile_time_shape,
-        [&](const xla::Shape& compile_time_shape,
+        [&](const xla::Shape& sub_shape,
             const xla::ShapeIndex& index) -> Status {
-          if (compile_time_shape.IsTuple() || compile_time_shape.is_static()) {
+          if (sub_shape.IsTuple() || sub_shape.is_static()) {
             return Status::OK();
           }
-          const xla::Shape& runtime_shape = xla::ShapeUtil::GetSubshape(
-              runtime_input->on_device_shape(), index);
-          TF_RET_CHECK(!runtime_shape.IsTuple());
-          TF_RET_CHECK(xla::ShapeUtil::DynamicShapeIsCompatible(
-              runtime_shape, compile_time_shape));
-          se::DeviceMemoryBase* static_input =
-              runtime_input->buffers().mutable_element(index);
           TF_ASSIGN_OR_RETURN(
-              auto dynamic_input,
+              const xla::Shape* runtime_shape,
+              xla::ShapeUtil::TryGetSubshape(execution_input->shape(), index));
+          TF_RET_CHECK(!runtime_shape->IsTuple());
+          TF_RET_CHECK(xla::ShapeUtil::DynamicArrayShapeIsCompatible(
+              *runtime_shape, sub_shape));
+          TF_ASSIGN_OR_RETURN(
+              se::OwningDeviceMemory dynamic_input,
               allocator->Allocate(stream->parent()->device_ordinal(),
-                                  shape_size_fn(compile_time_shape)));
-          new_allocations.emplace_back(std::move(dynamic_input));
-          se::DeviceMemory* dynamic_input_base =
-              new_allocations.back().ptr();
+                                  shape_size_fn(sub_shape)));
+
+          se::DeviceMemoryBase static_input =
+              execution_input->Buffer(index).AsDeviceMemoryBase();
+          se::DeviceMemory* dynamic_input_base = dynamic_input.ptr();
           // Send the original data to the new location.
-          stream->ThenMemcpyD2D(dynamic_input_base, *static_input,
-                                static_input->size());
+          stream->ThenMemcpyD2D(dynamic_input_base, static_input,
+                                static_input.size());
           TF_RETURN_IF_ERROR(UpdateMetadata(stream, dynamic_input_base,
-                                            compile_time_shape, runtime_shape));
+                                            sub_shape, *runtime_shape));
           // Modify the memory location in the input shape tree to point to the
           // new input.
-          runtime_input->set_buffer(*dynamic_input_base, index);
+          execution_input->SetBuffer(
+              index, xla::MaybeOwningDeviceMemory(std::move(dynamic_input)));
+          execution_input->ClearUnownedIndex(index);
           element_modified = true;
           return Status::OK();
         }));
     if (element_modified) {
-      runtime_input->set_shapes(compile_time_shape, compile_time_shape);
+      TF_RETURN_IF_ERROR(execution_input->SetDynamicShape(compile_time_shape));
+      TF_ASSIGN_OR_RETURN(xla::ShapedBuffer shaped_buffer,
+                          execution_input->ToShapedBuffer(
+                              allocator, stream->parent()->device_ordinal()));
       // The input location has been modified, need to fix tuple table to
       // point to the correct address.
       TF_ASSIGN_OR_RETURN(
           auto transfer_manager,
           xla::TransferManager::GetForPlatform(stream->parent()->platform()));
       TF_RETURN_IF_ERROR(
-          transfer_manager->WriteTupleIndexTablesAsync(stream, *runtime_input));
+          transfer_manager->WriteTupleIndexTablesAsync(stream, shaped_buffer));
     }
   }
-  return std::move(new_allocations);
+  return Status::OK();
 }
 
 xla::StatusOr ReadMetadataLiteral(
-    se::Stream* stream, se::DeviceMemoryBase* buffer,
+    se::Stream* stream, se::DeviceMemoryBase buffer,
     const xla::Shape& buffer_shape, xla::TransferManager* transfer_manager) {
   TF_ASSIGN_OR_RETURN(auto compiler, xla::Compiler::GetForPlatform(
                                          stream->parent()->platform()));
@@ -302,7 +275,7 @@ xla::StatusOr ReadMetadataLiteral(
   const int64 offset = shape_size_fn(buffer_shape_static);
   int64 metadata_size = shape_size_fn(buffer_shape) - offset;
   TF_RET_CHECK(metadata_size != 0);
-  auto buffer_8 = se::DeviceMemory(*buffer);
+  auto buffer_8 = se::DeviceMemory(buffer);
   auto metadata_buffer =
       stream->parent()->GetSubBuffer(&buffer_8, offset, metadata_size);
   return transfer_manager->TransferArrayFromDevice(
@@ -315,7 +288,7 @@ xla::StatusOr ReadMetadataLiteral(
 // dimension sizes from the metadata, and update output shapes. The result shape
 // is a static and concrete shape.
 xla::Status UpdateDynamicOutputs(se::Stream* stream,
-                                 xla::ShapedBuffer* shaped_buffer,
+                                 const xla::ShapedBuffer& shaped_buffer,
                                  xla::Shape* output_host_shape,
                                  xla::Shape* output_device_shape) {
   DCHECK(output_device_shape->is_dynamic());
@@ -323,8 +296,8 @@ xla::Status UpdateDynamicOutputs(se::Stream* stream,
       auto transfer_manager,
       xla::TransferManager::GetForPlatform(stream->parent()->platform()));
   TF_RETURN_IF_ERROR(stream->BlockHostUntilDone());
-  TF_RETURN_IF_ERROR(shaped_buffer->buffers().ForEachMutableElementWithStatus(
-      [&](const xla::ShapeIndex& index, se::DeviceMemoryBase* buffer) {
+  TF_RETURN_IF_ERROR(shaped_buffer.buffers().ForEachElementWithStatus(
+      [&](const xla::ShapeIndex& index, const se::DeviceMemoryBase& buffer) {
         const xla::Shape& buffer_shape =
             xla::ShapeUtil::GetSubshape(*output_device_shape, index);
         if (buffer_shape.IsTuple()) {
@@ -352,19 +325,18 @@ xla::Status UpdateDynamicOutputs(se::Stream* stream,
   return Status::OK();
 }
 
-// Create output tuple from run_result.
 xla::StatusOr> CreateOutputTuple(
-    se::Stream* stream, xla::ScopedShapedBuffer run_result,
-    xla::Backend* backend, int device_ordinal) {
+    se::Stream* stream, xla::ExecutionOutput run_result, xla::Backend* backend,
+    int device_ordinal) {
   XRTTupleAllocation* output_tuple;
-  xla::ShapedBuffer shaped_buffer = run_result.release();
+  const xla::ScopedShapedBuffer& shaped_buffer = run_result.Result();
   if (shaped_buffer.on_device_shape().is_dynamic()) {
     // Update dynamic shapes from output buffer, and create a XRT tensor with
     // dimension sizes read from metadata.
     xla::Shape output_host_shape = shaped_buffer.on_host_shape();
     xla::Shape output_device_shape = shaped_buffer.on_device_shape();
     TF_RETURN_IF_ERROR(UpdateDynamicOutputs(
-        stream, &shaped_buffer, &output_host_shape, &output_device_shape));
+        stream, shaped_buffer, &output_host_shape, &output_device_shape));
     TF_RETURN_IF_ERROR(XRTTupleAllocation::CreateFromBuffer(
         shaped_buffer, output_host_shape, output_device_shape, backend,
         device_ordinal, &output_tuple));
@@ -373,15 +345,27 @@ xla::StatusOr> CreateOutputTuple(
     TF_RETURN_IF_ERROR(XRTTupleAllocation::CreateFromBuffer(
         shaped_buffer, backend, device_ordinal, &output_tuple));
   }
+  // After the output tuple is created, we can release the output result
+  // buffers, to make sure they won't be cleared by its destructor.
+  (void)run_result.ConsumeResult().release();
   return RefPtr(output_tuple);
 }
 
 xla::StatusOr> RunExecutable(
     OpKernelContext* context, XRTGenericDeviceAccessor::ScopedRef* device_ref,
-    xla::LocalExecutable* executable, const InputBuffers& input_buffers,
-    se::Stream* stream, int rng_seed,
+    xla::LocalExecutable* executable,
+    absl::Span> input_tuples,
+    bool release_inputs, se::Stream* stream, int rng_seed,
     const xrt::CommonExecutionConfig& config) {
-  VLOG(2) << "Executing computation.";
+  const xla::ComputationLayout& computation_layout =
+      executable->executable()->module_config().entry_computation_layout();
+  std::vector input_is_dynamic = GetDynamicInputInfo(computation_layout);
+  TF_ASSIGN_OR_RETURN(
+      std::vector execution_inputs,
+      GetArgumentsBuffers(
+          executable->executable()->module().input_output_alias_config(),
+          input_tuples, input_is_dynamic, release_inputs));
+
   xla::ExecutableRunOptions run_options;
   run_options.set_stream(stream);
   run_options.set_allocator(device_ref->backend()->memory_allocator());
@@ -419,51 +403,28 @@ xla::StatusOr> RunExecutable(
   }
   run_options.set_gpu_executable_run_options(&gpu_options);
 
-  Env* env = Env::Default();
-  auto start_time = env->NowMicros();
   const std::vector& shape_layouts =
       executable->executable()
           ->module_config()
           .entry_computation_layout()
           .parameter_layouts();
-  TF_ASSIGN_OR_RETURN(auto new_allocations,
-                      UpdateDynamicInputs(stream, run_options.allocator(),
-                                          input_buffers.input_pointers,
-                                          shape_layouts));
-  auto new_allocations_ptr =
-      std::make_shared>(
-          std::move(new_allocations));
+  TF_RETURN_IF_ERROR(UpdateDynamicInputs(stream, run_options.allocator(),
+                                         &execution_inputs, shape_layouts));
   TF_ASSIGN_OR_RETURN(
-      xla::ScopedShapedBuffer run_result,
-      executable->Run(input_buffers.input_pointers, run_options));
-  // Retain the new allocation for input memory until the end of execution.
-  stream->ThenDoHostCallback([new_allocations_ptr]() { return Status::OK(); });
-
-  auto elapsed = env->NowMicros() - start_time;
-  VLOG(2) << "Elapsed time: " << elapsed << "us";
+      xla::ExecutionOutput run_result,
+      executable->Run(std::move(execution_inputs), run_options));
 
   TF_ASSIGN_OR_RETURN(
       RefPtr output_tuple_ptr,
       CreateOutputTuple(stream, std::move(run_result), device_ref->backend(),
                         device_ref->device_ordinal()));
-
   // The ScopedShapedBuffer returned by the executable Run() API, in case of
   // input/output buffer aliasing, might have holes in it, which need to be
   // filled using the proper input tuples buffers which are the source of
   // aliasing.
-  const xla::HloInputOutputAliasConfig& input_output_alias =
-      executable->executable()->module().input_output_alias_config();
-  auto alias_function =
-      [&](const xla::ShapeIndex& output_index,
-          const xla::HloInputOutputAliasConfig::Alias& alias) -> Status {
-    TF_RET_CHECK(alias.parameter_number < input_buffers.input_tuples.size());
-    return alias.kind == xla::HloInputOutputAliasConfig::AliasKind::kUserAlias
-               ? output_tuple_ptr->AliasBufferFrom(
-                     *input_buffers.input_tuples[alias.parameter_number],
-                     alias.parameter_index, output_index)
-               : Status::OK();
-  };
-  TF_RETURN_IF_ERROR(input_output_alias.ForEachAliasWithStatus(alias_function));
+  TF_RETURN_IF_ERROR(RebuildOutputAliases(
+      output_tuple_ptr, input_tuples,
+      executable->executable()->module().input_output_alias_config()));
 
   return std::move(output_tuple_ptr);
 }
@@ -471,12 +432,13 @@ xla::StatusOr> RunExecutable(
 xla::StatusOr> ExecuteComputation(
     OpKernelContext* context, XRTMemoryManager* memory_manager,
     XRTGenericDeviceAccessor::ScopedRef* device_ref,
-    xla::LocalExecutable* executable, const InputBuffers& input_buffers,
-    se::Stream* stream, int rng_seed,
+    xla::LocalExecutable* executable,
+    absl::Span> input_tuples,
+    bool release_inputs, se::Stream* stream, int rng_seed,
     const xrt::CommonExecutionConfig& config) {
   auto runfn = [&]() {
-    return RunExecutable(context, device_ref, executable, input_buffers, stream,
-                         rng_seed, config);
+    return RunExecutable(context, device_ref, executable, input_tuples,
+                         release_inputs, stream, rng_seed, config);
   };
 
   // We pass zero as requested_free_size as there is no simple way to get the
@@ -495,12 +457,13 @@ xla::StatusOr> ExecuteComputation(
     se::Stream* stream, int rng_seed,
     const xrt::CommonExecutionConfig& config) {
   XRTMemoryManager::WorkingSet working_set(memory_manager);
-  TF_ASSIGN_OR_RETURN(InputBuffers input_buffers,
-                      GetInputBuffers(&working_set, device_ref->backend(),
-                                      input_coords, release_inputs));
+  TF_ASSIGN_OR_RETURN(
+      std::vector> input_tuples,
+      GetInputTuples(executable, &working_set, device_ref->backend(),
+                     input_coords, release_inputs));
   return ExecuteComputation(context, memory_manager.get(), device_ref,
-                            executable, input_buffers, stream, rng_seed,
-                            config);
+                            executable, input_tuples, release_inputs, stream,
+                            rng_seed, config);
 }
 
 // XRTExecuteOp
@@ -653,16 +616,16 @@ Status XRTExecuteChainedOp::DoWork(OpKernelContext* context) {
   auto execute_op = [&](const xrt::XRTChainedExecuteOp& op,
                         absl::Span> op_inputs)
       -> xla::StatusOr> {
-    TF_ASSIGN_OR_RETURN(InputBuffers input_buffers,
-                        GetChainedOpInputs(op, op_inputs));
-
     std::unique_ptr entry;
     TF_RETURN_IF_ERROR(cache->Lookup(op.computation_handle(), &entry));
     xla::LocalExecutable* executable = entry->get().get_executable();
 
-    return ExecuteComputation(context, memory_manager.get(), &device_ref,
-                              executable, input_buffers, stream, rng_seed,
-                              config.common_config());
+    TF_ASSIGN_OR_RETURN(std::vector> input_tuples,
+                        GetChainedOpInputTuples(op, op_inputs));
+
+    return ExecuteComputation(
+        context, memory_manager.get(), &device_ref, executable, input_tuples,
+        /*release_inputs=*/false, stream, rng_seed, config.common_config());
   };
 
   return ExecuteChained(context, memory_manager, device_ref.backend(),
diff --git a/tensorflow/compiler/xrt/xrt_util.cc b/tensorflow/compiler/xrt/xrt_util.cc
index b8a0afc92c5..926ba23c7af 100644
--- a/tensorflow/compiler/xrt/xrt_util.cc
+++ b/tensorflow/compiler/xrt/xrt_util.cc
@@ -221,6 +221,140 @@ xla::StatusOr> GetComputationInputs(
   return std::move(input_coords);
 }
 
+bool InputShapeMatches(const xla::Shape& parameter_shape,
+                       const xla::Shape& input_shape) {
+  auto shape_checker = [&](const xla::Shape& pshape,
+                           const xla::ShapeIndex& index) {
+    if (pshape.IsArray()) {
+      TF_ASSIGN_OR_RETURN(const xla::Shape* ishape,
+                          xla::ShapeUtil::TryGetSubshape(input_shape, index));
+      if (pshape.rank() != ishape->rank() ||
+          pshape.element_type() != ishape->element_type()) {
+        return errors::InvalidArgument("Mismatching shapes");
+      }
+      if (pshape.is_static() && pshape.layout() != ishape->layout()) {
+        return errors::InvalidArgument("Mismatching layouts");
+      }
+      for (int64 dim = 0; dim < pshape.rank(); ++dim) {
+        if (pshape.is_dynamic_dimension(dim)) {
+          if (pshape.dimensions(dim) < ishape->dimensions(dim)) {
+            return errors::InvalidArgument("Mismatching shapes");
+          }
+        } else if (pshape.dimensions(dim) != ishape->dimensions(dim)) {
+          return errors::InvalidArgument("Mismatching shapes");
+        }
+      }
+    }
+    return Status::OK();
+  };
+  return xla::ShapeUtil::ForEachSubshapeWithStatus(parameter_shape,
+                                                   shape_checker)
+      .ok();
+}
+
+xla::StatusOr>> GetInputTupleAllocations(
+    const std::vector& input_coords,
+    XRTMemoryManager::WorkingSet* working_set, xla::Backend* backend,
+    int64 num_input_shapes,
+    const std::function& shape_getter, bool release_inputs) {
+  if (input_coords.size() != num_input_shapes) {
+    return errors::InvalidArgument(
+        "Number of inputs does not match executable proto input shapes: ",
+        input_coords.size(), " vs. ", num_input_shapes);
+  }
+  std::vector> input_tuples;
+  input_tuples.reserve(input_coords.size());
+  for (size_t i = 0; i < input_coords.size(); ++i) {
+    TF_RETURN_IF_ERROR(
+        working_set->LookupAndPin(backend, input_coords[i].handle));
+    auto tuple = working_set->PinnedTuples().back();
+    if (release_inputs) {
+      // We are holding a reference to the tuple, so we can safely delete it
+      // from the resource manager here.
+      TF_RETURN_IF_ERROR(
+          working_set->MemoryManager()->Release(input_coords[i].handle));
+      VLOG(2) << "Released allocation handle " << input_coords[i].handle;
+    }
+    xla::Shape input_shape = shape_getter(i);
+    if (!InputShapeMatches(input_shape, tuple->on_host_shape())) {
+      return errors::InvalidArgument(
+          "Run-time shape mismatch for XRTExecute argument[", i, "] (",
+          input_coords[i].handle, "). Expected ", input_shape.DebugString(),
+          "; got ", tuple->on_host_shape().DebugString());
+    }
+    if (input_coords[i].index.empty()) {
+      input_tuples.emplace_back(std::move(tuple));
+    } else {
+      XRTTupleAllocation* sub_tuple;
+      TF_RETURN_IF_ERROR(XRTTupleAllocation::MakeSubBuffer(
+          tuple.get(), input_coords[i].index, &sub_tuple,
+          /*alias_parent_allocation=*/true));
+      input_tuples.emplace_back(sub_tuple);
+    }
+  }
+  return std::move(input_tuples);
+}
+
+Status RebuildOutputAliases(
+    const RefPtr& output_tuple,
+    absl::Span> input_tuples,
+    const xla::HloInputOutputAliasConfig& input_output_alias) {
+  auto alias_function =
+      [&](const xla::ShapeIndex& output_index,
+          const xla::HloInputOutputAliasConfig::Alias& alias) -> Status {
+    TF_RET_CHECK(alias.parameter_number < input_tuples.size());
+    return alias.kind == xla::HloInputOutputAliasConfig::AliasKind::kUserAlias
+               ? output_tuple->AliasBufferFrom(
+                     *input_tuples[alias.parameter_number],
+                     alias.parameter_index, output_index)
+               : Status::OK();
+  };
+  return input_output_alias.ForEachAliasWithStatus(alias_function);
+}
+
+xla::StatusOr> GetArgumentsBuffers(
+    const xla::HloInputOutputAliasConfig& input_output_alias,
+    absl::Span> input_tuples,
+    const std::vector& input_is_dynamic, bool release_inputs) {
+  auto is_dynamic = [&](size_t arg) {
+    return arg < input_is_dynamic.size() && input_is_dynamic[arg];
+  };
+  std::vector arguments;
+  // Don't alias dynamic input -- Due to the underlying implementation,
+  // aliased inputs have two owners: XRTAllocation and return value of
+  // this function. If an argument is dynamic and the ownership is
+  // released to output of this function, TPUExecute will free it and
+  // reallocate a new one, which creates a double freeing issue where
+  // XRTAllocation also attempts to release the buffer.
+  bool alias_outputs = release_inputs && input_tuples.size() == 1 &&
+                       input_tuples[0]->IsExclusiveOwner() && !is_dynamic(0);
+  arguments.reserve(input_tuples.size());
+  for (int64 i = 0; i < input_tuples.size(); ++i) {
+    auto alias_checker =
+        [&](const xla::ShapeIndex& index) -> xla::StatusOr {
+      // Only the buffers which the caller explicitly marked as aliased
+      // (kUserAlias), should create aliases.
+      // The XLA compiler might create opportunistic aliases (kSystemAlias)
+      // which need a different handling. With a system alias we know that XLA
+      // is going to reuse a given input parameter buffer for a given output, so
+      // unless it is known at call site that the input buffer has no more uses,
+      // a copy needs to be made at call site. With user specified alias the
+      // caller tells us that he expects a given output to land over the buffers
+      // of a given parametter.
+      if (input_output_alias.ParameterAliasKind(i, index) ==
+          xla::HloInputOutputAliasConfig::AliasKind::kUserAlias) {
+        TF_RET_CHECK(!is_dynamic(i));
+        return true;
+      }
+      return alias_outputs;
+    };
+    TF_ASSIGN_OR_RETURN(xla::ExecutionInput exec_input,
+                        input_tuples[i]->ToExecutionInput(alias_checker));
+    arguments.emplace_back(std::move(exec_input));
+  }
+  return std::move(arguments);
+}
+
 Status CreateExecuteOutput(OpKernelContext* context,
                            XRTMemoryManager* memory_manager,
                            RefPtr output_tuple,
diff --git a/tensorflow/compiler/xrt/xrt_util.h b/tensorflow/compiler/xrt/xrt_util.h
index cc1480fdb00..832c106621f 100644
--- a/tensorflow/compiler/xrt/xrt_util.h
+++ b/tensorflow/compiler/xrt/xrt_util.h
@@ -23,6 +23,8 @@ limitations under the License.
 #include 
 
 #include "tensorflow/compiler/xla/service/backend.h"
+#include "tensorflow/compiler/xla/service/hlo_input_output_alias_config.h"
+#include "tensorflow/compiler/xla/shape.h"
 #include "tensorflow/compiler/xla/shape_util.h"
 #include "tensorflow/compiler/xla/statusor.h"
 #include "tensorflow/compiler/xla/xla.pb.h"
@@ -69,6 +71,25 @@ xla::DebugOptions BuildXlaDebugOptions(const xla::DebugOptions& ref_options);
 xla::StatusOr> GetComputationInputs(
     OpKernelContext* context, const char* input_name);
 
+bool InputShapeMatches(const xla::Shape& parameter_shape,
+                       const xla::Shape& input_shape);
+
+xla::StatusOr>> GetInputTupleAllocations(
+    const std::vector& input_coords,
+    XRTMemoryManager::WorkingSet* working_set, xla::Backend* backend,
+    int64 num_input_shapes,
+    const std::function& shape_getter, bool release_inputs);
+
+Status RebuildOutputAliases(
+    const RefPtr& output_tuple,
+    absl::Span> input_tuples,
+    const xla::HloInputOutputAliasConfig& input_output_alias);
+
+xla::StatusOr> GetArgumentsBuffers(
+    const xla::HloInputOutputAliasConfig& input_output_alias,
+    absl::Span> input_tuples,
+    const std::vector& input_is_dynamic, bool release_inputs);
+
 // Create the XRT execute output tensor given the computation result
 // (output_tuple). The return_exploded_tuple tells whether a tuple result should
 // be returned as vector of handles representing each tuple child.

From 8cf97846290cf7d8b95256fe3123abaaa8c8e553 Mon Sep 17 00:00:00 2001
From: "A. Unique TensorFlower" 
Date: Mon, 22 Jun 2020 11:45:59 -0700
Subject: [PATCH 452/867] Update Eigen to:
 https://gitlab.com/libeigen/eigen/-/commit/386d809bde475c65b7940f290efe80e6a05878c4

PiperOrigin-RevId: 317701612
Change-Id: I115babafbfc7fbad62ac4408da765dc1cf190610
---
 tensorflow/workspace.bzl | 8 ++++----
 1 file changed, 4 insertions(+), 4 deletions(-)

diff --git a/tensorflow/workspace.bzl b/tensorflow/workspace.bzl
index 5af8c8272b8..0f591ba8b90 100755
--- a/tensorflow/workspace.bzl
+++ b/tensorflow/workspace.bzl
@@ -237,11 +237,11 @@ def tf_repositories(path_prefix = "", tf_repo_name = ""):
         name = "eigen_archive",
         build_file = clean_dep("//third_party:eigen.BUILD"),
         patch_file = clean_dep("//third_party/eigen3:gpu_packet_math.patch"),
-        sha256 = "d26ada177ed9b696a9447fc85d209932a032c8ffc51630cf15eea8629b29dad6",  # SHARED_EIGEN_SHA
-        strip_prefix = "eigen-6b9c92fe7eff0dedb031cec38004c9c3667f3057",
+        sha256 = "f632d82e43ffc46adfac9043beace700b0265748075e7edc0701d81380258038",  # SHARED_EIGEN_SHA
+        strip_prefix = "eigen-386d809bde475c65b7940f290efe80e6a05878c4",
         urls = [
-            "https://storage.googleapis.com/mirror.tensorflow.org/gitlab.com/libeigen/eigen/-/archive/6b9c92fe7eff0dedb031cec38004c9c3667f3057/eigen-6b9c92fe7eff0dedb031cec38004c9c3667f3057.tar.gz",
-            "https://gitlab.com/libeigen/eigen/-/archive/6b9c92fe7eff0dedb031cec38004c9c3667f3057/eigen-6b9c92fe7eff0dedb031cec38004c9c3667f3057.tar.gz",
+            "https://storage.googleapis.com/mirror.tensorflow.org/gitlab.com/libeigen/eigen/-/archive/386d809bde475c65b7940f290efe80e6a05878c4/eigen-386d809bde475c65b7940f290efe80e6a05878c4.tar.gz",
+            "https://gitlab.com/libeigen/eigen/-/archive/386d809bde475c65b7940f290efe80e6a05878c4/eigen-386d809bde475c65b7940f290efe80e6a05878c4.tar.gz",
         ],
     )
 

From 39d080e8b9c0b1aa3896d4dca11169760d9ba506 Mon Sep 17 00:00:00 2001
From: Yujing Zhang 
Date: Mon, 22 Jun 2020 11:48:57 -0700
Subject: [PATCH 453/867] Use the same CompositeDevice name on remote workers
 as the one on a client.

PiperOrigin-RevId: 317702206
Change-Id: I7068efb25eb930252f89a167108ed59c69c2078f
---
 .../core/common_runtime/composite_device.cc   | 17 ++++++----
 .../core/common_runtime/composite_device.h    |  5 +++
 .../common_runtime/composite_device_test.cc   | 16 ++++++++++
 .../core/common_runtime/eager/context.cc      | 24 +++++++++-----
 .../core/common_runtime/eager/context.h       |  7 +++--
 .../core/common_runtime/eager/context_test.cc | 31 +++++++++++++++++++
 .../core/common_runtime/eager/execute.cc      |  2 +-
 .../common_runtime/eager/tensor_handle.cc     |  9 +++---
 .../core/common_runtime/eager/tensor_handle.h |  2 +-
 .../eager/eager_service_impl.cc               |  2 +-
 .../eager/eager_service_impl_test.cc          |  4 +++
 .../eager/remote_copy_node.cc                 |  1 +
 tensorflow/core/protobuf/eager_service.proto  |  2 ++
 13 files changed, 99 insertions(+), 23 deletions(-)

diff --git a/tensorflow/core/common_runtime/composite_device.cc b/tensorflow/core/common_runtime/composite_device.cc
index 7fd41e00a04..d4548946cbf 100644
--- a/tensorflow/core/common_runtime/composite_device.cc
+++ b/tensorflow/core/common_runtime/composite_device.cc
@@ -25,6 +25,16 @@ const char* const kCompositeDeviceType = "COMPOSITE";
 std::unique_ptr CompositeDevice::MakeDevice(
     const std::vector& underlying_devices, const int unique_device_id,
     const DeviceNameUtils::ParsedName& host_name, Status* status) {
+  DeviceNameUtils::ParsedName parsed_name = host_name;
+  parsed_name.type = kCompositeDeviceType;
+  parsed_name.id = unique_device_id;
+  const string device_name = DeviceNameUtils::ParsedNameToString(parsed_name);
+  return CompositeDevice::MakeDevice(underlying_devices, device_name, status);
+}
+
+std::unique_ptr CompositeDevice::MakeDevice(
+    const std::vector& underlying_devices, const string& device_name,
+    Status* status) {
   if (underlying_devices.empty()) {
     status->Update(
         errors::InvalidArgument("underlying_devices should not be empty."));
@@ -63,13 +73,8 @@ std::unique_ptr CompositeDevice::MakeDevice(
     }
   }
 
-  DeviceNameUtils::ParsedName parsed_composite_name = host_name;
   DeviceAttributes device_attributes;
-  parsed_composite_name.type = kCompositeDeviceType;
-  parsed_composite_name.id = unique_device_id;
-  const string composite_name =
-      DeviceNameUtils::ParsedNameToString(parsed_composite_name);
-  device_attributes.set_name(composite_name);
+  device_attributes.set_name(device_name);
   device_attributes.set_device_type(kCompositeDeviceType);
 
   return absl::WrapUnique(
diff --git a/tensorflow/core/common_runtime/composite_device.h b/tensorflow/core/common_runtime/composite_device.h
index 850eae55e8d..c68c395198a 100644
--- a/tensorflow/core/common_runtime/composite_device.h
+++ b/tensorflow/core/common_runtime/composite_device.h
@@ -48,6 +48,11 @@ class CompositeDevice : public Device {
       const std::vector& underlying_devices, const int unique_device_id,
       const DeviceNameUtils::ParsedName& host_name, Status* status);
 
+  // Helper for creating a CompositeDevice with the given device name.
+  static std::unique_ptr MakeDevice(
+      const std::vector& underlying_devices, const string& device_name,
+      Status* status);
+
  private:
   CompositeDevice(const DeviceAttributes& device_attributes,
                   const std::vector& underlying_devices)
diff --git a/tensorflow/core/common_runtime/composite_device_test.cc b/tensorflow/core/common_runtime/composite_device_test.cc
index 73a6ae44912..7d195a7a08e 100644
--- a/tensorflow/core/common_runtime/composite_device_test.cc
+++ b/tensorflow/core/common_runtime/composite_device_test.cc
@@ -80,4 +80,20 @@ TEST(CompositeDeviceTest, Basic) {
   }
 }
 
+TEST(CompositeDeviceTest, DeviceName) {
+  const string composite_device_name =
+      "/job:localhost/replica:0/task:0/device:CPU:10";
+  std::vector underlying_devices;
+  underlying_devices.push_back("/job:worker/replica:0/task:0/device:CPU:0");
+  underlying_devices.push_back("/job:worker/replica:0/task:0/device:CPU:1");
+  Status status;
+  std::unique_ptr composite_device =
+      CompositeDevice::MakeDevice(underlying_devices, composite_device_name,
+                                  &status);
+  TF_ASSERT_OK(status);
+  EXPECT_EQ(composite_device->name(), composite_device_name);
+  EXPECT_EQ(composite_device->device_type(), kCompositeDeviceType);
+  EXPECT_EQ(underlying_devices, *composite_device->underlying_devices());
+}
+
 }  // namespace tensorflow
diff --git a/tensorflow/core/common_runtime/eager/context.cc b/tensorflow/core/common_runtime/eager/context.cc
index 6dc0a3a8200..7ca40fc6cf6 100644
--- a/tensorflow/core/common_runtime/eager/context.cc
+++ b/tensorflow/core/common_runtime/eager/context.cc
@@ -893,7 +893,7 @@ Status EagerContext::FindDeviceFromName(const char* device_name,
 }
 
 Status EagerContext::FindCompositeDeviceFromName(
-    const char* device_name, CompositeDevice** device) const {
+    StringPiece device_name, CompositeDevice** device) const {
   tf_shared_lock l(composite_devices_mu_);
   for (const auto& d : composite_devices_) {
     if (d.second->name() == device_name) {
@@ -939,8 +939,13 @@ Status EagerContext::RegisterCustomDevice(
 }
 
 Status EagerContext::FindOrCreateCompositeDevice(
-    const std::vector& underlying_devices,
+    const std::vector& underlying_devices, const string& device_name,
     CompositeDevice** composite_device) {
+  if (!device_name.empty() &&
+      FindCompositeDeviceFromName(device_name, composite_device).ok()) {
+    return Status::OK();
+  }
+
   const uint64 hash_key = Fingerprint64(absl::StrJoin(underlying_devices, ","));
 
   mutex_lock l(composite_devices_mu_);
@@ -951,11 +956,16 @@ Status EagerContext::FindOrCreateCompositeDevice(
   }
 
   Status s;
-  // Create a CompositeDevice on the same task as the host CPU, in order to
-  // trigger packed TensorHandle copy from a client to a remote worker.
-  auto device =
-      CompositeDevice::MakeDevice(underlying_devices, composite_devices_.size(),
-                                  HostCPU()->parsed_name(), &s);
+  std::unique_ptr device;
+  if (device_name.empty()) {
+    // Create a CompositeDevice on the same task as the host CPU, in order to
+    // trigger packed TensorHandle copy from a client to a remote worker.
+    device = CompositeDevice::MakeDevice(underlying_devices,
+                                         composite_devices_.size(),
+                                         HostCPU()->parsed_name(), &s);
+  } else {
+    device = CompositeDevice::MakeDevice(underlying_devices, device_name, &s);
+  }
   TF_RETURN_IF_ERROR(s);
   *composite_device = device.get();
   pflr_->AddCompositeDevice(*composite_device);
diff --git a/tensorflow/core/common_runtime/eager/context.h b/tensorflow/core/common_runtime/eager/context.h
index 3dab7c08d77..c16e1f0f4ad 100644
--- a/tensorflow/core/common_runtime/eager/context.h
+++ b/tensorflow/core/common_runtime/eager/context.h
@@ -486,7 +486,7 @@ class EagerContext : public ImmediateExecutionContext, public core::RefCounted {
 
   Status FindDeviceFromName(const char* device_name, Device** device) const;
 
-  Status FindCompositeDeviceFromName(const char* device_name,
+  Status FindCompositeDeviceFromName(StringPiece device_name,
                                      CompositeDevice** device) const;
 
   Status FindCustomDeviceFromName(const string& device_name,
@@ -495,9 +495,10 @@ class EagerContext : public ImmediateExecutionContext, public core::RefCounted {
   Status RegisterCustomDevice(const string& name,
                               std::unique_ptr device);
 
-  // Find or create a composite device with the given `underlying_devices`.
+  // Find or create a composite device with the given `underlying_devices` and
+  // `device_name` (if not empty).
   Status FindOrCreateCompositeDevice(
-      const std::vector& underlying_devices,
+      const std::vector& underlying_devices, const string& device_name,
       CompositeDevice** composite_device);
 
   bool OnSameTask(const Device* first, const Device* second) const;
diff --git a/tensorflow/core/common_runtime/eager/context_test.cc b/tensorflow/core/common_runtime/eager/context_test.cc
index c6ed61c80c4..7f34884b4db 100644
--- a/tensorflow/core/common_runtime/eager/context_test.cc
+++ b/tensorflow/core/common_runtime/eager/context_test.cc
@@ -177,6 +177,7 @@ TEST_F(EagerContextTest, CompositeDevice) {
       "/job:worker/replica:0/task:0/device:CPU:1"};
   CompositeDevice* composite_device_0 = nullptr;
   TF_ASSERT_OK(context()->FindOrCreateCompositeDevice(underlying_devices,
+                                                      /*device_name=*/"",
                                                       &composite_device_0));
   EXPECT_EQ(composite_device_0->name(),
             "/job:localhost/replica:0/task:0/device:COMPOSITE:0");
@@ -186,11 +187,13 @@ TEST_F(EagerContextTest, CompositeDevice) {
   EXPECT_EQ(device, composite_device_0);
   CompositeDevice* composite_device_1 = nullptr;
   TF_ASSERT_OK(context()->FindOrCreateCompositeDevice(underlying_devices,
+                                                      /*device_name=*/"",
                                                       &composite_device_1));
   EXPECT_EQ(composite_device_1, composite_device_0);
   underlying_devices.push_back("/job:worker/replica:0/task:0/device:CPU:2");
   CompositeDevice* composite_device_2 = nullptr;
   TF_ASSERT_OK(context()->FindOrCreateCompositeDevice(underlying_devices,
+                                                      /*device_name=*/"",
                                                       &composite_device_2));
   EXPECT_EQ(composite_device_2->name(),
             "/job:localhost/replica:0/task:0/device:COMPOSITE:1");
@@ -202,5 +205,33 @@ TEST_F(EagerContextTest, CompositeDevice) {
       "/job:localhost/replica:0/task:0/device:COMPOSITE:2", &device)));
 }
 
+TEST_F(EagerContextTest, CompositeDeviceWithGivenName) {
+  InitContext(SessionOptions(), DEVICE_PLACEMENT_EXPLICIT);
+  const std::vector underlying_devices_0 = {
+      "/job:worker/replica:0/task:0/device:CPU:0",
+      "/job:worker/replica:0/task:0/device:CPU:1"};
+  const string composite_device_name =
+      "/job:worker1/replica:0/task:0/device:COMPOSITE:5";
+  // Create a CompositeDevice with the given name.
+  CompositeDevice* composite_device_0 = nullptr;
+  TF_ASSERT_OK(context()->FindOrCreateCompositeDevice(
+      underlying_devices_0, composite_device_name, &composite_device_0));
+  EXPECT_EQ(composite_device_0->name(), composite_device_name);
+
+  CompositeDevice* device = nullptr;
+  TF_EXPECT_OK(
+      context()->FindCompositeDeviceFromName(composite_device_name, &device));
+  EXPECT_EQ(device, composite_device_0);
+
+  std::vector underlying_devices_1 = {
+      "/job:worker/replica:0/task:0/device:CPU:1",
+      "/job:worker/replica:0/task:0/device:CPU:2"};
+  // Find a CompositeDevice with the given name.
+  CompositeDevice* composite_device_1 = nullptr;
+  TF_ASSERT_OK(context()->FindOrCreateCompositeDevice(
+      underlying_devices_1, composite_device_0->name(), &composite_device_1));
+  EXPECT_EQ(composite_device_1, composite_device_0);
+}
+
 }  // namespace
 }  // namespace tensorflow
diff --git a/tensorflow/core/common_runtime/eager/execute.cc b/tensorflow/core/common_runtime/eager/execute.cc
index 5be0ad8db57..40b737c8ccf 100644
--- a/tensorflow/core/common_runtime/eager/execute.cc
+++ b/tensorflow/core/common_runtime/eager/execute.cc
@@ -415,7 +415,7 @@ Status GetOrCreateKernelAndDevice(
       TF_RETURN_IF_ERROR(GetDeviceForInput(ctx, input, &input_device));
       input_dev_ptrs.push_back(input_device);
       CompositeDevice* composite_device = nullptr;
-      if (ctx.FindCompositeDeviceFromName(input_device->name().c_str(),
+      if (ctx.FindCompositeDeviceFromName(input_device->name(),
                                           &composite_device)
               .ok()) {
         composite_devices[input_device->name()] =
diff --git a/tensorflow/core/common_runtime/eager/tensor_handle.cc b/tensorflow/core/common_runtime/eager/tensor_handle.cc
index 9e607c97683..ef3e7a3cd28 100644
--- a/tensorflow/core/common_runtime/eager/tensor_handle.cc
+++ b/tensorflow/core/common_runtime/eager/tensor_handle.cc
@@ -320,6 +320,7 @@ TensorHandle::TensorHandle(Device* d, Device* op_device,
 Status TensorHandle::CreatePackedHandle(std::vector&& handles,
                                         const tensorflow::DataType dtype,
                                         const tensorflow::TensorShape& shape,
+                                        const string& device_name,
                                         EagerContext* ctx,
                                         TensorHandle** packed_handle) {
   if (handles.empty()) {
@@ -343,8 +344,8 @@ Status TensorHandle::CreatePackedHandle(std::vector&& handles,
   }
 
   CompositeDevice* composite_device = nullptr;
-  TF_RETURN_IF_ERROR(
-      ctx->FindOrCreateCompositeDevice(devices, &composite_device));
+  TF_RETURN_IF_ERROR(ctx->FindOrCreateCompositeDevice(devices, device_name,
+                                                      &composite_device));
   *packed_handle =
       new TensorHandle(std::move(handles), composite_device, dtype, shape, ctx);
   (*packed_handle)->SetResourceHandleInfo(std::move(resource_handle_info));
@@ -363,8 +364,8 @@ Status TensorHandle::CreatePackedHandle(std::vector&& handles,
   tensorflow::DataType dtype = handles.at(0)->dtype;
   tensorflow::TensorShape shape;
   TF_RETURN_IF_ERROR(handles.at(0)->Shape(&shape));
-  return CreatePackedHandle(std::move(handles), dtype, shape, ctx,
-                            packed_handle);
+  return CreatePackedHandle(std::move(handles), dtype, shape,
+                            /*device_name*/ "", ctx, packed_handle);
 }
 
 TensorHandle::TensorHandle(std::vector&& handles, Device* device,
diff --git a/tensorflow/core/common_runtime/eager/tensor_handle.h b/tensorflow/core/common_runtime/eager/tensor_handle.h
index a14df475e0f..6ac48bdac26 100644
--- a/tensorflow/core/common_runtime/eager/tensor_handle.h
+++ b/tensorflow/core/common_runtime/eager/tensor_handle.h
@@ -94,7 +94,7 @@ class TensorHandle : public ImmediateExecutionTensorHandle,
   static Status CreatePackedHandle(std::vector&& handles,
                                    const tensorflow::DataType dtype,
                                    const tensorflow::TensorShape& shape,
-                                   EagerContext* ctx,
+                                   const string& device_name, EagerContext* ctx,
                                    TensorHandle** packed_handle);
   static Status CreatePackedHandle(std::vector&& handles,
                                    EagerContext* ctx,
diff --git a/tensorflow/core/distributed_runtime/eager/eager_service_impl.cc b/tensorflow/core/distributed_runtime/eager/eager_service_impl.cc
index f1e70d53757..fd0606538c4 100644
--- a/tensorflow/core/distributed_runtime/eager/eager_service_impl.cc
+++ b/tensorflow/core/distributed_runtime/eager/eager_service_impl.cc
@@ -685,7 +685,7 @@ Status EagerServiceImpl::SendPackedHandle(
   // Create a unshaped packed TensorHandle.
   TF_RETURN_IF_ERROR(TensorHandle::CreatePackedHandle(
       std::move(handles_to_pack), handles.at(0)->dtype, TensorShape(),
-      eager_context, &packed_handle));
+      send_packed_handle.device_name(), eager_context, &packed_handle));
 
   for (auto* h : handles) {
     // Unref handle since it has a ref in the packed handle now.
diff --git a/tensorflow/core/distributed_runtime/eager/eager_service_impl_test.cc b/tensorflow/core/distributed_runtime/eager/eager_service_impl_test.cc
index 7a315ca1ea5..a2412eb9625 100644
--- a/tensorflow/core/distributed_runtime/eager/eager_service_impl_test.cc
+++ b/tensorflow/core/distributed_runtime/eager/eager_service_impl_test.cc
@@ -1071,6 +1071,8 @@ TEST_F(EagerServiceImplTest, SendPackedHandleTest) {
   const string device0 = "/job:localhost/replica:0/task:0/device:CPU:0";
   const string device1 = "/job:localhost/replica:0/task:1/device:CPU:0";
   const string device2 = "/job:localhost/replica:0/task:2/device:CPU:0";
+  const string composite_device =
+      "/job:localhost/replica:0/task:0/device:COMPOSITE:0";
 
   uint64 context_id = random::New64();
   CreateContextRequest request;
@@ -1125,6 +1127,8 @@ TEST_F(EagerServiceImplTest, SendPackedHandleTest) {
 
   EXPECT_EQ(packed_handle->Type(), TensorHandle::PACKED);
   EXPECT_EQ(packed_handle->NumPackedHandles(), 3);
+  EXPECT_EQ(absl::get(packed_handle->device())->name(),
+            composite_device);
 
   TensorHandle* handle0 = nullptr;
   TF_ASSERT_OK(packed_handle->ExtractPackedHandle(0, &handle0));
diff --git a/tensorflow/core/distributed_runtime/eager/remote_copy_node.cc b/tensorflow/core/distributed_runtime/eager/remote_copy_node.cc
index 41027d43188..a5eeed6a0b6 100644
--- a/tensorflow/core/distributed_runtime/eager/remote_copy_node.cc
+++ b/tensorflow/core/distributed_runtime/eager/remote_copy_node.cc
@@ -298,6 +298,7 @@ Status SerializePackedHandle(const uint64 op_id, TensorHandle* packed_handle,
                              const Device* target_device, EagerContext* ctx,
                              SendPackedHandleOp* op) {
   op->set_op_id(op_id);
+  op->set_device_name(VariantDeviceName(packed_handle->DeviceOrHostCPU(*ctx)));
   for (int i = 0; i < packed_handle->NumPackedHandles(); ++i) {
     TensorHandle* h = nullptr;
     TF_RETURN_IF_ERROR(packed_handle->ExtractPackedHandle(i, &h));
diff --git a/tensorflow/core/protobuf/eager_service.proto b/tensorflow/core/protobuf/eager_service.proto
index 3fe2bd486ba..179ef19f805 100644
--- a/tensorflow/core/protobuf/eager_service.proto
+++ b/tensorflow/core/protobuf/eager_service.proto
@@ -258,6 +258,8 @@ message SendPackedHandleOp {
   }
 
   repeated Handle handles = 2;
+
+  string device_name = 3;
 }
 
 ////////////////////////////////////////////////////////////////////////////////

From 4a340a291feaf53274df8baedbc3ae85d8891eb5 Mon Sep 17 00:00:00 2001
From: Frank Chen 
Date: Mon, 22 Jun 2020 11:51:03 -0700
Subject: [PATCH 454/867] Add snappy compression support to RecordReader,
 RecordWriter, and tf.data snapshot

PiperOrigin-RevId: 317702626
Change-Id: Ic4c583387fb39646ac9e195ff96a3c5052da07c9
---
 tensorflow/core/BUILD                         |  1 +
 .../data/experimental/snapshot_util_test.cc   |  6 +++
 tensorflow/core/lib/io/BUILD                  | 18 ++++++++
 tensorflow/core/lib/io/record_reader.cc       | 36 ++++++++-------
 tensorflow/core/lib/io/record_reader.h        | 14 ++++--
 .../core/lib/io/record_reader_writer_test.cc  | 38 ++++++++++++++++
 tensorflow/core/lib/io/record_writer.cc       | 44 +++++++++++--------
 tensorflow/core/lib/io/record_writer.h        | 15 +++++--
 .../io/snappy/snappy_compression_options.h    | 36 +++++++++++++++
 9 files changed, 166 insertions(+), 42 deletions(-)
 create mode 100644 tensorflow/core/lib/io/snappy/snappy_compression_options.h

diff --git a/tensorflow/core/BUILD b/tensorflow/core/BUILD
index 0f709750897..695035c91e9 100644
--- a/tensorflow/core/BUILD
+++ b/tensorflow/core/BUILD
@@ -1864,6 +1864,7 @@ cc_library(
         "//tensorflow/core/lib/io:random_inputstream",
         "//tensorflow/core/lib/io:record_reader",
         "//tensorflow/core/lib/io:record_writer",
+        "//tensorflow/core/lib/io:snappy_compression_options",
         "//tensorflow/core/lib/io:snappy_inputbuffer",
         "//tensorflow/core/lib/io:snappy_inputstream",
         "//tensorflow/core/lib/io:snappy_outputbuffer",
diff --git a/tensorflow/core/kernels/data/experimental/snapshot_util_test.cc b/tensorflow/core/kernels/data/experimental/snapshot_util_test.cc
index aedc0e194d7..e253014bf94 100644
--- a/tensorflow/core/kernels/data/experimental/snapshot_util_test.cc
+++ b/tensorflow/core/kernels/data/experimental/snapshot_util_test.cc
@@ -88,6 +88,7 @@ TEST(SnapshotUtilTest, CombinationRoundTripTest) {
 
   SnapshotRoundTrip(io::compression::kNone, 2);
   SnapshotRoundTrip(io::compression::kGzip, 2);
+  SnapshotRoundTrip(io::compression::kSnappy, 2);
 }
 
 void SnapshotReaderBenchmarkLoop(int iters, std::string compression_type,
@@ -195,11 +196,16 @@ void SnapshotTFRecordWriterGzipBenchmark(int iters) {
   SnapshotWriterBenchmarkLoop(iters, io::compression::kGzip, 2);
 }
 
+void SnapshotTFRecordWriterSnappyBenchmark(int iters) {
+  SnapshotWriterBenchmarkLoop(iters, io::compression::kSnappy, 2);
+}
+
 BENCHMARK(SnapshotCustomWriterNoneBenchmark);
 BENCHMARK(SnapshotCustomWriterGzipBenchmark);
 BENCHMARK(SnapshotCustomWriterSnappyBenchmark);
 BENCHMARK(SnapshotTFRecordWriterNoneBenchmark);
 BENCHMARK(SnapshotTFRecordWriterGzipBenchmark);
+BENCHMARK(SnapshotTFRecordWriterSnappyBenchmark);
 
 }  // namespace
 }  // namespace snapshot_util
diff --git a/tensorflow/core/lib/io/BUILD b/tensorflow/core/lib/io/BUILD
index 5e1704a50c1..797e9ad1a4b 100644
--- a/tensorflow/core/lib/io/BUILD
+++ b/tensorflow/core/lib/io/BUILD
@@ -145,6 +145,8 @@ cc_library(
         ":compression",
         ":inputstream_interface",
         ":random_inputstream",
+        ":snappy_compression_options",
+        ":snappy_inputstream",
         ":zlib_compression_options",
         ":zlib_inputstream",
         "//tensorflow/core/lib/core:coding",
@@ -164,6 +166,8 @@ cc_library(
     hdrs = ["record_writer.h"],
     deps = [
         ":compression",
+        ":snappy_compression_options",
+        ":snappy_outputbuffer",
         ":zlib_compression_options",
         ":zlib_outputbuffer",
         "//tensorflow/core/lib/core:coding",
@@ -221,6 +225,15 @@ cc_library(
     alwayslink = True,
 )
 
+cc_library(
+    name = "snappy_compression_options",
+    hdrs = ["snappy/snappy_compression_options.h"],
+    deps = [
+        "//tensorflow/core/platform:types",
+    ],
+    alwayslink = True,
+)
+
 cc_library(
     name = "cache",
     srcs = [
@@ -336,6 +349,9 @@ filegroup(
         "random_inputstream.h",
         "record_reader.cc",
         "record_reader.h",
+        "snappy/snappy_compression_options.h",
+        "snappy/snappy_inputstream.cc",
+        "snappy/snappy_inputstream.h",
         "table.cc",
         "table.h",
         "table_builder.cc",
@@ -366,6 +382,7 @@ filegroup(
         "random_inputstream.h",
         "record_reader.h",
         "record_writer.h",
+        "snappy/snappy_compression_options.h",
         "snappy/snappy_inputbuffer.h",
         "snappy/snappy_inputstream.h",
         "snappy/snappy_outputbuffer.h",
@@ -422,6 +439,7 @@ filegroup(
     srcs = [
         "inputbuffer.h",
         "iterator.h",
+        "snappy/snappy_compression_options.h",
         "snappy/snappy_inputbuffer.h",
         "snappy/snappy_inputstream.h",
         "snappy/snappy_outputbuffer.h",
diff --git a/tensorflow/core/lib/io/record_reader.cc b/tensorflow/core/lib/io/record_reader.cc
index 1af81bd902c..40e516f5ef9 100644
--- a/tensorflow/core/lib/io/record_reader.cc
+++ b/tensorflow/core/lib/io/record_reader.cc
@@ -31,26 +31,26 @@ namespace io {
 RecordReaderOptions RecordReaderOptions::CreateRecordReaderOptions(
     const string& compression_type) {
   RecordReaderOptions options;
+
+#if defined(IS_SLIM_BUILD)
+  if (compression_type != compression::kNone) {
+    LOG(ERROR) << "Compression is not supported but compression_type is set."
+               << " No compression will be used.";
+  }
+#else
   if (compression_type == compression::kZlib) {
     options.compression_type = io::RecordReaderOptions::ZLIB_COMPRESSION;
-#if defined(IS_SLIM_BUILD)
-    LOG(ERROR) << "Compression is not supported but compression_type is set."
-               << " No compression will be used.";
-#else
     options.zlib_options = io::ZlibCompressionOptions::DEFAULT();
-#endif  // IS_SLIM_BUILD
   } else if (compression_type == compression::kGzip) {
     options.compression_type = io::RecordReaderOptions::ZLIB_COMPRESSION;
-#if defined(IS_SLIM_BUILD)
-    LOG(ERROR) << "Compression is not supported but compression_type is set."
-               << " No compression will be used.";
-#else
     options.zlib_options = io::ZlibCompressionOptions::GZIP();
-#endif  // IS_SLIM_BUILD
+  } else if (compression_type == compression::kSnappy) {
+    options.compression_type = io::RecordReaderOptions::SNAPPY_COMPRESSION;
   } else if (compression_type != compression::kNone) {
     LOG(ERROR) << "Unsupported compression_type:" << compression_type
                << ". No compression will be used.";
   }
+#endif
   return options;
 }
 
@@ -63,20 +63,26 @@ RecordReader::RecordReader(RandomAccessFile* file,
     input_stream_.reset(new BufferedInputStream(input_stream_.release(),
                                                 options.buffer_size, true));
   }
-  if (options.compression_type == RecordReaderOptions::ZLIB_COMPRESSION) {
-// We don't have zlib available on all embedded platforms, so fail.
 #if defined(IS_SLIM_BUILD)
-    LOG(FATAL) << "Zlib compression is unsupported on mobile platforms.";
-#else   // IS_SLIM_BUILD
+  if (options.compression_type != RecordReaderOptions::NONE) {
+    LOG(FATAL) << "Compression is unsupported on mobile platforms.";
+  }
+#else
+  if (options.compression_type == RecordReaderOptions::ZLIB_COMPRESSION) {
     input_stream_.reset(new ZlibInputStream(
         input_stream_.release(), options.zlib_options.input_buffer_size,
         options.zlib_options.output_buffer_size, options.zlib_options, true));
-#endif  // IS_SLIM_BUILD
+  } else if (options.compression_type ==
+             RecordReaderOptions::SNAPPY_COMPRESSION) {
+    input_stream_.reset(
+        new SnappyInputStream(input_stream_.release(),
+                              options.snappy_options.output_buffer_size, true));
   } else if (options.compression_type == RecordReaderOptions::NONE) {
     // Nothing to do.
   } else {
     LOG(FATAL) << "Unrecognized compression type :" << options.compression_type;
   }
+#endif
 }
 
 // Read n+4 bytes from file, verify that checksum of first n bytes is
diff --git a/tensorflow/core/lib/io/record_reader.h b/tensorflow/core/lib/io/record_reader.h
index dd7def79f05..07709990a64 100644
--- a/tensorflow/core/lib/io/record_reader.h
+++ b/tensorflow/core/lib/io/record_reader.h
@@ -20,6 +20,8 @@ limitations under the License.
 #include "tensorflow/core/lib/core/stringpiece.h"
 #include "tensorflow/core/lib/io/inputstream_interface.h"
 #if !defined(IS_SLIM_BUILD)
+#include "tensorflow/core/lib/io/snappy/snappy_compression_options.h"
+#include "tensorflow/core/lib/io/snappy/snappy_inputstream.h"
 #include "tensorflow/core/lib/io/zlib_compression_options.h"
 #include "tensorflow/core/lib/io/zlib_inputstream.h"
 #endif  // IS_SLIM_BUILD
@@ -32,9 +34,12 @@ class RandomAccessFile;
 
 namespace io {
 
-class RecordReaderOptions {
- public:
-  enum CompressionType { NONE = 0, ZLIB_COMPRESSION = 1 };
+struct RecordReaderOptions {
+  enum CompressionType {
+    NONE = 0,
+    ZLIB_COMPRESSION = 1,
+    SNAPPY_COMPRESSION = 2
+  };
   CompressionType compression_type = NONE;
 
   // If buffer_size is non-zero, then all reads must be sequential, and no
@@ -46,8 +51,9 @@ class RecordReaderOptions {
       const string& compression_type);
 
 #if !defined(IS_SLIM_BUILD)
-  // Options specific to zlib compression.
+  // Options specific to compression.
   ZlibCompressionOptions zlib_options;
+  SnappyCompressionOptions snappy_options;
 #endif  // IS_SLIM_BUILD
 };
 
diff --git a/tensorflow/core/lib/io/record_reader_writer_test.cc b/tensorflow/core/lib/io/record_reader_writer_test.cc
index 373c0d8b664..486b238bd29 100644
--- a/tensorflow/core/lib/io/record_reader_writer_test.cc
+++ b/tensorflow/core/lib/io/record_reader_writer_test.cc
@@ -158,6 +158,44 @@ TEST(RecordReaderWriterTest, TestBasics) {
   }
 }
 
+TEST(RecordReaderWriterTest, TestSnappy) {
+  Env* env = Env::Default();
+  string fname = testing::TmpDir() + "/record_reader_writer_snappy_test";
+
+  for (auto buf_size : BufferSizes()) {
+    // Snappy compression needs output buffer size > 1.
+    if (buf_size == 1) continue;
+    {
+      std::unique_ptr file;
+      TF_CHECK_OK(env->NewWritableFile(fname, &file));
+
+      io::RecordWriterOptions options;
+      options.compression_type = io::RecordWriterOptions::SNAPPY_COMPRESSION;
+      options.zlib_options.output_buffer_size = buf_size;
+      io::RecordWriter writer(file.get(), options);
+      TF_EXPECT_OK(writer.WriteRecord("abc"));
+      TF_EXPECT_OK(writer.WriteRecord("defg"));
+      TF_CHECK_OK(writer.Flush());
+    }
+
+    {
+      std::unique_ptr read_file;
+      // Read it back with the RecordReader.
+      TF_CHECK_OK(env->NewRandomAccessFile(fname, &read_file));
+      io::RecordReaderOptions options;
+      options.compression_type = io::RecordReaderOptions::SNAPPY_COMPRESSION;
+      options.zlib_options.input_buffer_size = buf_size;
+      io::RecordReader reader(read_file.get(), options);
+      uint64 offset = 0;
+      tstring record;
+      TF_CHECK_OK(reader.ReadRecord(&offset, &record));
+      EXPECT_EQ("abc", record);
+      TF_CHECK_OK(reader.ReadRecord(&offset, &record));
+      EXPECT_EQ("defg", record);
+    }
+  }
+}
+
 TEST(RecordReaderWriterTest, TestZlib) {
   Env* env = Env::Default();
   string fname = testing::TmpDir() + "/record_reader_writer_zlib_test";
diff --git a/tensorflow/core/lib/io/record_writer.cc b/tensorflow/core/lib/io/record_writer.cc
index 52d0ef9a358..c82963d40c2 100644
--- a/tensorflow/core/lib/io/record_writer.cc
+++ b/tensorflow/core/lib/io/record_writer.cc
@@ -23,45 +23,49 @@ limitations under the License.
 namespace tensorflow {
 namespace io {
 namespace {
-bool IsZlibCompressed(RecordWriterOptions options) {
+bool IsZlibCompressed(const RecordWriterOptions& options) {
   return options.compression_type == RecordWriterOptions::ZLIB_COMPRESSION;
 }
+
+bool IsSnappyCompressed(const RecordWriterOptions& options) {
+  return options.compression_type == RecordWriterOptions::SNAPPY_COMPRESSION;
+}
 }  // namespace
 
 RecordWriterOptions RecordWriterOptions::CreateRecordWriterOptions(
     const string& compression_type) {
   RecordWriterOptions options;
+#if defined(IS_SLIM_BUILD)
+  if (compression_type != compression::kNone) {
+    LOG(ERROR) << "Compression is not supported but compression_type is set."
+               << " No compression will be used.";
+  }
+#else
   if (compression_type == compression::kZlib) {
     options.compression_type = io::RecordWriterOptions::ZLIB_COMPRESSION;
-#if defined(IS_SLIM_BUILD)
-    LOG(ERROR) << "Compression is not supported but compression_type is set."
-               << " No compression will be used.";
-#else
     options.zlib_options = io::ZlibCompressionOptions::DEFAULT();
-#endif  // IS_SLIM_BUILD
   } else if (compression_type == compression::kGzip) {
     options.compression_type = io::RecordWriterOptions::ZLIB_COMPRESSION;
-#if defined(IS_SLIM_BUILD)
-    LOG(ERROR) << "Compression is not supported but compression_type is set."
-               << " No compression will be used.";
-#else
     options.zlib_options = io::ZlibCompressionOptions::GZIP();
-#endif  // IS_SLIM_BUILD
+  } else if (compression_type == compression::kSnappy) {
+    options.compression_type = io::RecordWriterOptions::SNAPPY_COMPRESSION;
   } else if (compression_type != compression::kNone) {
     LOG(ERROR) << "Unsupported compression_type:" << compression_type
                << ". No compression will be used.";
   }
+#endif
   return options;
 }
 
 RecordWriter::RecordWriter(WritableFile* dest,
                            const RecordWriterOptions& options)
     : dest_(dest), options_(options) {
-  if (IsZlibCompressed(options)) {
-// We don't have zlib available on all embedded platforms, so fail.
 #if defined(IS_SLIM_BUILD)
-    LOG(FATAL) << "Zlib compression is unsupported on mobile platforms.";
-#else   // IS_SLIM_BUILD
+  if (compression_type != compression::kNone) {
+    LOG(FATAL) << "Compression is unsupported on mobile platforms.";
+  }
+#else
+  if (IsZlibCompressed(options)) {
     ZlibOutputBuffer* zlib_output_buffer = new ZlibOutputBuffer(
         dest, options.zlib_options.input_buffer_size,
         options.zlib_options.output_buffer_size, options.zlib_options);
@@ -71,12 +75,16 @@ RecordWriter::RecordWriter(WritableFile* dest,
                  << s.ToString();
     }
     dest_ = zlib_output_buffer;
-#endif  // IS_SLIM_BUILD
+  } else if (IsSnappyCompressed(options)) {
+    dest_ =
+        new SnappyOutputBuffer(dest, options.snappy_options.input_buffer_size,
+                               options.snappy_options.output_buffer_size);
   } else if (options.compression_type == RecordWriterOptions::NONE) {
     // Nothing to do
   } else {
     LOG(FATAL) << "Unspecified compression type :" << options.compression_type;
   }
+#endif
 }
 
 RecordWriter::~RecordWriter() {
@@ -130,14 +138,12 @@ Status RecordWriter::WriteRecord(const absl::Cord& data) {
 
 Status RecordWriter::Close() {
   if (dest_ == nullptr) return Status::OK();
-#if !defined(IS_SLIM_BUILD)
-  if (IsZlibCompressed(options_)) {
+  if (IsZlibCompressed(options_) || IsSnappyCompressed(options_)) {
     Status s = dest_->Close();
     delete dest_;
     dest_ = nullptr;
     return s;
   }
-#endif  // IS_SLIM_BUILD
   return Status::OK();
 }
 
diff --git a/tensorflow/core/lib/io/record_writer.h b/tensorflow/core/lib/io/record_writer.h
index 012c2fbbc91..243dc847ec5 100644
--- a/tensorflow/core/lib/io/record_writer.h
+++ b/tensorflow/core/lib/io/record_writer.h
@@ -21,6 +21,8 @@ limitations under the License.
 #include "tensorflow/core/lib/core/stringpiece.h"
 #include "tensorflow/core/lib/hash/crc32c.h"
 #if !defined(IS_SLIM_BUILD)
+#include "tensorflow/core/lib/io/snappy/snappy_compression_options.h"
+#include "tensorflow/core/lib/io/snappy/snappy_outputbuffer.h"
 #include "tensorflow/core/lib/io/zlib_compression_options.h"
 #include "tensorflow/core/lib/io/zlib_outputbuffer.h"
 #endif  // IS_SLIM_BUILD
@@ -34,17 +36,22 @@ class WritableFile;
 
 namespace io {
 
-class RecordWriterOptions {
+struct RecordWriterOptions {
  public:
-  enum CompressionType { NONE = 0, ZLIB_COMPRESSION = 1 };
+  enum CompressionType {
+    NONE = 0,
+    ZLIB_COMPRESSION = 1,
+    SNAPPY_COMPRESSION = 2
+  };
   CompressionType compression_type = NONE;
 
   static RecordWriterOptions CreateRecordWriterOptions(
       const string& compression_type);
 
-// Options specific to zlib compression.
 #if !defined(IS_SLIM_BUILD)
+  // Options specific to compression.
   tensorflow::io::ZlibCompressionOptions zlib_options;
+  tensorflow::io::SnappyCompressionOptions snappy_options;
 #endif  // IS_SLIM_BUILD
 };
 
@@ -70,7 +77,7 @@ class RecordWriter {
   // implicit Close() call in the destructor.
   ~RecordWriter();
 
-  Status WriteRecord(StringPiece slice);
+  Status WriteRecord(StringPiece data);
 
 #if defined(PLATFORM_GOOGLE)
   Status WriteRecord(const absl::Cord& data);
diff --git a/tensorflow/core/lib/io/snappy/snappy_compression_options.h b/tensorflow/core/lib/io/snappy/snappy_compression_options.h
new file mode 100644
index 00000000000..d3d798bfa8f
--- /dev/null
+++ b/tensorflow/core/lib/io/snappy/snappy_compression_options.h
@@ -0,0 +1,36 @@
+/* Copyright 2020 The TensorFlow Authors. All Rights Reserved.
+
+Licensed under the Apache License, Version 2.0 (the "License");
+you may not use this file except in compliance with the License.
+You may obtain a copy of the License at
+
+    http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software
+distributed under the License is distributed on an "AS IS" BASIS,
+WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+See the License for the specific language governing permissions and
+limitations under the License.
+==============================================================================*/
+
+#ifndef TENSORFLOW_CORE_LIB_IO_SNAPPY_SNAPPY_COMPRESSION_OPTIONS_H_
+#define TENSORFLOW_CORE_LIB_IO_SNAPPY_SNAPPY_COMPRESSION_OPTIONS_H_
+
+#include "tensorflow/core/platform/types.h"
+
+namespace tensorflow {
+namespace io {
+
+struct SnappyCompressionOptions {
+  // Size of the buffer used for caching the data read from source file.
+  int64 input_buffer_size = 256 << 10;
+
+  // Size of the sink buffer where the compressed/decompressed data produced by
+  // snappy is cached.
+  int64 output_buffer_size = 256 << 10;
+};
+
+}  // namespace io
+}  // namespace tensorflow
+
+#endif  // TENSORFLOW_CORE_LIB_IO_SNAPPY_SNAPPY_COMPRESSION_OPTIONS_H_

From 0a946a0069d1964d19b6b42f0ca35161bf4c7843 Mon Sep 17 00:00:00 2001
From: Yi Situ 
Date: Mon, 22 Jun 2020 12:02:30 -0700
Subject: [PATCH 455/867] Misc ClangTidy cleanups.

PiperOrigin-RevId: 317705032
Change-Id: I94c55b6d083c97f9962567ee28c2e952958a5ea8
---
 tensorflow/core/profiler/internal/gpu/cupti_tracer.cc  | 7 ++++---
 tensorflow/core/profiler/internal/gpu/device_tracer.cc | 2 +-
 2 files changed, 5 insertions(+), 4 deletions(-)

diff --git a/tensorflow/core/profiler/internal/gpu/cupti_tracer.cc b/tensorflow/core/profiler/internal/gpu/cupti_tracer.cc
index c9c626b7289..931801427e7 100644
--- a/tensorflow/core/profiler/internal/gpu/cupti_tracer.cc
+++ b/tensorflow/core/profiler/internal/gpu/cupti_tracer.cc
@@ -18,6 +18,7 @@ limitations under the License.
 #include "absl/container/flat_hash_map.h"
 #include "absl/container/flat_hash_set.h"
 #include "absl/container/node_hash_map.h"
+#include "absl/container/node_hash_set.h"
 #include "tensorflow/core/platform/env.h"
 #include "tensorflow/core/platform/errors.h"
 #include "tensorflow/core/platform/host_info.h"
@@ -683,7 +684,7 @@ Status CreateAndRecordEvent(CUevent *event, CUstream stream) {
 // Note: cuStreamGetCtx only available after CUDA 9.2.
 class ScopedCudaContext {
  public:
-  ScopedCudaContext(CUstream stream) : stream_(stream) {
+  explicit ScopedCudaContext(CUstream stream) : stream_(stream) {
     CuptiApiTracingDisabler disabler;  // don't trace cuda call in this func.
     CUcontext context;
     if (cuStreamGetCtx(stream, &context) != CUDA_SUCCESS) return;
@@ -1244,7 +1245,7 @@ class CuptiDriverApiHookWithCudaEvent : public CuptiDriverApiHook {
   // However there is no guarantee that we receive such callbacks in pairs, we
   // maintain a on-going API calls to make sure no memory leaks.
   struct CuptiApiCallbackContext {
-    CuptiApiCallbackContext(std::vector &&r)
+    explicit CuptiApiCallbackContext(std::vector &&r)
         : record_indices(std::move(r)) {}
     std::vector record_indices;
   };
@@ -1252,7 +1253,7 @@ class CuptiDriverApiHookWithCudaEvent : public CuptiDriverApiHook {
   const CuptiTracerOptions option_;
   CuptiInterface *cupti_interface_;
   CuptiTraceCollector *collector_;
-  std::set callback_contexts_;
+  absl::node_hash_set callback_contexts_;
   std::vector> cuda_event_recorders_;
   TF_DISALLOW_COPY_AND_ASSIGN(CuptiDriverApiHookWithCudaEvent);
 };
diff --git a/tensorflow/core/profiler/internal/gpu/device_tracer.cc b/tensorflow/core/profiler/internal/gpu/device_tracer.cc
index 9c3e2d67bf0..3c0ac04caf2 100644
--- a/tensorflow/core/profiler/internal/gpu/device_tracer.cc
+++ b/tensorflow/core/profiler/internal/gpu/device_tracer.cc
@@ -241,7 +241,7 @@ class CuptiTraceCollectorImpl : public CuptiTraceCollector {
   std::string ReportDroppedEvents() {
     absl::MutexLock lock(&mutex_);
     string result;
-    for (const auto dropped : dropped_events_) {
+    for (const auto& dropped : dropped_events_) {
       absl::StrAppend(&result, " ", dropped.second, " events dropped because ",
                       dropped.first, ";");
     }

From 9ca89a201cf56b475b97197bb5605126e225038f Mon Sep 17 00:00:00 2001
From: George Karpenkov 
Date: Mon, 22 Jun 2020 12:07:25 -0700
Subject: [PATCH 456/867] [TF2XLA] [NFC] Refactor datastructures for resource
 variables to not require snapshotting for compilation

Previously, `BuildXlaCompilerArguments` required taking a snapshot of all
resource variables in order to start compiling.
In this CL it can operate with a span of pointers to resource variables instead
(we CHECK at runtime that the lock is held).
That refactoring allows to launch the XLA compilation without creating an extra
reference to the underlying Tensor of the passed resource variables.

PiperOrigin-RevId: 317706126
Change-Id: I37a97601a08f165b23b4745e1b032bf91c21c313
---
 tensorflow/compiler/jit/kernels/xla_ops.cc    | 49 ++++++----
 .../compiler/jit/xla_compile_on_demand_op.cc  | 55 ++++++-----
 .../compiler/jit/xla_compile_on_demand_op.h   |  5 +-
 tensorflow/compiler/jit/xla_launch_util.cc    | 91 ++++++++++---------
 tensorflow/compiler/jit/xla_launch_util.h     | 61 +++++++------
 5 files changed, 141 insertions(+), 120 deletions(-)

diff --git a/tensorflow/compiler/jit/kernels/xla_ops.cc b/tensorflow/compiler/jit/kernels/xla_ops.cc
index 0fc1a349adc..e3542586c89 100644
--- a/tensorflow/compiler/jit/kernels/xla_ops.cc
+++ b/tensorflow/compiler/jit/kernels/xla_ops.cc
@@ -108,8 +108,7 @@ class XlaExecutableClosure {
   explicit XlaExecutableClosure(
       xla::LocalClient* client, xla::LocalExecutable* executable,
       const XlaCompiler::CompilationResult* compilation_result,
-      std::map resource_var_snapshots,
-      int num_constant_args)
+      ResourceVarsSnapshot resource_var_snapshots, int num_constant_args)
       : client_(client),
         executable_(executable),
         compilation_result_(compilation_result),
@@ -124,7 +123,7 @@ class XlaExecutableClosure {
   const XlaCompiler::CompilationResult* compilation_result() const {
     return compilation_result_;
   }
-  const std::map& resource_var_snapshots() const {
+  const ResourceVarsSnapshot& resource_var_snapshots() const {
     return resource_var_snapshots_;
   }
   int num_constant_args() const { return num_constant_args_; }
@@ -133,7 +132,7 @@ class XlaExecutableClosure {
   xla::LocalClient* client_;
   xla::LocalExecutable* executable_;
   const XlaCompiler::CompilationResult* compilation_result_;
-  std::map resource_var_snapshots_;
+  ResourceVarsSnapshot resource_var_snapshots_;
   int num_constant_args_;
 
   TF_DISALLOW_COPY_AND_ASSIGN(XlaExecutableClosure);
@@ -276,10 +275,10 @@ static Status BuildCompilationCache(OpKernelContext* ctx,
 
 static Status CompileToLocalExecutable(
     OpKernelContext* ctx, const NameAttrList& function, bool has_ref_vars,
-    const XlaPlatformInfo& platform_info, absl::Span resources,
+    const XlaPlatformInfo& platform_info,
+    absl::Span variable_infos,
     absl::Span constants, bool lazy, xla::LocalClient** client,
-    std::map* variables,
-    const XlaCompiler::CompilationResult** kernel,
+    const XlaCompiler::CompilationResult** compilation_result,
     xla::LocalExecutable** executable) {
   // We store information about the JIT-compiled XLA computation
   // in the ResourceMgr.
@@ -299,7 +298,6 @@ static Status CompileToLocalExecutable(
   // this is more obviously correct.)
   core::ScopedUnref cache_ref(cache);
 
-  TF_RETURN_IF_ERROR(SnapshotResourceVariables(ctx, resources, variables));
   *client = static_cast(cache->client());
 
   absl::optional tf_allocator_adapter;
@@ -337,11 +335,11 @@ static Status CompileToLocalExecutable(
 
   std::vector args;
   TF_RETURN_IF_ERROR(XlaComputationLaunchContext::BuildXlaCompilerArguments(
-      constant_args, *variables, ctx, &args));
+      constant_args, variable_infos, ctx, &args));
   return cache->Compile(options, function, args, compile_options,
                         lazy ? XlaCompilationCache::CompileMode::kLazy
                              : XlaCompilationCache::CompileMode::kStrict,
-                        kernel, executable);
+                        compilation_result, executable);
 }
 
 void XlaLocalLaunchBase::Compute(OpKernelContext* ctx) {
@@ -349,16 +347,22 @@ void XlaLocalLaunchBase::Compute(OpKernelContext* ctx) {
           << Canonicalize(function_.name(), AttrSlice(&function_.attr()));
 
   xla::LocalClient* client;
-  const XlaCompiler::CompilationResult* kernel;
+  const XlaCompiler::CompilationResult* compilation_result;
   xla::LocalExecutable* executable;
-  std::map variables;
 
+  ResourceVarsSnapshot variables;
   {
+    std::vector variable_infos;
+    OP_REQUIRES_OK(
+        ctx, GetVariableInfosFromCtxInputs(ctx, resources_, &variable_infos));
+    OP_REQUIRES_OK(ctx, LockVariables(absl::MakeSpan(variable_infos)));
     Status s = CompileToLocalExecutable(
         ctx, function_, /*has_ref_vars=*/has_ref_vars_, platform_info_,
-        resources_, constants_, /*lazy=*/false, &client, &variables, &kernel,
-        &executable);
+        variable_infos, constants_, /*lazy=*/false, &client,
+        &compilation_result, &executable);
     OP_REQUIRES_OK(ctx, s);
+    OP_REQUIRES_OK(ctx, SnapshotResourceVariables(ctx, resources_,
+                                                  variable_infos, &variables));
   }
 
   se::Stream* stream =
@@ -373,7 +377,7 @@ void XlaLocalLaunchBase::Compute(OpKernelContext* ctx) {
       client, allocator,
       /*allocate_xla_tensors=*/platform_info_.is_on_xla_device(),
       platform_info_.UseMultipleStreams());
-  launch_context.PopulateInputs(ctx, kernel, variables,
+  launch_context.PopulateInputs(ctx, compilation_result, variables,
                                 /*missing_ctx_input_prefix=*/0);
 
   // Execute the computation.
@@ -413,7 +417,7 @@ void XlaLocalLaunchBase::Compute(OpKernelContext* ctx) {
       executable->executable()->module().input_output_alias_config();
   OP_REQUIRES_OK(
       ctx, launch_context.PopulateOutputs(
-               ctx, kernel, run_result.ConsumeValueOrDie(),
+               ctx, compilation_result, run_result.ConsumeValueOrDie(),
                /*missing_ctx_input_prefix=*/0, input_output_alias, variables));
   VLOG(1) << "Done";
 }
@@ -494,7 +498,7 @@ void XlaCompileOp::Compute(OpKernelContext* ctx) {
   xla::LocalClient* client;
   const XlaCompiler::CompilationResult* kernel;
   xla::LocalExecutable* executable;
-  std::map variables;
+  ResourceVarsSnapshot variables;
 
   bool cannot_compile_cluster;
   {
@@ -506,9 +510,16 @@ void XlaCompileOp::Compute(OpKernelContext* ctx) {
       cannot_compile_cluster) {
     executable = nullptr;
   } else {
+    std::vector variable_infos;
+    OP_REQUIRES_OK(
+        ctx, GetVariableInfosFromCtxInputs(ctx, resources_, &variable_infos));
+    OP_REQUIRES_OK(ctx, LockVariables(absl::MakeSpan(variable_infos)));
     Status status = CompileToLocalExecutable(
-        ctx, function_, has_ref_vars_, platform_info_, resources_, constants_,
-        /*lazy=*/!must_compile_, &client, &variables, &kernel, &executable);
+        ctx, function_, has_ref_vars_, platform_info_, variable_infos,
+        constants_,
+        /*lazy=*/!must_compile_, &client, &kernel, &executable);
+    OP_REQUIRES_OK(ctx, SnapshotResourceVariables(ctx, resources_,
+                                                  variable_infos, &variables));
     if (must_compile_ || status.code() != error::UNIMPLEMENTED) {
       OP_REQUIRES_OK(ctx, status);
     }
diff --git a/tensorflow/compiler/jit/xla_compile_on_demand_op.cc b/tensorflow/compiler/jit/xla_compile_on_demand_op.cc
index e1ad0e8c5af..afaee614f02 100644
--- a/tensorflow/compiler/jit/xla_compile_on_demand_op.cc
+++ b/tensorflow/compiler/jit/xla_compile_on_demand_op.cc
@@ -28,32 +28,23 @@ limitations under the License.
 
 namespace tensorflow {
 
-namespace {
-std::map GetVariables(OpKernelContext* ctx) {
-  std::map variables;
-  for (int64 i = 0; i < ctx->num_inputs(); ++i) {
+// Returns argument indices corresponding to the resource variable inputs of
+// kernel context `ctx`.
+static std::vector GetResourceVariableIndices(OpKernelContext* ctx) {
+  std::vector out;
+  for (int64 i = 0; i < ctx->num_inputs(); i++) {
     if (ctx->input(i).dtype() == DT_RESOURCE) {
-      core::RefCountPtr variable;
-      ResourceHandle handle = HandleFromInput(ctx, i);
-      OptionalTensor& optional = variables[i];
-      optional.name = handle.name();
-      if (LookupResource(ctx, handle, &variable).ok()) {
-        tf_shared_lock lock(*variable->mu());
-        optional.present = true;
-        optional.value = *variable->tensor();
-      }
+      out.push_back(i);
     }
   }
-  return variables;
+  return out;
 }
-}  // namespace
 
 Status XlaCompileOnDemandOp::Run(OpKernelContext* ctx,
                                  const XlaDevice::Metadata& metadata,
                                  const XlaCompiler::CompilationResult* result,
-                                 xla::LocalExecutable* executable) {
-  std::map variables = GetVariables(ctx);
-
+                                 xla::LocalExecutable* executable,
+                                 const ResourceVarsSnapshot& variable_args) {
   xla::LocalClient* client = metadata.client();
 
   // Builds an XLA allocator for the device.
@@ -62,7 +53,7 @@ Status XlaCompileOnDemandOp::Run(OpKernelContext* ctx,
       /*allocate_xla_tensors=*/true,
       /*use_multiple_streams=*/metadata.UseMultipleStreams());
 
-  launch_context.PopulateInputs(ctx, result, variables,
+  launch_context.PopulateInputs(ctx, result, variable_args,
                                 /*missing_ctx_input_prefix=*/0);
 
   se::Stream* stream =
@@ -87,7 +78,7 @@ Status XlaCompileOnDemandOp::Run(OpKernelContext* ctx,
       executable->executable()->module().input_output_alias_config();
   TF_RETURN_IF_ERROR(launch_context.PopulateOutputs(
       ctx, result, run_result.ConsumeValueOrDie(),
-      /*missing_ctx_input_prefix=*/0, input_output_alias, variables));
+      /*missing_ctx_input_prefix=*/0, input_output_alias, variable_args));
   return Status::OK();
 }
 
@@ -115,7 +106,7 @@ Status XlaCompileOnDemandOp::ShouldArgumentBeConstant(
 Status XlaCompileOnDemandOp::Compile(
     OpKernelContext* ctx, const XlaDevice::Metadata& metadata,
     const XlaCompiler::CompilationResult** result,
-    xla::LocalExecutable** executable) {
+    ResourceVarsSnapshot* variable_args, xla::LocalExecutable** executable) {
   std::map constant_arguments;
   for (int64 i = 0; i < ctx->num_inputs(); ++i) {
     const Tensor& device_tensor = ctx->input(i);
@@ -190,12 +181,18 @@ Status XlaCompileOnDemandOp::Compile(
   // rather than a one-element tuple.
   compile_options.always_return_tuple = false;
 
-  std::map variable_args = GetVariables(ctx);
-
+  std::vector variables_indices = GetResourceVariableIndices(ctx);
   std::vector args;
-
-  TF_RETURN_IF_ERROR(XlaComputationLaunchContext::BuildXlaCompilerArguments(
-      constant_arguments, variable_args, ctx, &args));
+  {
+    std::vector variable_infos;
+    TF_RETURN_IF_ERROR(
+        GetVariableInfosFromCtxInputs(ctx, variables_indices, &variable_infos));
+    TF_RETURN_IF_ERROR(LockVariables(absl::MakeSpan(variable_infos)));
+    TF_RETURN_IF_ERROR(SnapshotResourceVariables(
+        ctx, variables_indices, variable_infos, variable_args));
+    TF_RETURN_IF_ERROR(XlaComputationLaunchContext::BuildXlaCompilerArguments(
+        constant_arguments, variable_infos, ctx, &args));
+  }
 
   return cache->CompileSingleOp(options, args, ctx, compile_options, result,
                                 executable);
@@ -206,8 +203,10 @@ void XlaCompileOnDemandOp::Compute(OpKernelContext* ctx) {
   xla::LocalExecutable* executable;
   const XlaDevice::Metadata* metadata;
   OP_REQUIRES_OK(ctx, XlaDevice::GetMetadata(ctx, &metadata));
-  OP_REQUIRES_OK(ctx, Compile(ctx, *metadata, &result, &executable));
-  OP_REQUIRES_OK(ctx, Run(ctx, *metadata, result, executable));
+  ResourceVarsSnapshot variable_args;
+  OP_REQUIRES_OK(ctx,
+                 Compile(ctx, *metadata, &result, &variable_args, &executable));
+  OP_REQUIRES_OK(ctx, Run(ctx, *metadata, result, executable, variable_args));
 }
 
 }  // namespace tensorflow
diff --git a/tensorflow/compiler/jit/xla_compile_on_demand_op.h b/tensorflow/compiler/jit/xla_compile_on_demand_op.h
index 98f634db98f..cc5f2f1e42f 100644
--- a/tensorflow/compiler/jit/xla_compile_on_demand_op.h
+++ b/tensorflow/compiler/jit/xla_compile_on_demand_op.h
@@ -20,6 +20,7 @@ limitations under the License.
 #define TENSORFLOW_COMPILER_JIT_XLA_COMPILE_ON_DEMAND_OP_H_
 
 #include "tensorflow/compiler/jit/xla_device.h"
+#include "tensorflow/compiler/jit/xla_launch_util.h"
 #include "tensorflow/compiler/tf2xla/xla_compiler.h"
 #include "tensorflow/compiler/xla/client/local_client.h"
 #include "tensorflow/core/framework/function.h"
@@ -47,10 +48,12 @@ class XlaCompileOnDemandOp : public OpKernel {
                                 bool* result);
   Status Compile(OpKernelContext* ctx, const XlaDevice::Metadata& metadata,
                  const XlaCompiler::CompilationResult** result,
+                 ResourceVarsSnapshot* variable_args,
                  xla::LocalExecutable** executable);
   Status Run(OpKernelContext* ctx, const XlaDevice::Metadata& metadata,
              const XlaCompiler::CompilationResult* result,
-             xla::LocalExecutable* executable);
+             xla::LocalExecutable* executable,
+             const ResourceVarsSnapshot& variable_args);
 };
 
 }  // namespace tensorflow
diff --git a/tensorflow/compiler/jit/xla_launch_util.cc b/tensorflow/compiler/jit/xla_launch_util.cc
index 25eed134e35..fc0ff8d9445 100644
--- a/tensorflow/compiler/jit/xla_launch_util.cc
+++ b/tensorflow/compiler/jit/xla_launch_util.cc
@@ -52,7 +52,8 @@ const char kPossibleNonVariableResourceHintMessage[] =
     "resource inputs to XLA.";
 }  // anonymous namespace
 
-VariableInfo::VariableInfo(int index, Var* var) : index_(index), var_(var) {}
+VariableInfo::VariableInfo(int index, absl::string_view name, Var* var)
+    : index_(index), name_(name), var_(var) {}
 VariableInfo::VariableInfo(VariableInfo&& other)
     : index_(other.index_), var_(other.var_), lock_held_(other.lock_held_) {
   other.index_ = -1;
@@ -87,16 +88,15 @@ VariableInfo::~VariableInfo() {
 // Returns a vector of VariableInfo instances for the resource variable inputs
 // to the kernel with context `ctx`.  The input indices for the resource
 // variable inputs are in `variable_indices`.
-static Status GetVariableInfosFromCtxInputs(
-    OpKernelContext* ctx, absl::Span variable_indices,
-    std::vector* result) {
+Status GetVariableInfosFromCtxInputs(OpKernelContext* ctx,
+                                     absl::Span variable_indices,
+                                     std::vector* result) {
   std::vector resource_handles;
   absl::c_transform(
       variable_indices, std::back_inserter(resource_handles),
       [&](int variable_idx) { return &HandleFromInput(ctx, variable_idx); });
 
   std::vector> variables;
-
   Status s = LookupResources(ctx, resource_handles, &variables);
   if (!s.ok()) {
     errors::AppendToMessage(&s, kPossibleNonVariableResourceHintMessage);
@@ -109,7 +109,9 @@ static Status GetVariableInfosFromCtxInputs(
     // *Release* the variable because we're going to unref it later in
     // ~VariableInfo.
     Var* variable = variables[i].release();
-    result->emplace_back(variable_indices[i], variable);
+    int input_idx = variable_indices[i];
+    std::string var_name = HandleFromInput(ctx, input_idx).name();
+    result->emplace_back(input_idx, var_name, variable);
   }
 
   return Status::OK();
@@ -162,21 +164,12 @@ Status LockVariables(absl::Span variables) {
 
 Status SnapshotResourceVariables(OpKernelContext* ctx,
                                  absl::Span variable_indices,
-                                 std::map* result) {
-  std::vector variable_infos;
-  TF_RETURN_IF_ERROR(
-      GetVariableInfosFromCtxInputs(ctx, variable_indices, &variable_infos));
-  TF_RETURN_IF_ERROR(LockVariables(absl::MakeSpan(variable_infos)));
-
+                                 absl::Span variable_infos,
+                                 ResourceVarsSnapshot* result) {
   for (int i = 0; i < variable_indices.size(); i++) {
-    if (variable_infos[i].var()) {
-      OptionalTensor& tensor = (*result)[variable_indices[i]];
-      tensor.name = HandleFromInput(ctx, variable_indices[i]).name();
-      tensor.present = true;
-      tensor.value = *variable_infos[i].var()->tensor();
-    } else {
-      (*result)[variable_indices[i]] = OptionalTensor();
-    }
+    Var* var = variable_infos[i].var();
+    (*result)[variable_indices[i]] =
+        var ? absl::make_optional(*var->tensor()) : absl::nullopt;
   }
   return Status::OK();
 }
@@ -197,8 +190,7 @@ XlaComputationLaunchContext::XlaComputationLaunchContext(
 void XlaComputationLaunchContext::PopulateInputs(
     OpKernelContext* ctx,
     const XlaCompiler::CompilationResult* compilation_result,
-    const std::map& variables,
-    int missing_ctx_input_prefix) {
+    const ResourceVarsSnapshot& variables, int missing_ctx_input_prefix) {
   // Build ShapedBuffers that point directly to the Tensor buffers.
   arg_ptrs_ =
       std::vector(compilation_result->xla_input_shapes.size());
@@ -210,7 +202,7 @@ void XlaComputationLaunchContext::PopulateInputs(
     CHECK_GE(arg_num, missing_ctx_input_prefix);
     const xla::Shape& shape = compilation_result->xla_input_shapes[i];
     const Tensor* t = variables.count(arg_num)
-                          ? &(variables.at(arg_num).value)
+                          ? &(variables.at(arg_num).value())
                           : &(ctx->input(arg_num - missing_ctx_input_prefix));
     CHECK(t);
 
@@ -262,7 +254,7 @@ static const Tensor* FindAliasedTensorForOutput(
     int output_num, OpKernelContext* ctx, int missing_ctx_input_prefix,
     const xla::HloInputOutputAliasConfig& input_output_alias,
     absl::Span input_mapping,
-    const std::map& resource_var_snapshots) {
+    const ResourceVarsSnapshot& resource_var_snapshots) {
   if (MustAliasOutput(input_output_alias, output_num)) {
     int xla_param = input_output_alias.GetAliasedParameter({output_num})
                         .value()
@@ -274,8 +266,8 @@ static const Tensor* FindAliasedTensorForOutput(
     // entry time.
     if (input_tensor->dtype() == DT_RESOURCE) {
       auto& v = resource_var_snapshots.at(missing_ctx_input_prefix + tf_param);
-      CHECK(v.present);
-      return &v.value;
+      CHECK(v.has_value());
+      return &v.value();
     }
     return input_tensor;
   }
@@ -298,9 +290,9 @@ static Tensor GetOrCreateTensorForOutput(
     int output_num, OpKernelContext* ctx, int missing_ctx_input_prefix,
     const xla::HloInputOutputAliasConfig& input_output_alias,
     absl::Span input_mapping,
-    const std::map& resource_var_snapshots,
-    DataType output_dtype, const TensorShape& output_shape,
-    se::DeviceMemoryBase output_buffer, Allocator* output_allocator) {
+    const ResourceVarsSnapshot& resource_var_snapshots, DataType output_dtype,
+    const TensorShape& output_shape, se::DeviceMemoryBase output_buffer,
+    Allocator* output_allocator) {
   if (const Tensor* aliased_tensor = FindAliasedTensorForOutput(
           output_num, ctx, missing_ctx_input_prefix, input_output_alias,
           input_mapping, resource_var_snapshots)) {
@@ -431,13 +423,13 @@ static xla::StatusOr> GatherVariableInfo(
     // TODO(b/35625933): tensorflow::Var should contain a PersistentTensor,
     // not a Tensor.
     Var* variable = nullptr;
-    TF_RETURN_IF_ERROR(LookupOrCreateResource(
-        ctx, HandleFromInput(ctx, actual_input_index), &variable,
-        [&write](Var** ptr) {
-          *ptr = new Var(write.type);
-          return Status::OK();
-        }));
-    variable_infos.emplace_back(actual_input_index, variable);
+    const ResourceHandle handle = HandleFromInput(ctx, actual_input_index);
+    TF_RETURN_IF_ERROR(LookupOrCreateResource(ctx, handle, &variable,
+                                                   [&write](Var** ptr) {
+                                                     *ptr = new Var(write.type);
+                                                     return Status::OK();
+                                                   }));
+    variable_infos.emplace_back(actual_input_index, handle.name(), variable);
   }
   return variable_infos;
 }
@@ -447,7 +439,7 @@ Status XlaComputationLaunchContext::PopulateOutputs(
     const XlaCompiler::CompilationResult* compilation_result,
     ScopedShapedBuffer output, int missing_ctx_input_prefix,
     const xla::HloInputOutputAliasConfig& input_output_alias,
-    const std::map& resource_var_snapshots) {
+    const ResourceVarsSnapshot& resource_var_snapshots) {
   se::Stream* stream =
       ctx->op_device_context() ? ctx->op_device_context()->stream() : nullptr;
   Allocator* allocator = ctx->device()->GetAllocator({});
@@ -564,12 +556,21 @@ Status XlaComputationLaunchContext::PopulateOutputs(
 
 Status XlaComputationLaunchContext::BuildXlaCompilerArguments(
     const std::map& constant_args,
-    const std::map& variable_args, OpKernelContext* ctx,
+    absl::Span variable_args, OpKernelContext* ctx,
     std::vector* args) {
   args->resize(ctx->num_inputs());
 
+  absl::flat_hash_map variable_info_lookup;
+  for (const VariableInfo& info : variable_args) {
+    CHECK(!info.var() || info.lock_held())
+        << "Need to hold the lock on resource variables "
+           "before calling BuildXlaCompilerArguments";
+    variable_info_lookup.emplace(info.index(), &info);
+  }
+
   for (int64 input_num = 0; input_num < ctx->num_inputs(); ++input_num) {
     XlaCompiler::Argument& arg = (*args)[input_num];
+
     if (constant_args.count(input_num) > 0) {
       // Handles compile-time constants.
       const Tensor& input = constant_args.at(input_num);
@@ -578,7 +579,7 @@ Status XlaComputationLaunchContext::BuildXlaCompilerArguments(
       arg.type = input.dtype();
       arg.shape = input.shape();
       arg.constant_value = input;
-    } else if (variable_args.count(input_num) == 0) {
+    } else if (variable_info_lookup.count(input_num) == 0) {
       // Handles the non-constant arguments.
       const Tensor& input = ctx->input(input_num);
       TF_RET_CHECK(input.dtype() != DT_RESOURCE);
@@ -594,14 +595,14 @@ Status XlaComputationLaunchContext::BuildXlaCompilerArguments(
       // Handles resource variables.
       const Tensor& input = ctx->input(input_num);
       TF_RET_CHECK(input.dtype() == DT_RESOURCE);
-      const OptionalTensor& variable = variable_args.at(input_num);
-      arg.name = variable.name;
+      const VariableInfo& variable = *variable_info_lookup[input_num];
+      arg.name = std::string(variable.name());
       arg.kind = XlaCompiler::Argument::kResource;
       arg.resource_kind = XlaResource::kVariable;
-      if (variable.present) {
-        const Tensor& value = variable.value;
-        arg.type = value.dtype();
-        arg.shape = value.shape();
+      if (variable.var()) {
+        const Tensor* value = variable.var()->tensor();
+        arg.type = value->dtype();
+        arg.shape = value->shape();
         arg.initialized = true;
       } else {
         // The values of uninitialized variables are not passed as inputs, since
diff --git a/tensorflow/compiler/jit/xla_launch_util.h b/tensorflow/compiler/jit/xla_launch_util.h
index 9a7f20cb310..92b6c4c8a08 100644
--- a/tensorflow/compiler/jit/xla_launch_util.h
+++ b/tensorflow/compiler/jit/xla_launch_util.h
@@ -34,36 +34,17 @@ limitations under the License.
 
 namespace tensorflow {
 
-// Struct that represents a possibly-absent Tensor.
-struct OptionalTensor {
-  string name;           // A descriptive name
-  bool present = false;  // Is the tensor present?
-  Tensor value;          // If present, what is the Tensor's value?
-};
-
-// Takes a snapshot of the values of resource variable arguments, whose indices
-// are specified in `variable_indices` argument. We snapshot tensors that back
-// resource variables since concurrent updates may modify the shape, and it is
-// important that the shapes used for compilation match the true shapes of the
-// buffers.
-//
-// We snapshot the entire set of resource variables as one atomic operation.
-// This models Read->* dependencies between resource variable operations.  See
-// jit/resource_operation_safety_analysis for details.
-//
-// Returns a map of TensorFlow argument index to resource variable. If a
-// resource variable is not initialized, the corresponding OptionalTensor
-// will have its `present` field set to false.
-Status SnapshotResourceVariables(OpKernelContext* ctx,
-                                 absl::Span variable_indices,
-                                 std::map* result);
+// Snapshot of resource variables for a TF kernel invocation, mapping from
+// parameter number to values at execution time. If the resource variable is not
+// initialized, the value will not be present.
+using ResourceVarsSnapshot = absl::flat_hash_map>;
 
 // Information about the state of a variable passed as input to the _XlaCompile
 // and _XlaRun operators.  Unlocks the resource variable and decrements its
 // refcount on destruction.
 class VariableInfo {
  public:
-  explicit VariableInfo(int index, Var* var);
+  explicit VariableInfo(int index, absl::string_view name, Var* var);
   VariableInfo(VariableInfo&& other);
 
   VariableInfo& operator=(VariableInfo&& other);
@@ -79,6 +60,9 @@ class VariableInfo {
   // "empty", i.e. it does not track a resource variable.
   Var* var() const { return var_; }
 
+  // Returns the variable name.
+  absl::string_view name() const { return name_; }
+
   // Returns true if the resource variable lock was successfully acquired by
   // this thread.
   bool lock_held() const { return lock_held_; }
@@ -88,6 +72,7 @@ class VariableInfo {
 
  private:
   int index_;
+  std::string name_;
   Var* var_;
 
   // We can't use a optional here because it confuses the compiler's
@@ -96,6 +81,20 @@ class VariableInfo {
   bool lock_held_ = false;
 };
 
+// Takes a snapshot of the values of resource variable arguments, whose indices
+// are specified in `variable_indices` argument. We snapshot tensors that back
+// resource variables since concurrent updates may modify the shape, and it is
+// important that the shapes used for compilation match the true shapes of the
+// buffers.
+//
+// We snapshot the entire set of resource variables as one atomic operation.
+// This models Read->* dependencies between resource variable operations.  See
+// jit/resource_operation_safety_analysis for details.
+Status SnapshotResourceVariables(OpKernelContext* ctx,
+                                 absl::Span variable_indices,
+                                 absl::Span variable_infos,
+                                 ResourceVarsSnapshot* result);
+
 // Acquires the mutexes for all the variables in `variables` using a
 // deadlock-safe protocol (acquire the mutexes in increasing-address order).
 //
@@ -104,6 +103,13 @@ class VariableInfo {
 Status LockVariables(absl::Span variables)
     TF_EXCLUSIVE_LOCK_FUNCTION();
 
+// Returns a vector of VariableInfo instances for the resource variable inputs
+// to the kernel with context `ctx`.  The input indices for the resource
+// variable inputs are in `variable_indices`.
+Status GetVariableInfosFromCtxInputs(OpKernelContext* ctx,
+                                     absl::Span variable_indices,
+                                     std::vector* result);
+
 // Helper class to perform the marshalling of TensorFlow inputs and outputs to
 // ShapedBuffers suitable for passing to an XLA computation.
 class XlaComputationLaunchContext {
@@ -123,9 +129,10 @@ class XlaComputationLaunchContext {
 
   // Builds a XlaCompiler::Argument vector from the arguments to an XlaLaunch
   // op.
+  // Precondition: variables in `variable_args` are locked.
   static Status BuildXlaCompilerArguments(
       const std::map& constant_args,
-      const std::map& variable_args, OpKernelContext* ctx,
+      absl::Span variable_args, OpKernelContext* ctx,
       std::vector* args);
 
   // Add all inputs within `ctx` as XLA arguments (returned by arguments()).
@@ -137,7 +144,7 @@ class XlaComputationLaunchContext {
   // (in other words, no inputs actually required by the kernel can be missing).
   void PopulateInputs(OpKernelContext* ctx,
                       const XlaCompiler::CompilationResult* compilation_result,
-                      const std::map& variables,
+                      const ResourceVarsSnapshot& variables,
                       int missing_ctx_input_prefix);
 
   // Given the XLA output in `output`, populate all outputs of `ctx`.  Also
@@ -155,7 +162,7 @@ class XlaComputationLaunchContext {
       const XlaCompiler::CompilationResult* compilation_result,
       xla::ScopedShapedBuffer output, int missing_ctx_input_prefix,
       const xla::HloInputOutputAliasConfig& input_output_alias,
-      const std::map& resource_var_snapshots);
+      const ResourceVarsSnapshot& resource_var_snapshots);
 
   // Return the argument list. Only valid after PopulateInputs() has been
   // called.

From af8f596d21f03372291384227cae75d95fa6e468 Mon Sep 17 00:00:00 2001
From: Sung Jin Hwang 
Date: Mon, 22 Jun 2020 12:13:10 -0700
Subject: [PATCH 457/867] Add _SwitchN op to the handled op lists in
 convert_variables_to_constants_v2.

Grappler transforms tf.switch_case() to use _SwitchN instead of Case op, and
because _SwitchN is currently not handled convert_variables_to_constants_v2()
fails when function contains tf.switch_case.

PiperOrigin-RevId: 317707465
Change-Id: I3524c0c50f0acd02c2347c3c4f49ece5a5d671f7
---
 .../python/framework/convert_to_constants.py  |  3 +-
 .../framework/convert_to_constants_test.py    | 34 +++++++++++++++++++
 2 files changed, 36 insertions(+), 1 deletion(-)

diff --git a/tensorflow/python/framework/convert_to_constants.py b/tensorflow/python/framework/convert_to_constants.py
index 87c74c3263d..4c3cbb06bf1 100644
--- a/tensorflow/python/framework/convert_to_constants.py
+++ b/tensorflow/python/framework/convert_to_constants.py
@@ -235,7 +235,8 @@ class _Node(_Convertible):
       return _If(node, function, enclosing_graph)
     elif node.op in ["While", "StatelessWhile"]:
       return _While(node, function, enclosing_graph)
-    elif node.op in ["Enter", "Exit", "Identity", "NextIteration", "Switch"]:
+    elif node.op in [
+        "Enter", "Exit", "Identity", "NextIteration", "Switch", "_SwitchN"]:
       return _Intermediate(node, function, enclosing_graph)
     else:
       return _Node(node, function, enclosing_graph)
diff --git a/tensorflow/python/framework/convert_to_constants_test.py b/tensorflow/python/framework/convert_to_constants_test.py
index b1e11003939..7252082d084 100644
--- a/tensorflow/python/framework/convert_to_constants_test.py
+++ b/tensorflow/python/framework/convert_to_constants_test.py
@@ -486,6 +486,40 @@ class VariablesToConstantsTest(test.TestCase):
     root, output_func = self._freezeModel(model)
     self._testConvertedFunction(root, root.f, output_func, input_data)
 
+  @test_util.run_v2_only
+  def testSwitchCase(self):
+    """Test a switch_case statement."""
+    input_data = {
+        "i": constant_op.constant(np.random.randint(0, 3, dtype=np.int32)),
+        "x": constant_op.constant(
+            np.asarray(np.random.random_sample((10, 3)), dtype=np.float32)),
+    }
+
+    w0 = variables.Variable(np.random.random_sample((3, 4)), dtype=np.float32)
+    w1 = variables.Variable(np.random.random_sample((3, 4)), dtype=np.float32)
+    w2 = variables.Variable(np.random.random_sample((4,)), dtype=np.float32)
+
+    def branch0(x):
+      return math_ops.matmul(x, w0)
+
+    def branch1(x):
+      return math_ops.matmul(x, w1)
+
+    def branch2(x):
+      x = array_ops.pad(x, [[0, 0], [0, 1]])
+      return x + w2
+
+    @def_function.function(input_signature=[
+        tensor_spec.TensorSpec(shape=[], dtype=dtypes.int32),
+        tensor_spec.TensorSpec(shape=[10, 3], dtype=dtypes.float32),
+    ])
+    def model(i, x):
+      return control_flow_ops.switch_case(i, [
+          lambda: branch0(x), lambda: branch1(x), lambda: branch2(x)])
+
+    root, output_func = self._freezeModel(model)
+    self._testConvertedFunction(root, root.f, output_func, input_data)
+
 
 class ConvertVariablesToConstantsSessionTest(test.TestCase):
 

From dfd21eaec6e3da7210bca9bfc951fcffc117a56c Mon Sep 17 00:00:00 2001
From: Akshay Modi 
Date: Mon, 22 Jun 2020 12:16:17 -0700
Subject: [PATCH 458/867] Undo some of the composite tensor changes in gradient
 tape code

Its not clear that we should always take gradients with respect to component tensors.

PiperOrigin-RevId: 317708164
Change-Id: I8a0cdddd705497e5539857afcbb60aaa38821e0c
---
 tensorflow/python/eager/backprop.py           |  9 ++-
 tensorflow/python/eager/backprop_test.py      | 69 -------------------
 tensorflow/python/ops/numpy_ops/np_arrays.py  |  4 ++
 .../python/ops/numpy_ops/np_interop_test.py   |  5 +-
 4 files changed, 11 insertions(+), 76 deletions(-)

diff --git a/tensorflow/python/eager/backprop.py b/tensorflow/python/eager/backprop.py
index 5c2deb9c0f2..8da3f71360a 100644
--- a/tensorflow/python/eager/backprop.py
+++ b/tensorflow/python/eager/backprop.py
@@ -1024,7 +1024,7 @@ class GradientTape(object):
             "derivatives.", 1)
 
     flat_targets = []
-    for t in nest.flatten(target, expand_composites=True):
+    for t in nest.flatten(target):
       if not backprop_util.IsTrainable(t):
         logging.vlog(
             logging.WARN, "The dtype of the target tensor must be "
@@ -1035,7 +1035,7 @@ class GradientTape(object):
           t = ops.convert_to_tensor(t)
       flat_targets.append(t)
 
-    flat_sources = nest.flatten(sources, expand_composites=True)
+    flat_sources = nest.flatten(sources)
     flat_sources_raw = flat_sources
     flat_sources = [_handle_or_self(x) for x in flat_sources]
     for t in flat_sources_raw:
@@ -1051,8 +1051,7 @@ class GradientTape(object):
 
     if output_gradients is not None:
       output_gradients = [None if x is None else ops.convert_to_tensor(x)
-                          for x in nest.flatten(
-                              output_gradients, expand_composites=True)]
+                          for x in nest.flatten(output_gradients)]
 
     flat_grad = imperative_grad.imperative_grad(
         self._tape,
@@ -1067,7 +1066,7 @@ class GradientTape(object):
       self._watched_variables = self._tape.watched_variables()
       self._tape = None
 
-    grad = nest.pack_sequence_as(sources, flat_grad, expand_composites=True)
+    grad = nest.pack_sequence_as(sources, flat_grad)
     return grad
 
   def jacobian(self,
diff --git a/tensorflow/python/eager/backprop_test.py b/tensorflow/python/eager/backprop_test.py
index abdac526ce4..a0f98fc0a44 100644
--- a/tensorflow/python/eager/backprop_test.py
+++ b/tensorflow/python/eager/backprop_test.py
@@ -28,7 +28,6 @@ from tensorflow.python.eager import def_function
 from tensorflow.python.eager import function
 from tensorflow.python.eager import tape as tape_lib
 from tensorflow.python.eager import test
-from tensorflow.python.framework import composite_tensor
 from tensorflow.python.framework import constant_op
 from tensorflow.python.framework import dtypes
 from tensorflow.python.framework import errors_impl
@@ -37,7 +36,6 @@ from tensorflow.python.framework import sparse_tensor
 from tensorflow.python.framework import tensor_shape
 from tensorflow.python.framework import tensor_util
 from tensorflow.python.framework import test_util
-from tensorflow.python.framework import type_spec
 from tensorflow.python.framework.memory_checker import MemoryChecker
 from tensorflow.python.layers.pooling import max_pooling3d
 from tensorflow.python.ops import array_ops
@@ -54,44 +52,6 @@ from tensorflow.python.ops import resource_variable_ops
 from tensorflow.python.ops import sparse_ops
 from tensorflow.python.ops import variables
 from tensorflow.python.training import training
-from tensorflow.python.util import nest
-
-
-# TODO(nareshmodi): This is copied from composite_tensor_test.py. Extract it out
-# to a common library to avoid duplication.
-class CTSpec(type_spec.TypeSpec):
-  """A generic CompositeTensor TypeSpec, used for constructing tests."""
-
-  def __init__(self, component_specs):
-    self.component_specs = component_specs
-
-  value_type = property(lambda self: CT)
-  _component_specs = property(lambda self: self.component_specs)
-
-  def _serialize(self):
-    return (self.component_specs,)
-
-  def _to_components(self, value):
-    return value.components
-
-  def _from_components(self, tensor_list):
-    return CT(tensor_list)
-
-
-class CT(composite_tensor.CompositeTensor):
-  """A generic CompositeTensor, used for constructing tests."""
-  _type_spec_class = CTSpec
-
-  def __init__(self, components):
-    if isinstance(components, list):
-      components = tuple(components)
-    self.components = components
-
-  @property
-  def _type_spec(self):
-    component_specs = nest.map_structure(type_spec.type_spec_from_value,
-                                         self.components)
-    return self._type_spec_class(component_specs)
 
 
 class BackpropTest(test.TestCase, parameterized.TestCase):
@@ -1621,35 +1581,6 @@ class BackpropTest(test.TestCase, parameterized.TestCase):
     memory_checker.report()
     memory_checker.assert_no_leak_if_all_possibly_except_one()
 
-  def testCompositeTensorAsSource(self):
-    t = CT([constant_op.constant(3.), constant_op.constant(2.)])
-    with backprop.GradientTape() as gt:
-      gt.watch(t)
-      y = CT([t.components[0] * 2, t.components[1] * 3])
-
-    grad = gt.gradient(y, t)
-    expected_grad = CT([constant_op.constant(2.), constant_op.constant(3.)])
-
-    flat_grads = nest.flatten(grad, expand_composites=True)
-    flat_expected_grads = nest.flatten(expected_grad, expand_composites=True)
-
-    self.assertAllClose(flat_grads, flat_expected_grads)
-
-  def testCompositeTensorAsOutputGradients(self):
-    t = CT([constant_op.constant(3.), constant_op.constant(2.)])
-    with backprop.GradientTape() as gt:
-      gt.watch(t)
-      y = CT([t.components[0] * 2, t.components[1] * 3])
-
-    output_gradients = CT([constant_op.constant(5.), constant_op.constant(10.)])
-    grad = gt.gradient(y, t, output_gradients=output_gradients)
-    expected_grad = CT([constant_op.constant(10.), constant_op.constant(30.)])
-
-    flat_grads = nest.flatten(grad, expand_composites=True)
-    flat_expected_grads = nest.flatten(expected_grad, expand_composites=True)
-
-    self.assertAllClose(flat_grads, flat_expected_grads)
-
 
 class JacobianTest(test.TestCase):
 
diff --git a/tensorflow/python/ops/numpy_ops/np_arrays.py b/tensorflow/python/ops/numpy_ops/np_arrays.py
index d5ad87a887e..46252dd169c 100644
--- a/tensorflow/python/ops/numpy_ops/np_arrays.py
+++ b/tensorflow/python/ops/numpy_ops/np_arrays.py
@@ -306,6 +306,10 @@ class ndarray(composite_tensor.CompositeTensor):  # pylint: disable=invalid-name
   def __repr__(self):
     return 'ndarray<{}>'.format(self.data.__repr__())
 
+  @property
+  def _id(self):
+    return self.data._id  # pylint: disable=protected-access
+
 
 def tensor_to_ndarray(tensor):
   return ndarray.from_tensor(tensor)
diff --git a/tensorflow/python/ops/numpy_ops/np_interop_test.py b/tensorflow/python/ops/numpy_ops/np_interop_test.py
index f6882e968b5..5c7560b12d9 100644
--- a/tensorflow/python/ops/numpy_ops/np_interop_test.py
+++ b/tensorflow/python/ops/numpy_ops/np_interop_test.py
@@ -64,8 +64,9 @@ class InteropTest(test.TestCase):
 
     dx, dy = t.gradient([xx, yy], [x, y])
 
-    self.assertIsInstance(dx, np_arrays.ndarray)
-    self.assertIsInstance(dy, np_arrays.ndarray)
+    # # TODO(nareshmodi): Figure out a way to rewrap ndarray as tensors.
+    # self.assertIsInstance(dx, np_arrays.ndarray)
+    # self.assertIsInstance(dy, np_arrays.ndarray)
     self.assertAllClose(dx, 2.0)
     self.assertAllClose(dy, 3.0)
 

From 780c0a29fe66b93a8fd67fd8cc1c63ad6ea60a17 Mon Sep 17 00:00:00 2001
From: Haitang Hu 
Date: Mon, 22 Jun 2020 12:21:45 -0700
Subject: [PATCH 459/867] Update device_id to be int32 rather than int64.

PiperOrigin-RevId: 317709385
Change-Id: I577dd469d223cc05c50dbdb6a8bd908e2e757344
---
 .../c/eager/parallel_device/parallel_device_lib.cc     | 10 +++++-----
 .../c/eager/parallel_device/parallel_device_testlib.cc |  4 ++--
 2 files changed, 7 insertions(+), 7 deletions(-)

diff --git a/tensorflow/c/eager/parallel_device/parallel_device_lib.cc b/tensorflow/c/eager/parallel_device/parallel_device_lib.cc
index d0149b29c08..768f686bd88 100644
--- a/tensorflow/c/eager/parallel_device/parallel_device_lib.cc
+++ b/tensorflow/c/eager/parallel_device/parallel_device_lib.cc
@@ -262,14 +262,14 @@ std::unique_ptr ParallelDevice::DeviceIDs(
   components.reserve(underlying_devices_.size());
   for (int device_index = 0; device_index < underlying_devices_.size();
        ++device_index) {
-    int64_t* device_id = new int64_t;
+    int32_t* device_id = new int32_t;
     *device_id = device_index;
     std::unique_ptr tensor(
         TF_NewTensor(
-            TF_INT64, /*dims=*/nullptr, /*num_dims=*/0, device_id,
-            sizeof(int64_t),
+            TF_INT32, /*dims=*/nullptr, /*num_dims=*/0, device_id,
+            sizeof(int32_t),
             [](void* data, size_t, void* arg) {
-              delete reinterpret_cast(data);
+              delete reinterpret_cast(data);
             },
             nullptr),
         TF_DeleteTensor);
@@ -283,7 +283,7 @@ std::unique_ptr ParallelDevice::DeviceIDs(
     if (TF_GetCode(status) != TF_OK) return nullptr;
     TFE_OpSetAttrTensor(const_op.get(), "value", tensor.get(), status);
     if (TF_GetCode(status) != TF_OK) return nullptr;
-    TFE_OpSetAttrType(const_op.get(), "dtype", TF_INT64);
+    TFE_OpSetAttrType(const_op.get(), "dtype", TF_INT32);
     TFE_TensorHandle* device_handle;
     int num_outputs = 1;
     TFE_Execute(const_op.get(), &device_handle, &num_outputs, status);
diff --git a/tensorflow/c/eager/parallel_device/parallel_device_testlib.cc b/tensorflow/c/eager/parallel_device/parallel_device_testlib.cc
index fba47865c36..828dcbae093 100644
--- a/tensorflow/c/eager/parallel_device/parallel_device_testlib.cc
+++ b/tensorflow/c/eager/parallel_device/parallel_device_testlib.cc
@@ -296,8 +296,8 @@ void BasicTestsForTwoDevices(TFE_Context* context, const char* first_device,
     TFE_DeleteTensorHandle(result_handle);
     ASSERT_TRUE(TF_GetCode(status.get()) == TF_OK) << TF_Message(status.get());
 
-    ExpectScalarEq(components[0].get(), 0);
-    ExpectScalarEq(components[1].get(), 1);
+    ExpectScalarEq(components[0].get(), 0);
+    ExpectScalarEq(components[1].get(), 1);
     std::string first_device =
         TFE_TensorHandleBackingDeviceName(components[0].get(), status.get());
     ASSERT_EQ(underlying_devices[0], first_device);

From aac1dd5788000f05f19f247e3bacacd81810d72b Mon Sep 17 00:00:00 2001
From: Tomer Kaftan 
Date: Mon, 22 Jun 2020 12:24:15 -0700
Subject: [PATCH 460/867] Updates Keras layer `__call__` to always set
 `training`, with the following priority order:     # Training mode for
 `Layer.__call__` is set via (in order of priority):     # (1) The `training`
 argument passed to this `Layer.__call__`, if it is not None     # (2) The
 training mode of an outer `Layer.__call__`.     # (3) The default mode set by
 `tf.keras.backed.set_learning_phase` (if set).     # (4) Any non-None default
 value for `training` specified in the `call`     #  signature     # (5) False
 (treating the layer as if it's in inference)

Previously (4) and (5) were missing, leading to crashes for layers that do not provide a default argument for `training`.

Note that (4) is fragile to reflection issues, and may get confused by decorators.

PiperOrigin-RevId: 317709904
Change-Id: I58039a4d9e5106bcb27f4cfbf65e6762f1b40807
---
 tensorflow/python/keras/engine/base_layer.py  | 58 +++++++++---
 .../python/keras/engine/base_layer_test.py    | 91 +++++++++++++++++++
 .../python/keras/engine/functional_test.py    | 68 ++++++++++----
 .../keras/layers/rnn_cell_wrapper_v2.py       |  5 +
 .../keras/tests/custom_training_loop_test.py  |  2 +-
 5 files changed, 192 insertions(+), 32 deletions(-)

diff --git a/tensorflow/python/keras/engine/base_layer.py b/tensorflow/python/keras/engine/base_layer.py
index fbec5382a08..97eb0447a69 100644
--- a/tensorflow/python/keras/engine/base_layer.py
+++ b/tensorflow/python/keras/engine/base_layer.py
@@ -943,10 +943,14 @@ class Layer(module.Module, version_utils.LayerVersionSelector):
       kwargs['mask'] = input_masks
 
     # Training mode for `Layer.call` is set via (in order of priority):
-    # (1) The `training` argument passed to this `Layer.call`.
+    # (1) The `training` argument passed to this `Layer.call`, if it is not None
     # (2) The training mode of an outer `Layer.call`.
-    # (3) The default mode set by `tf.keras.backed.set_learning_phase` (if set).
-    training_mode = self._set_training_mode(args, kwargs, call_context)
+    # (3) The default mode set by `tf.keras.backend.set_learning_phase` (if set)
+    # (4) Any non-None default value for `training` specified in the call
+    #  signature
+    # (5) False (treating the layer as if it's in inference)
+    args, kwargs, training_mode = self._set_training_mode(
+        args, kwargs, call_context)
 
     # Losses are cleared for all sublayers on the outermost `Layer.call`.
     # Losses are not cleared on inner `Layer.call`s, because sublayers can be
@@ -1020,7 +1024,7 @@ class Layer(module.Module, version_utils.LayerVersionSelector):
     # propagate `training` value from this layer's calling layer.
     training_value = None
     training_arg_passed_by_framework = False
-    # Priority 1: `training` was explicitly passed.
+    # Priority 1: `training` was explicitly passed a non-None value.
     if self._call_arg_was_passed('training', args, kwargs):
       training_value = self._get_call_arg_value('training', args, kwargs)
       if not self._expects_training_arg:
@@ -1030,17 +1034,23 @@ class Layer(module.Module, version_utils.LayerVersionSelector):
       # Priority 2: `training` was passed to a parent layer.
       if call_context.training is not None:
         training_value = call_context.training
-      # Priority 3a: `learning_phase()` has been set.
+      # Priority 3: `learning_phase()` has been set.
       elif backend.global_learning_phase_is_set():
         training_value = backend.learning_phase()
-
-      if self._expects_training_arg and training_value is not None:
         # Force the training_value to be bool type which matches to the contract
         # for layer/model call args.
         if tensor_util.is_tensor(training_value):
           training_value = math_ops.cast(training_value, dtypes.bool)
         else:
           training_value = bool(training_value)
+      # Priority 4: trace layer with the default training argument specified
+      # in the `call` signature (or in inference mode if the `call` signature
+      # specifies no non-None default).
+      else:
+        training_value = self._default_training_arg
+      # In cases (2), (3), (4) the training argument is passed automatically
+      # by the framework, and will not be hard-coded into the model.
+      if self._expects_training_arg:
         args, kwargs = self._set_call_arg_value('training', training_value,
                                                 args, kwargs)
         training_arg_passed_by_framework = True
@@ -1150,6 +1160,8 @@ class Layer(module.Module, version_utils.LayerVersionSelector):
       # (1) `training` was passed to this `Layer.call`.
       if self._call_arg_was_passed('training', args, kwargs):
         training_mode = self._get_call_arg_value('training', args, kwargs)
+      # If no `training` arg was passed, or `None` was explicitly passed,
+      # the framework will make a decision about the training mode is.
       if training_mode is None:
         call_ctx_training = call_context.training
         # (2) `training` mode is inferred from an outer `Layer.call`.
@@ -1165,10 +1177,15 @@ class Layer(module.Module, version_utils.LayerVersionSelector):
             training_mode = math_ops.cast(training_mode, dtypes.bool)
           else:
             training_mode = bool(training_mode)
+        # (4) We default to using `call`'s default value for `training`,
+        # or treating the layer as if it is in inference if no non-None default
+        # is specified in the `call` signature.
+        else:
+          training_mode = self._default_training_arg
 
-        # For case (2) or (3), `training` arg is passed by framework.
-        if training_mode is not None:
-          kwargs['training'] = training_mode
+        # For case (2), (3), (4) `training` arg is passed by framework.
+        args, kwargs = self._set_call_arg_value('training', training_mode, args,
+                                                kwargs)
     else:
       if 'training' in kwargs:
         # `training` was passed to this `Layer` but is not needed for
@@ -1178,7 +1195,7 @@ class Layer(module.Module, version_utils.LayerVersionSelector):
         # Grab the current `training` mode from any outer `Layer.call`.
         training_mode = call_context.training
 
-    return training_mode
+    return args, kwargs, training_mode
 
   def _autographed_call(self):
     # Wrapping `call` function in autograph to allow for dynamic control
@@ -2529,7 +2546,7 @@ class Layer(module.Module, version_utils.LayerVersionSelector):
       if len(args) > arg_pos:
         args = list(args)
         args[arg_pos] = new_value
-        return args, kwargs
+        return tuple(args), kwargs
     if new_value is None and pop_kwarg_if_none:
       kwargs.pop(arg_name, None)
     else:
@@ -2873,6 +2890,10 @@ class Layer(module.Module, version_utils.LayerVersionSelector):
     call_fn_args = self._call_fn_args
     self._expects_training_arg = ('training' in call_fn_args or
                                   self._call_accepts_kwargs)
+    # The default training arg will be any (non-None) default specified in the
+    # method signature, or `False` if no non-None default is specified.
+    self._default_training_arg = self._call_fn_arg_defaults.get(
+        'training') or False
     self._expects_mask_arg = ('mask' in call_fn_args or
                               self._call_accepts_kwargs)
 
@@ -2892,6 +2913,19 @@ class Layer(module.Module, version_utils.LayerVersionSelector):
       return all_args[1:]
     return all_args
 
+  @property
+  @tracking.cached_per_instance
+  def _call_fn_arg_defaults(self):
+    call_fn_args = self._call_fn_args
+    call_fn_defaults = self._call_full_argspec.defaults or []
+    defaults = dict()
+
+    # The call arg defaults are an n-tuple of the last n elements of the args
+    # list. (n = # of elements that have a default argument)
+    for i in range(-1 * len(call_fn_defaults), 0):
+      defaults[call_fn_args[i]] = call_fn_defaults[i]
+    return defaults
+
   @property
   @tracking.cached_per_instance
   def _call_fn_arg_positions(self):
diff --git a/tensorflow/python/keras/engine/base_layer_test.py b/tensorflow/python/keras/engine/base_layer_test.py
index b861d7e4b5b..58a0799329a 100644
--- a/tensorflow/python/keras/engine/base_layer_test.py
+++ b/tensorflow/python/keras/engine/base_layer_test.py
@@ -629,6 +629,96 @@ class BaseLayerTest(keras_parameterized.TestCase):
     self.assertTrue(layer.built)
     self.assertEqual([None, 3], layer._build_input_shape.as_list())
 
+  @combinations.generate(combinations.combine(mode=['eager']))
+  def custom_layer_training_arg(self):
+    class CustomLayerNoTrainingArg(base_layer.Layer):
+
+      def __init__(self, nested_layer=None):
+        self._nested_layer = nested_layer or array_ops.identity
+
+      def call(self, inputs):
+        return self._nested_layer(inputs)
+
+    class CustomLayerDefaultTrainingMissing(base_layer.Layer):
+
+      def __init__(self, nested_layer=None):
+        self._nested_layer = nested_layer or array_ops.identity
+
+      def call(self, inputs, training):
+        if training:
+          return self._nested_layer(inputs)
+        else:
+          return self._nested_layer(inputs) * 0.5
+
+    class CustomLayerDefaultTrainingFalse(base_layer.Layer):
+
+      def __init__(self, nested_layer=None):
+        self._nested_layer = nested_layer or array_ops.identity
+
+      def call(self, inputs, training=False):
+        if training:
+          return self._nested_layer(inputs)
+        else:
+          return self._nested_layer(inputs) * 0.5
+
+    class CustomLayerDefaultTrainingTrue(base_layer.Layer):
+
+      def __init__(self, nested_layer=None):
+        self._nested_layer = nested_layer or array_ops.identity
+
+      def call(self, inputs, training=True):
+        if training:
+          return self._nested_layer(inputs)
+        else:
+          return self._nested_layer(inputs) * 0.5
+
+    x = array_ops.ones(shape=(1, 1))
+
+    # If the layer signature doesn't specify a default training arg,
+    # run it in inference mode when to training arg is passed
+    # to __call__
+    layer = CustomLayerDefaultTrainingMissing()
+    self.assertAllEqual(layer(x), x * 0.5)
+    self.assertAllEqual(layer(x, training=False), x * 0.5)
+    self.assertAllEqual(layer(x, training=True), x)
+
+    # If the layer signature specifies `False` as the default training arg,
+    # run it in inference mode when no training arg is passed
+    # to __call__
+    layer = CustomLayerDefaultTrainingFalse()
+    self.assertAllEqual(layer(x), x * 0.5)
+    self.assertAllEqual(layer(x, training=False), x * 0.5)
+    self.assertAllEqual(layer(x, training=True), x)
+
+    # If the layer signature specifies `True` as the default training arg,
+    # explicitly run it in training mode when no training arg is passed
+    # to __call__
+    layer = CustomLayerDefaultTrainingTrue()
+    self.assertAllEqual(layer(x), x)
+    self.assertAllEqual(layer(x, training=False), x * 0.5)
+    self.assertAllEqual(layer(x, training=True), x)
+
+    # Outer layers/models should set the training context implicitly for all
+    # nested layers, respecting whatever mode the outer layer was run with.
+    layer = CustomLayerDefaultTrainingTrue(CustomLayerDefaultTrainingFalse())
+    self.assertAllEqual(layer(x), x)
+    self.assertAllEqual(layer(x, training=False), x * 0.25)
+    self.assertAllEqual(layer(x, training=True), x)
+
+    layer = CustomLayerDefaultTrainingFalse(CustomLayerDefaultTrainingTrue())
+    self.assertAllEqual(layer(x), x * 0.25)
+    self.assertAllEqual(layer(x, training=False), x * 0.25)
+    self.assertAllEqual(layer(x, training=True), x)
+
+    # If the outer layer `call` doesn't take a training argument at all,
+    # it'll set the nested scope as inference when no training arg is passed in.
+    # If a training arg is passed in it won't use it directly in `call`, but
+    # it will set the nested training mode.
+    layer = CustomLayerNoTrainingArg(CustomLayerDefaultTrainingTrue())
+    self.assertAllEqual(layer(x), x * 0.5)
+    self.assertAllEqual(layer(x, training=False), x * 0.5)
+    self.assertAllEqual(layer(x, training=True), x)
+
   def test_activity_regularizer_string(self):
 
     class MyLayer(base_layer.Layer):
@@ -1387,6 +1477,7 @@ class DTypeTest(keras_parameterized.TestCase):
     class IdentityLayerWithArgs(base_layer.Layer):
 
       def call(self, inputs, *args, **kwargs):
+        kwargs.pop('training', None)
         return nest.flatten([inputs, args, kwargs])
 
     layer = IdentityLayerWithArgs(dtype='float64')
diff --git a/tensorflow/python/keras/engine/functional_test.py b/tensorflow/python/keras/engine/functional_test.py
index a7e314d4a49..3c14411deb9 100644
--- a/tensorflow/python/keras/engine/functional_test.py
+++ b/tensorflow/python/keras/engine/functional_test.py
@@ -2036,43 +2036,73 @@ class CacheCorrectnessTest(keras_parameterized.TestCase):
 
   def test_training_passed_during_construction(self):
 
+    def _call(inputs, training):
+      if training is None:
+        return inputs * -1.0
+      elif training:
+        return inputs
+      else:
+        return inputs * 0.0
+
     class MyLayer(base_layer.Layer):
 
-      def call(self, x, training=None):
-        if training is None:
-          return x * -1.0
-        elif training:
-          return x
-        else:
-          return x * 0.0
+      def call(self, inputs, training=True):
+        return _call(inputs, training)
 
     my_layer = MyLayer()
     x = np.ones((1, 10))
 
+    # Hard-coded `true` value passed during construction is respected.
     inputs = input_layer_lib.Input(10)
     outputs = my_layer(inputs, training=True)
     network = functional.Functional(inputs, outputs)
+    self.assertAllEqual(network(x, training=True), _call(x, True))
+    self.assertAllEqual(network(x, training=False), _call(x, True))
+    self.assertAllEqual(network(x), _call(x, True))
 
-    # Hard-coded value passed during construction is respected.
-    self.assertAllEqual(network(x, training=False), x)
-
+    # Hard-coded `false` value passed during construction is respected.
     inputs = input_layer_lib.Input(10)
     outputs = my_layer(inputs, training=False)
     network = functional.Functional(inputs, outputs)
+    self.assertAllEqual(network(x, training=True), _call(x, False))
+    self.assertAllEqual(network(x, training=False), _call(x, False))
+    self.assertAllEqual(network(x), _call(x, False))
 
-    network(x, training=True)
-    # Hard-coded value passed during construction is respected.
-    self.assertAllEqual(network(x, training=True), x * 0.0)
+    if context.executing_eagerly():
+      # In v2, construction still works when no `training` is specified
+      # When no value passed during construction, it uses the runtime value.
+      inputs = input_layer_lib.Input(10)
+      outputs = my_layer(inputs)
+      network = functional.Functional(inputs, outputs)
+      self.assertAllEqual(network(x, training=True), _call(x, True))
+      self.assertAllEqual(network(x, training=False), _call(x, False))
+      self.assertAllEqual(network(x), _call(x, False))
 
+    # `None` value passed positionally during construction is ignored at runtime
+    inputs = input_layer_lib.Input(10)
+    outputs = my_layer(inputs, None)
+    network = functional.Functional(inputs, outputs)
+    self.assertAllEqual(network(x, training=True), _call(x, True))
+    self.assertAllEqual(network(x, training=False), _call(x, False))
+    if context.executing_eagerly():
+      self.assertAllEqual(network(x), _call(x, False))
+    else:
+      # in v1 training would have defaulted to using the `None` inside the layer
+      # if training is not passed at runtime
+      self.assertAllEqual(network(x), _call(x, None))
+
+    # `None` value passed as kwarg during construction is ignored at runtime.
     inputs = input_layer_lib.Input(10)
     outputs = my_layer(inputs, training=None)
     network = functional.Functional(inputs, outputs)
-
-    # `None` value passed during construction is overridden.
-    self.assertAllEqual(network(x, training=True), x)
-    # `None` value passed during construction is overridden.
-    self.assertAllEqual(network(x, training=False), x * 0.0)
-
+    self.assertAllEqual(network(x, training=True), _call(x, True))
+    self.assertAllEqual(network(x, training=False), _call(x, False))
+    if context.executing_eagerly():
+      self.assertAllEqual(network(x), _call(x, False))
+    else:
+      # in v1 training would have defaulted to using the `None` inside the layer
+      # if training is not passed at runtime
+      self.assertAllEqual(network(x), _call(x, None))
 
 if __name__ == '__main__':
   test.main()
diff --git a/tensorflow/python/keras/layers/rnn_cell_wrapper_v2.py b/tensorflow/python/keras/layers/rnn_cell_wrapper_v2.py
index 4356244b292..d387a375aa2 100644
--- a/tensorflow/python/keras/layers/rnn_cell_wrapper_v2.py
+++ b/tensorflow/python/keras/layers/rnn_cell_wrapper_v2.py
@@ -27,6 +27,7 @@ from __future__ import print_function
 
 from tensorflow.python.keras.layers import recurrent
 from tensorflow.python.ops import rnn_cell_wrapper_impl
+from tensorflow.python.util import tf_inspect
 from tensorflow.python.util.tf_export import tf_export
 
 
@@ -41,6 +42,10 @@ class _RNNCellWrapperV2(recurrent.AbstractRNNCell):
   def __init__(self, cell, *args, **kwargs):
     super(_RNNCellWrapperV2, self).__init__(*args, **kwargs)
     self.cell = cell
+    cell_call_spec = tf_inspect.getfullargspec(cell.call)
+    self._expects_training_arg = ("training" in cell_call_spec.args) or (
+        cell_call_spec.varkw is not None
+    )
 
   def call(self, inputs, state, **kwargs):
     """Runs the RNN cell step computation.
diff --git a/tensorflow/python/keras/tests/custom_training_loop_test.py b/tensorflow/python/keras/tests/custom_training_loop_test.py
index 5b3310b2b40..6291933ac99 100644
--- a/tensorflow/python/keras/tests/custom_training_loop_test.py
+++ b/tensorflow/python/keras/tests/custom_training_loop_test.py
@@ -186,7 +186,7 @@ class CustomTrainingLoopTest(keras_parameterized.TestCase):
 
     def train_step(x):
       no_learning_phase_out = model(x)
-      self.assertIsNone(model.layer.training)
+      self.assertFalse(model.layer.training)
       with keras.backend.learning_phase_scope(0):
         inf_learning_phase_out = model(x)
       self.assertEqual(model.layer.training, 0)

From 38d95ad2d87a94b9260ff9ef242a9d5f12640147 Mon Sep 17 00:00:00 2001
From: Yujing Zhang 
Date: Mon, 22 Jun 2020 12:29:26 -0700
Subject: [PATCH 461/867] [Cleanup] Remove allowed_devices of ResourceHandle
 since it's no longer used.

PiperOrigin-RevId: 317710941
Change-Id: Ib1920c5ee25d405290f852b725d693ee5ea09766
---
 .../base_api/api_def_VarHandleOp.pbtxt        |  4 +-
 .../core/common_runtime/eager/execute.cc      | 32 +++---------
 .../common_runtime/eager/tensor_handle.cc     | 51 +++++--------------
 .../core/common_runtime/eager/tensor_handle.h | 20 +++-----
 .../eager/tensor_handle_test.cc               | 17 +++----
 .../distributed_runtime/eager/remote_mgr.cc   | 12 ++---
 tensorflow/core/framework/resource_handle.cc  |  6 ---
 tensorflow/core/framework/resource_handle.h   | 13 +----
 .../core/framework/resource_handle.proto      |  4 +-
 tensorflow/core/framework/resource_mgr.cc     | 39 ++------------
 tensorflow/core/framework/resource_mgr.h      | 28 +++++-----
 .../core/framework/resource_mgr_test.cc       | 47 -----------------
 .../core/kernels/resource_variable_ops.cc     |  8 +--
 .../core/kernels/resource_variable_ops.h      |  4 --
 .../resource_variable_ops_test.py             | 37 --------------
 15 files changed, 63 insertions(+), 259 deletions(-)

diff --git a/tensorflow/core/api_def/base_api/api_def_VarHandleOp.pbtxt b/tensorflow/core/api_def/base_api/api_def_VarHandleOp.pbtxt
index 39606a07184..29ffcdaad6b 100644
--- a/tensorflow/core/api_def/base_api/api_def_VarHandleOp.pbtxt
+++ b/tensorflow/core/api_def/base_api/api_def_VarHandleOp.pbtxt
@@ -28,8 +28,8 @@ END
   attr {
     name: "allowed_devices"
     description: <GetResourceHandleInfo(&resource_handle_info));
-        std::vector* resource_dtypes_and_shapes =
-            &resource_handle_info.dtypes_and_shapes;
-        if (!resource_dtypes_and_shapes->empty()) {
+        std::vector resource_dtypes_and_shapes;
+        TF_RETURN_IF_ERROR(input->GetResourceHandleDtypesAndShapes(
+            &resource_dtypes_and_shapes));
+        if (!resource_dtypes_and_shapes.empty()) {
           const DtypeAndPartialTensorShape& dtype_and_shape =
-              resource_dtypes_and_shapes->at(0);
+              resource_dtypes_and_shapes.at(0);
           input_resource_variable_dtypes_and_shapes[i] = dtype_and_shape;
 
           // Add _Arg index, dtype and shape to "cache_key".
@@ -695,13 +694,8 @@ Status StoreResourceDtypesAndShapes(const eager::Operation& remote_op,
     TF_RETURN_IF_ERROR(attr_slice.Find("dtype", &dtype));
     const AttrValue* shape;
     TF_RETURN_IF_ERROR(attr_slice.Find("shape", &shape));
-    TensorHandle::ResourceHandleInfo resource_handle_info = {
-        {DtypeAndPartialTensorShape{dtype->type(), shape->shape()}}, {}};
-    // "allowed_devices" is set only when the output represents a
-    // per-replica/partitioned resource variable.
-    TryGetNodeAttr(attr_slice, "allowed_devices",
-                   &resource_handle_info.allowed_devices);
-    retvals[0]->SetResourceHandleInfo(std::move(resource_handle_info));
+    retvals[0]->SetResourceHandleDtypeAndShape(
+        {DtypeAndPartialTensorShape{dtype->type(), shape->shape()}});
   }
   return Status::OK();
 }
@@ -985,18 +979,6 @@ Status MaybeUpdateOpDevice(EagerOperation* op) {
       // is a resource we must pin it to prevent different device selection.
       // TODO(iga): null device can mean "unspecified" or "CPU". Clean this up.
       if (resource_device != op_device || op->Device() == kVariantDeviceNull) {
-        std::vector allowed_devices;
-        TF_RETURN_IF_ERROR(
-            tensor_handle->GetResourceAllowedDevices(&allowed_devices));
-        if (!allowed_devices.empty()) {
-          // TODO(b/145922293): Support allowed_devices specified in wildcard
-          // patterns.
-          if (std::find(allowed_devices.begin(), allowed_devices.end(),
-                        op->DeviceName()) != allowed_devices.end()) {
-            TF_RETURN_IF_ERROR(ctx.FindDeviceFromName(op->DeviceName().c_str(),
-                                                      &resource_device));
-          }
-        }
         DVLOG(1) << (resource_device != op_device ? "Changing " : "Setting ")
                  << "device of operation " << op->Name() << " to "
                  << resource_device->name() << " because input #" << i
diff --git a/tensorflow/core/common_runtime/eager/tensor_handle.cc b/tensorflow/core/common_runtime/eager/tensor_handle.cc
index ef3e7a3cd28..0cd55959924 100644
--- a/tensorflow/core/common_runtime/eager/tensor_handle.cc
+++ b/tensorflow/core/common_runtime/eager/tensor_handle.cc
@@ -145,13 +145,13 @@ Status TensorHandle::PackedTensorHandleData::ExtractPackedHandle(
   return Status::OK();
 }
 
-void TensorHandle::SetResourceHandleInfo(
-    ResourceHandleInfo&& resource_handle_info) {
-  resource_handle_info_ = std::move(resource_handle_info);
+void TensorHandle::SetResourceHandleDtypeAndShape(
+    std::vector dtypes_and_shapes) {
+  handle_dtypes_and_shapes_ = std::move(dtypes_and_shapes);
 }
 
-Status TensorHandle::GetResourceHandleInfoImpl(
-    std::function set_resource_info) {
+Status TensorHandle::GetResourceHandleDtypesAndShapes(
+    std::vector* result) {
   if (dtype != DT_RESOURCE) {
     return errors::InvalidArgument(
         "TensorHandle::GetResourceDtypeAndShape should be called on tensor "
@@ -160,7 +160,7 @@ Status TensorHandle::GetResourceHandleInfoImpl(
   }
 
   if (Type() != LOCAL) {
-    set_resource_info();
+    *result = handle_dtypes_and_shapes_;
     return Status::OK();
   }
 
@@ -170,32 +170,10 @@ Status TensorHandle::GetResourceHandleInfoImpl(
   auto& data = absl::get(data_);
   TF_RETURN_IF_ERROR(data.WaitReady("TensorHandle::GetResourceHandleInfo"));
 
-  set_resource_info();
+  *result = handle_dtypes_and_shapes_;
   return Status::OK();
 }
 
-Status TensorHandle::GetResourceHandleInfo(ResourceHandleInfo* result) {
-  auto get_resource_info = [result, this]() {
-    *result = resource_handle_info_;
-  };
-  return GetResourceHandleInfoImpl(get_resource_info);
-}
-
-Status TensorHandle::GetResourceHandleDtypesAndShapes(
-    std::vector* result) {
-  auto get_resource_info = [result, this]() {
-    *result = resource_handle_info_.dtypes_and_shapes;
-  };
-  return GetResourceHandleInfoImpl(get_resource_info);
-}
-
-Status TensorHandle::GetResourceAllowedDevices(std::vector* result) {
-  auto get_resource_info = [result, this]() {
-    *result = resource_handle_info_.allowed_devices;
-  };
-  return GetResourceHandleInfoImpl(get_resource_info);
-}
-
 int TensorHandle::NumPackedHandles() const {
   if (Type() != PACKED) {
     return 0;
@@ -270,9 +248,8 @@ TensorHandle::TensorHandle(tensorflow::Tensor&& t, Device* d, Device* op_device,
       resource_remote_device_incarnation_(
           GetRemoteDeviceIncarnation(resource_device_)),
       ctx_(ctx),
-      resource_handle_info_(
-          {t.flat()(0).dtypes_and_shapes(),
-           t.flat()(0).allowed_devices()}),
+      handle_dtypes_and_shapes_(
+          t.flat()(0).dtypes_and_shapes()),
       data_(absl::in_place_type, std::move(t)) {
   DVLOG(3) << "Creating Local TensorHandle: " << this
            << " device: " << VariantDeviceDebugString(device_)
@@ -327,10 +304,10 @@ Status TensorHandle::CreatePackedHandle(std::vector&& handles,
     return errors::InvalidArgument("Handles should not be empty.");
   }
 
-  ResourceHandleInfo resource_handle_info;
+  std::vector dtypes_and_shapes;
   if (dtype == DT_RESOURCE) {
     TF_RETURN_IF_ERROR(
-        handles.at(0)->GetResourceHandleInfo(&resource_handle_info));
+        handles.at(0)->GetResourceHandleDtypesAndShapes(&dtypes_and_shapes));
   }
   std::vector devices;
   for (auto* handle : handles) {
@@ -348,7 +325,8 @@ Status TensorHandle::CreatePackedHandle(std::vector&& handles,
                                                       &composite_device));
   *packed_handle =
       new TensorHandle(std::move(handles), composite_device, dtype, shape, ctx);
-  (*packed_handle)->SetResourceHandleInfo(std::move(resource_handle_info));
+  (*packed_handle)
+      ->SetResourceHandleDtypeAndShape(std::move(dtypes_and_shapes));
   return Status::OK();
 }
 
@@ -898,8 +876,7 @@ Status TensorHandle::SetTensor(tensorflow::Tensor&& t, const Device* d) {
 
     if (t.dtype() == DT_RESOURCE && t.NumElements() > 0) {
       auto& resource_handle = t.flat()(0);
-      resource_handle_info_ = {resource_handle.dtypes_and_shapes(),
-                               resource_handle.allowed_devices()};
+      handle_dtypes_and_shapes_ = resource_handle.dtypes_and_shapes();
     }
     auto& data = absl::get(data_);
     return data.SetTensor(std::move(t));
diff --git a/tensorflow/core/common_runtime/eager/tensor_handle.h b/tensorflow/core/common_runtime/eager/tensor_handle.h
index 6ac48bdac26..8ef482cd82c 100644
--- a/tensorflow/core/common_runtime/eager/tensor_handle.h
+++ b/tensorflow/core/common_runtime/eager/tensor_handle.h
@@ -226,19 +226,13 @@ class TensorHandle : public ImmediateExecutionTensorHandle,
 
   string DebugString() const;
 
-  struct ResourceHandleInfo {
-    std::vector dtypes_and_shapes;
-    std::vector allowed_devices;
-  };
-
-  void SetResourceHandleInfo(ResourceHandleInfo&& resource_handle_info);
+  void SetResourceHandleDtypeAndShape(
+      std::vector dtypes_and_shapes);
 
   // If this TensorHandle is 1) a local tensor, and 2) a resource handle,
-  // return data types, shapes and allowed devices of the underlying resource.
-  Status GetResourceHandleInfo(ResourceHandleInfo* result);
+  // return data types and shapes of the underlying resource.
   Status GetResourceHandleDtypesAndShapes(
       std::vector* result);
-  Status GetResourceAllowedDevices(std::vector* result);
 
   // Returns the number of packed handles. 0 if the handle type is not PACKED.
   int NumPackedHandles() const;
@@ -261,8 +255,6 @@ class TensorHandle : public ImmediateExecutionTensorHandle,
   // with a ready version of the tensor handle data.
   bool IsReady() const;
 
-  Status GetResourceHandleInfoImpl(std::function set_resource_info);
-
   VariantDevice const device_;
 
   // Device in which the op producing this tensor was executed. Equals to
@@ -308,9 +300,9 @@ class TensorHandle : public ImmediateExecutionTensorHandle,
   Status is_poisoned_;
 
   // If this TensorHandle 1) is a local tensor, and 2) is a resource handle or
-  // refers to a remote resource handle, we store data types, shapes and allowed
-  // devices for the underlying resource.
-  ResourceHandleInfo resource_handle_info_;
+  // refers to a remote resource handle, we store data types and shapes for
+  // the underlying resource.
+  std::vector handle_dtypes_and_shapes_;
 
   // A handle data which refers to multiple TensorHandles of the same dtype and
   // shape.
diff --git a/tensorflow/core/common_runtime/eager/tensor_handle_test.cc b/tensorflow/core/common_runtime/eager/tensor_handle_test.cc
index 28092c0a604..40cec3fcc49 100644
--- a/tensorflow/core/common_runtime/eager/tensor_handle_test.cc
+++ b/tensorflow/core/common_runtime/eager/tensor_handle_test.cc
@@ -150,13 +150,13 @@ TEST_F(PackedTensorHandleTest, PackedHandle) {
   Device* d0 = ListDevices().at(0);
   TensorHandle* h0 =
       TensorHandle::CreateLocalHandle(std::move(t0), d0, d0, d0, context());
-  h0->SetResourceHandleInfo({{dtype_and_shape}, {}});
+  h0->SetResourceHandleDtypeAndShape({dtype_and_shape});
   handles.push_back(h0);
   Tensor t1(dtype, shape);
   Device* d1 = ListDevices().at(1);
   TensorHandle* h1 =
       TensorHandle::CreateLocalHandle(std::move(t1), d1, d1, d1, context());
-  h1->SetResourceHandleInfo({{dtype_and_shape}, {}});
+  h1->SetResourceHandleDtypeAndShape({dtype_and_shape});
   handles.push_back(h1);
 
   // Create 2 remote TensorHandles (not ready).
@@ -185,13 +185,12 @@ TEST_F(PackedTensorHandleTest, PackedHandle) {
   TensorShape packed_shape;
   TF_ASSERT_OK(packed_handle->Shape(&packed_shape));
   EXPECT_EQ(packed_shape, shape);
-  TensorHandle::ResourceHandleInfo resource_handle_info;
-  TF_ASSERT_OK(packed_handle->GetResourceHandleInfo(&resource_handle_info));
-  EXPECT_EQ(resource_handle_info.dtypes_and_shapes.size(), 1);
-  EXPECT_EQ(resource_handle_info.dtypes_and_shapes.at(0).dtype, DT_FLOAT);
-  EXPECT_EQ(
-      resource_handle_info.dtypes_and_shapes.at(0).shape.IsIdenticalTo({2, 2}),
-      true);
+  std::vector dtypes_and_shapes;
+  TF_ASSERT_OK(
+      packed_handle->GetResourceHandleDtypesAndShapes(&dtypes_and_shapes));
+  EXPECT_EQ(dtypes_and_shapes.size(), 1);
+  EXPECT_EQ(dtypes_and_shapes.at(0).dtype, DT_FLOAT);
+  EXPECT_EQ(dtypes_and_shapes.at(0).shape.IsIdenticalTo({2, 2}), true);
 
   CompositeDevice* device = reinterpret_cast(
       absl::get(packed_handle->device()));
diff --git a/tensorflow/core/distributed_runtime/eager/remote_mgr.cc b/tensorflow/core/distributed_runtime/eager/remote_mgr.cc
index 94a4f199337..9003f2b3f17 100644
--- a/tensorflow/core/distributed_runtime/eager/remote_mgr.cc
+++ b/tensorflow/core/distributed_runtime/eager/remote_mgr.cc
@@ -167,24 +167,22 @@ Status RemoteMgr::DeserializeRemoteTensorHandle(const RemoteTensorHandle& in,
         parent_->FindDeviceFromName(device_name.c_str(), &device));
     *out = TensorHandle::CreateLazyRemoteHandle(in.op_id(), in.output_num(),
                                                 in.dtype(), device, parent_);
-    TensorHandle::ResourceHandleInfo resource_handle_info;
-    std::vector* dtypes_and_shapes =
-        &resource_handle_info.dtypes_and_shapes;
+    std::vector dtypes_and_shapes;
     if (!GetMirroredResourceShape(RemoteTensorHandleInternal(in),
-                                  dtypes_and_shapes)
+                                  &dtypes_and_shapes)
              .ok()) {
       for (const auto& dtype_and_shape_proto :
            in.resource_dtypes_and_shapes()) {
-        dtypes_and_shapes->push_back(DtypeAndPartialTensorShape{
+        dtypes_and_shapes.push_back(DtypeAndPartialTensorShape{
             dtype_and_shape_proto.dtype(),
             TensorShape(dtype_and_shape_proto.shape())});
       }
       mutex_lock l(mirrored_resource_shape_mu_);
       mirrored_resource_shape_map_.emplace(
           RemoteTensorHandleInternal(in.op_id(), in.output_num()),
-          *dtypes_and_shapes);
+          dtypes_and_shapes);
     }
-    (*out)->SetResourceHandleInfo(std::move(resource_handle_info));
+    (*out)->SetResourceHandleDtypeAndShape(std::move(dtypes_and_shapes));
   }
 
   return Status::OK();
diff --git a/tensorflow/core/framework/resource_handle.cc b/tensorflow/core/framework/resource_handle.cc
index 2db5cfa301c..e7f4c2afc90 100644
--- a/tensorflow/core/framework/resource_handle.cc
+++ b/tensorflow/core/framework/resource_handle.cc
@@ -38,9 +38,6 @@ void ResourceHandle::AsProto(ResourceHandleProto* proto) const {
     dtype_and_shape->set_dtype(dtype_and_shape_pair.dtype);
     dtype_and_shape_pair.shape.AsProto(dtype_and_shape->mutable_shape());
   }
-  for (const string& device : allowed_devices_) {
-    *proto->add_allowed_devices() = device;
-  }
 }
 
 void ResourceHandle::FromProto(const ResourceHandleProto& proto) {
@@ -56,9 +53,6 @@ void ResourceHandle::FromProto(const ResourceHandleProto& proto) {
     dtypes_and_shapes.push_back(DtypeAndPartialTensorShape{dtype, shape});
   }
   dtypes_and_shapes_ = std::move(dtypes_and_shapes);
-  for (const string& device : proto.allowed_devices()) {
-    allowed_devices_.push_back(device);
-  }
 }
 
 string ResourceHandle::SerializeAsString() const {
diff --git a/tensorflow/core/framework/resource_handle.h b/tensorflow/core/framework/resource_handle.h
index 88c9f9da190..9acb94b6e79 100644
--- a/tensorflow/core/framework/resource_handle.h
+++ b/tensorflow/core/framework/resource_handle.h
@@ -39,14 +39,8 @@ class ResourceHandle {
 
   // Unique name for the device containing the resource.
   const std::string& device() const { return device_; }
-  // Names of the devices containing the resource.
-  const std::vector& allowed_devices() const {
-    return allowed_devices_;
-  }
+
   void set_device(const std::string& device) { device_ = device; }
-  void set_allowed_devices(const std::vector& devices) {
-    allowed_devices_ = devices;
-  }
 
   // Container in which this resource is placed.
   const std::string& container() const { return container_; }
@@ -93,12 +87,7 @@ class ResourceHandle {
       "cd2c89b7-88b7-44c8-ad83-06c2a9158347";
 
  public:
-  // The default device containing the resource, where the ResourceHandle is
-  // initially created.
   std::string device_;
-  // A set of devices containing the resource. If empty, the resource only
-  // exists on device_. Can be represented in wildcard patterns.
-  std::vector allowed_devices_;
   std::string container_;
   std::string name_;
   uint64 hash_code_ = 0;
diff --git a/tensorflow/core/framework/resource_handle.proto b/tensorflow/core/framework/resource_handle.proto
index eb0d1631c2f..5a41750475d 100644
--- a/tensorflow/core/framework/resource_handle.proto
+++ b/tensorflow/core/framework/resource_handle.proto
@@ -41,7 +41,5 @@ message ResourceHandleProto {
   // Data types and shapes for the underlying resource.
   repeated DtypeAndShape dtypes_and_shapes = 6;
 
-  // A set of devices containing the resource. If empty, the resource only
-  // exists on `device`.
-  repeated string allowed_devices = 7;
+  reserved 7;
 }
diff --git a/tensorflow/core/framework/resource_mgr.cc b/tensorflow/core/framework/resource_mgr.cc
index fd524b05bb9..e6ecfbb9190 100644
--- a/tensorflow/core/framework/resource_mgr.cc
+++ b/tensorflow/core/framework/resource_mgr.cc
@@ -36,8 +36,7 @@ static std::atomic current_id_;
 ResourceHandle MakeResourceHandle(
     const string& container, const string& name, const DeviceBase& device,
     const TypeIndex& type_index,
-    const std::vector& dtypes_and_shapes,
-    const std::vector& allowed_devices) {
+    const std::vector& dtypes_and_shapes) {
   ResourceHandle result;
   result.set_device(device.name());
   result.set_container(container);
@@ -49,7 +48,6 @@ ResourceHandle MakeResourceHandle(
   result.set_hash_code(type_index.hash_code());
   result.set_maybe_type_name(type_index.name());
   result.set_dtypes_and_shapes(dtypes_and_shapes);
-  result.set_allowed_devices(allowed_devices);
   return result;
 }
 
@@ -67,39 +65,12 @@ Status MakeResourceHandleToOutput(OpKernelContext* context, int output_index,
 namespace internal {
 
 Status ValidateDevice(OpKernelContext* ctx, const ResourceHandle& p) {
-  const string& current_device_name = ctx->device()->attributes().name();
-  if (current_device_name == p.device()) {
-    return Status::OK();
-  }
-  DeviceNameUtils::ParsedName parsed_current_device_name;
-  if (!DeviceNameUtils::ParseFullName(current_device_name,
-                                      &parsed_current_device_name)) {
+  if (ctx->device()->attributes().name() != p.device()) {
     return errors::InvalidArgument(
-        "Cannot parse device name in OpKernelContext: ", current_device_name);
+        "Trying to access resource ", p.name(), " located in device ",
+        p.device(), " from device ", ctx->device()->attributes().name());
   }
-
-  for (const string& device : p.allowed_devices()) {
-    DeviceNameUtils::ParsedName parsed;
-    if (!DeviceNameUtils::ParseFullName(device, &parsed)) {
-      return errors::InvalidArgument("Cannot parse allowed device name: ",
-                                     device);
-    }
-    if (DeviceNameUtils::IsCompleteSpecification(parsed,
-                                                 parsed_current_device_name)) {
-      return Status::OK();
-    }
-  }
-  string error_message = strings::StrCat("Trying to access resource ", p.name(),
-                                         " located in device ", p.device(),
-                                         " from device ", current_device_name);
-  if (!p.allowed_devices().empty()) {
-    absl::StrAppend(&error_message, " (allowed devices: ");
-    for (const string& device : p.allowed_devices()) {
-      absl::StrAppend(&error_message, device, ", ");
-    }
-    absl::StrAppend(&error_message, ") ");
-  }
-  return errors::InvalidArgument(error_message);
+  return Status::OK();
 }
 
 }  // end namespace internal
diff --git a/tensorflow/core/framework/resource_mgr.h b/tensorflow/core/framework/resource_mgr.h
index 3a9b97c7831..b0e4eace16e 100644
--- a/tensorflow/core/framework/resource_mgr.h
+++ b/tensorflow/core/framework/resource_mgr.h
@@ -291,31 +291,27 @@ class ResourceMgr {
 ResourceHandle MakeResourceHandle(
     const string& container, const string& name, const DeviceBase& device,
     const TypeIndex& type_index,
-    const std::vector& dtypes_and_shapes = {},
-    const std::vector& allowed_devices = {}) TF_MUST_USE_RESULT;
+    const std::vector& dtypes_and_shapes = {})
+    TF_MUST_USE_RESULT;
 
 template 
 ResourceHandle MakeResourceHandle(
     OpKernelContext* ctx, const string& container, const string& name,
-    const std::vector& dtypes_and_shapes = {},
-    const std::vector& allowed_devices = {}) {
-  return MakeResourceHandle(container.empty()
-                                ? ctx->resource_manager()->default_container()
-                                : container,
-                            name, *ctx->device(), MakeTypeIndex(),
-                            dtypes_and_shapes, allowed_devices);
+    const std::vector& dtypes_and_shapes = {}) {
+  return MakeResourceHandle(
+      container.empty() ? ctx->resource_manager()->default_container()
+                        : container,
+      name, *ctx->device(), MakeTypeIndex(), dtypes_and_shapes);
 }
 
 template 
 ResourceHandle MakeResourceHandle(
     OpKernelConstruction* ctx, const string& container, const string& name,
-    const std::vector& dtypes_and_shapes = {},
-    const std::vector& allowed_devices = {}) {
-  return MakeResourceHandle(container.empty()
-                                ? ctx->resource_manager()->default_container()
-                                : container,
-                            name, *ctx->device(), MakeTypeIndex(),
-                            dtypes_and_shapes, allowed_devices);
+    const std::vector& dtypes_and_shapes = {}) {
+  return MakeResourceHandle(
+      container.empty() ? ctx->resource_manager()->default_container()
+                        : container,
+      name, *ctx->device(), MakeTypeIndex(), dtypes_and_shapes);
 }
 
 Status MakeResourceHandleToOutput(OpKernelContext* context, int output_index,
diff --git a/tensorflow/core/framework/resource_mgr_test.cc b/tensorflow/core/framework/resource_mgr_test.cc
index a48024123a6..f524ff77c11 100644
--- a/tensorflow/core/framework/resource_mgr_test.cc
+++ b/tensorflow/core/framework/resource_mgr_test.cc
@@ -352,51 +352,4 @@ TEST(ResourceHandleTest, DeleteUsingResourceHandle) {
   EXPECT_NE(LookupResource(&ctx, p, &lookup_r).ok(), true);
 }
 
-TEST(ResourceHandleTest, AllowedDevices) {
-  const std::vector device_names = {
-      "/job:worker/replica:0/task:0/device:CPU:0",
-      "/job:worker/replica:0/task:0/device:CPU:2",
-      "/job:worker/replica:1/task:3/device:CPU:5"};
-  std::vector devices;
-  for (const string& name : device_names) {
-    devices.emplace_back(name);
-  }
-
-  std::vector params(device_names.size());
-  std::vector> resource_mgrs;
-  std::vector> ctxs;
-  for (int i = 0; i < device_names.size(); ++i) {
-    resource_mgrs.emplace_back(
-        absl::make_unique(/* default_container= */ ""));
-    params[i].resource_manager = resource_mgrs[i].get();
-    params[i].device = &(devices[i]);
-    ctxs.emplace_back(
-        absl::make_unique(&(params[i]), /* num_outputs= */ 0));
-  }
-
-  const string partially_specified_name =
-      "/job:worker/replica:0/task:0/device:CPU:*";
-  const string& fully_specified_name = device_names.at(2);
-  const std::vector allowed_devices = {partially_specified_name,
-                                               fully_specified_name};
-  // Create a ResourceHandle on device 0.
-  ResourceHandle p = MakeResourceHandle(
-      ctxs[0].get(), "container", "name",
-      /* dtypes_and_shapes= */ {}, allowed_devices);
-
-  std::vector resources;
-  for (const auto& ctx : ctxs) {
-    StubResource* r = new StubResource;
-    TF_EXPECT_OK(CreateResource(ctx.get(), p, r));
-    resources.push_back(r);
-  }
-
-  for (int i = 0; i < ctxs.size(); ++i) {
-    core::RefCountPtr lookup_r;
-    TF_EXPECT_OK(LookupResource(ctxs[i].get(), p, &lookup_r));
-    EXPECT_EQ(lookup_r.get(), resources[i]);
-    TF_EXPECT_OK(DeleteResource(ctxs[i].get(), p));
-  }
-}
-
 }  // end namespace tensorflow
diff --git a/tensorflow/core/kernels/resource_variable_ops.cc b/tensorflow/core/kernels/resource_variable_ops.cc
index 0fc1d53749f..b9c883c7e2f 100644
--- a/tensorflow/core/kernels/resource_variable_ops.cc
+++ b/tensorflow/core/kernels/resource_variable_ops.cc
@@ -222,8 +222,6 @@ VarHandleOp::VarHandleOp(OpKernelConstruction* context) : OpKernel(context) {
   OP_REQUIRES_OK(context, context->GetAttr("dtype", &dtype_and_shape_.dtype));
   PartialTensorShape shape;
   OP_REQUIRES_OK(context, context->GetAttr("shape", &dtype_and_shape_.shape));
-  OP_REQUIRES_OK(context,
-                 context->GetAttr("allowed_devices", &allowed_devices_));
 
   is_anonymous_ = name_ == ResourceHandle::ANONYMOUS_NAME;
 
@@ -234,8 +232,7 @@ VarHandleOp::VarHandleOp(OpKernelConstruction* context) : OpKernel(context) {
                                                    &resource_, attr));
     resource_.scalar()() = MakeResourceHandle(
         context, container_, name_,
-        std::vector{dtype_and_shape_},
-        allowed_devices_);
+        std::vector{dtype_and_shape_});
   }
 }
 
@@ -248,8 +245,7 @@ void VarHandleOp::Compute(OpKernelContext* ctx) {
         ctx, ctx->allocate_temp(DT_RESOURCE, TensorShape({}), &handle, attr));
     handle.scalar()() = MakeResourceHandle(
         ctx, container_, name_,
-        std::vector{dtype_and_shape_},
-        allowed_devices_);
+        std::vector{dtype_and_shape_});
     ctx->set_output(0, handle);
   } else {
     ctx->set_output(0, resource_);
diff --git a/tensorflow/core/kernels/resource_variable_ops.h b/tensorflow/core/kernels/resource_variable_ops.h
index 5935fa91d21..1bb70b537c1 100644
--- a/tensorflow/core/kernels/resource_variable_ops.h
+++ b/tensorflow/core/kernels/resource_variable_ops.h
@@ -36,10 +36,6 @@ class VarHandleOp : public OpKernel {
   Tensor resource_;
 
   DtypeAndPartialTensorShape dtype_and_shape_;
-
-  // A set of devices containing the resource variable. Set when the output
-  // ResourceHandle represents a per-replica/partitioned resource variable.
-  std::vector allowed_devices_;
 };
 
 class ReadVariableOp : public OpKernel {
diff --git a/tensorflow/python/kernel_tests/resource_variable_ops_test.py b/tensorflow/python/kernel_tests/resource_variable_ops_test.py
index b45e9dfb2bc..fb172fbcb10 100644
--- a/tensorflow/python/kernel_tests/resource_variable_ops_test.py
+++ b/tensorflow/python/kernel_tests/resource_variable_ops_test.py
@@ -30,7 +30,6 @@ from tensorflow.core.framework import tensor_pb2
 from tensorflow.python.eager import backprop
 from tensorflow.python.eager import context
 from tensorflow.python.eager import def_function
-from tensorflow.python.framework import config
 from tensorflow.python.framework import constant_op
 from tensorflow.python.framework import cpp_shape_inference_pb2
 from tensorflow.python.framework import dtypes
@@ -1513,41 +1512,5 @@ class ResourceVariableOpsTest(test_util.TensorFlowTestCase,
     self.assertAllEqual(expected, result)
 
 
-class PerReplicaResourceHandleTest(test_util.TensorFlowTestCase):
-
-  def setUp(self):
-    super(PerReplicaResourceHandleTest, self).setUp()
-    cpus = config.list_physical_devices("CPU")
-    # Set 2 virtual CPUs
-    config.set_logical_device_configuration(cpus[0], [
-        context.LogicalDeviceConfiguration(),
-        context.LogicalDeviceConfiguration(),
-    ])
-
-  @test_util.disable_tfrt("Multiple device support. b/154956430")
-  def testAllowedDevices(self):
-    device0 = "/job:localhost/replica:0/task:0/device:CPU:0"
-    device1 = "/job:localhost/replica:0/task:0/device:CPU:1"
-    value0 = 1
-    value1 = 2
-    with context.eager_mode():
-      handle = resource_variable_ops.var_handle_op(
-          dtype=dtypes.int32, shape=[], allowed_devices=[device0, device1])
-      with ops.device(device0):
-        assign0 = resource_variable_ops.assign_variable_op(handle, value0)
-      with ops.device(device1):
-        assign1 = resource_variable_ops.assign_variable_op(handle, value1)
-      with ops.control_dependencies([assign0, assign1]):
-        with ops.device(device0):
-          read0 = resource_variable_ops.read_variable_op(
-              handle, dtype=dtypes.int32)
-        with ops.device(device1):
-          read1 = resource_variable_ops.read_variable_op(
-              handle, dtype=dtypes.int32)
-
-      self.assertAllEqual(value0, read0)
-      self.assertAllEqual(value1, read1)
-
-
 if __name__ == "__main__":
   test.main()

From c27b834b49002cca86e8ad2bcbb98a71dfa7538a Mon Sep 17 00:00:00 2001
From: Bruce Fontaine 
Date: Mon, 22 Jun 2020 13:09:21 -0700
Subject: [PATCH 462/867] Wrap save/restore logic in tf.function when in eager
 mode. This allows parallel saving and restoring when using multiple devices.

PiperOrigin-RevId: 317719780
Change-Id: Ifb7e34f708da4121b49fb38d8dad046d45fedc42
---
 .../grappler/optimizers/function_optimizer.cc |   8 +-
 .../parallel_device/parallel_device_test.py   |   4 +
 .../python/framework/auto_control_deps.py     |   2 +-
 tensorflow/python/training/saving/BUILD       |   1 +
 .../training/saving/functional_saver.py       | 111 ++++++++++++------
 .../training/saving/functional_saver_test.py  |  17 ++-
 6 files changed, 101 insertions(+), 42 deletions(-)

diff --git a/tensorflow/core/grappler/optimizers/function_optimizer.cc b/tensorflow/core/grappler/optimizers/function_optimizer.cc
index a66e645e04b..0e156aaa84c 100644
--- a/tensorflow/core/grappler/optimizers/function_optimizer.cc
+++ b/tensorflow/core/grappler/optimizers/function_optimizer.cc
@@ -837,7 +837,6 @@ const bool IsExemptFromSideEffectsExecutionValidation(const string& op) {
        "ParameterizedTruncatedNormal", "TruncatedNormal", "RandomShuffle",
        "Multinomial", "RandomGamma", "RandomGammaGrad", "RandomPoisson",
        "RandomPoissonV2",
-       // LINT.ThenChange(//tensorflow/python/framework/auto_control_deps.py)
 
        // ReadVariableOp marked as stateful because it consumes DT_RESOURCE,
        // but it can't generate any observable side-effect.
@@ -851,7 +850,12 @@ const bool IsExemptFromSideEffectsExecutionValidation(const string& op) {
        // the same device_ordinal on the same host.
        "EnqueueTPUEmbeddingSparseBatch", "EnqueueTPUEmbeddingIntegerBatch",
        "EnqueueTPUEmbeddingSparseTensorBatch",
-       "EnqueueTPUEmbeddingRaggedTensorBatch"});
+       "EnqueueTPUEmbeddingRaggedTensorBatch",
+
+       // SaveV2 and RestoreV2 should be allowed to operate in parallel on
+       // multiple hosts.
+       "SaveV2", "RestoreV2"});
+  // LINT.ThenChange(//tensorflow/python/framework/auto_control_deps.py)
   return exemption->contains(op);
 }
 
diff --git a/tensorflow/python/distribute/parallel_device/parallel_device_test.py b/tensorflow/python/distribute/parallel_device/parallel_device_test.py
index 8fc3dcb5816..1429c522aba 100644
--- a/tensorflow/python/distribute/parallel_device/parallel_device_test.py
+++ b/tensorflow/python/distribute/parallel_device/parallel_device_test.py
@@ -172,6 +172,8 @@ class ParallelDeviceTests(_VirtualDeviceTestCase):
       config.set_synchronous_execution(previous)
 
   def test_checkpointing(self):
+    self.skipTest(
+        "Disable saving until SaveableObject's methods are traceable.")
     prefix = os.path.join(self.get_temp_dir(), "ckpt")
     with self.device.scope():
       different_values = self.device.pack(
@@ -263,6 +265,8 @@ class LayerTests(_VirtualDeviceTestCase):
     self.assertIn(self.device.components[1], final_kernels[1].backing_device)
 
   def test_training_loop(self):
+    self.skipTest(
+        "Disable saving until SaveableObject's methods are traceable.")
     for _ in range(5):
       layer = _Dense(5)
       checkpoint = tracking.Checkpoint(layer=layer)
diff --git a/tensorflow/python/framework/auto_control_deps.py b/tensorflow/python/framework/auto_control_deps.py
index 51dcb248b11..4b47735e0bf 100644
--- a/tensorflow/python/framework/auto_control_deps.py
+++ b/tensorflow/python/framework/auto_control_deps.py
@@ -100,7 +100,7 @@ _ORDER_INSENSITIVE_STATEFUL_OPS = [
     "CudnnRNNV2", "CudnnRNNV3", "CudnnRNNBackpropV2", "CudnnRNNBackpropV3",
     "EnqueueTPUEmbeddingSparseBatch", "EnqueueTPUEmbeddingIntegerBatch",
     "EnqueueTPUEmbeddingSparseTensorBatch",
-    "EnqueueTPUEmbeddingRaggedTensorBatch"
+    "EnqueueTPUEmbeddingRaggedTensorBatch", "RestoreV2", "SaveV2"
 ]
 # LINT.ThenChange(//tensorflow/core/grappler/optimizers/function_optimizer.cc)
 
diff --git a/tensorflow/python/training/saving/BUILD b/tensorflow/python/training/saving/BUILD
index 670a4c35c6f..12940840309 100644
--- a/tensorflow/python/training/saving/BUILD
+++ b/tensorflow/python/training/saving/BUILD
@@ -43,6 +43,7 @@ cuda_py_test(
         ":checkpoint_options",
         ":functional_saver",
         ":saveable_hook",
+        "//tensorflow/python/eager:remote",
         "//tensorflow/python/eager:test",
     ],
 )
diff --git a/tensorflow/python/training/saving/functional_saver.py b/tensorflow/python/training/saving/functional_saver.py
index c4334e096df..3a9b565470d 100644
--- a/tensorflow/python/training/saving/functional_saver.py
+++ b/tensorflow/python/training/saving/functional_saver.py
@@ -21,6 +21,7 @@ from __future__ import print_function
 import uuid
 
 from tensorflow.core.protobuf import saver_pb2
+from tensorflow.python.eager import context
 from tensorflow.python.eager import def_function
 from tensorflow.python.framework import constant_op
 from tensorflow.python.framework import dtypes
@@ -161,7 +162,8 @@ class MultiDeviceSaver(object):
         self._after_restore_callbacks.append(saveable.after_restore)
 
       if is_saveable:
-        saveables_by_device.setdefault(saveable.device, []).append(saveable)
+        host_device = saveable_object_util.set_cpu0(saveable.device)
+        saveables_by_device.setdefault(host_device, []).append(saveable)
 
     self._single_device_savers = {
         device: _SingleDeviceSaver(saveables)
@@ -247,33 +249,50 @@ class MultiDeviceSaver(object):
       tmp_checkpoint_prefix = string_ops.string_join(
           [file_prefix, sharded_suffix])
 
-    num_shards = len(self._single_device_savers)
-    sharded_saves = []
-    sharded_prefixes = []
-    num_shards_tensor = constant_op.constant(num_shards, name="num_shards")
-    last_device = None
-    for shard, (device, saver) in enumerate(
-        sorted(self._single_device_savers.items())):
-      last_device = device
-      with ops.device(saveable_object_util.set_cpu0(device)):
-        shard_prefix = sharded_filename(tmp_checkpoint_prefix, shard,
-                                        num_shards_tensor)
-      sharded_prefixes.append(shard_prefix)
-      with ops.device(device):
-        # _SingleDeviceSaver will use the CPU device when necessary, but initial
-        # read operations should be placed on the SaveableObject's device.
-        sharded_saves.append(saver.save(shard_prefix, options))
+    def save_fn():
+      num_shards = len(self._single_device_savers)
+      sharded_saves = []
+      sharded_prefixes = []
+      num_shards_tensor = constant_op.constant(num_shards, name="num_shards")
+      last_device = None
+      for shard, (device, saver) in enumerate(
+          sorted(self._single_device_savers.items())):
+        last_device = device
+        with ops.device(saveable_object_util.set_cpu0(device)):
+          shard_prefix = sharded_filename(tmp_checkpoint_prefix, shard,
+                                          num_shards_tensor)
+        sharded_prefixes.append(shard_prefix)
+        with ops.device(device):
+          # _SingleDeviceSaver will use the CPU device when necessary, but
+          # initial read operations should be placed on the SaveableObject's
+          # device.
+          sharded_saves.append(saver.save(shard_prefix, options))
 
-    with ops.control_dependencies(sharded_saves):
-      # Merge on the io_device if specified, otherwise co-locates the merge op
-      # with the last device used.
-      merge_device = (options.experimental_io_device or
-                      saveable_object_util.set_cpu0(last_device))
-      with ops.device(merge_device):
-        # V2 format write path consists of a metadata merge step.  Once merged,
-        # attempts to delete the temporary directory, "_temp".
-        return gen_io_ops.merge_v2_checkpoints(
-            sharded_prefixes, file_prefix, delete_old_dirs=True)
+      with ops.control_dependencies(sharded_saves):
+        # Merge on the io_device if specified, otherwise co-locates the merge op
+        # with the last device used.
+        merge_device = (
+            options.experimental_io_device or
+            saveable_object_util.set_cpu0(last_device))
+        with ops.device(merge_device):
+          # V2 format write path consists of a metadata merge step.  Once
+          # merged, attempts to delete the temporary directory,
+          # "_temp".
+          return gen_io_ops.merge_v2_checkpoints(
+              sharded_prefixes, file_prefix, delete_old_dirs=True)
+
+    # Since this will causes a function re-trace on each save, limit this to the
+    # cases where it is needed: eager and when there are multiple tasks/single
+    # device savers. Note that the retrace is needed to ensure we pickup the
+    # latest values of options like experimental_io_device.
+    if context.executing_eagerly() and len(self._single_device_savers) > 1:
+      # Explicitly place the identity op on the first device.
+      @def_function.function(experimental_compile=False)
+      def tf_function_save():
+        save_fn()
+      tf_function_save()
+    else:
+      return save_fn()
 
   def restore(self, file_prefix, options=None):
     """Restore the saveable objects from a checkpoint with `file_prefix`.
@@ -287,12 +306,38 @@ class MultiDeviceSaver(object):
       A dictionary mapping from SaveableObject names to restore operations.
     """
     options = options or checkpoint_options.CheckpointOptions()
-    restore_ops = {}
-    # Sort by device name to avoid propagating non-deterministic dictionary
-    # ordering in some Python versions.
-    for device, saver in sorted(self._single_device_savers.items()):
-      with ops.device(device):
-        restore_ops.update(saver.restore(file_prefix, options))
+
+    def restore_fn():
+      restore_ops = {}
+      # Sort by device name to avoid propagating non-deterministic dictionary
+      # ordering in some Python versions.
+      for device, saver in sorted(self._single_device_savers.items()):
+        with ops.device(device):
+          restore_ops.update(saver.restore(file_prefix, options))
+
+      return restore_ops
+
+    # Since this will causes a function re-trace on each save, limit this to the
+    # cases where it is needed: eager and when there are multiple tasks/single
+    # device savers. Note that the retrace is needed to ensure we pickup the
+    # latest values of options like experimental_io_device.
+    if context.executing_eagerly() and len(self._single_device_savers) > 1:
+      first_device, _ = list(self._single_device_savers.items())[0]
+      @def_function.function(experimental_compile=False)
+      def tf_function_restore():
+        restore_ops = restore_fn()
+        restore_tensors = {}
+        # tf.functions must return tensors, thus we use control dependencies so
+        # that we can return a tensor which depends on the given op.
+        with ops.device(saveable_object_util.set_cpu0(first_device)):
+          for name, op in restore_ops.items():
+            with ops.control_dependencies([op]):
+              restore_tensors[name] = array_ops.identity(file_prefix)
+        return restore_tensors
+
+      restore_ops = tf_function_restore()
+    else:
+      restore_ops = restore_fn()
 
     for callback in self._after_restore_callbacks:
       callback()
diff --git a/tensorflow/python/training/saving/functional_saver_test.py b/tensorflow/python/training/saving/functional_saver_test.py
index 7db32ff72d7..8f3eef4fb9c 100644
--- a/tensorflow/python/training/saving/functional_saver_test.py
+++ b/tensorflow/python/training/saving/functional_saver_test.py
@@ -21,6 +21,7 @@ from __future__ import print_function
 import os
 
 from tensorflow.python.eager import context
+from tensorflow.python.eager import remote
 from tensorflow.python.eager import test
 from tensorflow.python.eager import wrap_function
 from tensorflow.python.framework import config
@@ -29,6 +30,7 @@ from tensorflow.python.framework import ops
 from tensorflow.python.framework import test_util
 from tensorflow.python.ops import resource_variable_ops
 from tensorflow.python.platform import gfile
+from tensorflow.python.training import server_lib
 from tensorflow.python.training.saving import checkpoint_options
 from tensorflow.python.training.saving import functional_saver
 from tensorflow.python.training.saving import saveable_hook
@@ -126,13 +128,16 @@ class SaverTest(test.TestCase):
     second_saver.restore(save_path)
     self.assertEqual(2., self.evaluate(v2))
 
-  @test_util.run_in_graph_and_eager_modes
-  def test_checkpoint_is_sharded_by_device(self):
-    with ops.device("cpu:0"):
+  def test_checkpoint_is_sharded_by_task(self):
+    servers = [server_lib.Server.create_local_server() for _ in range(3)]
+    cluster_spec = server_lib.ClusterSpec({
+        "worker": [s.target[len("grpc://"):] for s in servers]})
+    remote.connect_to_cluster(cluster_spec)
+    with ops.device("/job:worker/task:0/cpu:0"):
       v0 = resource_variable_ops.ResourceVariable(0.)
-    with ops.device("cpu:1"):
+    with ops.device("/job:worker/task:1/cpu:0"):
       v1 = resource_variable_ops.ResourceVariable(1.)
-    with ops.device("cpu:2"):
+    with ops.device("/job:worker/task:2/cpu:0"):
       v2 = resource_variable_ops.ResourceVariable(2.)
 
     self.evaluate([v0.initializer, v1.initializer, v2.initializer])
@@ -167,7 +172,7 @@ class SaverTest(test.TestCase):
         list(saveable_object_util.saveable_objects_for_op(v2, "v2")))
     prefix = os.path.join(self.get_temp_dir(), "ckpt")
     self.evaluate(saver.save(constant_op.constant(prefix), self.local_options))
-    self.assertEqual(4, len(gfile.Glob(prefix + "*")))
+    self.assertEqual(2, len(gfile.Glob(prefix + "*")))
     self.evaluate(v0.assign(-1.))
     self.evaluate(v1.assign(-1.))
     self.evaluate(v2.assign(-1.))

From 31acd85aee129cd7172c1498a72d2f136048a79e Mon Sep 17 00:00:00 2001
From: Feng Liu 
Date: Mon, 22 Jun 2020 13:39:29 -0700
Subject: [PATCH 463/867] Add quantization support for tfl.transpose_conv

PiperOrigin-RevId: 317725618
Change-Id: If98427ba8b732edc7e1bd30f9c0f11d565d08a84
---
 tensorflow/compiler/mlir/lite/ir/tfl_ops.td   |  4 ++++
 .../lite/tests/prepare-quantize-signed.mlir   | 23 +++++++++++++++++++
 2 files changed, 27 insertions(+)

diff --git a/tensorflow/compiler/mlir/lite/ir/tfl_ops.td b/tensorflow/compiler/mlir/lite/ir/tfl_ops.td
index 33281cc58fb..f379b241f9d 100644
--- a/tensorflow/compiler/mlir/lite/ir/tfl_ops.td
+++ b/tensorflow/compiler/mlir/lite/ir/tfl_ops.td
@@ -571,6 +571,8 @@ def TFL_TransposeConvOp: TFL_Op<"transpose_conv", [
     TFL_OperandHasRank<2, 4>,
     PredOpTrait<"input and output must have same element type",
       TFL_TCresVTEtIsSameAsOp<0, 2>>,
+    AccumulatorUniformScale<3, 1, 2>,
+    TFL_ChannelDimIndexInterface, AffineOpCoefficient<0, 2>,
     TFL_GpuTargetOp,
     TFL_SparseOp]> {
   let summary = "Transpose convolution operator";
@@ -596,6 +598,8 @@ def TFL_TransposeConvOp: TFL_Op<"transpose_conv", [
   let verifier = [{ return Verify(*this); }];
 
   let extraClassDeclaration = [{
+    // ChannelDimIndexInterface:
+    int GetChannelDimIndex() { return 0; }
     // SparseOpInterface:
     std::vector GetSparseOperands() { return {1}; }
     std::vector> GetFloatBlockSize() { return {}; }
diff --git a/tensorflow/compiler/mlir/lite/tests/prepare-quantize-signed.mlir b/tensorflow/compiler/mlir/lite/tests/prepare-quantize-signed.mlir
index e1f496b91f4..4a83616408e 100644
--- a/tensorflow/compiler/mlir/lite/tests/prepare-quantize-signed.mlir
+++ b/tensorflow/compiler/mlir/lite/tests/prepare-quantize-signed.mlir
@@ -70,6 +70,7 @@ func @prepareAdd(%arg0: tensor<2x2xf32>) -> tensor<2x2xf32> {
 }
 
 // CHECK-LABEL: prepareConv2DSplat
+// PerTensor-LABEL: prepareConv2DSplat
 func @prepareConv2DSplat(%arg0: tensor<1x5x5x3xf32>) -> tensor<1x5x5x3xf32> {
   %w = constant dense<127.0> : tensor<3x3x3x3xf32>
   %b = constant dense<0.0> : tensor<3xf32>
@@ -89,6 +90,7 @@ func @prepareConv2DSplat(%arg0: tensor<1x5x5x3xf32>) -> tensor<1x5x5x3xf32> {
 }
 
 // CHECK-LABEL: prepareConv2D
+// PerTensor-LABEL: prepareConv2D
 func @prepareConv2D(%arg0: tensor<1x5x5x1xf32>) -> tensor<1x5x5x3xf32> {
   %w = constant dense<[[[[0.0]]], [[[127.0]]], [[[-127.0]]]]> : tensor<3x1x1x1xf32>
   %b = constant dense<0.0> : tensor<3xf32>
@@ -108,6 +110,7 @@ func @prepareConv2D(%arg0: tensor<1x5x5x1xf32>) -> tensor<1x5x5x3xf32> {
 }
 
 // CHECK-LABEL: prepareDepthwiseConv2D
+// PerTensor-LABEL: prepareDepthwiseConv2D
 func @prepareDepthwiseConv2D(%arg0: tensor<1x224x224x3xf32>) -> tensor<1x112x112x32xf32> {
   %w = constant dense<127.0> : tensor<32x3x3x3xf32>
   %b = constant dense<0.0> : tensor<32xf32>
@@ -127,6 +130,7 @@ func @prepareDepthwiseConv2D(%arg0: tensor<1x224x224x3xf32>) -> tensor<1x112x112
 }
 
 // CHECK-LABEL: QuantizeFullyConnected
+// PerTensor-LABEL: QuantizeFullyConnected
 func @QuantizeFullyConnected(%arg0: tensor<1x224x224x3xf32>) -> tensor<1x112x112x32xf32> {
   %w = constant dense<127.0> : tensor<32x12xf32>
   %b = constant dense<0.0> : tensor<32xf32>
@@ -143,3 +147,22 @@ func @QuantizeFullyConnected(%arg0: tensor<1x224x224x3xf32>) -> tensor<1x112x112
 // PerTensor: %[[dq:.*]] = "tfl.dequantize"(%0) : (tensor<32x12x!quant.uniform:f32, 1.000000e+00>>) -> tensor<32x12xf32>
 // PerTensor: "tfl.fully_connected"(%arg0, %[[dq]]
 }
+
+// CHECK-LABEL: QuantizeTransposeConv
+// PerTensor-LABEL: QuantizeTransposeConv
+func @QuantizeTransposeConv(%arg0: tensor<32x4x4x128xf32>, %arg1: tensor<4xi32>) -> tensor<1x32x42x128xf32> {
+  %w = constant dense<127.0> : tensor<1x32x42x128xf32>
+  %b = constant dense<0.0> : tensor<1x32x42x128xf32>
+  %tc = "tfl.transpose_conv"(%arg1, %arg0, %w, %b) {padding = "SAME", stride_h = 2 : i32, stride_w = 2 : i32} : (tensor<4xi32>, tensor<32x4x4x128xf32>, tensor<1x32x42x128xf32>, tensor<1x32x42x128xf32>) -> tensor<1x32x42x128xf32>
+  return %tc : tensor<1x32x42x128xf32>
+
+// CHECK: %[[CST:.*]] = constant dense<1.270000e+02> : tensor<1x32x42x128xf32>
+// CHECK: %[[QUANTIZE:.*]] = "tfl.quantize"(%[[CST]]) {qtype = tensor<1x32x42x128x!quant.uniform:f32:0, {1.000000e+00}>>, volatile}
+// CHECK: %[[DEQUANTIZE:.*]] = "tfl.dequantize"(%[[QUANTIZE]]) : (tensor<1x32x42x128x!quant.uniform:f32:0, {1.000000e+00}>>) -> tensor<1x32x42x128xf32>
+// CHECK: "tfl.transpose_conv"(%arg1, %arg0, %[[DEQUANTIZE]]
+
+// PerTensor: %[[CST:.*]] = constant dense<1.270000e+02> : tensor<1x32x42x128xf32>
+// PerTensor: %[[QUANTIZE:.*]] = "tfl.quantize"(%[[CST]]) {qtype = tensor<1x32x42x128x!quant.uniform:f32, 1.000000e+00>>, volatile}
+// PerTensor: %[[DEQUANTIZE:.*]] = "tfl.dequantize"(%[[QUANTIZE]]) : (tensor<1x32x42x128x!quant.uniform:f32, 1.000000e+00>>) -> tensor<1x32x42x128xf32>
+// PerTensor: "tfl.transpose_conv"(%arg1, %arg0, %[[DEQUANTIZE]]
+}

From a3dc11ea136770d70d27c4ef02d01ee744d68bd4 Mon Sep 17 00:00:00 2001
From: Nick Kreeger 
Date: Mon, 22 Jun 2020 14:00:22 -0700
Subject: [PATCH 464/867] Add endian-aware flatbuffer conversions to
 MicroAllocator.

Currently, MicroAllocator manually maps TfLite*Array struct values directly to flatbuffer values. This change cleans up other instances inside MicroAllocator that are not endian-aware.

This works only on little-endian (LE) architecture systems because of the layout of TfLite*Array:

struct TfLiteIntArray {
  int size;
  int data[];
}

The compiler maintains mapping, but |size| and |data| are laid out as the following in LE:

[lowest-order-byte(e.g. data) .... highest-order-byte(e.g. size)]

Casting and remapping work on LE because the vector is written in the lowest-order-byte sequence. On BE systems, this memory savings trick does not work and requires a malloc from the arena and manual copying of values from the flatbuffer.

PiperOrigin-RevId: 317730072
Change-Id: I1baff898356e3d82b2faed6468a50ae44acd3082
---
 tensorflow/lite/micro/micro_allocator.cc | 96 +++++++++++++-----------
 1 file changed, 53 insertions(+), 43 deletions(-)

diff --git a/tensorflow/lite/micro/micro_allocator.cc b/tensorflow/lite/micro/micro_allocator.cc
index d6c192c3ed3..bf9e38d1050 100644
--- a/tensorflow/lite/micro/micro_allocator.cc
+++ b/tensorflow/lite/micro/micro_allocator.cc
@@ -392,28 +392,48 @@ TfLiteStatus CommitPlan(ErrorReporter* error_reporter, MemoryPlanner* planner,
 
 namespace internal {
 
-// Allocate a TfLiteIntArray and copy the contents of a FlatBuffers Vector
-// into it.
-template 
-TfLiteStatus FlatBufferIntArrayToTfLiteIntArray(
+// Handles architecture safe mapping of flatbuffer vectors to a TfLite*Array
+// struct. Matching types are required (e.g. float and TfLiteFloatArray).
+template 
+TfLiteStatus FlatBufferVectorToTfLiteTypeArray(
     SimpleMemoryAllocator* allocator, ErrorReporter* error_reporter,
-    const flatbuffers::Vector* flat_array, TfLiteIntArray** result) {
-  TfLiteIntArray* ret =
-      reinterpret_cast(allocator->AllocateFromTail(
-          TfLiteIntArrayGetSizeInBytes(flat_array->Length()),
-          alignof(TfLiteIntArray)));
-  if (nullptr == ret) {
-    TF_LITE_REPORT_ERROR(
-        error_reporter,
-        "Failed to allocate %d bytes of memory to copy an array.",
-        TfLiteIntArrayGetSizeInBytes(flat_array->Length()));
-    return kTfLiteError;
+    const flatbuffers::Vector* flatbuffer_array,
+    kTfLiteArrayType** result) {
+  TFLITE_DCHECK(error_reporter != nullptr);
+  TFLITE_DCHECK(flatbuffer_array != nullptr);
+  // Only two conversions are supported - float and int32 - ensure that these
+  // match at compile time instead of duplicating functions here:
+  static_assert((std::is_same() &&
+                 std::is_same()) ||
+                (std::is_same() &&
+                 std::is_same()));
+  if (FLATBUFFERS_LITTLEENDIAN) {
+    // On little-endian machines, TfLite*Array happens to have the same memory
+    // layout as flatbuffers:Vector, so we can
+    // reinterpret_cast the flatbuffer vector and avoid a copy and malloc.
+    *result = const_cast(
+        reinterpret_cast(flatbuffer_array));
+  } else {
+    // Big-endian architecture can not use the same memory layout as
+    // flatbuffers::Vector. Allocate from the tail and
+    // copy values from the flatbuffer into the newly allocated chunk.
+    kTfLiteArrayType* array =
+        reinterpret_cast(allocator->AllocateFromTail(
+            TfLiteIntArrayGetSizeInBytes(flatbuffer_array->Length()),
+            alignof(kTfLiteArrayType)));
+    if (array == nullptr) {
+      TF_LITE_REPORT_ERROR(
+          error_reporter,
+          "Failed to allocate %d bytes of memory to copy an array.",
+          TfLiteIntArrayGetSizeInBytes(flatbuffer_array->Length()));
+      return kTfLiteError;
+    }
+    array->size = flatbuffer_array->Length();
+    for (int i = 0; i < array->size; ++i) {
+      array->data[i] = flatbuffer_array->Get(i);
+    }
+    *result = array;
   }
-  ret->size = flat_array->Length();
-  for (int64_t i = 0; i < static_cast(flat_array->Length()); i++) {
-    ret->data[i] = flat_array->Get(i);
-  }
-  *result = ret;
   return kTfLiteOk;
 }
 
@@ -469,27 +489,17 @@ TfLiteStatus InitializeTfLiteTensorFromFlatbuffer(
   TF_LITE_ENSURE_STATUS(BytesRequiredForTensor(
       flatbuffer_tensor, &result->bytes, &type_size, error_reporter));
 
-  // TODO(b/159043126): Cleanup endian casting by doing all endian casting in
-  // one spot:
   if (flatbuffer_tensor.shape() == nullptr) {
     // flatbuffer_tensor.shape() can return a nullptr in the case of a scalar
     // tensor.
     result->dims = const_cast(&kZeroLengthIntArray);
-  } else if (!FLATBUFFERS_LITTLEENDIAN) {
-    // Big-endian architecture. Copy and byte-swap the little-endian shape
-    // data.
-    TF_LITE_ENSURE_STATUS(FlatBufferIntArrayToTfLiteIntArray(
-        allocator, error_reporter, flatbuffer_tensor.shape(), &(result->dims)));
   } else {
-    // On little-endian machines, TfLiteIntArray happens to have the same
-    // memory layout as flatbuffers:Vector, so we can reinterpret_cast the
-    // tensor shape vector and avoid a copy.
     // TFLM doesn't allow reshaping the tensor which requires dynamic memory
     // allocation so it is safe to drop the const qualifier. In the future, if
     // we really want to update the tensor shape, we can always pass in a new
-    // TfLiteIntArray - especially we have to do so if the dimension is changed.
-    result->dims = const_cast(
-        reinterpret_cast(flatbuffer_tensor.shape()));
+    // TfLiteIntArray - especially we have to do so if the dimension is
+    TF_LITE_ENSURE_STATUS(FlatBufferVectorToTfLiteTypeArray(
+        allocator, error_reporter, flatbuffer_tensor.shape(), &(result->dims)));
   }
 
   // Copy the quantization information from the serialized data.
@@ -531,9 +541,9 @@ TfLiteStatus InitializeTfLiteTensorFromFlatbuffer(
       return kTfLiteError;
     }
 
-    // TODO(b/159043126): Check for big endian before casting flatbuffer values.
-    quantization->scale = const_cast(
-        reinterpret_cast(src_quantization->scale()));
+    TF_LITE_ENSURE_STATUS(FlatBufferVectorToTfLiteTypeArray(
+        allocator, error_reporter, src_quantization->scale(),
+        &quantization->scale));
 
     quantization->zero_point->size = channels;
     int* zero_point_data = quantization->zero_point->data;
@@ -815,13 +825,13 @@ TfLiteStatus MicroAllocator::PrepareNodeAndRegistrationDataFromFlatbuffer(
                                    (void**)(&builtin_data)));
     }
 
-    // Disregard const qualifier to workaround with existing API.
-    // TODO(b/159043126): Check for big endian before casting flatbuffer values.
-    TfLiteIntArray* inputs_array = const_cast(
-        reinterpret_cast(op->inputs()));
-    // TODO(b/159043126): Check for big endian before casting flatbuffer values.
-    TfLiteIntArray* outputs_array = const_cast(
-        reinterpret_cast(op->outputs()));
+    TfLiteIntArray* inputs_array;
+    TF_LITE_ENSURE_STATUS(internal::FlatBufferVectorToTfLiteTypeArray(
+        memory_allocator_, error_reporter_, op->inputs(), &inputs_array));
+
+    TfLiteIntArray* outputs_array;
+    TF_LITE_ENSURE_STATUS(internal::FlatBufferVectorToTfLiteTypeArray(
+        memory_allocator_, error_reporter_, op->outputs(), &outputs_array));
 
     TfLiteNode* node = &(node_and_registrations[i].node);
     *node = {};

From 17fb3fcf7801df02c5d7696fb08e1dd55c9d0fe1 Mon Sep 17 00:00:00 2001
From: Rick Chao 
Date: Mon, 22 Jun 2020 14:21:25 -0700
Subject: [PATCH 465/867] MultiProcessRunner: Add UnexpectedSubprocessExitError
 to be raised if the exit code from a subprocess in unexpected. This results
 in subprocess having segfault failing the test, which would not have before
 this change.

PiperOrigin-RevId: 317734369
Change-Id: I008f9bd4dd5b7a68e8e4a7f17b7e5d490a30c09a
---
 .../python/distribute/multi_process_runner.py | 38 +++++++++++++++++--
 .../distribute/multi_process_runner_test.py   | 35 +++++++++++++++++
 2 files changed, 70 insertions(+), 3 deletions(-)

diff --git a/tensorflow/python/distribute/multi_process_runner.py b/tensorflow/python/distribute/multi_process_runner.py
index 8699e59b410..af527b67b4b 100644
--- a/tensorflow/python/distribute/multi_process_runner.py
+++ b/tensorflow/python/distribute/multi_process_runner.py
@@ -463,14 +463,28 @@ class MultiProcessRunner(object):
     process_statuses = self._queue_to_list(self._process_status_queue)
     if not self._all_forced_terminated and len(
         process_statuses) != self._outstanding_subprocess_count:
-      raise RuntimeError(
-          'missing statuses from %d subproceses.' %
-          (self._outstanding_subprocess_count - len(process_statuses)))
+      raise UnexpectedSubprocessExitError(
+          'Missing status(es) from %d subprocess(es). See logs for details.' %
+          (self._outstanding_subprocess_count - len(process_statuses)),
+          self._get_mpr_result(process_statuses))
     for process_status in process_statuses:
       assert isinstance(process_status, _ProcessStatusInfo)
       if not process_status.is_successful:
         six.reraise(*process_status.exc_info)
 
+    # Checking all the processes that are expected to exit properly.
+    for (task_type, task_id), p in self._processes.items():
+      if self._dependence_on_chief and task_type != 'chief':
+        # If _dependence_on_chief, other processes may have been
+        # forced-terminated, which is expected.
+        continue
+      # Successfully exiting process has exit code 0.
+      if p.exitcode > 0:
+        raise UnexpectedSubprocessExitError(
+            'Subprocess %s-%d exited with exit code %d. See logs for details.' %
+            (task_type, task_id, p.exitcode),
+            self._get_mpr_result(process_statuses))
+
     logging.info('Joining log reading threads.')
     for thread in self._reading_threads:
       thread.join()
@@ -506,6 +520,8 @@ class MultiProcessRunner(object):
     for (task_type, task_id), p in self._processes.items():
       try:
         os.kill(p.pid, sig)
+        logging.info('%s-%d terminated with signal %r.', task_type, task_id,
+                     sig)
       except ProcessLookupError:
         logging.info('Attempting to kill %s-%d but it does not exist.',
                      task_type, task_id)
@@ -658,6 +674,9 @@ class _ProcFunc(object):
       self._resources.process_status_queue.put(info)
       self._close_streaming()
 
+    # Exit with code 0 as it's considered successful exit at this point.
+    sys.exit(0)
+
 
 class SubprocessTimeoutError(RuntimeError):
   """An error that indicates there is at least one subprocess timing out.
@@ -672,6 +691,19 @@ class SubprocessTimeoutError(RuntimeError):
     self.mpr_result = mpr_result
 
 
+class UnexpectedSubprocessExitError(RuntimeError):
+  """An error indicating there is at least one subprocess with unexpected exit.
+
+  When this is raised, a `MultiProcessRunnerResult` object can be retrieved by
+  `UnexpectedSubprocessExitError`'s mpr_result attribute. See
+  `MultiProcessRunner.join()` for more information.
+  """
+
+  def __init__(self, msg, mpr_result):
+    super(UnexpectedSubprocessExitError, self).__init__(msg)
+    self.mpr_result = mpr_result
+
+
 def _set_tf_config(task_type, task_id, cluster_spec, rpc_layer=None):
   """Set TF_CONFIG environment variable."""
   tf_config_dict = {
diff --git a/tensorflow/python/distribute/multi_process_runner_test.py b/tensorflow/python/distribute/multi_process_runner_test.py
index aeba43b6b7c..6194ac527d5 100644
--- a/tensorflow/python/distribute/multi_process_runner_test.py
+++ b/tensorflow/python/distribute/multi_process_runner_test.py
@@ -18,6 +18,7 @@ from __future__ import absolute_import
 from __future__ import division
 from __future__ import print_function
 
+import ctypes
 import json
 import os
 import threading
@@ -324,6 +325,40 @@ class MultiProcessRunnerTest(test.TestCase):
             any('(logging) {}-0, i: {}'.format(job, iteration) in line
                 for line in list_to_assert))
 
+  def test_seg_fault_raises_error(self):
+
+    def proc_func_expected_to_seg_fault():
+      ctypes.string_at(0)  # Intentionally made seg fault.
+
+    with self.assertRaises(
+        multi_process_runner.UnexpectedSubprocessExitError) as cm:
+      multi_process_runner.run(
+          proc_func_expected_to_seg_fault,
+          multi_worker_test_base.create_cluster_spec(num_workers=1),
+          list_stdout=True)
+    self.assertIn('Missing status(es) from 1 subprocess(es).',
+                  str(cm.exception))
+    list_to_assert = cm.exception.mpr_result.stdout
+    self.assertTrue(any('SIGSEGV' in line for line in list_to_assert))
+
+  def test_seg_fault_in_chief_raises_error(self):
+
+    def proc_func_expected_to_seg_fault():
+      if multi_worker_test_base.get_task_type() == 'worker':
+        time.sleep(10000)
+      ctypes.string_at(0)  # Intentionally made seg fault.
+
+    with self.assertRaises(
+        multi_process_runner.UnexpectedSubprocessExitError) as cm:
+      multi_process_runner.run(
+          proc_func_expected_to_seg_fault,
+          multi_worker_test_base.create_cluster_spec(
+              has_chief=True, num_workers=1),
+          list_stdout=True)
+    self.assertIn('Subprocess chief-0 exited with exit code',
+                  str(cm.exception))
+    list_to_assert = cm.exception.mpr_result.stdout
+    self.assertTrue(any('SIGSEGV' in line for line in list_to_assert))
 
 if __name__ == '__main__':
   multi_process_runner.test_main()

From 0c227aed65e62f741a88c9915923d262710fc8c9 Mon Sep 17 00:00:00 2001
From: Scott Wegner 
Date: Mon, 22 Jun 2020 14:26:34 -0700
Subject: [PATCH 466/867] Special case tfnp ndarrays in some places, and minor
 improvements to tfnp array functions

PiperOrigin-RevId: 317735476
Change-Id: I58123c5fb64257975e220df4e57f2a1f74856da7
---
 tensorflow/python/eager/BUILD                  |  1 -
 tensorflow/python/eager/function.py            | 18 ++----------------
 tensorflow/python/ops/numpy_ops/np_arrays.py   | 17 ++---------------
 .../python/ops/numpy_ops/np_interop_test.py    |  4 +---
 4 files changed, 5 insertions(+), 35 deletions(-)

diff --git a/tensorflow/python/eager/BUILD b/tensorflow/python/eager/BUILD
index f51bd97e488..af5f3d16408 100644
--- a/tensorflow/python/eager/BUILD
+++ b/tensorflow/python/eager/BUILD
@@ -560,7 +560,6 @@ py_library(
         "//tensorflow/python:graph_to_function_def",
         "//tensorflow/python:pywrap_tf_session",
         "//tensorflow/python:util",
-        "//tensorflow/python/ops/numpy_ops:numpy",
         "//third_party/py/numpy",
         "@six_archive//:six",
     ],
diff --git a/tensorflow/python/eager/function.py b/tensorflow/python/eager/function.py
index 7b235749533..4a08e66b441 100644
--- a/tensorflow/python/eager/function.py
+++ b/tensorflow/python/eager/function.py
@@ -81,9 +81,6 @@ from tensorflow.python.util import tf_inspect
 ag_ctx = lazy_loader.LazyLoader(
     "ag_ctx", globals(),
     "tensorflow.python.autograph.core.ag_ctx")
-np_arrays = lazy_loader.LazyLoader(
-    "np_arrays", globals(),
-    "tensorflow.python.ops.numpy_ops.np_arrays")
 
 
 FORWARD_FUNCTION_ATTRIBUTE_NAME = "forward_function_name"
@@ -1487,11 +1484,6 @@ class ConcreteFunction(object):
     self._func_graph = func_graph
     self._captured_inputs = self._func_graph.external_captures
     self._captured_closures = self._func_graph.deferred_external_captures
-    structured_outputs = self._func_graph.structured_outputs
-    self._ndarrays_list = (
-        isinstance(structured_outputs, (list, tuple)) and
-        all([isinstance(o, np_arrays.ndarray) for o in structured_outputs]))
-    self._ndarray_singleton = isinstance(structured_outputs, np_arrays.ndarray)
 
     # function_spec defines the structured signature.
     self._set_function_spec(function_spec)
@@ -2158,15 +2150,9 @@ class ConcreteFunction(object):
     if self._func_graph.structured_outputs is None:
       return result
 
-    if result:
-      if self._ndarrays_list:
-        return [np_arrays.tensor_to_ndarray(o) for o in result]
-      elif self._ndarray_singleton:
-        return np_arrays.tensor_to_ndarray(result[0])
-
     # Replace outputs with results, skipping over any 'None' values.
-    outputs_list = nest.flatten(
-        self._func_graph.structured_outputs, expand_composites=True)
+    outputs_list = nest.flatten(self._func_graph.structured_outputs,
+                                expand_composites=True)
     j = 0
     for i, o in enumerate(outputs_list):
       if o is not None:
diff --git a/tensorflow/python/ops/numpy_ops/np_arrays.py b/tensorflow/python/ops/numpy_ops/np_arrays.py
index fd26318bea9..40ccfa9dc50 100644
--- a/tensorflow/python/ops/numpy_ops/np_arrays.py
+++ b/tensorflow/python/ops/numpy_ops/np_arrays.py
@@ -149,23 +149,10 @@ class ndarray(composite_tensor.CompositeTensor):  # pylint: disable=invalid-name
     if dtype and dtype != buffer.dtype:
       buffer = array_ops.bitcast(buffer, dtype)
     self._data = buffer
-    self._type_spec_internal = None
-
-  @classmethod
-  def from_tensor(cls, tensor):
-    o = cls.__new__(cls, None)
-    # pylint: disable=protected-access
-    o._data = tensor
-    o._type_spec_internal = None
-    # pylint: enable=protected-access
-    return o
 
   @property
   def _type_spec(self):
-    if self._type_spec_internal is None:
-      self._type_spec_internal = NdarraySpec(
-          type_spec.type_spec_from_value(self._data))
-    return self._type_spec_internal
+    return NdarraySpec(type_spec.type_spec_from_value(self._data))
 
   @property
   def data(self):
@@ -312,7 +299,7 @@ class ndarray(composite_tensor.CompositeTensor):  # pylint: disable=invalid-name
 
 
 def tensor_to_ndarray(tensor):
-  return ndarray.from_tensor(tensor)
+  return ndarray(tensor._shape_tuple(), dtype=tensor.dtype, buffer=tensor)  # pylint: disable=protected-access
 
 
 def ndarray_to_tensor(arr, dtype=None, name=None, as_ref=False):
diff --git a/tensorflow/python/ops/numpy_ops/np_interop_test.py b/tensorflow/python/ops/numpy_ops/np_interop_test.py
index 5c7560b12d9..9580b787202 100644
--- a/tensorflow/python/ops/numpy_ops/np_interop_test.py
+++ b/tensorflow/python/ops/numpy_ops/np_interop_test.py
@@ -64,9 +64,7 @@ class InteropTest(test.TestCase):
 
     dx, dy = t.gradient([xx, yy], [x, y])
 
-    # # TODO(nareshmodi): Figure out a way to rewrap ndarray as tensors.
-    # self.assertIsInstance(dx, np_arrays.ndarray)
-    # self.assertIsInstance(dy, np_arrays.ndarray)
+    # TODO(nareshmodi): Gradient tape returns tensors. Is it possible to rewrap?
     self.assertAllClose(dx, 2.0)
     self.assertAllClose(dy, 3.0)
 

From 20572f9ed0f4c6b02b0226281809c63868773a0e Mon Sep 17 00:00:00 2001
From: Geoffrey Martin-Noble 
Date: Mon, 22 Jun 2020 14:27:36 -0700
Subject: [PATCH 467/867] Rename and alias MLIR targets to match CMake

These differ from the CMake library names somewhat arbitrarily, which just creates confusion.

PiperOrigin-RevId: 317735676
Change-Id: Ia9d3d1645467b1af72f917b01d1a3f6061bceb34
---
 third_party/mlir/BUILD | 117 +++++++++++++++++++++++++----------------
 1 file changed, 71 insertions(+), 46 deletions(-)

diff --git a/third_party/mlir/BUILD b/third_party/mlir/BUILD
index a0d10667d23..2be59fc44b1 100644
--- a/third_party/mlir/BUILD
+++ b/third_party/mlir/BUILD
@@ -282,7 +282,7 @@ cc_library(
     deps = [
         ":AVX512IncGen",
         ":IR",
-        ":SideEffects",
+        ":SideEffectInterfaces",
         ":VectorOps",
         "@llvm-project//llvm:Core",
         "@llvm-project//llvm:Support",
@@ -305,9 +305,9 @@ cc_library(
         ":IR",
         ":LLVMAVX512",
         ":LLVMDialect",
-        ":LLVMTransforms",
         ":Pass",
         ":StandardOps",
+        ":StandardToLLVM",
         ":Support",
         ":Transforms",
         ":VectorOps",
@@ -489,7 +489,7 @@ cc_library(
         ":EDSC",
         ":IR",
         ":LoopLikeInterface",
-        ":SideEffects",
+        ":SideEffectInterfaces",
         ":StandardOps",
         ":Support",
         "@llvm-project//llvm:Support",
@@ -571,7 +571,7 @@ gentbl(
 )
 
 cc_library(
-    name = "AffineToStandardTransforms",
+    name = "AffineToStandard",
     srcs = glob([
         "lib/Conversion/AffineToStandard/*.cpp",
         "lib/Conversion/AffineToStandard/*.h",
@@ -591,6 +591,11 @@ cc_library(
     ],
 )
 
+alias(
+    name = "AffineToStandardTransforms",
+    actual = "AffineToStandard",
+)
+
 # SDBM dialect only contains attribute components that can be constructed given
 # a dialect object, so whenever it is used it must also be registered. Therefore
 # we don't split out the registration library for it.
@@ -631,7 +636,7 @@ cc_library(
         ":IR",
         ":LoopLikeInterface",
         ":SCFIncGen",
-        ":SideEffects",
+        ":SideEffectInterfaces",
         ":StandardOps",
         ":Support",
         "@llvm-project//llvm:Support",
@@ -719,7 +724,7 @@ cc_library(
         ":InferTypeOpInterface",
         ":MLIRShapeCanonicalizationIncGen",
         ":ShapeOpsIncGen",
-        ":SideEffects",
+        ":SideEffectInterfaces",
         ":Support",
         "@llvm-project//llvm:Support",
     ],
@@ -833,7 +838,7 @@ cc_library(
         ":ControlFlowInterfaces",
         ":EDSC",
         ":IR",
-        ":SideEffects",
+        ":SideEffectInterfaces",
         ":StandardOpsIncGen",
         ":Support",
         ":ViewLikeInterface",
@@ -895,7 +900,7 @@ cc_library(
         ":DialectUtils",
         ":EDSC",
         ":IR",
-        ":SideEffects",
+        ":SideEffectInterfaces",
         ":StandardOps",
         ":Support",
         ":VectorOpsIncGen",
@@ -1070,7 +1075,7 @@ cc_library(
         ":ControlFlowInterfaces",
         ":IR",
         ":LLVMOpsIncGen",
-        ":SideEffects",
+        ":SideEffectInterfaces",
         ":Support",
         "@llvm-project//llvm:AsmParser",
         "@llvm-project//llvm:BitReader",
@@ -1193,7 +1198,7 @@ cc_library(
         ":GPUOpsIncGen",
         ":IR",
         ":LLVMDialect",
-        ":SideEffects",
+        ":SideEffectInterfaces",
         ":StandardOps",
         ":Support",
     ],
@@ -1271,8 +1276,8 @@ cc_library(
         ":GPUDialect",
         ":IR",
         ":LLVMDialect",
-        ":LLVMTransforms",
         ":StandardOps",
+        ":StandardToLLVM",
         "@llvm-project//llvm:Support",
     ],
 )
@@ -1311,9 +1316,9 @@ cc_library(
         ":GPUToNVVMGen",
         ":GPUTransforms",
         ":IR",
-        ":LLVMTransforms",
         ":NVVMDialect",
         ":Pass",
+        ":StandardToLLVM",
         ":Transforms",
         "@llvm-project//llvm:Support",
     ],
@@ -1333,10 +1338,10 @@ cc_library(
         ":ConversionPassIncGen",
         ":GPUDialect",
         ":LLVMDialect",
-        ":LLVMTransforms",
         ":Pass",
         ":ROCDLDialect",
         ":StandardOps",
+        ":StandardToLLVM",
         ":Transforms",
         ":VectorOps",
     ],
@@ -1375,9 +1380,9 @@ cc_library(
         ":GPUDialect",
         ":GPUToROCDLTGen",
         ":GPUTransforms",
-        ":LLVMTransforms",
         ":Pass",
         ":ROCDLDialect",
+        ":StandardToLLVM",
         ":Transforms",
         ":VectorOps",
         ":VectorToLLVM",
@@ -1475,7 +1480,7 @@ cc_library(
         ":SCFDialect",
         ":SPIRVDialect",
         ":SPIRVLowering",
-        ":StandardToSPIRVConversions",
+        ":StandardToSPIRVTransforms",
         ":Support",
         ":Transforms",
     ],
@@ -1496,10 +1501,10 @@ cc_library(
         ":ConversionPassIncGen",
         ":IR",
         ":LLVMDialect",
-        ":LLVMTransforms",
         ":Pass",
         ":SPIRVDialect",
         ":StandardOps",
+        ":StandardToLLVM",
         ":Support",
         ":Transforms",
     ],
@@ -1574,7 +1579,7 @@ cc_library(
         ":IR",
         ":LLVMDialect",
         ":NVVMOpsIncGen",
-        ":SideEffects",
+        ":SideEffectInterfaces",
         ":StandardOps",
         ":Support",
         "@llvm-project//llvm:AsmParser",
@@ -1646,7 +1651,7 @@ cc_library(
         ":IR",
         ":LLVMDialect",
         ":ROCDLOpsIncGen",
-        ":SideEffects",
+        ":SideEffectInterfaces",
         ":StandardOps",
         ":Support",
         "@llvm-project//llvm:AsmParser",
@@ -1894,7 +1899,7 @@ cc_library(
         ":SPIRVOpsIncGen",
         ":SPIRVSerializationGen",
         ":SPIRVTargetAndABIStructGen",
-        ":SideEffects",
+        ":SideEffectInterfaces",
         ":Support",
         ":Transforms",
         "@llvm-project//llvm:Support",
@@ -1947,7 +1952,7 @@ cc_library(
 )
 
 cc_library(
-    name = "StandardToSPIRVConversions",
+    name = "StandardToSPIRVTransforms",
     srcs = glob([
         "lib/Conversion/StandardToSPIRV/*.cpp",
         "lib/Conversion/StandardToSPIRV/*.h",
@@ -1972,6 +1977,11 @@ cc_library(
     ],
 )
 
+alias(
+    name = "StandardToSPIRVConversions",
+    actual = "StandardToSPIRVTransforms",
+)
+
 cc_library(
     name = "SPIRVSerialization",
     srcs = glob(
@@ -2033,7 +2043,7 @@ cc_library(
         ":IR",
         ":LoopLikeInterface",
         ":SCFDialect",
-        ":SideEffects",
+        ":SideEffectInterfaces",
         ":StandardOps",
         ":Support",
         "@llvm-project//llvm:Support",
@@ -2152,7 +2162,7 @@ cc_library(
         ":LoopLikeInterface",
         ":Pass",
         ":SCFDialect",
-        ":SideEffects",
+        ":SideEffectInterfaces",
         ":StandardOps",
         ":Support",
         ":TransformUtils",
@@ -2182,7 +2192,7 @@ cc_library(
     includes = ["include"],
     deps = [
         ":Affine",
-        ":AffineToStandardTransforms",
+        ":AffineToStandard",
         ":ConversionPassIncGen",
         ":GPUDialect",
         ":GPUTransforms",
@@ -2222,7 +2232,7 @@ cc_library(
 )
 
 cc_library(
-    name = "CFGTransforms",
+    name = "SCFToStandard",
     srcs = [
         "lib/Conversion/PassDetail.h",
         "lib/Conversion/SCFToStandard/SCFToStandard.cpp",
@@ -2244,8 +2254,13 @@ cc_library(
     ],
 )
 
+alias(
+    name = "CFGTransforms",
+    actual = "SCFToStandard",
+)
+
 cc_library(
-    name = "LLVMTransforms",
+    name = "StandardToLLVM",
     srcs = [
         "lib/Conversion/PassDetail.h",
         "lib/Conversion/StandardToLLVM/StandardToLLVM.cpp",
@@ -2269,6 +2284,11 @@ cc_library(
     ],
 )
 
+alias(
+    name = "LLVMTransforms",
+    actual = "StandardToLLVM",
+)
+
 gentbl(
     name = "CallOpInterfacesIncGen",
     strip_include_prefix = "include",
@@ -2401,7 +2421,7 @@ gentbl(
 )
 
 cc_library(
-    name = "SideEffects",
+    name = "SideEffectInterfaces",
     srcs = [
         "lib/Interfaces/SideEffectInterfaces.cpp",
     ],
@@ -2417,6 +2437,11 @@ cc_library(
     ],
 )
 
+alias(
+    name = "SideEffects",
+    actual = "SideEffectInterfaces",
+)
+
 cc_library(
     name = "Analysis",
     srcs = glob(
@@ -2627,7 +2652,6 @@ cc_library(
         ":GPUTransforms",
         ":IR",
         ":LLVMDialect",
-        ":LLVMTransforms",
         ":LinalgToLLVM",
         ":LinalgToSPIRV",
         ":LinalgToStandard",
@@ -2639,7 +2663,8 @@ cc_library(
         ":ShapeToStandard",
         ":ShapeTransforms",
         ":StandardOpsTransforms",
-        ":StandardToSPIRVConversions",
+        ":StandardToLLVM",
+        ":StandardToSPIRVTransforms",
         ":Support",
         ":Transforms",
         ":VectorToLLVM",
@@ -2699,7 +2724,6 @@ cc_library(
         ":Affine",
         ":AffinePassIncGen",
         ":AffineTransforms",
-        ":CFGTransforms",
         ":ConversionPassIncGen",
         ":GPUDialect",
         ":GPUPassIncGen",
@@ -2714,7 +2738,6 @@ cc_library(
         ":LLVMDialect",
         ":LLVMIRTransforms",
         ":LLVMPassIncGen",
-        ":LLVMTransforms",
         ":LinalgOps",
         ":LinalgPassIncGen",
         ":LinalgToLLVM",
@@ -2729,6 +2752,7 @@ cc_library(
         ":ROCDLDialect",
         ":SCFDialect",
         ":SCFToGPUPass",
+        ":SCFToStandard",
         ":SCFTransforms",
         ":SDBM",
         ":SPIRVDialect",
@@ -2743,7 +2767,8 @@ cc_library(
         ":StandardOps",
         ":StandardOpsTransforms",
         ":StandardOpsTransformsPassIncGen",
-        ":StandardToSPIRVConversions",
+        ":StandardToLLVM",
+        ":StandardToSPIRVTransforms",
         ":Transforms",
         ":TransformsPassIncGen",
         ":VectorOps",
@@ -2809,13 +2834,13 @@ cc_library(
     includes = ["include"],
     deps = [
         ":AllPassesAndDialectsNoRegistration",
-        ":CFGTransforms",
         ":ExecutionEngine",
         ":ExecutionEngineUtils",
         ":IR",
         ":LLVMDialect",
         ":Parser",
         ":Pass",
+        ":SCFToStandard",
         ":Support",
         "@llvm-project//llvm:Core",
         "@llvm-project//llvm:OrcJIT",
@@ -2885,7 +2910,7 @@ cc_library(
         ":IR",
         ":Pass",
         ":SPIRVDialect",
-        ":SideEffects",
+        ":SideEffectInterfaces",
         ":StandardOps",
         ":Support",
         "@llvm-project//llvm:Support",
@@ -2918,10 +2943,10 @@ cc_binary(
         ":GPUTransforms",
         ":IR",
         ":LLVMDialect",
-        ":LLVMTransforms",
         ":MlirJitRunner",
         ":NVVMDialect",
         ":Pass",
+        ":StandardToLLVM",
         ":TargetNVVMIR",
         ":Transforms",
         "//devtools/build/runtime:get_runfiles_dir",
@@ -2945,11 +2970,11 @@ cc_binary(
         ":GPUToSPIRVTransforms",
         ":GPUToVulkanTransforms",
         ":GPUTransforms",
-        ":LLVMTransforms",
         ":MlirJitRunner",
         ":Pass",
         ":SPIRVDialect",
-        ":StandardToSPIRVConversions",
+        ":StandardToLLVM",
+        ":StandardToSPIRVTransforms",
         "@llvm-project//llvm:Support",
     ],
 )
@@ -3159,7 +3184,7 @@ cc_library(
         ":Pass",
         ":QuantOpsIncGen",
         ":QuantPassIncGen",
-        ":SideEffects",
+        ":SideEffectInterfaces",
         ":StandardOps",
         "@llvm-project//llvm:Support",
     ],
@@ -3294,18 +3319,18 @@ cc_library(
     ]),
     includes = ["include"],
     deps = [
-        ":AffineToStandardTransforms",
+        ":AffineToStandard",
         ":Analysis",
-        ":CFGTransforms",
         ":ConversionPassIncGen",
         ":EDSC",
         ":IR",
         ":LLVMDialect",
-        ":LLVMTransforms",
         ":LinalgOps",
         ":LinalgTransforms",
         ":Pass",
+        ":SCFToStandard",
         ":StandardOps",
+        ":StandardToLLVM",
         ":Support",
         ":Transforms",
         ":VectorToLLVM",
@@ -3385,7 +3410,7 @@ cc_library(
         ":LinalgOpsIncGen",
         ":LinalgStructuredOpsIncGen",
         ":Parser",
-        ":SideEffects",
+        ":SideEffectInterfaces",
         ":StandardOps",
         ":Support",
         ":ViewLikeInterface",
@@ -3431,21 +3456,21 @@ cc_library(
     includes = ["include"],
     deps = [
         ":Affine",
-        ":AffineToStandardTransforms",
+        ":AffineToStandard",
         ":Analysis",
-        ":CFGTransforms",
         ":DialectUtils",
         ":EDSC",
         ":IR",
         ":LLVMDialect",
-        ":LLVMTransforms",
         ":LinalgOps",
         ":LinalgPassIncGen",
         ":LinalgStructuredOpsIncGen",
         ":Pass",
         ":SCFDialect",
+        ":SCFToStandard",
         ":SCFTransforms",
         ":StandardOps",
+        ":StandardToLLVM",
         ":Support",
         ":TransformUtils",
         ":Transforms",
@@ -3537,9 +3562,9 @@ cc_library(
         ":EDSC",
         ":IR",
         ":LLVMDialect",
-        ":LLVMTransforms",
         ":Pass",
         ":StandardOps",
+        ":StandardToLLVM",
         ":Support",
         ":Transforms",
         ":VectorOps",
@@ -3564,10 +3589,10 @@ cc_library(
         ":EDSC",
         ":IR",
         ":LLVMDialect",
-        ":LLVMTransforms",
         ":Pass",
         ":SCFDialect",
         ":StandardOps",
+        ":StandardToLLVM",
         ":Support",
         ":Transforms",
         ":VectorOps",

From e6ada6a6b4b2a98cd8a8e361ae7e4c6ae1d09645 Mon Sep 17 00:00:00 2001
From: Smit Hinsu 
Date: Mon, 22 Jun 2020 14:28:48 -0700
Subject: [PATCH 468/867] Add unit test for Prelu in XNNPACK delegate

- Add PreluTester class and unit test for XNNPACK-delegated Prelu operator
- Relax restrictions on the number of input/output dimensions in delegated
  Prelu operators

PiperOrigin-RevId: 317735911
Change-Id: Iddf727f5f916b142412a1be44efa1f367dc31d49
---
 tensorflow/lite/delegates/xnnpack/BUILD       |  32 -
 .../delegates/xnnpack/leaky_relu_tester.h     |   1 +
 .../lite/delegates/xnnpack/prelu_test.cc      | 565 ------------------
 .../lite/delegates/xnnpack/prelu_tester.cc    | 237 --------
 .../lite/delegates/xnnpack/prelu_tester.h     |  88 ---
 .../delegates/xnnpack/xnnpack_delegate.cc     |   6 +-
 6 files changed, 3 insertions(+), 926 deletions(-)
 delete mode 100644 tensorflow/lite/delegates/xnnpack/prelu_test.cc
 delete mode 100644 tensorflow/lite/delegates/xnnpack/prelu_tester.cc
 delete mode 100644 tensorflow/lite/delegates/xnnpack/prelu_tester.h

diff --git a/tensorflow/lite/delegates/xnnpack/BUILD b/tensorflow/lite/delegates/xnnpack/BUILD
index e0d3d39f719..eaf7d8f6f03 100644
--- a/tensorflow/lite/delegates/xnnpack/BUILD
+++ b/tensorflow/lite/delegates/xnnpack/BUILD
@@ -180,23 +180,6 @@ cc_library(
     ],
 )
 
-cc_library(
-    name = "prelu_tester",
-    testonly = 1,
-    srcs = ["prelu_tester.cc"],
-    hdrs = ["prelu_tester.h"],
-    deps = [
-        "//tensorflow/lite:framework",
-        "//tensorflow/lite:schema_fbs_version",
-        "//tensorflow/lite/c:common",
-        "//tensorflow/lite/kernels:builtin_ops",
-        "//tensorflow/lite/schema:schema_fbs",
-        "@FP16",
-        "@com_google_googletest//:gtest",
-        "@flatbuffers",
-    ],
-)
-
 cc_library(
     name = "reduce_tester",
     testonly = 1,
@@ -544,21 +527,6 @@ cc_test(
     ],
 )
 
-cc_test(
-    name = "prelu_test",
-    srcs = ["prelu_test.cc"],
-    linkopts = select({
-        "//tensorflow:emscripten": EMSCRIPTEN_LINKOPTS,
-        "//conditions:default": [],
-    }),
-    deps = [
-        ":prelu_tester",
-        ":test_main",
-        ":xnnpack_delegate_test_mode",
-        "@com_google_googletest//:gtest",
-    ],
-)
-
 cc_test(
     name = "relu_test",
     srcs = ["relu_test.cc"],
diff --git a/tensorflow/lite/delegates/xnnpack/leaky_relu_tester.h b/tensorflow/lite/delegates/xnnpack/leaky_relu_tester.h
index 191dc938e89..f1d9efd7209 100644
--- a/tensorflow/lite/delegates/xnnpack/leaky_relu_tester.h
+++ b/tensorflow/lite/delegates/xnnpack/leaky_relu_tester.h
@@ -21,6 +21,7 @@ limitations under the License.
 
 #include 
 #include "tensorflow/lite/c/common.h"
+#include "tensorflow/lite/schema/schema_generated.h"
 
 namespace tflite {
 namespace xnnpack {
diff --git a/tensorflow/lite/delegates/xnnpack/prelu_test.cc b/tensorflow/lite/delegates/xnnpack/prelu_test.cc
deleted file mode 100644
index ae40c032f32..00000000000
--- a/tensorflow/lite/delegates/xnnpack/prelu_test.cc
+++ /dev/null
@@ -1,565 +0,0 @@
-/* Copyright 2020 The TensorFlow Authors. All Rights Reserved.
-
-Licensed under the Apache License, Version 2.0 (the "License");
-you may not use this file except in compliance with the License.
-You may obtain a copy of the License at
-
-    http://www.apache.org/licenses/LICENSE-2.0
-
-Unless required by applicable law or agreed to in writing, software
-distributed under the License is distributed on an "AS IS" BASIS,
-WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-See the License for the specific language governing permissions and
-limitations under the License.
-==============================================================================*/
-
-#include 
-#include 
-#include 
-#include 
-
-#include 
-#include "tensorflow/lite/delegates/xnnpack/prelu_tester.h"
-#include "tensorflow/lite/delegates/xnnpack/xnnpack_delegate.h"
-
-namespace tflite {
-namespace xnnpack {
-
-TEST(Prelu, DISABLED_4DBy4D) {
-  std::unique_ptr
-      xnnpack_delegate(TfLiteXNNPackDelegateCreate(nullptr),
-                       TfLiteXNNPackDelegateDelete);
-
-  std::random_device random_device;
-  auto rng = std::mt19937(random_device());
-  auto shape_rng =
-      std::bind(std::uniform_int_distribution(2, 5), std::ref(rng));
-  const auto batch = shape_rng();
-  const auto height = shape_rng();
-  const auto width = shape_rng();
-  const auto channels = shape_rng();
-
-  PreluTester()
-      .InputShape({batch, height, width, channels})
-      .SlopeShape({batch, height, width, channels})
-      .Test(xnnpack_delegate.get());
-}
-
-TEST(Prelu, 4DBy4DBroadcastChannels) {
-  std::unique_ptr
-      xnnpack_delegate(TfLiteXNNPackDelegateCreate(nullptr),
-                       TfLiteXNNPackDelegateDelete);
-
-  std::random_device random_device;
-  auto rng = std::mt19937(random_device());
-  auto shape_rng =
-      std::bind(std::uniform_int_distribution(2, 5), std::ref(rng));
-  const auto batch = shape_rng();
-  const auto height = shape_rng();
-  const auto width = shape_rng();
-  const auto channels = shape_rng();
-
-  PreluTester()
-      .InputShape({batch, height, width, channels})
-      .SlopeShape({1, 1, 1, channels})
-      .Test(xnnpack_delegate.get());
-}
-
-TEST(Prelu, DISABLED_4DBy4DBroadcastWidth) {
-  std::unique_ptr
-      xnnpack_delegate(TfLiteXNNPackDelegateCreate(nullptr),
-                       TfLiteXNNPackDelegateDelete);
-
-  std::random_device random_device;
-  auto rng = std::mt19937(random_device());
-  auto shape_rng =
-      std::bind(std::uniform_int_distribution(2, 5), std::ref(rng));
-  const auto batch = shape_rng();
-  const auto height = shape_rng();
-  const auto width = shape_rng();
-  const auto channels = shape_rng();
-
-  PreluTester()
-      .InputShape({batch, height, width, channels})
-      .SlopeShape({1, 1, width, 1})
-      .Test(xnnpack_delegate.get());
-}
-
-TEST(Prelu, DISABLED_4DBy4DBroadcastHeight) {
-  std::unique_ptr
-      xnnpack_delegate(TfLiteXNNPackDelegateCreate(nullptr),
-                       TfLiteXNNPackDelegateDelete);
-
-  std::random_device random_device;
-  auto rng = std::mt19937(random_device());
-  auto shape_rng =
-      std::bind(std::uniform_int_distribution(2, 5), std::ref(rng));
-  const auto batch = shape_rng();
-  const auto height = shape_rng();
-  const auto width = shape_rng();
-  const auto channels = shape_rng();
-
-  PreluTester()
-      .InputShape({batch, height, width, channels})
-      .SlopeShape({1, height, 1, 1})
-      .Test(xnnpack_delegate.get());
-}
-
-TEST(Prelu, DISABLED_4DBy4DBroadcastBatch) {
-  std::unique_ptr
-      xnnpack_delegate(TfLiteXNNPackDelegateCreate(nullptr),
-                       TfLiteXNNPackDelegateDelete);
-
-  std::random_device random_device;
-  auto rng = std::mt19937(random_device());
-  auto shape_rng =
-      std::bind(std::uniform_int_distribution(2, 5), std::ref(rng));
-  const auto batch = shape_rng();
-  const auto height = shape_rng();
-  const auto width = shape_rng();
-  const auto channels = shape_rng();
-
-  PreluTester()
-      .InputShape({batch, height, width, channels})
-      .SlopeShape({batch, 1, 1, 1})
-      .Test(xnnpack_delegate.get());
-}
-
-TEST(Prelu, DISABLED_4DBy4DBroadcastHeightWidthChannels) {
-  std::unique_ptr
-      xnnpack_delegate(TfLiteXNNPackDelegateCreate(nullptr),
-                       TfLiteXNNPackDelegateDelete);
-
-  std::random_device random_device;
-  auto rng = std::mt19937(random_device());
-  auto shape_rng =
-      std::bind(std::uniform_int_distribution(2, 5), std::ref(rng));
-  const auto batch = shape_rng();
-  const auto height = shape_rng();
-  const auto width = shape_rng();
-  const auto channels = shape_rng();
-
-  PreluTester()
-      .InputShape({batch, height, width, channels})
-      .SlopeShape({1, height, width, channels})
-      .Test(xnnpack_delegate.get());
-}
-
-TEST(Prelu, DISABLED_4DBy3D) {
-  std::unique_ptr
-      xnnpack_delegate(TfLiteXNNPackDelegateCreate(nullptr),
-                       TfLiteXNNPackDelegateDelete);
-
-  std::random_device random_device;
-  auto rng = std::mt19937(random_device());
-  auto shape_rng =
-      std::bind(std::uniform_int_distribution(2, 5), std::ref(rng));
-  const auto batch = shape_rng();
-  const auto height = shape_rng();
-  const auto width = shape_rng();
-  const auto channels = shape_rng();
-
-  PreluTester()
-      .InputShape({batch, height, width, channels})
-      .SlopeShape({height, width, channels})
-      .Test(xnnpack_delegate.get());
-}
-
-TEST(Prelu, DISABLED_4DBy2D) {
-  std::unique_ptr
-      xnnpack_delegate(TfLiteXNNPackDelegateCreate(nullptr),
-                       TfLiteXNNPackDelegateDelete);
-
-  std::random_device random_device;
-  auto rng = std::mt19937(random_device());
-  auto shape_rng =
-      std::bind(std::uniform_int_distribution(2, 5), std::ref(rng));
-  const auto batch = shape_rng();
-  const auto height = shape_rng();
-  const auto width = shape_rng();
-  const auto channels = shape_rng();
-
-  PreluTester()
-      .InputShape({batch, height, width, channels})
-      .SlopeShape({width, channels})
-      .Test(xnnpack_delegate.get());
-}
-
-TEST(Prelu, 4DBy1D) {
-  std::unique_ptr
-      xnnpack_delegate(TfLiteXNNPackDelegateCreate(nullptr),
-                       TfLiteXNNPackDelegateDelete);
-
-  std::random_device random_device;
-  auto rng = std::mt19937(random_device());
-  auto shape_rng =
-      std::bind(std::uniform_int_distribution(2, 5), std::ref(rng));
-  const auto batch = shape_rng();
-  const auto height = shape_rng();
-  const auto width = shape_rng();
-  const auto channels = shape_rng();
-
-  PreluTester()
-      .InputShape({batch, height, width, channels})
-      .SlopeShape({channels})
-      .Test(xnnpack_delegate.get());
-}
-
-TEST(Prelu, DISABLED_4DBy0D) {
-  std::unique_ptr
-      xnnpack_delegate(TfLiteXNNPackDelegateCreate(nullptr),
-                       TfLiteXNNPackDelegateDelete);
-
-  std::random_device random_device;
-  auto rng = std::mt19937(random_device());
-  auto shape_rng =
-      std::bind(std::uniform_int_distribution(2, 5), std::ref(rng));
-  const auto batch = shape_rng();
-  const auto height = shape_rng();
-  const auto width = shape_rng();
-  const auto channels = shape_rng();
-
-  PreluTester()
-      .InputShape({batch, height, width, channels})
-      .SlopeShape({})
-      .Test(xnnpack_delegate.get());
-}
-
-TEST(Prelu, DISABLED_3DBy3D) {
-  std::unique_ptr
-      xnnpack_delegate(TfLiteXNNPackDelegateCreate(nullptr),
-                       TfLiteXNNPackDelegateDelete);
-
-  std::random_device random_device;
-  auto rng = std::mt19937(random_device());
-  auto shape_rng =
-      std::bind(std::uniform_int_distribution(2, 5), std::ref(rng));
-  const auto batch = shape_rng();
-  const auto width = shape_rng();
-  const auto channels = shape_rng();
-
-  PreluTester()
-      .InputShape({batch, width, channels})
-      .SlopeShape({batch, width, channels})
-      .Test(xnnpack_delegate.get());
-}
-
-TEST(Prelu, 3DBy3DBroadcastChannels) {
-  std::unique_ptr
-      xnnpack_delegate(TfLiteXNNPackDelegateCreate(nullptr),
-                       TfLiteXNNPackDelegateDelete);
-
-  std::random_device random_device;
-  auto rng = std::mt19937(random_device());
-  auto shape_rng =
-      std::bind(std::uniform_int_distribution(2, 5), std::ref(rng));
-  const auto batch = shape_rng();
-  const auto width = shape_rng();
-  const auto channels = shape_rng();
-
-  PreluTester()
-      .InputShape({batch, width, channels})
-      .SlopeShape({1, 1, channels})
-      .Test(xnnpack_delegate.get());
-}
-
-TEST(Prelu, DISABLED_3DBy3DBroadcastWidth) {
-  std::unique_ptr
-      xnnpack_delegate(TfLiteXNNPackDelegateCreate(nullptr),
-                       TfLiteXNNPackDelegateDelete);
-
-  std::random_device random_device;
-  auto rng = std::mt19937(random_device());
-  auto shape_rng =
-      std::bind(std::uniform_int_distribution(2, 5), std::ref(rng));
-  const auto batch = shape_rng();
-  const auto width = shape_rng();
-  const auto channels = shape_rng();
-
-  PreluTester()
-      .InputShape({batch, width, channels})
-      .SlopeShape({1, width, 1})
-      .Test(xnnpack_delegate.get());
-}
-
-TEST(Prelu, DISABLED_3DBy3DBroadcastBatch) {
-  std::unique_ptr
-      xnnpack_delegate(TfLiteXNNPackDelegateCreate(nullptr),
-                       TfLiteXNNPackDelegateDelete);
-
-  std::random_device random_device;
-  auto rng = std::mt19937(random_device());
-  auto shape_rng =
-      std::bind(std::uniform_int_distribution(2, 5), std::ref(rng));
-  const auto batch = shape_rng();
-  const auto width = shape_rng();
-  const auto channels = shape_rng();
-
-  PreluTester()
-      .InputShape({batch, width, channels})
-      .SlopeShape({batch, 1, 1})
-      .Test(xnnpack_delegate.get());
-}
-
-TEST(Prelu, DISABLED_3DBy3DBroadcastWidthChannels) {
-  std::unique_ptr
-      xnnpack_delegate(TfLiteXNNPackDelegateCreate(nullptr),
-                       TfLiteXNNPackDelegateDelete);
-
-  std::random_device random_device;
-  auto rng = std::mt19937(random_device());
-  auto shape_rng =
-      std::bind(std::uniform_int_distribution(2, 5), std::ref(rng));
-  const auto batch = shape_rng();
-  const auto width = shape_rng();
-  const auto channels = shape_rng();
-
-  PreluTester()
-      .InputShape({batch, width, channels})
-      .SlopeShape({1, width, channels})
-      .Test(xnnpack_delegate.get());
-}
-
-TEST(Prelu, DISABLED_3DBy2D) {
-  std::unique_ptr
-      xnnpack_delegate(TfLiteXNNPackDelegateCreate(nullptr),
-                       TfLiteXNNPackDelegateDelete);
-
-  std::random_device random_device;
-  auto rng = std::mt19937(random_device());
-  auto shape_rng =
-      std::bind(std::uniform_int_distribution(2, 5), std::ref(rng));
-  const auto batch = shape_rng();
-  const auto width = shape_rng();
-  const auto channels = shape_rng();
-
-  PreluTester()
-      .InputShape({batch, width, channels})
-      .SlopeShape({width, channels})
-      .Test(xnnpack_delegate.get());
-}
-
-TEST(Prelu, 3DBy1D) {
-  std::unique_ptr
-      xnnpack_delegate(TfLiteXNNPackDelegateCreate(nullptr),
-                       TfLiteXNNPackDelegateDelete);
-
-  std::random_device random_device;
-  auto rng = std::mt19937(random_device());
-  auto shape_rng =
-      std::bind(std::uniform_int_distribution(2, 5), std::ref(rng));
-  const auto batch = shape_rng();
-  const auto width = shape_rng();
-  const auto channels = shape_rng();
-
-  PreluTester()
-      .InputShape({batch, width, channels})
-      .SlopeShape({channels})
-      .Test(xnnpack_delegate.get());
-}
-
-TEST(Prelu, DISABLED_3DBy0D) {
-  std::unique_ptr
-      xnnpack_delegate(TfLiteXNNPackDelegateCreate(nullptr),
-                       TfLiteXNNPackDelegateDelete);
-
-  std::random_device random_device;
-  auto rng = std::mt19937(random_device());
-  auto shape_rng =
-      std::bind(std::uniform_int_distribution(2, 5), std::ref(rng));
-  const auto batch = shape_rng();
-  const auto width = shape_rng();
-  const auto channels = shape_rng();
-
-  PreluTester()
-      .InputShape({batch, width, channels})
-      .SlopeShape({})
-      .Test(xnnpack_delegate.get());
-}
-
-TEST(Prelu, DISABLED_2DBy2D) {
-  std::unique_ptr
-      xnnpack_delegate(TfLiteXNNPackDelegateCreate(nullptr),
-                       TfLiteXNNPackDelegateDelete);
-
-  std::random_device random_device;
-  auto rng = std::mt19937(random_device());
-  auto shape_rng =
-      std::bind(std::uniform_int_distribution(2, 5), std::ref(rng));
-  const auto batch = shape_rng();
-  const auto channels = shape_rng();
-
-  PreluTester()
-      .InputShape({batch, channels})
-      .SlopeShape({batch, channels})
-      .Test(xnnpack_delegate.get());
-}
-
-TEST(Prelu, 2DBy2DBroadcastChannels) {
-  std::unique_ptr
-      xnnpack_delegate(TfLiteXNNPackDelegateCreate(nullptr),
-                       TfLiteXNNPackDelegateDelete);
-
-  std::random_device random_device;
-  auto rng = std::mt19937(random_device());
-  auto shape_rng =
-      std::bind(std::uniform_int_distribution(2, 5), std::ref(rng));
-  const auto batch = shape_rng();
-  const auto channels = shape_rng();
-
-  PreluTester()
-      .InputShape({batch, channels})
-      .SlopeShape({1, channels})
-      .Test(xnnpack_delegate.get());
-}
-
-TEST(Prelu, DISABLED_2DBy2DBroadcastBatch) {
-  std::unique_ptr
-      xnnpack_delegate(TfLiteXNNPackDelegateCreate(nullptr),
-                       TfLiteXNNPackDelegateDelete);
-
-  std::random_device random_device;
-  auto rng = std::mt19937(random_device());
-  auto shape_rng =
-      std::bind(std::uniform_int_distribution(2, 5), std::ref(rng));
-  const auto batch = shape_rng();
-  const auto channels = shape_rng();
-
-  PreluTester()
-      .InputShape({batch, channels})
-      .SlopeShape({batch, 1})
-      .Test(xnnpack_delegate.get());
-}
-
-TEST(Prelu, 2DBy1D) {
-  std::unique_ptr
-      xnnpack_delegate(TfLiteXNNPackDelegateCreate(nullptr),
-                       TfLiteXNNPackDelegateDelete);
-
-  std::random_device random_device;
-  auto rng = std::mt19937(random_device());
-  auto shape_rng =
-      std::bind(std::uniform_int_distribution(2, 5), std::ref(rng));
-  const auto batch = shape_rng();
-  const auto channels = shape_rng();
-
-  PreluTester()
-      .InputShape({batch, channels})
-      .SlopeShape({channels})
-      .Test(xnnpack_delegate.get());
-}
-
-TEST(Prelu, DISABLED_2DBy0D) {
-  std::unique_ptr
-      xnnpack_delegate(TfLiteXNNPackDelegateCreate(nullptr),
-                       TfLiteXNNPackDelegateDelete);
-
-  std::random_device random_device;
-  auto rng = std::mt19937(random_device());
-  auto shape_rng =
-      std::bind(std::uniform_int_distribution(2, 5), std::ref(rng));
-  const auto batch = shape_rng();
-  const auto channels = shape_rng();
-
-  PreluTester()
-      .InputShape({batch, channels})
-      .SlopeShape({})
-      .Test(xnnpack_delegate.get());
-}
-
-TEST(Prelu, 1DBy1D) {
-  std::unique_ptr
-      xnnpack_delegate(TfLiteXNNPackDelegateCreate(nullptr),
-                       TfLiteXNNPackDelegateDelete);
-
-  std::random_device random_device;
-  auto rng = std::mt19937(random_device());
-  auto shape_rng =
-      std::bind(std::uniform_int_distribution(2, 5), std::ref(rng));
-  const auto batch = shape_rng();
-
-  PreluTester().InputShape({batch}).SlopeShape({batch}).Test(
-      xnnpack_delegate.get());
-}
-
-TEST(Prelu, DISABLED_1DBy0D) {
-  std::unique_ptr
-      xnnpack_delegate(TfLiteXNNPackDelegateCreate(nullptr),
-                       TfLiteXNNPackDelegateDelete);
-
-  std::random_device random_device;
-  auto rng = std::mt19937(random_device());
-  auto shape_rng =
-      std::bind(std::uniform_int_distribution(2, 5), std::ref(rng));
-  const auto batch = shape_rng();
-
-  PreluTester().InputShape({batch}).SlopeShape({}).Test(xnnpack_delegate.get());
-}
-
-TEST(Prelu, FP16Weights) {
-  std::unique_ptr
-      xnnpack_delegate(TfLiteXNNPackDelegateCreate(nullptr),
-                       TfLiteXNNPackDelegateDelete);
-
-  std::random_device random_device;
-  auto rng = std::mt19937(random_device());
-  auto shape_rng =
-      std::bind(std::uniform_int_distribution(2, 5), std::ref(rng));
-  const auto batch = shape_rng();
-  const auto height = shape_rng();
-  const auto width = shape_rng();
-  const auto channels = shape_rng();
-
-  PreluTester()
-      .InputShape({batch, height, width, channels})
-      .SlopeShape({channels})
-      .FP16Weights()
-      .Test(xnnpack_delegate.get());
-}
-
-TEST(Prelu, SparseWeights) {
-  std::unique_ptr
-      xnnpack_delegate(TfLiteXNNPackDelegateCreate(nullptr),
-                       TfLiteXNNPackDelegateDelete);
-
-  std::random_device random_device;
-  auto rng = std::mt19937(random_device());
-  auto shape_rng =
-      std::bind(std::uniform_int_distribution(2, 5), std::ref(rng));
-  const auto batch = shape_rng();
-  const auto height = shape_rng();
-  const auto width = shape_rng();
-  const auto channels = shape_rng();
-
-  PreluTester()
-      .InputShape({batch, height, width, channels})
-      .SlopeShape({channels})
-      .SparseWeights()
-      .Test(xnnpack_delegate.get());
-}
-
-TEST(Prelu, MultiThreading) {
-  TfLiteXNNPackDelegateOptions delegate_options =
-      TfLiteXNNPackDelegateOptionsDefault();
-  delegate_options.num_threads = 2;
-  std::unique_ptr
-      xnnpack_delegate(TfLiteXNNPackDelegateCreate(&delegate_options),
-                       TfLiteXNNPackDelegateDelete);
-
-  std::random_device random_device;
-  auto rng = std::mt19937(random_device());
-  auto shape_rng =
-      std::bind(std::uniform_int_distribution(2, 5), std::ref(rng));
-  const auto batch = shape_rng();
-  const auto height = shape_rng();
-  const auto width = shape_rng();
-  const auto channels = shape_rng();
-
-  PreluTester()
-      .InputShape({batch, height, width, channels})
-      .SlopeShape({channels})
-      .Test(xnnpack_delegate.get());
-}
-
-}  // namespace xnnpack
-}  // namespace tflite
diff --git a/tensorflow/lite/delegates/xnnpack/prelu_tester.cc b/tensorflow/lite/delegates/xnnpack/prelu_tester.cc
deleted file mode 100644
index 90fa0fb67c0..00000000000
--- a/tensorflow/lite/delegates/xnnpack/prelu_tester.cc
+++ /dev/null
@@ -1,237 +0,0 @@
-/* Copyright 2020 The TensorFlow Authors. All Rights Reserved.
-
-Licensed under the Apache License, Version 2.0 (the "License");
-you may not use this file except in compliance with the License.
-You may obtain a copy of the License at
-
-    http://www.apache.org/licenses/LICENSE-2.0
-
-Unless required by applicable law or agreed to in writing, software
-distributed under the License is distributed on an "AS IS" BASIS,
-WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-See the License for the specific language governing permissions and
-limitations under the License.
-==============================================================================*/
-
-#include "tensorflow/lite/delegates/xnnpack/prelu_tester.h"
-
-#include 
-#include 
-#include 
-#include 
-#include 
-#include 
-
-#include 
-#include "third_party/FP16/include/fp16.h"
-#include "flatbuffers/flatbuffers.h"  // from @flatbuffers
-#include "tensorflow/lite/interpreter.h"
-#include "tensorflow/lite/kernels/register.h"
-#include "tensorflow/lite/model.h"
-#include "tensorflow/lite/schema/schema_generated.h"
-#include "tensorflow/lite/version.h"
-
-namespace tflite {
-namespace xnnpack {
-
-void PreluTester::Test(TfLiteDelegate* delegate) const {
-  std::random_device random_device;
-  auto rng = std::mt19937(random_device());
-  auto input_rng = std::bind(std::uniform_real_distribution(-1.0f, 1.0f),
-                             std::ref(rng));
-
-  std::vector buffer = CreateTfLiteModel();
-  const Model* model = GetModel(buffer.data());
-
-  std::unique_ptr delegate_interpreter;
-  ASSERT_EQ(
-      InterpreterBuilder(model, ::tflite::ops::builtin::BuiltinOpResolver())(
-          &delegate_interpreter),
-      kTfLiteOk);
-  std::unique_ptr default_interpreter;
-  ASSERT_EQ(
-      InterpreterBuilder(model, ::tflite::ops::builtin::BuiltinOpResolver())(
-          &default_interpreter),
-      kTfLiteOk);
-
-  ASSERT_TRUE(delegate_interpreter);
-  ASSERT_TRUE(default_interpreter);
-
-  ASSERT_EQ(delegate_interpreter->inputs().size(), 1);
-  ASSERT_EQ(default_interpreter->inputs().size(), 1);
-
-  ASSERT_EQ(delegate_interpreter->outputs().size(), 1);
-  ASSERT_EQ(default_interpreter->outputs().size(), 1);
-
-  ASSERT_EQ(delegate_interpreter->AllocateTensors(), kTfLiteOk);
-  ASSERT_EQ(default_interpreter->AllocateTensors(), kTfLiteOk);
-
-  ASSERT_EQ(delegate_interpreter->ModifyGraphWithDelegate(delegate), kTfLiteOk);
-
-  float* default_input_data = default_interpreter->typed_tensor(
-      default_interpreter->inputs()[0]);
-  std::generate(default_input_data,
-                default_input_data + ComputeSize(InputShape()),
-                std::ref(input_rng));
-
-  float* xnnpack_input_data = delegate_interpreter->typed_tensor(
-      delegate_interpreter->inputs()[0]);
-  std::copy(default_input_data, default_input_data + ComputeSize(InputShape()),
-            xnnpack_input_data);
-
-  ASSERT_EQ(default_interpreter->Invoke(), kTfLiteOk);
-  ASSERT_EQ(delegate_interpreter->Invoke(), kTfLiteOk);
-
-  float* default_output_data = default_interpreter->typed_tensor(
-      default_interpreter->outputs()[0]);
-  float* xnnpack_output_data = delegate_interpreter->typed_tensor(
-      delegate_interpreter->outputs()[0]);
-
-  for (size_t i = 0; i < ComputeSize(OutputShape()); i++) {
-    ASSERT_EQ(default_output_data[i], xnnpack_output_data[i]);
-  }
-}
-
-std::vector PreluTester::CreateTfLiteModel() const {
-  std::random_device random_device;
-  auto rng = std::mt19937(random_device());
-  auto slope_rng = std::bind(std::uniform_real_distribution(0.25f, 0.5f),
-                             std::ref(rng));
-
-  flatbuffers::FlatBufferBuilder builder;
-  std::vector> operator_codes{
-      {CreateOperatorCode(builder, BuiltinOperator_PRELU)}};
-  if (FP16Weights()) {
-    operator_codes.emplace_back(
-        CreateOperatorCode(builder, BuiltinOperator_DEQUANTIZE));
-  } else if (SparseWeights()) {
-    operator_codes.emplace_back(
-        CreateOperatorCode(builder, BuiltinOperator_DENSIFY));
-  }
-
-  std::vector> buffers{{
-      CreateBuffer(builder, builder.CreateVector({})),
-  }};
-
-  if (FP16Weights()) {
-    std::vector slope_data(ComputeSize(SlopeShape()));
-    std::generate(slope_data.begin(), slope_data.end(),
-                  std::bind(fp16_ieee_from_fp32_value, slope_rng));
-
-    buffers.push_back(CreateBuffer(
-        builder, builder.CreateVector(
-                     reinterpret_cast(slope_data.data()),
-                     sizeof(uint16_t) * slope_data.size())));
-  } else {
-    std::vector slope_data(ComputeSize(SlopeShape()));
-    std::generate(slope_data.begin(), slope_data.end(), slope_rng);
-
-    buffers.push_back(CreateBuffer(
-        builder, builder.CreateVector(
-                     reinterpret_cast(slope_data.data()),
-                     sizeof(float) * slope_data.size())));
-  }
-
-  std::vector> tensors;
-  std::vector> operators;
-  if (FP16Weights()) {
-    tensors.emplace_back(CreateTensor(
-        builder,
-        builder.CreateVector(SlopeShape().data(), SlopeShape().size()),
-        TensorType_FLOAT16, /*buffer=*/1));
-  } else if (SparseWeights()) {
-    const int dims_count = SlopeShape().size();
-    std::vector> dim_metadata(
-        dims_count);
-    std::vector traversal_order(dims_count);
-    for (int i = 0; i < dims_count; i++) {
-      traversal_order[i] = i;
-      dim_metadata[i] = CreateDimensionMetadata(builder, DimensionType_DENSE,
-                                                SlopeShape()[i]);
-    }
-    const flatbuffers::Offset sparsity_param =
-        CreateSparsityParameters(builder, builder.CreateVector(traversal_order),
-                                 0, builder.CreateVector(dim_metadata));
-    tensors.emplace_back(CreateTensor(
-        builder,
-        builder.CreateVector(SlopeShape().data(), SlopeShape().size()),
-        TensorType_FLOAT32, /*buffer=*/1, /*name=*/0, /*quantization=*/0,
-        /*is_variable=*/false, /*sparsity=*/sparsity_param));
-  }
-  if (FP16Weights()) {
-    const std::array dequantize_inputs{{0}};
-    const std::array dequantize_outputs{{2}};
-    operators.emplace_back(CreateOperator(
-        builder, /*opcode_index=*/1,
-        builder.CreateVector(dequantize_inputs.data(),
-                                      dequantize_inputs.size()),
-        builder.CreateVector(dequantize_outputs.data(),
-                                      dequantize_outputs.size())));
-  } else if (SparseWeights()) {
-    const std::array densify_inputs{{0}};
-    const std::array densify_outputs{{2}};
-    operators.emplace_back(
-        CreateOperator(builder, /*opcode_index=*/1,
-                       builder.CreateVector(densify_inputs.data(),
-                                                     densify_inputs.size()),
-                       builder.CreateVector(densify_outputs.data(),
-                                                     densify_outputs.size())));
-  }
-  tensors.emplace_back(CreateTensor(
-      builder,
-      builder.CreateVector(InputShape().data(), InputShape().size()),
-      TensorType_FLOAT32));
-  tensors.emplace_back(CreateTensor(
-      builder,
-      builder.CreateVector(SlopeShape().data(), SlopeShape().size()),
-      TensorType_FLOAT32,
-      /*buffer=*/(FP16Weights() || SparseWeights()) ? 0 : 1));
-  tensors.emplace_back(CreateTensor(
-      builder,
-      builder.CreateVector(OutputShape().data(), OutputShape().size()),
-      TensorType_FLOAT32));
-
-  const std::array op_inputs{
-      {static_cast(tensors.size()) - 3,
-       static_cast(tensors.size()) - 2}};
-  const std::array op_outputs{
-      {static_cast(tensors.size()) - 1}};
-  operators.emplace_back(CreateOperator(
-      builder, /*opcode_index=*/0,
-      builder.CreateVector(op_inputs.data(), op_inputs.size()),
-      builder.CreateVector(op_outputs.data(), op_outputs.size())));
-
-  const std::array subgraph_inputs{
-      {static_cast(tensors.size() - 3)}};
-  const std::array subgraph_outputs{
-      {static_cast(tensors.size()) - 1}};
-  flatbuffers::Offset subgraph = CreateSubGraph(
-      builder, builder.CreateVector(tensors.data(), tensors.size()),
-      builder.CreateVector(subgraph_inputs.data(),
-                                    subgraph_inputs.size()),
-      builder.CreateVector(subgraph_outputs.data(),
-                                    subgraph_outputs.size()),
-      builder.CreateVector(operators.data(), operators.size()));
-
-  flatbuffers::Offset description =
-      builder.CreateString("PReLU model");
-
-  flatbuffers::Offset model_buffer = CreateModel(
-      builder, TFLITE_SCHEMA_VERSION,
-      builder.CreateVector(operator_codes.data(), operator_codes.size()),
-      builder.CreateVector(&subgraph, 1), description,
-      builder.CreateVector(buffers.data(), buffers.size()));
-
-  builder.Finish(model_buffer);
-
-  return std::vector(builder.GetBufferPointer(),
-                           builder.GetBufferPointer() + builder.GetSize());
-}
-
-int32_t PreluTester::ComputeSize(const std::vector& shape) {
-  return std::accumulate(shape.cbegin(), shape.cend(), 1,
-                         std::multiplies());
-}
-
-}  // namespace xnnpack
-}  // namespace tflite
diff --git a/tensorflow/lite/delegates/xnnpack/prelu_tester.h b/tensorflow/lite/delegates/xnnpack/prelu_tester.h
deleted file mode 100644
index e89bae6029b..00000000000
--- a/tensorflow/lite/delegates/xnnpack/prelu_tester.h
+++ /dev/null
@@ -1,88 +0,0 @@
-/* Copyright 2020 The TensorFlow Authors. All Rights Reserved.
-
-Licensed under the Apache License, Version 2.0 (the "License");
-you may not use this file except in compliance with the License.
-You may obtain a copy of the License at
-
-    http://www.apache.org/licenses/LICENSE-2.0
-
-Unless required by applicable law or agreed to in writing, software
-distributed under the License is distributed on an "AS IS" BASIS,
-WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-See the License for the specific language governing permissions and
-limitations under the License.
-==============================================================================*/
-
-#ifndef TENSORFLOW_LITE_DELEGATES_XNNPACK_PRELU_TESTER_H_
-#define TENSORFLOW_LITE_DELEGATES_XNNPACK_PRELU_TESTER_H_
-
-#include 
-#include 
-
-#include 
-#include "tensorflow/lite/c/common.h"
-
-namespace tflite {
-namespace xnnpack {
-
-class PreluTester {
- public:
-  PreluTester() = default;
-  PreluTester(const PreluTester&) = delete;
-  PreluTester& operator=(const PreluTester&) = delete;
-
-  inline PreluTester& InputShape(std::initializer_list shape) {
-    for (auto it = shape.begin(); it != shape.end(); ++it) {
-      EXPECT_GT(*it, 0);
-    }
-    input_shape_ = std::vector(shape.begin(), shape.end());
-    return *this;
-  }
-
-  inline const std::vector& InputShape() const { return input_shape_; }
-
-  inline PreluTester& SlopeShape(std::initializer_list shape) {
-    for (auto it = shape.begin(); it != shape.end(); ++it) {
-      EXPECT_GT(*it, 0);
-    }
-    slope_shape_ = std::vector(shape.begin(), shape.end());
-    return *this;
-  }
-
-  inline const std::vector& SlopeShape() const { return slope_shape_; }
-
-  inline const std::vector& OutputShape() const {
-    return InputShape();
-  }
-
-  inline PreluTester& FP16Weights() {
-    fp16_weights_ = true;
-    return *this;
-  }
-
-  inline bool FP16Weights() const { return fp16_weights_; }
-
-  inline PreluTester& SparseWeights() {
-    sparse_weights_ = true;
-    return *this;
-  }
-
-  inline bool SparseWeights() const { return sparse_weights_; }
-
-  void Test(TfLiteDelegate* delegate) const;
-
- private:
-  std::vector CreateTfLiteModel() const;
-
-  static int32_t ComputeSize(const std::vector& shape);
-
-  std::vector input_shape_;
-  std::vector slope_shape_;
-  bool fp16_weights_ = false;
-  bool sparse_weights_ = false;
-};
-
-}  // namespace xnnpack
-}  // namespace tflite
-
-#endif  // TENSORFLOW_LITE_DELEGATES_XNNPACK_PRELU_TESTER_H_
diff --git a/tensorflow/lite/delegates/xnnpack/xnnpack_delegate.cc b/tensorflow/lite/delegates/xnnpack/xnnpack_delegate.cc
index 31468ef7407..0afc9c32122 100644
--- a/tensorflow/lite/delegates/xnnpack/xnnpack_delegate.cc
+++ b/tensorflow/lite/delegates/xnnpack/xnnpack_delegate.cc
@@ -2266,8 +2266,7 @@ class Subgraph {
     const TfLiteTensor& input_tensor = tensors[node->inputs->data[0]];
     TF_LITE_ENSURE_STATUS(CheckTensorFloatType(
         logging_context, input_tensor, node->inputs->data[0], node_index));
-    TF_LITE_ENSURE_STATUS(CheckTensorShape(logging_context, input_tensor, 1,
-                                           XNN_MAX_TENSOR_DIMS,
+    TF_LITE_ENSURE_STATUS(CheckTensorShape(logging_context, input_tensor, 4,
                                            node->inputs->data[0]));
     TF_LITE_ENSURE_STATUS(CheckTensorNonDynamicAllocation(
         logging_context, input_tensor, node->inputs->data[0], node_index));
@@ -2285,8 +2284,7 @@ class Subgraph {
     const TfLiteTensor& output_tensor = tensors[node->outputs->data[0]];
     TF_LITE_ENSURE_STATUS(CheckTensorFloatType(
         logging_context, output_tensor, node->outputs->data[0], node_index));
-    TF_LITE_ENSURE_STATUS(CheckTensorShape(logging_context, output_tensor, 1,
-                                           XNN_MAX_TENSOR_DIMS,
+    TF_LITE_ENSURE_STATUS(CheckTensorShape(logging_context, output_tensor, 4,
                                            node->outputs->data[0]));
     TF_LITE_ENSURE_STATUS(CheckTensorNonDynamicAllocation(
         logging_context, output_tensor, node->outputs->data[0], node_index));

From 61c2e69663ee78de30431dc129d059d29d6d80ea Mon Sep 17 00:00:00 2001
From: Rick Chao 
Date: Mon, 22 Jun 2020 14:33:38 -0700
Subject: [PATCH 469/867] Make task_type and task_id standard properties in
 tf.distribute cluster resolvers.

PiperOrigin-RevId: 317736970
Change-Id: Ia9c76462afc4c2fcc42a149960b50b2cbcafd482
---
 .../cluster_resolver/cluster_resolver.py      | 121 +++++++++++++++++-
 .../gce_cluster_resolver_test.py              |  32 +++++
 ...e.cluster_resolver.-cluster-resolver.pbtxt |   8 ++
 ...esolver.-kubernetes-cluster-resolver.pbtxt |   8 ++
 ...ter_resolver.-slurm-cluster-resolver.pbtxt |   8 ++
 ...ter_resolver.-t-p-u-cluster-resolver.pbtxt |   8 ++
 ...e.cluster_resolver.-cluster-resolver.pbtxt |   8 ++
 ...esolver.-kubernetes-cluster-resolver.pbtxt |   8 ++
 ...ter_resolver.-slurm-cluster-resolver.pbtxt |   8 ++
 ...ter_resolver.-t-p-u-cluster-resolver.pbtxt |   8 ++
 10 files changed, 212 insertions(+), 5 deletions(-)

diff --git a/tensorflow/python/distribute/cluster_resolver/cluster_resolver.py b/tensorflow/python/distribute/cluster_resolver/cluster_resolver.py
index a8babc21af6..f43bfc9845f 100644
--- a/tensorflow/python/distribute/cluster_resolver/cluster_resolver.py
+++ b/tensorflow/python/distribute/cluster_resolver/cluster_resolver.py
@@ -71,11 +71,22 @@ class ClusterResolver(object):
   workers. This will eventually allow us to automatically recover from
   underlying machine failures and scale TensorFlow worker clusters up and down.
 
-  Note to Implementors: In addition to these abstract methods, you must also
-  implement the task_type, task_id, and rpc_layer attributes. You may choose
-  to implement them either as properties with getters or setters or directly
-  set the attributes. The task_type and task_id attributes are required by
-  `tf.distribute.experimental.MultiWorkerMirroredStrategy`.
+  Note to Implementors of `tf.distribute.cluster_resolver.ClusterResolver`
+  subclass: In addition to these abstract methods, when task_type, task_id, and
+  rpc_layer attributes are applicable, you should also implement them either as
+  properties with getters or setters, or directly set the attributes
+  `self._task_type`, `self._task_id`, or `self._rpc_layer` so the base class'
+  getters and setters are used. See
+  `tf.distribute.cluster_resolver.SimpleClusterResolver.__init__` for an
+  example.
+
+  In general, multi-client tf.distribute strategies such as
+  `tf.distribute.experimental.MultiWorkerMirroredStrategy` require task_type and
+  task_id properties to be available in the `ClusterResolver` they are using. On
+  the other hand, these concepts are not applicable in single-client strategies,
+  such as `tf.distribute.experimental.TPUStrategy`, because the program is only
+  expected to be run on one task, so there should not be a need to have code
+  branches according to task type and task id.
 
   - task_type is the name of the server's current named job (e.g. 'worker',
      'ps' in a distributed parameterized training job).
@@ -177,6 +188,106 @@ class ClusterResolver(object):
     """
     return ''
 
+  @property
+  def task_type(self):
+    """Returns the task type this `ClusterResolver` indicates.
+
+    In TensorFlow distributed environment, each job may have an applicable
+    task type. Valid task types in TensorFlow include
+    'chief': a worker that is designated with more responsibility,
+    'worker': a regular worker for training/evaluation,
+    'ps': a parameter server, or
+    'evaluator': an evaluator that evaluates the checkpoints for metrics.
+
+    See [Multi-worker configuration](
+    https://www.tensorflow.org/tutorials/distribute/multi_worker_with_keras#multi-worker_configuration)
+    for more information about 'chief' and 'worker' task type, which are most
+    commonly used.
+
+    Having access to such information is useful when user needs to run specific
+    code according to task types. For example,
+
+    ```python
+    cluster_spec = tf.train.ClusterSpec({
+        "ps": ["localhost:2222", "localhost:2223"],
+        "worker": ["localhost:2224", "localhost:2225", "localhost:2226"]
+    })
+
+    # SimpleClusterResolver is used here for illustration; other cluster
+    # resolvers may be used for other source of task type/id.
+    simple_resolver = SimpleClusterResolver(cluster_spec, task_type="worker",
+                                            task_id=1)
+
+    ...
+
+    if cluster_resolver.task_type == 'worker':
+      # Perform something that's only applicable on workers. This block
+      # will run on this particular instance since we've specified this task to
+      # be a worker in above cluster resolver.
+    elif cluster_resolver.task_type == 'ps':
+      # Perform something that's only applicable on parameter servers. This
+      # block will not run on this particular instance.
+    ```
+
+    Returns `None` if such information is not available or is not applicable
+    in the current distributed environment, such as training with
+    `tf.distribute.experimental.TPUStrategy`.
+
+    For more information, please see
+    `tf.distribute.cluster_resolver.ClusterResolver`'s class doc.
+    """
+    return getattr(self, '_task_type', None)
+
+  @property
+  def task_id(self):
+    """Returns the task id this `ClusterResolver` indicates.
+
+    In TensorFlow distributed environment, each job may have an applicable
+    task id, which is the index of the instance within its task type. This is
+    useful when user needs to run specific code according to task index. For
+    example,
+
+    ```python
+    cluster_spec = tf.train.ClusterSpec({
+        "ps": ["localhost:2222", "localhost:2223"],
+        "worker": ["localhost:2224", "localhost:2225", "localhost:2226"]
+    })
+
+    # SimpleClusterResolver is used here for illustration; other cluster
+    # resolvers may be used for other source of task type/id.
+    simple_resolver = SimpleClusterResolver(cluster_spec, task_type="worker",
+                                            task_id=0)
+
+    ...
+
+    if cluster_resolver.task_type == 'worker' and cluster_resolver.task_id == 0:
+      # Perform something that's only applicable on 'worker' type, id 0. This
+      # block will run on this particular instance since we've specified this
+      # task to be a 'worker', id 0 in above cluster resolver.
+    else:
+      # Perform something that's only applicable on other ids. This block will
+      # not run on this particular instance.
+    ```
+
+    Returns `None` if such information is not available or is not applicable
+    in the current distributed environment, such as training with
+    `tf.distribute.cluster_resolver.TPUClusterResolver`.
+
+    For more information, please see
+    `tf.distribute.cluster_resolver.ClusterResolver`'s class docstring.
+    """
+    return getattr(self, '_task_id', None)
+
+  @task_type.setter
+  def task_type(self, task_type):
+    """Setter of `task_type` property. See `task_type` property doc."""
+    self._task_type = task_type
+
+  @task_id.setter
+  def task_id(self, task_id):
+    """Setter of `task_id` property. See `task_type` property doc."""
+    self._task_id = task_id
+
 
 @tf_export('distribute.cluster_resolver.SimpleClusterResolver')
 class SimpleClusterResolver(ClusterResolver):
diff --git a/tensorflow/python/distribute/cluster_resolver/gce_cluster_resolver_test.py b/tensorflow/python/distribute/cluster_resolver/gce_cluster_resolver_test.py
index f39c86a0495..d8037497cb9 100644
--- a/tensorflow/python/distribute/cluster_resolver/gce_cluster_resolver_test.py
+++ b/tensorflow/python/distribute/cluster_resolver/gce_cluster_resolver_test.py
@@ -310,5 +310,37 @@ class GCEClusterResolverTest(test.TestCase):
     """
     self._verifyClusterSpecEquality(actual_cluster_spec, expected_proto)
 
+  def testSettingTaskTypeRaiseError(self):
+    name_to_ip = [
+        {
+            'name': 'instance1',
+            'ip': '10.1.2.3'
+        },
+        {
+            'name': 'instance2',
+            'ip': '10.2.3.4'
+        },
+        {
+            'name': 'instance3',
+            'ip': '10.3.4.5'
+        },
+    ]
+
+    gce_cluster_resolver = GCEClusterResolver(
+        project='test-project',
+        zone='us-east1-d',
+        instance_group='test-instance-group',
+        task_type='testworker',
+        port=8470,
+        credentials=None,
+        service=self.gen_standard_mock_service_client(name_to_ip))
+
+    with self.assertRaisesRegexp(
+        RuntimeError, 'You cannot reset the task_type '
+        'of the GCEClusterResolver after it has '
+        'been created.'):
+      gce_cluster_resolver.task_type = 'foobar'
+
+
 if __name__ == '__main__':
   test.main()
diff --git a/tensorflow/tools/api/golden/v1/tensorflow.distribute.cluster_resolver.-cluster-resolver.pbtxt b/tensorflow/tools/api/golden/v1/tensorflow.distribute.cluster_resolver.-cluster-resolver.pbtxt
index 0b35b61b4c0..c9dbca7368e 100644
--- a/tensorflow/tools/api/golden/v1/tensorflow.distribute.cluster_resolver.-cluster-resolver.pbtxt
+++ b/tensorflow/tools/api/golden/v1/tensorflow.distribute.cluster_resolver.-cluster-resolver.pbtxt
@@ -6,6 +6,14 @@ tf_class {
     name: "environment"
     mtype: ""
   }
+  member {
+    name: "task_id"
+    mtype: ""
+  }
+  member {
+    name: "task_type"
+    mtype: ""
+  }
   member_method {
     name: "__init__"
   }
diff --git a/tensorflow/tools/api/golden/v1/tensorflow.distribute.cluster_resolver.-kubernetes-cluster-resolver.pbtxt b/tensorflow/tools/api/golden/v1/tensorflow.distribute.cluster_resolver.-kubernetes-cluster-resolver.pbtxt
index 3220d68e054..2819ca85612 100644
--- a/tensorflow/tools/api/golden/v1/tensorflow.distribute.cluster_resolver.-kubernetes-cluster-resolver.pbtxt
+++ b/tensorflow/tools/api/golden/v1/tensorflow.distribute.cluster_resolver.-kubernetes-cluster-resolver.pbtxt
@@ -7,6 +7,14 @@ tf_class {
     name: "environment"
     mtype: ""
   }
+  member {
+    name: "task_id"
+    mtype: ""
+  }
+  member {
+    name: "task_type"
+    mtype: ""
+  }
   member_method {
     name: "__init__"
     argspec: "args=[\'self\', \'job_to_label_mapping\', \'tf_server_port\', \'rpc_layer\', \'override_client\'], varargs=None, keywords=None, defaults=[\'None\', \'8470\', \'grpc\', \'None\'], "
diff --git a/tensorflow/tools/api/golden/v1/tensorflow.distribute.cluster_resolver.-slurm-cluster-resolver.pbtxt b/tensorflow/tools/api/golden/v1/tensorflow.distribute.cluster_resolver.-slurm-cluster-resolver.pbtxt
index 30cfac0830c..a2c63243c86 100644
--- a/tensorflow/tools/api/golden/v1/tensorflow.distribute.cluster_resolver.-slurm-cluster-resolver.pbtxt
+++ b/tensorflow/tools/api/golden/v1/tensorflow.distribute.cluster_resolver.-slurm-cluster-resolver.pbtxt
@@ -7,6 +7,14 @@ tf_class {
     name: "environment"
     mtype: ""
   }
+  member {
+    name: "task_id"
+    mtype: ""
+  }
+  member {
+    name: "task_type"
+    mtype: ""
+  }
   member_method {
     name: "__init__"
     argspec: "args=[\'self\', \'jobs\', \'port_base\', \'gpus_per_node\', \'gpus_per_task\', \'tasks_per_node\', \'auto_set_gpu\', \'rpc_layer\'], varargs=None, keywords=None, defaults=[\'None\', \'8888\', \'None\', \'None\', \'None\', \'True\', \'grpc\'], "
diff --git a/tensorflow/tools/api/golden/v1/tensorflow.distribute.cluster_resolver.-t-p-u-cluster-resolver.pbtxt b/tensorflow/tools/api/golden/v1/tensorflow.distribute.cluster_resolver.-t-p-u-cluster-resolver.pbtxt
index 8eee489df93..165f8d43ace 100644
--- a/tensorflow/tools/api/golden/v1/tensorflow.distribute.cluster_resolver.-t-p-u-cluster-resolver.pbtxt
+++ b/tensorflow/tools/api/golden/v1/tensorflow.distribute.cluster_resolver.-t-p-u-cluster-resolver.pbtxt
@@ -7,6 +7,14 @@ tf_class {
     name: "environment"
     mtype: ""
   }
+  member {
+    name: "task_id"
+    mtype: ""
+  }
+  member {
+    name: "task_type"
+    mtype: ""
+  }
   member_method {
     name: "__init__"
     argspec: "args=[\'self\', \'tpu\', \'zone\', \'project\', \'job_name\', \'coordinator_name\', \'coordinator_address\', \'credentials\', \'service\', \'discovery_url\'], varargs=None, keywords=None, defaults=[\'None\', \'None\', \'None\', \'worker\', \'None\', \'None\', \'default\', \'None\', \'None\'], "
diff --git a/tensorflow/tools/api/golden/v2/tensorflow.distribute.cluster_resolver.-cluster-resolver.pbtxt b/tensorflow/tools/api/golden/v2/tensorflow.distribute.cluster_resolver.-cluster-resolver.pbtxt
index 0b35b61b4c0..c9dbca7368e 100644
--- a/tensorflow/tools/api/golden/v2/tensorflow.distribute.cluster_resolver.-cluster-resolver.pbtxt
+++ b/tensorflow/tools/api/golden/v2/tensorflow.distribute.cluster_resolver.-cluster-resolver.pbtxt
@@ -6,6 +6,14 @@ tf_class {
     name: "environment"
     mtype: ""
   }
+  member {
+    name: "task_id"
+    mtype: ""
+  }
+  member {
+    name: "task_type"
+    mtype: ""
+  }
   member_method {
     name: "__init__"
   }
diff --git a/tensorflow/tools/api/golden/v2/tensorflow.distribute.cluster_resolver.-kubernetes-cluster-resolver.pbtxt b/tensorflow/tools/api/golden/v2/tensorflow.distribute.cluster_resolver.-kubernetes-cluster-resolver.pbtxt
index 3220d68e054..2819ca85612 100644
--- a/tensorflow/tools/api/golden/v2/tensorflow.distribute.cluster_resolver.-kubernetes-cluster-resolver.pbtxt
+++ b/tensorflow/tools/api/golden/v2/tensorflow.distribute.cluster_resolver.-kubernetes-cluster-resolver.pbtxt
@@ -7,6 +7,14 @@ tf_class {
     name: "environment"
     mtype: ""
   }
+  member {
+    name: "task_id"
+    mtype: ""
+  }
+  member {
+    name: "task_type"
+    mtype: ""
+  }
   member_method {
     name: "__init__"
     argspec: "args=[\'self\', \'job_to_label_mapping\', \'tf_server_port\', \'rpc_layer\', \'override_client\'], varargs=None, keywords=None, defaults=[\'None\', \'8470\', \'grpc\', \'None\'], "
diff --git a/tensorflow/tools/api/golden/v2/tensorflow.distribute.cluster_resolver.-slurm-cluster-resolver.pbtxt b/tensorflow/tools/api/golden/v2/tensorflow.distribute.cluster_resolver.-slurm-cluster-resolver.pbtxt
index 30cfac0830c..a2c63243c86 100644
--- a/tensorflow/tools/api/golden/v2/tensorflow.distribute.cluster_resolver.-slurm-cluster-resolver.pbtxt
+++ b/tensorflow/tools/api/golden/v2/tensorflow.distribute.cluster_resolver.-slurm-cluster-resolver.pbtxt
@@ -7,6 +7,14 @@ tf_class {
     name: "environment"
     mtype: ""
   }
+  member {
+    name: "task_id"
+    mtype: ""
+  }
+  member {
+    name: "task_type"
+    mtype: ""
+  }
   member_method {
     name: "__init__"
     argspec: "args=[\'self\', \'jobs\', \'port_base\', \'gpus_per_node\', \'gpus_per_task\', \'tasks_per_node\', \'auto_set_gpu\', \'rpc_layer\'], varargs=None, keywords=None, defaults=[\'None\', \'8888\', \'None\', \'None\', \'None\', \'True\', \'grpc\'], "
diff --git a/tensorflow/tools/api/golden/v2/tensorflow.distribute.cluster_resolver.-t-p-u-cluster-resolver.pbtxt b/tensorflow/tools/api/golden/v2/tensorflow.distribute.cluster_resolver.-t-p-u-cluster-resolver.pbtxt
index 8eee489df93..165f8d43ace 100644
--- a/tensorflow/tools/api/golden/v2/tensorflow.distribute.cluster_resolver.-t-p-u-cluster-resolver.pbtxt
+++ b/tensorflow/tools/api/golden/v2/tensorflow.distribute.cluster_resolver.-t-p-u-cluster-resolver.pbtxt
@@ -7,6 +7,14 @@ tf_class {
     name: "environment"
     mtype: ""
   }
+  member {
+    name: "task_id"
+    mtype: ""
+  }
+  member {
+    name: "task_type"
+    mtype: ""
+  }
   member_method {
     name: "__init__"
     argspec: "args=[\'self\', \'tpu\', \'zone\', \'project\', \'job_name\', \'coordinator_name\', \'coordinator_address\', \'credentials\', \'service\', \'discovery_url\'], varargs=None, keywords=None, defaults=[\'None\', \'None\', \'None\', \'worker\', \'None\', \'None\', \'default\', \'None\', \'None\'], "

From 08180e68453306e767f36c755ada377c688fa6ee Mon Sep 17 00:00:00 2001
From: Raman Sarokin 
Date: Mon, 22 Jun 2020 14:38:34 -0700
Subject: [PATCH 470/867] Padding converted to new style.

PiperOrigin-RevId: 317737917
Change-Id: Ie21457323a2cdb0376c16065fc4b809a5cc9a1d3
---
 .../lite/delegates/gpu/cl/kernels/padding.cc  | 134 +++++++++---------
 1 file changed, 64 insertions(+), 70 deletions(-)

diff --git a/tensorflow/lite/delegates/gpu/cl/kernels/padding.cc b/tensorflow/lite/delegates/gpu/cl/kernels/padding.cc
index 48edcb448a1..883067a2c2d 100644
--- a/tensorflow/lite/delegates/gpu/cl/kernels/padding.cc
+++ b/tensorflow/lite/delegates/gpu/cl/kernels/padding.cc
@@ -26,20 +26,21 @@ namespace gpu {
 namespace cl {
 namespace {
 
-std::string GetPaddingCode(
-    const OperationDef& op_def,
-    const std::vector& linked_operations,
-    const PadAttributes& attr) {
-  TensorCodeGenerator src_tensor(
-      "src_data",
-      WHSBPoint{"src_size.x", "src_size.y", "src_size.z", "src_size.w"},
-      op_def.src_tensors[0]);
-  TensorCodeGenerator dst_tensor(
-      "dst_data",
-      WHSBPoint{"dst_size.x", "dst_size.y", "dst_size.z", "dst_size.w"},
-      op_def.dst_tensors[0]);
+std::string GetPaddingCode(const OperationDef& op_def,
+                           const PadAttributes& attr, Arguments* args) {
+  args->AddObjectRef(
+      "src_tensor", AccessType::READ,
+      absl::make_unique(op_def.src_tensors[0]));
+  args->AddObjectRef(
+      "dst_tensor", AccessType::WRITE,
+      absl::make_unique(op_def.dst_tensors[0]));
+  args->AddInt("prepended_x");
+  args->AddInt("prepended_y");
+  args->AddInt("prepended_z");
+  args->AddInt("prepended_w");
 
-  const std::string dst_batch = op_def.IsBatchSupported() ? "B" : "";
+  const std::string dst_batch =
+      op_def.dst_tensors[0].HasAxis(Axis::BATCH) ? "B" : "0";
   std::string c = GetCommonDefines(op_def.precision);
   const std::string channels[] = {".x", ".y", ".z", ".w"};
 
@@ -51,76 +52,70 @@ std::string GetPaddingCode(
   }
 
   c += "__kernel void main_function(\n";
-  c += src_tensor.GetDeclaration(AccessType::READ);
-  c += GetArgsDeclaration(linked_operations);
-  c += dst_tensor.GetDeclaration(AccessType::WRITE) + ",\n";
-  c += "    int4 src_size,      \n";
-  c += "    int src_channels,   \n";
-  c += "    int4 dst_size,      \n";
-  c += "    int4 prepended      \n";
-  c += ") {\n";
-  if (op_def.IsBatchSupported()) {
+  c += "$0) {\n";
+  if (op_def.dst_tensors[0].HasAxis(Axis::BATCH)) {
     c += "  int linear_id = get_global_id(0);\n";
-    c += "  int X = linear_id / dst_size.w;\n";
-    c += "  int B = linear_id % dst_size.w;\n";
+    c += "  int X = linear_id / args.dst_tensor.Batch();\n";
+    c += "  int B = linear_id % args.dst_tensor.Batch();\n";
+    c += "  args.dst_tensor.SetBatchRef(B);\n";
   } else {
     c += "  int X = get_global_id(0);\n";
   }
   c += "  int Y = get_global_id(1);\n";
   c += "  int Z = get_global_id(2);\n";
-  c += "  if (X >= dst_size.x || Y >= dst_size.y || Z >= dst_size.z) return;\n";
+  c += "  if (X >= args.dst_tensor.Width() || Y >= args.dst_tensor.Height() || "
+       "Z >= args.dst_tensor.Slices()) { \n";
+  c += "    return; \n";
+  c += "  } \n";
   c += "  FLT4 result = (FLT4)(0.0);\n";
-  c += "  int s_x = X - prepended.x;\n";
-  c += "  int s_y = Y - prepended.y;\n";
-  if (op_def.IsBatchSupported()) {
-    c += "  int s_b = B - prepended.w;\n";
+  c += "  int s_x = X - args.prepended_x;\n";
+  c += "  int s_y = Y - args.prepended_y;\n";
+  if (op_def.src_tensors[0].HasAxis(Axis::BATCH)) {
+    c += "  int s_b = " + dst_batch + " - args.prepended_w;\n";
+    c += "  args.src_tensor.SetBatchRef(s_b);\n";
   }
-  const std::string src_batch = op_def.IsBatchSupported() ? "s_b" : "";
   if (attr.type == PaddingContentType::REFLECT) {
-    c += "  s_x = reflect(s_x, src_size.x);\n";
-    c += "  s_y = reflect(s_y, src_size.y);\n";
-    if (op_def.IsBatchSupported()) {
-      c += "  int s_b = reflect(s_b, src_size.w);\n";
+    c += "  s_x = reflect(s_x, args.src_tensor.Width());\n";
+    c += "  s_y = reflect(s_y, args.src_tensor.Height());\n";
+    if (op_def.src_tensors[0].HasAxis(Axis::BATCH)) {
+      c += "  int s_b = reflect(s_b, args.src_tensor.Batch());\n";
     }
     if (attr.prepended.c == 0 && attr.appended.c == 0) {
       // optimized case
-      c += "  result = " + src_tensor.ReadWHSB("s_x", "s_y", "Z", src_batch) +
-           ";\n";
+      c += "  result = args.src_tensor.Read(s_x, s_y, Z);\n";
     } else {
       c += "  int start_channel = Z * 4;\n";
       for (int i = 0; i < 4; ++i) {
         const auto& s = channels[i];
         c += "  {\n";
         c += "    int channel = start_channel + " + std::to_string(i) + ";\n";
-        c += "    int s_z = channel - prepended.z;\n";
+        c += "    int s_z = channel - args.prepended_z;\n";
         // We need additional clamp for z, so that we use alignment for channels
         // and can proceed extra channels that can lead to reading out of
         // resource.
-        c += "    s_z = clamp(reflect(s_z, src_channels), 0, src_channels - "
+        c += "    s_z = clamp(reflect(s_z, args.src_tensor.Channels()), 0, "
+             "args.src_tensor.Channels() - "
              "1);\n";
-        c += "    FLT4 t = " +
-             src_tensor.ReadWHSB("s_x", "s_y", "s_z / 4", src_batch) + ";\n";
+        c += "    FLT4 t = args.src_tensor.Read(s_x, s_y, s_z / 4);\n";
         c += "    FLT t_ar[4] = {t.x, t.y, t.z, t.w};\n";
         c += "    result" + s + " = t_ar[s_z % 4];\n";
         c += "  }\n";
       }
     }
   } else {
-    c += "  bool inside_x = s_x >= 0 && s_x < src_size.x;\n";
-    c += "  bool inside_y = s_y >= 0 && s_y < src_size.y;\n";
-    if (op_def.IsBatchSupported()) {
-      c += "  inside_y &= (s_b >= 0 && s_b < src_size.w);\n";
+    c += "  bool inside_x = s_x >= 0 && s_x < args.src_tensor.Width();\n";
+    c += "  bool inside_y = s_y >= 0 && s_y < args.src_tensor.Height();\n";
+    if (op_def.src_tensors[0].HasAxis(Axis::BATCH)) {
+      c += "  inside_y &= (s_b >= 0 && s_b < args.src_tensor.Batch());\n";
     }
     c += "  if (inside_x && inside_y) {\n";
     if (attr.prepended.c == 0 && attr.appended.c == 0) {
       // optimized case
-      c += "    result = " + src_tensor.ReadWHSB("s_x", "s_y", "Z", src_batch) +
-           ";\n";
+      c += "    result = args.src_tensor.Read(s_x, s_y, Z);\n";
     } else if (attr.prepended.c % 4 == 0) {
-      c += "    int s_z = Z - prepended.z / 4;\n";
-      c += "    if (s_z >= 0 && s_z < src_size.z) {\n";
-      c += "      result = " +
-           src_tensor.ReadWHSB("s_x", "s_y", "s_z", src_batch) + ";\n";
+      c += "    int s_z = Z - args.prepended_z / 4;\n";
+      c += "    if (s_z >= 0 && s_z < args.src_tensor.Slices()) {\n";
+      c += "      result = args.src_tensor.Read(s_x, s_y, s_z);\n";
       c += "    }\n";
     } else {
       c += "    int start_channel = Z * 4;\n";
@@ -128,10 +123,9 @@ std::string GetPaddingCode(
         const auto& s = channels[i];
         c += "    {\n";
         c += "    int channel = start_channel + " + std::to_string(i) + ";\n";
-        c += "    int s_z = channel - prepended.z;\n";
-        c += "    if (s_z >= 0 && s_z < src_channels) {\n";
-        c += "      FLT4 t = " +
-             src_tensor.ReadWHSB("s_x", "s_y", "s_z / 4", src_batch) + ";\n";
+        c += "    int s_z = channel - args.prepended_z;\n";
+        c += "    if (s_z >= 0 && s_z < args.src_tensor.Channels()) {\n";
+        c += "      FLT4 t = args.src_tensor.Read(s_x, s_y, s_z / 4);\n";
         c += "      FLT t_ar[4] = {t.x, t.y, t.z, t.w};\n";
         c += "      result" + s + " = t_ar[s_z % 4];\n";
         c += "    }\n";
@@ -140,10 +134,7 @@ std::string GetPaddingCode(
     }
     c += "  }\n";
   }
-  std::string x_3dcoord =
-      op_def.IsBatchSupported() ? "X * dst_size.w + B" : "X";
-  c += PostProcess(linked_operations, {"result", x_3dcoord, "Y", "Z"});
-  c += "  " + dst_tensor.WriteWHSB("result", "X", "Y", "Z", dst_batch);
+  c += "  args.dst_tensor.Write(result, X, Y, Z);\n";
   c += "}\n";
 
   return c;
@@ -170,24 +161,27 @@ Padding& Padding::operator=(Padding&& kernel) {
 }
 
 absl::Status Padding::Compile(const CreationContext& creation_context) {
-  const auto code =
-      GetPaddingCode(definition_, linked_operations_, attributes_);
+  std::string code = GetPaddingCode(definition_, attributes_, &args_);
+  std::string element_wise_code;
+  RETURN_IF_ERROR(
+      MergeOperations(linked_operations_, &args_, &element_wise_code));
+  RETURN_IF_ERROR(args_.TransformToCLCode(creation_context.device->GetInfo(),
+                                          {{"dst_tensor", element_wise_code}},
+                                          &code));
   return creation_context.cache->GetOrCreateCLKernel(
       code, "main_function", *creation_context.context,
       *creation_context.device, &kernel_);
 }
 
 absl::Status Padding::BindArguments() {
-  kernel_.ResetBindingCounter();
-  RETURN_IF_ERROR(kernel_.SetMemoryAuto(src_[0]->GetMemoryPtr()));
-  RETURN_IF_ERROR(BindArgs(&kernel_, linked_operations_));
-  RETURN_IF_ERROR(kernel_.SetMemoryAuto(dst_[0]->GetMemoryPtrForWriting()));
-  RETURN_IF_ERROR(kernel_.SetBytesAuto(src_[0]->GetWHSB()));
-  RETURN_IF_ERROR(kernel_.SetBytesAuto(src_[0]->Channels()));
-  RETURN_IF_ERROR(kernel_.SetBytesAuto(dst_[0]->GetWHSB()));
-  const auto& prep = attributes_.prepended;
-  RETURN_IF_ERROR(kernel_.SetBytesAuto(int4(prep.w, prep.h, prep.c, prep.b)));
-  return absl::OkStatus();
+  RETURN_IF_ERROR(args_.SetObjectRef("src_tensor", src_[0]));
+  RETURN_IF_ERROR(args_.SetObjectRef("dst_tensor", dst_[0]));
+  RETURN_IF_ERROR(args_.SetInt("prepended_x", attributes_.prepended.w));
+  RETURN_IF_ERROR(args_.SetInt("prepended_y", attributes_.prepended.h));
+  RETURN_IF_ERROR(args_.SetInt("prepended_z", attributes_.prepended.c));
+  RETURN_IF_ERROR(args_.SetInt("prepended_w", attributes_.prepended.b));
+  RETURN_IF_ERROR(SetArguments(linked_operations_, &args_));
+  return args_.Bind(kernel_.kernel());
 }
 
 int3 Padding::GetGridSize() const {

From 3a34bde6520c8d2643337e1db1a986a22943f27f Mon Sep 17 00:00:00 2001
From: "A. Unique TensorFlower" 
Date: Mon, 22 Jun 2020 14:41:02 -0700
Subject: [PATCH 471/867] Go: Update generated wrapper functions for TensorFlow
 ops.

PiperOrigin-RevId: 317738384
Change-Id: I2c74c34bb5fed2809a46f4ecdb95616f2f1c832d
---
 tensorflow/go/op/wrappers.go | 4 ++--
 1 file changed, 2 insertions(+), 2 deletions(-)

diff --git a/tensorflow/go/op/wrappers.go b/tensorflow/go/op/wrappers.go
index ace8e58fdcd..8bceeb78564 100644
--- a/tensorflow/go/op/wrappers.go
+++ b/tensorflow/go/op/wrappers.go
@@ -31850,8 +31850,8 @@ func VarHandleOpSharedName(value string) VarHandleOpAttr {
 
 // VarHandleOpAllowedDevices sets the optional allowed_devices attribute to value.
 //
-// value: The allowed devices containing the resource variable. Set when the output
-// ResourceHandle represents a per-replica/partitioned resource variable.
+// value: DEPRECATED. The allowed devices containing the resource variable. Set when the
+// output ResourceHandle represents a per-replica/partitioned resource variable.
 // If not specified, defaults to <>
 func VarHandleOpAllowedDevices(value []string) VarHandleOpAttr {
 	return func(m optionalAttr) {

From b2f092894012e9c2612cb46c332140f28d91ced2 Mon Sep 17 00:00:00 2001
From: Yanhua Sun 
Date: Mon, 22 Jun 2020 14:53:59 -0700
Subject: [PATCH 472/867] Add DeviceIndex xla op.

DeviceIndex op: given a list of device names, this operation returns the index of the device this op runs.  In the case of XLA, we are not executing on any device, we return the length of the list.

PiperOrigin-RevId: 317740778
Change-Id: I0679aa0adc5508b83502eee0d2044584577ed5b4
---
 .../compiler/jit/mark_for_compilation_pass.cc |  2 +-
 tensorflow/compiler/tf2xla/kernels/BUILD      |  1 +
 .../tf2xla/kernels/device_index_op.cc         | 51 +++++++++++++++++++
 .../base_api/api_def_DeviceIndex.pbtxt        |  6 +++
 .../python/ops/control_flow_ops_test.py       | 20 ++++++++
 5 files changed, 79 insertions(+), 1 deletion(-)
 create mode 100644 tensorflow/compiler/tf2xla/kernels/device_index_op.cc

diff --git a/tensorflow/compiler/jit/mark_for_compilation_pass.cc b/tensorflow/compiler/jit/mark_for_compilation_pass.cc
index 9f5723f4fa4..dc5df94e963 100644
--- a/tensorflow/compiler/jit/mark_for_compilation_pass.cc
+++ b/tensorflow/compiler/jit/mark_for_compilation_pass.cc
@@ -1837,7 +1837,7 @@ absl::flat_hash_map>* GetWhitelistTable() {
       "ConcatOffset", "Const", "MirrorPad", "Pack", "Pad", "PadV2", "Reverse",
       "ReverseV2", "ReverseSequence", "Slice", "Split", "SplitV",
       "StridedSlice", "StridedSliceGrad", "ResourceStridedSliceAssign",
-      "Tile", "Transpose", "InvertPermutation", "Unpack"}}};
+      "Tile", "Transpose", "InvertPermutation", "Unpack", "DeviceIndex"}}};
   // clang-format on
   return result;
 }
diff --git a/tensorflow/compiler/tf2xla/kernels/BUILD b/tensorflow/compiler/tf2xla/kernels/BUILD
index bdaeeafd295..e072225566d 100644
--- a/tensorflow/compiler/tf2xla/kernels/BUILD
+++ b/tensorflow/compiler/tf2xla/kernels/BUILD
@@ -32,6 +32,7 @@ tf_kernel_library(
         "data_format_ops.cc",
         "depthtospace_op.cc",
         "dequantize_op.cc",
+        "device_index_op.cc",
         "diag_op.cc",
         "dynamic_slice_ops.cc",
         "dynamic_stitch_op.cc",
diff --git a/tensorflow/compiler/tf2xla/kernels/device_index_op.cc b/tensorflow/compiler/tf2xla/kernels/device_index_op.cc
new file mode 100644
index 00000000000..ff058f92cd7
--- /dev/null
+++ b/tensorflow/compiler/tf2xla/kernels/device_index_op.cc
@@ -0,0 +1,51 @@
+/* Copyright 2020 The TensorFlow Authors. All Rights Reserved.
+
+Licensed under the Apache License, Version 2.0 (the "License");
+you may not use this file except in compliance with the License.
+You may obtain a copy of the License at
+
+    http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software
+distributed under the License is distributed on an "AS IS" BASIS,
+WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+See the License for the specific language governing permissions and
+limitations under the License.
+==============================================================================*/
+
+#include "absl/container/flat_hash_map.h"
+#include "absl/strings/string_view.h"
+#include "tensorflow/compiler/tf2xla/xla_helpers.h"
+#include "tensorflow/compiler/tf2xla/xla_op_kernel.h"
+#include "tensorflow/compiler/tf2xla/xla_op_registry.h"
+#include "tensorflow/compiler/xla/client/client_library.h"
+#include "tensorflow/compiler/xla/client/lib/arithmetic.h"
+#include "tensorflow/compiler/xla/client/lib/constants.h"
+#include "tensorflow/compiler/xla/client/lib/math.h"
+#include "tensorflow/compiler/xla/client/xla_builder.h"
+#include "tensorflow/core/framework/kernel_def_builder.h"
+
+namespace tensorflow {
+namespace {
+
+class DeviceIndexOp : public XlaOpKernel {
+ public:
+  explicit DeviceIndexOp(OpKernelConstruction* ctx) : XlaOpKernel(ctx) {
+    OP_REQUIRES_OK(ctx, ctx->GetAttr("device_names", &device_names_));
+  }
+
+  void Compile(XlaOpKernelContext* ctx) override {
+    // When compiling we are not executing on any physical device, so we return
+    // a sentinel value (size of the list of devices).
+    ctx->SetOutput(
+        0, xla::ConstantR0(ctx->builder(), device_names_.size()));
+  }
+
+ private:
+  std::vector device_names_;
+};
+
+REGISTER_XLA_OP(Name("DeviceIndex"), DeviceIndexOp);
+
+}  // namespace
+}  // namespace tensorflow
diff --git a/tensorflow/core/api_def/base_api/api_def_DeviceIndex.pbtxt b/tensorflow/core/api_def/base_api/api_def_DeviceIndex.pbtxt
index 9a4e5abd110..87c146910ff 100644
--- a/tensorflow/core/api_def/base_api/api_def_DeviceIndex.pbtxt
+++ b/tensorflow/core/api_def/base_api/api_def_DeviceIndex.pbtxt
@@ -2,4 +2,10 @@ op {
   graph_op_name: "DeviceIndex"
   visibility: HIDDEN
   summary: "Return the index of device the op runs."
+  description: <
Date: Mon, 22 Jun 2020 14:55:44 -0700
Subject: [PATCH 473/867] Special case tfnp ndarrays in some places, and minor
 improvements to tfnp array functions

PiperOrigin-RevId: 317741132
Change-Id: I0c1ba744b0df7a648879da58bd600e81e26184ee
---
 tensorflow/python/eager/BUILD                 |  1 +
 tensorflow/python/eager/function.py           | 19 +++++++++++++++++--
 tensorflow/python/ops/numpy_ops/np_arrays.py  | 17 +++++++++++++++--
 .../python/ops/numpy_ops/np_interop_test.py   |  4 +++-
 4 files changed, 36 insertions(+), 5 deletions(-)

diff --git a/tensorflow/python/eager/BUILD b/tensorflow/python/eager/BUILD
index af5f3d16408..f51bd97e488 100644
--- a/tensorflow/python/eager/BUILD
+++ b/tensorflow/python/eager/BUILD
@@ -560,6 +560,7 @@ py_library(
         "//tensorflow/python:graph_to_function_def",
         "//tensorflow/python:pywrap_tf_session",
         "//tensorflow/python:util",
+        "//tensorflow/python/ops/numpy_ops:numpy",
         "//third_party/py/numpy",
         "@six_archive//:six",
     ],
diff --git a/tensorflow/python/eager/function.py b/tensorflow/python/eager/function.py
index 4a08e66b441..e2f5d86fbd2 100644
--- a/tensorflow/python/eager/function.py
+++ b/tensorflow/python/eager/function.py
@@ -81,6 +81,9 @@ from tensorflow.python.util import tf_inspect
 ag_ctx = lazy_loader.LazyLoader(
     "ag_ctx", globals(),
     "tensorflow.python.autograph.core.ag_ctx")
+np_arrays = lazy_loader.LazyLoader(
+    "np_arrays", globals(),
+    "tensorflow.python.ops.numpy_ops.np_arrays")
 
 
 FORWARD_FUNCTION_ATTRIBUTE_NAME = "forward_function_name"
@@ -1484,6 +1487,12 @@ class ConcreteFunction(object):
     self._func_graph = func_graph
     self._captured_inputs = self._func_graph.external_captures
     self._captured_closures = self._func_graph.deferred_external_captures
+    structured_outputs = self._func_graph.structured_outputs
+    self._ndarrays_list = (
+        isinstance(structured_outputs, (list, tuple)) and
+        structured_outputs and
+        all([isinstance(o, np_arrays.ndarray) for o in structured_outputs]))
+    self._ndarray_singleton = isinstance(structured_outputs, np_arrays.ndarray)
 
     # function_spec defines the structured signature.
     self._set_function_spec(function_spec)
@@ -2150,9 +2159,15 @@ class ConcreteFunction(object):
     if self._func_graph.structured_outputs is None:
       return result
 
+    if result:
+      if self._ndarrays_list:
+        return [np_arrays.tensor_to_ndarray(o) for o in result]
+      elif self._ndarray_singleton:
+        return np_arrays.tensor_to_ndarray(result[0])
+
     # Replace outputs with results, skipping over any 'None' values.
-    outputs_list = nest.flatten(self._func_graph.structured_outputs,
-                                expand_composites=True)
+    outputs_list = nest.flatten(
+        self._func_graph.structured_outputs, expand_composites=True)
     j = 0
     for i, o in enumerate(outputs_list):
       if o is not None:
diff --git a/tensorflow/python/ops/numpy_ops/np_arrays.py b/tensorflow/python/ops/numpy_ops/np_arrays.py
index 40ccfa9dc50..fd26318bea9 100644
--- a/tensorflow/python/ops/numpy_ops/np_arrays.py
+++ b/tensorflow/python/ops/numpy_ops/np_arrays.py
@@ -149,10 +149,23 @@ class ndarray(composite_tensor.CompositeTensor):  # pylint: disable=invalid-name
     if dtype and dtype != buffer.dtype:
       buffer = array_ops.bitcast(buffer, dtype)
     self._data = buffer
+    self._type_spec_internal = None
+
+  @classmethod
+  def from_tensor(cls, tensor):
+    o = cls.__new__(cls, None)
+    # pylint: disable=protected-access
+    o._data = tensor
+    o._type_spec_internal = None
+    # pylint: enable=protected-access
+    return o
 
   @property
   def _type_spec(self):
-    return NdarraySpec(type_spec.type_spec_from_value(self._data))
+    if self._type_spec_internal is None:
+      self._type_spec_internal = NdarraySpec(
+          type_spec.type_spec_from_value(self._data))
+    return self._type_spec_internal
 
   @property
   def data(self):
@@ -299,7 +312,7 @@ class ndarray(composite_tensor.CompositeTensor):  # pylint: disable=invalid-name
 
 
 def tensor_to_ndarray(tensor):
-  return ndarray(tensor._shape_tuple(), dtype=tensor.dtype, buffer=tensor)  # pylint: disable=protected-access
+  return ndarray.from_tensor(tensor)
 
 
 def ndarray_to_tensor(arr, dtype=None, name=None, as_ref=False):
diff --git a/tensorflow/python/ops/numpy_ops/np_interop_test.py b/tensorflow/python/ops/numpy_ops/np_interop_test.py
index 9580b787202..5c7560b12d9 100644
--- a/tensorflow/python/ops/numpy_ops/np_interop_test.py
+++ b/tensorflow/python/ops/numpy_ops/np_interop_test.py
@@ -64,7 +64,9 @@ class InteropTest(test.TestCase):
 
     dx, dy = t.gradient([xx, yy], [x, y])
 
-    # TODO(nareshmodi): Gradient tape returns tensors. Is it possible to rewrap?
+    # # TODO(nareshmodi): Figure out a way to rewrap ndarray as tensors.
+    # self.assertIsInstance(dx, np_arrays.ndarray)
+    # self.assertIsInstance(dy, np_arrays.ndarray)
     self.assertAllClose(dx, 2.0)
     self.assertAllClose(dy, 3.0)
 

From 00acf333e283661dd1c3a310b124da86b6ba579a Mon Sep 17 00:00:00 2001
From: "A. Unique TensorFlower" 
Date: Mon, 22 Jun 2020 15:03:37 -0700
Subject: [PATCH 474/867] [Profiler] Add eager-mode related stats and advices
 to the Overview Page.

PiperOrigin-RevId: 317742833
Change-Id: Ib78b44b673b29dd3ff0b81fa00068f64198feb30
---
 .../convert/op_stats_to_overview_page.cc      | 40 +++++++++++++++++++
 .../convert/op_stats_to_overview_page.h       | 10 +++++
 .../profiler/protobuf/overview_page.proto     | 11 ++++-
 3 files changed, 60 insertions(+), 1 deletion(-)

diff --git a/tensorflow/core/profiler/convert/op_stats_to_overview_page.cc b/tensorflow/core/profiler/convert/op_stats_to_overview_page.cc
index 96bbcc24fff..a92902b6cf7 100644
--- a/tensorflow/core/profiler/convert/op_stats_to_overview_page.cc
+++ b/tensorflow/core/profiler/convert/op_stats_to_overview_page.cc
@@ -130,11 +130,13 @@ void SetCommonRecommendation(absl::string_view input_classification,
                              absl::string_view output_statement,
                              HardwareType hardware_type,
                              absl::string_view tf_function_statement_html,
+                             absl::string_view eager_statement_html,
                              OverviewPageRecommendation* re) {
   re->set_bottleneck(std::string(input_classification));
   re->set_statement(std::string(input_statement));
   re->set_output_statement(std::string(output_statement));
   re->set_tf_function_statement_html(std::string(tf_function_statement_html));
+  re->set_eager_statement_html(std::string(eager_statement_html));
   ComputeHostTips(re);
   ComputeDeviceTips(hardware_type, re);
   ComputeDocumentationTips(re);
@@ -188,13 +190,26 @@ OverviewPageAnalysis ComputeAnalysisResult(const OpStats& op_stats) {
       SafeDivide(
           op_stats.device_op_metrics_db().precision_stats().compute_32bit_ps(),
           total_device_compute_ps));
+
   uint64 num_host_tf_ops = 0;
+  uint64 total_host_op_time_ps_exclude_idle = 0;
+  uint64 eager_host_op_time_ps = 0;
   for (const OpMetrics& metrics : op_stats.host_op_metrics_db().metrics_db()) {
     num_host_tf_ops += metrics.occurrences();
+    if (!IsIdleOp(metrics)) {
+      total_host_op_time_ps_exclude_idle += metrics.self_time_ps();
+      if (metrics.is_eager()) eager_host_op_time_ps += metrics.self_time_ps();
+    }
   }
   uint64 num_device_tf_ops = 0;
+  uint64 total_device_op_time_ps_exclude_idle = 0;
+  uint64 eager_device_op_time_ps = 0;
   for (const OpMetrics& metrics : device_tf_op_metrics_db.metrics_db()) {
     num_device_tf_ops += metrics.occurrences();
+    if (!IsIdleOp(metrics)) {
+      total_device_op_time_ps_exclude_idle += metrics.self_time_ps();
+      if (metrics.is_eager()) eager_device_op_time_ps += metrics.self_time_ps();
+    }
   }
   uint64 num_total_tf_ops = num_host_tf_ops + num_device_tf_ops;
   analysis.set_host_tf_op_percent(
@@ -202,6 +217,12 @@ OverviewPageAnalysis ComputeAnalysisResult(const OpStats& op_stats) {
   analysis.set_device_tf_op_percent(
       100.0 * SafeDivide(num_device_tf_ops, num_total_tf_ops));
   analysis.set_host_trace_level(op_stats.run_environment().host_trace_level());
+  analysis.set_host_op_time_eager_percent(
+      100.0 *
+      SafeDivide(eager_host_op_time_ps, total_host_op_time_ps_exclude_idle));
+  analysis.set_device_op_time_eager_percent(
+      100.0 * SafeDivide(eager_device_op_time_ps,
+                         total_device_op_time_ps_exclude_idle));
   return analysis;
 }
 
@@ -279,6 +300,22 @@ std::string TfFunctionRecommendationHtml(const TfFunctionDb& tf_function_db) {
                       ") due to either retracing or eager execution.");
 }
 
+std::string EagerRecommendationHtml(double host_op_time_eager_percent,
+                                    double device_op_time_eager_percent) {
+  std::string recommendation = "";
+  if (host_op_time_eager_percent > kEagerReportThresholdInPercent)
+    absl::StrAppend(&recommendation, host_op_time_eager_percent,
+                    "% of Op time on the host used eager execution. ");
+  if (device_op_time_eager_percent > kEagerReportThresholdInPercent)
+    absl::StrAppend(&recommendation, device_op_time_eager_percent,
+                    "% of Op time on the device used eager execution. ");
+  if (!recommendation.empty())
+    absl::StrAppend(&recommendation, "Performance could be improved with ",
+                    AnchorElement("https://www.tensorflow.org/guide/function",
+                                  "tf.function."));
+  return recommendation;
+}
+
 OverviewPage ConvertOpStatsToOverviewPage(const OpStats& op_stats,
                                           HardwareType hardware_type) {
   OverviewPage overview_page;
@@ -295,6 +332,9 @@ OverviewPage ConvertOpStatsToOverviewPage(const OpStats& op_stats,
   SetCommonRecommendation(
       bottleneck.input_classification(), bottleneck.input_statement(), "",
       hardware_type, TfFunctionRecommendationHtml(op_stats.tf_function_db()),
+      EagerRecommendationHtml(
+          overview_page.analysis().host_op_time_eager_percent(),
+          overview_page.analysis().device_op_time_eager_percent()),
       overview_page.mutable_recommendation());
   PopulateOverviewDiagnostics(op_stats, overview_page.mutable_diagnostics());
   return overview_page;
diff --git a/tensorflow/core/profiler/convert/op_stats_to_overview_page.h b/tensorflow/core/profiler/convert/op_stats_to_overview_page.h
index 098185b8feb..0d49ae492fc 100644
--- a/tensorflow/core/profiler/convert/op_stats_to_overview_page.h
+++ b/tensorflow/core/profiler/convert/op_stats_to_overview_page.h
@@ -32,11 +32,17 @@ namespace profiler {
 // the tf-functions profiled.
 const double kTfFunctionReportThresholdInPercent = 20;
 
+// Reports eager-mode optimization opportunity in the Overview Page if the
+// percent of Op time on host (or device) that is spent on eager mode is over
+// this threshold.
+const double kEagerReportThresholdInPercent = 10;
+
 void SetCommonRecommendation(absl::string_view input_classification,
                              absl::string_view input_statement,
                              absl::string_view output_statement,
                              HardwareType hardware_type,
                              absl::string_view tf_function_statement_html,
+                             absl::string_view eager_statement_html,
                              OverviewPageRecommendation* re);
 
 OverviewPageRecommendation ComputeGenericRecommendation(
@@ -54,6 +60,10 @@ OverviewPage ConvertOpStatsToOverviewPage(const OpStats& op_stats,
 // Returns a html which provides tf-function related recommendation.
 std::string TfFunctionRecommendationHtml(const TfFunctionDb& tf_function_db);
 
+// Returns a html which provides eager-mode related recommendation.
+std::string EagerRecommendationHtml(double host_op_time_eager_percent,
+                                    double device_op_time_eager_percent);
+
 }  // namespace profiler
 }  // namespace tensorflow
 
diff --git a/tensorflow/core/profiler/protobuf/overview_page.proto b/tensorflow/core/profiler/protobuf/overview_page.proto
index cbef05d4d9f..5621ad92a0d 100644
--- a/tensorflow/core/profiler/protobuf/overview_page.proto
+++ b/tensorflow/core/profiler/protobuf/overview_page.proto
@@ -49,6 +49,12 @@ message OverviewPageAnalysis {
   double device_tf_op_percent = 12;
   // Host trace level.
   uint32 host_trace_level = 13;
+  // Percentage of TF-op execution time on the host (excluding the idle time)
+  // that are in eager mode.
+  double host_op_time_eager_percent = 14;
+  // Percentage of TF-op execution time on the device (excluding the idle time)
+  // that are in eager mode.
+  double device_op_time_eager_percent = 15;
 }
 
 // Overview result for a performance tip to users.
@@ -87,8 +93,11 @@ message OverviewPageRecommendation {
   // A statement for output that recommends the next steps for investigating the
   // bottleneck.
   string output_statement = 9;
+  // A statement that recommends the next steps for investigating eager-mode
+  // related bottleneck (it is a html so that it can link to other tools/docs.)
+  string eager_statement_html = 12;
   // A statement that recommends the next steps for investigating tf-function
-  // related bottleneck (it is a html so that it can link to other tools/docs.
+  // related bottleneck (it is a html so that it can link to other tools/docs.)
   string tf_function_statement_html = 10;
   // A list of tips for improving host performance.
   repeated OverviewPageTip host_tips = 3;

From 8c08608ff87d56de5ea5351414f0aadac458dba7 Mon Sep 17 00:00:00 2001
From: "A. Unique TensorFlower" 
Date: Mon, 22 Jun 2020 15:04:34 -0700
Subject: [PATCH 475/867] fix the numpy version to 1.16.0

PiperOrigin-RevId: 317743030
Change-Id: I8ac4e0c30cb67cfc611d351ba405e2b94e57c04a
---
 tensorflow/tools/ci_build/release/common_win.bat | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/tensorflow/tools/ci_build/release/common_win.bat b/tensorflow/tools/ci_build/release/common_win.bat
index 0918f4c43b0..fa577fcfc33 100644
--- a/tensorflow/tools/ci_build/release/common_win.bat
+++ b/tensorflow/tools/ci_build/release/common_win.bat
@@ -30,7 +30,7 @@ SET PATH=%PATH%;C:\%PYTHON_DIRECTORY%
 %PIP_EXE% install future>=0.17.1 --no-deps
 %PIP_EXE% install --ignore-installed --force-reinstall --upgrade tf-estimator-nightly --no-deps
 %PIP_EXE% install tb-nightly --no-deps
-%PIP_EXE% install numpy --upgrade --no-deps
+%PIP_EXE% install numpy==1.16.0 --upgrade --no-deps
 %PIP_EXE% install opt_einsum --upgrade
 %PIP_EXE% install pandas --upgrade --no-deps
 %PIP_EXE% install protobuf --upgrade --no-deps

From cf15e5cf153e02dc96d87374c26d4aa21df80f21 Mon Sep 17 00:00:00 2001
From: Zhenyu Tan 
Date: Mon, 22 Jun 2020 15:11:04 -0700
Subject: [PATCH 476/867] Add distribution test for image preprocessing.

PiperOrigin-RevId: 317744390
Change-Id: I62557f99761043c3e3c83de4f4faee0cb66d94a8
---
 .../python/keras/layers/preprocessing/BUILD   | 16 +++++
 .../preprocessing/image_preprocessing.py      | 12 ++++
 .../image_preprocessing_distribution_test.py  | 67 +++++++++++++++++++
 3 files changed, 95 insertions(+)
 create mode 100644 tensorflow/python/keras/layers/preprocessing/image_preprocessing_distribution_test.py

diff --git a/tensorflow/python/keras/layers/preprocessing/BUILD b/tensorflow/python/keras/layers/preprocessing/BUILD
index af7f6392219..6916712d52c 100644
--- a/tensorflow/python/keras/layers/preprocessing/BUILD
+++ b/tensorflow/python/keras/layers/preprocessing/BUILD
@@ -361,6 +361,22 @@ distribute_py_test(
     ],
 )
 
+distribute_py_test(
+    name = "image_preprocessing_distribution_test",
+    srcs = ["image_preprocessing_distribution_test.py"],
+    main = "image_preprocessing_distribution_test.py",
+    python_version = "PY3",
+    tags = [
+        "multi_and_single_gpu",
+    ],
+    deps = [
+        ":category_crossing",
+        "//tensorflow/python/distribute:combinations",
+        "//tensorflow/python/distribute:strategy_combinations",
+        "//tensorflow/python/keras",
+    ],
+)
+
 tf_py_test(
     name = "discretization_test",
     size = "small",
diff --git a/tensorflow/python/keras/layers/preprocessing/image_preprocessing.py b/tensorflow/python/keras/layers/preprocessing/image_preprocessing.py
index 074652a0b79..dd741c8c72c 100644
--- a/tensorflow/python/keras/layers/preprocessing/image_preprocessing.py
+++ b/tensorflow/python/keras/layers/preprocessing/image_preprocessing.py
@@ -1306,6 +1306,18 @@ class _RandomGenerator(stateful_random_ops.Generator):
   numbers.
   """
 
+  # TODO(b/157995497): Temporarily use primary variable handle inside cross
+  # replica context.
+  @property
+  def state(self):
+    """The internal state of the RNG."""
+    state_var = self._state_var
+    try:
+      _ = getattr(state_var, 'handle')
+      return state_var
+    except ValueError:
+      return state_var.values[0]
+
   def _create_variable(self, *args, **kwargs):
     # This function does the same thing as the base class's namesake, except
     # that it skips the distribution-strategy check. When we are inside a
diff --git a/tensorflow/python/keras/layers/preprocessing/image_preprocessing_distribution_test.py b/tensorflow/python/keras/layers/preprocessing/image_preprocessing_distribution_test.py
new file mode 100644
index 00000000000..0b93c1d57c6
--- /dev/null
+++ b/tensorflow/python/keras/layers/preprocessing/image_preprocessing_distribution_test.py
@@ -0,0 +1,67 @@
+# Copyright 2019 The TensorFlow Authors. All Rights Reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+# ==============================================================================
+"""Tests for keras.layers.preprocessing.normalization."""
+
+from __future__ import absolute_import
+from __future__ import division
+from __future__ import print_function
+
+import numpy as np
+
+from tensorflow.python import keras
+from tensorflow.python.data.ops import dataset_ops
+from tensorflow.python.distribute import combinations
+from tensorflow.python.distribute import strategy_combinations
+from tensorflow.python.framework import dtypes
+from tensorflow.python.keras import keras_parameterized
+from tensorflow.python.keras.layers.preprocessing import image_preprocessing
+from tensorflow.python.keras.layers.preprocessing import preprocessing_test_utils
+from tensorflow.python.platform import test
+
+
+@combinations.generate(
+    combinations.combine(
+        distribution=strategy_combinations.all_strategies,
+        mode=["eager", "graph"]))
+class ImagePreprocessingDistributionTest(
+    keras_parameterized.TestCase,
+    preprocessing_test_utils.PreprocessingLayerTest):
+
+  def test_distribution(self, distribution):
+    np_images = np.random.random((1000, 32, 32, 3)).astype(np.float32)
+    image_dataset = dataset_ops.Dataset.from_tensor_slices(np_images).batch(
+        32, drop_remainder=True)
+
+    with distribution.scope():
+      input_data = keras.Input(shape=(32, 32, 3), dtype=dtypes.float32)
+      image_preprocessor = keras.Sequential([
+          image_preprocessing.Resizing(height=256, width=256),
+          image_preprocessing.RandomCrop(height=224, width=224),
+          image_preprocessing.RandomTranslation(.1, .1),
+          image_preprocessing.RandomRotation(.2),
+          image_preprocessing.RandomFlip(),
+          image_preprocessing.RandomZoom(.2, .2)])
+      preprocessed_image = image_preprocessor(input_data)
+      flatten_layer = keras.layers.Flatten(data_format="channels_last")
+      output = flatten_layer(preprocessed_image)
+      cls_layer = keras.layers.Dense(units=1, activation="sigmoid")
+      output = cls_layer(output)
+      model = keras.Model(inputs=input_data, outputs=preprocessed_image)
+    model.compile(loss="binary_crossentropy")
+    _ = model.predict(image_dataset)
+
+
+if __name__ == "__main__":
+  test.main()

From 763a82c83f492cbf4ee387a8e1e24ce66ef5535f Mon Sep 17 00:00:00 2001
From: Raman Sarokin 
Date: Mon, 22 Jun 2020 15:22:15 -0700
Subject: [PATCH 477/867] FullyConnected converted to new style. Buffer
 implements GPUObject.

PiperOrigin-RevId: 317746356
Change-Id: I715d37a924298e54e9fe68e5c7b719ab592237a1
---
 tensorflow/lite/delegates/gpu/cl/BUILD        |  1 +
 tensorflow/lite/delegates/gpu/cl/arguments.cc |  3 +
 tensorflow/lite/delegates/gpu/cl/buffer.cc    | 38 ++++++++
 tensorflow/lite/delegates/gpu/cl/buffer.h     | 19 +++-
 .../gpu/cl/kernels/fully_connected.cc         | 87 +++++++++----------
 .../gpu/cl/kernels/fully_connected.h          | 25 ++++--
 6 files changed, 120 insertions(+), 53 deletions(-)

diff --git a/tensorflow/lite/delegates/gpu/cl/BUILD b/tensorflow/lite/delegates/gpu/cl/BUILD
index e6e7e4747c4..fda57ce6da0 100644
--- a/tensorflow/lite/delegates/gpu/cl/BUILD
+++ b/tensorflow/lite/delegates/gpu/cl/BUILD
@@ -81,6 +81,7 @@ cc_library(
     deps = [
         ":cl_command_queue",
         ":cl_context",
+        ":gpu_object",
         ":opencl_wrapper",
         ":util",
         "//tensorflow/lite/delegates/gpu/common:status",
diff --git a/tensorflow/lite/delegates/gpu/cl/arguments.cc b/tensorflow/lite/delegates/gpu/cl/arguments.cc
index 6955681a366..d26c869957c 100644
--- a/tensorflow/lite/delegates/gpu/cl/arguments.cc
+++ b/tensorflow/lite/delegates/gpu/cl/arguments.cc
@@ -746,6 +746,9 @@ absl::Status Arguments::ResolveSelectorsPass(
       size_t close_bracket_pos;
       RETURN_IF_ERROR(ParseArgsInsideBrackets(*code, next_position,
                                               &close_bracket_pos, &args));
+      for (auto& arg : args) {
+        RETURN_IF_ERROR(ResolveSelectorsPass({}, &arg));
+      }
       std::string patch;
       RETURN_IF_ERROR(ResolveSelector(linkables, object_name, selector_name,
                                       args, template_args, &patch));
diff --git a/tensorflow/lite/delegates/gpu/cl/buffer.cc b/tensorflow/lite/delegates/gpu/cl/buffer.cc
index 207cdec5122..68e85593e5d 100644
--- a/tensorflow/lite/delegates/gpu/cl/buffer.cc
+++ b/tensorflow/lite/delegates/gpu/cl/buffer.cc
@@ -44,6 +44,38 @@ absl::Status CreateBuffer(size_t size_in_bytes, bool gpu_read_only,
 }
 }  // namespace
 
+GPUResources BufferDescriptor::GetGPUResources(AccessType access_type) const {
+  GPUResources resources;
+  GPUBufferDescriptor desc;
+  desc.data_type = element_type;
+  desc.access_type = access_type;
+  desc.element_size = element_size;
+  resources.buffers.push_back({"buffer", desc});
+  return resources;
+}
+
+absl::Status BufferDescriptor::PerformSelector(
+    const std::string& selector, const std::vector& args,
+    const std::vector& template_args, std::string* result) const {
+  if (selector == "Read") {
+    return PerformReadSelector(args, result);
+  } else {
+    return absl::NotFoundError(absl::StrCat(
+        "BufferDescriptor don't have selector with name - ", selector));
+  }
+}
+
+absl::Status BufferDescriptor::PerformReadSelector(
+    const std::vector& args, std::string* result) const {
+  if (args.size() != 1) {
+    return absl::NotFoundError(
+        absl::StrCat("BufferDescriptor Read require one argument, but ",
+                     args.size(), " was passed"));
+  }
+  *result = absl::StrCat("buffer[", args[0], "]");
+  return absl::OkStatus();
+}
+
 Buffer::Buffer(cl_mem buffer, size_t size_in_bytes)
     : buffer_(buffer), size_(size_in_bytes) {}
 
@@ -71,6 +103,12 @@ void Buffer::Release() {
   }
 }
 
+GPUResourcesWithValue Buffer::GetGPUResources(AccessType access_type) const {
+  GPUResourcesWithValue resources;
+  resources.buffers.push_back({"buffer", buffer_});
+  return resources;
+}
+
 absl::Status CreateReadOnlyBuffer(size_t size_in_bytes, CLContext* context,
                                   Buffer* result) {
   return CreateBuffer(size_in_bytes, true, nullptr, context, result);
diff --git a/tensorflow/lite/delegates/gpu/cl/buffer.h b/tensorflow/lite/delegates/gpu/cl/buffer.h
index 84c3292084b..771aae9e002 100644
--- a/tensorflow/lite/delegates/gpu/cl/buffer.h
+++ b/tensorflow/lite/delegates/gpu/cl/buffer.h
@@ -20,6 +20,7 @@ limitations under the License.
 #include "absl/types/span.h"
 #include "tensorflow/lite/delegates/gpu/cl/cl_command_queue.h"
 #include "tensorflow/lite/delegates/gpu/cl/cl_context.h"
+#include "tensorflow/lite/delegates/gpu/cl/gpu_object.h"
 #include "tensorflow/lite/delegates/gpu/cl/opencl_wrapper.h"
 #include "tensorflow/lite/delegates/gpu/cl/util.h"
 #include "tensorflow/lite/delegates/gpu/common/status.h"
@@ -28,9 +29,23 @@ namespace tflite {
 namespace gpu {
 namespace cl {
 
+struct BufferDescriptor : public GPUObjectDescriptor {
+  DataType element_type;  // FLOAT32 or FLOAT16
+  int element_size;
+
+  absl::Status PerformSelector(const std::string& selector,
+                               const std::vector& args,
+                               const std::vector& template_args,
+                               std::string* result) const override;
+
+  GPUResources GetGPUResources(AccessType access_type) const override;
+  absl::Status PerformReadSelector(const std::vector& args,
+                                   std::string* result) const;
+};
+
 // Buffer represent linear GPU data storage with arbitrary data format.
 // Buffer is moveable but not copyable.
-class Buffer {
+class Buffer : public GPUObject {
  public:
   Buffer() {}  // just for using Buffer as a class members
   Buffer(cl_mem buffer, size_t size_in_bytes);
@@ -57,6 +72,8 @@ class Buffer {
   template 
   absl::Status ReadData(CLCommandQueue* queue, std::vector* result) const;
 
+  GPUResourcesWithValue GetGPUResources(AccessType access_type) const override;
+
  private:
   void Release();
 
diff --git a/tensorflow/lite/delegates/gpu/cl/kernels/fully_connected.cc b/tensorflow/lite/delegates/gpu/cl/kernels/fully_connected.cc
index f93648f82fc..bc287ec2fee 100644
--- a/tensorflow/lite/delegates/gpu/cl/kernels/fully_connected.cc
+++ b/tensorflow/lite/delegates/gpu/cl/kernels/fully_connected.cc
@@ -31,14 +31,15 @@ namespace {
 // Good results for ~1024 x 1024 sizes, for other can be written more
 // optimized shaders
 
-std::string GetFullyConnectedKernelCode(
-    const OperationDef& op_def, const LinearStorage& biases,
-    const std::vector& linked_operations,
-    const int3& work_group_size) {
-  TensorCodeGenerator src_tensor("src_data", WHSPoint{"1", "1", "depthes.x"},
-                                 op_def.src_tensors[0]);
-  TensorCodeGenerator dst_tensor("dst_data", WHSPoint{"1", "1", "depthes.y"},
-                                 op_def.dst_tensors[0]);
+std::string GetFullyConnectedKernelCode(const OperationDef& op_def,
+                                        const int3& work_group_size,
+                                        Arguments* args) {
+  args->AddObjectRef(
+      "src_tensor", AccessType::READ,
+      absl::make_unique(op_def.src_tensors[0]));
+  args->AddObjectRef(
+      "dst_tensor", AccessType::WRITE,
+      absl::make_unique(op_def.dst_tensors[0]));
 
   std::string c = GetCommonDefines(op_def.precision);
   switch (op_def.precision) {
@@ -54,21 +55,16 @@ std::string GetFullyConnectedKernelCode(
   const std::string wg_x = std::to_string(work_group_size.x);
   const std::string wg_y = std::to_string(work_group_size.y);
   c += "__kernel void main_function(\n";
-  c += src_tensor.GetDeclaration(AccessType::READ) + ",\n";
-  c += "    __global FLT16* filters,      \n";
-  c += biases.GetDeclaration();
-  c += GetArgsDeclaration(linked_operations);
-  c += dst_tensor.GetDeclaration(AccessType::WRITE) + ",\n";
-  c += "    int2 depthes                  \n";
-  c += ") {\n";
+  c += "$0) {\n";
   c += "  int gid = get_global_id(0);\n";
-  c += "  bool inside = gid < depthes.y;\n";
-  c += "  gid = min(gid, depthes.y - 1);\n";
+  c += "  bool inside = gid < args.dst_tensor.Slices();\n";
+  c += "  gid = min(gid, args.dst_tensor.Slices() - 1);\n";
   c += "  int2 tid = (int2)(get_local_id(0), get_local_id(1));\n";
   c += "  ACCUM_FLT4 s = (ACCUM_FLT4)(0.0f);\n";
-  c += "  for (uint c = tid.y; c < depthes.x; c += " + wg_y + ") {\n";
-  c += "    FLT4 v = " + src_tensor.ReadWHS("0", "0", "c") + ";\n";
-  c += "    FLT16 w = filters[c * depthes.y + gid];\n";
+  c += "  for (uint c = tid.y; c < args.src_tensor.Slices(); c += " + wg_y +
+       ") {\n";
+  c += "    FLT4 v = args.src_tensor.Read(0, 0, c);\n";
+  c += "    FLT16 w = args.weights.Read(c * args.dst_tensor.Slices() + gid);\n";
   c += "    s.x += dot(v, w.s0123);\n";
   c += "    s.y += dot(v, w.s4567);\n";
   c += "    s.z += dot(v, w.s89ab);\n";
@@ -81,10 +77,8 @@ std::string GetFullyConnectedKernelCode(
   for (int i = 1; i < work_group_size.y; ++i) {
     c += "    s += temp[tid.x][" + std::to_string(i) + "];\n";
   }
-  c += "    FLT4 r0 = TO_FLT4(s) + " + biases.ReadLinearFLT4("gid") + ";\n";
-  const LinkingContext context{"r0", "0", "0", "gid"};
-  c += PostProcess(linked_operations, context);
-  c += "  " + dst_tensor.WriteWHS("r0", "0", "0", "gid") + "\n";
+  c += "    FLT4 r0 = TO_FLT4(s) + args.biases.Read(gid);\n";
+  c += "    args.dst_tensor.Write(r0, 0, 0, gid);\n";
   c += "  }\n";
   c += "}\n";
 
@@ -97,15 +91,11 @@ FullyConnected::FullyConnected(const OperationDef& definition)
 
 FullyConnected::FullyConnected(FullyConnected&& kernel)
     : GPUOperation(std::move(kernel)),
-      weights_(std::move(kernel.weights_)),
-      biases_(std::move(kernel.biases_)),
       kernel_(std::move(kernel.kernel_)),
       work_group_size_(kernel.work_group_size_) {}
 
 FullyConnected& FullyConnected::operator=(FullyConnected&& kernel) {
   if (this != &kernel) {
-    weights_ = std::move(kernel.weights_);
-    biases_ = std::move(kernel.biases_);
     kernel_ = std::move(kernel.kernel_);
     std::swap(work_group_size_, kernel.work_group_size_);
     GPUOperation::operator=(std::move(kernel));
@@ -120,8 +110,14 @@ absl::Status FullyConnected::Compile(const CreationContext& creation_context) {
   do {
     work_group_size_ = {wg_width, wg_height, 1};
     wg_width /= 2;
-    const auto code = GetFullyConnectedKernelCode(
-        definition_, biases_, linked_operations_, work_group_size_);
+    std::string code =
+        GetFullyConnectedKernelCode(definition_, work_group_size_, &args_);
+    std::string element_wise_code;
+    RETURN_IF_ERROR(
+        MergeOperations(linked_operations_, &args_, &element_wise_code));
+    RETURN_IF_ERROR(args_.TransformToCLCode(creation_context.device->GetInfo(),
+                                            {{"dst_tensor", element_wise_code}},
+                                            &code));
     auto status = creation_context.cache->GetOrCreateCLKernel(
         code, "main_function", *creation_context.context,
         *creation_context.device, &kernel_);
@@ -138,14 +134,10 @@ absl::Status FullyConnected::Compile(const CreationContext& creation_context) {
 }
 
 absl::Status FullyConnected::AddToQueue(CLCommandQueue* queue) {
-  kernel_.ResetBindingCounter();
-  RETURN_IF_ERROR(kernel_.SetMemoryAuto(src_[0]->GetMemoryPtr()));
-  RETURN_IF_ERROR(kernel_.SetMemoryAuto(weights_.GetMemoryPtr()));
-  RETURN_IF_ERROR(kernel_.SetMemoryAuto(biases_.GetMemoryPtr()));
-  RETURN_IF_ERROR(BindArgs(&kernel_, linked_operations_));
-  RETURN_IF_ERROR(kernel_.SetMemoryAuto(dst_[0]->GetMemoryPtrForWriting()));
-  RETURN_IF_ERROR(
-      kernel_.SetBytesAuto(int2(src_[0]->Slices(), dst_[0]->Slices())));
+  RETURN_IF_ERROR(args_.SetObjectRef("src_tensor", src_[0]));
+  RETURN_IF_ERROR(args_.SetObjectRef("dst_tensor", dst_[0]));
+  RETURN_IF_ERROR(SetArguments(linked_operations_, &args_));
+  RETURN_IF_ERROR(args_.Bind(kernel_.kernel()));
   return queue->DispatchImplicit(kernel_, {dst_[0]->Slices(), 1, 1},
                                  work_group_size_);
 }
@@ -157,13 +149,18 @@ absl::Status CreateFullyConnected(const CreationContext& creation_context,
   *result = FullyConnected(definition);
   RETURN_IF_ERROR(
       result->UploadWeights(attr.weights, creation_context.context));
-  LinearStorageCreateInfo create_info;
-  create_info.storage_type = LinearStorageType::TEXTURE_2D;
-  create_info.data_type = definition.GetDataType();
-  create_info.name = "biases";
-  create_info.aligned_size = attr.weights.shape.o;
-  RETURN_IF_ERROR(CreateLinearStorage(
-      create_info, attr.bias, creation_context.context, &result->biases_));
+
+  TensorLinearDescriptor desc;
+  desc.storage_type = LinearStorageType::TEXTURE_2D;
+  desc.element_type = definition.GetDataType();
+
+  LinearStorage lt;
+  RETURN_IF_ERROR(
+      CreateLinearStorage(desc, attr.bias, creation_context.context, <));
+  result->args_.AddObject("biases", AccessType::READ,
+                          absl::make_unique(std::move(lt)),
+                          absl::make_unique(desc));
+
   return absl::OkStatus();
 }
 
diff --git a/tensorflow/lite/delegates/gpu/cl/kernels/fully_connected.h b/tensorflow/lite/delegates/gpu/cl/kernels/fully_connected.h
index 0be5288ed3a..83490b281ab 100644
--- a/tensorflow/lite/delegates/gpu/cl/kernels/fully_connected.h
+++ b/tensorflow/lite/delegates/gpu/cl/kernels/fully_connected.h
@@ -61,8 +61,6 @@ class FullyConnected : public GPUOperation {
   void RearrangeWeights(const tflite::gpu::Tensor& weights,
                         absl::Span dst);
 
-  Buffer weights_;
-  LinearStorage biases_;
   CLKernel kernel_;
   int3 work_group_size_ = int3(0, 0, 0);
 };
@@ -78,17 +76,30 @@ absl::Status FullyConnected::UploadWeights(
 
   const int float4_size = f32_weights ? 16 : 8;
 
-  if (definition_.GetDataType() == DataType::FLOAT32) {
+  BufferDescriptor desc;
+  desc.element_type = f32_weights ? DataType::FLOAT32 : DataType::FLOAT16;
+  desc.element_size = 16;
+
+  Buffer weights_buffer;
+  if (f32_weights) {
     std::vector gpu_data(dst_depth * src_depth * 4);
     RearrangeWeights(weights, absl::MakeSpan(gpu_data));
-    return CreateReadOnlyBuffer(float4_size * elements_count, gpu_data.data(),
-                                context, &weights_);
+    RETURN_IF_ERROR(CreateReadOnlyBuffer(float4_size * elements_count,
+                                         gpu_data.data(), context,
+                                         &weights_buffer));
   } else {
     std::vector gpu_data(dst_depth * src_depth * 4);
     RearrangeWeights(weights, absl::MakeSpan(gpu_data));
-    return CreateReadOnlyBuffer(float4_size * elements_count, gpu_data.data(),
-                                context, &weights_);
+    RETURN_IF_ERROR(CreateReadOnlyBuffer(float4_size * elements_count,
+                                         gpu_data.data(), context,
+                                         &weights_buffer));
   }
+
+  args_.AddObject("weights", AccessType::READ,
+                  absl::make_unique(std::move(weights_buffer)),
+                  absl::make_unique(desc));
+
+  return absl::OkStatus();
 }
 
 template 

From 090f88dcb4451aabcf2792b3caf7022e673baccb Mon Sep 17 00:00:00 2001
From: "A. Unique TensorFlower" 
Date: Mon, 22 Jun 2020 15:22:24 -0700
Subject: [PATCH 478/867] Add tests for code snippets in README.

PiperOrigin-RevId: 317746384
Change-Id: I437965917d7d392b906cecd145f87ced4635a592
---
 tensorflow/python/ops/numpy_ops/BUILD         |   4 +-
 .../python/ops/numpy_ops/np_interop_test.py   | 162 +++++++++++-------
 2 files changed, 99 insertions(+), 67 deletions(-)

diff --git a/tensorflow/python/ops/numpy_ops/BUILD b/tensorflow/python/ops/numpy_ops/BUILD
index 704e5a27b48..a70e3f3918d 100644
--- a/tensorflow/python/ops/numpy_ops/BUILD
+++ b/tensorflow/python/ops/numpy_ops/BUILD
@@ -112,7 +112,7 @@ cuda_py_test(
     srcs = ["np_interop_test.py"],
     deps = [
         ":numpy",
-        "//tensorflow/python:platform",
-        "@absl_py//absl/testing:parameterized",
+        "//tensorflow:tensorflow_py",
+        "//third_party/py/numpy",
     ],
 )
diff --git a/tensorflow/python/ops/numpy_ops/np_interop_test.py b/tensorflow/python/ops/numpy_ops/np_interop_test.py
index 5c7560b12d9..836bce61494 100644
--- a/tensorflow/python/ops/numpy_ops/np_interop_test.py
+++ b/tensorflow/python/ops/numpy_ops/np_interop_test.py
@@ -18,44 +18,73 @@ from __future__ import absolute_import
 from __future__ import division
 from __future__ import print_function
 
-
 import numpy as onp
+import tensorflow.compat.v2 as tf
+
+import tensorflow.python.ops.numpy_ops as np
 
 
-from tensorflow.python.distribute import distribution_strategy_context
-from tensorflow.python.distribute import mirrored_strategy
-from tensorflow.python.distribute import reduce_util
-from tensorflow.python.eager import backprop
-from tensorflow.python.eager import context
-from tensorflow.python.eager import def_function
-from tensorflow.python.framework import config
-from tensorflow.python.framework import constant_op
-from tensorflow.python.framework import ops
-from tensorflow.python.ops import control_flow_ops
-from tensorflow.python.ops import math_ops
-from tensorflow.python.ops.numpy_ops import np_array_ops
-from tensorflow.python.ops.numpy_ops import np_arrays
-from tensorflow.python.ops.numpy_ops import np_math_ops
-from tensorflow.python.platform import test
+# Tests for code snippet put in README.md
+class ReadmeTest(tf.test.TestCase):
+
+  def testBroadcastAdd(self):
+    x_np = np.ones([2, 1]) + np.ones([1, 2])
+    x_onp = onp.ones([2, 1]) + onp.ones([1, 2])
+    self.assertAllClose(x_onp, x_np)
+
+  def testTypePromotion(self):
+    x_np = np.ones([1, 2], dtype=np.int16) + np.ones([2, 1], dtype=np.uint8)
+    x_onp = np.ones([1, 2], dtype=np.int16) + np.ones([2, 1], dtype=np.uint8)
+    self.assertEqual(x_onp.dtype, x_np.dtype)
+    self.assertAllClose(x_onp, x_np)
+
+  def testTFInterop(self):
+    x_np = np.sum(np.ones([1, 2]) + tf.ones([2, 1]))
+    x_onp = onp.sum(onp.ones([1, 2]) + onp.ones([2, 1]))
+    self.assertAllClose(x_onp, x_np)
+
+  def testOnpInterop(self):
+    x_np = onp.sum(np.ones([1, 2]) + onp.ones([2, 1]))
+    x_onp = onp.sum(onp.ones([1, 2]) + onp.ones([2, 1]))
+    self.assertAllClose(x_onp, x_np)
+
+  def testDevice(self):
+    if tf.test.is_gpu_available():
+      with tf.device('GPU:0'):
+        x = np.ones([1, 2])
+      self.assertIn('GPU', tf.convert_to_tensor(x).device)
+    with tf.device('CPU:0'):
+      x = np.ones([1, 2])
+    self.assertIn('CPU', tf.convert_to_tensor(x).device)
+
+  def testFunction(self):
+
+    @tf.function
+    def f(x, y):
+      return np.sum(x + y)
+
+    x_np = f(np.ones([1, 2]), tf.ones([2, 1]))
+    x_onp = onp.sum(onp.ones([1, 2]) + onp.ones([2, 1]))
+    self.assertAllClose(x_onp, x_np)
 
 
-class InteropTest(test.TestCase):
+class InteropTest(tf.test.TestCase):
 
   def setUp(self):
     super(InteropTest, self).setUp()
-    physical_devices = config.list_physical_devices('CPU')
-    configs = config.get_logical_device_configuration(physical_devices[0])
+    physical_devices = tf.config.list_physical_devices('CPU')
+    configs = tf.config.get_logical_device_configuration(physical_devices[0])
     if configs is None:
       logical_devices = [
-          context.LogicalDeviceConfiguration() for _ in range(3)
+          tf.config.LogicalDeviceConfiguration() for _ in range(3)
       ]
-      config.set_logical_device_configuration(physical_devices[0],
-                                              logical_devices)
+      tf.config.set_logical_device_configuration(physical_devices[0],
+                                                 logical_devices)
 
   def testGradientTapeInterop(self):
-    with backprop.GradientTape() as t:
-      x = np_array_ops.asarray(3.0)
-      y = np_array_ops.asarray(2.0)
+    with tf.GradientTape() as t:
+      x = np.asarray(3.0)
+      y = np.asarray(2.0)
 
       t.watch([x, y])
 
@@ -71,30 +100,30 @@ class InteropTest(test.TestCase):
     self.assertAllClose(dy, 3.0)
 
   def testFunctionInterop(self):
-    x = np_array_ops.asarray(3.0)
-    y = np_array_ops.asarray(2.0)
+    x = np.asarray(3.0)
+    y = np.asarray(2.0)
 
     add = lambda x, y: x + y
-    add_fn = def_function.function(add)
+    add_fn = tf.function(add)
 
     raw_result = add(x, y)
     fn_result = add_fn(x, y)
 
-    self.assertIsInstance(raw_result, np_arrays.ndarray)
-    self.assertIsInstance(fn_result, np_arrays.ndarray)
+    self.assertIsInstance(raw_result, np.ndarray)
+    self.assertIsInstance(fn_result, np.ndarray)
     self.assertAllClose(raw_result, fn_result)
 
   def testCondInterop(self):
-    x = np_array_ops.asarray(3.0)
+    x = np.asarray(3.0)
 
     def fn(x):
-      x_plus_1 = control_flow_ops.cond(x > 0, lambda: x+1, lambda: x+2)
-      x_plus_2 = control_flow_ops.cond(x < 0, lambda: x+1, lambda: x+2)
+      x_plus_1 = tf.cond(x > 0, lambda: x + 1, lambda: x + 2)
+      x_plus_2 = tf.cond(x < 0, lambda: x + 1, lambda: x + 2)
 
       return x_plus_1, x_plus_2
 
     raw_x_plus_1, raw_x_plus_2 = fn(x)
-    fn_x_plus_1, fn_x_plus_2 = def_function.function(fn)(x)
+    fn_x_plus_1, fn_x_plus_2 = tf.function(fn)(x)
 
     self.assertAllClose(raw_x_plus_1, x + 1)
     self.assertAllClose(raw_x_plus_2, x + 2)
@@ -103,44 +132,45 @@ class InteropTest(test.TestCase):
     self.assertAllClose(fn_x_plus_2, x + 2)
 
   def testWhileInterop(self):
+
     def fn():
-      x = np_array_ops.asarray(0)
+      x = np.asarray(0)
       c = lambda x: x < 10000
       b = lambda x: [x + 1]
-      return control_flow_ops.while_loop_v2(c, b, [x], parallel_iterations=20)
+      return tf.while_loop(c, b, [x], parallel_iterations=20)
 
     self.assertEqual(10000, fn()[0])
-    self.assertEqual(10000, def_function.function(fn)()[0])
+    self.assertEqual(10000, tf.function(fn)()[0])
 
   def testTensorTFNPArrayInterop(self):
-    arr = np_array_ops.asarray(0.)
-    t = constant_op.constant(10.)
+    arr = np.asarray(0.)
+    t = tf.constant(10.)
 
     arr_plus_t = arr + t
     t_plus_arr = t + arr
 
-    self.assertIsInstance(arr_plus_t, ops.Tensor)
-    self.assertIsInstance(t_plus_arr, ops.Tensor)
+    self.assertIsInstance(arr_plus_t, tf.Tensor)
+    self.assertIsInstance(t_plus_arr, tf.Tensor)
     self.assertEqual(10., arr_plus_t.numpy())
     self.assertEqual(10., t_plus_arr.numpy())
 
   def testTensorTFNPOp(self):
-    t = constant_op.constant(10.)
+    t = tf.constant(10.)
 
-    sq = np_math_ops.square(t)
-    self.assertIsInstance(sq, np_arrays.ndarray)
+    sq = np.square(t)
+    self.assertIsInstance(sq, np.ndarray)
     self.assertEqual(100., sq)
 
   def testTFNPArrayTFOpInterop(self):
-    arr = np_array_ops.asarray(10.)
+    arr = np.asarray(10.)
 
     # TODO(nareshmodi): Test more ops.
-    sq = math_ops.square(arr)
-    self.assertIsInstance(sq, ops.Tensor)
+    sq = tf.square(arr)
+    self.assertIsInstance(sq, tf.Tensor)
     self.assertEqual(100., sq.numpy())
 
   def testTFNPArrayNPOpInterop(self):
-    arr = np_array_ops.asarray([10.])
+    arr = np.asarray([10.])
 
     # TODO(nareshmodi): Test more ops.
     sq = onp.square(arr)
@@ -149,47 +179,49 @@ class InteropTest(test.TestCase):
 
     # TODO(nareshmodi): Fails since the autopacking code doesn't use
     # nest.flatten.
+
+
 #   def testAutopacking(self):
-#     arr1 = np_array_ops.asarray(1.)
-#     arr2 = np_array_ops.asarray(2.)
-#     arr3 = np_array_ops.asarray(3.)
+#     arr1 = np.asarray(1.)
+#     arr2 = np.asarray(2.)
+#     arr3 = np.asarray(3.)
 #     t = ops.convert_to_tensor_v2([arr1, arr2, arr3])
 
 #     self.assertEqual(t.numpy(), [1., 2., 3.])
 
   def testDistStratInterop(self):
-    strategy = mirrored_strategy.MirroredStrategy(
+    strategy = tf.distribute.MirroredStrategy(
         devices=['CPU:0', 'CPU:1', 'CPU:2'])
 
-    multiplier = np_array_ops.asarray(5.)
+    multiplier = np.asarray(5.)
 
     with strategy.scope():
-      @def_function.function
+
+      @tf.function
       def run():
-        ctx = distribution_strategy_context.get_replica_context()
-        val = np_array_ops.asarray(ctx.replica_id_in_sync_group)
+        ctx = tf.distribute.get_replica_context()
+        val = np.asarray(ctx.replica_id_in_sync_group)
         return val * multiplier
 
       distributed_values = strategy.run(run)
-      reduced = strategy.reduce(reduce_util.ReduceOp.SUM,
-                                distributed_values, axis=None)
+      reduced = strategy.reduce(
+          tf.distribute.ReduceOp.SUM, distributed_values, axis=None)
 
     values = distributed_values.values
 
     # Note that this should match the number of virtual CPUs.
     self.assertLen(values, 3)
-    self.assertIsInstance(values[0], np_arrays.ndarray)
-    self.assertIsInstance(values[1], np_arrays.ndarray)
-    self.assertIsInstance(values[2], np_arrays.ndarray)
+    self.assertIsInstance(values[0], np.ndarray)
+    self.assertIsInstance(values[1], np.ndarray)
+    self.assertIsInstance(values[2], np.ndarray)
     self.assertAllClose(values[0], 0)
     self.assertAllClose(values[1], 5)
     self.assertAllClose(values[2], 10)
 
     # "strategy.reduce" doesn't rewrap in ndarray.
-    # self.assertIsInstance(reduced, np_arrays.ndarray)
+    # self.assertIsInstance(reduced, np.ndarray)
     self.assertAllClose(reduced, 15)
 
-
 if __name__ == '__main__':
-  ops.enable_eager_execution()
-  test.main()
+  tf.compat.v1.enable_eager_execution()
+  tf.test.main()

From 188ba91f08a44b18ea503bada5cbf7d1f8a4bb1b Mon Sep 17 00:00:00 2001
From: Raman Sarokin 
Date: Mon, 22 Jun 2020 15:23:18 -0700
Subject: [PATCH 479/867] MaxUnpooling converted to new style. Merged 2D/3D
 versions.

PiperOrigin-RevId: 317746537
Change-Id: I69961e9ae1412ccff177a44493020ef97ee0921f
---
 .../delegates/gpu/cl/kernels/max_unpooling.cc | 364 +++++++-----------
 .../delegates/gpu/cl/kernels/max_unpooling.h  |  39 +-
 2 files changed, 139 insertions(+), 264 deletions(-)

diff --git a/tensorflow/lite/delegates/gpu/cl/kernels/max_unpooling.cc b/tensorflow/lite/delegates/gpu/cl/kernels/max_unpooling.cc
index 56109fc713b..dc16837102a 100644
--- a/tensorflow/lite/delegates/gpu/cl/kernels/max_unpooling.cc
+++ b/tensorflow/lite/delegates/gpu/cl/kernels/max_unpooling.cc
@@ -25,178 +25,136 @@ namespace gpu {
 namespace cl {
 namespace {
 
-std::string GetMaxUnpoolingKernelCode(
-    const OperationDef& op_def, const CLDevice& device,
-    const std::vector& linked_operations) {
-  TensorCodeGenerator src("src_data",
-                          WHSPoint{"src_size.x", "src_size.y", "src_size.z"},
-                          op_def.src_tensors[0]);
-  TensorCodeGenerator src_ind(
-      "src_data_indices", WHSPoint{"src_size.x", "src_size.y", "src_size.z"},
-      op_def.src_tensors[1]);
-  TensorCodeGenerator dst("dst_data",
-                          WHSPoint{"dst_size.x", "dst_size.y", "dst_size.z"},
-                          op_def.dst_tensors[0]);
-
-  const auto address_mode = GetFastestZeroMode(device);
+std::string GetMaxUnpoolingKernelCode(const OperationDef& op_def,
+                                      const CLDevice& device, Arguments* args) {
+  auto src_desc = absl::make_unique(op_def.src_tensors[0]);
+  src_desc->SetTextureAddressMode(GetFastestZeroMode(device));
+  if (op_def.IsBatchSupported()) {
+    src_desc->SetStateVar("BatchedWidth", "true");
+  }
+  args->AddObjectRef("src_tensor", AccessType::READ, std::move(src_desc));
+  auto src_ind_desc =
+      absl::make_unique(op_def.src_tensors[1]);
+  src_ind_desc->SetTextureAddressMode(GetFastestZeroMode(device));
+  if (op_def.IsBatchSupported()) {
+    src_ind_desc->SetStateVar("BatchedWidth", "true");
+  }
+  args->AddObjectRef("src_indices", AccessType::READ, std::move(src_ind_desc));
+  auto dst_desc = absl::make_unique(op_def.dst_tensors[0]);
+  if (op_def.IsBatchSupported()) {
+    dst_desc->SetStateVar("BatchedWidth", "true");
+  }
+  args->AddObjectRef("dst_tensor", AccessType::WRITE, std::move(dst_desc));
+  if (op_def.dst_tensors[0].HasAxis(Axis::WIDTH)) {
+    args->AddInt("kernel_size_x");
+    args->AddInt("padding_x");
+    args->AddInt("stride_x");
+  }
+  if (op_def.dst_tensors[0].HasAxis(Axis::HEIGHT)) {
+    args->AddInt("kernel_size_y");
+    args->AddInt("padding_y");
+    args->AddInt("stride_y");
+  }
+  if (op_def.dst_tensors[0].HasAxis(Axis::DEPTH)) {
+    args->AddInt("kernel_size_z");
+    args->AddInt("padding_z");
+    args->AddInt("stride_z");
+  }
 
   std::string c = GetCommonDefines(op_def.precision);
-
   c += "__kernel void main_function(\n";
-  c += src.GetDeclaration(AccessType::READ) + ",\n";
-  c += src_ind.GetDeclaration(AccessType::READ);
-  c += GetArgsDeclaration(linked_operations);
-  c += dst.GetDeclaration(AccessType::WRITE) + ",\n";
-  c += "    int4 src_size,      \n";
-  c += "    int4 dst_size,      \n";
-  c += "    int2 kernel_size,   \n";
-  c += "    int2 padding,       \n";
-  c += "    int2 stride         \n";
-  c += ") {\n";
+  c += "$0) {\n";
   c += "  int X = get_global_id(0);\n";
-  c += "  int Y = get_global_id(1);\n";
-  c += "  int Z = get_global_id(2);\n";
-  c += "  if (X >= dst_size.x || Y >= dst_size.y || Z >= dst_size.z) return;\n";
-  if (op_def.IsBatchSupported()) {
-    c += "  int linear_id = get_global_id(0);\n";
-    c += "  int X0 = linear_id / dst_size.w;\n";
-    c += "  int B = linear_id % dst_size.w;\n";
-    c += "  int src_x0 = (X0 + padding.x) / stride.x;\n";
-    c += "  int src_x = src_x0 * dst_size.w + B;\n";
+  if (op_def.dst_tensors[0].HasAxis(Axis::DEPTH)) {
+    c += "  int linear_id_1 = get_global_id(1);\n";
+    c += "  int Y = linear_id_1 / args.dst_tensor.Depth();\n";
+    c += "  int Z = linear_id_1 % args.dst_tensor.Depth();\n";
+    c += "  int src_z = (Z + args.padding_z) / args.stride_z;\n";
   } else {
-    c += "  int src_x = (X + padding.x) / stride.x;\n";
+    c += "  int Y = get_global_id(1);\n";
   }
-  c += "  int src_y = (Y + padding.y) / stride.y;\n";
-  c += "  " + src.GetAddressWHS("src_adr", "src_x", "src_y", "Z") + "\n";
+  c += "  int S = get_global_id(2);\n";
+  c += "  if (X >= args.dst_tensor.Width() || Y >= args.dst_tensor.Height() || "
+       "S >= args.dst_tensor.Slices()) { \n";
+  c += "    return; \n";
+  c += "  } \n";
+  if (op_def.dst_tensors[0].HasAxis(Axis::BATCH)) {
+    c += "  int linear_id_0 = get_global_id(0);\n";
+    c += "  int X0 = linear_id_0 / args.dst_tensor.Batch();\n";
+    c += "  int B = linear_id_0 % args.dst_tensor.Batch();\n";
+    c += "  int src_x0 = (X0 + args.padding_x) / args.stride_x;\n";
+    c += "  int src_x = src_x0 * args.dst_tensor.Batch() + B;\n";
+  } else {
+    c += "  int src_x = (X + args.padding_x) / args.stride_x;\n";
+  }
+  c += "  int src_y = (Y + args.padding_y) / args.stride_y;\n";
+  std::string src_args = op_def.dst_tensors[0].HasAxis(Axis::DEPTH)
+                             ? "src_x, src_y, src_z, S"
+                             : "src_x, src_y, S";
   if (op_def.src_tensors[0].storage_type == TensorStorageType::BUFFER) {
-    c += "  bool outside = src_x < 0 || src_y < 0 ||";
-    c += "  src_x >= src_size.x || src_y >= src_size.y;\n";
+    if (op_def.dst_tensors[0].HasAxis(Axis::DEPTH)) {
+      c += "  bool outside = src_x < 0 || src_y < 0 || src_z < 0 || src_x >= "
+           "args.src_tensor.Width() || src_y >= args.src_tensor.Height() || "
+           "src_z >= args.src_tensor.Depth();\n";
+    } else {
+      c += "  bool outside = src_x < 0 || src_y < 0 || src_x >= "
+           "args.src_tensor.Width() || src_y >= args.src_tensor.Height();\n";
+    }
     c += "  FLT4 src = (FLT4)(0.0f);\n";
     c += "  int4 ind = (int4)(0);\n";
     c += "  if (!outside) {\n";
-    c += "    src = " + src.Read("src_adr") + ";\n";
-    c += "    ind = convert_int4(" + src_ind.Read("src_adr") + ");\n";
+    c += "    src = args.src_tensor.Read(" + src_args + ");\n";
+    c += "    ind = convert_int4(args.src_indices.Read(" + src_args + "));\n";
     c += "  }\n";
   } else {
-    c += "  FLT4 src = " + src.Read("src_adr", address_mode) + ";\n";
-    c += "  int4 ind = convert_int4(" + src_ind.Read("src_adr", address_mode) +
-         ");\n";
+    c += "  FLT4 src = args.src_tensor.Read(" + src_args + ");\n";
+    c +=
+        "  int4 ind = convert_int4(args.src_indices.Read(" + src_args + "));\n";
   }
-  if (op_def.IsBatchSupported()) {
-    c += "  int t_x = X0 - (src_x0 * stride.x - padding.x);\n";
+  if (op_def.dst_tensors[0].HasAxis(Axis::BATCH)) {
+    c += "  int t_x = X0 - (src_x0 * args.stride_x - args.padding_x);\n";
   } else {
-    c += "  int t_x = X - (src_x * stride.x - padding.x);\n";
+    c += "  int t_x = X - (src_x * args.stride_x - args.padding_x);\n";
+  }
+  c += "  int t_y = Y - (src_y * args.stride_y - args.padding_y);\n";
+  if (op_def.dst_tensors[0].HasAxis(Axis::DEPTH)) {
+    c += "  int t_z = Z - (src_z * args.stride_z - args.padding_z);\n";
+    c += "  int t_index = (t_y * args.kernel_size_x + t_x) * "
+         "args.kernel_size_z + t_z;\n";
+  } else {
+    c += "  int t_index = t_y * args.kernel_size_x + t_x;\n";
   }
-  c += "  int t_y = Y - (src_y * stride.y - padding.y);\n";
-  c += "  int t_index = t_y * kernel_size.x + t_x;\n";
   c += "  FLT4 result;\n";
   const std::string channels[] = {".x", ".y", ".z", ".w"};
   for (int i = 0; i < 4; ++i) {
     const auto& s = channels[i];
     c += "  result" + s + "= t_index == ind" + s + "? src" + s + ": 0.0f;\n";
   }
-  c += PostProcess(linked_operations, {"result", "X", "Y", "Z"});
-  c += "  " + dst.WriteWHS("result", "X", "Y", "Z");
+  if (op_def.dst_tensors[0].HasAxis(Axis::DEPTH)) {
+    c += "  args.dst_tensor.Write(result, X, Y, Z, S);\n";
+  } else {
+    c += "  args.dst_tensor.Write(result, X, Y, S);\n";
+  }
   c += "}\n";
 
   return c;
 }
-
-std::string GetMaxUnpooling3DKernelCode(
-    const OperationDef& op_def, const CLDevice& device,
-    const std::vector& linked_operations) {
-  TensorCodeGenerator src(
-      "src_data",
-      WHDSPoint{"src_size.x", "src_size.y", "src_size.z", "src_size.w"},
-      op_def.src_tensors[0]);
-  TensorCodeGenerator src_ind(
-      "src_data_indices",
-      WHDSPoint{"src_size.x", "src_size.y", "src_size.z", "src_size.w"},
-      op_def.src_tensors[1]);
-  TensorCodeGenerator dst(
-      "dst_data",
-      WHDSPoint{"dst_size.x", "dst_size.y", "dst_size.z", "dst_size.w"},
-      op_def.dst_tensors[0]);
-
-  const auto address_mode = GetFastestZeroMode(device);
-
-  std::string c = GetCommonDefines(op_def.precision);
-
-  c += "__kernel void main_function(\n";
-  c += src.GetDeclaration(AccessType::READ) + ",\n";
-  c += src_ind.GetDeclaration(AccessType::READ);
-  c += GetArgsDeclaration(linked_operations);
-  c += dst.GetDeclaration(AccessType::WRITE) + ",\n";
-  c += "    int4 src_size,      \n";
-  c += "    int4 dst_size,      \n";
-  if (op_def.IsBatchSupported()) {
-    c += "    int batch_size,          \n";
-  }
-  c += "    int4 kernel_size,   \n";
-  c += "    int4 padding,       \n";
-  c += "    int4 stride         \n";
-  c += ") {\n";
-  c += "  int X = get_global_id(0);\n";
-  c += "  int Y = get_global_id(1);\n";
-  c += "  int linear_id_z = get_global_id(2);\n";
-  c += "  int S = linear_id_z % dst_size.w;\n";
-  c += "  int Z = linear_id_z / dst_size.w;\n";
-  c += "  if (X >= dst_size.x || Y >= dst_size.y || Z >= dst_size.z) return;\n";
-  if (op_def.IsBatchSupported()) {
-    c += "  int linear_id = get_global_id(0);\n";
-    c += "  int X0 = linear_id / batch_size;\n";
-    c += "  int B = linear_id % batch_size;\n";
-    c += "  int src_x0 = (X0 + padding.x) / stride.x;\n";
-    c += "  int src_x = src_x0 * batch_size + B;\n";
-  } else {
-    c += "  int src_x = (X + padding.x) / stride.x;\n";
-  }
-  c += "  int src_y = (Y + padding.y) / stride.y;\n";
-  c += "  int src_z = (Z + padding.z) / stride.z;\n";
-  c += "  " + src.GetAddressWHDS("src_adr", "src_x", "src_y", "src_z", "S") +
-       "\n";
-  if (op_def.src_tensors[0].storage_type == TensorStorageType::BUFFER) {
-    c += "  bool outside = src_x < 0 || src_y < 0 || src_z < 0 || ";
-    c += "  src_x >= src_size.x || src_y >= src_size.y || src_z >= "
-         "src_size.z;\n";
-    c += "  FLT4 src = (FLT4)(0.0f);\n";
-    c += "  int4 ind = (int4)(0);\n";
-    c += "  if (!outside) {\n";
-    c += "    src = " + src.Read("src_adr") + ";\n";
-    c += "    ind = convert_int4(" + src_ind.Read("src_adr") + ");\n";
-    c += "  }\n";
-  } else {
-    c += "  FLT4 src = " + src.Read("src_adr", address_mode) + ";\n";
-    c += "  int4 ind = convert_int4(" + src_ind.Read("src_adr", address_mode) +
-         ");\n";
-  }
-  if (op_def.IsBatchSupported()) {
-    c += "  int t_x = X0 - (src_x0 * stride.x - padding.x);\n";
-  } else {
-    c += "  int t_x = X - (src_x * stride.x - padding.x);\n";
-  }
-  c += "  int t_y = Y - (src_y * stride.y - padding.y);\n";
-  c += "  int t_z = Z - (src_z * stride.z - padding.z);\n";
-  c += "  int t_index = (t_y * kernel_size.x + t_x) * kernel_size.z + t_z;\n";
-  c += "  FLT4 result;\n";
-  const std::string channels[] = {".x", ".y", ".z", ".w"};
-  for (int i = 0; i < 4; ++i) {
-    const auto& s = channels[i];
-    c += "  result" + s + " = t_index == ind" + s + " ? src" + s + ": 0.0f;\n";
-  }
-  c += PostProcess(linked_operations, {"result", "X", "Y", "S"});
-  c += "  " + dst.WriteWHDS("result", "X", "Y", "Z", "S");
-  c += "}\n";
-  return c;
-}
 }  // namespace
 
 MaxUnpooling::MaxUnpooling(const OperationDef& definition,
                            const MaxUnpooling2DAttributes& attr)
     : GPUOperation(definition),
-      stride_(attr.strides.w, attr.strides.h),
-      padding_(attr.padding.appended.w, attr.padding.appended.h),
-      kernel_size_(attr.kernel.w, attr.kernel.h) {}
+      stride_(attr.strides.w, attr.strides.h, 0, 0),
+      padding_(attr.padding.appended.w, attr.padding.appended.h, 0, 0),
+      kernel_size_(attr.kernel.w, attr.kernel.h, 0, 0) {}
+
+MaxUnpooling::MaxUnpooling(const OperationDef& definition,
+                           const MaxUnpooling3DAttributes& attr)
+    : GPUOperation(definition),
+      stride_(attr.strides.w, attr.strides.h, attr.strides.d, 0),
+      padding_(attr.padding.appended.w, attr.padding.appended.h,
+               attr.padding.appended.d, 0),
+      kernel_size_(attr.kernel.w, attr.kernel.h, attr.kernel.d, 0) {}
 
 MaxUnpooling::MaxUnpooling(MaxUnpooling&& kernel)
     : GPUOperation(std::move(kernel)),
@@ -219,30 +177,45 @@ MaxUnpooling& MaxUnpooling::operator=(MaxUnpooling&& kernel) {
 }
 
 absl::Status MaxUnpooling::Compile(const CreationContext& creation_context) {
-  const auto code = GetMaxUnpoolingKernelCode(
-      definition_, *creation_context.device, linked_operations_);
+  std::string code =
+      GetMaxUnpoolingKernelCode(definition_, *creation_context.device, &args_);
+  std::string element_wise_code;
+  RETURN_IF_ERROR(
+      MergeOperations(linked_operations_, &args_, &element_wise_code));
+  RETURN_IF_ERROR(args_.TransformToCLCode(creation_context.device->GetInfo(),
+                                          {{"dst_tensor", element_wise_code}},
+                                          &code));
   return creation_context.cache->GetOrCreateCLKernel(
       code, "main_function", *creation_context.context,
       *creation_context.device, &kernel_);
 }
 
 absl::Status MaxUnpooling::BindArguments() {
-  kernel_.ResetBindingCounter();
-  RETURN_IF_ERROR(kernel_.SetMemoryAuto(src_[0]->GetMemoryPtr()));
-  RETURN_IF_ERROR(kernel_.SetMemoryAuto(src_[1]->GetMemoryPtr()));
-  RETURN_IF_ERROR(BindArgs(&kernel_, linked_operations_));
-  RETURN_IF_ERROR(kernel_.SetMemoryAuto(dst_[0]->GetMemoryPtrForWriting()));
-  RETURN_IF_ERROR(kernel_.SetBytesAuto(src_[0]->GetWBatchedHSB()));
-  RETURN_IF_ERROR(kernel_.SetBytesAuto(dst_[0]->GetWBatchedHSB()));
-  RETURN_IF_ERROR(kernel_.SetBytesAuto(kernel_size_));
-  RETURN_IF_ERROR(kernel_.SetBytesAuto(padding_));
-  RETURN_IF_ERROR(kernel_.SetBytesAuto(stride_));
-  return absl::OkStatus();
+  RETURN_IF_ERROR(args_.SetObjectRef("src_tensor", src_[0]));
+  RETURN_IF_ERROR(args_.SetObjectRef("src_indices", src_[1]));
+  RETURN_IF_ERROR(args_.SetObjectRef("dst_tensor", dst_[0]));
+  if (definition_.dst_tensors[0].HasAxis(Axis::WIDTH)) {
+    RETURN_IF_ERROR(args_.SetInt("stride_x", stride_.x));
+    RETURN_IF_ERROR(args_.SetInt("padding_x", padding_.x * src_[0]->Batch()));
+    RETURN_IF_ERROR(args_.SetInt("kernel_size_x", kernel_size_.x));
+  }
+  if (definition_.dst_tensors[0].HasAxis(Axis::HEIGHT)) {
+    RETURN_IF_ERROR(args_.SetInt("stride_y", stride_.y));
+    RETURN_IF_ERROR(args_.SetInt("padding_y", padding_.y));
+    RETURN_IF_ERROR(args_.SetInt("kernel_size_y", kernel_size_.y));
+  }
+  if (definition_.dst_tensors[0].HasAxis(Axis::DEPTH)) {
+    RETURN_IF_ERROR(args_.SetInt("stride_z", stride_.z));
+    RETURN_IF_ERROR(args_.SetInt("padding_z", padding_.z));
+    RETURN_IF_ERROR(args_.SetInt("kernel_size_z", kernel_size_.z));
+  }
+  RETURN_IF_ERROR(SetArguments(linked_operations_, &args_));
+  return args_.Bind(kernel_.kernel());
 }
 
 int3 MaxUnpooling::GetGridSize() const {
   const int grid_x = dst_[0]->Width() * dst_[0]->Batch();
-  const int grid_y = dst_[0]->Height();
+  const int grid_y = dst_[0]->Height() * dst_[0]->Depth();
   const int grid_z = dst_[0]->Slices();
   return int3(grid_x, grid_y, grid_z);
 }
@@ -262,82 +235,9 @@ MaxUnpooling CreateMaxUnpooling(const OperationDef& definition,
   return MaxUnpooling(definition, attr);
 }
 
-MaxUnpooling3D::MaxUnpooling3D(const OperationDef& definition,
-                               const MaxUnpooling3DAttributes& attr)
-    : GPUOperation(definition),
-      stride_(attr.strides.w, attr.strides.h, attr.strides.d),
-      padding_(attr.padding.appended.w, attr.padding.appended.h,
-               attr.padding.appended.d),
-      kernel_size_(attr.kernel.w, attr.kernel.h, attr.kernel.d) {}
-
-MaxUnpooling3D::MaxUnpooling3D(MaxUnpooling3D&& kernel)
-    : GPUOperation(std::move(kernel)),
-      stride_(kernel.stride_),
-      padding_(kernel.padding_),
-      kernel_size_(kernel.kernel_size_),
-      kernel_(std::move(kernel.kernel_)),
-      work_group_size_(kernel.work_group_size_) {}
-
-MaxUnpooling3D& MaxUnpooling3D::operator=(MaxUnpooling3D&& kernel) {
-  if (this != &kernel) {
-    std::swap(stride_, kernel.stride_);
-    std::swap(padding_, kernel.padding_);
-    std::swap(kernel_size_, kernel.kernel_size_);
-    kernel_ = std::move(kernel.kernel_);
-    std::swap(work_group_size_, kernel.work_group_size_);
-    GPUOperation::operator=(std::move(kernel));
-  }
-  return *this;
-}
-
-absl::Status MaxUnpooling3D::Compile(const CreationContext& creation_context) {
-  const auto code = GetMaxUnpooling3DKernelCode(
-      definition_, *creation_context.device, linked_operations_);
-  return creation_context.cache->GetOrCreateCLKernel(
-      code, "main_function", *creation_context.context,
-      *creation_context.device, &kernel_);
-}
-
-absl::Status MaxUnpooling3D::BindArguments() {
-  kernel_.ResetBindingCounter();
-  RETURN_IF_ERROR(kernel_.SetMemoryAuto(src_[0]->GetMemoryPtr()));
-  RETURN_IF_ERROR(kernel_.SetMemoryAuto(src_[1]->GetMemoryPtr()));
-  RETURN_IF_ERROR(BindArgs(&kernel_, linked_operations_));
-  RETURN_IF_ERROR(kernel_.SetMemoryAuto(dst_[0]->GetMemoryPtrForWriting()));
-  RETURN_IF_ERROR(kernel_.SetBytesAuto(src_[0]->GetWBatchedHDS()));
-  RETURN_IF_ERROR(kernel_.SetBytesAuto(dst_[0]->GetWBatchedHDS()));
-  if (definition_.IsBatchSupported()) {
-    RETURN_IF_ERROR(kernel_.SetBytesAuto(src_[0]->Batch()));
-  }
-  RETURN_IF_ERROR(kernel_.SetBytesAuto(
-      int4(kernel_size_.x, kernel_size_.y, kernel_size_.z, 1)));
-  RETURN_IF_ERROR(
-      kernel_.SetBytesAuto(int4(padding_.x, padding_.y, padding_.z, 1)));
-  RETURN_IF_ERROR(
-      kernel_.SetBytesAuto(int4(stride_.x, stride_.y, stride_.z, 1)));
-  return absl::OkStatus();
-}
-
-int3 MaxUnpooling3D::GetGridSize() const {
-  const int grid_x = dst_[0]->Width() * dst_[0]->Batch();
-  const int grid_y = dst_[0]->Height();
-  const int grid_z = dst_[0]->Slices() * dst_[0]->Depth();
-  return int3(grid_x, grid_y, grid_z);
-}
-
-absl::Status MaxUnpooling3D::Tune(const TuningParameters& params) {
-  RETURN_IF_ERROR(BindArguments());
-  return GetBestWorkGroup(params, kernel_, GetGridSize(), &work_group_size_);
-}
-
-absl::Status MaxUnpooling3D::AddToQueue(CLCommandQueue* queue) {
-  RETURN_IF_ERROR(BindArguments());
-  return queue->DispatchImplicit(kernel_, GetGridSize(), work_group_size_);
-}
-
-MaxUnpooling3D CreateMaxUnpooling3D(const OperationDef& definition,
-                                    const MaxUnpooling3DAttributes& attr) {
-  return MaxUnpooling3D(definition, attr);
+MaxUnpooling CreateMaxUnpooling(const OperationDef& definition,
+                                const MaxUnpooling3DAttributes& attr) {
+  return MaxUnpooling(definition, attr);
 }
 
 }  // namespace cl
diff --git a/tensorflow/lite/delegates/gpu/cl/kernels/max_unpooling.h b/tensorflow/lite/delegates/gpu/cl/kernels/max_unpooling.h
index 19184ee1e89..24b8c4bbfe3 100644
--- a/tensorflow/lite/delegates/gpu/cl/kernels/max_unpooling.h
+++ b/tensorflow/lite/delegates/gpu/cl/kernels/max_unpooling.h
@@ -29,6 +29,8 @@ class MaxUnpooling : public GPUOperation {
  public:
   MaxUnpooling(const OperationDef& definition,
                const MaxUnpooling2DAttributes& attr);
+  MaxUnpooling(const OperationDef& definition,
+               const MaxUnpooling3DAttributes& attr);
   absl::Status AddToQueue(CLCommandQueue* queue) override;
   absl::Status Tune(const TuningParameters& params) override;
 
@@ -44,9 +46,9 @@ class MaxUnpooling : public GPUOperation {
   absl::Status BindArguments();
   int3 GetGridSize() const;
 
-  int2 stride_;
-  int2 padding_;
-  int2 kernel_size_;
+  int4 stride_;
+  int4 padding_;
+  int4 kernel_size_;
 
   CLKernel kernel_;
   int3 work_group_size_ = int3(8, 4, 1);
@@ -55,35 +57,8 @@ class MaxUnpooling : public GPUOperation {
 MaxUnpooling CreateMaxUnpooling(const OperationDef& definition,
                                 const MaxUnpooling2DAttributes& attr);
 
-class MaxUnpooling3D : public GPUOperation {
- public:
-  MaxUnpooling3D(const OperationDef& definition,
-                 const MaxUnpooling3DAttributes& attr);
-  absl::Status AddToQueue(CLCommandQueue* queue) override;
-  absl::Status Tune(const TuningParameters& params) override;
-
-  absl::Status Compile(const CreationContext& creation_context) override;
-
-  // Move only
-  MaxUnpooling3D(MaxUnpooling3D&& kernel);
-  MaxUnpooling3D& operator=(MaxUnpooling3D&& kernel);
-  MaxUnpooling3D(const MaxUnpooling3D&) = delete;
-  MaxUnpooling3D& operator=(const MaxUnpooling3D&) = delete;
-
- private:
-  absl::Status BindArguments();
-  int3 GetGridSize() const;
-
-  int3 stride_;
-  int3 padding_;
-  int3 kernel_size_;
-
-  CLKernel kernel_;
-  int3 work_group_size_ = int3(8, 4, 1);
-};
-
-MaxUnpooling3D CreateMaxUnpooling3D(const OperationDef& definition,
-                                    const MaxUnpooling3DAttributes& attr);
+MaxUnpooling CreateMaxUnpooling(const OperationDef& definition,
+                                const MaxUnpooling3DAttributes& attr);
 
 }  // namespace cl
 }  // namespace gpu

From 13deeb095cbc18960b45eff52c7403f6c1963fd0 Mon Sep 17 00:00:00 2001
From: Raman Sarokin 
Date: Mon, 22 Jun 2020 15:23:37 -0700
Subject: [PATCH 480/867] Mean converted to new style.

PiperOrigin-RevId: 317746588
Change-Id: I9bd2eb57a3a39db01636f0e12ab812d5c21bf251
---
 .../lite/delegates/gpu/cl/kernels/mean.cc     | 80 ++++++++++---------
 1 file changed, 44 insertions(+), 36 deletions(-)

diff --git a/tensorflow/lite/delegates/gpu/cl/kernels/mean.cc b/tensorflow/lite/delegates/gpu/cl/kernels/mean.cc
index f79a30e33dd..3f8fb5ee648 100644
--- a/tensorflow/lite/delegates/gpu/cl/kernels/mean.cc
+++ b/tensorflow/lite/delegates/gpu/cl/kernels/mean.cc
@@ -27,41 +27,46 @@ namespace gpu {
 namespace cl {
 namespace {
 
-std::string GetMeanKernelCode(
-    const OperationDef& op_def,
-    const std::vector& linked_operations,
-    const int3& work_group_size) {
-  TensorCodeGenerator src_tensor(
-      "src_data", WHSPoint{"src_size.x", "src_size.y", "src_size.z"},
-      op_def.src_tensors[0]);
-  TensorCodeGenerator dst_tensor("dst_data", WHSPoint{"1", "1", "src_size.z"},
-                                 op_def.dst_tensors[0]);
+std::string GetMeanKernelCode(const OperationDef& op_def,
+                              const int3& work_group_size, Arguments* args) {
+  args->AddObjectRef(
+      "src_tensor", AccessType::READ,
+      absl::make_unique(op_def.src_tensors[0]));
+  args->AddObjectRef(
+      "dst_tensor", AccessType::WRITE,
+      absl::make_unique(op_def.dst_tensors[0]));
+  args->AddFloat("inv_multiplier_1");
+  args->AddFloat("inv_multiplier_2");
 
   std::string c = GetCommonDefines(op_def.precision);
   const std::string wg_x = std::to_string(work_group_size.x);
   const std::string wg_y = std::to_string(work_group_size.y);
   c += "__kernel void main_function(\n";
-  c += src_tensor.GetDeclaration(AccessType::READ);
-  c += GetArgsDeclaration(linked_operations);
-  c += dst_tensor.GetDeclaration(AccessType::WRITE) + ",\n";
-  c += "    int4 src_size,           \n";
-  c += "    float2 inv_multipliers   \n";
-  c += ") {\n";
+  c += "$0) {\n";
   c += "  __local float4 accum[" +
        std::to_string(work_group_size.x * work_group_size.y) + "];\n";
   c += "  int local_x = get_local_id(0);\n";
   c += "  int local_y = get_local_id(1);\n";
   c += "  int local_id = local_y * " + wg_x + " + local_x;\n";
-  c += "  int S = get_global_id(2);\n";
-  c += "  if (S >= src_size.z) return;\n";
+  if (op_def.dst_tensors[0].HasAxis(Axis::BATCH)) {
+    c += "  int linear_id_2 = get_global_id(2);\n";
+    c += "  int S = linear_id_2 / args.dst_tensor.Batch();\n";
+    c += "  int B = linear_id_2 % args.dst_tensor.Batch();\n";
+    c += "  args.dst_tensor.SetBatchRef(B);\n";
+    c += "  args.src_tensor.SetBatchRef(B);\n";
+  } else {
+    c += "  int S = get_global_id(2);\n";
+  }
+  c += "  if (S >= args.dst_tensor.Slices()) return;\n";
   c += "  accum[local_id] = (float4)(0.0f);\n";
-  c += "  for (int s_y = local_y; s_y < src_size.y; s_y += " + wg_y + ") {\n";
-  c += "    for (int s_x = local_x; s_x < src_size.x; s_x += " + wg_x + ") {\n";
-  c += "        accum[local_id] += " +
-       src_tensor.ReadAsFloatWHS("s_x", "s_y", "S") + ";\n";
+  c += "  for (int s_y = local_y; s_y < args.src_tensor.Height(); s_y += " +
+       wg_y + ") {\n";
+  c += "    for (int s_x = local_x; s_x < args.src_tensor.Width(); s_x += " +
+       wg_x + ") {\n";
+  c += "      accum[local_id] += args.src_tensor.Read(s_x, s_y, S);\n";
   c += "    }\n";
   c += "  }\n";
-  c += "  accum[local_id] *= inv_multipliers.x;\n";
+  c += "  accum[local_id] *= args.inv_multiplier_1;\n";
   c += "  barrier(CLK_LOCAL_MEM_FENCE);\n";
   const int total_size = work_group_size.x * work_group_size.y;
   int offset = 1;
@@ -81,9 +86,8 @@ std::string GetMeanKernelCode(
   for (int i = 1; i < reminder; ++i) {
     c += "  sum += accum[" + std::to_string(offset * i) + "];\n";
   }
-  c += "  FLT4 result = TO_FLT4(sum * inv_multipliers.y);\n";
-  c += PostProcess(linked_operations, {"result", "0", "0", "S"});
-  c += "  " + dst_tensor.WriteWHS("result", "0", "0", "S");
+  c += "  FLT4 result = TO_FLT4(sum * args.inv_multiplier_2);\n";
+  c += "  args.dst_tensor.Write(result, 0, 0, S);\n";
   c += "}\n";
   return c;
 }
@@ -107,30 +111,34 @@ absl::Status Mean::Compile(const CreationContext& creation_context) {
   if (creation_context.device->IsAdreno3xx()) {
     work_group_size_ = int3(16, 8, 1);
   }
-  const auto code =
-      GetMeanKernelCode(definition_, linked_operations_, work_group_size_);
+  std::string code = GetMeanKernelCode(definition_, work_group_size_, &args_);
+  std::string element_wise_code;
+  RETURN_IF_ERROR(
+      MergeOperations(linked_operations_, &args_, &element_wise_code));
+  RETURN_IF_ERROR(args_.TransformToCLCode(creation_context.device->GetInfo(),
+                                          {{"dst_tensor", element_wise_code}},
+                                          &code));
   return creation_context.cache->GetOrCreateCLKernel(
       code, "main_function", *creation_context.context,
       *creation_context.device, &kernel_);
 }
 
 absl::Status Mean::BindArguments() {
-  kernel_.ResetBindingCounter();
-  RETURN_IF_ERROR(kernel_.SetMemoryAuto(src_[0]->GetMemoryPtr()));
-  RETURN_IF_ERROR(BindArgs(&kernel_, linked_operations_));
-  RETURN_IF_ERROR(kernel_.SetMemoryAuto(dst_[0]->GetMemoryPtrForWriting()));
-  RETURN_IF_ERROR(kernel_.SetBytesAuto(src_[0]->GetWBatchedHSB()));
+  RETURN_IF_ERROR(args_.SetObjectRef("src_tensor", src_[0]));
+  RETURN_IF_ERROR(args_.SetObjectRef("dst_tensor", dst_[0]));
   const double total_size = src_[0]->Width() * src_[0]->Height();
   const double size_0 = work_group_size_.x * work_group_size_.y;
   const double size_1 = total_size / size_0;
-  RETURN_IF_ERROR(kernel_.SetBytesAuto(float2(1.0 / size_1, 1.0 / size_0)));
-  return absl::OkStatus();
+  RETURN_IF_ERROR(args_.SetFloat("inv_multiplier_1", 1.0 / size_1));
+  RETURN_IF_ERROR(args_.SetFloat("inv_multiplier_2", 1.0 / size_0));
+  RETURN_IF_ERROR(SetArguments(linked_operations_, &args_));
+  return args_.Bind(kernel_.kernel());
 }
 
 int3 Mean::GetGridSize() const {
-  const int grid_x = work_group_size_.x * dst_[0]->Batch();
+  const int grid_x = work_group_size_.x;
   const int grid_y = work_group_size_.y;
-  const int grid_z = dst_[0]->Slices();
+  const int grid_z = dst_[0]->Slices() * dst_[0]->Batch();
   return int3(grid_x, grid_y, grid_z);
 }
 

From 1c12a84e20a626854cad2651da8a23126910ee3f Mon Sep 17 00:00:00 2001
From: Raman Sarokin 
Date: Mon, 22 Jun 2020 15:24:03 -0700
Subject: [PATCH 481/867] LSTM converted to new style.

PiperOrigin-RevId: 317746681
Change-Id: I8471f352e2f842d40d71c2733564ab346d4eb69c
---
 .../lite/delegates/gpu/cl/kernels/lstm.cc     | 75 +++++++++----------
 1 file changed, 34 insertions(+), 41 deletions(-)

diff --git a/tensorflow/lite/delegates/gpu/cl/kernels/lstm.cc b/tensorflow/lite/delegates/gpu/cl/kernels/lstm.cc
index 77eea07f278..4732d35e987 100644
--- a/tensorflow/lite/delegates/gpu/cl/kernels/lstm.cc
+++ b/tensorflow/lite/delegates/gpu/cl/kernels/lstm.cc
@@ -26,39 +26,34 @@ namespace gpu {
 namespace cl {
 namespace {
 
-std::string GetLSTMCode(const OperationDef& op_def, const CLDevice& device) {
-  const WHSBPoint state_size{"1", "1", "state_size.z", "state_size.w"};
-  const WHSBPoint src_size{"1", "1", "src_size.z", "src_size.w"};
-
-  TensorCodeGenerator intermediate("src_data", src_size, op_def.src_tensors[0]);
-  TensorCodeGenerator prev_state("prev_state", state_size,
-                                 op_def.src_tensors[1]);
-
-  TensorCodeGenerator activation("dst_data", state_size, op_def.dst_tensors[0]);
-  TensorCodeGenerator new_state("new_state", state_size, op_def.dst_tensors[1]);
+std::string GetLSTMCode(const OperationDef& op_def, const CLDevice& device,
+                        Arguments* args) {
+  args->AddObjectRef(
+      "intermediate", AccessType::READ,
+      absl::make_unique(op_def.src_tensors[0]));
+  args->AddObjectRef(
+      "prev_state", AccessType::READ,
+      absl::make_unique(op_def.src_tensors[1]));
+  args->AddObjectRef(
+      "new_state", AccessType::WRITE,
+      absl::make_unique(op_def.dst_tensors[0]));
+  args->AddObjectRef(
+      "activation", AccessType::WRITE,
+      absl::make_unique(op_def.dst_tensors[1]));
 
   std::string c = GetCommonDefines(op_def.precision);
-
   c += "__kernel void main_function(\n";
-  c += intermediate.GetDeclaration(AccessType::READ) + ",\n";
-  c += prev_state.GetDeclaration(AccessType::READ) + ",\n";
-  c += new_state.GetDeclaration(AccessType::WRITE) + ",\n";
-  c += activation.GetDeclaration(AccessType::WRITE) + ",\n";
-  c += "    int4 src_size,             \n";
-  c += "    int4 state_size,           \n";
-  c += "    int BATCH_SIZE             \n";
-  c += ") {\n";
+  c += "$0) {\n";
   c += "  int B = get_global_id(0);\n";
   c += "  int Z = get_global_id(1);\n";
-  c += "  if (Z >= state_size.z || B >= state_size.w) return;\n";
-  c += "  FLT4 prev_st = " + prev_state.ReadWHSB("0", "0", "Z", "B") + ";\n";
-  c += "  FLT4 r0 = " + intermediate.ReadWHSB("0", "0", "Z", "B") + ";\n";
-  c += "  FLT4 r1 = " +
-       intermediate.ReadWHSB("0", "0", "Z + state_size.z", "B") + ";\n";
-  c += "  FLT4 r2 = " +
-       intermediate.ReadWHSB("0", "0", "Z + state_size.z * 2", "B") + ";\n";
-  c += "  FLT4 r3 = " +
-       intermediate.ReadWHSB("0", "0", "Z + state_size.z * 3", "B") + ";\n";
+  c += "  if (Z >= args.activation.Slices() || B >= args.activation.Batch()) "
+       "return;\n";
+  c += "  FLT4 prev_st = args.prev_state.Read(0, 0, Z, B);\n";
+  c += "  FLT4 r0 = args.intermediate.Read(0, 0, Z, B);\n";
+  c += "  int state_stride = args.activation.Slices();\n";
+  c += "  FLT4 r1 = args.intermediate.Read(0, 0, Z + state_stride, B);\n";
+  c += "  FLT4 r2 = args.intermediate.Read(0, 0, Z + state_stride * 2, B);\n";
+  c += "  FLT4 r3 = args.intermediate.Read(0, 0, Z + state_stride * 3, B);\n";
   if (op_def.precision != CalculationsPrecision::F32 && device.IsAdreno()) {
     c += "  FLT4 input_gate;\n";
     c += "  FLT4 new_input;\n";
@@ -97,9 +92,9 @@ std::string GetLSTMCode(const OperationDef& op_def, const CLDevice& device) {
         "* r3));\n";
   }
   c += "  FLT4 new_st = input_gate * new_input + forget_gate * prev_st;\n";
-  c += "  FLT4 activation = output_gate * tanh(new_st);\n";
-  c += "  " + activation.WriteWHSB("activation", "0", "0", "Z", "B");
-  c += "  " + new_state.WriteWHSB("new_st", "0", "0", "Z", "B");
+  c += "  FLT4 act_value = output_gate * tanh(new_st);\n";
+  c += "  args.activation.Write(act_value, 0, 0, Z, B);\n";
+  c += "  args.new_state.Write(new_st, 0, 0, Z, B);\n";
   c += "}\n";
   return c;
 }
@@ -122,22 +117,20 @@ LSTM& LSTM::operator=(LSTM&& kernel) {
 }
 
 absl::Status LSTM::Compile(const CreationContext& creation_context) {
-  const auto code = GetLSTMCode(definition_, *creation_context.device);
+  std::string code = GetLSTMCode(definition_, *creation_context.device, &args_);
+  RETURN_IF_ERROR(
+      args_.TransformToCLCode(creation_context.device->GetInfo(), {}, &code));
   return creation_context.cache->GetOrCreateCLKernel(
       code, "main_function", *creation_context.context,
       *creation_context.device, &kernel_);
 }
 
 absl::Status LSTM::BindArguments() {
-  kernel_.ResetBindingCounter();
-  RETURN_IF_ERROR(kernel_.SetMemoryAuto(src_[0]->GetMemoryPtr()));
-  RETURN_IF_ERROR(kernel_.SetMemoryAuto(src_[1]->GetMemoryPtr()));
-  RETURN_IF_ERROR(kernel_.SetMemoryAuto(dst_[0]->GetMemoryPtrForWriting()));
-  RETURN_IF_ERROR(kernel_.SetMemoryAuto(dst_[1]->GetMemoryPtrForWriting()));
-  RETURN_IF_ERROR(kernel_.SetBytesAuto(src_[0]->GetWHSB()));
-  RETURN_IF_ERROR(kernel_.SetBytesAuto(dst_[0]->GetWHSB()));
-  RETURN_IF_ERROR(kernel_.SetBytesAuto(dst_[0]->Batch()));
-  return absl::OkStatus();
+  RETURN_IF_ERROR(args_.SetObjectRef("intermediate", src_[0]));
+  RETURN_IF_ERROR(args_.SetObjectRef("prev_state", src_[1]));
+  RETURN_IF_ERROR(args_.SetObjectRef("new_state", dst_[0]));
+  RETURN_IF_ERROR(args_.SetObjectRef("activation", dst_[1]));
+  return args_.Bind(kernel_.kernel());
 }
 
 int3 LSTM::GetGridSize() const {

From 5f9ae657edf18cedbf514a325727264a823b0136 Mon Sep 17 00:00:00 2001
From: Rick Chao 
Date: Mon, 22 Jun 2020 15:28:57 -0700
Subject: [PATCH 482/867] Fix tsan failure in multi_process_runner_test.

PiperOrigin-RevId: 317747609
Change-Id: I8bf2e493431a69a0cf581012666045df2879055e
---
 tensorflow/python/distribute/BUILD            |  2 +-
 .../python/distribute/multi_process_runner.py | 12 +++++
 .../distribute/multi_process_runner_test.py   | 46 ++++---------------
 3 files changed, 23 insertions(+), 37 deletions(-)

diff --git a/tensorflow/python/distribute/BUILD b/tensorflow/python/distribute/BUILD
index f0f3766afe1..74d80b63e12 100644
--- a/tensorflow/python/distribute/BUILD
+++ b/tensorflow/python/distribute/BUILD
@@ -1794,7 +1794,7 @@ py_test(
     name = "multi_process_runner_test",
     srcs = ["multi_process_runner_test.py"],
     python_version = "PY3",
-    tags = ["notsan"],  # TODO(b/158874970)
+    shard_count = 12,
     deps = [
         ":multi_process_runner",
         ":multi_worker_test_base",
diff --git a/tensorflow/python/distribute/multi_process_runner.py b/tensorflow/python/distribute/multi_process_runner.py
index af527b67b4b..cb460c8fff5 100644
--- a/tensorflow/python/distribute/multi_process_runner.py
+++ b/tensorflow/python/distribute/multi_process_runner.py
@@ -423,6 +423,18 @@ class MultiProcessRunner(object):
   def join(self, timeout=_DEFAULT_TIMEOUT_SEC):
     """Joins all the processes with timeout.
 
+    If any of the subprocesses does not exit approximately after `timeout`
+    seconds has passed after `join` call, this raises a
+    `SubprocessTimeoutError`.
+
+    Note: At timeout, it uses SIGTERM to terminate the subprocesses, in order to
+    log the stack traces of the subprocesses when they exit. However, this
+    results in timeout when the test runs with tsan (thread sanitizer); if tsan
+    is being run on the test targets that rely on timeout to assert information,
+    `MultiProcessRunner.terminate_all()` must be called after `join()`, before
+    the test exits, so the subprocesses are terminated with SIGKILL, and data
+    race is removed.
+
     Args:
       timeout: if set and not all processes report status within roughly
         `timeout` seconds, a `SubprocessTimeoutError` exception will be raised.
diff --git a/tensorflow/python/distribute/multi_process_runner_test.py b/tensorflow/python/distribute/multi_process_runner_test.py
index 6194ac527d5..529d7fd91a5 100644
--- a/tensorflow/python/distribute/multi_process_runner_test.py
+++ b/tensorflow/python/distribute/multi_process_runner_test.py
@@ -124,24 +124,6 @@ class MultiProcessRunnerTest(test.TestCase):
                   std_stream_results)
     self.assertIn('This is returned data.', return_value)
 
-  def test_process_that_exits(self):
-
-    def func_to_exit_in_25_sec():
-      logging.error('foo')
-      time.sleep(100)
-      logging.error('bar')
-
-    mpr = multi_process_runner.MultiProcessRunner(
-        func_to_exit_in_25_sec,
-        multi_worker_test_base.create_cluster_spec(num_workers=1),
-        list_stdout=True,
-        max_run_time=25)
-
-    mpr.start()
-    stdout = mpr.join().stdout
-    self.assertLen([msg for msg in stdout if 'foo' in msg], 1)
-    self.assertLen([msg for msg in stdout if 'bar' in msg], 0)
-
   def test_termination(self):
 
     def proc_func():
@@ -301,29 +283,21 @@ class MultiProcessRunnerTest(test.TestCase):
   def test_stdout_available_when_timeout(self):
 
     def proc_func():
-      for i in range(50):
-        logging.info('(logging) %s-%d, i: %d',
-                     multi_worker_test_base.get_task_type(), self._worker_idx(),
-                     i)
-        time.sleep(1)
+      logging.info('something printed')
+      time.sleep(10000)  # Intentionally make the test timeout.
 
     with self.assertRaises(multi_process_runner.SubprocessTimeoutError) as cm:
-      multi_process_runner.run(
+      mpr = multi_process_runner.MultiProcessRunner(
           proc_func,
-          multi_worker_test_base.create_cluster_spec(num_workers=1, num_ps=1),
-          list_stdout=True,
-          timeout=5)
+          multi_worker_test_base.create_cluster_spec(num_workers=1),
+          list_stdout=True)
+      mpr.start()
+      mpr.join(timeout=60)
+    mpr.terminate_all()
 
     list_to_assert = cm.exception.mpr_result.stdout
-    # We should see 5 iterations from worker and ps, however sometime on TAP
-    # due to CPU throttling and slugginess of msan/asan build, this became
-    # flaky. Therefore we allow more margin of errors to only check the first
-    # 3 iterations.
-    for job in ['worker', 'ps']:
-      for iteration in range(0, 3):
-        self.assertTrue(
-            any('(logging) {}-0, i: {}'.format(job, iteration) in line
-                for line in list_to_assert))
+    self.assertTrue(
+        any('something printed' in line for line in list_to_assert))
 
   def test_seg_fault_raises_error(self):
 

From 1d0b5f87672f90c6afdcaa665a88fc9f3216503c Mon Sep 17 00:00:00 2001
From: "A. Unique TensorFlower" 
Date: Mon, 22 Jun 2020 15:29:33 -0700
Subject: [PATCH 483/867] Some edits to README.md. Adds some interoperation
 testing for ndarray methods. Improves support for tf.Variable. Points
 np.__doc__ to README.md

PiperOrigin-RevId: 317747715
Change-Id: Ia19bc25959dd5ddb00f828c5aaa17ac5a35c8131
---
 tensorflow/python/ops/numpy_ops/README.md   | 121 +++++++++++---------
 tensorflow/python/ops/numpy_ops/__init__.py | 101 +---------------
 tensorflow/python/ops/numpy_ops/np_utils.py |  10 +-
 3 files changed, 79 insertions(+), 153 deletions(-)

diff --git a/tensorflow/python/ops/numpy_ops/README.md b/tensorflow/python/ops/numpy_ops/README.md
index 812f6b946c0..111de75044f 100644
--- a/tensorflow/python/ops/numpy_ops/README.md
+++ b/tensorflow/python/ops/numpy_ops/README.md
@@ -1,44 +1,52 @@
-# tf.experimental.numpy
+# NumPy API on TensorFlow
 
-This module provides a subset of numpy API, built on top of TensorFlow
-operations. APIs are based on numpy 1.16 version.
+This module provides a subset of NumPy API, built on top of TensorFlow
+operations. APIs are based on and have been tested with NumPy 1.16 version.
 
 The set of supported APIs may be expanded over time. Also future releases may
-change the baseline version of numpy API being supported. A list of some
-systematic differences with numpy are listed later in the "Differences with
-Numpy" section.
+change the baseline version of NumPy API being supported. A list of some
+systematic differences with NumPy are listed later in the "Differences with
+NumPy" section.
 
 ## Getting Started
 
 ```python
 import tensorflow as tf
-from tf.experimental import numpy as np
+from tensorflow.python.ops import numpy_ops as np
 print(np.ones([2,1]) + np.ones([1, 2]))
 ```
 
 ## Types
 
-The module provide an `ndarray` class which wraps an immutable `tf.Tensor`.
+The module provides an `ndarray` class which wraps an immutable `tf.Tensor`.
 Additional functions are provided which accept array-like objects. Here
 array-like objects includes `ndarrays` as defined by this module, as well as
-`tf.Tensor`, in addition to types accepted by `numpy`.
+`tf.Tensor`, in addition to types accepted by NumPy.
 
-A subset of `numpy` dtypes are supported. Type promotion follows numpy
+A subset of NumPy dtypes are supported. Type promotion follows NumPy
 semantics.
 
 ```python
 print(np.ones([1, 2], dtype=np.int16) + np.ones([2, 1], dtype=np.uint8))
 ```
 
-## Interop
+## Array Interface
 
-The numpy API calls can be interleaved with TensorFlow calls without incurring
-Tensor data copies. This is true even if the `ndarray` or `tf.Tensor` is placed
-on a non-CPU device.
+The `ndarray` class implements the `__array__` interface. This should allow
+these objects to be passed into contexts that expect a NumPy or array-like
+object (e.g. matplotlib).
 
-Additionally, one could put these calls in a `with tf.GradientTape()` context to
-compute gradients through the numpy API calls. Similarly, code vectorization can
-be done using `tf.vectorized_map()`.
+```python
+import numpy as onp
+onp.sum(np.ones([1, 2]) + onp.ones([2, 1]))
+```
+
+
+## TF Interoperability
+
+The TF-NumPy API calls can be interleaved with TensorFlow calls
+without incurring Tensor data copies. This is true even if the `ndarray` or
+`tf.Tensor` is placed on a non-CPU device.
 
 In general, the expected behavior should be on par with that of code involving
 `tf.Tensor` and running stateless TensorFlow functions on them.
@@ -47,23 +55,24 @@ In general, the expected behavior should be on par with that of code involving
 np.sum(np.ones([1, 2]) + tf.ones([2, 1]))
 ```
 
-## Array Interface
+Note that the `__array_priority__` is currently chosen to be lower than
+`tf.Tensor`. Hence the `+` operator above returns a `tf.Tensor`.
+
+Additional examples of interopability include:
+
+*  using `with tf.GradientTape()` scope to compute gradients through the
+  TF-NumPy API calls.
+*  using `tf.distribution.Strategy` scope for distributed execution
+*  using `tf.vectorized_map()` for speeding up code using auto-vectorization
 
-The `ndarray` class implements the `__array__` interface. This should allow
-these objects to be passed into contexts that expect a `numpy` or array-like
-object (e.g. matplotlib).
 
-```python
-import numpy as onp
-onp.sum(np.ones([1, 2]) + onp.ones([2, 1]))
-```
 
 ## Device Support
 
 Given that `ndarray` and functions wrap TensorFlow constructs, the code will
-have GPU and TPU support on par with TensorFlow. Also the code can be wrapped
-with `tf.function` and XLA compiled. Device placement can be controlled by using
-`with tf.device` scopes.
+have GPU and TPU support on par with TensorFlow. Device placement can be
+controlled by using `with tf.device` scopes. Note that these devices could
+be local or remote.
 
 ```python
 with tf.device("GPU:0"):
@@ -73,10 +82,10 @@ print(tf.convert_to_tensor(x).device)
 
 ## Graph and Eager Modes
 
-Eager mode execution should typically match numpy semantics of executing
+Eager mode execution should typically match NumPy semantics of executing
 op-by-op. However the same code can be executed in graph mode, by putting it
-inside a `tf.function`. The function body can contain numpy code, and the inputs
-can be ndarray as well.
+inside a `tf.function`. The function body can contain NumPy code, and the inputs
+can be `ndarray` as well.
 
 ```python
 @tf.function
@@ -85,45 +94,51 @@ def f(x, y):
 
 f(np.ones([1, 2]), tf.ones([2, 1]))
 ```
+Python control flow based on `ndarray` values will be translated by
+[autograph](https://www.tensorflow.org/code/tensorflow/python/autograph/g3doc/reference/index.md)
+into `tf.cond` and `tf.while_loop` constructs. The code can be XLA compiled
+for further optimizations.
 
-Note that this can change behavior of certain operations since symbolic
-execution may not have information that is computed during runtime.
+However, note that graph mode execution can change behavior of certain
+operations since symbolic execution may not have information that is computed
+during runtime. Some differences are:
 
-Some differences are:
+*   Shapes can be incomplete or unknown in graph mode. This means that
+    `ndarray.shape`, `ndarray.size` and `ndarray.ndim` can return `ndarray`
+    objects instead of returning integer (or tuple of integer) values.
+*   `__len__`, `__iter__` and `__index__` properties of `ndarray`
+    may similarly not be supported in graph mode. Code using these
+    may need to change to explicit shape operations or control flow
+    constructs.
+*   Also note the [autograph limitations](https://www.tensorflow.org/code/tensorflow/python/autograph/g3doc/reference/limitations.md).
 
-*   Shapes can be incomplete or unknown. This means that `ndarray.shape`,
-    `ndarray.size` and `ndarray.ndim` can return `ndarray` objects instead of
-    returning integer (or tuple of integer) values.
-*   Python control flow based on `ndarray` values may not work and may have to
-    be rewritten to use `tf.cond` or `tf.while_loop`. Note that autograph
-    conversion as part of `tf.function` should still work.
-*   `__len__`, `__iter__` and `__index__` properties of `ndarray` may similarly
-    not work in graph mode.
 
 ## Mutation and Variables
 
-`ndarrays` currently wrap immutable `tf.Tensor`. Also currently mutation
+`ndarrays` currently wrap immutable `tf.Tensor`. Hence mutation
 operations like slice assigns are not supported. This may change in the future.
+Note however that one can directly construct a `tf.Variable` and use that with
+the TF-NumPy APIs.
 
-There is currently no explict construct on par with `tf.Variable`. However one
-can directly construct a `tf.Variable` and use that with the numpy APIs in this
-module. See section on Interop.
+```python
+tf_var = tf.Variable(2.0)
+tf_var.assign_add(np.square(tf_var))
+```
 
-## Differences with Numpy
+## Differences with NumPy
 
 Here is a non-exhaustive list of differences:
 
 *   Not all dtypes are currently supported. e.g. `np.float96`, `np.float128`.
     `np.object`, `np.str`, `np.recarray` types are not supported.
-*   `ndarray` storage is in C order only. Fortran order, views, stride_tricks
+*   `ndarray` storage is in C order only. Fortran order, views, `stride_tricks`
     are not supported.
-*   Only a subset of functions and modules are supported. This set would be
+*   Only a subset of functions and modules are supported. This set will be
     expanded over time. For supported functions, some arguments or argument
-    values may not be supported. This differences are listed in the function
-    comments.
+    values may not be supported. This differences are generally provide in the
+    function comments. Full `ufunc` support is also not provided.
 *   Buffer mutation is currently not supported. `ndarrays` wrap immutable
     tensors. This means that output buffer arguments (e..g `out` in ufuncs) are
     not supported
-*   full `ufunc` support is not provided.
-*   Numpy C API is not supported. Numpy's Cython and Swig integration are not
+*   NumPy C API is not supported. NumPy's Cython and Swig integration are not
     supported.
diff --git a/tensorflow/python/ops/numpy_ops/__init__.py b/tensorflow/python/ops/numpy_ops/__init__.py
index c520ece7843..0e87036584e 100644
--- a/tensorflow/python/ops/numpy_ops/__init__.py
+++ b/tensorflow/python/ops/numpy_ops/__init__.py
@@ -12,103 +12,14 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 # ==============================================================================
-"""numpy_ops.
+"""tf.experimental.numpy: Numpy API on top of TensorFlow.
 
-This module provides a subset of numpy API, built on top of TensorFlow
-operations. APIs are based on numpy 1.16 version.
-
-The set of supported APIs may be expanded over time. Also future releases may
-change the baseline version of numpy API being supported. A list of some
-systematic differences with numpy are listed later in the "Differences with
-Numpy" section.
-
-Types
------
-
-The module provide an `ndarray` class which wraps an immutable `tf.Tensor`.
-Additional functions are provided which accept array-like objects. Here
-array-like objects includes `ndarrays` as defined by this module, as well as
-`tf.Tensor`, in addition to types accepted by `numpy`.
-
-A subset of `numpy` dtypes are supported, along with `tf.bfloat16`.
-
-Device Support
--------------
-
-Given that `ndarray` and functions wrap TensorFlow constructs, the code will
-have GPU and TPU support on par with TensorFlow. Also the code can be wrapped
-with `tf.function` and XLA compiled. Device placement can be controlled by using
-`with tf.device` scopes.
-
-Graph and Eager Modes
---------------------
-
-Eager mode execution should typically match numpy semantics of executing
-op-by-op. However the same code can be executed in graph mode, by putting it
-inside a `tf.function`. This can change behavior of certain operations since
-symbolic execution may not have information that is computed during runtime.
-
-Some differences are:
-  * Shapes can be incomplete or unknown. This means that `ndarray.shape`,
-    `ndarray.size` and `ndarray.ndim` can return `ndarray` objects instead of
-    returning integer (or tuple of integer) values.
-  * Python control flow based on `ndarray` values may not work and may have to
-    be rewritten to use `tf.cond` or `tf.while_loop`. Note that autograph
-    conversion as part of `tf.function` should still work.
-  * `__len__`, `__iter__` and `__index__` properties of `ndarray` may similarly
-    not work in graph mode.
-
-Mutation and Variables
----------------------
-
-`ndarrays` currently wrap immutable `tf.Tensor`. Also currently mutation
-operations like slice assigns are not supported. This may change in the future.
-
-There is currently no explict construct on par with tf.Variable. However one can
-directly construct a `tf.Variable` and use that with the numpy APIs in this
-module. See section on Interop.
-
-Interop
-------
-
-The numpy API calls can be interleaved with TensorFlow calls without incurring
-Tensor data copies. This is true even if the `ndarray` or `tf.Tensor` is placed
-on a non-CPU device.
-
-Additionally, one could put these calls in a `with tf.GradientTape()` context to
-compute gradients through the numpy API calls. Similarly, code vectorization can
-be done using `tf.vectorized_map()`.
-
-In general, the expected behavior should be on par with that of code involving
-`tf.Tensor` and running stateless TensorFlow functions on them.
-
-Array Interface
---------------
-
-The `ndarray` class implements the `__array__ interface. This should allow these
-objects to be passed into contexts that expect a `numpy` or array-like object
-(e.g. matplotlib).
-
-
-Differences with Numpy
----------------------
-
-Here is a non-exhaustive list of differences:
-  * Not all dtypes are currently supported. e.g. `np.float96`, `np.float128`.
-    `np.object`, `np.str`, `np.recarray` types are not supported.
-  * `ndarray` storage is in C order only. Fortran order, views, stride_tricks
-    are not supported.
-  * Only a subset of functions and modules are supported. This set would be
-    expanded over time. For supported functions, some arguments or argument
-    values may not be supported. This differences are listed in the function
-    comments.
-  * Buffer mutation is currently not supported. `ndarrays` wrap immutable
-    tensors. This means that output buffer arguments (e..g `out` in ufuncs) are
-    not supported
-  * full `ufunc` support is not provided.
-  * Numpy C API is not supported. Numpy's Cython and Swig integration are not
-    supported.
+This module provides a subset of numpy APIs, built on top of TensorFlow
+operations. Please see documentation here:
+https://github.com/tensorflow/tensorflow/tree/master/tensorflow/python/ops/numpy_ops.
 """
+# TODO(wangpeng): Append `np_export`ed symbols to the comments above.
+
 # pylint: disable=g-direct-tensorflow-import
 
 from __future__ import absolute_import
diff --git a/tensorflow/python/ops/numpy_ops/np_utils.py b/tensorflow/python/ops/numpy_ops/np_utils.py
index 04ec38d611c..4a7d5f8fea7 100644
--- a/tensorflow/python/ops/numpy_ops/np_utils.py
+++ b/tensorflow/python/ops/numpy_ops/np_utils.py
@@ -24,13 +24,13 @@ import numpy as np
 
 from tensorflow.python.framework import dtypes
 from tensorflow.python.framework import indexed_slices
-from tensorflow.python.framework import ops
 from tensorflow.python.framework import tensor_util
 from tensorflow.python.ops import array_ops
 from tensorflow.python.ops import control_flow_ops
 from tensorflow.python.ops import math_ops
 from tensorflow.python.ops.numpy_ops import np_arrays
 from tensorflow.python.ops.numpy_ops import np_dtypes
+from tensorflow.python.types import core
 from tensorflow.python.util import nest
 
 tensor_to_ndarray = np_arrays.tensor_to_ndarray
@@ -43,7 +43,7 @@ def _canonicalize_axis(axis, rank):
 def _canonicalize_axes(axes, rank):
   rank = _maybe_static(rank)
 
-  if isinstance(rank, ops.Tensor):
+  if isinstance(rank, core.Tensor):
     canonicalizer = (
         lambda axis: cond(axis < 0, lambda: axis + rank, lambda: axis))
   else:
@@ -102,7 +102,7 @@ def isscalar(val):
   """Returns whether `val` is a scalar value or scalar Tensor."""
   if isinstance(val, np_arrays.ndarray):
     val = val.data
-  if isinstance(val, ops.Tensor):
+  if isinstance(val, core.Tensor):
     ndims = val.shape.ndims
     if ndims is not None:
       return ndims == 0
@@ -127,7 +127,7 @@ def result_type(*arrays_and_dtypes):
     # Don't put np.ndarray in this list, because np.result_type looks at the
     # value (not just dtype) of np.ndarray to decide the result type.
     if isinstance(
-        x, (np_arrays.ndarray, ops.Tensor, indexed_slices.IndexedSlices)):
+        x, (np_arrays.ndarray, core.Tensor, indexed_slices.IndexedSlices)):
       return _to_numpy_type(x.dtype)
     elif isinstance(x, dtypes.DType):
       return _to_numpy_type(x)
@@ -393,7 +393,7 @@ def get_static_value(x):
     Same as `tf.get_static_value`, except that it returns None when `x` has a
     float dtype.
   """
-  if isinstance(x, ops.Tensor) and (x.dtype.is_floating or x.dtype.is_complex):
+  if isinstance(x, core.Tensor) and (x.dtype.is_floating or x.dtype.is_complex):
     return None
   return tensor_util.constant_value(x)
 

From dc756d14ca1363a434b30397fd9ffa4d65133b68 Mon Sep 17 00:00:00 2001
From: Andy Ly 
Date: Mon, 22 Jun 2020 15:33:04 -0700
Subject: [PATCH 484/867] Sync TensorFlow MLIR ODS with TensorFlow op registry.

tf.BesselI0e and tf.BesselI1e are moved to tf_ops.td as they are now python generated ops, so they do not have summaries and descriptions in the op registry. int8 and int16 support has been added to tf.Acos, tf.Atan, tf.Inv, tf.Reciprocal, tf.Round, and tf.Tan.

PiperOrigin-RevId: 317748438
Change-Id: Icb86560d5118c38c69819fd02daa3a3841e113b2
---
 .../mlir/tensorflow/ir/tf_generated_ops.td    | 70 ++++---------------
 .../compiler/mlir/tensorflow/ir/tf_ops.td     | 42 +++++++++++
 2 files changed, 56 insertions(+), 56 deletions(-)

diff --git a/tensorflow/compiler/mlir/tensorflow/ir/tf_generated_ops.td b/tensorflow/compiler/mlir/tensorflow/ir/tf_generated_ops.td
index 1e5dad345f8..ecbe63ac525 100644
--- a/tensorflow/compiler/mlir/tensorflow/ir/tf_generated_ops.td
+++ b/tensorflow/compiler/mlir/tensorflow/ir/tf_generated_ops.td
@@ -53,11 +53,11 @@ def TF_AcosOp : TF_Op<"Acos", [NoSideEffect, SameOperandsAndResultType]> {
   let summary = "Computes acos of x element-wise.";
 
   let arguments = (ins
-    TensorOf<[BF16, F16, F32, F64, I32, I64, TF_Complex128, TF_Complex64]>:$x
+    TensorOf<[BF16, F16, F32, F64, I16, I32, I64, I8, TF_Complex128, TF_Complex64]>:$x
   );
 
   let results = (outs
-    TensorOf<[BF16, F16, F32, F64, I32, I64, TF_Complex128, TF_Complex64]>:$y
+    TensorOf<[BF16, F16, F32, F64, I16, I32, I64, I8, TF_Complex128, TF_Complex64]>:$y
   );
 
   TF_DerivedOperandTypeAttr T = TF_DerivedOperandTypeAttr<0>;
@@ -505,11 +505,11 @@ tf.math.asin(y) # [1.047, 0.785] = x
   }];
 
   let arguments = (ins
-    TensorOf<[BF16, F16, F32, F64, I32, I64, TF_Complex128, TF_Complex64]>:$x
+    TensorOf<[BF16, F16, F32, F64, I16, I32, I64, I8, TF_Complex128, TF_Complex64]>:$x
   );
 
   let results = (outs
-    TensorOf<[BF16, F16, F32, F64, I32, I64, TF_Complex128, TF_Complex64]>:$y
+    TensorOf<[BF16, F16, F32, F64, I16, I32, I64, I8, TF_Complex128, TF_Complex64]>:$y
   );
 
   TF_DerivedOperandTypeAttr T = TF_DerivedOperandTypeAttr<0>;
@@ -638,11 +638,11 @@ tf.math.atan(y) # [1.047, 0.785] = x
   }];
 
   let arguments = (ins
-    TensorOf<[BF16, F16, F32, F64, I32, I64, TF_Complex128, TF_Complex64]>:$x
+    TensorOf<[BF16, F16, F32, F64, I16, I32, I64, I8, TF_Complex128, TF_Complex64]>:$x
   );
 
   let results = (outs
-    TensorOf<[BF16, F16, F32, F64, I32, I64, TF_Complex128, TF_Complex64]>:$y
+    TensorOf<[BF16, F16, F32, F64, I16, I32, I64, I8, TF_Complex128, TF_Complex64]>:$y
   );
 
   TF_DerivedOperandTypeAttr T = TF_DerivedOperandTypeAttr<0>;
@@ -921,48 +921,6 @@ reverse of SpaceToBatch.  See below for a precise description.
   TF_DerivedOperandTypeAttr Tblock_shape = TF_DerivedOperandTypeAttr<1>;
 }
 
-def TF_BesselI0eOp : TF_Op<"BesselI0e", [NoSideEffect, SameOperandsAndResultType]> {
-  let summary = "Computes the Bessel i0e function of `x` element-wise.";
-
-  let description = [{
-Exponentially scaled modified Bessel function of order 0 defined as
-`bessel_i0e(x) = exp(-abs(x)) bessel_i0(x)`.
-
-This function is faster and numerically stabler than `bessel_i0(x)`.
-  }];
-
-  let arguments = (ins
-    TF_FpTensor:$x
-  );
-
-  let results = (outs
-    TF_FpTensor:$y
-  );
-
-  TF_DerivedOperandTypeAttr T = TF_DerivedOperandTypeAttr<0>;
-}
-
-def TF_BesselI1eOp : TF_Op<"BesselI1e", [NoSideEffect, SameOperandsAndResultType]> {
-  let summary = "Computes the Bessel i1e function of `x` element-wise.";
-
-  let description = [{
-Exponentially scaled modified Bessel function of order 0 defined as
-`bessel_i1e(x) = exp(-abs(x)) bessel_i1(x)`.
-
-This function is faster and numerically stabler than `bessel_i1(x)`.
-  }];
-
-  let arguments = (ins
-    TF_FpTensor:$x
-  );
-
-  let results = (outs
-    TF_FpTensor:$y
-  );
-
-  TF_DerivedOperandTypeAttr T = TF_DerivedOperandTypeAttr<0>;
-}
-
 def TF_BiasAddOp : TF_Op<"BiasAdd", [NoSideEffect]> {
   let summary = "Adds `bias` to `value`.";
 
@@ -4225,11 +4183,11 @@ I.e., \\(y = 1 / x\\).
   }];
 
   let arguments = (ins
-    TensorOf<[BF16, F16, F32, F64, I32, I64, TF_Complex128, TF_Complex64]>:$x
+    TensorOf<[BF16, F16, F32, F64, I16, I32, I64, I8, TF_Complex128, TF_Complex64]>:$x
   );
 
   let results = (outs
-    TensorOf<[BF16, F16, F32, F64, I32, I64, TF_Complex128, TF_Complex64]>:$y
+    TensorOf<[BF16, F16, F32, F64, I16, I32, I64, I8, TF_Complex128, TF_Complex64]>:$y
   );
 
   TF_DerivedOperandTypeAttr T = TF_DerivedOperandTypeAttr<0>;
@@ -7004,11 +6962,11 @@ I.e., \\(y = 1 / x\\).
   }];
 
   let arguments = (ins
-    TensorOf<[BF16, F16, F32, F64, I32, I64, TF_Complex128, TF_Complex64]>:$x
+    TensorOf<[BF16, F16, F32, F64, I16, I32, I64, I8, TF_Complex128, TF_Complex64]>:$x
   );
 
   let results = (outs
-    TensorOf<[BF16, F16, F32, F64, I32, I64, TF_Complex128, TF_Complex64]>:$y
+    TensorOf<[BF16, F16, F32, F64, I16, I32, I64, I8, TF_Complex128, TF_Complex64]>:$y
   );
 
   TF_DerivedOperandTypeAttr T = TF_DerivedOperandTypeAttr<0>;
@@ -7734,11 +7692,11 @@ according to the current system rounding mode use std::cint.
   }];
 
   let arguments = (ins
-    TensorOf<[BF16, F16, F32, F64, I32, I64, TF_Complex128, TF_Complex64]>:$x
+    TensorOf<[BF16, F16, F32, F64, I16, I32, I64, I8, TF_Complex128, TF_Complex64]>:$x
   );
 
   let results = (outs
-    TensorOf<[BF16, F16, F32, F64, I32, I64, TF_Complex128, TF_Complex64]>:$y
+    TensorOf<[BF16, F16, F32, F64, I16, I32, I64, I8, TF_Complex128, TF_Complex64]>:$y
   );
 
   TF_DerivedOperandTypeAttr T = TF_DerivedOperandTypeAttr<0>;
@@ -9532,11 +9490,11 @@ Given an input tensor, this function computes tangent of every
   }];
 
   let arguments = (ins
-    TensorOf<[BF16, F16, F32, F64, I32, I64, TF_Complex128, TF_Complex64]>:$x
+    TensorOf<[BF16, F16, F32, F64, I16, I32, I64, I8, TF_Complex128, TF_Complex64]>:$x
   );
 
   let results = (outs
-    TensorOf<[BF16, F16, F32, F64, I32, I64, TF_Complex128, TF_Complex64]>:$y
+    TensorOf<[BF16, F16, F32, F64, I16, I32, I64, I8, TF_Complex128, TF_Complex64]>:$y
   );
 
   TF_DerivedOperandTypeAttr T = TF_DerivedOperandTypeAttr<0>;
diff --git a/tensorflow/compiler/mlir/tensorflow/ir/tf_ops.td b/tensorflow/compiler/mlir/tensorflow/ir/tf_ops.td
index 24e88b0e966..f5d8fbae46a 100644
--- a/tensorflow/compiler/mlir/tensorflow/ir/tf_ops.td
+++ b/tensorflow/compiler/mlir/tensorflow/ir/tf_ops.td
@@ -1041,4 +1041,46 @@ operation create / operate on a copy of `x`.
   TF_DerivedOperandTypeAttr T = TF_DerivedOperandTypeAttr<0>;
 }
 
+def TF_BesselI0eOp : TF_Op<"BesselI0e", [NoSideEffect, SameOperandsAndResultType]> {
+  let summary = "Computes the Bessel i0e function of `x` element-wise.";
+
+  let description = [{
+Exponentially scaled modified Bessel function of order 0 defined as
+`bessel_i0e(x) = exp(-abs(x)) bessel_i0(x)`.
+
+This function is faster and numerically stabler than `bessel_i0(x)`.
+  }];
+
+  let arguments = (ins
+    TF_FpTensor:$x
+  );
+
+  let results = (outs
+    TF_FpTensor:$y
+  );
+
+  TF_DerivedOperandTypeAttr T = TF_DerivedOperandTypeAttr<0>;
+}
+
+def TF_BesselI1eOp : TF_Op<"BesselI1e", [NoSideEffect, SameOperandsAndResultType]> {
+  let summary = "Computes the Bessel i1e function of `x` element-wise.";
+
+  let description = [{
+Exponentially scaled modified Bessel function of order 0 defined as
+`bessel_i1e(x) = exp(-abs(x)) bessel_i1(x)`.
+
+This function is faster and numerically stabler than `bessel_i1(x)`.
+  }];
+
+  let arguments = (ins
+    TF_FpTensor:$x
+  );
+
+  let results = (outs
+    TF_FpTensor:$y
+  );
+
+  TF_DerivedOperandTypeAttr T = TF_DerivedOperandTypeAttr<0>;
+}
+
 #endif // TF_OPS

From e653495365a1d50bb7660686f02e55594ea4b62c Mon Sep 17 00:00:00 2001
From: Tiezhen WANG 
Date: Mon, 22 Jun 2020 15:33:55 -0700
Subject: [PATCH 485/867] Update document to use latest TF API - gfile has been
 moved to tf.io

PiperOrigin-RevId: 317748628
Change-Id: I6822e5991d892d238acb0976966e90f60e130961
---
 tensorflow/lite/g3doc/convert/python_api.md | 4 ++--
 1 file changed, 2 insertions(+), 2 deletions(-)

diff --git a/tensorflow/lite/g3doc/convert/python_api.md b/tensorflow/lite/g3doc/convert/python_api.md
index ff14c4b92e7..3171306af13 100644
--- a/tensorflow/lite/g3doc/convert/python_api.md
+++ b/tensorflow/lite/g3doc/convert/python_api.md
@@ -53,7 +53,7 @@ converter = tf.lite.TFLiteConverter.from_saved_model(export_dir)
 tflite_model = converter.convert()
 
 # Save the TF Lite model.
-with tf.gfile.GFile('model.tflite', 'wb') as f:
+with tf.io.gfile.GFile('model.tflite', 'wb') as f:
   f.write(tflite_model)
 ```
 
@@ -125,7 +125,7 @@ converter = tf.lite.TFLiteConverter.from_concrete_functions([concrete_func])
 tflite_model = converter.convert()
 
 # Save the TF Lite model.
-with tf.gfile.GFile('model.tflite', 'wb') as f:
+with tf.io.gfile.GFile('model.tflite', 'wb') as f:
   f.write(tflite_model)
 ```
 

From 3980537192303fc7e7bd10157857d0b0dbd82619 Mon Sep 17 00:00:00 2001
From: Tim Shen 
Date: Mon, 22 Jun 2020 15:38:24 -0700
Subject: [PATCH 486/867] [MLIR] Remove TupleSelectOp from LHLO.

LHLO uses output-parameter, but TupleSelectOp outputs into a tuple on the device. The current type constraints are wrong, and there is not enough expressiveness in LHLO to define a device-memory representation that can be passed in between kernel launches.

PiperOrigin-RevId: 317749542
Change-Id: I6d3350cbf9decf006f239a7208f6bbef0175ac61
---
 tensorflow/compiler/mlir/xla/ir/lhlo_ops.td   |  9 ---------
 .../compiler/mlir/xla/tests/lhlo_ops.mlir     | 20 -------------------
 2 files changed, 29 deletions(-)

diff --git a/tensorflow/compiler/mlir/xla/ir/lhlo_ops.td b/tensorflow/compiler/mlir/xla/ir/lhlo_ops.td
index aed7c83570e..95ad97118ef 100644
--- a/tensorflow/compiler/mlir/xla/ir/lhlo_ops.td
+++ b/tensorflow/compiler/mlir/xla/ir/lhlo_ops.td
@@ -760,15 +760,6 @@ def LHLO_SortOp: LHLO_Op<"sort", []>, BASE_HLO_SortOp {
   let regions = (region SizedRegion<1>:$comparator);
 }
 
-def LHLO_TupleSelectOp: LHLO_Op<"tuple_select", [SameOperandsShape]> {
-  let arguments = (ins
-    Arg:$pred,
-    Arg:$on_true,
-    Arg:$on_false,
-    Arg:$output
-  );
-}
-
 //===----------------------------------------------------------------------===//
 // Late operations
 //===----------------------------------------------------------------------===//
diff --git a/tensorflow/compiler/mlir/xla/tests/lhlo_ops.mlir b/tensorflow/compiler/mlir/xla/tests/lhlo_ops.mlir
index 0ed8b36466e..1e803da4ac6 100644
--- a/tensorflow/compiler/mlir/xla/tests/lhlo_ops.mlir
+++ b/tensorflow/compiler/mlir/xla/tests/lhlo_ops.mlir
@@ -964,23 +964,3 @@ func @sort_memrefs(%arg0: memref<16x16xf32>, %arg1: memref<16x16xf16>,
   }) : (memref<16x16xf32>, memref<16x16xf16>, tuple, memref<16x16xf16>>) -> ()
   return
 }
-
-// -----
-
-// CHECK-LABEL: func @tuple_select_memrefs
-func @tuple_select_memrefs(%pred: memref<20xi1>, %true_values: memref<20xf32>,
-                           %false_values: memref<20xf32>, %arg_out: memref<20xf32>) -> () {
-  "xla_lhlo.tuple_select"(%pred, %true_values, %false_values, %arg_out)
-      : (memref<20xi1>, memref<20xf32>, memref<20xf32>, memref<20xf32>) -> ()
-  return
-}
-
-// -----
-
-func @tuple_select_memrefs(%pred: memref<10xi1>, %true_values: memref<20xf32>,
-                           %false_values: memref<20xf32>, %arg_out: memref<20xf32>) -> () {
-  // expected-error@+1{{requires the same shape for all operands}}
-  "xla_lhlo.tuple_select"(%pred, %true_values, %false_values, %arg_out)
-      : (memref<10xi1>, memref<20xf32>, memref<20xf32>, memref<20xf32>) -> ()
-  return
-}

From 752ce6670f01fa536adcf67c06905ab7a7347add Mon Sep 17 00:00:00 2001
From: Raman Sarokin 
Date: Mon, 22 Jun 2020 15:50:15 -0700
Subject: [PATCH 487/867] ConcatZ converted to new style. Added support of
 BHWDC/HWDC layouts.

PiperOrigin-RevId: 317751701
Change-Id: I0cb91a6a2317e13624189081eb321a1562810b15
---
 .../lite/delegates/gpu/cl/kernels/concat_z.cc | 135 ++++++++----------
 1 file changed, 60 insertions(+), 75 deletions(-)

diff --git a/tensorflow/lite/delegates/gpu/cl/kernels/concat_z.cc b/tensorflow/lite/delegates/gpu/cl/kernels/concat_z.cc
index f1970cef645..08c18907c78 100644
--- a/tensorflow/lite/delegates/gpu/cl/kernels/concat_z.cc
+++ b/tensorflow/lite/delegates/gpu/cl/kernels/concat_z.cc
@@ -36,73 +36,61 @@ bool IsAllChannelsX4(const std::vector& channels) {
   return true;
 }
 
-std::string GetSrcDepthSizeVar(int src_index) {
-  return "src_size_" + std::to_string(src_index) + "_depth";
-}
-
-std::string GetConcatKernelCode(
-    const OperationDef& op_def, const std::vector& channels,
-    const std::vector& linked_operations) {
-  std::vector srcs(channels.size());
-  for (int i = 0; i < channels.size(); ++i) {
-    const std::string tensor_name = "src_data_" + std::to_string(i);
-    srcs[i] = TensorCodeGenerator(
-        tensor_name,
-        WHSPoint{"dst_size.x", "dst_size.y", GetSrcDepthSizeVar(i)},
-        op_def.src_tensors[i]);
+std::string GetConcatKernelCode(const OperationDef& op_def,
+                                const std::vector& channels,
+                                Arguments* args) {
+  std::vector tensor_names(op_def.src_tensors.size());
+  for (int i = 0; i < op_def.src_tensors.size(); ++i) {
+    tensor_names[i] = "src_tensor_" + std::to_string(i);
+    auto src_desc = absl::make_unique(op_def.src_tensors[i]);
+    if (op_def.IsBatchSupported()) {
+      src_desc->SetStateVar("BatchedWidth", "true");
+    }
+    args->AddObjectRef(tensor_names[i], AccessType::READ, std::move(src_desc));
   }
-  TensorCodeGenerator dst("dst_data",
-                          WHSPoint{"dst_size.x", "dst_size.y", "dst_size.z"},
-                          op_def.dst_tensors[0]);
+  auto dst_desc = absl::make_unique(op_def.dst_tensors[0]);
+  if (op_def.IsBatchSupported()) {
+    dst_desc->SetStateVar("BatchedWidth", "true");
+  }
+  args->AddObjectRef("dst_tensor", AccessType::WRITE, std::move(dst_desc));
 
   std::string c = GetCommonDefines(op_def.precision);
-  const std::string postfix[] = {".x", ".y", ".z", ".w"};
-
   c += "__kernel void main_function(\n";
-  for (const auto& src : srcs) {
-    c += src.GetDeclaration(AccessType::READ) + ",\n";
-  }
-  c += dst.GetDeclaration(AccessType::WRITE);
-  c += GetArgsDeclaration(linked_operations);
-  for (int i = 0; i < channels.size(); ++i) {
-    c += "    int " + GetSrcDepthSizeVar(i) + ",\n";
-  }
-  c += "    int4 dst_size\n";
-  c += ") {\n";
+  c += "$0) {\n";
   c += "  int X = get_global_id(0);\n";
   c += "  int Y = get_global_id(1);\n";
-  c += "  if (X >= dst_size.x || Y >= dst_size.y) return; \n";
+  std::string coords = "X, Y";
+  if (op_def.dst_tensors[0].HasAxis(Axis::DEPTH)) {
+    c += "  int Z = get_global_id(2);\n";
+    c += "  if (Z >= args.dst_tensor.Depth()) return;\n";
+    coords = "X, Y, Z";
+  }
+  c += "  if (X >= args.dst_tensor.Width() || Y >= args.dst_tensor.Height()) "
+       "return; \n";
 
   if (IsAllChannelsX4(channels)) {
     // When all channels % 4 == 0 we can read/assign/write FLT4 elements easily.
     // Also it is easy to write a loop in this case, to prevent long kernel
     // generation.
-    c += "  int Z = 0;\n";
+    c += "  int S = 0;\n";
     for (int i = 0; i < channels.size(); ++i) {
+      std::string t_name = "args." + tensor_names[i];
       const int depth = DivideRoundUp(channels[i], 4);
       if (depth % 2 == 0) {
         // We can read more at once inside of loop in case depth % 2 == 0
         // it should be better for reading latency hiding
-        c += "  for (int i = 0; i < " + GetSrcDepthSizeVar(i) + "; i += 2) {\n";
-        c += "    FLT4 result0 = " + srcs[i].ReadWHS("X", "Y", "i") + ";\n";
-        c += "    FLT4 result1 = " + srcs[i].ReadWHS("X", "Y", "i + 1") + ";\n";
-        c += "    " + dst.GetAddressWHS("dst_adr0", "X", "Y", "Z") + "\n";
-        c += "    " + dst.GetAddressWHS("dst_adr1", "X", "Y", "Z + 1") + "\n";
-        const LinkingContext context_0{"result0", "X", "Y", "Z"};
-        const LinkingContext context_1{"result1", "X", "Y", "Z + 1"};
-        c += PostProcess(linked_operations, context_0);
-        c += PostProcess(linked_operations, context_1);
-        c += "    " + dst.WriteWHS("result0", "X", "Y", "Z");
-        c += "    " + dst.WriteWHS("result1", "X", "Y", "Z + 1");
-        c += "    Z += 2;\n";
+        c += "  for (int i = 0; i < " + t_name + ".Slices(); i += 2) {\n";
+        c += "    FLT4 result0 = " + t_name + ".Read(" + coords + ", i);\n";
+        c += "    FLT4 result1 = " + t_name + ".Read(" + coords + ", i + 1);\n";
+        c += "    args.dst_tensor.Write(result0, " + coords + ", S);\n";
+        c += "    args.dst_tensor.Write(result1, " + coords + ", S + 1);\n";
+        c += "    S += 2;\n";
         c += "  }\n";
       } else {
-        c += "  for (int i = 0; i < " + GetSrcDepthSizeVar(i) + "; ++i) {\n";
-        c += "    FLT4 result = " + srcs[i].ReadWHS("X", "Y", "i") + ";\n";
-        const LinkingContext context{"result", "X", "Y", "Z"};
-        c += PostProcess(linked_operations, context);
-        c += "    " + dst.WriteWHS("result", "X", "Y", "Z");
-        c += "    Z++;\n";
+        c += "  for (int i = 0; i < " + t_name + ".Slices(); ++i) {\n";
+        c += "    FLT4 result = " + t_name + ".Read(" + coords + ", i);\n";
+        c += "    args.dst_tensor.Write(result, " + coords + ", S);\n";
+        c += "    S++;\n";
         c += "  }\n";
       }
     }
@@ -111,24 +99,23 @@ std::string GetConcatKernelCode(
     int out_channel = 0;
     int read_index = 0;
     int z = 0;
+    const std::string postfix[] = {".x", ".y", ".z", ".w"};
     for (int i = 0; i < channels.size(); ++i) {
+      std::string tensor_name = "args." + tensor_names[i];
       const int depth = DivideRoundUp(channels[i], 4);
       for (int d = 0; d < depth; ++d) {
         const int channels_in_group = std::min(4, channels[i] - d * 4);
         const std::string temp_name = "t" + std::to_string(read_index);
-        c += "  FLT4 " + temp_name + " = ";
-        c += srcs[i].ReadWHS("X", "Y", std::to_string(d)) + ";\n";
+        c += "  FLT4 " + temp_name + " = " + tensor_name + ".Read(" + coords +
+             ", " + std::to_string(d) + ");\n";
         for (int ch = 0; ch < channels_in_group; ++ch) {
           c += "  result" + postfix[out_channel] + " = ";
           c += temp_name + postfix[ch] + ";\n";
           out_channel++;
           if (out_channel == 4) {
             out_channel = 0;
-            c += "  {\n";
-            const LinkingContext context{"result", "X", "Y", std::to_string(z)};
-            c += PostProcess(linked_operations, context);
-            c += "  " + dst.WriteWHS("result", "X", "Y", std::to_string(z));
-            c += "  }\n";
+            c += "  args.dst_tensor.Write(result, " + coords + ", " +
+                 std::to_string(z) + ");\n";
             z++;
           }
         }
@@ -136,11 +123,8 @@ std::string GetConcatKernelCode(
       }
     }
     if (out_channel != 0) {
-      c += "  {\n";
-      const LinkingContext context{"result", "X", "Y", std::to_string(z)};
-      c += PostProcess(linked_operations, context);
-      c += "  " + dst.WriteWHS("result", "X", "Y", std::to_string(z));
-      c += "  }\n";
+      c += "  args.dst_tensor.Write(result, " + coords + ", " +
+           std::to_string(z) + ");\n";
     }
   }
   c += "}\n";
@@ -166,8 +150,7 @@ ConcatZ& ConcatZ::operator=(ConcatZ&& kernel) {
 }
 
 absl::Status ConcatZ::Compile(const CreationContext& creation_context) {
-  const auto code =
-      GetConcatKernelCode(definition_, channels_, linked_operations_);
+  std::string code = GetConcatKernelCode(definition_, channels_, &args_);
   std::vector options;
   if (creation_context.device->IsPowerVR() &&
       definition_.precision == CalculationsPrecision::F32 &&
@@ -179,32 +162,34 @@ absl::Status ConcatZ::Compile(const CreationContext& creation_context) {
       definition_.precision != CalculationsPrecision::F32 &&
       definition_.src_tensors[0].storage_type != TensorStorageType::BUFFER &&
       !IsAllChannelsX4(channels_)) {
-    // BUG, some AMD gpus crashe without it
+    // BUG, some AMD gpus crash without it
     options.push_back(CompilerOptions::CL_OPT_DISABLE);
   }
+  std::string element_wise_code;
+  RETURN_IF_ERROR(
+      MergeOperations(linked_operations_, &args_, &element_wise_code));
+  RETURN_IF_ERROR(args_.TransformToCLCode(creation_context.device->GetInfo(),
+                                          {{"dst_tensor", element_wise_code}},
+                                          &code));
   return creation_context.cache->GetOrCreateCLKernel(
       code, "main_function", options, *creation_context.context,
       *creation_context.device, &kernel_);
 }
 
 absl::Status ConcatZ::BindArguments() {
-  kernel_.ResetBindingCounter();
-  for (int i = 0; i < channels_.size(); ++i) {
-    RETURN_IF_ERROR(kernel_.SetMemoryAuto(src_[i]->GetMemoryPtr()));
+  for (int i = 0; i < definition_.src_tensors.size(); ++i) {
+    RETURN_IF_ERROR(
+        args_.SetObjectRef("src_tensor_" + std::to_string(i), src_[i]));
   }
-  RETURN_IF_ERROR(kernel_.SetMemoryAuto(dst_[0]->GetMemoryPtrForWriting()));
-  RETURN_IF_ERROR(BindArgs(&kernel_, linked_operations_));
-  for (int i = 0; i < channels_.size(); ++i) {
-    RETURN_IF_ERROR(kernel_.SetBytesAuto(src_[i]->Slices()));
-  }
-  RETURN_IF_ERROR(kernel_.SetBytesAuto(dst_[0]->GetWBatchedHSB()));
-  return absl::OkStatus();
+  RETURN_IF_ERROR(args_.SetObjectRef("dst_tensor", dst_[0]));
+  RETURN_IF_ERROR(SetArguments(linked_operations_, &args_));
+  return args_.Bind(kernel_.kernel());
 }
 
 int3 ConcatZ::GetGridSize() const {
   const int grid_x = dst_[0]->Width() * dst_[0]->Batch();
   const int grid_y = dst_[0]->Height();
-  const int grid_z = 1;
+  const int grid_z = dst_[0]->Depth();
   return int3(grid_x, grid_y, grid_z);
 }
 

From afeac170f0d3035ddea8ff4cac637b47074fb775 Mon Sep 17 00:00:00 2001
From: Raman Sarokin 
Date: Mon, 22 Jun 2020 15:52:35 -0700
Subject: [PATCH 488/867] ConverterToConvWeights converted to new style.

PiperOrigin-RevId: 317752091
Change-Id: I4e23ab1a9a943c197db7a7be25e5272114fd43c3
---
 .../gpu/cl/kernels/conv_weights_converter.cc  | 95 ++++++++++---------
 .../lite/delegates/gpu/cl/tensor_type.cc      | 17 ++++
 .../lite/delegates/gpu/cl/tensor_type.h       |  3 +
 3 files changed, 70 insertions(+), 45 deletions(-)

diff --git a/tensorflow/lite/delegates/gpu/cl/kernels/conv_weights_converter.cc b/tensorflow/lite/delegates/gpu/cl/kernels/conv_weights_converter.cc
index 71559ab587a..18a6886dc89 100644
--- a/tensorflow/lite/delegates/gpu/cl/kernels/conv_weights_converter.cc
+++ b/tensorflow/lite/delegates/gpu/cl/kernels/conv_weights_converter.cc
@@ -26,51 +26,51 @@ namespace cl {
 namespace {
 
 std::string GetConverterToConvWeightsCode(
-    const OperationDef& op_def,
-    const ConvWeightsDescription& conv_weights_desc) {
-  TensorCodeGenerator src_tensor(
-      "src_data",
-      WHSBPoint{"src_size.x", "src_size.y", "src_size.z", "src_size.w"},
-      op_def.src_tensors[0]);
-  TensorCodeGenerator dst_tensor(
-      "dst_data",
-      WHSBPoint{"dst_size.x", "dst_size.y", "dst_size.z", "dst_size.w"},
-      op_def.dst_tensors[0]);
+    const OperationDef& op_def, const ConvWeightsDescription& conv_weights_desc,
+    Arguments* args) {
+  args->AddObjectRef(
+      "src_tensor", AccessType::READ,
+      absl::make_unique(op_def.src_tensors[0]));
+  args->AddObjectRef(
+      "dst_tensor", AccessType::WRITE,
+      absl::make_unique(op_def.dst_tensors[0]));
+  args->AddFloat("mask_x");
+  args->AddFloat("mask_y");
+  args->AddFloat("mask_z");
+  args->AddFloat("mask_w");
 
   std::string c = GetCommonDefines(op_def.precision);
   c += "__kernel void main_function(\n";
-  c += src_tensor.GetDeclaration(AccessType::READ) + ",\n";
-  c += dst_tensor.GetDeclaration(AccessType::WRITE) + ",\n";
-  c += "    int4 src_size,              \n";
-  c += "    float4 mask\n";
-  c += ") {\n";
+  c += "$0) {\n";
   c += "  int GROUP_SIZE = " +
        std::to_string(conv_weights_desc.output_group_size) + ";\n";
   c += "  int O = get_global_id(0) * 4;\n";
   c += "  int I = get_global_id(1);\n";
   c += "  int Z = get_global_id(2);\n";
-  c += "  int W = Z % src_size.x;\n";
-  c += "  int H = Z / src_size.x;\n";
-  c += "  if (O >= src_size.w || I >= src_size.z || H >= src_size.y) return;\n";
-  c += "  FLT4 v0 =" + src_tensor.ReadWHSB("W", "H", "I", "O + 0") + ";\n";
+  c += "  int W = Z % args.src_tensor.Width();\n";
+  c += "  int H = Z / args.src_tensor.Width();\n";
+  c += "  if (O >= args.src_tensor.Batch() || I >= args.src_tensor.Slices() || "
+       "H >= args.src_tensor.Height()) return;\n";
+  c += "  FLT4 v0 = args.src_tensor.Read(W, H, I, O + 0);\n";
   c += "  FLT4 v1 = (FLT4)(0.0f, 0.0f, 0.0f, 0.0f);\n";
   c += "  FLT4 v2 = (FLT4)(0.0f, 0.0f, 0.0f, 0.0f);\n";
   c += "  FLT4 v3 = (FLT4)(0.0f, 0.0f, 0.0f, 0.0f);\n";
-  c += "  if (O + 1 < src_size.w) {\n";
-  c += "    v1 =" + src_tensor.ReadWHSB("W", "H", "I", "O + 1") + ";\n";
+  c += "  if (O + 1 < args.src_tensor.Batch()) {\n";
+  c += "    v1 = args.src_tensor.Read(W, H, I, O + 1);\n";
   c += "  }\n";
-  c += "  if (O + 2 < src_size.w) {\n";
-  c += "    v2 =" + src_tensor.ReadWHSB("W", "H", "I", "O + 2") + ";\n";
+  c += "  if (O + 2 < args.src_tensor.Batch()) {\n";
+  c += "    v2 = args.src_tensor.Read(W, H, I, O + 2);\n";
   c += "  }\n";
-  c += "  if (O + 3 < src_size.w) {\n";
-  c += "    v3 =" + src_tensor.ReadWHSB("W", "H", "I", "O + 3") + ";\n";
+  c += "  if (O + 3 < args.src_tensor.Batch()) {\n";
+  c += "    v3 = args.src_tensor.Read(W, H, I, O + 3);\n";
   c += "  }\n";
-  c += "  if (I == src_size.z - 1) {\n";
-  c += "    FLT4 mask_t = TO_FLT4(mask);\n";
-  c += "    v0 *= mask_t;\n";
-  c += "    v1 *= mask_t;\n";
-  c += "    v2 *= mask_t;\n";
-  c += "    v3 *= mask_t;\n";
+  c += "  if (I == args.src_tensor.Slices() - 1) {\n";
+  c += "    FLT4 mask = (FLT4)(args.mask_x, args.mask_y, args.mask_z, "
+       "args.mask_w);\n";
+  c += "    v0 *= mask;\n";
+  c += "    v1 *= mask;\n";
+  c += "    v2 *= mask;\n";
+  c += "    v3 *= mask;\n";
   c += "  }\n";
   c += "  FLT4 r0 = (FLT4)(v0.x, v1.x, v2.x, v3.x);\n";
   c += "  FLT4 r1 = (FLT4)(v0.y, v1.y, v2.y, v3.y);\n";
@@ -78,17 +78,18 @@ std::string GetConverterToConvWeightsCode(
   c += "  FLT4 r3 = (FLT4)(v0.w, v1.w, v2.w, v3.w);\n";
   c += "  int d_index = O / (GROUP_SIZE * 4);\n";
   c += "  int k_index = (O % (GROUP_SIZE * 4)) / 4;\n";
-  c += "  int dst_offset = (((d_index * src_size.y + H) * src_size.x + W) * "
-       "src_size.z + I) * GROUP_SIZE + "
+  c += "  int dst_offset = (((d_index * args.src_tensor.Height() + H) * "
+       "args.src_tensor.Width() + W) * "
+       "args.src_tensor.Slices() + I) * GROUP_SIZE + "
        "k_index;\n";
   c += "  int address0 = dst_offset * 4 + 0;\n";
   c += "  int address1 = dst_offset * 4 + 1;\n";
   c += "  int address2 = dst_offset * 4 + 2;\n";
   c += "  int address3 = dst_offset * 4 + 3;\n";
-  c += "  " + dst_tensor.Write("r0", "address0");
-  c += "  " + dst_tensor.Write("r1", "address1");
-  c += "  " + dst_tensor.Write("r2", "address2");
-  c += "  " + dst_tensor.Write("r3", "address3");
+  c += "  args.dst_tensor.WriteLinear(r0, dst_offset * 4 + 0)\n;";
+  c += "  args.dst_tensor.WriteLinear(r1, dst_offset * 4 + 1)\n;";
+  c += "  args.dst_tensor.WriteLinear(r2, dst_offset * 4 + 2)\n;";
+  c += "  args.dst_tensor.WriteLinear(r3, dst_offset * 4 + 3)\n;";
   c += "}\n";
   return c;
 }
@@ -115,20 +116,24 @@ ConverterToConvWeights& ConverterToConvWeights::operator=(
 absl::Status ConverterToConvWeights::Compile(
     const CreationContext& creation_context) {
   std::string code =
-      GetConverterToConvWeightsCode(definition_, conv_weights_desc_);
+      GetConverterToConvWeightsCode(definition_, conv_weights_desc_, &args_);
+  RETURN_IF_ERROR(
+      args_.TransformToCLCode(creation_context.device->GetInfo(), {}, &code));
   return creation_context.cache->GetOrCreateCLKernel(
       code, "main_function", *creation_context.context,
       *creation_context.device, &kernel_);
 }
 
 absl::Status ConverterToConvWeights::BindArguments() {
-  kernel_.ResetBindingCounter();
-  RETURN_IF_ERROR(kernel_.SetMemoryAuto(src_[0]->GetMemoryPtr()));
-  RETURN_IF_ERROR(kernel_.SetMemoryAuto(dst_[0]->GetMemoryPtrForWriting()));
-  RETURN_IF_ERROR(kernel_.SetBytesAuto(src_[0]->GetWHSB()));
-  RETURN_IF_ERROR(
-      kernel_.SetBytesAuto(GetMaskForLastPlane(src_[0]->Channels())));
-  return absl::OkStatus();
+  RETURN_IF_ERROR(args_.SetObjectRef("src_tensor", src_[0]));
+  RETURN_IF_ERROR(args_.SetObjectRef("dst_tensor", dst_[0]));
+  float4 mask = GetMaskForLastPlane(src_[0]->Channels());
+  RETURN_IF_ERROR(args_.SetFloat("mask_x", mask.x));
+  RETURN_IF_ERROR(args_.SetFloat("mask_y", mask.y));
+  RETURN_IF_ERROR(args_.SetFloat("mask_z", mask.z));
+  RETURN_IF_ERROR(args_.SetFloat("mask_w", mask.w));
+  RETURN_IF_ERROR(SetArguments(linked_operations_, &args_));
+  return args_.Bind(kernel_.kernel());
 }
 
 int3 ConverterToConvWeights::GetGridSize() const {
diff --git a/tensorflow/lite/delegates/gpu/cl/tensor_type.cc b/tensorflow/lite/delegates/gpu/cl/tensor_type.cc
index 8e048675697..0c3a1e3508c 100644
--- a/tensorflow/lite/delegates/gpu/cl/tensor_type.cc
+++ b/tensorflow/lite/delegates/gpu/cl/tensor_type.cc
@@ -168,6 +168,8 @@ absl::Status TensorDescriptor::PerformSelector(
     return PerformReadSelector(args, template_args, result);
   } else if (selector == "Write") {
     return PerformWriteSelector(args, result);
+  } else if (selector == "WriteLinear") {
+    return PerformWriteLinearSelector(args, result);
   } else if (selector == "GetAddress") {
     return PerformGetAddressSelector(args, result);
   } else {
@@ -253,6 +255,21 @@ absl::Status TensorDescriptor::PerformWriteSelector(
   return absl::OkStatus();
 }
 
+absl::Status TensorDescriptor::PerformWriteLinearSelector(
+    const std::vector& args, std::string* result) const {
+  if (storage_type != TensorStorageType::BUFFER &&
+      storage_type != TensorStorageType::IMAGE_BUFFER) {
+    return absl::InvalidArgumentError(
+        "WriteLinear selector can be used only with linear "
+        "storages(BUFFER/IMAGE_BUFFER)");
+  }
+  if (args.size() != 2) {
+    return absl::NotFoundError("Unrecognized WriteLinear selector");
+  }
+  *result = Write(args[0], "(" + args[1] + ")");
+  return absl::OkStatus();
+}
+
 std::string TensorDescriptor::Read(DataType read_as_type,
                                    const std::string& global_address) const {
   const std::string read_as =
diff --git a/tensorflow/lite/delegates/gpu/cl/tensor_type.h b/tensorflow/lite/delegates/gpu/cl/tensor_type.h
index 7e173753217..3a1d7abb01a 100644
--- a/tensorflow/lite/delegates/gpu/cl/tensor_type.h
+++ b/tensorflow/lite/delegates/gpu/cl/tensor_type.h
@@ -93,6 +93,9 @@ struct TensorDescriptor : public GPUObjectDescriptor {
   absl::Status PerformWriteSelector(const std::vector& args,
                                     std::string* result) const;
 
+  absl::Status PerformWriteLinearSelector(const std::vector& args,
+                                          std::string* result) const;
+
   std::string Read(DataType read_as_type,
                    const std::string& global_address) const;
   std::string Write(const std::string& var_name,

From 26421826a0b43d6148ece23d2fc6d0a4d0442fa0 Mon Sep 17 00:00:00 2001
From: Yujing Zhang 
Date: Mon, 22 Jun 2020 15:52:37 -0700
Subject: [PATCH 489/867] Support packed variable for tf data captured
 function.

If a packed variable is passed to a DatasetOp through a captured function,
1) function instantiation (graph expansion) would insert a PackOp between per-replica arg nodes and the DatasetOp.
2) function execution would feed unpacked ResourceHandles to corresponding sub-functions.

PiperOrigin-RevId: 317752101
Change-Id: I53f7f1d6d075fd40a76a32f329f2dcbbf1db6494
---
 .../process_function_library_runtime.cc       | 19 ++++++++---
 .../process_function_library_runtime_test.cc  | 32 +++++++++++++-----
 .../core/kernels/data/captured_function.cc    | 33 +++++++++++++++++--
 .../custom_training_loop_input_test.py        |  4 ---
 tensorflow/python/eager/function_test.py      |  7 ++++
 tensorflow/python/framework/func_graph.py     |  7 ++++
 6 files changed, 83 insertions(+), 19 deletions(-)

diff --git a/tensorflow/core/common_runtime/process_function_library_runtime.cc b/tensorflow/core/common_runtime/process_function_library_runtime.cc
index 160b9dd88f4..5ee6546f6be 100644
--- a/tensorflow/core/common_runtime/process_function_library_runtime.cc
+++ b/tensorflow/core/common_runtime/process_function_library_runtime.cc
@@ -1356,11 +1356,22 @@ void ProcessFunctionLibraryRuntime::Run(
       // "Index"s of _Arg nodes are unique when all arguments are local Tensors.
       for (const auto& it : comp_data.arg_indices) {
         if (it.sub_index >= 0) {
-          return errors::InvalidArgument("Got unexpected sub_index ",
-                                         it.sub_index, " for argument ",
-                                         it.index);
+          const Tensor& t = args[it.index];
+          if (t.dtype() != DT_RESOURCE) {
+            return errors::InvalidArgument("Got unexpected sub_index ",
+                                           it.sub_index, " for argument ",
+                                           it.index);
+          }
+          const auto& handles = t.flat();
+          if (it.sub_index >= handles.size()) {
+            return errors::InvalidArgument(
+                "Sub_index ", it.sub_index, "is out of range [0,",
+                handles.size(), ") for argument ", it.index);
+          }
+          comp_args->args.push_back(Tensor(handles(it.sub_index)));
+        } else {
+          comp_args->args.push_back(args[it.index]);
         }
-        comp_args->args.push_back(args[it.index]);
       }
       return Status::OK();
     };
diff --git a/tensorflow/core/common_runtime/process_function_library_runtime_test.cc b/tensorflow/core/common_runtime/process_function_library_runtime_test.cc
index a007501fc82..6e17cdf4316 100644
--- a/tensorflow/core/common_runtime/process_function_library_runtime_test.cc
+++ b/tensorflow/core/common_runtime/process_function_library_runtime_test.cc
@@ -28,6 +28,7 @@ limitations under the License.
 #include "tensorflow/core/framework/resource_var.h"
 #include "tensorflow/core/framework/tensor_testutil.h"
 #include "tensorflow/core/framework/type_index.h"
+#include "tensorflow/core/framework/types.pb.h"
 #include "tensorflow/core/lib/core/errors.h"
 #include "tensorflow/core/lib/core/status_test_util.h"
 #include "tensorflow/core/lib/core/threadpool.h"
@@ -867,14 +868,29 @@ TEST_F(ProcessFunctionLibraryRuntimeTest, MultiDevice_CompositeDevice) {
   inst_opts.input_resource_dtypes_and_shapes[0] = {
       initial_resource_value0.dtype(), initial_resource_value0.shape()};
 
-  gtl::InlinedVector handles;
-  handles.push_back(TensorValue(&resource_handle0));
-  handles.push_back(TensorValue(&resource_handle1));
-  TestFunctionPackedArgs args(0, std::move(handles));
-  Tensor ret;
-  TF_CHECK_OK(RunWithPackedArgs("AddVarAcrossDevices", opts, {{"T", DT_FLOAT}},
-                                inst_opts, args, {&ret}));
-  test::ExpectTensorEqual(ret, test::AsTensor({40, 60}));
+  // Packed TensorHandle
+  {
+    gtl::InlinedVector handles;
+    handles.push_back(TensorValue(&resource_handle0));
+    handles.push_back(TensorValue(&resource_handle1));
+    TestFunctionPackedArgs args(0, std::move(handles));
+    Tensor ret;
+    TF_CHECK_OK(RunWithPackedArgs("AddVarAcrossDevices", opts,
+                                  {{"T", DT_FLOAT}}, inst_opts, args, {&ret}));
+    test::ExpectTensorEqual(ret, test::AsTensor({40, 60}));
+  }
+
+  // Packed Tensor
+  {
+    Tensor arg(DT_RESOURCE, TensorShape({2}));
+    arg.flat()(0) = resource_handle0.scalar()();
+    arg.flat()(1) = resource_handle1.scalar()();
+
+    Tensor ret;
+    TF_CHECK_OK(Run("AddVarAcrossDevices", opts, {{"T", DT_FLOAT}}, inst_opts,
+                    {arg}, {&ret}));
+    test::ExpectTensorEqual(ret, test::AsTensor({40, 60}));
+  }
 }
 
 TEST_F(ProcessFunctionLibraryRuntimeTest, MultiDevice_ResourceOutput_GPU) {
diff --git a/tensorflow/core/kernels/data/captured_function.cc b/tensorflow/core/kernels/data/captured_function.cc
index f740d7ff1ad..07e5a5b1273 100644
--- a/tensorflow/core/kernels/data/captured_function.cc
+++ b/tensorflow/core/kernels/data/captured_function.cc
@@ -571,6 +571,9 @@ Status CapturedFunction::Instantiate(
   DCHECK(lib->device() != nullptr);
   inst_opts.target = lib->device()->name();
 
+  // Maps from a CompositeDevice name to underlying physical device names.
+  absl::flat_hash_map> composite_devices;
+
   if (inst_opts.is_multi_device_function) {
     // Compute devices of non-captured inputs.
     //
@@ -596,9 +599,29 @@ Status CapturedFunction::Instantiate(
       const auto& input = captured_inputs_[i];
       DataType dtype = input.dtype();
       if (dtype == DT_RESOURCE) {
-        const ResourceHandle& handle = input.flat()(0);
-        inst_opts.input_devices.push_back(handle.device());
-        const auto& dtypes_and_shapes = handle.dtypes_and_shapes();
+        const auto& handles = input.flat();
+        const ResourceHandle& handle0 = handles(0);
+        string composite_device;
+        auto iter = fdef->arg_attr().find(num_non_captured_inputs + i);
+        if (iter != fdef->arg_attr().end()) {
+          auto arg_attr = iter->second.attr().find("_composite_device");
+          if (arg_attr != iter->second.attr().end()) {
+            composite_device = arg_attr->second.s();
+          }
+        }
+        if (!composite_device.empty()) {
+          if (composite_devices.find(composite_device) ==
+              composite_devices.end()) {
+            for (int i = 0; i < handles.size(); ++i) {
+              composite_devices[composite_device].push_back(
+                  handles(i).device());
+            }
+          }
+          inst_opts.input_devices.push_back(composite_device);
+        } else {
+          inst_opts.input_devices.push_back(handle0.device());
+        }
+        const auto& dtypes_and_shapes = handle0.dtypes_and_shapes();
         // Set dtypes and shapes for resource variable inputs.
         if (!dtypes_and_shapes.empty()) {
           input_resource_variable_dtypes_and_shapes[num_non_captured_inputs +
@@ -613,6 +636,10 @@ Status CapturedFunction::Instantiate(
       }
     }
 
+    for (const auto& it : composite_devices) {
+      inst_opts.composite_devices[it.first] = &it.second;
+    }
+
     for (size_t i = 0; i < fdef->signature().output_arg_size(); ++i) {
       inst_opts.output_devices.push_back(inst_opts.target);
     }
diff --git a/tensorflow/python/distribute/custom_training_loop_input_test.py b/tensorflow/python/distribute/custom_training_loop_input_test.py
index 748cb7834fc..e2c4076f3f1 100644
--- a/tensorflow/python/distribute/custom_training_loop_input_test.py
+++ b/tensorflow/python/distribute/custom_training_loop_input_test.py
@@ -796,10 +796,6 @@ class InputIterationTest(test.TestCase, parameterized.TestCase,
           mode=["eager"]
       ))
   def testMultiDeviceDataCapturedFunction(self, distribution):
-    if getattr(distribution, "_enable_packed_variable_in_eager_mode", False):
-      self.skipTest(
-          "Dataset captured function doesn't support packed tensors yet "
-          "(b/145922293).")
     inputs = constant_op.constant([2., 3.])
     dataset = lambda _: dataset_ops.Dataset.from_tensor_slices(inputs).repeat(5)
     input_iterator = iter(
diff --git a/tensorflow/python/eager/function_test.py b/tensorflow/python/eager/function_test.py
index 2c49795ba8a..3c42d95e437 100644
--- a/tensorflow/python/eager/function_test.py
+++ b/tensorflow/python/eager/function_test.py
@@ -224,6 +224,13 @@ class FunctionTest(test.TestCase, parameterized.TestCase):
 
       return read0, read1, read2, read3
 
+    arg_attrs = read_var.get_concrete_function().function_def.arg_attr
+    self.assertLen(arg_attrs, 2)
+    self.assertEqual(arg_attrs[0].attr['_composite_device'].s,
+                     compat.as_bytes(packed_var_0.device))
+    self.assertEqual(arg_attrs[1].attr['_composite_device'].s,
+                     compat.as_bytes(packed_var_1.device))
+
     self.assertAllEqual(read_var(), (1 + 5, 2 + 5, 3 + 6, 4 + 6))
 
   def testImplementsAttributeBasic(self):
diff --git a/tensorflow/python/framework/func_graph.py b/tensorflow/python/framework/func_graph.py
index b0f8821b17f..e8e8fcbf081 100644
--- a/tensorflow/python/framework/func_graph.py
+++ b/tensorflow/python/framework/func_graph.py
@@ -646,6 +646,13 @@ class FuncGraph(ops.Graph):
     if capture is None:
       placeholder = _create_substitute_placeholder(
           tensor, name=name, dtype=tensor.dtype, shape=shape)
+      # Record the composite device as an attribute to the placeholder.
+      # This attribute would be propogated into the arg_attr of the FunctionDef.
+      # Currently, a packed eager tensor is always placed on a CompositeDevice.
+      if isinstance(tensor, ops.EagerTensor) and tensor.is_packed:
+        placeholder.op._set_attr(  # pylint: disable=protected-access
+            "_composite_device",
+            attr_value_pb2.AttrValue(s=compat.as_bytes(tensor.device)))
       self.add_capture(tensor, placeholder)
     else:
       placeholder = capture[1]

From 23d5a2e00ac5f93234d4cba91108c0bc3d15b112 Mon Sep 17 00:00:00 2001
From: Raman Sarokin 
Date: Mon, 22 Jun 2020 15:53:37 -0700
Subject: [PATCH 490/867] ConvTexture converted to new style. Texture2D
 implements GPUObject.

PiperOrigin-RevId: 317752276
Change-Id: Id1c66b2f6ca7b70475cc82abc422935fc3f1a251
---
 tensorflow/lite/delegates/gpu/cl/BUILD        |   1 +
 .../delegates/gpu/cl/kernels/conv_texture.cc  | 208 +++++++++---------
 .../delegates/gpu/cl/kernels/conv_texture.h   |  85 ++++---
 tensorflow/lite/delegates/gpu/cl/texture2d.cc |  41 ++++
 tensorflow/lite/delegates/gpu/cl/texture2d.h  |  18 +-
 5 files changed, 218 insertions(+), 135 deletions(-)

diff --git a/tensorflow/lite/delegates/gpu/cl/BUILD b/tensorflow/lite/delegates/gpu/cl/BUILD
index fda57ce6da0..3e10f669d70 100644
--- a/tensorflow/lite/delegates/gpu/cl/BUILD
+++ b/tensorflow/lite/delegates/gpu/cl/BUILD
@@ -523,6 +523,7 @@ cc_library(
     deps = [
         ":cl_command_queue",
         ":cl_context",
+        ":gpu_object",
         ":opencl_wrapper",
         ":tensor_type",
         ":util",
diff --git a/tensorflow/lite/delegates/gpu/cl/kernels/conv_texture.cc b/tensorflow/lite/delegates/gpu/cl/kernels/conv_texture.cc
index e92cc13706d..12765b11fa5 100644
--- a/tensorflow/lite/delegates/gpu/cl/kernels/conv_texture.cc
+++ b/tensorflow/lite/delegates/gpu/cl/kernels/conv_texture.cc
@@ -31,18 +31,32 @@ namespace gpu {
 namespace cl {
 namespace {
 
-std::string GenerateConvCode(
-    const OperationDef& op_def, const int3& block_size, bool is1x1,
-    bool adreno4xx_optimization, bool stride_correction,
-    bool different_weights_for_height, const CLDevice& device,
-    const std::vector& linked_operations) {
-  std::string c = GetCommonDefines(op_def.precision);
-  TensorCodeGenerator src_tensor(
-      "src_data", WHSPoint{"src_size.x", "src_size.y", "src_size.z"},
-      op_def.src_tensors[0]);
-  TensorCodeGenerator dst_tensor(
-      "dst_data", WHSPoint{"dst_size.x", "dst_size.y", "dst_size.z"},
-      op_def.dst_tensors[0]);
+std::string GenerateConvCode(const OperationDef& op_def, const int3& block_size,
+                             bool is1x1, bool adreno4xx_optimization,
+                             bool stride_correction,
+                             bool different_weights_for_height,
+                             const CLDevice& device, Arguments* args) {
+  auto src_desc = absl::make_unique(op_def.src_tensors[0]);
+  src_desc->SetTextureAddressMode(GetFastestZeroMode(device));
+  if (op_def.IsBatchSupported()) {
+    src_desc->SetStateVar("BatchedWidth", "true");
+  }
+  args->AddObjectRef("src_tensor", AccessType::READ, std::move(src_desc));
+  auto dst_desc = absl::make_unique(op_def.dst_tensors[0]);
+  if (op_def.IsBatchSupported()) {
+    dst_desc->SetStateVar("BatchedWidth", "true");
+  }
+  args->AddObjectRef("dst_tensor", AccessType::WRITE, std::move(dst_desc));
+  if (!is1x1) {
+    args->AddInt("kernel_size_x");
+    args->AddInt("kernel_size_y");
+    args->AddInt("dilation_x");
+    args->AddInt("dilation_y");
+  }
+  args->AddInt("stride_x");
+  args->AddInt("stride_y");
+  args->AddInt("padding_x");
+  args->AddInt("padding_y");
 
   const auto src_tensor_type = op_def.src_tensors[0].storage_type;
   const bool is_buffer = src_tensor_type == TensorStorageType::IMAGE_BUFFER ||
@@ -63,6 +77,7 @@ std::string GenerateConvCode(
     zs[z] = std::to_string(z);
   }
 
+  std::string c = GetCommonDefines(op_def.precision);
   for (int z = 0; z < block_size.z; ++z) {
     const std::string f0 = std::to_string(z * 4 + 0);
     const std::string f1 = std::to_string(z * 4 + 1);
@@ -86,43 +101,29 @@ std::string GenerateConvCode(
   }
 
   c += "__kernel void main_function(\n";
-  c += src_tensor.GetDeclaration(AccessType::READ) + ",\n";
-  c += "    __read_only image2d_t filters0,   \n";
-  c += "    __read_only image2d_t filters1,   \n";
-  c += "    __read_only image2d_t filters2,   \n";
-  c += "    __read_only image2d_t filters3,   \n";
-  c += "    __read_only image2d_t biases";
-  c += GetArgsDeclaration(linked_operations);
-  c += dst_tensor.GetDeclaration(AccessType::WRITE) + ",\n";
-  c += "    int4 src_size,                   \n";
-  c += "    int4 dst_size,                   \n";
-  if (!is1x1) {
-    c += "    int2 kernel_size,              \n";
-    c += "    int2 dilation,                 \n";
-  }
-  c += "    int2 stride,                     \n";
-  c += "    int2 padding                     \n";
-  c += ") {\n";
+  c += "$0) {\n";
   c += "  int X = get_global_id(0) * " + std::to_string(block_size.x) + ";\n";
   c += "  int Y = get_global_id(1) * " + std::to_string(block_size.y) + ";\n";
   c += "  int Z = get_global_id(2) * " + std::to_string(block_size.z) + ";\n";
-  c += "  if (X >= dst_size.x || Y >= dst_size.y || Z >= dst_size.z) return;\n";
+  c += "  if (X >= args.dst_tensor.Width() || Y >= args.dst_tensor.Height() "
+       "|| Z >= args.dst_tensor.Slices()) return;\n";
   std::vector s_x(block_size.x);
   std::vector s_y(block_size.y);
   for (int x = 0; x < block_size.x; ++x) {
     if (stride_correction) {
       c += "  int xc" + xs[x] + " = " +
-           GetXStrideCorrected("X + " + xs[x], "src_size.w", "stride.x",
-                               "padding.x") +
+           GetXStrideCorrected("X + " + xs[x], "args.src_tensor.Batch()",
+                               "args.stride_x", "args.padding_x") +
            ";\n";
     } else {
       c += "  int xc" + xs[x] + " = (X +" + xs[x] +
-           ") * stride.x + padding.x;\n";
+           ") * args.stride_x + args.padding_x;\n";
     }
     s_x[x] = is1x1 ? "xc" + xs[x] : "cx" + xs[x];
   }
   for (int y = 0; y < block_size.y; ++y) {
-    c += "  int yc" + ys[y] + " = (Y +" + ys[y] + ") * stride.y + padding.y;\n";
+    c += "  int yc" + ys[y] + " = (Y +" + ys[y] +
+         ") * args.stride_y + args.padding_y;\n";
     s_y[y] = is1x1 ? "yc" + ys[y] : "cy" + ys[y];
   }
   for (int i = 0; i < block_size.x * block_size.y * block_size.z; ++i) {
@@ -131,7 +132,7 @@ std::string GenerateConvCode(
   }
   std::string f_y = is1x1 ? "s" : "filter_offset";
   if (different_weights_for_height) {
-    f_y = "Y * src_size.z + s";
+    f_y = "Y * args.src_tensor.Slices() + s";
   }
   if (!is1x1) {
     for (int x = 0; x < block_size.x; ++x) {
@@ -141,31 +142,31 @@ std::string GenerateConvCode(
       c += "  int cy" + ys[y] + ";\n";
     }
     c += "  int filter_offset = 0;\n";
-    c += "  for (int y = 0; y < kernel_size.y; ++y) {\n";
+    c += "  for (int y = 0; y < args.kernel_size_y; ++y) {\n";
     for (int y = 0; y < block_size.y; ++y) {
-      c += "  cy" + ys[y] + " = y * dilation.y + yc" + ys[y] + ";\n";
+      c += "  cy" + ys[y] + " = y * args.dilation_y + yc" + ys[y] + ";\n";
     }
     if (is_buffer) {
       for (int y = 0; y < block_size.y; ++y) {
         c += "  bool in_y" + ys[y] + " = cy" + ys[y] + " >= 0 && cy" + ys[y] +
-             " < src_size.y;\n";
+             " < args.src_tensor.Height();\n";
         if (src_tensor_type == TensorStorageType::BUFFER) {
           c += "    cy" + ys[y] + " = clamp(cy" + ys[y] +
-               ", 0, src_size.y - 1);\n";
+               ", 0, args.src_tensor.Height() - 1);\n";
         }
       }
     }
-    c += "  for (int x = 0; x < kernel_size.x; ++x) {\n";
+    c += "  for (int x = 0; x < args.kernel_size_x; ++x) {\n";
     for (int x = 0; x < block_size.x; ++x) {
-      c += "  cx" + xs[x] + " = x * dilation.x + xc" + xs[x] + ";\n";
+      c += "  cx" + xs[x] + " = x * args.dilation_x + xc" + xs[x] + ";\n";
     }
     if (is_buffer) {
       for (int x = 0; x < block_size.x; ++x) {
         c += "  bool in_x" + xs[x] + " = cx" + xs[x] + " >= 0 && cx" + xs[x] +
-             " < src_size.x;\n";
+             " < args.src_tensor.Width();\n";
         if (src_tensor_type == TensorStorageType::BUFFER) {
           c += "    cx" + xs[x] + " = clamp(cx" + xs[x] +
-               ", 0, src_size.x - 1);\n";
+               ", 0, args.src_tensor.Width() - 1);\n";
         }
       }
       for (int x = 0; x < block_size.x; ++x) {
@@ -173,90 +174,95 @@ std::string GenerateConvCode(
           const std::string id = std::to_string(y * block_size.x + x);
           if (src_tensor_type == TensorStorageType::IMAGE_BUFFER) {
             c += absl::Substitute(
-                "  int addr_$0 = select(-1, cy$2 * src_size.x + cx$1, (in_x$1 "
+                "  int addr_$0 = select(-1, cy$2 * args.src_tensor.Width() + "
+                "cx$1, (in_x$1 "
                 "&& "
                 "in_y$2));\n",
                 y * block_size.x + x, x, y);
             c += absl::Substitute(
-                "  int dz_$0 = select(0, src_size.x * src_size.y, (in_x$1 && "
+                "  int dz_$0 = select(0, args.src_tensor.Width() * "
+                "args.src_tensor.Height(), (in_x$1 && "
                 "in_y$2));\n",
                 y * block_size.x + x, x, y);
           } else {
-            c += absl::Substitute("  int addr_$0 = cy$2 * src_size.x + cx$1;\n",
-                                  y * block_size.x + x, x, y);
+            c += absl::Substitute(
+                "  int addr_$0 = cy$2 * args.src_tensor.Width() + cx$1;\n",
+                y * block_size.x + x, x, y);
           }
         }
       }
       if (src_tensor_type == TensorStorageType::BUFFER) {
-        c += "  int dz = src_size.x * src_size.y;\n";
+        c += "  int dz = args.src_tensor.Width() * args.src_tensor.Height();\n";
       }
     }
   } else if (is_buffer) {
     for (int y = 0; y < block_size.y; ++y) {
       c += "  bool in_y" + ys[y] + " = yc" + ys[y] + " >= 0 && yc" + ys[y] +
-           " < src_size.y;\n";
+           " < args.src_tensor.Height();\n";
     }
     for (int x = 0; x < block_size.x; ++x) {
       c += "  bool in_x" + xs[x] + " = xc" + xs[x] + " >= 0 && xc" + xs[x] +
-           " < src_size.x;\n";
+           " < args.src_tensor.Width();\n";
     }
     for (int x = 0; x < block_size.x; ++x) {
       for (int y = 0; y < block_size.y; ++y) {
         const std::string id = std::to_string(y * block_size.x + x);
         if (src_tensor_type == TensorStorageType::IMAGE_BUFFER) {
           c += absl::Substitute(
-              "  int addr_$0 = select(-1, yc$2 * src_size.x + xc$1, (in_x$1 && "
+              "  int addr_$0 = select(-1, yc$2 * args.src_tensor.Width() + "
+              "xc$1, (in_x$1 && "
               "in_y$2));\n",
               y * block_size.x + x, x, y);
           c += absl::Substitute(
-              "  int dz_$0 = select(0, src_size.x * src_size.y, (in_x$1 && "
+              "  int dz_$0 = select(0, args.src_tensor.Width() * "
+              "args.src_tensor.Height(), (in_x$1 && "
               "in_y$2));\n",
               y * block_size.x + x, x, y);
         } else {
-          c += absl::Substitute("  int addr_$0 = yc$2 * src_size.x + xc$1;\n",
-                                y * block_size.x + x, x, y);
+          c += absl::Substitute(
+              "  int addr_$0 = yc$2 * args.src_tensor.Width() + xc$1;\n",
+              y * block_size.x + x, x, y);
         }
       }
     }
     if (src_tensor_type == TensorStorageType::BUFFER) {
-      c += "  int dz = src_size.x * src_size.y;\n";
+      c += "  int dz = args.src_tensor.Width() * args.src_tensor.Height();\n";
     }
   }
-  c += "  for (int s = 0; s < src_size.z; ++s) {\n";
+  c += "  for (int s = 0; s < args.src_tensor.Slices(); ++s) {\n";
   if (is_buffer) {
     if (src_tensor_type == TensorStorageType::IMAGE_BUFFER) {
       for (int index = 0; index < block_size.x * block_size.y; ++index) {
         const std::string id = std::to_string(index);
         c +=
-            "    FLT4 src" + id + " = " + src_tensor.Read("addr_" + id) + ";\n";
+            "    FLT4 src" + id + " = args.src_tensor.Read(addr_" + id + ");\n";
       }
     } else {
       for (int x = 0; x < block_size.x; ++x) {
         for (int y = 0; y < block_size.y; ++y) {
           const std::string id = std::to_string(y * block_size.x + x);
-          c += "    FLT4 src" + id + " = " + src_tensor.Read("addr_" + id) +
-               " * (FLT)(in_x" + xs[x] + " && in_y" + ys[y] + "); addr_" + id +
+          c += "    FLT4 src" + id + " = args.src_tensor.Read(addr_" + id +
+               ") * (FLT)(in_x" + xs[x] + " && in_y" + ys[y] + "); addr_" + id +
                " += dz;\n";
         }
       }
     }
   }
   for (int z = 0; z < block_size.z; ++z) {
-    const std::string fc = "(int2)(Z + " + zs[z] + ", " + f_y + ")";
-    c += absl::Substitute(R"(    FLT4 f$1 = READ_IMAGE(filters0, smp_none, $0);
-    FLT4 f$2 = READ_IMAGE(filters1, smp_none, $0);
-    FLT4 f$3 = READ_IMAGE(filters2, smp_none, $0);
-    FLT4 f$4 = READ_IMAGE(filters3, smp_none, $0);
+    c += absl::Substitute(R"(    FLT4 f$2 = args.weights0.Read($0, $1);
+    FLT4 f$3 = args.weights1.Read($0, $1);
+    FLT4 f$4 = args.weights2.Read($0, $1);
+    FLT4 f$5 = args.weights3.Read($0, $1);
 )",
-                          fc, z * 4 + 0, z * 4 + 1, z * 4 + 2, z * 4 + 3);
+                          "Z + " + zs[z], f_y, z * 4 + 0, z * 4 + 1, z * 4 + 2,
+                          z * 4 + 3);
   }
   if (!is_buffer) {
-    const auto mode = GetFastestZeroMode(device);
     for (int x = 0; x < block_size.x; ++x) {
       for (int y = 0; y < block_size.y; ++y) {
         const std::string id = std::to_string(y * block_size.x + x);
-        c += "    FLT4 src" + id + " = " +
-             src_tensor.ReadWHS(s_x[x], s_y[y], "s", mode) + ";\n";
+        c += "    FLT4 src" + id + " = args.src_tensor.Read(" + s_x[x] + ", " +
+             s_y[y] + ", s);\n";
       }
     }
   }
@@ -278,17 +284,17 @@ std::string GenerateConvCode(
       }
     }
   }
-  c += "  }\n";  // src_size.z
+  c += "  }\n";  // args.src_tensor.Slices()
   if (!is1x1) {
-    c += "  }\n";  // kernel_size.x
-    c += "  }\n";  // kernel_size.y
+    c += "  }\n";  // kernel_size_x
+    c += "  }\n";  // kernel_size_y
   }
   // when is1x1 && adreno4xx_optimization is true, xc0 == X and yc0 == Y
   std::string dst_x = is1x1 && adreno4xx_optimization ? "xc0" : "X";
   std::string dst_y = is1x1 && adreno4xx_optimization ? "yc0" : "Y";
   for (int z = 0; z < block_size.z; ++z) {
-    c += "  if (Z < dst_size.z) {\n";
-    c += "    FLT4 bias_val = READ_IMAGE(biases, smp_none, (int2)(Z, 0));\n";
+    c += "  if (Z < args.dst_tensor.Slices()) {\n";
+    c += "    FLT4 bias_val = args.biases.Read(Z);\n";
     for (int y = 0; y < block_size.y; ++y) {
       for (int x = 0; x < block_size.x; ++x) {
         const std::string id =
@@ -296,11 +302,10 @@ std::string GenerateConvCode(
         c += "    {\n";
         c += "      int xc = " + dst_x + " + " + xs[x] + ";\n";
         c += "      int yc = " + dst_y + " + " + ys[y] + ";\n";
-        c += "      if (xc < dst_size.x && yc < dst_size.y) {\n";
+        c += "      if (xc < args.dst_tensor.Width() && yc < "
+             "args.dst_tensor.Height()) {\n";
         c += "        FLT4 res = TO_FLT4(r" + id + ") + bias_val;\n";
-        const LinkingContext context{"res", "xc", "yc", "Z"};
-        c += PostProcess(linked_operations, context);
-        c += "        " + dst_tensor.WriteWHS("res", "xc", "yc", "Z") + "\n";
+        c += "        args.dst_tensor.Write(res, xc, yc, Z);\n";
         c += "      }\n";
         c += "    }\n";
       }
@@ -350,11 +355,6 @@ ConvTexture::ConvTexture(const OperationDef& definition)
 
 ConvTexture::ConvTexture(ConvTexture&& operation)
     : GPUOperation(std::move(operation)),
-      weights_0_(std::move(operation.weights_0_)),
-      weights_1_(std::move(operation.weights_1_)),
-      weights_2_(std::move(operation.weights_2_)),
-      weights_3_(std::move(operation.weights_3_)),
-      biases_(std::move(operation.biases_)),
       kernel_size_(operation.kernel_size_),
       stride_(operation.stride_),
       padding_(operation.padding_),
@@ -366,11 +366,6 @@ ConvTexture::ConvTexture(ConvTexture&& operation)
 
 ConvTexture& ConvTexture::operator=(ConvTexture&& operation) {
   if (this != &operation) {
-    weights_0_ = std::move(operation.weights_0_);
-    weights_1_ = std::move(operation.weights_1_);
-    weights_2_ = std::move(operation.weights_2_);
-    weights_3_ = std::move(operation.weights_3_);
-    biases_ = std::move(operation.biases_);
     std::swap(kernel_size_, operation.kernel_size_);
     std::swap(stride_, operation.stride_);
     std::swap(padding_, operation.padding_);
@@ -395,10 +390,16 @@ absl::Status ConvTexture::Compile(const CreationContext& creation_context) {
       definition_.precision == CalculationsPrecision::F16;
   const bool stride_correction =
       definition_.IsBatchSupported() && stride_.x != 1;
-  const std::string code =
+  std::string code =
       GenerateConvCode(definition_, block_size_, is1x1, adreno4xx_optimization,
                        stride_correction, different_weights_for_height_,
-                       *creation_context.device, linked_operations_);
+                       *creation_context.device, &args_);
+  std::string element_wise_code;
+  RETURN_IF_ERROR(
+      MergeOperations(linked_operations_, &args_, &element_wise_code));
+  RETURN_IF_ERROR(args_.TransformToCLCode(creation_context.device->GetInfo(),
+                                          {{"dst_tensor", element_wise_code}},
+                                          &code));
   std::vector options;
   if (UseFP16SIMD(*creation_context.device, definition_.precision, is1x1)) {
     options.push_back(CompilerOptions::ADRENO_FULL_SIMD_LINE);
@@ -409,25 +410,20 @@ absl::Status ConvTexture::Compile(const CreationContext& creation_context) {
 }
 
 absl::Status ConvTexture::BindArguments() {
-  kernel_.ResetBindingCounter();
-  RETURN_IF_ERROR(kernel_.SetMemoryAuto(src_[0]->GetMemoryPtr()));
-  RETURN_IF_ERROR(kernel_.SetMemoryAuto(weights_0_.GetMemoryPtr()));
-  RETURN_IF_ERROR(kernel_.SetMemoryAuto(weights_1_.GetMemoryPtr()));
-  RETURN_IF_ERROR(kernel_.SetMemoryAuto(weights_2_.GetMemoryPtr()));
-  RETURN_IF_ERROR(kernel_.SetMemoryAuto(weights_3_.GetMemoryPtr()));
-  RETURN_IF_ERROR(kernel_.SetMemoryAuto(biases_.GetMemoryPtr()));
-  RETURN_IF_ERROR(BindArgs(&kernel_, linked_operations_));
-  RETURN_IF_ERROR(kernel_.SetMemoryAuto(dst_[0]->GetMemoryPtrForWriting()));
-  RETURN_IF_ERROR(kernel_.SetBytesAuto(src_[0]->GetWBatchedHSB()));
-  RETURN_IF_ERROR(kernel_.SetBytesAuto(dst_[0]->GetWBatchedHSB()));
+  RETURN_IF_ERROR(args_.SetObjectRef("src_tensor", src_[0]));
+  RETURN_IF_ERROR(args_.SetObjectRef("dst_tensor", dst_[0]));
   if (!(kernel_size_.x == 1 && kernel_size_.y == 1)) {
-    RETURN_IF_ERROR(kernel_.SetBytesAuto(kernel_size_));
-    RETURN_IF_ERROR(kernel_.SetBytesAuto(
-        int2(dilation_.x * src_[0]->Batch(), dilation_.y)));
+    RETURN_IF_ERROR(args_.SetInt("kernel_size_x", kernel_size_.x));
+    RETURN_IF_ERROR(args_.SetInt("kernel_size_y", kernel_size_.y));
+    RETURN_IF_ERROR(args_.SetInt("dilation_x", dilation_.x * src_[0]->Batch()));
+    RETURN_IF_ERROR(args_.SetInt("dilation_y", dilation_.y));
   }
-  RETURN_IF_ERROR(kernel_.SetBytesAuto(stride_));
-  RETURN_IF_ERROR(
-      kernel_.SetBytesAuto(int2(padding_.x * src_[0]->Batch(), padding_.y)));
+  RETURN_IF_ERROR(args_.SetInt("stride_x", stride_.x));
+  RETURN_IF_ERROR(args_.SetInt("stride_y", stride_.y));
+  RETURN_IF_ERROR(args_.SetInt("padding_x", padding_.x * src_[0]->Batch()));
+  RETURN_IF_ERROR(args_.SetInt("padding_y", padding_.y));
+  RETURN_IF_ERROR(SetArguments(linked_operations_, &args_));
+  RETURN_IF_ERROR(args_.Bind(kernel_.kernel()));
   return absl::OkStatus();
 }
 
diff --git a/tensorflow/lite/delegates/gpu/cl/kernels/conv_texture.h b/tensorflow/lite/delegates/gpu/cl/kernels/conv_texture.h
index 42f7ecd51af..31c2a72021e 100644
--- a/tensorflow/lite/delegates/gpu/cl/kernels/conv_texture.h
+++ b/tensorflow/lite/delegates/gpu/cl/kernels/conv_texture.h
@@ -92,12 +92,6 @@ class ConvTexture : public GPUOperation {
   absl::Status BindArguments();
   int3 GetGridSize() const;
 
-  Texture2D weights_0_;
-  Texture2D weights_1_;
-  Texture2D weights_2_;
-  Texture2D weights_3_;
-  LinearStorage biases_;
-
   int2 kernel_size_;
   int2 stride_;
   int2 padding_;
@@ -119,11 +113,16 @@ absl::Status ConvTexture::UploadData(
     const tflite::gpu::Tensor& weights,
     const tflite::gpu::Tensor& biases, CLContext* context) {
   RETURN_IF_ERROR(UploadWeights(weights, context));
-  LinearStorageCreateInfo create_info;
-  create_info.storage_type = LinearStorageType::TEXTURE_2D;
-  create_info.data_type = definition_.GetDataType();
-  create_info.aligned_size = weights.shape.o;
-  RETURN_IF_ERROR(CreateLinearStorage(create_info, biases, context, &biases_));
+
+  TensorLinearDescriptor desc;
+  desc.storage_type = LinearStorageType::TEXTURE_2D;
+  desc.element_type = definition_.GetDataType();
+
+  LinearStorage lt;
+  RETURN_IF_ERROR(CreateLinearStorage(desc, biases, context, <));
+  args_.AddObject("biases", AccessType::READ,
+                  absl::make_unique(std::move(lt)),
+                  absl::make_unique(desc));
   return absl::OkStatus();
 }
 
@@ -135,14 +134,19 @@ absl::Status ConvTexture::UploadDataForWinograd4x4To6x6(
   RearrangeWeightsToWinograd4x4To6x6Weights(weights, &wino_weights);
   RETURN_IF_ERROR(UploadWeights(wino_weights, context));
 
-  LinearStorageCreateInfo create_info;
-  create_info.storage_type = LinearStorageType::TEXTURE_2D;
-  create_info.data_type = definition_.GetDataType();
-  create_info.aligned_size = 1;
   tflite::gpu::Tensor bias;
   bias.shape = Linear(1);
   bias.data = {0.0f};
-  return CreateLinearStorage(create_info, bias, context, &biases_);
+  TensorLinearDescriptor desc;
+  desc.storage_type = LinearStorageType::TEXTURE_2D;
+  desc.element_type = definition_.GetDataType();
+
+  LinearStorage lt;
+  RETURN_IF_ERROR(CreateLinearStorage(desc, bias, context, <));
+  args_.AddObject("biases", AccessType::READ,
+                  absl::make_unique(std::move(lt)),
+                  absl::make_unique(desc));
+  return absl::OkStatus();
 }
 
 template 
@@ -157,11 +161,20 @@ absl::Status ConvTexture::UploadWeights(
   int texture_width = dst_depth;
   int texture_height = src_depth * kernel_x * kernel_y;
 
-  DataType data_type = definition_.GetDataType();
+  const bool f32_weights = definition_.precision == CalculationsPrecision::F32;
+  DataType data_type = f32_weights ? DataType::FLOAT32 : DataType::FLOAT16;
 
   const int elements_count = texture_width * texture_height;
 
-  if (data_type == DataType::FLOAT32) {
+  Texture2DDescriptor desc;
+  desc.element_type = data_type;
+
+  Texture2D weights_0;
+  Texture2D weights_1;
+  Texture2D weights_2;
+  Texture2D weights_3;
+
+  if (f32_weights) {
     std::vector gpu_data_0(elements_count);
     std::vector gpu_data_1(elements_count);
     std::vector gpu_data_2(elements_count);
@@ -171,15 +184,16 @@ absl::Status ConvTexture::UploadWeights(
                          absl::MakeSpan(gpu_data_3));
     RETURN_IF_ERROR(CreateTexture2DRGBA(data_type, texture_width,
                                         texture_height, gpu_data_0.data(),
-                                        context, &weights_0_));
+                                        context, &weights_0));
     RETURN_IF_ERROR(CreateTexture2DRGBA(data_type, texture_width,
                                         texture_height, gpu_data_1.data(),
-                                        context, &weights_1_));
+                                        context, &weights_1));
     RETURN_IF_ERROR(CreateTexture2DRGBA(data_type, texture_width,
                                         texture_height, gpu_data_2.data(),
-                                        context, &weights_2_));
-    return CreateTexture2DRGBA(data_type, texture_width, texture_height,
-                               gpu_data_3.data(), context, &weights_3_);
+                                        context, &weights_2));
+    RETURN_IF_ERROR(CreateTexture2DRGBA(data_type, texture_width,
+                                        texture_height, gpu_data_3.data(),
+                                        context, &weights_3));
   } else {
     std::vector gpu_data_0(elements_count);
     std::vector gpu_data_1(elements_count);
@@ -190,16 +204,31 @@ absl::Status ConvTexture::UploadWeights(
                          absl::MakeSpan(gpu_data_3));
     RETURN_IF_ERROR(CreateTexture2DRGBA(data_type, texture_width,
                                         texture_height, gpu_data_0.data(),
-                                        context, &weights_0_));
+                                        context, &weights_0));
     RETURN_IF_ERROR(CreateTexture2DRGBA(data_type, texture_width,
                                         texture_height, gpu_data_1.data(),
-                                        context, &weights_1_));
+                                        context, &weights_1));
     RETURN_IF_ERROR(CreateTexture2DRGBA(data_type, texture_width,
                                         texture_height, gpu_data_2.data(),
-                                        context, &weights_2_));
-    return CreateTexture2DRGBA(data_type, texture_width, texture_height,
-                               gpu_data_3.data(), context, &weights_3_);
+                                        context, &weights_2));
+    RETURN_IF_ERROR(CreateTexture2DRGBA(data_type, texture_width,
+                                        texture_height, gpu_data_3.data(),
+                                        context, &weights_3));
   }
+
+  args_.AddObject("weights0", AccessType::READ,
+                  absl::make_unique(std::move(weights_0)),
+                  absl::make_unique(desc));
+  args_.AddObject("weights1", AccessType::READ,
+                  absl::make_unique(std::move(weights_1)),
+                  absl::make_unique(desc));
+  args_.AddObject("weights2", AccessType::READ,
+                  absl::make_unique(std::move(weights_2)),
+                  absl::make_unique(desc));
+  args_.AddObject("weights3", AccessType::READ,
+                  absl::make_unique(std::move(weights_3)),
+                  absl::make_unique(desc));
+  return absl::OkStatus();
 }
 
 template 
diff --git a/tensorflow/lite/delegates/gpu/cl/texture2d.cc b/tensorflow/lite/delegates/gpu/cl/texture2d.cc
index 022c15660ce..1b774c40862 100644
--- a/tensorflow/lite/delegates/gpu/cl/texture2d.cc
+++ b/tensorflow/lite/delegates/gpu/cl/texture2d.cc
@@ -59,6 +59,41 @@ absl::Status CreateTexture2D(int width, int height, cl_channel_type type,
 }
 }  // namespace
 
+GPUResources Texture2DDescriptor::GetGPUResources(
+    AccessType access_type) const {
+  GPUResources resources;
+  GPUImage2DDescriptor desc;
+  desc.data_type = element_type;
+  desc.access_type = access_type;
+  resources.images2d.push_back({"tex2d", desc});
+  return resources;
+}
+
+absl::Status Texture2DDescriptor::PerformSelector(
+    const std::string& selector, const std::vector& args,
+    const std::vector& template_args, std::string* result) const {
+  if (selector == "Read") {
+    return PerformReadSelector(args, result);
+  } else {
+    return absl::NotFoundError(absl::StrCat(
+        "TensorLinearDescriptor don't have selector with name - ", selector));
+  }
+}
+
+absl::Status Texture2DDescriptor::PerformReadSelector(
+    const std::vector& args, std::string* result) const {
+  if (args.size() != 2) {
+    return absl::NotFoundError(
+        absl::StrCat("Texture2DDescriptor Read require one argument, but ",
+                     args.size(), " was passed"));
+  }
+  const std::string read =
+      element_type == DataType::FLOAT16 ? "read_imageh" : "read_imagef";
+  *result = absl::StrCat(read, "(tex2d, smp_none, (int2)(", args[0],
+                         ", " + args[1] + "))");
+  return absl::OkStatus();
+}
+
 Texture2D::Texture2D(cl_mem texture, int width, int height,
                      cl_channel_type type)
     : texture_(texture), width_(width), height_(height), channel_type_(type) {}
@@ -95,6 +130,12 @@ void Texture2D::Release() {
   }
 }
 
+GPUResourcesWithValue Texture2D::GetGPUResources(AccessType access_type) const {
+  GPUResourcesWithValue resources;
+  resources.images2d.push_back({"tex2d", texture_});
+  return resources;
+}
+
 // Creates new 4-channel 2D texture with f32 elements
 absl::Status CreateTexture2DRGBA32F(int width, int height, CLContext* context,
                                     Texture2D* result) {
diff --git a/tensorflow/lite/delegates/gpu/cl/texture2d.h b/tensorflow/lite/delegates/gpu/cl/texture2d.h
index c12d8a2836c..cd41bb60aee 100644
--- a/tensorflow/lite/delegates/gpu/cl/texture2d.h
+++ b/tensorflow/lite/delegates/gpu/cl/texture2d.h
@@ -20,6 +20,7 @@ limitations under the License.
 #include "absl/types/span.h"
 #include "tensorflow/lite/delegates/gpu/cl/cl_command_queue.h"
 #include "tensorflow/lite/delegates/gpu/cl/cl_context.h"
+#include "tensorflow/lite/delegates/gpu/cl/gpu_object.h"
 #include "tensorflow/lite/delegates/gpu/cl/opencl_wrapper.h"
 #include "tensorflow/lite/delegates/gpu/cl/tensor_type.h"
 #include "tensorflow/lite/delegates/gpu/cl/util.h"
@@ -30,9 +31,22 @@ namespace tflite {
 namespace gpu {
 namespace cl {
 
+struct Texture2DDescriptor : public GPUObjectDescriptor {
+  DataType element_type;  // FLOAT32 or FLOAT16
+
+  absl::Status PerformSelector(const std::string& selector,
+                               const std::vector& args,
+                               const std::vector& template_args,
+                               std::string* result) const override;
+
+  GPUResources GetGPUResources(AccessType access_type) const override;
+  absl::Status PerformReadSelector(const std::vector& args,
+                                   std::string* result) const;
+};
+
 // Texture2D represent formatted GPU data storage.
 // Texture2D is moveable but not copyable.
-class Texture2D {
+class Texture2D : public GPUObject {
  public:
   Texture2D() {}  // just for using Texture2D as a class members
   Texture2D(cl_mem texture, int width, int height, cl_channel_type type);
@@ -56,6 +70,8 @@ class Texture2D {
   template 
   absl::Status ReadData(CLCommandQueue* queue, std::vector* result) const;
 
+  GPUResourcesWithValue GetGPUResources(AccessType access_type) const override;
+
  private:
   void Release();
 

From 094340c9d4147a5be79bd37be71acc5a687e7f41 Mon Sep 17 00:00:00 2001
From: Dan Moldovan 
Date: Mon, 22 Jun 2020 15:58:38 -0700
Subject: [PATCH 491/867] Add a workaround for slowness of tf.while_loop in the
 default executor when maximum_iterations is set. Fixes #40517.

PiperOrigin-RevId: 317753132
Change-Id: I7b5767019ceebd3f21975a990e7c2c05dd878ca6
---
 .../python/autograph/operators/control_flow.py    | 15 +++++++++++----
 1 file changed, 11 insertions(+), 4 deletions(-)

diff --git a/tensorflow/python/autograph/operators/control_flow.py b/tensorflow/python/autograph/operators/control_flow.py
index 77db7579ece..8b4e97c238c 100644
--- a/tensorflow/python/autograph/operators/control_flow.py
+++ b/tensorflow/python/autograph/operators/control_flow.py
@@ -80,6 +80,7 @@ from tensorflow.python.framework import func_graph
 from tensorflow.python.framework import ops
 from tensorflow.python.framework import tensor_util
 from tensorflow.python.ops import control_flow_ops
+from tensorflow.python.ops import control_flow_util
 from tensorflow.python.ops import math_ops
 from tensorflow.python.ops import tensor_array_ops
 from tensorflow.python.ops.ragged import ragged_tensor
@@ -429,7 +430,9 @@ def _known_len_tf_for_stmt(
       return control_flow_ops.cond(main_test, extra_test, lambda: False)
     return main_test
 
-  opts['maximum_iterations'] = n
+  # TODO(b/159186914): Remove.
+  if not control_flow_util.GraphOrParentsInXlaContext(ops.get_default_graph()):
+    opts['maximum_iterations'] = n
 
   _tf_while_stmt(
       aug_test,
@@ -475,7 +478,9 @@ def _tf_ragged_for_stmt(
       return control_flow_ops.cond(main_test, extra_test, lambda: False)
     return main_test
 
-  opts['maximum_iterations'] = n
+  # TODO(b/159186914): Remove.
+  if not control_flow_util.GraphOrParentsInXlaContext(ops.get_default_graph()):
+    opts['maximum_iterations'] = n
 
   _tf_while_stmt(
       aug_test,
@@ -524,8 +529,10 @@ def _tf_range_for_stmt(
       return control_flow_ops.cond(main_test, extra_test, lambda: False)
     return main_test
 
-  opts['maximum_iterations'] = math_ops.cast(
-      misc.get_range_len(start, limit, delta), dtypes.int32)
+  # TODO(b/159186914): Remove.
+  if not control_flow_util.GraphOrParentsInXlaContext(ops.get_default_graph()):
+    opts['maximum_iterations'] = math_ops.cast(
+        misc.get_range_len(start, limit, delta), dtypes.int32)
 
   _tf_while_stmt(
       aug_test,

From 8d99ecdbd50face521441389f6ee02c8d682ce68 Mon Sep 17 00:00:00 2001
From: "Ahmed S. Taei" 
Date: Mon, 22 Jun 2020 16:02:10 -0700
Subject: [PATCH 492/867] Add AArch64/GISel/*.cpp to AArch64CodeGen srcs.

PiperOrigin-RevId: 317753780
Change-Id: I33c44d6355cca3b9a054773afad47b6d00ba6d9f
---
 third_party/llvm/llvm.autogenerated.BUILD | 1 +
 1 file changed, 1 insertion(+)

diff --git a/third_party/llvm/llvm.autogenerated.BUILD b/third_party/llvm/llvm.autogenerated.BUILD
index 50b71a3686f..c70ff559165 100644
--- a/third_party/llvm/llvm.autogenerated.BUILD
+++ b/third_party/llvm/llvm.autogenerated.BUILD
@@ -698,6 +698,7 @@ cc_library(
         "lib/Target/AArch64/*.c",
         "lib/Target/AArch64/*.cpp",
         "lib/Target/AArch64/*.inc",
+        "lib/Target/AArch64/GISel/*.cpp",
     ]),
     hdrs = glob([
         "include/llvm/Target/AArch64/*.h",

From b5adbbcf0e2722d47b0533c665d18df96e91c56e Mon Sep 17 00:00:00 2001
From: "A. Unique TensorFlower" 
Date: Mon, 22 Jun 2020 16:05:37 -0700
Subject: [PATCH 493/867] tf numpy: Interop test for tf.Variable.

PiperOrigin-RevId: 317754511
Change-Id: I6bcabccaa626039c53269bf7e8837fef972bd537
---
 .../python/ops/numpy_ops/np_interop_test.py   | 78 +++++++++++++++----
 1 file changed, 62 insertions(+), 16 deletions(-)

diff --git a/tensorflow/python/ops/numpy_ops/np_interop_test.py b/tensorflow/python/ops/numpy_ops/np_interop_test.py
index 836bce61494..17a3bf81a01 100644
--- a/tensorflow/python/ops/numpy_ops/np_interop_test.py
+++ b/tensorflow/python/ops/numpy_ops/np_interop_test.py
@@ -94,25 +94,11 @@ class InteropTest(tf.test.TestCase):
     dx, dy = t.gradient([xx, yy], [x, y])
 
     # # TODO(nareshmodi): Figure out a way to rewrap ndarray as tensors.
-    # self.assertIsInstance(dx, np_arrays.ndarray)
-    # self.assertIsInstance(dy, np_arrays.ndarray)
+    # self.assertIsInstance(dx, np.ndarray)
+    # self.assertIsInstance(dy, np.ndarray)
     self.assertAllClose(dx, 2.0)
     self.assertAllClose(dy, 3.0)
 
-  def testFunctionInterop(self):
-    x = np.asarray(3.0)
-    y = np.asarray(2.0)
-
-    add = lambda x, y: x + y
-    add_fn = tf.function(add)
-
-    raw_result = add(x, y)
-    fn_result = add_fn(x, y)
-
-    self.assertIsInstance(raw_result, np.ndarray)
-    self.assertIsInstance(fn_result, np.ndarray)
-    self.assertAllClose(raw_result, fn_result)
-
   def testCondInterop(self):
     x = np.asarray(3.0)
 
@@ -222,6 +208,66 @@ class InteropTest(tf.test.TestCase):
     # self.assertIsInstance(reduced, np.ndarray)
     self.assertAllClose(reduced, 15)
 
+
+class FunctionTest(InteropTest):
+
+  def testFunctionInterop(self):
+    x = np.asarray(3.0)
+    y = np.asarray(2.0)
+
+    add = lambda x, y: x + y
+    add_fn = tf.function(add)
+
+    raw_result = add(x, y)
+    fn_result = add_fn(x, y)
+
+    self.assertIsInstance(raw_result, np.ndarray)
+    self.assertIsInstance(fn_result, np.ndarray)
+    self.assertAllClose(raw_result, fn_result)
+
+  def testLen(self):
+
+    @tf.function
+    def f(x):
+      # Note that shape of input to len is data dependent.
+      return len(np.where(x)[0])
+
+    t = np.asarray([True, False, True])
+    with self.assertRaises(TypeError):
+      f(t)
+
+  def testIter(self):
+
+    @tf.function
+    def f(x):
+      y, z = x
+      return y, z
+
+    with self.assertRaises(TypeError):
+      f(np.asarray([3, 4]))
+
+  def testIndex(self):
+
+    @tf.function
+    def f(x):
+      return [0, 1][x]
+
+    with self.assertRaises(TypeError):
+      f(np.asarray([1]))
+
+
+class VariableTest(InteropTest):
+
+  def test(self):
+    tf_var = tf.Variable(2.0)
+    value = np.square(tf_var)
+    self.assertIsInstance(value, np.ndarray)
+    self.assertAllClose(4.0, value)
+    with tf.control_dependencies([tf_var.assign_add(value)]):
+      tf_var_value = tf_var.read_value()
+    self.assertAllClose(6.0, tf_var_value)
+
+
 if __name__ == '__main__':
   tf.compat.v1.enable_eager_execution()
   tf.test.main()

From 8ad4b3c4d5ad0e6a38fb0bc60651e3a0032ef52e Mon Sep 17 00:00:00 2001
From: Lukas Geiger 
Date: Tue, 23 Jun 2020 01:54:00 +0200
Subject: [PATCH 494/867] Call Context.ensure_initialized() method directly

---
 tensorflow/python/eager/context.py | 4 ++--
 1 file changed, 2 insertions(+), 2 deletions(-)

diff --git a/tensorflow/python/eager/context.py b/tensorflow/python/eager/context.py
index e1f3c58496b..96481d8bd48 100644
--- a/tensorflow/python/eager/context.py
+++ b/tensorflow/python/eager/context.py
@@ -890,13 +890,13 @@ class Context(object):
 
   @property
   def executor(self):
-    ensure_initialized()
+    self.ensure_initialized()
     return executor.Executor(
         pywrap_tfe.TFE_ContextGetExecutorForThread(self._context_handle))
 
   @executor.setter
   def executor(self, e):
-    ensure_initialized()
+    self.ensure_initialized()
     pywrap_tfe.TFE_ContextSetExecutorForThread(self._context_handle, e.handle())
 
   @property

From 76ac3e6a5ff4a1f60bc560dd069f5ba15eb57de9 Mon Sep 17 00:00:00 2001
From: Peng Wang 
Date: Mon, 22 Jun 2020 16:08:32 -0700
Subject: [PATCH 495/867] Adds NumPy LICENSE and lists of NumPy API symbols
 that `tf.experimental.numpy` implements to tensorflow/third_party/py/numpy/

PiperOrigin-RevId: 317755065
Change-Id: I0efd21b4bf917b7f14ec30e0b5710954844de448
---
 third_party/py/numpy/LICENSE                  |  60 ++
 third_party/py/numpy/README.md                |   4 +
 .../tf_numpy_api/numpy_ops.ndarray.pbtxt      |  58 ++
 .../py/numpy/tf_numpy_api/numpy_ops.pbtxt     | 903 ++++++++++++++++++
 .../numpy/tf_numpy_api/numpy_ops.random.pbtxt |  27 +
 5 files changed, 1052 insertions(+)
 create mode 100644 third_party/py/numpy/LICENSE
 create mode 100644 third_party/py/numpy/README.md
 create mode 100644 third_party/py/numpy/tf_numpy_api/numpy_ops.ndarray.pbtxt
 create mode 100644 third_party/py/numpy/tf_numpy_api/numpy_ops.pbtxt
 create mode 100644 third_party/py/numpy/tf_numpy_api/numpy_ops.random.pbtxt

diff --git a/third_party/py/numpy/LICENSE b/third_party/py/numpy/LICENSE
new file mode 100644
index 00000000000..b9731f734f5
--- /dev/null
+++ b/third_party/py/numpy/LICENSE
@@ -0,0 +1,60 @@
+Copyright (c) 2005-2019, NumPy Developers.
+All rights reserved.
+
+Redistribution and use in source and binary forms, with or without
+modification, are permitted provided that the following conditions are
+met:
+
+    * Redistributions of source code must retain the above copyright
+       notice, this list of conditions and the following disclaimer.
+
+    * Redistributions in binary form must reproduce the above
+       copyright notice, this list of conditions and the following
+       disclaimer in the documentation and/or other materials provided
+       with the distribution.
+
+    * Neither the name of the NumPy Developers nor the names of any
+       contributors may be used to endorse or promote products derived
+       from this software without specific prior written permission.
+
+THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
+"AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
+LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
+A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
+OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
+SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
+LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
+DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
+THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
+(INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
+OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+
+
+
+The NumPy repository and source distributions bundle several libraries that are
+compatibly licensed.  We list these here.
+
+Name: Numpydoc
+Files: doc/sphinxext/numpydoc/*
+License: 2-clause BSD
+  For details, see doc/sphinxext/LICENSE.txt
+
+Name: scipy-sphinx-theme
+Files: doc/scipy-sphinx-theme/*
+License: 3-clause BSD, PSF and Apache 2.0
+  For details, see doc/scipy-sphinx-theme/LICENSE.txt
+
+Name: lapack-lite
+Files: numpy/linalg/lapack_lite/*
+License: 3-clause BSD
+  For details, see numpy/linalg/lapack_lite/LICENSE.txt
+
+Name: tempita
+Files: tools/npy_tempita/*
+License: BSD derived
+  For details, see tools/npy_tempita/license.txt
+
+Name: dragon4
+Files: numpy/core/src/multiarray/dragon4.c
+License: One of a kind
+  For license text, see numpy/core/src/multiarray/dragon4.c
diff --git a/third_party/py/numpy/README.md b/third_party/py/numpy/README.md
new file mode 100644
index 00000000000..4e58b9df87b
--- /dev/null
+++ b/third_party/py/numpy/README.md
@@ -0,0 +1,4 @@
+# numpy_ops
+
+The folder tf_numpy_api/ contains lists of NumPy API symbols that the
+`numpy_ops` internal module in TensorFlow implements.
diff --git a/third_party/py/numpy/tf_numpy_api/numpy_ops.ndarray.pbtxt b/third_party/py/numpy/tf_numpy_api/numpy_ops.ndarray.pbtxt
new file mode 100644
index 00000000000..8492a30d81b
--- /dev/null
+++ b/third_party/py/numpy/tf_numpy_api/numpy_ops.ndarray.pbtxt
@@ -0,0 +1,58 @@
+path: "numpy_ops.ndarray"
+tf_class {
+  is_instance: ""
+  is_instance: ""
+  is_instance: ""
+  member {
+    name: "T"
+    mtype: ""
+  }
+  member {
+    name: "data"
+    mtype: ""
+  }
+  member {
+    name: "dtype"
+    mtype: ""
+  }
+  member {
+    name: "ndim"
+    mtype: ""
+  }
+  member {
+    name: "shape"
+    mtype: ""
+  }
+  member {
+    name: "size"
+    mtype: ""
+  }
+  member_method {
+    name: "__init__"
+    argspec: "args=[\'self\', \'shape\', \'dtype\', \'buffer\'], varargs=None, keywords=None, defaults=[\"\", \'None\'], "
+  }
+  member_method {
+    name: "astype"
+    argspec: "args=[\'self\', \'dtype\'], varargs=None, keywords=None, defaults=None"
+  }
+  member_method {
+    name: "from_tensor"
+    argspec: "args=[\'cls\', \'tensor\'], varargs=None, keywords=None, defaults=None"
+  }
+  member_method {
+    name: "ravel"
+    argspec: "args=[\'a\'], varargs=None, keywords=None, defaults=None"
+  }
+  member_method {
+    name: "reshape"
+    argspec: "args=[\'a\'], varargs=newshape, keywords=kwargs, defaults=None"
+  }
+  member_method {
+    name: "tolist"
+    argspec: "args=[\'self\'], varargs=None, keywords=None, defaults=None"
+  }
+  member_method {
+    name: "transpose"
+    argspec: "args=[\'a\', \'axes\'], varargs=None, keywords=None, defaults=[\'None\'], "
+  }
+}
diff --git a/third_party/py/numpy/tf_numpy_api/numpy_ops.pbtxt b/third_party/py/numpy/tf_numpy_api/numpy_ops.pbtxt
new file mode 100644
index 00000000000..30913665f14
--- /dev/null
+++ b/third_party/py/numpy/tf_numpy_api/numpy_ops.pbtxt
@@ -0,0 +1,903 @@
+path: "numpy_ops"
+tf_module {
+  member {
+    name: "bool_"
+    mtype: ""
+  }
+  member {
+    name: "complex128"
+    mtype: ""
+  }
+  member {
+    name: "complex64"
+    mtype: ""
+  }
+  member {
+    name: "complex_"
+    mtype: ""
+  }
+  member {
+    name: "e"
+    mtype: ""
+  }
+  member {
+    name: "float16"
+    mtype: ""
+  }
+  member {
+    name: "float32"
+    mtype: ""
+  }
+  member {
+    name: "float64"
+    mtype: ""
+  }
+  member {
+    name: "float_"
+    mtype: ""
+  }
+  member {
+    name: "iinfo"
+    mtype: ""
+  }
+  member {
+    name: "inexact"
+    mtype: ""
+  }
+  member {
+    name: "inf"
+    mtype: ""
+  }
+  member {
+    name: "int16"
+    mtype: ""
+  }
+  member {
+    name: "int32"
+    mtype: ""
+  }
+  member {
+    name: "int64"
+    mtype: ""
+  }
+  member {
+    name: "int8"
+    mtype: ""
+  }
+  member {
+    name: "int_"
+    mtype: ""
+  }
+  member {
+    name: "ndarray"
+    mtype: ""
+  }
+  member {
+    name: "object_"
+    mtype: ""
+  }
+  member {
+    name: "pi"
+    mtype: ""
+  }
+  member {
+    name: "random"
+    mtype: ""
+  }
+  member {
+    name: "string_"
+    mtype: ""
+  }
+  member {
+    name: "uint16"
+    mtype: ""
+  }
+  member {
+    name: "uint32"
+    mtype: ""
+  }
+  member {
+    name: "uint64"
+    mtype: ""
+  }
+  member {
+    name: "uint8"
+    mtype: ""
+  }
+  member {
+    name: "unicode_"
+    mtype: ""
+  }
+  member_method {
+    name: "abs"
+    argspec: "args=[\'x\'], varargs=None, keywords=None, defaults=None"
+  }
+  member_method {
+    name: "absolute"
+    argspec: "args=[\'x\'], varargs=None, keywords=None, defaults=None"
+  }
+  member_method {
+    name: "add"
+    argspec: "args=[\'x1\', \'x2\'], varargs=None, keywords=None, defaults=None"
+  }
+  member_method {
+    name: "all"
+    argspec: "args=[\'a\', \'axis\', \'keepdims\'], varargs=None, keywords=None, defaults=[\'None\', \'None\'], "
+  }
+  member_method {
+    name: "allclose"
+    argspec: "args=[\'a\', \'b\', \'rtol\', \'atol\', \'equal_nan\'], varargs=None, keywords=None, defaults=[\'1e-05\', \'1e-08\', \'False\'], "
+  }
+  member_method {
+    name: "amax"
+    argspec: "args=[\'a\', \'axis\', \'keepdims\'], varargs=None, keywords=None, defaults=[\'None\', \'None\'], "
+  }
+  member_method {
+    name: "amin"
+    argspec: "args=[\'a\', \'axis\', \'keepdims\'], varargs=None, keywords=None, defaults=[\'None\', \'None\'], "
+  }
+  member_method {
+    name: "angle"
+    argspec: "args=[\'z\', \'deg\'], varargs=None, keywords=None, defaults=[\'False\'], "
+  }
+  member_method {
+    name: "any"
+    argspec: "args=[\'a\', \'axis\', \'keepdims\'], varargs=None, keywords=None, defaults=[\'None\', \'None\'], "
+  }
+  member_method {
+    name: "append"
+    argspec: "args=[\'arr\', \'values\', \'axis\'], varargs=None, keywords=None, defaults=[\'None\'], "
+  }
+  member_method {
+    name: "arange"
+    argspec: "args=[\'start\', \'stop\', \'step\', \'dtype\'], varargs=None, keywords=None, defaults=[\'None\', \'1\', \'None\'], "
+  }
+  member_method {
+    name: "arccos"
+    argspec: "args=[\'x\'], varargs=None, keywords=None, defaults=None"
+  }
+  member_method {
+    name: "arccosh"
+    argspec: "args=[\'x\'], varargs=None, keywords=None, defaults=None"
+  }
+  member_method {
+    name: "arcsin"
+    argspec: "args=[\'x\'], varargs=None, keywords=None, defaults=None"
+  }
+  member_method {
+    name: "arcsinh"
+    argspec: "args=[\'x\'], varargs=None, keywords=None, defaults=None"
+  }
+  member_method {
+    name: "arctan"
+    argspec: "args=[\'x\'], varargs=None, keywords=None, defaults=None"
+  }
+  member_method {
+    name: "arctan2"
+    argspec: "args=[\'x1\', \'x2\'], varargs=None, keywords=None, defaults=None"
+  }
+  member_method {
+    name: "arctanh"
+    argspec: "args=[\'x\'], varargs=None, keywords=None, defaults=None"
+  }
+  member_method {
+    name: "argmax"
+    argspec: "args=[\'a\', \'axis\'], varargs=None, keywords=None, defaults=[\'None\'], "
+  }
+  member_method {
+    name: "argmin"
+    argspec: "args=[\'a\', \'axis\'], varargs=None, keywords=None, defaults=[\'None\'], "
+  }
+  member_method {
+    name: "argsort"
+    argspec: "args=[\'a\', \'axis\', \'kind\', \'order\'], varargs=None, keywords=None, defaults=[\'-1\', \'quicksort\', \'None\'], "
+  }
+  member_method {
+    name: "around"
+    argspec: "args=[\'a\', \'decimals\'], varargs=None, keywords=None, defaults=[\'0\'], "
+  }
+  member_method {
+    name: "array"
+    argspec: "args=[\'val\', \'dtype\', \'copy\', \'ndmin\'], varargs=None, keywords=None, defaults=[\'None\', \'True\', \'0\'], "
+  }
+  member_method {
+    name: "array_equal"
+    argspec: "args=[\'a1\', \'a2\'], varargs=None, keywords=None, defaults=None"
+  }
+  member_method {
+    name: "asanyarray"
+    argspec: "args=[\'a\', \'dtype\'], varargs=None, keywords=None, defaults=[\'None\'], "
+  }
+  member_method {
+    name: "asarray"
+    argspec: "args=[\'a\', \'dtype\'], varargs=None, keywords=None, defaults=[\'None\'], "
+  }
+  member_method {
+    name: "ascontiguousarray"
+    argspec: "args=[\'a\', \'dtype\'], varargs=None, keywords=None, defaults=[\'None\'], "
+  }
+  member_method {
+    name: "atleast_1d"
+    argspec: "args=[], varargs=arys, keywords=None, defaults=None"
+  }
+  member_method {
+    name: "atleast_2d"
+    argspec: "args=[], varargs=arys, keywords=None, defaults=None"
+  }
+  member_method {
+    name: "atleast_3d"
+    argspec: "args=[], varargs=arys, keywords=None, defaults=None"
+  }
+  member_method {
+    name: "average"
+    argspec: "args=[\'a\', \'axis\', \'weights\', \'returned\'], varargs=None, keywords=None, defaults=[\'None\', \'None\', \'False\'], "
+  }
+  member_method {
+    name: "bitwise_and"
+    argspec: "args=[\'x1\', \'x2\'], varargs=None, keywords=None, defaults=None"
+  }
+  member_method {
+    name: "bitwise_not"
+    argspec: "args=[\'x\'], varargs=None, keywords=None, defaults=None"
+  }
+  member_method {
+    name: "bitwise_or"
+    argspec: "args=[\'x1\', \'x2\'], varargs=None, keywords=None, defaults=None"
+  }
+  member_method {
+    name: "bitwise_xor"
+    argspec: "args=[\'x1\', \'x2\'], varargs=None, keywords=None, defaults=None"
+  }
+  member_method {
+    name: "broadcast_arrays"
+    argspec: "args=[], varargs=args, keywords=kwargs, defaults=None"
+  }
+  member_method {
+    name: "broadcast_to"
+    argspec: "args=[\'array\', \'shape\'], varargs=None, keywords=None, defaults=None"
+  }
+  member_method {
+    name: "cbrt"
+    argspec: "args=[\'x\'], varargs=None, keywords=None, defaults=None"
+  }
+  member_method {
+    name: "ceil"
+    argspec: "args=[\'x\'], varargs=None, keywords=None, defaults=None"
+  }
+  member_method {
+    name: "clip"
+    argspec: "args=[\'a\', \'a_min\', \'a_max\'], varargs=None, keywords=None, defaults=None"
+  }
+  member_method {
+    name: "compress"
+    argspec: "args=[\'condition\', \'a\', \'axis\'], varargs=None, keywords=None, defaults=[\'None\'], "
+  }
+  member_method {
+    name: "concatenate"
+    argspec: "args=[\'arys\', \'axis\'], varargs=None, keywords=None, defaults=[\'0\'], "
+  }
+  member_method {
+    name: "conj"
+    argspec: "args=[\'x\'], varargs=None, keywords=None, defaults=None"
+  }
+  member_method {
+    name: "conjugate"
+    argspec: "args=[\'x\'], varargs=None, keywords=None, defaults=None"
+  }
+  member_method {
+    name: "copy"
+    argspec: "args=[\'a\'], varargs=None, keywords=None, defaults=None"
+  }
+  member_method {
+    name: "cos"
+    argspec: "args=[\'x\'], varargs=None, keywords=None, defaults=None"
+  }
+  member_method {
+    name: "cosh"
+    argspec: "args=[\'x\'], varargs=None, keywords=None, defaults=None"
+  }
+  member_method {
+    name: "count_nonzero"
+    argspec: "args=[\'a\', \'axis\'], varargs=None, keywords=None, defaults=[\'None\'], "
+  }
+  member_method {
+    name: "cross"
+    argspec: "args=[\'a\', \'b\', \'axisa\', \'axisb\', \'axisc\', \'axis\'], varargs=None, keywords=None, defaults=[\'-1\', \'-1\', \'-1\', \'None\'], "
+  }
+  member_method {
+    name: "cumprod"
+    argspec: "args=[\'a\', \'axis\', \'dtype\'], varargs=None, keywords=None, defaults=[\'None\', \'None\'], "
+  }
+  member_method {
+    name: "cumsum"
+    argspec: "args=[\'a\', \'axis\', \'dtype\'], varargs=None, keywords=None, defaults=[\'None\', \'None\'], "
+  }
+  member_method {
+    name: "deg2rad"
+    argspec: "args=[\'x\'], varargs=None, keywords=None, defaults=None"
+  }
+  member_method {
+    name: "diag"
+    argspec: "args=[\'v\', \'k\'], varargs=None, keywords=None, defaults=[\'0\'], "
+  }
+  member_method {
+    name: "diag_indices"
+    argspec: "args=[\'n\', \'ndim\'], varargs=None, keywords=None, defaults=[\'2\'], "
+  }
+  member_method {
+    name: "diagflat"
+    argspec: "args=[\'v\', \'k\'], varargs=None, keywords=None, defaults=[\'0\'], "
+  }
+  member_method {
+    name: "diagonal"
+    argspec: "args=[\'a\', \'offset\', \'axis1\', \'axis2\'], varargs=None, keywords=None, defaults=[\'0\', \'0\', \'1\'], "
+  }
+  member_method {
+    name: "diff"
+    argspec: "args=[\'a\', \'n\', \'axis\'], varargs=None, keywords=None, defaults=[\'1\', \'-1\'], "
+  }
+  member_method {
+    name: "divide"
+    argspec: "args=[\'x1\', \'x2\'], varargs=None, keywords=None, defaults=None"
+  }
+  member_method {
+    name: "divmod"
+    argspec: "args=[\'x1\', \'x2\'], varargs=None, keywords=None, defaults=None"
+  }
+  member_method {
+    name: "dot"
+    argspec: "args=[\'a\', \'b\'], varargs=None, keywords=None, defaults=None"
+  }
+  member_method {
+    name: "dsplit"
+    argspec: "args=[\'ary\', \'indices_or_sections\'], varargs=None, keywords=None, defaults=None"
+  }
+  member_method {
+    name: "dstack"
+    argspec: "args=[\'tup\'], varargs=None, keywords=None, defaults=None"
+  }
+  member_method {
+    name: "einsum"
+    argspec: "args=[\'subscripts\'], varargs=operands, keywords=kwargs, defaults=None"
+  }
+  member_method {
+    name: "empty"
+    argspec: "args=[\'shape\', \'dtype\'], varargs=None, keywords=None, defaults=[\"\"], "
+  }
+  member_method {
+    name: "empty_like"
+    argspec: "args=[\'a\', \'dtype\'], varargs=None, keywords=None, defaults=[\'None\'], "
+  }
+  member_method {
+    name: "equal"
+    argspec: "args=[\'x1\', \'x2\'], varargs=None, keywords=None, defaults=None"
+  }
+  member_method {
+    name: "exp"
+    argspec: "args=[\'x\'], varargs=None, keywords=None, defaults=None"
+  }
+  member_method {
+    name: "exp2"
+    argspec: "args=[\'x\'], varargs=None, keywords=None, defaults=None"
+  }
+  member_method {
+    name: "expand_dims"
+    argspec: "args=[\'a\', \'axis\'], varargs=None, keywords=None, defaults=None"
+  }
+  member_method {
+    name: "expm1"
+    argspec: "args=[\'x\'], varargs=None, keywords=None, defaults=None"
+  }
+  member_method {
+    name: "eye"
+    argspec: "args=[\'N\', \'M\', \'k\', \'dtype\'], varargs=None, keywords=None, defaults=[\'None\', \'0\', \"\"], "
+  }
+  member_method {
+    name: "fabs"
+    argspec: "args=[\'x\'], varargs=None, keywords=None, defaults=None"
+  }
+  member_method {
+    name: "finfo"
+    argspec: "args=[\'dtype\'], varargs=None, keywords=None, defaults=None"
+  }
+  member_method {
+    name: "fix"
+    argspec: "args=[\'x\'], varargs=None, keywords=None, defaults=None"
+  }
+  member_method {
+    name: "flip"
+    argspec: "args=[\'m\', \'axis\'], varargs=None, keywords=None, defaults=[\'None\'], "
+  }
+  member_method {
+    name: "fliplr"
+    argspec: "args=[\'m\'], varargs=None, keywords=None, defaults=None"
+  }
+  member_method {
+    name: "flipud"
+    argspec: "args=[\'m\'], varargs=None, keywords=None, defaults=None"
+  }
+  member_method {
+    name: "float_power"
+    argspec: "args=[\'x1\', \'x2\'], varargs=None, keywords=None, defaults=None"
+  }
+  member_method {
+    name: "floor"
+    argspec: "args=[\'x\'], varargs=None, keywords=None, defaults=None"
+  }
+  member_method {
+    name: "floor_divide"
+    argspec: "args=[\'x1\', \'x2\'], varargs=None, keywords=None, defaults=None"
+  }
+  member_method {
+    name: "full"
+    argspec: "args=[\'shape\', \'fill_value\', \'dtype\'], varargs=None, keywords=None, defaults=[\'None\'], "
+  }
+  member_method {
+    name: "full_like"
+    argspec: "args=[\'a\', \'fill_value\', \'dtype\', \'order\', \'subok\', \'shape\'], varargs=None, keywords=None, defaults=[\'None\', \'K\', \'True\', \'None\'], "
+  }
+  member_method {
+    name: "gcd"
+    argspec: "args=[\'x1\', \'x2\'], varargs=None, keywords=None, defaults=None"
+  }
+  member_method {
+    name: "geomspace"
+    argspec: "args=[\'start\', \'stop\', \'num\', \'endpoint\', \'dtype\', \'axis\'], varargs=None, keywords=None, defaults=[\'50\', \'True\', \'None\', \'0\'], "
+  }
+  member_method {
+    name: "greater"
+    argspec: "args=[\'x1\', \'x2\'], varargs=None, keywords=None, defaults=None"
+  }
+  member_method {
+    name: "greater_equal"
+    argspec: "args=[\'x1\', \'x2\'], varargs=None, keywords=None, defaults=None"
+  }
+  member_method {
+    name: "heaviside"
+    argspec: "args=[\'x1\', \'x2\'], varargs=None, keywords=None, defaults=None"
+  }
+  member_method {
+    name: "hsplit"
+    argspec: "args=[\'ary\', \'indices_or_sections\'], varargs=None, keywords=None, defaults=None"
+  }
+  member_method {
+    name: "hstack"
+    argspec: "args=[\'tup\'], varargs=None, keywords=None, defaults=None"
+  }
+  member_method {
+    name: "hypot"
+    argspec: "args=[\'x1\', \'x2\'], varargs=None, keywords=None, defaults=None"
+  }
+  member_method {
+    name: "identity"
+    argspec: "args=[\'n\', \'dtype\'], varargs=None, keywords=None, defaults=[\"\"], "
+  }
+  member_method {
+    name: "imag"
+    argspec: "args=[\'a\'], varargs=None, keywords=None, defaults=None"
+  }
+  member_method {
+    name: "inner"
+    argspec: "args=[\'a\', \'b\'], varargs=None, keywords=None, defaults=None"
+  }
+  member_method {
+    name: "isclose"
+    argspec: "args=[\'a\', \'b\', \'rtol\', \'atol\', \'equal_nan\'], varargs=None, keywords=None, defaults=[\'1e-05\', \'1e-08\', \'False\'], "
+  }
+  member_method {
+    name: "iscomplex"
+    argspec: "args=[\'x\'], varargs=None, keywords=None, defaults=None"
+  }
+  member_method {
+    name: "iscomplexobj"
+    argspec: "args=[\'x\'], varargs=None, keywords=None, defaults=None"
+  }
+  member_method {
+    name: "isfinite"
+    argspec: "args=[\'x\'], varargs=None, keywords=None, defaults=None"
+  }
+  member_method {
+    name: "isinf"
+    argspec: "args=[\'x\'], varargs=None, keywords=None, defaults=None"
+  }
+  member_method {
+    name: "isnan"
+    argspec: "args=[\'x\'], varargs=None, keywords=None, defaults=None"
+  }
+  member_method {
+    name: "isneginf"
+    argspec: "args=[\'x\'], varargs=None, keywords=None, defaults=None"
+  }
+  member_method {
+    name: "isposinf"
+    argspec: "args=[\'x\'], varargs=None, keywords=None, defaults=None"
+  }
+  member_method {
+    name: "isreal"
+    argspec: "args=[\'x\'], varargs=None, keywords=None, defaults=None"
+  }
+  member_method {
+    name: "isrealobj"
+    argspec: "args=[\'x\'], varargs=None, keywords=None, defaults=None"
+  }
+  member_method {
+    name: "isscalar"
+    argspec: "args=[\'a\'], varargs=None, keywords=None, defaults=None"
+  }
+  member_method {
+    name: "issubdtype"
+    argspec: "args=[\'arg1\', \'arg2\'], varargs=None, keywords=None, defaults=None"
+  }
+  member_method {
+    name: "ix_"
+    argspec: "args=[], varargs=args, keywords=None, defaults=None"
+  }
+  member_method {
+    name: "kron"
+    argspec: "args=[\'a\', \'b\'], varargs=None, keywords=None, defaults=None"
+  }
+  member_method {
+    name: "lcm"
+    argspec: "args=[\'x1\', \'x2\'], varargs=None, keywords=None, defaults=None"
+  }
+  member_method {
+    name: "less"
+    argspec: "args=[\'x1\', \'x2\'], varargs=None, keywords=None, defaults=None"
+  }
+  member_method {
+    name: "less_equal"
+    argspec: "args=[\'x1\', \'x2\'], varargs=None, keywords=None, defaults=None"
+  }
+  member_method {
+    name: "linspace"
+    argspec: "args=[\'start\', \'stop\', \'num\', \'endpoint\', \'retstep\', \'dtype\', \'axis\'], varargs=None, keywords=None, defaults=[\'50\', \'True\', \'False\', \"\", \'0\'], "
+  }
+  member_method {
+    name: "log"
+    argspec: "args=[\'x\'], varargs=None, keywords=None, defaults=None"
+  }
+  member_method {
+    name: "log10"
+    argspec: "args=[\'x\'], varargs=None, keywords=None, defaults=None"
+  }
+  member_method {
+    name: "log1p"
+    argspec: "args=[\'x\'], varargs=None, keywords=None, defaults=None"
+  }
+  member_method {
+    name: "log2"
+    argspec: "args=[\'x\'], varargs=None, keywords=None, defaults=None"
+  }
+  member_method {
+    name: "logaddexp"
+    argspec: "args=[\'x1\', \'x2\'], varargs=None, keywords=None, defaults=None"
+  }
+  member_method {
+    name: "logaddexp2"
+    argspec: "args=[\'x1\', \'x2\'], varargs=None, keywords=None, defaults=None"
+  }
+  member_method {
+    name: "logical_and"
+    argspec: "args=[\'x1\', \'x2\'], varargs=None, keywords=None, defaults=None"
+  }
+  member_method {
+    name: "logical_not"
+    argspec: "args=[\'x\'], varargs=None, keywords=None, defaults=None"
+  }
+  member_method {
+    name: "logical_or"
+    argspec: "args=[\'x1\', \'x2\'], varargs=None, keywords=None, defaults=None"
+  }
+  member_method {
+    name: "logical_xor"
+    argspec: "args=[\'x1\', \'x2\'], varargs=None, keywords=None, defaults=None"
+  }
+  member_method {
+    name: "logspace"
+    argspec: "args=[\'start\', \'stop\', \'num\', \'endpoint\', \'base\', \'dtype\', \'axis\'], varargs=None, keywords=None, defaults=[\'50\', \'True\', \'10.0\', \'None\', \'0\'], "
+  }
+  member_method {
+    name: "matmul"
+    argspec: "args=[\'x1\', \'x2\'], varargs=None, keywords=None, defaults=None"
+  }
+  member_method {
+    name: "max"
+    argspec: "args=[\'a\', \'axis\', \'keepdims\'], varargs=None, keywords=None, defaults=[\'None\', \'None\'], "
+  }
+  member_method {
+    name: "maximum"
+    argspec: "args=[\'x1\', \'x2\'], varargs=None, keywords=None, defaults=None"
+  }
+  member_method {
+    name: "mean"
+    argspec: "args=[\'a\', \'axis\', \'dtype\', \'keepdims\'], varargs=None, keywords=None, defaults=[\'None\', \'None\', \'None\'], "
+  }
+  member_method {
+    name: "meshgrid"
+    argspec: "args=[], varargs=xi, keywords=kwargs, defaults=None"
+  }
+  member_method {
+    name: "min"
+    argspec: "args=[\'a\', \'axis\', \'keepdims\'], varargs=None, keywords=None, defaults=[\'None\', \'None\'], "
+  }
+  member_method {
+    name: "minimum"
+    argspec: "args=[\'x1\', \'x2\'], varargs=None, keywords=None, defaults=None"
+  }
+  member_method {
+    name: "mod"
+    argspec: "args=[\'x1\', \'x2\'], varargs=None, keywords=None, defaults=None"
+  }
+  member_method {
+    name: "moveaxis"
+    argspec: "args=[\'a\', \'source\', \'destination\'], varargs=None, keywords=None, defaults=None"
+  }
+  member_method {
+    name: "multiply"
+    argspec: "args=[\'x1\', \'x2\'], varargs=None, keywords=None, defaults=None"
+  }
+  member_method {
+    name: "nanmean"
+    argspec: "args=[\'a\', \'axis\', \'dtype\', \'keepdims\'], varargs=None, keywords=None, defaults=[\'None\', \'None\', \'None\'], "
+  }
+  member_method {
+    name: "nanprod"
+    argspec: "args=[\'a\', \'axis\', \'dtype\', \'keepdims\'], varargs=None, keywords=None, defaults=[\'None\', \'None\', \'False\'], "
+  }
+  member_method {
+    name: "nansum"
+    argspec: "args=[\'a\', \'axis\', \'dtype\', \'keepdims\'], varargs=None, keywords=None, defaults=[\'None\', \'None\', \'False\'], "
+  }
+  member_method {
+    name: "ndim"
+    argspec: "args=[\'a\'], varargs=None, keywords=None, defaults=None"
+  }
+  member_method {
+    name: "negative"
+    argspec: "args=[\'x\'], varargs=None, keywords=None, defaults=None"
+  }
+  member_method {
+    name: "nextafter"
+    argspec: "args=[\'x1\', \'x2\'], varargs=None, keywords=None, defaults=None"
+  }
+  member_method {
+    name: "nonzero"
+    argspec: "args=[\'a\'], varargs=None, keywords=None, defaults=None"
+  }
+  member_method {
+    name: "not_equal"
+    argspec: "args=[\'x1\', \'x2\'], varargs=None, keywords=None, defaults=None"
+  }
+  member_method {
+    name: "ones"
+    argspec: "args=[\'shape\', \'dtype\'], varargs=None, keywords=None, defaults=[\"\"], "
+  }
+  member_method {
+    name: "ones_like"
+    argspec: "args=[\'a\', \'dtype\'], varargs=None, keywords=None, defaults=[\'None\'], "
+  }
+  member_method {
+    name: "outer"
+    argspec: "args=[\'a\', \'b\'], varargs=None, keywords=None, defaults=None"
+  }
+  member_method {
+    name: "pad"
+    argspec: "args=[\'ary\', \'pad_width\', \'mode\', \'constant_values\'], varargs=None, keywords=None, defaults=[\'0\'], "
+  }
+  member_method {
+    name: "polyval"
+    argspec: "args=[\'p\', \'x\'], varargs=None, keywords=None, defaults=None"
+  }
+  member_method {
+    name: "positive"
+    argspec: "args=[\'x\'], varargs=None, keywords=None, defaults=None"
+  }
+  member_method {
+    name: "power"
+    argspec: "args=[\'x1\', \'x2\'], varargs=None, keywords=None, defaults=None"
+  }
+  member_method {
+    name: "prod"
+    argspec: "args=[\'a\', \'axis\', \'dtype\', \'keepdims\'], varargs=None, keywords=None, defaults=[\'None\', \'None\', \'None\'], "
+  }
+  member_method {
+    name: "promote_types"
+    argspec: "args=[\'type1\', \'type2\'], varargs=None, keywords=None, defaults=None"
+  }
+  member_method {
+    name: "ptp"
+    argspec: "args=[\'a\', \'axis\', \'keepdims\'], varargs=None, keywords=None, defaults=[\'None\', \'None\'], "
+  }
+  member_method {
+    name: "rad2deg"
+    argspec: "args=[\'x\'], varargs=None, keywords=None, defaults=None"
+  }
+  member_method {
+    name: "ravel"
+    argspec: "args=[\'a\'], varargs=None, keywords=None, defaults=None"
+  }
+  member_method {
+    name: "real"
+    argspec: "args=[\'val\'], varargs=None, keywords=None, defaults=None"
+  }
+  member_method {
+    name: "reciprocal"
+    argspec: "args=[\'x\'], varargs=None, keywords=None, defaults=None"
+  }
+  member_method {
+    name: "remainder"
+    argspec: "args=[\'x1\', \'x2\'], varargs=None, keywords=None, defaults=None"
+  }
+  member_method {
+    name: "repeat"
+    argspec: "args=[\'a\', \'repeats\', \'axis\'], varargs=None, keywords=None, defaults=[\'None\'], "
+  }
+  member_method {
+    name: "reshape"
+    argspec: "args=[\'a\', \'newshape\', \'order\'], varargs=None, keywords=None, defaults=[\'C\'], "
+  }
+  member_method {
+    name: "result_type"
+    argspec: "args=[], varargs=arrays_and_dtypes, keywords=None, defaults=None"
+  }
+  member_method {
+    name: "roll"
+    argspec: "args=[\'a\', \'shift\', \'axis\'], varargs=None, keywords=None, defaults=[\'None\'], "
+  }
+  member_method {
+    name: "rot90"
+    argspec: "args=[\'m\', \'k\', \'axes\'], varargs=None, keywords=None, defaults=[\'1\', \'(0, 1)\'], "
+  }
+  member_method {
+    name: "round"
+    argspec: "args=[\'a\', \'decimals\'], varargs=None, keywords=None, defaults=[\'0\'], "
+  }
+  member_method {
+    name: "select"
+    argspec: "args=[\'condlist\', \'choicelist\', \'default\'], varargs=None, keywords=None, defaults=[\'0\'], "
+  }
+  member_method {
+    name: "shape"
+    argspec: "args=[\'a\'], varargs=None, keywords=None, defaults=None"
+  }
+  member_method {
+    name: "sign"
+    argspec: "args=[\'x\', \'out\', \'where\'], varargs=None, keywords=kwargs, defaults=[\'None\', \'None\'], "
+  }
+  member_method {
+    name: "signbit"
+    argspec: "args=[\'x\'], varargs=None, keywords=None, defaults=None"
+  }
+  member_method {
+    name: "sin"
+    argspec: "args=[\'x\'], varargs=None, keywords=None, defaults=None"
+  }
+  member_method {
+    name: "sinc"
+    argspec: "args=[\'x\'], varargs=None, keywords=None, defaults=None"
+  }
+  member_method {
+    name: "sinh"
+    argspec: "args=[\'x\'], varargs=None, keywords=None, defaults=None"
+  }
+  member_method {
+    name: "sort"
+    argspec: "args=[\'a\', \'axis\', \'kind\', \'order\'], varargs=None, keywords=None, defaults=[\'-1\', \'quicksort\', \'None\'], "
+  }
+  member_method {
+    name: "split"
+    argspec: "args=[\'ary\', \'indices_or_sections\', \'axis\'], varargs=None, keywords=None, defaults=[\'0\'], "
+  }
+  member_method {
+    name: "sqrt"
+    argspec: "args=[\'x\'], varargs=None, keywords=None, defaults=None"
+  }
+  member_method {
+    name: "square"
+    argspec: "args=[\'x\'], varargs=None, keywords=None, defaults=None"
+  }
+  member_method {
+    name: "squeeze"
+    argspec: "args=[\'a\', \'axis\'], varargs=None, keywords=None, defaults=[\'None\'], "
+  }
+  member_method {
+    name: "stack"
+    argspec: "args=[\'arrays\', \'axis\'], varargs=None, keywords=None, defaults=[\'0\'], "
+  }
+  member_method {
+    name: "std"
+    argspec: "args=[\'a\', \'axis\', \'keepdims\'], varargs=None, keywords=None, defaults=[\'None\', \'None\'], "
+  }
+  member_method {
+    name: "subtract"
+    argspec: "args=[\'x1\', \'x2\'], varargs=None, keywords=None, defaults=None"
+  }
+  member_method {
+    name: "sum"
+    argspec: "args=[\'a\', \'axis\', \'dtype\', \'keepdims\'], varargs=None, keywords=None, defaults=[\'None\', \'None\', \'None\'], "
+  }
+  member_method {
+    name: "swapaxes"
+    argspec: "args=[\'a\', \'axis1\', \'axis2\'], varargs=None, keywords=None, defaults=None"
+  }
+  member_method {
+    name: "take"
+    argspec: "args=[\'a\', \'indices\', \'axis\', \'out\', \'mode\'], varargs=None, keywords=None, defaults=[\'None\', \'None\', \'clip\'], "
+  }
+  member_method {
+    name: "take_along_axis"
+    argspec: "args=[\'arr\', \'indices\', \'axis\'], varargs=None, keywords=None, defaults=None"
+  }
+  member_method {
+    name: "tan"
+    argspec: "args=[\'x\'], varargs=None, keywords=None, defaults=None"
+  }
+  member_method {
+    name: "tanh"
+    argspec: "args=[\'x\'], varargs=None, keywords=None, defaults=None"
+  }
+  member_method {
+    name: "tensordot"
+    argspec: "args=[\'a\', \'b\', \'axes\'], varargs=None, keywords=None, defaults=[\'2\'], "
+  }
+  member_method {
+    name: "tile"
+    argspec: "args=[\'a\', \'reps\'], varargs=None, keywords=None, defaults=None"
+  }
+  member_method {
+    name: "trace"
+    argspec: "args=[\'a\', \'offset\', \'axis1\', \'axis2\', \'dtype\'], varargs=None, keywords=None, defaults=[\'0\', \'0\', \'1\', \'None\'], "
+  }
+  member_method {
+    name: "transpose"
+    argspec: "args=[\'a\', \'axes\'], varargs=None, keywords=None, defaults=[\'None\'], "
+  }
+  member_method {
+    name: "tri"
+    argspec: "args=[\'N\', \'M\', \'k\', \'dtype\'], varargs=None, keywords=None, defaults=[\'None\', \'0\', \'None\'], "
+  }
+  member_method {
+    name: "tril"
+    argspec: "args=[\'m\', \'k\'], varargs=None, keywords=None, defaults=[\'0\'], "
+  }
+  member_method {
+    name: "triu"
+    argspec: "args=[\'m\', \'k\'], varargs=None, keywords=None, defaults=[\'0\'], "
+  }
+  member_method {
+    name: "true_divide"
+    argspec: "args=[\'x1\', \'x2\'], varargs=None, keywords=None, defaults=None"
+  }
+  member_method {
+    name: "vander"
+    argspec: "args=[\'x\', \'N\', \'increasing\'], varargs=None, keywords=None, defaults=[\'None\', \'False\'], "
+  }
+  member_method {
+    name: "var"
+    argspec: "args=[\'a\', \'axis\', \'dtype\', \'out\', \'ddof\', \'keepdims\'], varargs=None, keywords=None, defaults=[\'None\', \'None\', \'None\', \'0\', \'None\'], "
+  }
+  member_method {
+    name: "vdot"
+    argspec: "args=[\'a\', \'b\'], varargs=None, keywords=None, defaults=None"
+  }
+  member_method {
+    name: "vsplit"
+    argspec: "args=[\'ary\', \'indices_or_sections\'], varargs=None, keywords=None, defaults=None"
+  }
+  member_method {
+    name: "vstack"
+    argspec: "args=[\'tup\'], varargs=None, keywords=None, defaults=None"
+  }
+  member_method {
+    name: "where"
+    argspec: "args=[\'condition\', \'x\', \'y\'], varargs=None, keywords=None, defaults=[\'None\', \'None\'], "
+  }
+  member_method {
+    name: "zeros"
+    argspec: "args=[\'shape\', \'dtype\'], varargs=None, keywords=None, defaults=[\"\"], "
+  }
+  member_method {
+    name: "zeros_like"
+    argspec: "args=[\'a\', \'dtype\'], varargs=None, keywords=None, defaults=[\'None\'], "
+  }
+}
diff --git a/third_party/py/numpy/tf_numpy_api/numpy_ops.random.pbtxt b/third_party/py/numpy/tf_numpy_api/numpy_ops.random.pbtxt
new file mode 100644
index 00000000000..3e6bb720e8c
--- /dev/null
+++ b/third_party/py/numpy/tf_numpy_api/numpy_ops.random.pbtxt
@@ -0,0 +1,27 @@
+path: "numpy_ops.random"
+tf_module {
+  member_method {
+    name: "rand"
+    argspec: "args=[], varargs=size, keywords=None, defaults=None"
+  }
+  member_method {
+    name: "randint"
+    argspec: "args=[\'low\', \'high\', \'size\', \'dtype\'], varargs=None, keywords=None, defaults=[\'None\', \'None\', \"\"], "
+  }
+  member_method {
+    name: "randn"
+    argspec: "args=[], varargs=args, keywords=None, defaults=None"
+  }
+  member_method {
+    name: "random"
+    argspec: "args=[\'size\'], varargs=None, keywords=None, defaults=[\'None\'], "
+  }
+  member_method {
+    name: "seed"
+    argspec: "args=[\'s\'], varargs=None, keywords=None, defaults=None"
+  }
+  member_method {
+    name: "uniform"
+    argspec: "args=[\'low\', \'high\', \'size\'], varargs=None, keywords=None, defaults=[\'0.0\', \'1.0\', \'None\'], "
+  }
+}

From 23fc134a1762ea29f1fb300d3d60e84bd8939a29 Mon Sep 17 00:00:00 2001
From: "A. Unique TensorFlower" 
Date: Mon, 22 Jun 2020 16:12:37 -0700
Subject: [PATCH 496/867] Go: Update generated wrapper functions for TensorFlow
 ops.

PiperOrigin-RevId: 317755766
Change-Id: I105b220b0760792d9f4fa7c1390c1477cdb1c627
---
 tensorflow/go/op/wrappers.go | 5 +++++
 1 file changed, 5 insertions(+)

diff --git a/tensorflow/go/op/wrappers.go b/tensorflow/go/op/wrappers.go
index 8bceeb78564..3675c26751c 100644
--- a/tensorflow/go/op/wrappers.go
+++ b/tensorflow/go/op/wrappers.go
@@ -11448,6 +11448,11 @@ func AssertNextDataset(scope *Scope, input_dataset tf.Output, transformations tf
 }
 
 // Return the index of device the op runs.
+//
+// Given a list of device names, this operation returns the index of the device
+// this op runs. The length of the list is returned in two cases:
+// (1) Device does not exist in the given device list.
+// (2) It is in XLA compilation.
 func DeviceIndex(scope *Scope, device_names []string) (index tf.Output) {
 	if scope.Err() != nil {
 		return

From a26044ac2f32da5b8e91e3ee4b1a3f49de08fee7 Mon Sep 17 00:00:00 2001
From: "A. Unique TensorFlower" 
Date: Mon, 22 Jun 2020 16:16:21 -0700
Subject: [PATCH 497/867] Make a couple of remaining tests in math_ops_test.py
 run in eager mode.

PiperOrigin-RevId: 317756434
Change-Id: I14911d5f460131f4b9bdf6031cc7011ddd6a1d53
---
 tensorflow/python/ops/math_ops_test.py | 40 +++++++++++++++++++-------
 1 file changed, 30 insertions(+), 10 deletions(-)

diff --git a/tensorflow/python/ops/math_ops_test.py b/tensorflow/python/ops/math_ops_test.py
index 940966741dc..9699f6d2b78 100644
--- a/tensorflow/python/ops/math_ops_test.py
+++ b/tensorflow/python/ops/math_ops_test.py
@@ -397,7 +397,6 @@ class AddNTest(test_util.TensorFlowTestCase):
         self.assertAllEqual(x[0] * num_inputs,
                             math_ops.add_n([tf_x[0]] * num_inputs))
 
-  @test_util.deprecated_graph_mode_only
   def testGrad(self):
     np.random.seed(42)
     for num_inputs in range(1, 10):
@@ -406,9 +405,16 @@ class AddNTest(test_util.TensorFlowTestCase):
             variables.Variable(10.0 * np.random.random())
             for _ in range(0, num_inputs)
         ]
-        addn = math_ops.add_n(input_vars)
         self.evaluate(variables.global_variables_initializer())
-        add_n_grad = gradients.gradients(addn, input_vars)
+        if context.executing_eagerly():
+          with backprop.GradientTape() as tape:
+            tape.watch(input_vars)
+            addn = math_ops.add_n(input_vars)
+            add_n_grad = tape.gradient(addn, input_vars)
+        else:
+          addn = math_ops.add_n(input_vars)
+          add_n_grad = gradients.gradients(addn, input_vars)
+
         self.assertAllEqual(
             np.repeat(1.0, num_inputs),  # d/dx (x + y + ...) = 1
             [self.evaluate(g) for g in add_n_grad])
@@ -515,18 +521,32 @@ class DivAndModTest(test_util.TensorFlowTestCase):
     _ = math_ops.divide(foo, 1.)
     _ = math_ops.div(foo, 2.)
 
-  @test_util.deprecated_graph_mode_only
   def testFloorDivGrad(self):
     a = variables.Variable(2.)
     b = variables.Variable(4.)
+    input_vars = [a, b]
     self.evaluate(variables.global_variables_initializer())
-    c_grad = gradients.gradients(math_ops.divide(a, b), [a, b])
-    self.assertAllEqual([self.evaluate(x) for x in c_grad], [.25, -.125])
-    c_grad = gradients.gradients(math_ops.div(a, b), [a, b])
-    self.assertAllEqual([self.evaluate(x) for x in c_grad], [.25, -.125])
-    c_grad = gradients.gradients(math_ops.floordiv(a, b), [a, b])
+    if context.executing_eagerly():
+      # TDOO(rmlarsen): Is there a more compact way of
+      # writing this for multiple expressions?
+      with backprop.GradientTape() as tape:
+        tape.watch(input_vars)
+        c_grad0 = tape.gradient(math_ops.divide(a, b), input_vars)
+      with backprop.GradientTape() as tape:
+        tape.watch(input_vars)
+        c_grad1 = tape.gradient(math_ops.div(a, b), input_vars)
+      with backprop.GradientTape() as tape:
+        tape.watch(input_vars)
+        c_grad2 = tape.gradient(math_ops.floordiv(a, b), input_vars)
+    else:
+      c_grad0 = gradients.gradients(math_ops.divide(a, b), input_vars)
+      c_grad1 = gradients.gradients(math_ops.div(a, b), input_vars)
+      c_grad2 = gradients.gradients(math_ops.floordiv(a, b), input_vars)
+    self.assertAllEqual([self.evaluate(x) for x in c_grad0], [.25, -.125])
+    self.assertAllEqual([self.evaluate(x) for x in c_grad1], [.25, -.125])
     self.assertAllEqual(
-        [None if x is None else self.evaluate(x) for x in c_grad], [None, None])
+        [None if x is None else self.evaluate(x) for x in c_grad2],
+        [None, None])
 
   def testConsistent(self):
     nums, divs = self.intTestData()

From 557916fa3fd9d4c7dbcfabeb982632e2f31c4f1d Mon Sep 17 00:00:00 2001
From: Feng Liu 
Date: Mon, 22 Jun 2020 16:17:13 -0700
Subject: [PATCH 498/867] Fuse the back to back tfr.cast with unranked input to
 tf.EnsureShape

After this step, all the tfr ops should be raised to the tf ops

PiperOrigin-RevId: 317756602
Change-Id: I15731f231caf6c47eac1719fc1c2ed3d01cff515
---
 .../mlir/tensorflow/ir/tf_generated_ops.td    | 21 +++++++++++++++++++
 1 file changed, 21 insertions(+)

diff --git a/tensorflow/compiler/mlir/tensorflow/ir/tf_generated_ops.td b/tensorflow/compiler/mlir/tensorflow/ir/tf_generated_ops.td
index ecbe63ac525..7f250392cb2 100644
--- a/tensorflow/compiler/mlir/tensorflow/ir/tf_generated_ops.td
+++ b/tensorflow/compiler/mlir/tensorflow/ir/tf_generated_ops.td
@@ -2806,6 +2806,27 @@ the corresponding feature.
   TF_DerivedOperandSizeAttr N = TF_DerivedOperandSizeAttr<0>;
 }
 
+def TF_EnsureShapeOp : TF_Op<"EnsureShape", [NoSideEffect]> {
+  let summary = "Ensures that the tensor's shape matches the expected shape.";
+
+  let description = [{
+Raises an error if the input tensor's shape does not match the specified shape.
+Returns the input tensor otherwise.
+  }];
+
+  let arguments = (ins
+    TF_Tensor:$input,
+
+    TF_ShapeAttr:$shape
+  );
+
+  let results = (outs
+    TF_Tensor:$output
+  );
+
+  TF_DerivedOperandTypeAttr T = TF_DerivedOperandTypeAttr<0>;
+}
+
 def TF_EqualOp : TF_Op<"Equal", [Commutative, NoSideEffect]> {
   let summary = "Returns the truth value of (x == y) element-wise.";
 

From d2b35a7955ae20829ad395999089f889cbbcc598 Mon Sep 17 00:00:00 2001
From: Xinyi Wang 
Date: Mon, 22 Jun 2020 16:39:57 -0700
Subject: [PATCH 499/867] Enable last partial batch for MWMS in TF2.x

PiperOrigin-RevId: 317760674
Change-Id: Ib7e0adbf4f8f013f21faef07ed4961c078806093
---
 .../collective_all_reduce_strategy.py         |  1 +
 .../collective_all_reduce_strategy_test.py    | 55 +++++++++--------
 tensorflow/python/distribute/input_lib.py     | 13 ++--
 .../python/distribute/input_lib_test.py       |  1 -
 .../python/distribute/strategy_common_test.py | 61 +++++++++++++++++--
 5 files changed, 91 insertions(+), 40 deletions(-)

diff --git a/tensorflow/python/distribute/collective_all_reduce_strategy.py b/tensorflow/python/distribute/collective_all_reduce_strategy.py
index f73474c4af4..e2b039ceb23 100644
--- a/tensorflow/python/distribute/collective_all_reduce_strategy.py
+++ b/tensorflow/python/distribute/collective_all_reduce_strategy.py
@@ -178,6 +178,7 @@ class CollectiveAllReduceExtended(mirrored_strategy.MirroredExtended):
     self._communication = communication
     self._initialize_strategy(self._cluster_resolver)
     self._cfer_fn_cache = weakref.WeakKeyDictionary()
+    self.experimental_enable_get_next_as_optional = True
     assert isinstance(self._cross_device_ops,
                       cross_device_ops_lib.CollectiveAllReduce)
 
diff --git a/tensorflow/python/distribute/collective_all_reduce_strategy_test.py b/tensorflow/python/distribute/collective_all_reduce_strategy_test.py
index a9f7bc74e9e..87212c85fc4 100644
--- a/tensorflow/python/distribute/collective_all_reduce_strategy_test.py
+++ b/tensorflow/python/distribute/collective_all_reduce_strategy_test.py
@@ -370,7 +370,8 @@ class CollectiveAllReduceStrategyTestBase(
         else:
           self.assertEqual(list(expected_value), list(computed_value))
 
-      with self.assertRaises(errors.OutOfRangeError):
+      # error raised by calling optional_get_value on an Optional of None
+      with self.assertRaises(errors.InvalidArgumentError):
         next_element = iterator.get_next()
         sess.run([distribute_utils.select_replica(r, next_element)
                   for r in range(len(devices))])
@@ -449,31 +450,35 @@ class DistributedCollectiveAllReduceStrategyTest(
       combinations.combine(
           mode=['graph'], required_gpus=[0, 1, 2], use_dataset=[True, False]))
   def testMakeInputFnIterator(self, required_gpus, use_dataset):
-    if use_dataset:
-      fn = lambda: dataset_ops.Dataset.range(100)
-    else:
-      def fn():
-        dataset = dataset_ops.Dataset.range(100)
-        it = dataset_ops.make_one_shot_iterator(dataset)
-        return it.get_next
-    # We use CPU as the device when required_gpus = 0
-    devices_per_worker = max(1, required_gpus)
-    expected_values = [[i+j for j in range(devices_per_worker)]
-                       for i in range(0, 100, devices_per_worker)]
+    def _worker_fn(task_type, task_id, required_gpus):
+      if use_dataset:
+        fn = lambda: dataset_ops.Dataset.range(100)
+      else:
+        def fn():
+          dataset = dataset_ops.Dataset.range(100)
+          it = dataset_ops.make_one_shot_iterator(dataset)
+          return it.get_next
+      # We use CPU as the device when required_gpus = 0
+      devices_per_worker = max(1, required_gpus)
+      expected_values = [[i+j for j in range(devices_per_worker)]
+                         for i in range(0, 100, devices_per_worker)]
 
-    input_fn = self._input_fn_to_test_input_context(
-        fn,
-        expected_num_replicas_in_sync=3*devices_per_worker,
-        expected_num_input_pipelines=3,
-        expected_input_pipeline_id=1)  # because task_id = 1
-    self._test_input_fn_iterator(
-        'worker',
-        1,
-        required_gpus,
-        input_fn,
-        expected_values,
-        test_reinitialize=use_dataset,
-        ignore_order=not use_dataset)
+      input_fn = self._input_fn_to_test_input_context(
+          fn,
+          expected_num_replicas_in_sync=3*devices_per_worker,
+          expected_num_input_pipelines=3,
+          expected_input_pipeline_id=task_id)
+      self._test_input_fn_iterator(
+          task_type,
+          task_id,
+          required_gpus,
+          input_fn,
+          expected_values,
+          test_reinitialize=use_dataset,
+          ignore_order=not use_dataset)
+
+    self._run_between_graph_clients(_worker_fn, self._cluster_spec,
+                                    required_gpus)
 
   @combinations.generate(combinations.combine(mode=['graph']))
   def testUpdateConfigProto(self):
diff --git a/tensorflow/python/distribute/input_lib.py b/tensorflow/python/distribute/input_lib.py
index e4a362a92c6..74268999de0 100644
--- a/tensorflow/python/distribute/input_lib.py
+++ b/tensorflow/python/distribute/input_lib.py
@@ -549,7 +549,7 @@ def _get_next_as_optional(iterator, strategy, name=None):
       # Collective all-reduce requires explicit devices for inputs.
       with ops.device("/cpu:0"):
         # Converting to integers for all-reduce.
-        worker_has_value = math_ops.cast(worker_has_value, dtypes.int32)
+        worker_has_value = math_ops.cast(worker_has_value, dtypes.int64)
         worker_devices.append(worker_has_value.device)
         worker_has_values.append(worker_has_value)
       # Make `replicas` a flat list of values across all replicas.
@@ -624,16 +624,12 @@ class DistributedIteratorBase(DistributedIteratorInterface):
     # get_next_as_optional(). And we only enable get_next_as_optional when the
     # output shapes are not static.
     #
-    # TODO(yuefengz): Currently `experimental_enable_get_next_as_optional` is
-    # always set to False in CollectiveAllReduceStrategy. We want to have a way
-    # to distinguish multi workers/single worker between graph, so we can enable
-    # the behavior in single worker case.
-    #
     # TODO(rxsang): We want to always enable the get_next_as_optional behavior
     # when user passed input_fn instead of dataset.
     if getattr(
         strategy.extended, "experimental_enable_get_next_as_optional", False):
-      self._enable_get_next_as_optional = not static_shape
+      self._enable_get_next_as_optional = (
+          not static_shape) or strategy.extended._in_multi_worker_mode()
     else:
       self._enable_get_next_as_optional = False
 
@@ -906,7 +902,8 @@ class DistributedIterator(DistributedIteratorBase,
       self._strategy = strategy
       if getattr(
           strategy.extended, "experimental_enable_get_next_as_optional", False):
-        self._enable_get_next_as_optional = not static_shape
+        self._enable_get_next_as_optional = (
+            not static_shape) or strategy.extended._in_multi_worker_mode()
       else:
         self._enable_get_next_as_optional = False
     else:
diff --git a/tensorflow/python/distribute/input_lib_test.py b/tensorflow/python/distribute/input_lib_test.py
index 7f02d0121d0..23397bf5070 100644
--- a/tensorflow/python/distribute/input_lib_test.py
+++ b/tensorflow/python/distribute/input_lib_test.py
@@ -1144,7 +1144,6 @@ class DistributedIteratorMultiWorkerTest(
           expected_values = [[[0, 1]], [[2, 3]], [[4]]]
           input_context = None
 
-        strategy.extended.experimental_enable_get_next_as_optional = True
         self._test_input_iteration(
             input_type,
             api_type,
diff --git a/tensorflow/python/distribute/strategy_common_test.py b/tensorflow/python/distribute/strategy_common_test.py
index ed52a4794ee..b823e2de331 100644
--- a/tensorflow/python/distribute/strategy_common_test.py
+++ b/tensorflow/python/distribute/strategy_common_test.py
@@ -19,6 +19,7 @@ from __future__ import division
 from __future__ import print_function
 
 from absl.testing import parameterized
+import numpy as np
 
 from tensorflow.python.data.ops import dataset_ops
 from tensorflow.python.distribute import combinations
@@ -29,7 +30,6 @@ from tensorflow.python.distribute import strategy_test_lib
 from tensorflow.python.eager import def_function
 from tensorflow.python.framework import dtypes
 from tensorflow.python.ops import array_ops
-from tensorflow.python.ops import math_ops
 from tensorflow.python.platform import test
 
 
@@ -81,7 +81,7 @@ class DistributedCollectiveAllReduceStrategyTest(
       return d.shard(input_context.num_input_pipelines,
                      input_context.input_pipeline_id)
 
-    expected_sum_on_workers = [10, 35]
+    expected_data_on_worker = [[0, 1, 2, 3, 4], [5, 6, 7, 8, 9]]
     input_iterator = iter(
         strategy.experimental_distribute_datasets_from_function(dataset_fn))
 
@@ -90,10 +90,59 @@ class DistributedCollectiveAllReduceStrategyTest(
       return strategy.experimental_local_results(iterator.get_next())
 
     result = run(input_iterator)
-    sum_value = math_ops.reduce_sum(result)
-    self.assertEqual(
-        sum_value.numpy(),
-        expected_sum_on_workers[multi_worker_test_base.get_task_index()])
+    self.assertTrue(
+        np.array_equal(
+            result[0].numpy(),
+            expected_data_on_worker[multi_worker_test_base.get_task_index()]))
+
+  def testSimpleInputFromDatasetLastPartialBatch(self, strategy):
+    global_batch_size = 8
+    dataset = dataset_ops.DatasetV2.range(14).batch(
+        global_batch_size, drop_remainder=False)
+    input_iterator = iter(strategy.experimental_distribute_dataset(dataset))
+
+    @def_function.function
+    def run(input_iterator):
+      return strategy.run(lambda x: x, args=(next(input_iterator),))
+
+    # Let the complete batch go.
+    run(input_iterator)
+
+    # `result` is an incomplete batch
+    result = run(input_iterator)
+    expected_data_on_worker = [[8, 9, 10], [11, 12, 13]]
+    self.assertTrue(
+        np.array_equal(
+            result.numpy(),
+            expected_data_on_worker[multi_worker_test_base.get_task_index()]))
+
+  def testSimpleInputFromFnLastPartialBatch(self, strategy):
+
+    def dataset_fn(input_context):
+      global_batch_size = 8
+      batch_size = input_context.get_per_replica_batch_size(global_batch_size)
+      dataset = dataset_ops.DatasetV2.range(14).batch(
+          batch_size, drop_remainder=False)
+      return dataset.shard(input_context.num_input_pipelines,
+                           input_context.input_pipeline_id)
+
+    input_iterator = iter(
+        strategy.experimental_distribute_datasets_from_function(dataset_fn))
+
+    @def_function.function
+    def run(input_iterator):
+      return strategy.run(lambda x: x, args=(next(input_iterator),))
+
+    # Let the complete batch go.
+    run(input_iterator)
+    # `result` is an incomplete batch
+    result = run(input_iterator)
+
+    expected_data_on_worker = [[8, 9, 10, 11], [12, 13]]
+    self.assertTrue(
+        np.array_equal(
+            result.numpy(), expected_data_on_worker[
+                multi_worker_test_base.get_task_index()]))
 
   def testReduceHostTensor(self, strategy):
     reduced = strategy.reduce(

From 2f9984ff838fb7636b78deff197c18c144d5532c Mon Sep 17 00:00:00 2001
From: Berkin Ilbeyi 
Date: Mon, 22 Jun 2020 16:42:36 -0700
Subject: [PATCH 500/867] Allocate input/output buffers at the correct memory
 space.

PiperOrigin-RevId: 317761146
Change-Id: Ia27b6765a547451575240e76ea419f705e29bd32
---
 tensorflow/compiler/xla/service/interpreter/executor.cc | 1 -
 tensorflow/compiler/xla/service/transfer_manager.cc     | 4 +++-
 2 files changed, 3 insertions(+), 2 deletions(-)

diff --git a/tensorflow/compiler/xla/service/interpreter/executor.cc b/tensorflow/compiler/xla/service/interpreter/executor.cc
index a2e46ba2afe..616fd031c47 100644
--- a/tensorflow/compiler/xla/service/interpreter/executor.cc
+++ b/tensorflow/compiler/xla/service/interpreter/executor.cc
@@ -35,7 +35,6 @@ XlaInterpreterExecutor::~XlaInterpreterExecutor() {}
 
 DeviceMemoryBase XlaInterpreterExecutor::Allocate(uint64 size,
                                                   int64 memory_space) {
-  CHECK_EQ(memory_space, 0);
   return DeviceMemoryBase(new char[size], size);
 }
 
diff --git a/tensorflow/compiler/xla/service/transfer_manager.cc b/tensorflow/compiler/xla/service/transfer_manager.cc
index 4658aebd571..ebb0226476f 100644
--- a/tensorflow/compiler/xla/service/transfer_manager.cc
+++ b/tensorflow/compiler/xla/service/transfer_manager.cc
@@ -355,7 +355,9 @@ StatusOr TransferManager::AllocateScopedShapedBuffer(
         ShapeUtil::GetSubshape(shaped_buffer.on_device_shape(), index);
     TF_ASSIGN_OR_RETURN(auto memory,
                         allocator->Allocate(shaped_buffer.device_ordinal(),
-                                            GetByteSizeRequirement(subshape)));
+                                            GetByteSizeRequirement(subshape),
+                                            /*retry_on_failure=*/true,
+                                            subshape.layout().memory_space()));
     // Move the allocated buffer into the ScopedShapedBuffer, which owns it.
     memory_base = memory.Release();
   }

From 2d8d440dbb0f9aaa4bef53c815a33c6cb36c3bdf Mon Sep 17 00:00:00 2001
From: Meghna Natraj 
Date: Mon, 22 Jun 2020 16:49:56 -0700
Subject: [PATCH 501/867] Update TFLite interpreter usage in TFLite Micro
 Speech example

PiperOrigin-RevId: 317762428
Change-Id: I075791cad44c6d9d4e747f76b0cb50a5d422fe87
---
 .../train/train_micro_speech_model.ipynb      | 600 +++++++++++++++++-
 1 file changed, 599 insertions(+), 1 deletion(-)

diff --git a/tensorflow/lite/micro/examples/micro_speech/train/train_micro_speech_model.ipynb b/tensorflow/lite/micro/examples/micro_speech/train/train_micro_speech_model.ipynb
index bfe75bdd9f7..1a1d5ff2c3b 100644
--- a/tensorflow/lite/micro/examples/micro_speech/train/train_micro_speech_model.ipynb
+++ b/tensorflow/lite/micro/examples/micro_speech/train/train_micro_speech_model.ipynb
@@ -1 +1,599 @@
-{"nbformat":4,"nbformat_minor":0,"metadata":{"colab":{"name":"train_micro_speech_model.ipynb","provenance":[{"file_id":"https://github.com/tensorflow/tensorflow/blob/master/tensorflow/lite/micro/examples/micro_speech/train/train_micro_speech_model.ipynb","timestamp":1587690382292}],"collapsed_sections":[],"toc_visible":true},"kernelspec":{"name":"python3","display_name":"Python 3"},"accelerator":"GPU"},"cells":[{"cell_type":"markdown","metadata":{"id":"pO4-CY_TCZZS","colab_type":"text"},"source":["# Train a Simple Audio Recognition Model"]},{"cell_type":"markdown","metadata":{"id":"BaFfr7DHRmGF","colab_type":"text"},"source":["This notebook demonstrates how to train a 20 kB [Simple Audio Recognition](https://www.tensorflow.org/tutorials/sequences/audio_recognition) model to recognize keywords in speech.\n","\n","The model created in this notebook is used in the [micro_speech](https://github.com/tensorflow/tensorflow/tree/master/tensorflow/lite/micro/examples/micro_speech) example for [TensorFlow Lite for MicroControllers](https://www.tensorflow.org/lite/microcontrollers/overview).\n","\n","
\n"," \n"," \n","
\n"," Run in Google Colab\n"," \n"," View source on GitHub\n","
\n"]},{"cell_type":"markdown","metadata":{"id":"XaVtYN4nlCft","colab_type":"text"},"source":["**Training is much faster using GPU acceleration.** Before you proceed, ensure you are using a GPU runtime by going to **Runtime -> Change runtime type** and set **Hardware accelerator: GPU**. Training 15,000 iterations will take 1.5 - 2 hours on a GPU runtime.\n","\n","## Configure Defaults\n","\n","**MODIFY** the following constants for your specific use case."]},{"cell_type":"code","metadata":{"id":"ludfxbNIaegy","colab_type":"code","colab":{}},"source":["# A comma-delimited list of the words you want to train for.\n","# The options are: yes,no,up,down,left,right,on,off,stop,go\n","# All the other words will be used to train an \"unknown\" label and silent\n","# audio data with no spoken words will be used to train a \"silence\" label.\n","WANTED_WORDS = \"yes,no\"\n","\n","# The number of steps and learning rates can be specified as comma-separated\n","# lists to define the rate at each stage. For example,\n","# TRAINING_STEPS=12000,3000 and LEARNING_RATE=0.001,0.0001\n","# will run 12,000 training loops in total, with a rate of 0.001 for the first\n","# 8,000, and 0.0001 for the final 3,000.\n","TRAINING_STEPS = \"12000,3000\"\n","LEARNING_RATE = \"0.001,0.0001\"\n","\n","# Calculate the total number of steps, which is used to identify the checkpoint\n","# file name.\n","TOTAL_STEPS = str(sum(map(lambda string: int(string), TRAINING_STEPS.split(\",\"))))\n","\n","# Print the configuration to confirm it\n","print(\"Training these words: %s\" % WANTED_WORDS)\n","print(\"Training steps in each stage: %s\" % TRAINING_STEPS)\n","print(\"Learning rate in each stage: %s\" % LEARNING_RATE)\n","print(\"Total number of training steps: %s\" % TOTAL_STEPS)"],"execution_count":0,"outputs":[]},{"cell_type":"markdown","metadata":{"id":"gCgeOpvY9pAi","colab_type":"text"},"source":["**DO NOT MODIFY** the following constants as they include filepaths used in this notebook and data that is shared during training and inference."]},{"cell_type":"code","metadata":{"id":"Nd1iM1o2ymvA","colab_type":"code","colab":{}},"source":["# Calculate the percentage of 'silence' and 'unknown' training samples required\n","# to ensure that we have equal number of samples for each label.\n","number_of_labels = WANTED_WORDS.count(',') + 1\n","number_of_total_labels = number_of_labels + 2 # for 'silence' and 'unknown' label\n","equal_percentage_of_training_samples = int(100.0/(number_of_total_labels))\n","SILENT_PERCENTAGE = equal_percentage_of_training_samples\n","UNKNOWN_PERCENTAGE = equal_percentage_of_training_samples\n","\n","# Constants which are shared during training and inference\n","PREPROCESS = 'micro'\n","WINDOW_STRIDE = 20\n","MODEL_ARCHITECTURE = 'tiny_conv' # Other options include: single_fc, conv,\n"," # low_latency_conv, low_latency_svdf, tiny_embedding_conv\n","\n","# Constants used during training only\n","VERBOSITY = 'WARN'\n","EVAL_STEP_INTERVAL = '1000'\n","SAVE_STEP_INTERVAL = '1000'\n","\n","# Constants for training directories and filepaths\n","DATASET_DIR = 'dataset/'\n","LOGS_DIR = 'logs/'\n","TRAIN_DIR = 'train/' # for training checkpoints and other files.\n","\n","# Constants for inference directories and filepaths\n","import os\n","MODELS_DIR = 'models'\n","if not os.path.exists(MODELS_DIR):\n"," os.mkdir(MODELS_DIR)\n","MODEL_TF = os.path.join(MODELS_DIR, 'model.pb')\n","MODEL_TFLITE = os.path.join(MODELS_DIR, 'model.tflite')\n","FLOAT_MODEL_TFLITE = os.path.join(MODELS_DIR, 'float_model.tflite')\n","MODEL_TFLITE_MICRO = os.path.join(MODELS_DIR, 'model.cc')\n","SAVED_MODEL = os.path.join(MODELS_DIR, 'saved_model')\n","\n","QUANT_INPUT_MIN = 0.0\n","QUANT_INPUT_MAX = 26.0\n","QUANT_INPUT_RANGE = QUANT_INPUT_MAX - QUANT_INPUT_MIN"],"execution_count":0,"outputs":[]},{"cell_type":"markdown","metadata":{"id":"6rLYpvtg9P4o","colab_type":"text"},"source":["## Setup Environment\n","\n","Install Dependencies"]},{"cell_type":"code","metadata":{"id":"ed_XpUrU5DvY","colab_type":"code","colab":{}},"source":["%tensorflow_version 1.x\n","import tensorflow as tf"],"execution_count":0,"outputs":[]},{"cell_type":"markdown","metadata":{"id":"T9Ty5mR58E4i","colab_type":"text"},"source":["**DELETE** any old data from previous runs\n"]},{"cell_type":"code","metadata":{"id":"APGx0fEh7hFF","colab_type":"code","colab":{}},"source":["!rm -rf {DATASET_DIR} {LOGS_DIR} {TRAIN_DIR} {MODELS_DIR}"],"execution_count":0,"outputs":[]},{"cell_type":"markdown","metadata":{"id":"GfEUlfFBizio","colab_type":"text"},"source":["Clone the TensorFlow Github Repository, which contains the relevant code required to run this tutorial."]},{"cell_type":"code","metadata":{"id":"yZArmzT85SLq","colab_type":"code","colab":{}},"source":["!git clone -q --depth 1 https://github.com/tensorflow/tensorflow"],"execution_count":0,"outputs":[]},{"cell_type":"markdown","metadata":{"id":"nS9swHLSi7Bi","colab_type":"text"},"source":["Load TensorBoard to visualize the accuracy and loss as training proceeds.\n"]},{"cell_type":"code","metadata":{"id":"q4qF1VxP3UE4","colab_type":"code","colab":{}},"source":["%load_ext tensorboard\n","%tensorboard --logdir {LOGS_DIR}"],"execution_count":0,"outputs":[]},{"cell_type":"markdown","metadata":{"id":"x1J96Ron-O4R","colab_type":"text"},"source":["## Training\n","\n","The following script downloads the dataset and begin training."]},{"cell_type":"code","metadata":{"id":"VJsEZx6lynbY","colab_type":"code","colab":{}},"source":["!python tensorflow/tensorflow/examples/speech_commands/train.py \\\n","--data_dir={DATASET_DIR} \\\n","--wanted_words={WANTED_WORDS} \\\n","--silence_percentage={SILENT_PERCENTAGE} \\\n","--unknown_percentage={UNKNOWN_PERCENTAGE} \\\n","--preprocess={PREPROCESS} \\\n","--window_stride={WINDOW_STRIDE} \\\n","--model_architecture={MODEL_ARCHITECTURE} \\\n","--how_many_training_steps={TRAINING_STEPS} \\\n","--learning_rate={LEARNING_RATE} \\\n","--train_dir={TRAIN_DIR} \\\n","--summaries_dir={LOGS_DIR} \\\n","--verbosity={VERBOSITY} \\\n","--eval_step_interval={EVAL_STEP_INTERVAL} \\\n","--save_step_interval={SAVE_STEP_INTERVAL}"],"execution_count":0,"outputs":[]},{"cell_type":"markdown","metadata":{"id":"UczQKtqLi7OJ","colab_type":"text"},"source":["# Skipping the training\n","\n","If you don't want to spend an hour or two training the model from scratch, you can download pretrained checkpoints by uncommenting the lines below (removing the '#'s at the start of each line) and running them."]},{"cell_type":"code","metadata":{"id":"RZw3VNlnla-J","colab_type":"code","colab":{}},"source":["#!curl -O \"https://storage.googleapis.com/download.tensorflow.org/models/tflite/speech_micro_train_2020_05_10.tgz\"\n","#!tar xzf speech_micro_train_2020_05_10.tgz"],"execution_count":0,"outputs":[]},{"cell_type":"markdown","metadata":{"id":"XQUJLrdS-ftl","colab_type":"text"},"source":["## Generate a TensorFlow Model for Inference\n","\n","Combine relevant training results (graph, weights, etc) into a single file for inference. This process is known as freezing a model and the resulting model is known as a frozen model/graph, as it cannot be further re-trained after this process."]},{"cell_type":"code","metadata":{"id":"xyc3_eLh9sAg","colab_type":"code","colab":{}},"source":["!rm -rf {SAVED_MODEL}\n","!python tensorflow/tensorflow/examples/speech_commands/freeze.py \\\n","--wanted_words=$WANTED_WORDS \\\n","--window_stride_ms=$WINDOW_STRIDE \\\n","--preprocess=$PREPROCESS \\\n","--model_architecture=$MODEL_ARCHITECTURE \\\n","--start_checkpoint=$TRAIN_DIR$MODEL_ARCHITECTURE'.ckpt-'{TOTAL_STEPS} \\\n","--save_format=saved_model \\\n","--output_file={SAVED_MODEL}"],"execution_count":0,"outputs":[]},{"cell_type":"markdown","metadata":{"id":"_DBGDxVI-nKG","colab_type":"text"},"source":["## Generate a TensorFlow Lite Model\n","\n","Convert the frozen graph into a TensorFlow Lite model, which is fully quantized for use with embedded devices.\n","\n","The following cell will also print the model size, which will be under 20 kilobytes."]},{"cell_type":"code","metadata":{"id":"RIitkqvGWmre","colab_type":"code","colab":{}},"source":["import sys\n","# We add this path so we can import the speech processing modules.\n","sys.path.append(\"/content/tensorflow/tensorflow/examples/speech_commands/\")\n","import input_data\n","import models\n","import numpy as np"],"execution_count":0,"outputs":[]},{"cell_type":"code","metadata":{"id":"kzqECqMxgBh4","colab_type":"code","colab":{}},"source":["SAMPLE_RATE = 16000\n","CLIP_DURATION_MS = 1000\n","WINDOW_SIZE_MS = 30.0\n","FEATURE_BIN_COUNT = 40\n","BACKGROUND_FREQUENCY = 0.8\n","BACKGROUND_VOLUME_RANGE = 0.1\n","TIME_SHIFT_MS = 100.0\n","\n","DATA_URL = 'https://storage.googleapis.com/download.tensorflow.org/data/speech_commands_v0.02.tar.gz'\n","VALIDATION_PERCENTAGE = 10\n","TESTING_PERCENTAGE = 10"],"execution_count":0,"outputs":[]},{"cell_type":"code","metadata":{"id":"rNQdAplJV1fz","colab_type":"code","colab":{}},"source":["model_settings = models.prepare_model_settings(\n"," len(input_data.prepare_words_list(WANTED_WORDS.split(','))),\n"," SAMPLE_RATE, CLIP_DURATION_MS, WINDOW_SIZE_MS,\n"," WINDOW_STRIDE, FEATURE_BIN_COUNT, PREPROCESS)\n","audio_processor = input_data.AudioProcessor(\n"," DATA_URL, DATASET_DIR,\n"," SILENT_PERCENTAGE, UNKNOWN_PERCENTAGE,\n"," WANTED_WORDS.split(','), VALIDATION_PERCENTAGE,\n"," TESTING_PERCENTAGE, model_settings, LOGS_DIR)"],"execution_count":0,"outputs":[]},{"cell_type":"code","metadata":{"id":"lBj_AyCh1cC0","colab_type":"code","colab":{}},"source":["with tf.Session() as sess:\n"," float_converter = tf.lite.TFLiteConverter.from_saved_model(SAVED_MODEL)\n"," float_tflite_model = float_converter.convert()\n"," float_tflite_model_size = open(FLOAT_MODEL_TFLITE, \"wb\").write(float_tflite_model)\n"," print(\"Float model is %d bytes\" % float_tflite_model_size)\n","\n"," converter = tf.lite.TFLiteConverter.from_saved_model(SAVED_MODEL)\n"," converter.optimizations = [tf.lite.Optimize.DEFAULT]\n"," converter.inference_input_type = tf.lite.constants.INT8\n"," converter.inference_output_type = tf.lite.constants.INT8\n"," def representative_dataset_gen():\n"," for i in range(100):\n"," data, _ = audio_processor.get_data(1, i*1, model_settings,\n"," BACKGROUND_FREQUENCY, \n"," BACKGROUND_VOLUME_RANGE,\n"," TIME_SHIFT_MS,\n"," 'testing',\n"," sess)\n"," flattened_data = np.array(data.flatten(), dtype=np.float32).reshape(1, 1960)\n"," yield [flattened_data]\n"," converter.representative_dataset = representative_dataset_gen\n"," tflite_model = converter.convert()\n"," tflite_model_size = open(MODEL_TFLITE, \"wb\").write(tflite_model)\n"," print(\"Quantized model is %d bytes\" % tflite_model_size)\n"],"execution_count":0,"outputs":[]},{"cell_type":"markdown","metadata":{"id":"EeLiDZTbLkzv","colab_type":"text"},"source":["# Testing the TensorFlow Lite model's accuracy\n","\n","Verify that the model we've exported is still accurate, using the TF Lite Python API and our test set."]},{"cell_type":"code","metadata":{"id":"wQsEteKRLryJ","colab_type":"code","colab":{}},"source":["with tf.Session() as sess:\n"," test_data, test_labels = audio_processor.get_data(\n"," -1, 0, model_settings, BACKGROUND_FREQUENCY, BACKGROUND_VOLUME_RANGE,\n"," TIME_SHIFT_MS, 'testing', sess)\n","\n","float_interpreter = tf.lite.Interpreter(FLOAT_MODEL_TFLITE)\n","float_interpreter.allocate_tensors()\n","\n","float_input_index = float_interpreter.get_input_details()[0][\"index\"]\n","\n","float_output_index = float_interpreter.get_output_details()[0][\"index\"]\n","float_model_output = float_interpreter.tensor(float_output_index)\n","\n","float_correct_predictions = 0\n","for i in range(len(test_data)):\n"," current_input = test_data[i]\n"," current_label = test_labels[i]\n"," flattened_input = np.array(current_input.flatten(), dtype=np.float32).reshape(1, 1960)\n"," float_interpreter.set_tensor(float_input_index, flattened_input)\n"," float_interpreter.invoke()\n"," top_prediction = float_model_output()[0].argmax()\n"," if top_prediction == current_label:\n"," float_correct_predictions += 1\n","\n","print('Float accuracy is %f%% (N=%d)' % ((float_correct_predictions * 100) / len(test_data), len(test_data)))\n","\n","interpreter = tf.lite.Interpreter(MODEL_TFLITE)\n","interpreter.allocate_tensors()\n","\n","input_index = interpreter.get_input_details()[0][\"index\"]\n","\n","output_index = interpreter.get_output_details()[0][\"index\"]\n","model_output = interpreter.tensor(output_index)\n","\n","with tf.Session() as sess:\n"," test_data, test_labels = audio_processor.get_data(\n"," -1, 0, model_settings, BACKGROUND_FREQUENCY, BACKGROUND_VOLUME_RANGE,\n"," TIME_SHIFT_MS, 'testing', sess)\n","\n","correct_predictions = 0\n","for i in range(len(test_data)):\n"," current_input = test_data[i]\n"," current_label = test_labels[i]\n"," quantized_input = np.zeros((1960), np.int8)\n"," for index, input_value in enumerate(current_input.flatten()):\n"," # These scaling values are derived from those used in input_data.py in the\n"," # training pipeline.\n"," value = ((input_value - QUANT_INPUT_MIN) * 256) / QUANT_INPUT_RANGE\n"," value -= 128\n"," if value < -128:\n"," value = -128\n"," if value > 127:\n"," value = 127\n"," quantized_input[index] = value\n"," flattened_input = np.array(quantized_input.flatten(), dtype=np.int8).reshape(1, 1960)\n"," interpreter.set_tensor(input_index, flattened_input)\n"," interpreter.invoke()\n"," top_prediction = model_output()[0].argmax()\n"," if top_prediction == current_label:\n"," correct_predictions += 1\n","\n","print('Quantized accuracy is %f%% (N=%d)' % ((correct_predictions * 100) / len(test_data), len(test_data)))\n"],"execution_count":0,"outputs":[]},{"cell_type":"markdown","metadata":{"id":"dt6Zqbxu-wIi","colab_type":"text"},"source":["## Generate a TensorFlow Lite for MicroControllers Model\n","Convert the TensorFlow Lite model into a C source file that can be loaded by TensorFlow Lite for Microcontrollers."]},{"cell_type":"code","metadata":{"id":"XohZOTjR8ZyE","colab_type":"code","colab":{}},"source":["# Install xxd if it is not available\n","!apt-get update && apt-get -qq install xxd\n","# Convert to a C source file\n","!xxd -i {MODEL_TFLITE} > {MODEL_TFLITE_MICRO}\n","# Update variable names\n","REPLACE_TEXT = MODEL_TFLITE.replace('/', '_').replace('.', '_')\n","!sed -i 's/'{REPLACE_TEXT}'/g_model/g' {MODEL_TFLITE_MICRO}"],"execution_count":0,"outputs":[]},{"cell_type":"markdown","metadata":{"id":"2pQnN0i_-0L2","colab_type":"text"},"source":["## Deploy to a Microcontroller\n","\n","Follow the instructions in the [micro_speech](https://github.com/tensorflow/tensorflow/tree/master/tensorflow/lite/micro/examples/micro_speech) README.md for [TensorFlow Lite for MicroControllers](https://www.tensorflow.org/lite/microcontrollers/overview) to deploy this model on a specific microcontroller.\n","\n","**Reference Model:** If you have not modified this notebook, you can follow the instructions as is, to deploy the model. Refer to the [`micro_speech/train/models`](https://github.com/tensorflow/tensorflow/blob/master/tensorflow/lite/micro/examples/micro_speech/train/models) directory to access the models generated in this notebook. \n","\n","**New Model:** If you have generated a new model to identify different words: (i) Update `kCategoryCount` and `kCategoryLabels` in [`micro_speech/micro_features/micro_model_settings.h`](https://github.com/tensorflow/tensorflow/blob/master/tensorflow/lite/micro/examples/micro_speech/micro_features/micro_model_settings.h) and (ii) Update the values assigned to the variables defined in [`micro_speech/micro_features/model.cc`](https://github.com/tensorflow/tensorflow/blob/master/tensorflow/lite/micro/examples/micro_speech/micro_features/model.cc) with values displayed after running the following cell."]},{"cell_type":"code","metadata":{"id":"eoYyh0VU8pca","colab_type":"code","colab":{}},"source":["# Print the C source file\n","!cat {MODEL_TFLITE_MICRO}"],"execution_count":0,"outputs":[]},{"cell_type":"code","metadata":{"id":"iYlIKpO2mkhv","colab_type":"code","colab":{}},"source":[""],"execution_count":0,"outputs":[]}]} \ No newline at end of file +{ + "nbformat": 4, + "nbformat_minor": 0, + "metadata": { + "colab": { + "name": "train_micro_speech_model.ipynb", + "provenance": [], + "collapsed_sections": [], + "toc_visible": true + }, + "kernelspec": { + "name": "python3", + "display_name": "Python 3" + }, + "accelerator": "GPU" + }, + "cells": [ + { + "cell_type": "markdown", + "metadata": { + "id": "pO4-CY_TCZZS", + "colab_type": "text" + }, + "source": [ + "# Train a Simple Audio Recognition Model" + ] + }, + { + "cell_type": "markdown", + "metadata": { + "id": "BaFfr7DHRmGF", + "colab_type": "text" + }, + "source": [ + "This notebook demonstrates how to train a 20 kB [Simple Audio Recognition](https://www.tensorflow.org/tutorials/sequences/audio_recognition) model to recognize keywords in speech.\n", + "\n", + "The model created in this notebook is used in the [micro_speech](https://github.com/tensorflow/tensorflow/tree/master/tensorflow/lite/micro/examples/micro_speech) example for [TensorFlow Lite for MicroControllers](https://www.tensorflow.org/lite/microcontrollers/overview).\n", + "\n", + "\n", + " \n", + " \n", + "
\n", + " Run in Google Colab\n", + " \n", + " View source on GitHub\n", + "
\n" + ] + }, + { + "cell_type": "markdown", + "metadata": { + "id": "XaVtYN4nlCft", + "colab_type": "text" + }, + "source": [ + "**Training is much faster using GPU acceleration.** Before you proceed, ensure you are using a GPU runtime by going to **Runtime -> Change runtime type** and set **Hardware accelerator: GPU**. Training 15,000 iterations will take 1.5 - 2 hours on a GPU runtime.\n", + "\n", + "## Configure Defaults\n", + "\n", + "**MODIFY** the following constants for your specific use case." + ] + }, + { + "cell_type": "code", + "metadata": { + "id": "ludfxbNIaegy", + "colab_type": "code", + "colab": {} + }, + "source": [ + "# A comma-delimited list of the words you want to train for.\n", + "# The options are: yes,no,up,down,left,right,on,off,stop,go\n", + "# All the other words will be used to train an \"unknown\" label and silent\n", + "# audio data with no spoken words will be used to train a \"silence\" label.\n", + "WANTED_WORDS = \"yes,no\"\n", + "\n", + "# The number of steps and learning rates can be specified as comma-separated\n", + "# lists to define the rate at each stage. For example,\n", + "# TRAINING_STEPS=12000,3000 and LEARNING_RATE=0.001,0.0001\n", + "# will run 12,000 training loops in total, with a rate of 0.001 for the first\n", + "# 8,000, and 0.0001 for the final 3,000.\n", + "TRAINING_STEPS = \"12000,3000\"\n", + "LEARNING_RATE = \"0.001,0.0001\"\n", + "\n", + "# Calculate the total number of steps, which is used to identify the checkpoint\n", + "# file name.\n", + "TOTAL_STEPS = str(sum(map(lambda string: int(string), TRAINING_STEPS.split(\",\"))))\n", + "\n", + "# Print the configuration to confirm it\n", + "print(\"Training these words: %s\" % WANTED_WORDS)\n", + "print(\"Training steps in each stage: %s\" % TRAINING_STEPS)\n", + "print(\"Learning rate in each stage: %s\" % LEARNING_RATE)\n", + "print(\"Total number of training steps: %s\" % TOTAL_STEPS)" + ], + "execution_count": null, + "outputs": [] + }, + { + "cell_type": "markdown", + "metadata": { + "id": "gCgeOpvY9pAi", + "colab_type": "text" + }, + "source": [ + "**DO NOT MODIFY** the following constants as they include filepaths used in this notebook and data that is shared during training and inference." + ] + }, + { + "cell_type": "code", + "metadata": { + "id": "Nd1iM1o2ymvA", + "colab_type": "code", + "colab": {} + }, + "source": [ + "# Calculate the percentage of 'silence' and 'unknown' training samples required\n", + "# to ensure that we have equal number of samples for each label.\n", + "number_of_labels = WANTED_WORDS.count(',') + 1\n", + "number_of_total_labels = number_of_labels + 2 # for 'silence' and 'unknown' label\n", + "equal_percentage_of_training_samples = int(100.0/(number_of_total_labels))\n", + "SILENT_PERCENTAGE = equal_percentage_of_training_samples\n", + "UNKNOWN_PERCENTAGE = equal_percentage_of_training_samples\n", + "\n", + "# Constants which are shared during training and inference\n", + "PREPROCESS = 'micro'\n", + "WINDOW_STRIDE = 20\n", + "MODEL_ARCHITECTURE = 'tiny_conv' # Other options include: single_fc, conv,\n", + " # low_latency_conv, low_latency_svdf, tiny_embedding_conv\n", + "\n", + "# Constants used during training only\n", + "VERBOSITY = 'WARN'\n", + "EVAL_STEP_INTERVAL = '1000'\n", + "SAVE_STEP_INTERVAL = '1000'\n", + "\n", + "# Constants for training directories and filepaths\n", + "DATASET_DIR = 'dataset/'\n", + "LOGS_DIR = 'logs/'\n", + "TRAIN_DIR = 'train/' # for training checkpoints and other files.\n", + "\n", + "# Constants for inference directories and filepaths\n", + "import os\n", + "MODELS_DIR = 'models'\n", + "if not os.path.exists(MODELS_DIR):\n", + " os.mkdir(MODELS_DIR)\n", + "MODEL_TF = os.path.join(MODELS_DIR, 'model.pb')\n", + "MODEL_TFLITE = os.path.join(MODELS_DIR, 'model.tflite')\n", + "FLOAT_MODEL_TFLITE = os.path.join(MODELS_DIR, 'float_model.tflite')\n", + "MODEL_TFLITE_MICRO = os.path.join(MODELS_DIR, 'model.cc')\n", + "SAVED_MODEL = os.path.join(MODELS_DIR, 'saved_model')\n", + "\n", + "QUANT_INPUT_MIN = 0.0\n", + "QUANT_INPUT_MAX = 26.0\n", + "QUANT_INPUT_RANGE = QUANT_INPUT_MAX - QUANT_INPUT_MIN" + ], + "execution_count": null, + "outputs": [] + }, + { + "cell_type": "markdown", + "metadata": { + "id": "6rLYpvtg9P4o", + "colab_type": "text" + }, + "source": [ + "## Setup Environment\n", + "\n", + "Install Dependencies" + ] + }, + { + "cell_type": "code", + "metadata": { + "id": "ed_XpUrU5DvY", + "colab_type": "code", + "colab": {} + }, + "source": [ + "%tensorflow_version 1.x\n", + "import tensorflow as tf" + ], + "execution_count": null, + "outputs": [] + }, + { + "cell_type": "markdown", + "metadata": { + "id": "T9Ty5mR58E4i", + "colab_type": "text" + }, + "source": [ + "**DELETE** any old data from previous runs\n" + ] + }, + { + "cell_type": "code", + "metadata": { + "id": "APGx0fEh7hFF", + "colab_type": "code", + "colab": {} + }, + "source": [ + "!rm -rf {DATASET_DIR} {LOGS_DIR} {TRAIN_DIR} {MODELS_DIR}" + ], + "execution_count": null, + "outputs": [] + }, + { + "cell_type": "markdown", + "metadata": { + "id": "GfEUlfFBizio", + "colab_type": "text" + }, + "source": [ + "Clone the TensorFlow Github Repository, which contains the relevant code required to run this tutorial." + ] + }, + { + "cell_type": "code", + "metadata": { + "id": "yZArmzT85SLq", + "colab_type": "code", + "colab": {} + }, + "source": [ + "!git clone -q --depth 1 https://github.com/tensorflow/tensorflow" + ], + "execution_count": null, + "outputs": [] + }, + { + "cell_type": "markdown", + "metadata": { + "id": "nS9swHLSi7Bi", + "colab_type": "text" + }, + "source": [ + "Load TensorBoard to visualize the accuracy and loss as training proceeds.\n" + ] + }, + { + "cell_type": "code", + "metadata": { + "id": "q4qF1VxP3UE4", + "colab_type": "code", + "colab": {} + }, + "source": [ + "%load_ext tensorboard\n", + "%tensorboard --logdir {LOGS_DIR}" + ], + "execution_count": null, + "outputs": [] + }, + { + "cell_type": "markdown", + "metadata": { + "id": "x1J96Ron-O4R", + "colab_type": "text" + }, + "source": [ + "## Training\n", + "\n", + "The following script downloads the dataset and begin training." + ] + }, + { + "cell_type": "code", + "metadata": { + "id": "VJsEZx6lynbY", + "colab_type": "code", + "colab": {} + }, + "source": [ + "!python tensorflow/tensorflow/examples/speech_commands/train.py \\\n", + "--data_dir={DATASET_DIR} \\\n", + "--wanted_words={WANTED_WORDS} \\\n", + "--silence_percentage={SILENT_PERCENTAGE} \\\n", + "--unknown_percentage={UNKNOWN_PERCENTAGE} \\\n", + "--preprocess={PREPROCESS} \\\n", + "--window_stride={WINDOW_STRIDE} \\\n", + "--model_architecture={MODEL_ARCHITECTURE} \\\n", + "--how_many_training_steps={TRAINING_STEPS} \\\n", + "--learning_rate={LEARNING_RATE} \\\n", + "--train_dir={TRAIN_DIR} \\\n", + "--summaries_dir={LOGS_DIR} \\\n", + "--verbosity={VERBOSITY} \\\n", + "--eval_step_interval={EVAL_STEP_INTERVAL} \\\n", + "--save_step_interval={SAVE_STEP_INTERVAL}" + ], + "execution_count": null, + "outputs": [] + }, + { + "cell_type": "markdown", + "metadata": { + "id": "UczQKtqLi7OJ", + "colab_type": "text" + }, + "source": [ + "# Skipping the training\n", + "\n", + "If you don't want to spend an hour or two training the model from scratch, you can download pretrained checkpoints by uncommenting the lines below (removing the '#'s at the start of each line) and running them." + ] + }, + { + "cell_type": "code", + "metadata": { + "id": "RZw3VNlnla-J", + "colab_type": "code", + "colab": {} + }, + "source": [ + "#!curl -O \"https://storage.googleapis.com/download.tensorflow.org/models/tflite/speech_micro_train_2020_05_10.tgz\"\n", + "#!tar xzf speech_micro_train_2020_05_10.tgz" + ], + "execution_count": null, + "outputs": [] + }, + { + "cell_type": "markdown", + "metadata": { + "id": "XQUJLrdS-ftl", + "colab_type": "text" + }, + "source": [ + "## Generate a TensorFlow Model for Inference\n", + "\n", + "Combine relevant training results (graph, weights, etc) into a single file for inference. This process is known as freezing a model and the resulting model is known as a frozen model/graph, as it cannot be further re-trained after this process." + ] + }, + { + "cell_type": "code", + "metadata": { + "id": "xyc3_eLh9sAg", + "colab_type": "code", + "colab": {} + }, + "source": [ + "!rm -rf {SAVED_MODEL}\n", + "!python tensorflow/tensorflow/examples/speech_commands/freeze.py \\\n", + "--wanted_words=$WANTED_WORDS \\\n", + "--window_stride_ms=$WINDOW_STRIDE \\\n", + "--preprocess=$PREPROCESS \\\n", + "--model_architecture=$MODEL_ARCHITECTURE \\\n", + "--start_checkpoint=$TRAIN_DIR$MODEL_ARCHITECTURE'.ckpt-'{TOTAL_STEPS} \\\n", + "--save_format=saved_model \\\n", + "--output_file={SAVED_MODEL}" + ], + "execution_count": null, + "outputs": [] + }, + { + "cell_type": "markdown", + "metadata": { + "id": "_DBGDxVI-nKG", + "colab_type": "text" + }, + "source": [ + "## Generate a TensorFlow Lite Model\n", + "\n", + "Convert the frozen graph into a TensorFlow Lite model, which is fully quantized for use with embedded devices.\n", + "\n", + "The following cell will also print the model size, which will be under 20 kilobytes." + ] + }, + { + "cell_type": "code", + "metadata": { + "id": "RIitkqvGWmre", + "colab_type": "code", + "colab": {} + }, + "source": [ + "import sys\n", + "# We add this path so we can import the speech processing modules.\n", + "sys.path.append(\"/content/tensorflow/tensorflow/examples/speech_commands/\")\n", + "import input_data\n", + "import models\n", + "import numpy as np" + ], + "execution_count": null, + "outputs": [] + }, + { + "cell_type": "code", + "metadata": { + "id": "kzqECqMxgBh4", + "colab_type": "code", + "colab": {} + }, + "source": [ + "SAMPLE_RATE = 16000\n", + "CLIP_DURATION_MS = 1000\n", + "WINDOW_SIZE_MS = 30.0\n", + "FEATURE_BIN_COUNT = 40\n", + "BACKGROUND_FREQUENCY = 0.8\n", + "BACKGROUND_VOLUME_RANGE = 0.1\n", + "TIME_SHIFT_MS = 100.0\n", + "\n", + "DATA_URL = 'https://storage.googleapis.com/download.tensorflow.org/data/speech_commands_v0.02.tar.gz'\n", + "VALIDATION_PERCENTAGE = 10\n", + "TESTING_PERCENTAGE = 10" + ], + "execution_count": null, + "outputs": [] + }, + { + "cell_type": "code", + "metadata": { + "id": "rNQdAplJV1fz", + "colab_type": "code", + "colab": {} + }, + "source": [ + "model_settings = models.prepare_model_settings(\n", + " len(input_data.prepare_words_list(WANTED_WORDS.split(','))),\n", + " SAMPLE_RATE, CLIP_DURATION_MS, WINDOW_SIZE_MS,\n", + " WINDOW_STRIDE, FEATURE_BIN_COUNT, PREPROCESS)\n", + "audio_processor = input_data.AudioProcessor(\n", + " DATA_URL, DATASET_DIR,\n", + " SILENT_PERCENTAGE, UNKNOWN_PERCENTAGE,\n", + " WANTED_WORDS.split(','), VALIDATION_PERCENTAGE,\n", + " TESTING_PERCENTAGE, model_settings, LOGS_DIR)" + ], + "execution_count": null, + "outputs": [] + }, + { + "cell_type": "code", + "metadata": { + "id": "lBj_AyCh1cC0", + "colab_type": "code", + "colab": {} + }, + "source": [ + "with tf.Session() as sess:\n", + " float_converter = tf.lite.TFLiteConverter.from_saved_model(SAVED_MODEL)\n", + " float_tflite_model = float_converter.convert()\n", + " float_tflite_model_size = open(FLOAT_MODEL_TFLITE, \"wb\").write(float_tflite_model)\n", + " print(\"Float model is %d bytes\" % float_tflite_model_size)\n", + "\n", + " converter = tf.lite.TFLiteConverter.from_saved_model(SAVED_MODEL)\n", + " converter.optimizations = [tf.lite.Optimize.DEFAULT]\n", + " converter.inference_input_type = tf.lite.constants.INT8\n", + " converter.inference_output_type = tf.lite.constants.INT8\n", + " def representative_dataset_gen():\n", + " for i in range(100):\n", + " data, _ = audio_processor.get_data(1, i*1, model_settings,\n", + " BACKGROUND_FREQUENCY, \n", + " BACKGROUND_VOLUME_RANGE,\n", + " TIME_SHIFT_MS,\n", + " 'testing',\n", + " sess)\n", + " flattened_data = np.array(data.flatten(), dtype=np.float32).reshape(1, 1960)\n", + " yield [flattened_data]\n", + " converter.representative_dataset = representative_dataset_gen\n", + " tflite_model = converter.convert()\n", + " tflite_model_size = open(MODEL_TFLITE, \"wb\").write(tflite_model)\n", + " print(\"Quantized model is %d bytes\" % tflite_model_size)\n" + ], + "execution_count": null, + "outputs": [] + }, + { + "cell_type": "markdown", + "metadata": { + "id": "EeLiDZTbLkzv", + "colab_type": "text" + }, + "source": [ + "## Testing the TensorFlow Lite model's accuracy\n", + "\n", + "Verify that the model we've exported is still accurate, using the TF Lite Python API and our test set." + ] + }, + { + "cell_type": "code", + "metadata": { + "id": "wQsEteKRLryJ", + "colab_type": "code", + "colab": {} + }, + "source": [ + "# Helper function to run inference\n", + "def run_tflite_inference(tflite_model_path, model_type=\"Float\"):\n", + " # Load test data\n", + " np.random.seed(0) # set random seed for reproducible test results.\n", + " with tf.Session() as sess:\n", + " test_data, test_labels = audio_processor.get_data(\n", + " -1, 0, model_settings, BACKGROUND_FREQUENCY, BACKGROUND_VOLUME_RANGE,\n", + " TIME_SHIFT_MS, 'testing', sess)\n", + " test_data = np.expand_dims(test_data, axis=1).astype(np.float32)\n", + "\n", + " # Initialize the interpreter\n", + " interpreter = tf.lite.Interpreter(tflite_model_path)\n", + " interpreter.allocate_tensors()\n", + "\n", + " input_details = interpreter.get_input_details()[0]\n", + " output_details = interpreter.get_output_details()[0]\n", + "\n", + " # For quantized models, manually quantize the input data from float to integer\n", + " if model_type == \"Quantized\":\n", + " input_scale, input_zero_point = input_details[\"quantization\"]\n", + " test_data = test_data / input_scale + input_zero_point\n", + " test_data = test_data.astype(input_details[\"dtype\"])\n", + "\n", + " correct_predictions = 0\n", + " for i in range(len(test_data)):\n", + " interpreter.set_tensor(input_details[\"index\"], test_data[i])\n", + " interpreter.invoke()\n", + " output = interpreter.get_tensor(output_details[\"index\"])[0]\n", + " top_prediction = output.argmax()\n", + " correct_predictions += (top_prediction == test_labels[i])\n", + "\n", + " print('%s model accuracy is %f%% (Number of test samples=%d)' % (\n", + " model_type, (correct_predictions * 100) / len(test_data), len(test_data)))" + ], + "execution_count": 110, + "outputs": [] + }, + { + "cell_type": "code", + "metadata": { + "id": "l-pD52Na6jRa", + "colab_type": "code", + "colab": {} + }, + "source": [ + "# Compute float model accuracy\n", + "run_tflite_inference(FLOAT_MODEL_TFLITE)\n", + "\n", + "# Compute quantized model accuracy\n", + "run_tflite_inference(MODEL_TFLITE, model_type='Quantized')" + ], + "execution_count": 111, + "outputs": [] + }, + { + "cell_type": "markdown", + "metadata": { + "id": "dt6Zqbxu-wIi", + "colab_type": "text" + }, + "source": [ + "## Generate a TensorFlow Lite for MicroControllers Model\n", + "Convert the TensorFlow Lite model into a C source file that can be loaded by TensorFlow Lite for Microcontrollers." + ] + }, + { + "cell_type": "code", + "metadata": { + "id": "XohZOTjR8ZyE", + "colab_type": "code", + "colab": {} + }, + "source": [ + "# Install xxd if it is not available\n", + "!apt-get update && apt-get -qq install xxd\n", + "# Convert to a C source file\n", + "!xxd -i {MODEL_TFLITE} > {MODEL_TFLITE_MICRO}\n", + "# Update variable names\n", + "REPLACE_TEXT = MODEL_TFLITE.replace('/', '_').replace('.', '_')\n", + "!sed -i 's/'{REPLACE_TEXT}'/g_model/g' {MODEL_TFLITE_MICRO}" + ], + "execution_count": null, + "outputs": [] + }, + { + "cell_type": "markdown", + "metadata": { + "id": "2pQnN0i_-0L2", + "colab_type": "text" + }, + "source": [ + "## Deploy to a Microcontroller\n", + "\n", + "Follow the instructions in the [micro_speech](https://github.com/tensorflow/tensorflow/tree/master/tensorflow/lite/micro/examples/micro_speech) README.md for [TensorFlow Lite for MicroControllers](https://www.tensorflow.org/lite/microcontrollers/overview) to deploy this model on a specific microcontroller.\n", + "\n", + "**Reference Model:** If you have not modified this notebook, you can follow the instructions as is, to deploy the model. Refer to the [`micro_speech/train/models`](https://github.com/tensorflow/tensorflow/blob/master/tensorflow/lite/micro/examples/micro_speech/train/models) directory to access the models generated in this notebook.\n", + "\n", + "**New Model:** If you have generated a new model to identify different words: (i) Update `kCategoryCount` and `kCategoryLabels` in [`micro_speech/micro_features/micro_model_settings.h`](https://github.com/tensorflow/tensorflow/blob/master/tensorflow/lite/micro/examples/micro_speech/micro_features/micro_model_settings.h) and (ii) Update the values assigned to the variables defined in [`micro_speech/micro_features/model.cc`](https://github.com/tensorflow/tensorflow/blob/master/tensorflow/lite/micro/examples/micro_speech/micro_features/model.cc) with values displayed after running the following cell." + ] + }, + { + "cell_type": "code", + "metadata": { + "id": "eoYyh0VU8pca", + "colab_type": "code", + "colab": {} + }, + "source": [ + "# Print the C source file\n", + "!cat {MODEL_TFLITE_MICRO}" + ], + "execution_count": null, + "outputs": [] + } + ] +} \ No newline at end of file From 5d4a29eaf590b4a3068ef4d0b7bea9d4f7bd9369 Mon Sep 17 00:00:00 2001 From: Akshay Modi Date: Mon, 22 Jun 2020 16:50:08 -0700 Subject: [PATCH 502/867] Special case wrapping of ndarrays in the gradient tape code. PiperOrigin-RevId: 317762474 Change-Id: Ie848ad90a88aff5b2faef4069c3f05887038c367 --- tensorflow/python/eager/BUILD | 1 + tensorflow/python/eager/backprop.py | 24 ++++++++++++++++++-- tensorflow/python/ops/numpy_ops/np_arrays.py | 8 ++----- 3 files changed, 25 insertions(+), 8 deletions(-) diff --git a/tensorflow/python/eager/BUILD b/tensorflow/python/eager/BUILD index f51bd97e488..408d784ae82 100644 --- a/tensorflow/python/eager/BUILD +++ b/tensorflow/python/eager/BUILD @@ -588,6 +588,7 @@ py_library( "//tensorflow/python:tensor_shape", "//tensorflow/python:unconnected_gradients", "//tensorflow/python:util", + "//tensorflow/python/ops/numpy_ops:numpy", "//tensorflow/python/ops/parallel_for:control_flow_ops", "@six_archive//:six", ], diff --git a/tensorflow/python/eager/backprop.py b/tensorflow/python/eager/backprop.py index 8da3f71360a..5800a51f89a 100644 --- a/tensorflow/python/eager/backprop.py +++ b/tensorflow/python/eager/backprop.py @@ -62,6 +62,9 @@ from tensorflow.python.util.tf_export import tf_export pfor_ops = LazyLoader( "pfor_ops", globals(), "tensorflow.python.ops.parallel_for.control_flow_ops") +np_arrays = LazyLoader( + "np_arrays", globals(), + "tensorflow.python.ops.numpy_ops.np_arrays") function = LazyLoader("function", globals(), "tensorflow.python.eager.function") @@ -721,9 +724,11 @@ pywrap_tfe.TFE_Py_RegisterVSpace(_default_vspace) def _handle_or_self(x): - """If x is ResourceVariable, return its handle, else x.""" + """Unwrap resource variable/ndarray to return tensors.""" if resource_variable_ops.is_resource_variable(x): - x = x.handle + return x.handle + if isinstance(x, np_arrays.ndarray): + return x.data return x @@ -1023,6 +1028,7 @@ class GradientTape(object): "gradient in order to compute higher order " "derivatives.", 1) + num_ndarrays = 0 flat_targets = [] for t in nest.flatten(target): if not backprop_util.IsTrainable(t): @@ -1033,7 +1039,12 @@ class GradientTape(object): if resource_variable_ops.is_resource_variable(t): with self: t = ops.convert_to_tensor(t) + elif isinstance(t, np_arrays.ndarray): + t = t.data + num_ndarrays += 1 flat_targets.append(t) + # Only rewrap if all targets are ndarray. If not, prefer tensors. + rewrap_as_ndarray = num_ndarrays == len(flat_targets) flat_sources = nest.flatten(sources) flat_sources_raw = flat_sources @@ -1066,6 +1077,9 @@ class GradientTape(object): self._watched_variables = self._tape.watched_variables() self._tape = None + if rewrap_as_ndarray: + flat_grad = nest.map_structure(np_arrays.tensor_to_ndarray, flat_grad) + grad = nest.pack_sequence_as(sources, flat_grad) return grad @@ -1120,6 +1134,10 @@ class GradientTape(object): ValueError: If vectorization of jacobian computation fails. """ flat_sources = nest.flatten(sources) + rewrap_as_ndarray = False + if isinstance(target, np_arrays.ndarray): + target = target.data + rewrap_as_ndarray = True target_static_shape = target.shape target_shape = array_ops.shape(target) # Note that we push and pop the tape here and below. This is needed since we @@ -1169,6 +1187,8 @@ class GradientTape(object): out = array_ops.reshape(out, new_shape) if context.executing_eagerly(): out.set_shape(target_static_shape.concatenate(flat_sources[i].shape)) + if rewrap_as_ndarray: + out = np_arrays.tensor_to_ndarray(out) output[i] = out return nest.pack_sequence_as(sources, output) diff --git a/tensorflow/python/ops/numpy_ops/np_arrays.py b/tensorflow/python/ops/numpy_ops/np_arrays.py index fd26318bea9..eca84421d1b 100644 --- a/tensorflow/python/ops/numpy_ops/np_arrays.py +++ b/tensorflow/python/ops/numpy_ops/np_arrays.py @@ -82,10 +82,10 @@ class NdarraySpec(type_spec.BatchableTypeSpec): return (self._data_spec,) def _batch(self, batch_size): - return NdarraySpec(self._data_spec.batch(batch_size)) + return NdarraySpec(self._data_spec._batch(batch_size)) # pylint: disable=protected-access def _unbatch(self): - return NdarraySpec(self._data_spec.unbatch()) + return NdarraySpec(self._data_spec._unbatch()) # pylint: disable=protected-access class ndarray(composite_tensor.CompositeTensor): # pylint: disable=invalid-name @@ -306,10 +306,6 @@ class ndarray(composite_tensor.CompositeTensor): # pylint: disable=invalid-name def __repr__(self): return 'ndarray<{}>'.format(self.data.__repr__()) - @property - def _id(self): - return self.data._id # pylint: disable=protected-access - def tensor_to_ndarray(tensor): return ndarray.from_tensor(tensor) From 6b0b4876f8fe024485d02080d0be2d541ece6b17 Mon Sep 17 00:00:00 2001 From: Peng Wang Date: Mon, 22 Jun 2020 16:54:12 -0700 Subject: [PATCH 503/867] [TF-numpy] Adds __matmul__ method to ndarray. PiperOrigin-RevId: 317763209 Change-Id: I3a2ab8e07b040144239ecee58d4ba9267f5b8977 --- tensorflow/python/ops/numpy_ops/np_math_ops.py | 1 + 1 file changed, 1 insertion(+) diff --git a/tensorflow/python/ops/numpy_ops/np_math_ops.py b/tensorflow/python/ops/numpy_ops/np_math_ops.py index b242e51a2e5..427aa96e5a4 100644 --- a/tensorflow/python/ops/numpy_ops/np_math_ops.py +++ b/tensorflow/python/ops/numpy_ops/np_math_ops.py @@ -954,6 +954,7 @@ setattr(np_arrays.ndarray, '__pow__', _wrap(power)) setattr(np_arrays.ndarray, '__rpow__', _wrap(power, True)) setattr(np_arrays.ndarray, '__truediv__', _wrap(true_divide)) setattr(np_arrays.ndarray, '__rtruediv__', _wrap(true_divide, True)) +setattr(np_arrays.ndarray, '__matmul__', _wrap(matmul)) def _comparison(tf_fun, x1, x2, cast_bool_to_int=False): From 23e387975e683a971dd0cbe2198cb60c5cf71fc4 Mon Sep 17 00:00:00 2001 From: Frank Chen Date: Mon, 22 Jun 2020 16:58:50 -0700 Subject: [PATCH 504/867] Rename tpu_load_library to tpu_api_dlsym_initializer. This better reflects what the load library module does. PiperOrigin-RevId: 317763973 Change-Id: I66169394c2f1d77b260323cb81d76f0f4e167579 --- tensorflow/core/BUILD | 2 +- tensorflow/core/framework/load_library.cc | 2 +- tensorflow/core/tpu/BUILD | 24 ++++++-- tensorflow/core/tpu/kernels/BUILD | 4 +- .../core/tpu/kernels/tpu_configuration_ops.cc | 2 +- .../tpu/kernels/tpu_mesh_state_interface.h | 2 +- tensorflow/core/tpu/tpu_api.cc | 57 +++++++++++++++++++ .../tpu/{tpu_library_loader.h => tpu_api.h} | 12 +--- ...loader.cc => tpu_api_dlsym_initializer.cc} | 40 +------------ .../core/tpu/tpu_api_dlsym_initializer.h | 38 +++++++++++++ ...c => tpu_api_dlsym_initializer_windows.cc} | 20 +------ tensorflow/stream_executor/tpu/BUILD | 12 ++-- .../stream_executor/tpu/tpu_executor.cc | 2 +- .../stream_executor/tpu/tpu_node_context.cc | 2 +- .../stream_executor/tpu/tpu_platform.cc | 2 +- tensorflow/stream_executor/tpu/tpu_stream.h | 2 +- tensorflow/stream_executor/tpu/tpu_timer.h | 2 +- .../tpu/tpu_transfer_manager.cc | 2 +- 18 files changed, 141 insertions(+), 86 deletions(-) create mode 100644 tensorflow/core/tpu/tpu_api.cc rename tensorflow/core/tpu/{tpu_library_loader.h => tpu_api.h} (77%) rename tensorflow/core/tpu/{tpu_library_loader.cc => tpu_api_dlsym_initializer.cc} (71%) create mode 100644 tensorflow/core/tpu/tpu_api_dlsym_initializer.h rename tensorflow/core/tpu/{tpu_library_loader_windows.cc => tpu_api_dlsym_initializer_windows.cc} (64%) diff --git a/tensorflow/core/BUILD b/tensorflow/core/BUILD index 695035c91e9..d0be6ee9597 100644 --- a/tensorflow/core/BUILD +++ b/tensorflow/core/BUILD @@ -2259,7 +2259,7 @@ tf_cuda_library( "//tensorflow/core/platform/default/build_config:platformlib", "//tensorflow/core/profiler/lib:annotated_traceme", "//tensorflow/core/profiler/lib:traceme", - "//tensorflow/core/tpu:tpu_library_loader", + "//tensorflow/core/tpu:tpu_api_dlsym_initializer", "//tensorflow/core/util:einsum_op_util", "//tensorflow/core/util:padding", "//tensorflow/core/util:port", diff --git a/tensorflow/core/framework/load_library.cc b/tensorflow/core/framework/load_library.cc index ab08d644074..22181e1c8be 100644 --- a/tensorflow/core/framework/load_library.cc +++ b/tensorflow/core/framework/load_library.cc @@ -22,7 +22,7 @@ limitations under the License. #include "tensorflow/core/platform/env.h" #include "tensorflow/core/platform/mem.h" #if !defined(IS_MOBILE_PLATFORM) -#include "tensorflow/core/tpu/tpu_library_loader.h" +#include "tensorflow/core/tpu/tpu_api_dlsym_initializer.h" #endif // IS_MOBILE_PLATFORM namespace tensorflow { diff --git a/tensorflow/core/tpu/BUILD b/tensorflow/core/tpu/BUILD index 9e89cd69235..aa811f23672 100644 --- a/tensorflow/core/tpu/BUILD +++ b/tensorflow/core/tpu/BUILD @@ -107,15 +107,31 @@ cc_library( ) cc_library( - name = "tpu_library_loader", + name = "tpu_api", + srcs = ["tpu_api.cc"], + hdrs = ["tpu_api.h"], + deps = [ + ":libtftpu_header", + ":tpu_config_c_api", + "//tensorflow/core/tpu/kernels:tpu_compile_c_api_hdrs", + "//tensorflow/core/tpu/kernels:tpu_mesh_state_c_api_hdrs", + "//tensorflow/core/tpu/kernels:tpu_util_c_api_hdrs", + "//tensorflow/stream_executor/tpu:tpu_executor_c_api_hdrs", + "//tensorflow/stream_executor/tpu:tpu_node_context_c_api_hdrs", + ], +) + +cc_library( + name = "tpu_api_dlsym_initializer", srcs = if_windows( - ["tpu_library_loader_windows.cc"], - otherwise = ["tpu_library_loader.cc"], + ["tpu_api_dlsym_initializer_windows.cc"], + otherwise = ["tpu_api_dlsym_initializer.cc"], ), - hdrs = ["tpu_library_loader.h"], + hdrs = ["tpu_api_dlsym_initializer.h"], visibility = ["//visibility:public"], deps = [ ":libtftpu_header", + ":tpu_api", ":tpu_config_c_api", ":tpu_library_init_fns", "//tensorflow/core/platform:errors", diff --git a/tensorflow/core/tpu/kernels/BUILD b/tensorflow/core/tpu/kernels/BUILD index a9f2202cd45..f69c97b81de 100644 --- a/tensorflow/core/tpu/kernels/BUILD +++ b/tensorflow/core/tpu/kernels/BUILD @@ -72,10 +72,10 @@ tf_kernel_library( "//tensorflow/core:framework", "//tensorflow/core:protos_all_cc", "//tensorflow/core/platform:refcount", + "//tensorflow/core/tpu:tpu_api", "//tensorflow/core/tpu:tpu_config_c_api", "//tensorflow/core/tpu:tpu_configuration", "//tensorflow/core/tpu:tpu_defs", - "//tensorflow/core/tpu:tpu_library_loader", "//tensorflow/stream_executor/tpu:proto_helper", ], alwayslink = 1, @@ -224,7 +224,7 @@ cc_library( "//tensorflow/compiler/xla/service", "//tensorflow/core:framework", "//tensorflow/core/protobuf/tpu:compile_metadata_proto_cc", - "//tensorflow/core/tpu:tpu_library_loader", + "//tensorflow/core/tpu:tpu_api", ], ) diff --git a/tensorflow/core/tpu/kernels/tpu_configuration_ops.cc b/tensorflow/core/tpu/kernels/tpu_configuration_ops.cc index 12a3256a44f..583f1aec207 100644 --- a/tensorflow/core/tpu/kernels/tpu_configuration_ops.cc +++ b/tensorflow/core/tpu/kernels/tpu_configuration_ops.cc @@ -24,10 +24,10 @@ limitations under the License. #include "tensorflow/core/framework/tensor_shape.h" #include "tensorflow/core/platform/refcount.h" #include "tensorflow/core/tpu/kernels/tpu_mesh_state_interface.h" +#include "tensorflow/core/tpu/tpu_api.h" #include "tensorflow/core/tpu/tpu_config_c_api.h" #include "tensorflow/core/tpu/tpu_configuration.h" #include "tensorflow/core/tpu/tpu_defs.h" -#include "tensorflow/core/tpu/tpu_library_loader.h" #include "tensorflow/stream_executor/tpu/proto_helper.h" namespace tensorflow { diff --git a/tensorflow/core/tpu/kernels/tpu_mesh_state_interface.h b/tensorflow/core/tpu/kernels/tpu_mesh_state_interface.h index 3eff3be4915..e2ac38b5f84 100644 --- a/tensorflow/core/tpu/kernels/tpu_mesh_state_interface.h +++ b/tensorflow/core/tpu/kernels/tpu_mesh_state_interface.h @@ -21,7 +21,7 @@ limitations under the License. #include "tensorflow/core/protobuf/tpu/compile_metadata.pb.h" #include "tensorflow/core/tpu/kernels/tpu_compile_c_api.h" #include "tensorflow/core/tpu/kernels/tpu_mesh_state_c_api.h" -#include "tensorflow/core/tpu/tpu_library_loader.h" +#include "tensorflow/core/tpu/tpu_api.h" namespace tensorflow { diff --git a/tensorflow/core/tpu/tpu_api.cc b/tensorflow/core/tpu/tpu_api.cc new file mode 100644 index 00000000000..8dad82b3029 --- /dev/null +++ b/tensorflow/core/tpu/tpu_api.cc @@ -0,0 +1,57 @@ +/* Copyright 2020 The TensorFlow Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. +==============================================================================*/ + +#include "tensorflow/core/tpu/tpu_api.h" + +namespace tensorflow { +namespace tpu { + +TfTpu_BaseFn* InitializeApiFn() { + static TfTpu_BaseFn base_fn; + return &base_fn; +} + +TfTpu_ConfigApiFn* ConfigApiFn() { + static TfTpu_ConfigApiFn config_api_fn; + return &config_api_fn; +} + +TfTpu_MeshStateApiFn* MeshStateApiFn() { + static TfTpu_MeshStateApiFn mesh_state_api_fn; + return &mesh_state_api_fn; +} + +TfTpu_CompileApiFn* CompileApiFn() { + static TfTpu_CompileApiFn compile_api_fn; + return &compile_api_fn; +} + +TfTpu_ExecutorApiFn* ExecutorApiFn() { + static TfTpu_ExecutorApiFn executor_api_fn; + return &executor_api_fn; +} + +TfTpu_NodeContextApiFn* NodeContextApiFn() { + static TfTpu_NodeContextApiFn node_context_api_fn; + return &node_context_api_fn; +} + +TfTpu_UtilApiFn* UtilApiFn() { + static TfTpu_UtilApiFn util_api_fn; + return &util_api_fn; +} + +} // namespace tpu +} // namespace tensorflow diff --git a/tensorflow/core/tpu/tpu_library_loader.h b/tensorflow/core/tpu/tpu_api.h similarity index 77% rename from tensorflow/core/tpu/tpu_library_loader.h rename to tensorflow/core/tpu/tpu_api.h index ba6c324707d..c47ace6601d 100644 --- a/tensorflow/core/tpu/tpu_library_loader.h +++ b/tensorflow/core/tpu/tpu_api.h @@ -13,10 +13,9 @@ See the License for the specific language governing permissions and limitations under the License. ==============================================================================*/ -#ifndef TENSORFLOW_CORE_TPU_TPU_LIBRARY_LOADER_H_ -#define TENSORFLOW_CORE_TPU_TPU_LIBRARY_LOADER_H_ +#ifndef TENSORFLOW_CORE_TPU_TPU_API_H_ +#define TENSORFLOW_CORE_TPU_TPU_API_H_ -#include "tensorflow/core/platform/status.h" #include "tensorflow/core/tpu/kernels/tpu_compile_c_api.h" #include "tensorflow/core/tpu/kernels/tpu_mesh_state_c_api.h" #include "tensorflow/core/tpu/kernels/tpu_util_c_api.h" @@ -25,13 +24,9 @@ limitations under the License. #include "tensorflow/stream_executor/tpu/tpu_executor_c_api.h" #include "tensorflow/stream_executor/tpu/tpu_node_context_c_api.h" -// LINT.IfChange namespace tensorflow { namespace tpu { -Status InitializeTpuLibrary(void* library_handle); - -// TODO(frankchn): Separate out API functions from the loader. TfTpu_BaseFn* InitializeApiFn(); TfTpu_ConfigApiFn* ConfigApiFn(); @@ -48,6 +43,5 @@ TfTpu_UtilApiFn* UtilApiFn(); } // namespace tpu } // namespace tensorflow -// LINT.ThenChange(//tensorflow/core/tpu/tpu_library_loader_windows.cc) -#endif // TENSORFLOW_CORE_TPU_TPU_LIBRARY_LOADER_H_ +#endif // TENSORFLOW_CORE_TPU_TPU_API_H_ diff --git a/tensorflow/core/tpu/tpu_library_loader.cc b/tensorflow/core/tpu/tpu_api_dlsym_initializer.cc similarity index 71% rename from tensorflow/core/tpu/tpu_library_loader.cc rename to tensorflow/core/tpu/tpu_api_dlsym_initializer.cc index 834b86e68a7..c6666421327 100644 --- a/tensorflow/core/tpu/tpu_library_loader.cc +++ b/tensorflow/core/tpu/tpu_api_dlsym_initializer.cc @@ -13,14 +13,13 @@ See the License for the specific language governing permissions and limitations under the License. ==============================================================================*/ -// TODO(frankchn): Rename to `tpu_api_dlsym_initializer` or similar. - -#include "tensorflow/core/tpu/tpu_library_loader.h" +#include "tensorflow/core/tpu/tpu_api_dlsym_initializer.h" #include #include "tensorflow/core/platform/errors.h" #include "tensorflow/core/platform/status.h" +#include "tensorflow/core/tpu/tpu_api.h" #include "tensorflow/stream_executor/tpu/tpu_node_context_c_api.h" #include "tensorflow/stream_executor/tpu/tpu_platform.h" @@ -39,41 +38,6 @@ namespace tpu { #include "tensorflow/core/tpu/tpu_library_init_fns.inc" -TfTpu_BaseFn* InitializeApiFn() { - static TfTpu_BaseFn base_fn; - return &base_fn; -} - -TfTpu_ConfigApiFn* ConfigApiFn() { - static TfTpu_ConfigApiFn config_api_fn; - return &config_api_fn; -} - -TfTpu_MeshStateApiFn* MeshStateApiFn() { - static TfTpu_MeshStateApiFn mesh_state_api_fn; - return &mesh_state_api_fn; -} - -TfTpu_CompileApiFn* CompileApiFn() { - static TfTpu_CompileApiFn compile_api_fn; - return &compile_api_fn; -} - -TfTpu_ExecutorApiFn* ExecutorApiFn() { - static TfTpu_ExecutorApiFn executor_api_fn; - return &executor_api_fn; -} - -TfTpu_NodeContextApiFn* NodeContextApiFn() { - static TfTpu_NodeContextApiFn node_context_api_fn; - return &node_context_api_fn; -} - -TfTpu_UtilApiFn* UtilApiFn() { - static TfTpu_UtilApiFn util_api_fn; - return &util_api_fn; -} - Status InitializeTpuLibrary(void* library_handle) { bool shared_object_loaded = true; if (library_handle == nullptr) { diff --git a/tensorflow/core/tpu/tpu_api_dlsym_initializer.h b/tensorflow/core/tpu/tpu_api_dlsym_initializer.h new file mode 100644 index 00000000000..257fa25ad37 --- /dev/null +++ b/tensorflow/core/tpu/tpu_api_dlsym_initializer.h @@ -0,0 +1,38 @@ +/* Copyright 2020 The TensorFlow Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. +==============================================================================*/ + +#ifndef TENSORFLOW_CORE_TPU_TPU_API_DLSYM_INITIALIZER_H_ +#define TENSORFLOW_CORE_TPU_TPU_API_DLSYM_INITIALIZER_H_ + +#include "tensorflow/core/platform/status.h" +#include "tensorflow/core/tpu/kernels/tpu_compile_c_api.h" +#include "tensorflow/core/tpu/kernels/tpu_mesh_state_c_api.h" +#include "tensorflow/core/tpu/kernels/tpu_util_c_api.h" +#include "tensorflow/core/tpu/libtftpu.h" +#include "tensorflow/core/tpu/tpu_config_c_api.h" +#include "tensorflow/stream_executor/tpu/tpu_executor_c_api.h" +#include "tensorflow/stream_executor/tpu/tpu_node_context_c_api.h" + +// LINT.IfChange +namespace tensorflow { +namespace tpu { + +Status InitializeTpuLibrary(void* library_handle); + +} // namespace tpu +} // namespace tensorflow +// LINT.ThenChange(//tensorflow/core/tpu/tpu_api_dlsym_initializer_windows.cc) + +#endif // TENSORFLOW_CORE_TPU_TPU_API_DLSYM_INITIALIZER_H_ diff --git a/tensorflow/core/tpu/tpu_library_loader_windows.cc b/tensorflow/core/tpu/tpu_api_dlsym_initializer_windows.cc similarity index 64% rename from tensorflow/core/tpu/tpu_library_loader_windows.cc rename to tensorflow/core/tpu/tpu_api_dlsym_initializer_windows.cc index 7cf1b5cdb1d..f453a98e558 100644 --- a/tensorflow/core/tpu/tpu_library_loader_windows.cc +++ b/tensorflow/core/tpu/tpu_api_dlsym_initializer_windows.cc @@ -15,28 +15,14 @@ limitations under the License. #include "tensorflow/core/platform/errors.h" #include "tensorflow/core/platform/status.h" -#include "tensorflow/core/tpu/tpu_library_loader.h" +#include "tensorflow/core/tpu/tpu_api_dlsym_initializer.h" -// Reminder: Update tpu_library_loader.cc if you are adding new publicly -// visible methods. +// Reminder: Update tpu_api_dlsym_initializer_windows.cc if you are adding new +// publicly visible methods. namespace tensorflow { namespace tpu { -TfTpu_BaseFn* InitializeApiFn() { return nullptr; } - -TfTpu_ConfigApiFn* ConfigApiFn() { return nullptr; } - -TfTpu_MeshStateApiFn* MeshStateApiFn() { return nullptr; } - -TfTpu_CompileApiFn* CompileApiFn() { return nullptr; } - -TfTpu_ExecutorApiFn* ExecutorApiFn() { return nullptr; } - -TfTpu_NodeContextApiFn* NodeContextApiFn() { return nullptr; } - -TfTpu_UtilApiFn* UtilApiFn() { return nullptr; } - Status InitializeTpuLibrary(void* library_handle) { return errors::Unimplemented( "Loading TPU library is not supported on Windows."); diff --git a/tensorflow/stream_executor/tpu/BUILD b/tensorflow/stream_executor/tpu/BUILD index bf88e9809d0..720ba6bc0c3 100644 --- a/tensorflow/stream_executor/tpu/BUILD +++ b/tensorflow/stream_executor/tpu/BUILD @@ -70,7 +70,7 @@ cc_library( ":status_helper", ":tpu_executor_c_api_hdrs", ":tpu_stream_interface", - "//tensorflow/core/tpu:tpu_library_loader", + "//tensorflow/core/tpu:tpu_api", "//tensorflow/stream_executor:stream", ], ) @@ -81,7 +81,7 @@ cc_library( deps = [ ":tpu_executor_c_api_hdrs", "//tensorflow/core/platform:types", - "//tensorflow/core/tpu:tpu_library_loader", + "//tensorflow/core/tpu:tpu_api", "//tensorflow/stream_executor:stream", ], ) @@ -101,7 +101,7 @@ cc_library( ":tpu_timer", "//tensorflow/c:tf_status", "//tensorflow/core:lib", - "//tensorflow/core/tpu:tpu_library_loader", + "//tensorflow/core/tpu:tpu_api", "//tensorflow/stream_executor:stream", "//tensorflow/stream_executor/lib", "@com_google_absl//absl/container:flat_hash_map", @@ -151,7 +151,7 @@ cc_library( "//tensorflow/compiler/xla/service:stream_pool", "//tensorflow/compiler/xla/service:transfer_manager", "//tensorflow/core:framework", - "//tensorflow/core/tpu:tpu_library_loader", + "//tensorflow/core/tpu:tpu_api", "//tensorflow/stream_executor:device_memory_allocator", "//tensorflow/stream_executor/lib", "@com_google_absl//absl/memory", @@ -169,7 +169,7 @@ cc_library( ":tpu_platform_interface", "//tensorflow/c:tf_status", "//tensorflow/core/platform:types", - "//tensorflow/core/tpu:tpu_library_loader", + "//tensorflow/core/tpu:tpu_api", "//tensorflow/stream_executor:stream", "@com_google_absl//absl/container:flat_hash_map", ], @@ -201,7 +201,7 @@ cc_library( "//tensorflow/compiler/xla:xla_data_proto_cc", "//tensorflow/compiler/xla/service:shaped_buffer", "//tensorflow/compiler/xla/service:transfer_manager", - "//tensorflow/core/tpu:tpu_library_loader", + "//tensorflow/core/tpu:tpu_api", "//tensorflow/stream_executor:stream", ], ) diff --git a/tensorflow/stream_executor/tpu/tpu_executor.cc b/tensorflow/stream_executor/tpu/tpu_executor.cc index cb1410880eb..95c32714732 100644 --- a/tensorflow/stream_executor/tpu/tpu_executor.cc +++ b/tensorflow/stream_executor/tpu/tpu_executor.cc @@ -17,7 +17,7 @@ limitations under the License. #include "tensorflow/c/tf_status.h" #include "tensorflow/core/lib/gtl/cleanup.h" -#include "tensorflow/core/tpu/tpu_library_loader.h" +#include "tensorflow/core/tpu/tpu_api.h" #include "tensorflow/stream_executor/device_memory.h" #include "tensorflow/stream_executor/lib/status.h" #include "tensorflow/stream_executor/tpu/c_api_conversions.h" diff --git a/tensorflow/stream_executor/tpu/tpu_node_context.cc b/tensorflow/stream_executor/tpu/tpu_node_context.cc index 356ede40fb3..b502264cfc7 100644 --- a/tensorflow/stream_executor/tpu/tpu_node_context.cc +++ b/tensorflow/stream_executor/tpu/tpu_node_context.cc @@ -17,7 +17,7 @@ limitations under the License. #include "tensorflow/compiler/xla/service/backend.h" #include "tensorflow/compiler/xla/service/platform_util.h" #include "tensorflow/compiler/xla/service/transfer_manager.h" -#include "tensorflow/core/tpu/tpu_library_loader.h" +#include "tensorflow/core/tpu/tpu_api.h" #include "tensorflow/stream_executor/device_memory_allocator.h" #include "tensorflow/stream_executor/tpu/tpu_executor_c_api.h" #include "tensorflow/stream_executor/tpu/tpu_node_context_c_api.h" diff --git a/tensorflow/stream_executor/tpu/tpu_platform.cc b/tensorflow/stream_executor/tpu/tpu_platform.cc index 4bccd822e91..97a97a63351 100644 --- a/tensorflow/stream_executor/tpu/tpu_platform.cc +++ b/tensorflow/stream_executor/tpu/tpu_platform.cc @@ -16,7 +16,7 @@ limitations under the License. #include "tensorflow/stream_executor/tpu/tpu_platform.h" #include "tensorflow/c/tf_status.h" -#include "tensorflow/core/tpu/tpu_library_loader.h" +#include "tensorflow/core/tpu/tpu_api.h" #include "tensorflow/stream_executor/platform.h" #include "tensorflow/stream_executor/tpu/status_helper.h" #include "tensorflow/stream_executor/tpu/tpu_executor.h" diff --git a/tensorflow/stream_executor/tpu/tpu_stream.h b/tensorflow/stream_executor/tpu/tpu_stream.h index e1aa1164248..209a624b462 100644 --- a/tensorflow/stream_executor/tpu/tpu_stream.h +++ b/tensorflow/stream_executor/tpu/tpu_stream.h @@ -16,7 +16,7 @@ limitations under the License. #ifndef TENSORFLOW_STREAM_EXECUTOR_TPU_TPU_STREAM_H_ #define TENSORFLOW_STREAM_EXECUTOR_TPU_TPU_STREAM_H_ -#include "tensorflow/core/tpu/tpu_library_loader.h" +#include "tensorflow/core/tpu/tpu_api.h" #include "tensorflow/stream_executor/stream_executor_internal.h" #include "tensorflow/stream_executor/tpu/c_api_conversions.h" #include "tensorflow/stream_executor/tpu/status_helper.h" diff --git a/tensorflow/stream_executor/tpu/tpu_timer.h b/tensorflow/stream_executor/tpu/tpu_timer.h index d7f8f660b37..0ad48ce8a80 100644 --- a/tensorflow/stream_executor/tpu/tpu_timer.h +++ b/tensorflow/stream_executor/tpu/tpu_timer.h @@ -17,7 +17,7 @@ limitations under the License. #define TENSORFLOW_STREAM_EXECUTOR_TPU_TPU_TIMER_H_ #include "tensorflow/core/platform/types.h" -#include "tensorflow/core/tpu/tpu_library_loader.h" +#include "tensorflow/core/tpu/tpu_api.h" #include "tensorflow/stream_executor/stream_executor_internal.h" #include "tensorflow/stream_executor/tpu/tpu_executor_c_api.h" diff --git a/tensorflow/stream_executor/tpu/tpu_transfer_manager.cc b/tensorflow/stream_executor/tpu/tpu_transfer_manager.cc index 934fabbf54d..c55af7d58b9 100644 --- a/tensorflow/stream_executor/tpu/tpu_transfer_manager.cc +++ b/tensorflow/stream_executor/tpu/tpu_transfer_manager.cc @@ -17,7 +17,7 @@ limitations under the License. #include "tensorflow/compiler/xla/shape_util.h" #include "tensorflow/compiler/xla/xla_data.pb.h" -#include "tensorflow/core/tpu/tpu_library_loader.h" +#include "tensorflow/core/tpu/tpu_api.h" #include "tensorflow/stream_executor/device_memory.h" #include "tensorflow/stream_executor/tpu/c_api_conversions.h" #include "tensorflow/stream_executor/tpu/proto_helper.h" From 3cc9264e307a1ec221680542c79e86e0f3d13807 Mon Sep 17 00:00:00 2001 From: Peng Wang Date: Mon, 22 Jun 2020 17:00:43 -0700 Subject: [PATCH 505/867] [TF-numpy] Adds an accessor class for numpy_ops, in order to be tf_exported'ed. PiperOrigin-RevId: 317764327 Change-Id: I7af7e421c89b0c754e836a2a84cb67a1a82687a1 --- tensorflow/python/ops/numpy_ops/BUILD | 1 - .../python/ops/numpy_ops/np_accessor.py | 32 ------------------- 2 files changed, 33 deletions(-) delete mode 100644 tensorflow/python/ops/numpy_ops/np_accessor.py diff --git a/tensorflow/python/ops/numpy_ops/BUILD b/tensorflow/python/ops/numpy_ops/BUILD index a70e3f3918d..c4203840c61 100644 --- a/tensorflow/python/ops/numpy_ops/BUILD +++ b/tensorflow/python/ops/numpy_ops/BUILD @@ -11,7 +11,6 @@ py_library( name = "numpy", srcs = [ "__init__.py", - "np_accessor.py", "np_array_ops.py", "np_arrays.py", "np_dtypes.py", diff --git a/tensorflow/python/ops/numpy_ops/np_accessor.py b/tensorflow/python/ops/numpy_ops/np_accessor.py deleted file mode 100644 index 64786d2c50a..00000000000 --- a/tensorflow/python/ops/numpy_ops/np_accessor.py +++ /dev/null @@ -1,32 +0,0 @@ -# Copyright 2020 The TensorFlow Authors. All Rights Reserved. -# -# Licensed under the Apache License, Version 2.0 (the "License"); -# you may not use this file except in compliance with the License. -# You may obtain a copy of the License at -# -# http://www.apache.org/licenses/LICENSE-2.0 -# -# Unless required by applicable law or agreed to in writing, software -# distributed under the License is distributed on an "AS IS" BASIS, -# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -# See the License for the specific language governing permissions and -# limitations under the License. -# ============================================================================== -"""An accessor class for numpy_ops contents.""" - -from __future__ import absolute_import -from __future__ import division -from __future__ import print_function - -from tensorflow.python.ops import numpy_ops - - -class Numpy: - """An accessor class that forwards attribute accesses to module `numpy_ops`. - """ - - def __getattr__(self, attr): - return getattr(numpy_ops, attr) - - -numpy = Numpy() From 1a3b7af373ab365780729d927ff0117951b3ddfb Mon Sep 17 00:00:00 2001 From: Shanqing Cai Date: Mon, 22 Jun 2020 17:05:24 -0700 Subject: [PATCH 506/867] [tfdbg2] Support reading multiple DebugEvent file sets from the same tfdbg run TensorFlow jobs that involve multiple hosts (e.g., parameter-server setups and TPU coordinator-worker setups) can generate >1 DebugEvent file sets when instrumented with tfdbg2's `tf.debugging.experimental.enable_dump_debug_info()`. This CL adds capability to load these multiple file sets belonging to the same tfdbg_run_id to DebugEventsReader and DebugDataReader. PiperOrigin-RevId: 317765159 Change-Id: Ifcf593bd8b404e3e1c3a6f3f3be70bd6b8b73555 --- .../python/debug/lib/debug_events_reader.py | 290 +++++++++++------- .../debug/lib/debug_events_writer_test.py | 87 +++++- .../python/debug/lib/debug_v2_ops_test.py | 19 +- .../debug/lib/dumping_callback_test_lib.py | 7 +- 4 files changed, 282 insertions(+), 121 deletions(-) diff --git a/tensorflow/python/debug/lib/debug_events_reader.py b/tensorflow/python/debug/lib/debug_events_reader.py index 743cea7103a..915c65f7594 100644 --- a/tensorflow/python/debug/lib/debug_events_reader.py +++ b/tensorflow/python/debug/lib/debug_events_reader.py @@ -46,28 +46,37 @@ class DebugEventsReader(object): # penalty. _READER_RELEASE_PER = 100 + _METADATA_SUFFIX = ".metadata" + _SOURCE_FILE_SUFFIX = ".source_files" + _STACK_FRAMES_SUFFIX = ".stack_frames" + _GRAPHS_SUFFIX = ".graphs" + _EXECUTION_SUFFIX = ".execution" + _GRAPH_EXECUTION_TRACES_SUFFIX = ".graph_execution_traces" + def __init__(self, dump_root): if not file_io.is_directory(dump_root): raise ValueError("Specified dump_root is not a directory: %s" % dump_root) - metadata_paths = file_io.get_matching_files( - os.path.join(dump_root, "*.metadata")) - if not metadata_paths: - raise ValueError("Cannot find any metadata file in directory: %s" % - dump_root) - elif len(metadata_paths) > 1: - raise ValueError( - "Unexpected: Found multiple (%d) metadata in directory: %s" % - (len(metadata_paths), dump_root)) - self._metadata_path = compat.as_bytes(metadata_paths[0]) - self._metadata_reader = None + self._dump_root = dump_root + self._metadata_paths = self._load_metadata_files() - prefix = metadata_paths[0][:-len(".metadata")] - self._source_files_path = compat.as_bytes("%s.source_files" % prefix) - self._stack_frames_path = compat.as_bytes("%s.stack_frames" % prefix) - self._graphs_path = compat.as_bytes("%s.graphs" % prefix) - self._execution_path = compat.as_bytes("%s.execution" % prefix) - self._graph_execution_traces_path = compat.as_bytes( - "%s.graph_execution_traces" % prefix) + prefixes = [ + metadata_path[:-len(self._METADATA_SUFFIX)] + for metadata_path in self._metadata_paths + ] + prefix = prefixes[0] # This is the prefix of the main file set. + self._source_files_path = compat.as_bytes(prefix + self._SOURCE_FILE_SUFFIX) + self._stack_frames_path = compat.as_bytes(prefix + + self._STACK_FRAMES_SUFFIX) + self._graphs_path = compat.as_bytes(prefix + self._GRAPHS_SUFFIX) + self._execution_path = compat.as_bytes(prefix + self._EXECUTION_SUFFIX) + # There can be multiple .graph_execution_trace files each belonging + # to a file set generated on an individual host, in the case of + # a distributed TensorFlow job. + # This is different from the other debug event files in the file set. + self._graph_execution_traces_paths = [ + compat.as_bytes(prefix + self._GRAPH_EXECUTION_TRACES_SUFFIX) + for prefix in prefixes + ] self._readers = dict() # A map from file path to reader. # A map from file path to current reading offset. self._reader_offsets = dict() @@ -78,6 +87,91 @@ class DebugEventsReader(object): self._offsets = dict() + def _load_metadata_files(self): + """Load and parse metadata files in the dump root. + + Check that all metadata files have a common tfdbg_run_id, and raise + a ValueError if their tfdbg_run_ids differ. + + Returns: + A list of metadata file paths in ascending order of their starting + wall_time timestamp. + """ + + metadata_paths = file_io.get_matching_files( + os.path.join(self._dump_root, "*%s" % self._METADATA_SUFFIX)) + if not metadata_paths: + raise ValueError("Cannot find any tfdbg metadata file in directory: %s" % + self._dump_root) + wall_times = [] + run_ids = [] + tensorflow_versions = [] + file_versions = [] + for metadata_path in metadata_paths: + reader = tf_record.tf_record_random_reader(metadata_path) + try: + record = reader.read(0)[0] + debug_event = debug_event_pb2.DebugEvent.FromString(record) + wall_times.append(debug_event.wall_time) + run_ids.append(debug_event.debug_metadata.tfdbg_run_id) + tensorflow_versions.append( + debug_event.debug_metadata.tensorflow_version) + file_versions.append(debug_event.debug_metadata.file_version) + finally: + reader.close() + self._starting_wall_time = wall_times[0] + self._tfdbg_run_id = run_ids[0] + self._tensorflow_version = tensorflow_versions[0] + self._file_version = file_versions[0] + if len(metadata_paths) == 1: + # Fast path for a common case (only one DebugEvent file set.) + return metadata_paths + + num_no_id = len([run_id for run_id in run_ids if not run_id]) + if num_no_id: + paths_without_run_id = [ + metadata_path + for metadata_path, run_id in zip(metadata_paths, run_ids) + if not run_id + ] + raise ValueError( + "Found %d tfdbg metadata files and %d of them do not " + "have tfdbg run ids. The metadata files without run ids are: %s" % + (len(run_ids), num_no_id, paths_without_run_id)) + elif len(set(run_ids)) != 1: + raise ValueError( + "Unexpected: Found multiple (%d) tfdbg2 runs in directory %s" % + (len(set(run_ids)), self._dump_root)) + # Return the metadata files in ascending order of their timestamps. + paths_and_timestamps = sorted( + zip(metadata_paths, wall_times), key=lambda t: t[1]) + self._starting_wall_time = paths_and_timestamps[0][1] + return [path[0] for path in paths_and_timestamps] + + def starting_wall_time(self): + """Get the starting timestamp of the instrumented TensorFlow program. + + When there are multiple hosts (i.e., multiple tfdbg file sets), the earliest + timestamp among the file sets is returned. It is assumed to be the job that + starts first (e.g., the coordinator). + + Returns: + Starting timestamp in seconds since the epoch, as a float. + """ + return self._starting_wall_time + + def tfdbg_run_id(self): + """Get the run ID of the instrumented TensorFlow program.""" + return self._tfdbg_run_id + + def tensorflow_version(self): + """Get the version string of TensorFlow that the debugged program ran on.""" + return self._tensorflow_version + + def tfdbg_file_version(self): + """Get the tfdbg file format version.""" + return self._file_version + def __enter__(self): return self @@ -139,9 +233,6 @@ class DebugEventsReader(object): self._reader_offsets[file_path] = 0 return self._readers[file_path] - def metadata_iterator(self): - return self._generic_iterator(self._metadata_path) - def source_files_iterator(self): return self._generic_iterator(self._source_files_path) @@ -193,14 +284,18 @@ class DebugEventsReader(object): proto_string = self._get_reader(self._execution_path).read(offset)[0] return debug_event_pb2.DebugEvent.FromString(proto_string) - def graph_execution_traces_iterator(self): - return self._generic_iterator(self._graph_execution_traces_path) + def graph_execution_traces_iterators(self): + return [ + self._generic_iterator(path) + for path in self._graph_execution_traces_paths + ] - def read_graph_execution_traces_event(self, offset): - """Read DebugEvent at given offset from .graph_execution_traces file. + def read_graph_execution_traces_event(self, locator): + """Read DebugEvent at given offset from given .graph_execution_traces file. Args: - offset: Offset to read the DebugEvent proto from. + locator: A (file_index, offset) tuple that locates the DebugEvent + containing the graph execution trace. Returns: A DebugEventProto. @@ -209,9 +304,11 @@ class DebugEventsReader(object): `errors.DataLossError` if offset is at a wrong location. `IndexError` if offset is out of range of the file. """ - with self._reader_read_locks[self._graph_execution_traces_path]: - proto_string = self._get_reader( - self._graph_execution_traces_path).read(offset)[0] + file_index, offset = locator + graph_execution_traces_path = self._graph_execution_traces_paths[file_index] + with self._reader_read_locks[graph_execution_traces_path]: + proto_string = self._get_reader(graph_execution_traces_path).read( + offset)[0] return debug_event_pb2.DebugEvent.FromString(proto_string) def close(self): @@ -227,21 +324,27 @@ class BaseDigest(object): Properties: wall_time: A timestamp for the digest as a `float` (unit: s). - offset: A offset number in the corresponding file that can be used for - fast random read access. + locator: A datum that allows tracng the digest to its original + location. It can be either of the two: + 1. Bytes offset from the beginning of the file as a single integer, + for the case of all digests of the same kind coming from the same + file. + 2. A tuple of a file index and a byte offset. This applies to case + in which the same type of debugger data may come from multple files, + e.g., graph execution traces. """ - def __init__(self, wall_time, offset): + def __init__(self, wall_time, locator): self._wall_time = wall_time - self._offset = offset + self._locator = locator @property def wall_time(self): return self._wall_time @property - def offset(self): - return self._offset + def locator(self): + return self._locator def to_json(self): return {"wall_time": self.wall_time} @@ -265,10 +368,10 @@ class ExecutionDigest(BaseDigest): def __init__(self, wall_time, - offset, + locator, op_type, output_tensor_device_ids=None): - super(ExecutionDigest, self).__init__(wall_time, offset) + super(ExecutionDigest, self).__init__(wall_time, locator) self._op_type = op_type self._output_tensor_device_ids = _tuple_or_none(output_tensor_device_ids) @@ -332,7 +435,7 @@ class Execution(ExecutionDigest): debug_tensor_values=None): super(Execution, self).__init__( execution_digest.wall_time, - execution_digest.offset, + execution_digest.locator, execution_digest.op_type, output_tensor_device_ids=execution_digest.output_tensor_device_ids) self._host_name = host_name @@ -556,7 +659,7 @@ class GraphOpCreationDigest(BaseDigest): def __init__(self, wall_time, - offset, + locator, graph_id, op_type, op_name, @@ -565,7 +668,7 @@ class GraphOpCreationDigest(BaseDigest): stack_frame_ids, input_names=None, device_name=None): - super(GraphOpCreationDigest, self).__init__(wall_time, offset) + super(GraphOpCreationDigest, self).__init__(wall_time, locator) self._graph_id = graph_id self._op_type = op_type self._op_name = op_name @@ -640,14 +743,9 @@ class GraphExecutionTraceDigest(BaseDigest): graph. """ - def __init__(self, - wall_time, - offset, - op_type, - op_name, - output_slot, + def __init__(self, wall_time, locator, op_type, op_name, output_slot, graph_id): - super(GraphExecutionTraceDigest, self).__init__(wall_time, offset) + super(GraphExecutionTraceDigest, self).__init__(wall_time, locator) self._op_type = op_type self._op_name = op_name self._output_slot = output_slot @@ -701,13 +799,13 @@ class GraphExecutionTrace(GraphExecutionTraceDigest): tensor_debug_mode, debug_tensor_value=None, device_name=None): - super(GraphExecutionTrace, self).__init__( - graph_execution_trace_digest.wall_time, - graph_execution_trace_digest.offset, - graph_execution_trace_digest.op_type, - graph_execution_trace_digest.op_name, - graph_execution_trace_digest.output_slot, - graph_execution_trace_digest.graph_id) + super(GraphExecutionTrace, + self).__init__(graph_execution_trace_digest.wall_time, + graph_execution_trace_digest.locator, + graph_execution_trace_digest.op_type, + graph_execution_trace_digest.op_name, + graph_execution_trace_digest.output_slot, + graph_execution_trace_digest.graph_id) self._graph_ids = tuple(graph_ids) self._tensor_debug_mode = tensor_debug_mode self._debug_tensor_value = debug_tensor_value @@ -780,17 +878,17 @@ def _parse_tensor_value(tensor_proto, return_list=False): return None -def _execution_digest_from_debug_event_proto(debug_event, offset): +def _execution_digest_from_debug_event_proto(debug_event, locator): """Convert a DebugEvent proto into an ExecutionDigest data object.""" return ExecutionDigest( debug_event.wall_time, - offset, + locator, debug_event.execution.op_type, - output_tensor_device_ids=( - debug_event.execution.output_tensor_device_ids or None)) + output_tensor_device_ids=(debug_event.execution.output_tensor_device_ids + or None)) -def _execution_from_debug_event_proto(debug_event, offset): +def _execution_from_debug_event_proto(debug_event, locator): """Convert a DebugEvent proto into an Execution data object.""" execution_proto = debug_event.execution @@ -806,7 +904,7 @@ def _execution_from_debug_event_proto(debug_event, offset): debug_tensor_values.append( _parse_tensor_value(tensor_proto, return_list=True)) return Execution( - _execution_digest_from_debug_event_proto(debug_event, offset), + _execution_digest_from_debug_event_proto(debug_event, locator), execution_proto.code_location.host_name, tuple(execution_proto.code_location.stack_frame_ids), execution_proto.tensor_debug_mode, @@ -832,7 +930,6 @@ class DebugDataReader(object): def __init__(self, dump_root): self._reader = DebugEventsReader(dump_root) - self._load_metadata() # TODO(cais): Implement pagination for memory constraints. self._execution_digests = [] @@ -858,13 +955,6 @@ class DebugDataReader(object): def _add_monitor(self, monitor): self._monitors.append(monitor) - def _load_metadata(self): - metadata_iter = self._reader.metadata_iterator() - debug_event = next(metadata_iter).debug_event - self._starting_wall_time = debug_event.wall_time - self._tensorflow_version = debug_event.debug_metadata.tensorflow_version - self._tfdbg_run_id = debug_event.debug_metadata.tfdbg_run_id - def _load_source_files(self): """Incrementally read the .source_files DebugEvent file.""" source_files_iter = self._reader.source_files_iterator() @@ -944,37 +1034,32 @@ class DebugDataReader(object): def _load_graph_execution_traces(self): """Incrementally load the .graph_execution_traces file.""" - traces_iter = self._reader.graph_execution_traces_iterator() - for debug_event, offset in traces_iter: - self._graph_execution_trace_digests.append( - self._graph_execution_trace_digest_from_debug_event_proto( - debug_event, offset)) - if self._monitors: - graph_execution_trace = ( - self._graph_execution_trace_from_debug_event_proto( - debug_event, offset)) - for monitor in self._monitors: - monitor.on_graph_execution_trace( - len(self._graph_execution_trace_digests) - 1, - graph_execution_trace) + for i, traces_iter in enumerate( + self._reader.graph_execution_traces_iterators()): + for debug_event, offset in traces_iter: + self._graph_execution_trace_digests.append( + self._graph_execution_trace_digest_from_debug_event_proto( + debug_event, (i, offset))) + if self._monitors: + graph_execution_trace = ( + self._graph_execution_trace_from_debug_event_proto( + debug_event, (i, offset))) + for monitor in self._monitors: + monitor.on_graph_execution_trace( + len(self._graph_execution_trace_digests) - 1, + graph_execution_trace) - def _graph_execution_trace_digest_from_debug_event_proto(self, - debug_event, - offset): + def _graph_execution_trace_digest_from_debug_event_proto( + self, debug_event, locator): trace_proto = debug_event.graph_execution_trace op_name = trace_proto.op_name op_type = self._lookup_op_type(trace_proto.tfdbg_context_id, op_name) return GraphExecutionTraceDigest( - debug_event.wall_time, - offset, - op_type, - op_name, + debug_event.wall_time, locator, op_type, op_name, trace_proto.output_slot, debug_event.graph_execution_trace.tfdbg_context_id) - def _graph_execution_trace_from_debug_event_proto(self, - debug_event, - offset): + def _graph_execution_trace_from_debug_event_proto(self, debug_event, locator): """Convert a DebugEvent proto into a GraphExecutionTrace data object.""" trace_proto = debug_event.graph_execution_trace graph_ids = [trace_proto.tfdbg_context_id] @@ -995,7 +1080,7 @@ class DebugDataReader(object): trace_proto.tensor_proto, return_list=True) return GraphExecutionTrace( self._graph_execution_trace_digest_from_debug_event_proto( - debug_event, offset), + debug_event, locator), graph_ids=graph_ids, tensor_debug_mode=trace_proto.tensor_debug_mode, debug_tensor_value=debug_tensor_value, @@ -1059,7 +1144,7 @@ class DebugDataReader(object): Returns: Stating wall time as seconds since the epoch, as a `float`. """ - return self._starting_wall_time + return self._reader.starting_wall_time() def tensorflow_version(self): """TensorFlow version used in the debugged TensorFlow program. @@ -1070,11 +1155,11 @@ class DebugDataReader(object): Returns: TensorFlow version used by the debugged program, as a `str`. """ - return self._tensorflow_version + return self._reader.tensorflow_version() def tfdbg_run_id(self): """Get the debugger run ID of the debugged TensorFlow program.""" - return self._tfdbg_run_id + return self._reader.tfdbg_run_id() def outermost_graphs(self): """Get the number of outer most graphs read so far.""" @@ -1171,9 +1256,9 @@ class DebugDataReader(object): def read_execution(self, execution_digest): """Read a detailed Execution object.""" - debug_event = self._reader.read_execution_event(execution_digest.offset) - return _execution_from_debug_event_proto( - debug_event, execution_digest.offset) + debug_event = self._reader.read_execution_event(execution_digest.locator) + return _execution_from_debug_event_proto(debug_event, + execution_digest.locator) def read_graph_execution_trace(self, graph_execution_trace_digest): """Read the detailed graph execution trace. @@ -1185,9 +1270,9 @@ class DebugDataReader(object): The corresponding `GraphExecutionTrace` object. """ debug_event = self._reader.read_graph_execution_traces_event( - graph_execution_trace_digest.offset) + graph_execution_trace_digest.locator) return self._graph_execution_trace_from_debug_event_proto( - debug_event, graph_execution_trace_digest.offset) + debug_event, graph_execution_trace_digest.locator) def read_execution_stack_trace(self, execution): """Read the stack trace of a given Execution object. @@ -1234,7 +1319,7 @@ class DebugDataReader(object): A list of numpy arrays representing the output tensor values of the execution event. """ - debug_event = self._reader.read_execution_event(execution.offset) + debug_event = self._reader.read_execution_event(execution.locator) return [_parse_tensor_value(tensor_proto) for tensor_proto in debug_event.execution.tensor_protos] @@ -1248,8 +1333,7 @@ class DebugDataReader(object): A numpy array representing the output tensor value of the intra-graph tensor execution event. """ - debug_event = self._reader.read_graph_execution_traces_event( - trace.offset) + debug_event = self._reader.read_graph_execution_traces_event(trace.locator) return _parse_tensor_value(debug_event.graph_execution_trace.tensor_proto) def symbolic_tensor_id(self, graph_id, op_name, output_slot): diff --git a/tensorflow/python/debug/lib/debug_events_writer_test.py b/tensorflow/python/debug/lib/debug_events_writer_test.py index 7b06bf772be..3f3f9179e5d 100644 --- a/tensorflow/python/debug/lib/debug_events_writer_test.py +++ b/tensorflow/python/debug/lib/debug_events_writer_test.py @@ -21,6 +21,7 @@ from __future__ import print_function import glob import json as json_lib import os +import re import threading import time @@ -264,14 +265,14 @@ class DebugEventsWriterTest(dumping_callback_test_lib.DumpingCallbackTestBase, writer.WriteGraphExecutionTrace(trace) with debug_events_reader.DebugEventsReader(self.dump_root) as reader: - actuals = list(reader.graph_execution_traces_iterator()) + actuals = list(reader.graph_execution_traces_iterators()[0]) # Before FlushExecutionFiles() is called. No data should have been written # to the file. self.assertEmpty(actuals) writer.FlushExecutionFiles() actuals = list(item.debug_event.graph_execution_trace - for item in reader.graph_execution_traces_iterator()) + for item in reader.graph_execution_traces_iterators()[0]) self.assertLen(actuals, debug_events_writer.DEFAULT_CIRCULAR_BUFFER_SIZE) for i in range(debug_events_writer.DEFAULT_CIRCULAR_BUFFER_SIZE): self.assertEqual( @@ -291,7 +292,7 @@ class DebugEventsWriterTest(dumping_callback_test_lib.DumpingCallbackTestBase, with debug_events_reader.DebugEventsReader(self.dump_root) as reader: actuals = list(item.debug_event.graph_execution_trace - for item in reader.graph_execution_traces_iterator()) + for item in reader.graph_execution_traces_iterators()[0]) self.assertLen(actuals, num_execution_events) for i in range(num_execution_events): self.assertEqual(actuals[i].op_name, "Op%d" % i) @@ -598,6 +599,86 @@ class DebugEventsWriterTest(dumping_callback_test_lib.DumpingCallbackTestBase, self.assertEqual(traces[-1].op_name, "Op_%d" % (expected_end - 1)) +class MultiSetReaderTest(dumping_callback_test_lib.DumpingCallbackTestBase): + """Test for DebugDataReader for multiple file sets under a dump root.""" + + def testReadingTwoFileSetsWithTheSameDumpRootSucceeds(self): + # To simulate a multi-host data dump, we first generate file sets in two + # different directories, with the same tfdbg_run_id, and then combine them. + tfdbg_run_id = "foo" + for i in range(2): + writer = debug_events_writer.DebugEventsWriter( + os.path.join(self.dump_root, str(i)), + tfdbg_run_id, + circular_buffer_size=-1) + if i == 0: + debugged_graph = debug_event_pb2.DebuggedGraph( + graph_id="graph1", graph_name="graph1") + writer.WriteDebuggedGraph(debugged_graph) + op_name = "Op_0" + graph_op_creation = debug_event_pb2.GraphOpCreation( + op_type="FooOp", op_name=op_name, graph_id="graph1") + writer.WriteGraphOpCreation(graph_op_creation) + op_name = "Op_1" + graph_op_creation = debug_event_pb2.GraphOpCreation( + op_type="FooOp", op_name=op_name, graph_id="graph1") + writer.WriteGraphOpCreation(graph_op_creation) + for _ in range(10): + trace = debug_event_pb2.GraphExecutionTrace( + op_name="Op_%d" % i, tfdbg_context_id="graph1") + writer.WriteGraphExecutionTrace(trace) + writer.FlushNonExecutionFiles() + writer.FlushExecutionFiles() + + # Move all files from the subdirectory /1 to subdirectory /0. + dump_root_0 = os.path.join(self.dump_root, "0") + src_paths = glob.glob(os.path.join(self.dump_root, "1", "*")) + for src_path in src_paths: + dst_path = os.path.join( + dump_root_0, + # Rename the file set to avoid file name collision. + re.sub(r"(tfdbg_events\.\d+)", r"\g<1>1", os.path.basename(src_path))) + os.rename(src_path, dst_path) + + with debug_events_reader.DebugDataReader(dump_root_0) as reader: + reader.update() + # Verify the content of the .graph_execution_traces file. + trace_digests = reader.graph_execution_traces(digest=True) + self.assertLen(trace_digests, 20) + for _ in range(10): + trace = reader.read_graph_execution_trace(trace_digests[i]) + self.assertEqual(trace.op_name, "Op_0") + for _ in range(10): + trace = reader.read_graph_execution_trace(trace_digests[i + 10]) + self.assertEqual(trace.op_name, "Op_1") + + def testReadingTwoFileSetsWithTheDifferentRootsLeadsToError(self): + # To simulate a multi-host data dump, we first generate file sets in two + # different directories, with different tfdbg_run_ids, and then combine + # them. + for i in range(2): + writer = debug_events_writer.DebugEventsWriter( + os.path.join(self.dump_root, str(i)), + "run_id_%d" % i, + circular_buffer_size=-1) + writer.FlushNonExecutionFiles() + writer.FlushExecutionFiles() + + # Move all files from the subdirectory /1 to subdirectory /0. + dump_root_0 = os.path.join(self.dump_root, "0") + src_paths = glob.glob(os.path.join(self.dump_root, "1", "*")) + for src_path in src_paths: + dst_path = os.path.join( + dump_root_0, + # Rename the file set to avoid file name collision. + re.sub(r"(tfdbg_events\.\d+)", r"\g<1>1", os.path.basename(src_path))) + os.rename(src_path, dst_path) + + with self.assertRaisesRegexp(ValueError, + r"Found multiple \(2\) tfdbg2 runs"): + debug_events_reader.DebugDataReader(dump_root_0) + + class DataObjectsTest(test_util.TensorFlowTestCase, parameterized.TestCase): def jsonRoundTripCheck(self, obj): diff --git a/tensorflow/python/debug/lib/debug_v2_ops_test.py b/tensorflow/python/debug/lib/debug_v2_ops_test.py index d715869f359..d70c505d3fc 100644 --- a/tensorflow/python/debug/lib/debug_v2_ops_test.py +++ b/tensorflow/python/debug/lib/debug_v2_ops_test.py @@ -92,16 +92,13 @@ class DebugIdentityV2OpTest(dumping_callback_test_lib.DumpingCallbackTestBase): write_debug_trace(x), [9.0 + np.sqrt(3.0), 16.0 + 2.0]) with debug_events_reader.DebugEventsReader(self.dump_root) as reader: - metadata_iter = reader.metadata_iterator() # Check that the .metadata DebugEvents data file has been created, even # before FlushExecutionFiles() is called. - debug_event = next(metadata_iter).debug_event - self.assertGreater(debug_event.wall_time, 0) - self.assertTrue(debug_event.debug_metadata.tensorflow_version) - self.assertTrue( - debug_event.debug_metadata.file_version.startswith("debug.Event:")) + self.assertGreater(reader.starting_wall_time(), 0) + self.assertTrue(reader.tensorflow_version()) + self.assertTrue(reader.tfdbg_file_version().startswith("debug.Event")) - graph_trace_iter = reader.graph_execution_traces_iterator() + graph_trace_iter = reader.graph_execution_traces_iterators()[0] # Before FlushExecutionFiles() is called, the .graph_execution_traces file # ought to be empty. with self.assertRaises(StopIteration): @@ -109,7 +106,7 @@ class DebugIdentityV2OpTest(dumping_callback_test_lib.DumpingCallbackTestBase): # Flush the circular buffer. self.writer.FlushExecutionFiles() - graph_trace_iter = reader.graph_execution_traces_iterator() + graph_trace_iter = reader.graph_execution_traces_iterators()[0] # The circular buffer has a size of 4. So only the data from the # last two iterations should have been written to self.dump_root. @@ -167,7 +164,7 @@ class DebugIdentityV2OpTest(dumping_callback_test_lib.DumpingCallbackTestBase): self.writer.FlushExecutionFiles() with debug_events_reader.DebugEventsReader(self.dump_root) as reader: - graph_trace_iter = reader.graph_execution_traces_iterator() + graph_trace_iter = reader.graph_execution_traces_iterators()[0] try: x_values = [] timestamp = 0 @@ -216,7 +213,7 @@ class DebugIdentityV2OpTest(dumping_callback_test_lib.DumpingCallbackTestBase): for debug_root in (self.dump_root, another_dump_root): with debug_events_reader.DebugEventsReader(debug_root) as reader: - graph_trace_iter = reader.graph_execution_traces_iterator() + graph_trace_iter = reader.graph_execution_traces_iterators()[0] debug_event = next(graph_trace_iter).debug_event trace = debug_event.graph_execution_trace @@ -272,7 +269,7 @@ class DebugIdentityV2OpUninitializedWriterTest( writer.FlushExecutionFiles() with debug_events_reader.DebugEventsReader(self.dump_root) as reader: - graph_trace_iter = reader.graph_execution_traces_iterator() + graph_trace_iter = reader.graph_execution_traces_iterators()[0] graph_execution_traces = [] while True: try: diff --git a/tensorflow/python/debug/lib/dumping_callback_test_lib.py b/tensorflow/python/debug/lib/dumping_callback_test_lib.py index 05bf3aeb6da..e58ffdbd79f 100644 --- a/tensorflow/python/debug/lib/dumping_callback_test_lib.py +++ b/tensorflow/python/debug/lib/dumping_callback_test_lib.py @@ -48,7 +48,6 @@ class DumpingCallbackTestBase(test_util.TensorFlowTestCase): def _readAndCheckMetadataFile(self): """Read and check the .metadata debug-events file.""" with debug_events_reader.DebugEventsReader(self.dump_root) as reader: - metadata_iter = reader.metadata_iterator() - metadata = next(metadata_iter).debug_event.debug_metadata - self.assertEqual(metadata.tensorflow_version, versions.__version__) - self.assertTrue(metadata.file_version.startswith("debug.Event")) + self.assertTrue(reader.tfdbg_run_id()) + self.assertEqual(reader.tensorflow_version(), versions.__version__) + self.assertTrue(reader.tfdbg_file_version().startswith("debug.Event")) From c692c45daef9e4a1d5937f705fe8a3977b04fddd Mon Sep 17 00:00:00 2001 From: "A. Unique TensorFlower" Date: Mon, 22 Jun 2020 17:10:15 -0700 Subject: [PATCH 507/867] tf numpy: some changes to ndarray constructor logic. PiperOrigin-RevId: 317765968 Change-Id: Iea4338ad18707ff36fc49b450d0defad5c13a6a2 --- tensorflow/python/ops/numpy_ops/np_arrays.py | 7 +++---- tensorflow/python/ops/numpy_ops/np_arrays_test.py | 14 ++++++++++++++ 2 files changed, 17 insertions(+), 4 deletions(-) diff --git a/tensorflow/python/ops/numpy_ops/np_arrays.py b/tensorflow/python/ops/numpy_ops/np_arrays.py index eca84421d1b..77157544e8f 100644 --- a/tensorflow/python/ops/numpy_ops/np_arrays.py +++ b/tensorflow/python/ops/numpy_ops/np_arrays.py @@ -141,13 +141,12 @@ class ndarray(composite_tensor.CompositeTensor): # pylint: disable=invalid-name raise ValueError('Unexpected type for `buffer` {}. Must be an ndarray,' ' Tensor or np.ndarray.'.format(type(buffer))) - if shape is not None and tuple(shape) != buffer._shape_tuple(): # pylint: disable=protected-access - # TODO(srbs): NumPy allows this. Investigate if/how to support this. - raise ValueError('shape arg must match buffer.shape.') + if shape is not None: + buffer.set_shape(shape) assert isinstance(buffer, ops.Tensor) if dtype and dtype != buffer.dtype: - buffer = array_ops.bitcast(buffer, dtype) + buffer = math_ops.cast(buffer, dtype) self._data = buffer self._type_spec_internal = None diff --git a/tensorflow/python/ops/numpy_ops/np_arrays_test.py b/tensorflow/python/ops/numpy_ops/np_arrays_test.py index 412addc0ad7..ab407d2bfcf 100644 --- a/tensorflow/python/ops/numpy_ops/np_arrays_test.py +++ b/tensorflow/python/ops/numpy_ops/np_arrays_test.py @@ -22,6 +22,7 @@ import collections import numpy as np +from tensorflow.python.framework import constant_op from tensorflow.python.framework import dtypes from tensorflow.python.framework import ops from tensorflow.python.ops import array_ops @@ -51,6 +52,19 @@ class ArrayTest(test.TestCase): self.assertIs(a.dtype.type, np.bool_) self.assertAllEqual([False, True], a) + def testConstructor(self): + t = constant_op.constant([[1], [1]]) + a = np_arrays.ndarray(shape=(2, 1), buffer=t) + self.assertAllEqual(t, a) + self.assertEqual(dtypes.float64, a.dtype) + + a = np_arrays.ndarray(shape=(2, 1), dtype=dtypes.int32, buffer=t) + self.assertAllEqual(t, a) + self.assertEqual(dtypes.int32, a.dtype) + + with self.assertRaises(ValueError): # bad shape + _ = np_arrays.ndarray((2, 2), buffer=t) + def testNeg(self): a = t2a(ops.convert_to_tensor(value=[1.0, 2.0])) self.assertAllEqual([-1.0, -2.0], -a) From fe6e64b09853ac9dbb234ce9b7d4b10da11c7fe9 Mon Sep 17 00:00:00 2001 From: Xiao Yu Date: Mon, 22 Jun 2020 17:15:42 -0700 Subject: [PATCH 508/867] Refactor eager placement logic into three util methods: - MaybePinSmallOpsToCpu - MaybePinToResourceDevice - MaybePinToCustomDevice We are going to reuse MaybePinSmallOpsToCpu in TFRT but not the other two. Because TFRT doesn't have native Resource neither Custom Device. PiperOrigin-RevId: 317766813 Change-Id: I43241b5786120ddf39dc4bfff6071239afdfd785 --- tensorflow/core/common_runtime/eager/BUILD | 31 +++ .../core/common_runtime/eager/context.h | 2 +- tensorflow/core/common_runtime/eager/core.cc | 22 ++ .../common_runtime/eager/eager_operation.h | 2 +- .../core/common_runtime/eager/execute.cc | 169 ------------- .../common_runtime/eager/placement_utils.cc | 228 ++++++++++++++++++ .../common_runtime/eager/placement_utils.h | 55 +++++ .../eager/eager_service_impl.cc | 6 +- tensorflow/lite/delegates/flex/kernel.cc | 9 +- 9 files changed, 349 insertions(+), 175 deletions(-) create mode 100644 tensorflow/core/common_runtime/eager/placement_utils.cc create mode 100644 tensorflow/core/common_runtime/eager/placement_utils.h diff --git a/tensorflow/core/common_runtime/eager/BUILD b/tensorflow/core/common_runtime/eager/BUILD index fb69bcb7ab5..911b59eed17 100644 --- a/tensorflow/core/common_runtime/eager/BUILD +++ b/tensorflow/core/common_runtime/eager/BUILD @@ -29,6 +29,7 @@ tf_cuda_library( ":context", ":eager_operation", ":execute", + ":placement_utils", ":tensor_handle", "//tensorflow/c:c_api_internal", "//tensorflow/c:tf_tensor_internal", @@ -489,6 +490,7 @@ cc_library( ":eager_op_rewrite_registry", ":eager_operation", ":kernel_and_device", + ":placement_utils", ":tensor_handle", "@com_google_absl//absl/container:flat_hash_map", "@com_google_absl//absl/container:inlined_vector", @@ -521,6 +523,35 @@ cc_library( }), ) +tf_cuda_library( + name = "placement_utils", + srcs = [ + "placement_utils.cc", + ], + hdrs = [ + "placement_utils.h", + ], + visibility = ["//tensorflow:internal"], + deps = [ + ":context", + ":attr_builder", + ":eager_operation", + "//tensorflow/c/eager:immediate_execution_tensor_handle", + ] + select({ + "//tensorflow:android": [ + "//tensorflow/core:portable_tensorflow_lib_lite", + ], + "//conditions:default": [ + "//tensorflow/core:core_cpu_lib", + "//tensorflow/core:framework", + "//tensorflow/core:framework_internal", + "//tensorflow/core:lib", + "//tensorflow/core:lib_internal", + "//tensorflow/core:protos_all_cc", + ], + }), +) + tf_cuda_library( name = "attr_builder", srcs = ["attr_builder.cc"], diff --git a/tensorflow/core/common_runtime/eager/context.h b/tensorflow/core/common_runtime/eager/context.h index c16e1f0f4ad..e6769279558 100644 --- a/tensorflow/core/common_runtime/eager/context.h +++ b/tensorflow/core/common_runtime/eager/context.h @@ -478,7 +478,7 @@ class EagerContext : public ImmediateExecutionContext, public core::RefCounted { // On mobile, it just cleans the caches. void WaitForAndCloseRemoteContexts(); - bool PinSmallOpsToCPU() { return pin_small_ops_to_cpu_; } + bool PinSmallOpsToCPU() const { return pin_small_ops_to_cpu_; } tensorflow::Env* TFEnv() const { return env_; } diff --git a/tensorflow/core/common_runtime/eager/core.cc b/tensorflow/core/common_runtime/eager/core.cc index 3d37250a4fe..77d2b665f5e 100644 --- a/tensorflow/core/common_runtime/eager/core.cc +++ b/tensorflow/core/common_runtime/eager/core.cc @@ -18,6 +18,7 @@ limitations under the License. #include "tensorflow/core/common_runtime/eager/context.h" #include "tensorflow/core/common_runtime/eager/eager_operation.h" #include "tensorflow/core/common_runtime/eager/execute.h" +#include "tensorflow/core/common_runtime/eager/placement_utils.h" #include "tensorflow/core/common_runtime/eager/tensor_handle.h" #include "tensorflow/core/platform/errors.h" @@ -187,6 +188,27 @@ Status EagerContext::RegisterFunction(AbstractFunction* f) { // eager_operation.cc we can avoid a circular dependency between them. Status EagerOperation::Execute(absl::Span retvals, int* num_retvals) { + // Run eager placement logic. + VariantDevice device; + TF_RETURN_IF_ERROR(eager::MaybePinToCustomDevice(&device, *this)); + if (device == kVariantDeviceNull) { + TF_RETURN_IF_ERROR(eager::MaybePinToResourceDevice(&device, *this)); + } + if (device == kVariantDeviceNull) { + bool pin_to_cpu; + TF_RETURN_IF_ERROR(eager::MaybePinSmallOpsToCpu( + &pin_to_cpu, op_name(), + absl::MakeSpan( + reinterpret_cast(inputs_.data()), + inputs_.size()), + ctx_)); + if (pin_to_cpu) { + device = ctx_.HostCPU(); + } + } + if (device != kVariantDeviceNull) { + SetDevice(device); + } return EagerExecute( this, reinterpret_cast(retvals.data()), num_retvals); diff --git a/tensorflow/core/common_runtime/eager/eager_operation.h b/tensorflow/core/common_runtime/eager/eager_operation.h index 963aed25733..fa245649d5c 100644 --- a/tensorflow/core/common_runtime/eager/eager_operation.h +++ b/tensorflow/core/common_runtime/eager/eager_operation.h @@ -126,7 +126,7 @@ class EagerOperation : public ImmediateExecutionOperation { bool is_function() const { return is_function_; } bool colocation_exempt() const { return colocation_exempt_; } - tensorflow::EagerContext& EagerContext() { return ctx_; } + tensorflow::EagerContext& EagerContext() const { return ctx_; } AttrBuilder* MutableAttrs() { return &attrs_; } const AttrBuilder& Attrs() const { return attrs_; } diff --git a/tensorflow/core/common_runtime/eager/execute.cc b/tensorflow/core/common_runtime/eager/execute.cc index a94c882b3b3..a030f4d0356 100644 --- a/tensorflow/core/common_runtime/eager/execute.cc +++ b/tensorflow/core/common_runtime/eager/execute.cc @@ -870,173 +870,6 @@ Status EagerRemoteExecute(EagerOperation* op, TensorHandle** retvals, } #endif // IS_MOBILE_PLATFORM -// These ops are not pinnable since they generate data. It can be slower to -// generate and then copy the data instead of just generating the data on the -// device directly. -bool IsPinnableOp(const string& op_type) { - static const gtl::FlatSet* unpinnable_ops = new gtl::FlatSet({ - "RandomUniform", - "RandomUniformInt", - "RandomStandardNormal", - "StatelessRandomUniform", - "StatelessRandomUniformInt", - "StatelessRandomUniformFullInt", - "StatelessRandomNormal", - }); - - // XRT ops refer to per-device handles that are not safe to move between - // devices. - return unpinnable_ops->find(op_type) == unpinnable_ops->end() && - !absl::StartsWith(op_type, "XRT"); -} - -// Validate if the remote device with the given incarnation is valid in the -// remote device manager of the current eager context. -Status ValidateTensorHandleRemoteDevice(EagerContext* ctx, - int64 device_incarnation) { - if (ctx->remote_device_mgr()->ContainsDevice(device_incarnation)) { - return Status::OK(); - } - return errors::InvalidArgument( - "Resource input tensor contains an invalid device. This might happen " - "when the client has connected to a different cluster, or some remote " - "workers have been restarted."); -} - -// The Op device may be updated if: -// - A resource touching input is specified: all resource-touching ops run in -// the device the resource is, regardless of anything else that has been -// specified. This is identical to the graph mode behavior. -// -// - All op inputs are on the CPU, small (<64 elements) and integers -// (int32/int64). This can be disabled by setting the environment variable -// "TF_EAGER_ENABLE_SMALL_TENSOR_CPU_PINNING" to "0" or "false". -// -// TODO(b/154234908): Unify placement logic. -Status MaybeUpdateOpDevice(EagerOperation* op) { - // If operation was already placed on a custom device, use it. - if (VariantDeviceIsCustom(op->Device())) { - return Status::OK(); - } - - // If all the inputs are on the same custom device, use that custom - // device. Otherwise, it is an error to have a custom device as an input. - if (!op->Inputs().empty()) { - // We keep track of what we've seen with devices instead of booleans to be - // able to provide a meaningful error message below. - VariantDevice first = op->Inputs()[0]->device(); - VariantDevice different = first; // A different input device, if any. - VariantDevice custom = first; // The first custom device seen, or an - // arbitrary non-custom device otherwise. - for (size_t i = 1; first == different && i < op->Inputs().size(); ++i) { - VariantDevice device = op->Inputs()[i]->device(); - if (device != first) { - different = device; - } - if (!VariantDeviceIsCustom(custom) && VariantDeviceIsCustom(device)) { - custom = device; - } - if (different != first && VariantDeviceIsCustom(custom)) { - return errors::InvalidArgument(absl::StrCat( - "If an operation has one of its inputs in a custom device, then " - "all inputs should be on that same device. Operation ", - op->Name(), " has one input in custom device ", - VariantDeviceName(custom), - " and at least one input in a different device ", - VariantDeviceName(custom == first ? different : first))); - } - } - if (different == first && VariantDeviceIsCustom(custom)) { - op->SetDevice(first); - return Status::OK(); - } - } - - if (op->colocation_exempt()) { - return Status::OK(); - } - EagerContext& ctx = op->EagerContext(); - bool all_inputs_eligible_for_cpu_pinning = - ctx.PinSmallOpsToCPU() && !op->is_function() && IsPinnableOp(op->Name()); - Device* op_device = op->Device() == kVariantDeviceNull - ? ctx.HostCPU() - : absl::get(op->Device()); - for (int i = 0; i < op->Inputs().size(); ++i) { - TensorHandle* tensor_handle = op->Inputs()[i]; - if (tensor_handle->dtype == DT_RESOURCE) { - if (tensor_handle->resource_remote_device_incarnation() != 0) { - TF_RETURN_IF_ERROR(ValidateTensorHandleRemoteDevice( - &ctx, tensor_handle->resource_remote_device_incarnation())); - } - Device* resource_device = tensor_handle->resource_device(); - DVLOG(2) << "for op " << op->Name() << " input " << i << " " - << DataTypeString(tensor_handle->dtype) - << " input device = " << resource_device->name() - << ", op device = " << op_device->name(); - // We check for `op->Device() == nullptr` because it can be later - // interpreted as unspecified device and a different device can - // be selected based on device priority. If any input to an op - // is a resource we must pin it to prevent different device selection. - // TODO(iga): null device can mean "unspecified" or "CPU". Clean this up. - if (resource_device != op_device || op->Device() == kVariantDeviceNull) { - DVLOG(1) << (resource_device != op_device ? "Changing " : "Setting ") - << "device of operation " << op->Name() << " to " - << resource_device->name() << " because input #" << i - << " is a resource in this device."; - op->SetDevice(resource_device); - } - all_inputs_eligible_for_cpu_pinning = false; - // No point in looking at other inputs. If there are other resources, - // they must have the same device and we already declared the op to be - // ineligible for CPU pinning. - break; - } else if (all_inputs_eligible_for_cpu_pinning) { - auto input_device_variant = tensor_handle->DeviceOrHostCPU(ctx); - if (VariantDeviceIsCustom(input_device_variant)) { - all_inputs_eligible_for_cpu_pinning = false; - continue; - } - Device* input_device = absl::get(input_device_variant); - DVLOG(2) << "for op " << op->Name() << " input " << i << " " - << DataTypeString(tensor_handle->dtype) - << " input device = " << input_device->name() - << ", op device = " << op_device->name(); - - // Input is on CPU. - if (input_device != ctx.HostCPU()) { - all_inputs_eligible_for_cpu_pinning = false; - continue; - } - - if (tensor_handle->dtype != DataType::DT_INT32 && - tensor_handle->dtype != DataType::DT_INT64) { - all_inputs_eligible_for_cpu_pinning = false; - continue; - } - - int64 num_elements; - TF_RETURN_IF_ERROR(tensor_handle->NumElements(&num_elements)); - if (num_elements > 64) { - all_inputs_eligible_for_cpu_pinning = false; - } - } - } - - // Ops without inputs are usually ops that generate a tensor in some way and - // usually require being present on whatever device they are scheduled on - // - for e.g. VarHandleOp or _Recv). - // TODO(nareshmodi): Is it possible there is no int32/int64 CPU kernel for - // an op, but there is a GPU kernel? - if (!op->Inputs().empty() && all_inputs_eligible_for_cpu_pinning) { - DVLOG(1) << "Forcing op " << op->Name() - << " to be on the CPU since all input tensors have an " - "int32/int64 dtype, and are small (less than 64 elements)."; - op->SetDevice(ctx.HostCPU()); - } - - return Status::OK(); -} - Status GetKernelOutputs(std::vector* outputs, int num_outputs, TensorHandle** retvals, EagerContext* ctx, KernelAndDevice* kernel) { @@ -1099,8 +932,6 @@ Status EagerExecute(EagerOperation* op, TensorHandle** retvals, [&] { return absl::StrCat("EagerExecute: ", op->Name()); }, profiler::TraceMeLevel::kInfo); - TF_RETURN_IF_ERROR(MaybeUpdateOpDevice(op)); - if (VariantDeviceIsCustom(op->Device())) { return absl::get(op->Device()) ->Execute(op, retvals, num_retvals); diff --git a/tensorflow/core/common_runtime/eager/placement_utils.cc b/tensorflow/core/common_runtime/eager/placement_utils.cc new file mode 100644 index 00000000000..8898516612f --- /dev/null +++ b/tensorflow/core/common_runtime/eager/placement_utils.cc @@ -0,0 +1,228 @@ +/* Copyright 2020 The TensorFlow Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. +==============================================================================*/ + +#include "tensorflow/core/common_runtime/eager/placement_utils.h" + +#include "tensorflow/c/eager/immediate_execution_tensor_handle.h" +#include "tensorflow/core/common_runtime/eager/attr_builder.h" +#include "tensorflow/core/common_runtime/eager/eager_operation.h" +#include "tensorflow/core/common_runtime/input_colocation_exemption_registry.h" +#include "tensorflow/core/framework/op_def.pb.h" +#include "tensorflow/core/framework/types.pb.h" +#include "tensorflow/core/platform/errors.h" + +namespace tensorflow { +namespace eager { + +// These ops are not pinnable since they generate data. It can be slower to +// generate and then copy the data instead of just generating the data on the +// device directly. +static bool IsPinnableOp(StringPiece op_name) { + static const gtl::FlatSet* unpinnable_ops = new gtl::FlatSet({ + "RandomUniform", + "RandomUniformInt", + "RandomStandardNormal", + "StatelessRandomUniform", + "StatelessRandomUniformInt", + "StatelessRandomUniformFullInt", + "StatelessRandomNormal", + }); + + // XRT ops refer to per-device handles that are not safe to move between + // devices. + return unpinnable_ops->find(string(op_name)) == unpinnable_ops->end() && + !absl::StartsWith(op_name, "XRT"); +} +// Validate if the remote device with the given incarnation is valid in the +// remote device manager of the current eager context. +static Status ValidateTensorHandleRemoteDevice(EagerContext* ctx, + int64 device_incarnation) { + if (ctx->remote_device_mgr()->ContainsDevice(device_incarnation)) { + return Status::OK(); + } + return errors::InvalidArgument( + "Resource input tensor contains an invalid device. This might happen " + "when the client has connected to a different cluster, or some remote " + "workers have been restarted."); +} + +bool IsColocationExempt(StringPiece op_name) { + const auto& exempt_ops = InputColocationExemptionRegistry::Global()->Get(); + return exempt_ops.find(string(op_name)) != exempt_ops.end(); +} + +bool IsFunction(StringPiece op_name) { + const OpDef* op_def = nullptr; + Status s = OpDefForOp(string(op_name), &op_def); + if (!s.ok()) { + if (!errors::IsNotFound(s)) { + LOG(WARNING) << "Looking up OpDef failed with error: " << s.ToString(); + } + // Cannot find OpDef, it is a function. + return true; + } + return false; +} + +bool IsCustomDevice(StringPiece device_name, const EagerContext& ctx) { + CustomDevice* custom_device; + return ctx.FindCustomDeviceFromName(string(device_name), &custom_device).ok(); +} + +Status MaybePinSmallOpsToCpu(bool* result, StringPiece op_name, + absl::Span args, + const EagerContext& ctx) { + if (!ctx.PinSmallOpsToCPU() || IsFunction(op_name) || + IsColocationExempt(op_name) || !IsPinnableOp(op_name)) { + *result = false; + return Status::OK(); + } + + // Ops without inputs are usually ops that generate a tensor in some way and + // usually require being present on whatever device they are scheduled on + // - for e.g. VarHandleOp or _Recv). + if (args.empty()) { + *result = false; + return Status::OK(); + } + + int i = 0; + for (auto* arg : args) { + Status s; + const char* device_name = arg->DeviceName(&s); + DataType dtype = arg->DataType(); + TF_RETURN_IF_ERROR(s); + if (IsCustomDevice(device_name, ctx)) { + *result = false; + return Status::OK(); + } + + DVLOG(2) << "for op " << op_name << " input " << i << " " + << DataTypeString(dtype) << " input device = " << device_name; + + // Input is on CPU. + if (device_name != ctx.HostCPU()->name()) { + *result = false; + return Status::OK(); + } + + if (dtype != DataType::DT_INT32 && dtype != DataType::DT_INT64) { + *result = false; + return Status::OK(); + } + + int64 num_elements; + TF_RETURN_IF_ERROR(arg->NumElements(&num_elements)); + if (num_elements > 64) { + *result = false; + return Status::OK(); + } + i++; + } + + // TODO(nareshmodi): Is it possible there is no int32/int64 CPU kernel for + // an op, but there is a GPU kernel? + DVLOG(1) << "Forcing op " << op_name + << " to be on the CPU since all input tensors have an " + "int32/int64 dtype, and are small (less than 64 elements)."; + *result = true; + return Status::OK(); +} + +Status MaybePinToResourceDevice(VariantDevice* device, + const EagerOperation& op) { + if (op.colocation_exempt()) { + return Status::OK(); + } + EagerContext& ctx = op.EagerContext(); + Device* op_device = op.Device() == kVariantDeviceNull + ? ctx.HostCPU() + : absl::get(op.Device()); + for (int i = 0; i < op.Inputs().size(); ++i) { + TensorHandle* tensor_handle = op.Inputs()[i]; + if (tensor_handle->dtype == DT_RESOURCE) { + if (tensor_handle->resource_remote_device_incarnation() != 0) { + TF_RETURN_IF_ERROR(ValidateTensorHandleRemoteDevice( + &ctx, tensor_handle->resource_remote_device_incarnation())); + } + Device* resource_device = tensor_handle->resource_device(); + DVLOG(2) << "for op " << op.Name() << " input " << i << " " + << DataTypeString(tensor_handle->dtype) + << " input device = " << resource_device->name() + << ", op device = " << op_device->name(); + // We check for `op->Device() == nullptr` because it can be later + // interpreted as unspecified device and a different device can + // be selected based on device priority. If any input to an op + // is a resource we must pin it to prevent different device selection. + // TODO(iga): null device can mean "unspecified" or "CPU". Clean this up. + if (resource_device != op_device || op.Device() == kVariantDeviceNull) { + DVLOG(1) << (resource_device != op_device ? "Changing " : "Setting ") + << "device of operation " << op.Name() << " to " + << resource_device->name() << " because input #" << i + << " is a resource in this device."; + *device = resource_device; + return Status::OK(); + // No point in looking at other inputs. If there are other resources, + // they must have the same device and we already declared the op to be + // ineligible for CPU pinning. + } + } + } + return Status::OK(); +} + +Status MaybePinToCustomDevice(VariantDevice* device, const EagerOperation& op) { + // If operation was already placed on a custom device, use it. + if (VariantDeviceIsCustom(op.Device())) { + *device = op.Device(); + return Status::OK(); + } + + if (!op.Inputs().empty()) { + // We keep track of what we've seen with devices instead of booleans to be + // able to provide a meaningful error message below. + VariantDevice first = op.Inputs()[0]->device(); + VariantDevice different = first; // A different input device, if any. + VariantDevice custom = first; // The first custom device seen, or an + // arbitrary non-custom device otherwise. + for (size_t i = 1; first == different && i < op.Inputs().size(); ++i) { + VariantDevice device = op.Inputs()[i]->device(); + if (device != first) { + different = device; + } + if (!VariantDeviceIsCustom(custom) && VariantDeviceIsCustom(device)) { + custom = device; + } + if (different != first && VariantDeviceIsCustom(custom)) { + return errors::InvalidArgument(absl::StrCat( + "If an operation has one of its inputs in a custom device, then " + "all inputs should be on that same device. Operation ", + op.Name(), " has one input in custom device ", + VariantDeviceName(custom), + " and at least one input in a different device ", + VariantDeviceName(custom == first ? different : first))); + } + } + if (different == first && VariantDeviceIsCustom(custom)) { + *device = first; + return Status::OK(); + } + } + + return Status::OK(); +} + +} // namespace eager +} // namespace tensorflow diff --git a/tensorflow/core/common_runtime/eager/placement_utils.h b/tensorflow/core/common_runtime/eager/placement_utils.h new file mode 100644 index 00000000000..d58bd304b27 --- /dev/null +++ b/tensorflow/core/common_runtime/eager/placement_utils.h @@ -0,0 +1,55 @@ +/* Copyright 2020 The TensorFlow Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. +==============================================================================*/ + +#ifndef TENSORFLOW_CORE_COMMON_RUNTIME_EAGER_PLACEMENT_UTILS_H_ +#define TENSORFLOW_CORE_COMMON_RUNTIME_EAGER_PLACEMENT_UTILS_H_ + +#include "tensorflow/core/common_runtime/eager/context.h" +#include "tensorflow/core/platform/status.h" +#include "tensorflow/core/platform/stringpiece.h" + +namespace tensorflow { +namespace eager { + +bool IsColocationExempt(StringPiece op_name); + +bool IsFunction(StringPiece op_name); + +bool IsCustomDevice(StringPiece device_name, const EagerContext& ctx); + +// TODO(b/154234908): Unify placement logic. +// TODO(b/159647422): Add C++ unit tests for placement logic. + +// Pin the op to cpu if all op inputs are on the CPU, small (<64 elements) and +// integers (int32/int64). This can be disabled by setting the environment +// variable "TF_EAGER_ENABLE_SMALL_TENSOR_CPU_PINNING" to "0" or "false". +Status MaybePinSmallOpsToCpu(bool* result, StringPiece op_name, + absl::Span args, + const EagerContext& ctx); + +// If a resource touching input is specified, all resource-touching ops run in +// the device the resource is, regardless of anything else that has been +// specified. This is identical to the graph mode behavior. +Status MaybePinToResourceDevice(VariantDevice* device, + const EagerOperation& op); + +// If all the inputs are on the same custom device, use that custom +// device. Otherwise, it is an error to have a custom device as an input. +Status MaybePinToCustomDevice(VariantDevice* device, const EagerOperation& op); + +} // namespace eager +} // namespace tensorflow + +#endif // TENSORFLOW_CORE_COMMON_RUNTIME_EAGER_PLACEMENT_UTILS_H_ diff --git a/tensorflow/core/distributed_runtime/eager/eager_service_impl.cc b/tensorflow/core/distributed_runtime/eager/eager_service_impl.cc index fd0606538c4..4735ff6eaf6 100644 --- a/tensorflow/core/distributed_runtime/eager/eager_service_impl.cc +++ b/tensorflow/core/distributed_runtime/eager/eager_service_impl.cc @@ -491,7 +491,11 @@ Status EagerServiceImpl::ExecuteOp(const Operation& operation, absl::FixedArray retvals(num_retvals); VLOG(3) << "ServerContext: Calling EagerExecute for op " << operation.id(); - TF_RETURN_IF_ERROR(EagerExecute(&op, retvals.data(), &num_retvals)); + TF_RETURN_IF_ERROR(op.Execute( + absl::MakeSpan( + reinterpret_cast(retvals.data()), + num_retvals), + &num_retvals)); return AddOpRetvalsToResponse( eager_context, operation.id(), num_retvals, retvals.data(), diff --git a/tensorflow/lite/delegates/flex/kernel.cc b/tensorflow/lite/delegates/flex/kernel.cc index e7705ecf3ce..b3e978908bd 100644 --- a/tensorflow/lite/delegates/flex/kernel.cc +++ b/tensorflow/lite/delegates/flex/kernel.cc @@ -331,9 +331,12 @@ tensorflow::Status ExecuteFlexOp(TfLiteContext* context, BufferMap* buffer_map, node_data->mutable_outputs()->ResetTensorHandles(); int num_retvals = node_data->NumOutputs(); TF_RETURN_WITH_CONTEXT_IF_ERROR( - EagerExecute(node_data->op(), - node_data->mutable_outputs()->GetTensorHandles()->data(), - &num_retvals), + node_data->op()->Execute( + absl::MakeSpan( + reinterpret_cast( + node_data->mutable_outputs()->GetTensorHandles()->data()), + num_retvals), + &num_retvals), " (while executing '", node_data->name(), "' via Eager)"); if (num_retvals != node_data->NumOutputs()) { From 34b4fab30a40efbdd22c5fb4532e818c4d8e441d Mon Sep 17 00:00:00 2001 From: Advait Jain Date: Mon, 22 Jun 2020 17:39:59 -0700 Subject: [PATCH 509/867] Remove deprecated AddBuiltin API from MicroMutableOpResolver. * Added new API hooks for all the OPs currently supported in TFLM. * These new APIs still need to be implemented with operator specific parse functions but this change allows us to remove the old API and incrementally update the implementations. PiperOrigin-RevId: 317770205 Change-Id: Idaaa687401f2bac5fbf9925e27c04bf536b154ea --- tensorflow/lite/micro/all_ops_resolver.cc | 120 +++---- .../image_recognition_test.cc | 12 +- .../image_recognition_experimental/main.cc | 12 +- .../examples/magic_wand/magic_wand_test.cc | 16 +- .../examples/magic_wand/main_functions.cc | 16 +- .../examples/micro_speech/main_functions.cc | 16 +- .../micro_speech/micro_speech_test.cc | 13 +- .../person_detection/main_functions.cc | 10 +- .../person_detection/person_detection_test.cc | 10 +- .../main_functions.cc | 16 +- .../person_detection_test.cc | 16 +- .../lite/micro/micro_mutable_op_resolver.h | 332 ++++++++++++++++-- .../micro/micro_mutable_op_resolver_test.cc | 33 +- 13 files changed, 408 insertions(+), 214 deletions(-) diff --git a/tensorflow/lite/micro/all_ops_resolver.cc b/tensorflow/lite/micro/all_ops_resolver.cc index b0021a2e771..e728a95360a 100644 --- a/tensorflow/lite/micro/all_ops_resolver.cc +++ b/tensorflow/lite/micro/all_ops_resolver.cc @@ -26,74 +26,60 @@ const char* GetString_ETHOSU(); AllOpsResolver::AllOpsResolver() { // Please keep this list of Builtin Operators in alphabetical order. - AddBuiltin(BuiltinOperator_ABS, tflite::ops::micro::Register_ABS()); - AddBuiltin(BuiltinOperator_ADD, tflite::ops::micro::Register_ADD()); - AddBuiltin(BuiltinOperator_ARG_MAX, tflite::ops::micro::Register_ARG_MAX()); - AddBuiltin(BuiltinOperator_ARG_MIN, tflite::ops::micro::Register_ARG_MIN()); - AddBuiltin(BuiltinOperator_AVERAGE_POOL_2D, - tflite::ops::micro::Register_AVERAGE_POOL_2D()); - AddBuiltin(BuiltinOperator_CEIL, tflite::ops::micro::Register_CEIL()); - AddBuiltin(BuiltinOperator_CONCATENATION, - tflite::ops::micro::Register_CONCATENATION()); - AddBuiltin(BuiltinOperator_CONV_2D, tflite::ops::micro::Register_CONV_2D()); - AddBuiltin(BuiltinOperator_COS, tflite::ops::micro::Register_COS()); - AddBuiltin(BuiltinOperator_DEPTHWISE_CONV_2D, - tflite::ops::micro::Register_DEPTHWISE_CONV_2D()); - AddBuiltin(BuiltinOperator_DEQUANTIZE, - tflite::ops::micro::Register_DEQUANTIZE()); - AddBuiltin(BuiltinOperator_EQUAL, tflite::ops::micro::Register_EQUAL()); - AddBuiltin(BuiltinOperator_FLOOR, tflite::ops::micro::Register_FLOOR()); - AddBuiltin(BuiltinOperator_FULLY_CONNECTED, - tflite::ops::micro::Register_FULLY_CONNECTED()); - AddBuiltin(BuiltinOperator_GREATER, tflite::ops::micro::Register_GREATER()); - AddBuiltin(BuiltinOperator_GREATER_EQUAL, - tflite::ops::micro::Register_GREATER_EQUAL()); - AddBuiltin(BuiltinOperator_L2_NORMALIZATION, - tflite::ops::micro::Register_L2_NORMALIZATION()); - AddBuiltin(BuiltinOperator_LESS, tflite::ops::micro::Register_LESS()); - AddBuiltin(BuiltinOperator_LESS_EQUAL, - tflite::ops::micro::Register_LESS_EQUAL()); - AddBuiltin(BuiltinOperator_LOG, tflite::ops::micro::Register_LOG()); - AddBuiltin(BuiltinOperator_LOGICAL_AND, - tflite::ops::micro::Register_LOGICAL_AND()); - AddBuiltin(BuiltinOperator_LOGICAL_NOT, - tflite::ops::micro::Register_LOGICAL_NOT()); - AddBuiltin(BuiltinOperator_LOGICAL_OR, - tflite::ops::micro::Register_LOGICAL_OR()); - AddBuiltin(BuiltinOperator_LOGISTIC, tflite::ops::micro::Register_LOGISTIC()); - AddBuiltin(BuiltinOperator_MAX_POOL_2D, - tflite::ops::micro::Register_MAX_POOL_2D()); - AddBuiltin(BuiltinOperator_MAXIMUM, tflite::ops::micro::Register_MAXIMUM()); - AddBuiltin(BuiltinOperator_MEAN, tflite::ops::micro::Register_MEAN()); - AddBuiltin(BuiltinOperator_MINIMUM, tflite::ops::micro::Register_MINIMUM()); - AddBuiltin(BuiltinOperator_MUL, tflite::ops::micro::Register_MUL()); - AddBuiltin(BuiltinOperator_NEG, tflite::ops::micro::Register_NEG()); - AddBuiltin(BuiltinOperator_NOT_EQUAL, - tflite::ops::micro::Register_NOT_EQUAL()); - AddBuiltin(BuiltinOperator_PACK, tflite::ops::micro::Register_PACK()); - AddBuiltin(BuiltinOperator_PAD, tflite::ops::micro::Register_PAD()); - AddBuiltin(BuiltinOperator_PADV2, tflite::ops::micro::Register_PADV2()); - AddBuiltin(BuiltinOperator_PRELU, tflite::ops::micro::Register_PRELU()); - AddBuiltin(BuiltinOperator_QUANTIZE, tflite::ops::micro::Register_QUANTIZE()); - AddBuiltin(BuiltinOperator_RELU, tflite::ops::micro::Register_RELU()); - AddBuiltin(BuiltinOperator_RELU6, tflite::ops::micro::Register_RELU6()); - AddBuiltin(BuiltinOperator_RESHAPE, tflite::ops::micro::Register_RESHAPE()); - AddBuiltin(BuiltinOperator_RESIZE_NEAREST_NEIGHBOR, - tflite::ops::micro::Register_RESIZE_NEAREST_NEIGHBOR()); - AddBuiltin(BuiltinOperator_ROUND, tflite::ops::micro::Register_ROUND()); - AddBuiltin(BuiltinOperator_RSQRT, tflite::ops::micro::Register_RSQRT()); - AddBuiltin(BuiltinOperator_SIN, tflite::ops::micro::Register_SIN()); - AddBuiltin(BuiltinOperator_SOFTMAX, tflite::ops::micro::Register_SOFTMAX()); - AddBuiltin(BuiltinOperator_SPLIT, tflite::ops::micro::Register_SPLIT()); - AddBuiltin(BuiltinOperator_SQRT, tflite::ops::micro::Register_SQRT()); - AddBuiltin(BuiltinOperator_SQUARE, tflite::ops::micro::Register_SQUARE()); - AddBuiltin(BuiltinOperator_STRIDED_SLICE, - tflite::ops::micro::Register_STRIDED_SLICE()); - AddBuiltin(BuiltinOperator_SUB, tflite::ops::micro::Register_SUB()); - AddBuiltin(BuiltinOperator_SVDF, tflite::ops::micro::Register_SVDF()); - AddBuiltin(BuiltinOperator_TANH, tflite::ops::micro::Register_TANH()); - AddBuiltin(BuiltinOperator_UNPACK, tflite::ops::micro::Register_UNPACK()); + AddAbs(); + AddAdd(); + AddArgMax(); + AddArgMin(); + AddAveragePool2D(); + AddCeil(); + AddConcatenation(); + AddConv2D(); + AddCos(); + AddDepthwiseConv2D(); + AddDequantize(); + AddEqual(); + AddFloor(); + AddFullyConnected(); + AddGreater(); + AddGreaterEqual(); + AddL2Normalization(); + AddLess(); + AddLessEqual(); + AddLog(); + AddLogicalAnd(); + AddLogicalNot(); + AddLogicalOr(); + AddLogistic(); + AddMaximum(); + AddMaxPool2D(); + AddMean(); + AddMinimum(); + AddMul(); + AddNeg(); + AddNotEqual(); + AddPack(); + AddPad(); + AddPadV2(); + AddPrelu(); + AddQuantize(); + AddRelu(); + AddRelu6(); + AddReshape(); + AddResizeNearestNeighbor(); + AddRound(); + AddRsqrt(); + AddSin(); + AddSoftmax(); + AddSplit(); + AddSqrt(); + AddSquare(); + AddStridedSlice(); + AddSub(); + AddSvdf(); + AddTanh(); + AddUnpack(); + // TODO(b/159644355): Figure out if custom Ops belong in AllOpsResolver. TfLiteRegistration* registration = tflite::ops::micro::custom::Register_ETHOSU(); if (registration) { diff --git a/tensorflow/lite/micro/examples/image_recognition_experimental/image_recognition_test.cc b/tensorflow/lite/micro/examples/image_recognition_experimental/image_recognition_test.cc index ac4de118834..5ad2fb2acbe 100644 --- a/tensorflow/lite/micro/examples/image_recognition_experimental/image_recognition_test.cc +++ b/tensorflow/lite/micro/examples/image_recognition_experimental/image_recognition_test.cc @@ -44,14 +44,10 @@ TF_LITE_MICRO_TEST(TestImageRecognitionInvoke) { tflite::MicroMutableOpResolver<4> micro_op_resolver; - micro_op_resolver.AddBuiltin(tflite::BuiltinOperator_CONV_2D, - tflite::ops::micro::Register_CONV_2D()); - micro_op_resolver.AddBuiltin(tflite::BuiltinOperator_MAX_POOL_2D, - tflite::ops::micro::Register_MAX_POOL_2D()); - micro_op_resolver.AddBuiltin(tflite::BuiltinOperator_FULLY_CONNECTED, - tflite::ops::micro::Register_FULLY_CONNECTED()); - micro_op_resolver.AddBuiltin(tflite::BuiltinOperator_SOFTMAX, - tflite::ops::micro::Register_SOFTMAX()); + micro_op_resolver.AddConv2D(); + micro_op_resolver.AddMaxPool2D(); + micro_op_resolver.AddFullyConnected(); + micro_op_resolver.AddSoftmax(); const int tensor_arena_size = 50 * 1024; uint8_t tensor_arena[tensor_arena_size]; diff --git a/tensorflow/lite/micro/examples/image_recognition_experimental/main.cc b/tensorflow/lite/micro/examples/image_recognition_experimental/main.cc index becdbdf1bd7..fcf7b41b827 100644 --- a/tensorflow/lite/micro/examples/image_recognition_experimental/main.cc +++ b/tensorflow/lite/micro/examples/image_recognition_experimental/main.cc @@ -58,14 +58,10 @@ int main(int argc, char** argv) { tflite::MicroMutableOpResolver<4> micro_op_resolver; - micro_op_resolver.AddBuiltin(tflite::BuiltinOperator_CONV_2D, - tflite::ops::micro::Register_CONV_2D()); - micro_op_resolver.AddBuiltin(tflite::BuiltinOperator_MAX_POOL_2D, - tflite::ops::micro::Register_MAX_POOL_2D()); - micro_op_resolver.AddBuiltin(tflite::BuiltinOperator_FULLY_CONNECTED, - tflite::ops::micro::Register_FULLY_CONNECTED()); - micro_op_resolver.AddBuiltin(tflite::BuiltinOperator_SOFTMAX, - tflite::ops::micro::Register_SOFTMAX()); + micro_op_resolver.AddConv2D(); + micro_op_resolver.AddFullyConnected(); + micro_op_resolver.AddMaxPool2D(); + micro_op_resolver.AddSoftmax(); constexpr int tensor_arena_size = 50 * 1024; uint8_t tensor_arena[tensor_arena_size]; diff --git a/tensorflow/lite/micro/examples/magic_wand/magic_wand_test.cc b/tensorflow/lite/micro/examples/magic_wand/magic_wand_test.cc index 88bfad860e2..fb75afee309 100644 --- a/tensorflow/lite/micro/examples/magic_wand/magic_wand_test.cc +++ b/tensorflow/lite/micro/examples/magic_wand/magic_wand_test.cc @@ -47,17 +47,11 @@ TF_LITE_MICRO_TEST(LoadModelAndPerformInference) { // incur some penalty in code space for op implementations that are not // needed by this graph. static tflite::MicroMutableOpResolver<5> micro_op_resolver; // NOLINT - micro_op_resolver.AddBuiltin( - tflite::BuiltinOperator_DEPTHWISE_CONV_2D, - tflite::ops::micro::Register_DEPTHWISE_CONV_2D()); - micro_op_resolver.AddBuiltin(tflite::BuiltinOperator_MAX_POOL_2D, - tflite::ops::micro::Register_MAX_POOL_2D()); - micro_op_resolver.AddBuiltin(tflite::BuiltinOperator_CONV_2D, - tflite::ops::micro::Register_CONV_2D()); - micro_op_resolver.AddBuiltin(tflite::BuiltinOperator_FULLY_CONNECTED, - tflite::ops::micro::Register_FULLY_CONNECTED()); - micro_op_resolver.AddBuiltin(tflite::BuiltinOperator_SOFTMAX, - tflite::ops::micro::Register_SOFTMAX()); + micro_op_resolver.AddConv2D(); + micro_op_resolver.AddDepthwiseConv2D(); + micro_op_resolver.AddFullyConnected(); + micro_op_resolver.AddMaxPool2D(); + micro_op_resolver.AddSoftmax(); // Create an area of memory to use for input, output, and intermediate arrays. // Finding the minimum value for your model may require some trial and error. diff --git a/tensorflow/lite/micro/examples/magic_wand/main_functions.cc b/tensorflow/lite/micro/examples/magic_wand/main_functions.cc index 26c2eb44747..8defeaad866 100644 --- a/tensorflow/lite/micro/examples/magic_wand/main_functions.cc +++ b/tensorflow/lite/micro/examples/magic_wand/main_functions.cc @@ -66,17 +66,11 @@ void setup() { // incur some penalty in code space for op implementations that are not // needed by this graph. static tflite::MicroMutableOpResolver<5> micro_op_resolver; // NOLINT - micro_op_resolver.AddBuiltin( - tflite::BuiltinOperator_DEPTHWISE_CONV_2D, - tflite::ops::micro::Register_DEPTHWISE_CONV_2D()); - micro_op_resolver.AddBuiltin(tflite::BuiltinOperator_MAX_POOL_2D, - tflite::ops::micro::Register_MAX_POOL_2D()); - micro_op_resolver.AddBuiltin(tflite::BuiltinOperator_CONV_2D, - tflite::ops::micro::Register_CONV_2D()); - micro_op_resolver.AddBuiltin(tflite::BuiltinOperator_FULLY_CONNECTED, - tflite::ops::micro::Register_FULLY_CONNECTED()); - micro_op_resolver.AddBuiltin(tflite::BuiltinOperator_SOFTMAX, - tflite::ops::micro::Register_SOFTMAX()); + micro_op_resolver.AddConv2D(); + micro_op_resolver.AddDepthwiseConv2D(); + micro_op_resolver.AddFullyConnected(); + micro_op_resolver.AddMaxPool2D(); + micro_op_resolver.AddSoftmax(); // Build an interpreter to run the model with. static tflite::MicroInterpreter static_interpreter( diff --git a/tensorflow/lite/micro/examples/micro_speech/main_functions.cc b/tensorflow/lite/micro/examples/micro_speech/main_functions.cc index 30c5022b2d6..d09c4c7af06 100644 --- a/tensorflow/lite/micro/examples/micro_speech/main_functions.cc +++ b/tensorflow/lite/micro/examples/micro_speech/main_functions.cc @@ -75,24 +75,16 @@ void setup() { // tflite::AllOpsResolver resolver; // NOLINTNEXTLINE(runtime-global-variables) static tflite::MicroMutableOpResolver<4> micro_op_resolver(error_reporter); - if (micro_op_resolver.AddBuiltin( - tflite::BuiltinOperator_DEPTHWISE_CONV_2D, - tflite::ops::micro::Register_DEPTHWISE_CONV_2D()) != kTfLiteOk) { + if (micro_op_resolver.AddDepthwiseConv2D() != kTfLiteOk) { return; } - if (micro_op_resolver.AddBuiltin( - tflite::BuiltinOperator_FULLY_CONNECTED, - tflite::ops::micro::Register_FULLY_CONNECTED()) != kTfLiteOk) { + if (micro_op_resolver.AddFullyConnected() != kTfLiteOk) { return; } - if (micro_op_resolver.AddBuiltin(tflite::BuiltinOperator_SOFTMAX, - tflite::ops::micro::Register_SOFTMAX()) != - kTfLiteOk) { + if (micro_op_resolver.AddSoftmax() != kTfLiteOk) { return; } - if (micro_op_resolver.AddBuiltin(tflite::BuiltinOperator_RESHAPE, - tflite::ops::micro::Register_RESHAPE()) != - kTfLiteOk) { + if (micro_op_resolver.AddReshape() != kTfLiteOk) { return; } diff --git a/tensorflow/lite/micro/examples/micro_speech/micro_speech_test.cc b/tensorflow/lite/micro/examples/micro_speech/micro_speech_test.cc index 2c442f955cc..0f6a2afd527 100644 --- a/tensorflow/lite/micro/examples/micro_speech/micro_speech_test.cc +++ b/tensorflow/lite/micro/examples/micro_speech/micro_speech_test.cc @@ -49,15 +49,10 @@ TF_LITE_MICRO_TEST(TestInvoke) { // // tflite::AllOpsResolver resolver; tflite::MicroMutableOpResolver<4> micro_op_resolver; - micro_op_resolver.AddBuiltin( - tflite::BuiltinOperator_DEPTHWISE_CONV_2D, - tflite::ops::micro::Register_DEPTHWISE_CONV_2D()); - micro_op_resolver.AddBuiltin(tflite::BuiltinOperator_FULLY_CONNECTED, - tflite::ops::micro::Register_FULLY_CONNECTED()); - micro_op_resolver.AddBuiltin(tflite::BuiltinOperator_SOFTMAX, - tflite::ops::micro::Register_SOFTMAX()); - micro_op_resolver.AddBuiltin(tflite::BuiltinOperator_RESHAPE, - tflite::ops::micro::Register_RESHAPE()); + micro_op_resolver.AddDepthwiseConv2D(); + micro_op_resolver.AddFullyConnected(); + micro_op_resolver.AddReshape(); + micro_op_resolver.AddSoftmax(); // Create an area of memory to use for input, output, and intermediate arrays. const int tensor_arena_size = 10 * 1024; diff --git a/tensorflow/lite/micro/examples/person_detection/main_functions.cc b/tensorflow/lite/micro/examples/person_detection/main_functions.cc index aa4d83a3334..d7e9f6826c4 100644 --- a/tensorflow/lite/micro/examples/person_detection/main_functions.cc +++ b/tensorflow/lite/micro/examples/person_detection/main_functions.cc @@ -66,13 +66,9 @@ void setup() { // tflite::AllOpsResolver resolver; // NOLINTNEXTLINE(runtime-global-variables) static tflite::MicroMutableOpResolver<3> micro_op_resolver; - micro_op_resolver.AddBuiltin( - tflite::BuiltinOperator_DEPTHWISE_CONV_2D, - tflite::ops::micro::Register_DEPTHWISE_CONV_2D()); - micro_op_resolver.AddBuiltin(tflite::BuiltinOperator_CONV_2D, - tflite::ops::micro::Register_CONV_2D()); - micro_op_resolver.AddBuiltin(tflite::BuiltinOperator_AVERAGE_POOL_2D, - tflite::ops::micro::Register_AVERAGE_POOL_2D()); + micro_op_resolver.AddAveragePool2D(); + micro_op_resolver.AddConv2D(); + micro_op_resolver.AddDepthwiseConv2D(); // Build an interpreter to run the model with. static tflite::MicroInterpreter static_interpreter( diff --git a/tensorflow/lite/micro/examples/person_detection/person_detection_test.cc b/tensorflow/lite/micro/examples/person_detection/person_detection_test.cc index bc53a8410da..7e706d49fcc 100644 --- a/tensorflow/lite/micro/examples/person_detection/person_detection_test.cc +++ b/tensorflow/lite/micro/examples/person_detection/person_detection_test.cc @@ -57,13 +57,9 @@ TF_LITE_MICRO_TEST(TestInvoke) { // // tflite::AllOpsResolver resolver; tflite::MicroMutableOpResolver<3> micro_op_resolver; - micro_op_resolver.AddBuiltin( - tflite::BuiltinOperator_DEPTHWISE_CONV_2D, - tflite::ops::micro::Register_DEPTHWISE_CONV_2D()); - micro_op_resolver.AddBuiltin(tflite::BuiltinOperator_CONV_2D, - tflite::ops::micro::Register_CONV_2D()); - micro_op_resolver.AddBuiltin(tflite::BuiltinOperator_AVERAGE_POOL_2D, - tflite::ops::micro::Register_AVERAGE_POOL_2D()); + micro_op_resolver.AddAveragePool2D(); + micro_op_resolver.AddConv2D(); + micro_op_resolver.AddDepthwiseConv2D(); // Build an interpreter to run the model with. tflite::MicroInterpreter interpreter(model, micro_op_resolver, tensor_arena, diff --git a/tensorflow/lite/micro/examples/person_detection_experimental/main_functions.cc b/tensorflow/lite/micro/examples/person_detection_experimental/main_functions.cc index ac47e36ff8f..09a9cb2c6c4 100644 --- a/tensorflow/lite/micro/examples/person_detection_experimental/main_functions.cc +++ b/tensorflow/lite/micro/examples/person_detection_experimental/main_functions.cc @@ -73,17 +73,11 @@ void setup() { // tflite::AllOpsResolver resolver; // NOLINTNEXTLINE(runtime-global-variables) static tflite::MicroMutableOpResolver<5> micro_op_resolver; - micro_op_resolver.AddBuiltin( - tflite::BuiltinOperator_DEPTHWISE_CONV_2D, - tflite::ops::micro::Register_DEPTHWISE_CONV_2D()); - micro_op_resolver.AddBuiltin(tflite::BuiltinOperator_CONV_2D, - tflite::ops::micro::Register_CONV_2D()); - micro_op_resolver.AddBuiltin(tflite::BuiltinOperator_AVERAGE_POOL_2D, - tflite::ops::micro::Register_AVERAGE_POOL_2D()); - micro_op_resolver.AddBuiltin(tflite::BuiltinOperator_RESHAPE, - tflite::ops::micro::Register_RESHAPE()); - micro_op_resolver.AddBuiltin(tflite::BuiltinOperator_SOFTMAX, - tflite::ops::micro::Register_SOFTMAX()); + micro_op_resolver.AddAveragePool2D(); + micro_op_resolver.AddConv2D(); + micro_op_resolver.AddDepthwiseConv2D(); + micro_op_resolver.AddReshape(); + micro_op_resolver.AddSoftmax(); // Build an interpreter to run the model with. // NOLINTNEXTLINE(runtime-global-variables) diff --git a/tensorflow/lite/micro/examples/person_detection_experimental/person_detection_test.cc b/tensorflow/lite/micro/examples/person_detection_experimental/person_detection_test.cc index ddec8951596..270a427b1df 100644 --- a/tensorflow/lite/micro/examples/person_detection_experimental/person_detection_test.cc +++ b/tensorflow/lite/micro/examples/person_detection_experimental/person_detection_test.cc @@ -53,17 +53,11 @@ TF_LITE_MICRO_TEST(TestInvoke) { // incur some penalty in code space for op implementations that are not // needed by this graph. tflite::MicroMutableOpResolver<5> micro_op_resolver; - micro_op_resolver.AddBuiltin( - tflite::BuiltinOperator_DEPTHWISE_CONV_2D, - tflite::ops::micro::Register_DEPTHWISE_CONV_2D()); - micro_op_resolver.AddBuiltin(tflite::BuiltinOperator_CONV_2D, - tflite::ops::micro::Register_CONV_2D()); - micro_op_resolver.AddBuiltin(tflite::BuiltinOperator_AVERAGE_POOL_2D, - tflite::ops::micro::Register_AVERAGE_POOL_2D()); - micro_op_resolver.AddBuiltin(tflite::BuiltinOperator_RESHAPE, - tflite::ops::micro::Register_RESHAPE()); - micro_op_resolver.AddBuiltin(tflite::BuiltinOperator_SOFTMAX, - tflite::ops::micro::Register_SOFTMAX()); + micro_op_resolver.AddAveragePool2D(); + micro_op_resolver.AddConv2D(); + micro_op_resolver.AddDepthwiseConv2D(); + micro_op_resolver.AddReshape(); + micro_op_resolver.AddSoftmax(); // Build an interpreter to run the model with. tflite::MicroInterpreter interpreter(model, micro_op_resolver, tensor_arena, diff --git a/tensorflow/lite/micro/micro_mutable_op_resolver.h b/tensorflow/lite/micro/micro_mutable_op_resolver.h index b9ce2bb4bba..1b76f440a61 100644 --- a/tensorflow/lite/micro/micro_mutable_op_resolver.h +++ b/tensorflow/lite/micro/micro_mutable_op_resolver.h @@ -104,39 +104,72 @@ class MicroMutableOpResolver : public MicroOpResolver { return kTfLiteOk; } - // Registers a Builtin Operator with the MicroOpResolver. - // - // Only the first call for a given BuiltinOperator enum will be successful. - // i.e. if this function is called again for a previously added - // BuiltinOperator, the MicroOpResolver will be unchanged and this function - // will return kTfLiteError. - // - // TODO(b/149408647): remove this API once the BuiltinOperator specific Add - // functions are fully implemented. - TfLiteStatus AddBuiltin(tflite::BuiltinOperator op, - TfLiteRegistration* registration) { - TFLITE_DCHECK(registration != nullptr); - // For code that is not switched over to the new selective registration of - // the parse function, we pass in ParseOpData. This allows for backwards - // compatibility. - return AddBuiltin(op, *registration, ParseOpData); - } - // The Add* functions below add the various Builtin operators to the // MicroMutableOpResolver object. - // - // This API is currently experimental (and only supported for a small subset - // of operators). It will soon be preferred over the AddBuiltin function for - // the following reason: - // * If all calls to AddBuiltin for an application use this API, the code - // size will be smaller by 5-8K (compared to the using the AddBuiltin - // override). + + TfLiteStatus AddAbs() { + // TODO(b/149408647): Replace ParseOpData with the operator specific parse + // function. + return AddBuiltin(BuiltinOperator_ABS, *tflite::ops::micro::Register_ABS(), + ParseOpData); + } + + TfLiteStatus AddAdd() { + // TODO(b/149408647): Replace ParseOpData with the operator specific parse + // function. + return AddBuiltin(BuiltinOperator_ADD, *tflite::ops::micro::Register_ADD(), + ParseOpData); + } + + TfLiteStatus AddArgMax() { + // TODO(b/149408647): Replace ParseOpData with the operator specific parse + // function. + return AddBuiltin(BuiltinOperator_ARG_MAX, + *tflite::ops::micro::Register_ARG_MAX(), ParseOpData); + } + + TfLiteStatus AddArgMin() { + // TODO(b/149408647): Replace ParseOpData with the operator specific parse + // function. + return AddBuiltin(BuiltinOperator_ARG_MIN, + *tflite::ops::micro::Register_ARG_MIN(), ParseOpData); + } + + TfLiteStatus AddAveragePool2D() { + // TODO(b/149408647): Replace ParseOpData with the operator specific parse + // function. + return AddBuiltin(BuiltinOperator_AVERAGE_POOL_2D, + *tflite::ops::micro::Register_AVERAGE_POOL_2D(), + ParseOpData); + } + + TfLiteStatus AddCeil() { + // TODO(b/149408647): Replace ParseOpData with the operator specific parse + // function. + return AddBuiltin(BuiltinOperator_CEIL, + *tflite::ops::micro::Register_CEIL(), ParseOpData); + } + + TfLiteStatus AddConcatenation() { + // TODO(b/149408647): Replace ParseOpData with the operator specific parse + // function. + return AddBuiltin(BuiltinOperator_CONCATENATION, + *tflite::ops::micro::Register_CONCATENATION(), + ParseOpData); + } TfLiteStatus AddConv2D() { return AddBuiltin(BuiltinOperator_CONV_2D, *tflite::ops::micro::Register_CONV_2D(), ParseConv2D); } + TfLiteStatus AddCos() { + // TODO(b/149408647): Replace ParseOpData with the operator specific parse + // function. + return AddBuiltin(BuiltinOperator_COS, *tflite::ops::micro::Register_COS(), + ParseOpData); + } + TfLiteStatus AddDepthwiseConv2D() { return AddBuiltin(BuiltinOperator_DEPTHWISE_CONV_2D, *tflite::ops::micro::Register_DEPTHWISE_CONV_2D(), @@ -149,12 +182,91 @@ class MicroMutableOpResolver : public MicroOpResolver { ParseDequantize); } + TfLiteStatus AddEqual() { + // TODO(b/149408647): Replace ParseOpData with the operator specific parse + // function. + return AddBuiltin(BuiltinOperator_EQUAL, + *tflite::ops::micro::Register_EQUAL(), ParseOpData); + } + + TfLiteStatus AddFloor() { + // TODO(b/149408647): Replace ParseOpData with the operator specific parse + // function. + return AddBuiltin(BuiltinOperator_FLOOR, + *tflite::ops::micro::Register_FLOOR(), ParseOpData); + } + TfLiteStatus AddFullyConnected() { return AddBuiltin(BuiltinOperator_FULLY_CONNECTED, *tflite::ops::micro::Register_FULLY_CONNECTED(), ParseFullyConnected); } + TfLiteStatus AddGreater() { + // TODO(b/149408647): Replace ParseOpData with the operator specific parse + // function. + return AddBuiltin(BuiltinOperator_GREATER, + *tflite::ops::micro::Register_GREATER(), ParseOpData); + } + + TfLiteStatus AddGreaterEqual() { + // TODO(b/149408647): Replace ParseOpData with the operator specific parse + // function. + return AddBuiltin(BuiltinOperator_GREATER_EQUAL, + *tflite::ops::micro::Register_GREATER_EQUAL(), + ParseOpData); + } + + TfLiteStatus AddL2Normalization() { + // TODO(b/149408647): Replace ParseOpData with the operator specific parse + // function. + return AddBuiltin(BuiltinOperator_L2_NORMALIZATION, + *tflite::ops::micro::Register_L2_NORMALIZATION(), + ParseOpData); + } + + TfLiteStatus AddLess() { + // TODO(b/149408647): Replace ParseOpData with the operator specific parse + // function. + return AddBuiltin(BuiltinOperator_LESS, + *tflite::ops::micro::Register_LESS(), ParseOpData); + } + + TfLiteStatus AddLessEqual() { + // TODO(b/149408647): Replace ParseOpData with the operator specific parse + // function. + return AddBuiltin(BuiltinOperator_LESS_EQUAL, + *tflite::ops::micro::Register_LESS_EQUAL(), ParseOpData); + } + + TfLiteStatus AddLog() { + // TODO(b/149408647): Replace ParseOpData with the operator specific parse + // function. + return AddBuiltin(BuiltinOperator_LOG, *tflite::ops::micro::Register_LOG(), + ParseOpData); + } + + TfLiteStatus AddLogicalAnd() { + // TODO(b/149408647): Replace ParseOpData with the operator specific parse + // function. + return AddBuiltin(BuiltinOperator_LOGICAL_AND, + *tflite::ops::micro::Register_LOGICAL_AND(), ParseOpData); + } + + TfLiteStatus AddLogicalNot() { + // TODO(b/149408647): Replace ParseOpData with the operator specific parse + // function. + return AddBuiltin(BuiltinOperator_LOGICAL_NOT, + *tflite::ops::micro::Register_LOGICAL_NOT(), ParseOpData); + } + + TfLiteStatus AddLogicalOr() { + // TODO(b/149408647): Replace ParseOpData with the operator specific parse + // function. + return AddBuiltin(BuiltinOperator_LOGICAL_OR, + *tflite::ops::micro::Register_LOGICAL_OR(), ParseOpData); + } + TfLiteStatus AddLogistic() { // TODO(b/149408647): Replace ParseOpData with the operator specific parse // function. @@ -162,26 +274,196 @@ class MicroMutableOpResolver : public MicroOpResolver { *tflite::ops::micro::Register_LOGISTIC(), ParseOpData); } + TfLiteStatus AddMaximum() { + // TODO(b/149408647): Replace ParseOpData with the operator specific parse + // function. + return AddBuiltin(BuiltinOperator_MAXIMUM, + *tflite::ops::micro::Register_MAXIMUM(), ParseOpData); + } + + TfLiteStatus AddMaxPool2D() { + // TODO(b/149408647): Replace ParseOpData with the operator specific parse + // function. + return AddBuiltin(BuiltinOperator_MAX_POOL_2D, + *tflite::ops::micro::Register_MAX_POOL_2D(), ParseOpData); + } + + TfLiteStatus AddMean() { + // TODO(b/149408647): Replace ParseOpData with the operator specific parse + // function. + return AddBuiltin(BuiltinOperator_MEAN, + *tflite::ops::micro::Register_MEAN(), ParseOpData); + } + + TfLiteStatus AddMinimum() { + // TODO(b/149408647): Replace ParseOpData with the operator specific parse + // function. + return AddBuiltin(BuiltinOperator_MINIMUM, + *tflite::ops::micro::Register_MINIMUM(), ParseOpData); + } + + TfLiteStatus AddMul() { + // TODO(b/149408647): Replace ParseOpData with the operator specific parse + // function. + return AddBuiltin(BuiltinOperator_MUL, *tflite::ops::micro::Register_MUL(), + ParseOpData); + } + + TfLiteStatus AddNeg() { + // TODO(b/149408647): Replace ParseOpData with the operator specific parse + // function. + return AddBuiltin(BuiltinOperator_NEG, *tflite::ops::micro::Register_NEG(), + ParseOpData); + } + + TfLiteStatus AddNotEqual() { + // TODO(b/149408647): Replace ParseOpData with the operator specific parse + // function. + return AddBuiltin(BuiltinOperator_NOT_EQUAL, + *tflite::ops::micro::Register_NOT_EQUAL(), ParseOpData); + } + + TfLiteStatus AddPack() { + // TODO(b/149408647): Replace ParseOpData with the operator specific parse + // function. + return AddBuiltin(BuiltinOperator_PACK, + *tflite::ops::micro::Register_PACK(), ParseOpData); + } + + TfLiteStatus AddPad() { + // TODO(b/149408647): Replace ParseOpData with the operator specific parse + // function. + return AddBuiltin(BuiltinOperator_PAD, *tflite::ops::micro::Register_PAD(), + ParseOpData); + } + + TfLiteStatus AddPadV2() { + // TODO(b/149408647): Replace ParseOpData with the operator specific parse + // function. + return AddBuiltin(BuiltinOperator_PADV2, + *tflite::ops::micro::Register_PADV2(), ParseOpData); + } + + TfLiteStatus AddPrelu() { + // TODO(b/149408647): Replace ParseOpData with the operator specific parse + // function. + return AddBuiltin(BuiltinOperator_PRELU, + *tflite::ops::micro::Register_PRELU(), ParseOpData); + } + TfLiteStatus AddQuantize() { return AddBuiltin(BuiltinOperator_QUANTIZE, *tflite::ops::micro::Register_QUANTIZE(), ParseQuantize); } + TfLiteStatus AddRelu() { + // TODO(b/149408647): Replace ParseOpData with the operator specific parse + // function. + return AddBuiltin(BuiltinOperator_RELU, + *tflite::ops::micro::Register_RELU(), ParseOpData); + } + + TfLiteStatus AddRelu6() { + // TODO(b/149408647): Replace ParseOpData with the operator specific parse + // function. + return AddBuiltin(BuiltinOperator_RELU6, + *tflite::ops::micro::Register_RELU6(), ParseOpData); + } + TfLiteStatus AddReshape() { return AddBuiltin(BuiltinOperator_RESHAPE, *tflite::ops::micro::Register_RESHAPE(), ParseReshape); } + TfLiteStatus AddResizeNearestNeighbor() { + // TODO(b/149408647): Replace ParseOpData with the operator specific parse + // function. + return AddBuiltin(BuiltinOperator_RESIZE_NEAREST_NEIGHBOR, + *tflite::ops::micro::Register_RESIZE_NEAREST_NEIGHBOR(), + ParseOpData); + } + + TfLiteStatus AddRound() { + // TODO(b/149408647): Replace ParseOpData with the operator specific parse + // function. + return AddBuiltin(BuiltinOperator_ROUND, + *tflite::ops::micro::Register_ROUND(), ParseOpData); + } + + TfLiteStatus AddRsqrt() { + // TODO(b/149408647): Replace ParseOpData with the operator specific parse + // function. + return AddBuiltin(BuiltinOperator_RSQRT, + *tflite::ops::micro::Register_RSQRT(), ParseOpData); + } + + TfLiteStatus AddSin() { + // TODO(b/149408647): Replace ParseOpData with the operator specific parse + // function. + return AddBuiltin(BuiltinOperator_SIN, *tflite::ops::micro::Register_SIN(), + ParseOpData); + } + TfLiteStatus AddSoftmax() { return AddBuiltin(BuiltinOperator_SOFTMAX, *tflite::ops::micro::Register_SOFTMAX(), ParseSoftmax); } + TfLiteStatus AddSplit() { + // TODO(b/149408647): Replace ParseOpData with the operator specific parse + // function. + return AddBuiltin(BuiltinOperator_SPLIT, + *tflite::ops::micro::Register_SPLIT(), ParseOpData); + } + + TfLiteStatus AddSqrt() { + // TODO(b/149408647): Replace ParseOpData with the operator specific parse + // function. + return AddBuiltin(BuiltinOperator_SQRT, + *tflite::ops::micro::Register_SQRT(), ParseOpData); + } + + TfLiteStatus AddSquare() { + // TODO(b/149408647): Replace ParseOpData with the operator specific parse + // function. + return AddBuiltin(BuiltinOperator_SQUARE, + *tflite::ops::micro::Register_SQUARE(), ParseOpData); + } + + TfLiteStatus AddStridedSlice() { + // TODO(b/149408647): Replace ParseOpData with the operator specific parse + // function. + return AddBuiltin(BuiltinOperator_STRIDED_SLICE, + *tflite::ops::micro::Register_STRIDED_SLICE(), + ParseOpData); + } + + TfLiteStatus AddSub() { + // TODO(b/149408647): Replace ParseOpData with the operator specific parse + // function. + return AddBuiltin(BuiltinOperator_SUB, *tflite::ops::micro::Register_SUB(), + ParseOpData); + } + TfLiteStatus AddSvdf() { return AddBuiltin(BuiltinOperator_SVDF, *tflite::ops::micro::Register_SVDF(), ParseSvdf); } + TfLiteStatus AddTanh() { + // TODO(b/149408647): Replace ParseOpData with the operator specific parse + // function. + return AddBuiltin(BuiltinOperator_TANH, + *tflite::ops::micro::Register_TANH(), ParseOpData); + } + + TfLiteStatus AddUnpack() { + // TODO(b/149408647): Replace ParseOpData with the operator specific parse + // function. + return AddBuiltin(BuiltinOperator_UNPACK, + *tflite::ops::micro::Register_UNPACK(), ParseOpData); + } + unsigned int GetRegistrationLength() { return registrations_len_; } private: diff --git a/tensorflow/lite/micro/micro_mutable_op_resolver_test.cc b/tensorflow/lite/micro/micro_mutable_op_resolver_test.cc index ff5dfdf3a9a..fe9c8de5959 100644 --- a/tensorflow/lite/micro/micro_mutable_op_resolver_test.cc +++ b/tensorflow/lite/micro/micro_mutable_op_resolver_test.cc @@ -68,14 +68,7 @@ TF_LITE_MICRO_TEST(TestOperations) { static TfLiteRegistration r = {tflite::MockInit, tflite::MockFree, tflite::MockPrepare, tflite::MockInvoke}; - MicroMutableOpResolver<2> micro_op_resolver; - TF_LITE_MICRO_EXPECT_EQ( - kTfLiteOk, micro_op_resolver.AddBuiltin(BuiltinOperator_CONV_2D, &r)); - - // Only one AddBuiltin per operator should return kTfLiteOk. - TF_LITE_MICRO_EXPECT_EQ( - kTfLiteError, micro_op_resolver.AddBuiltin(BuiltinOperator_CONV_2D, &r)); - + MicroMutableOpResolver<1> micro_op_resolver; TF_LITE_MICRO_EXPECT_EQ(kTfLiteOk, micro_op_resolver.AddCustom("mock_custom", &r)); @@ -85,16 +78,10 @@ TF_LITE_MICRO_TEST(TestOperations) { tflite::MicroOpResolver* resolver = µ_op_resolver; + TF_LITE_MICRO_EXPECT_EQ(1, micro_op_resolver.GetRegistrationLength()); + const TfLiteRegistration* registration = - resolver->FindOp(BuiltinOperator_CONV_2D); - TF_LITE_MICRO_EXPECT_NE(nullptr, registration); - TF_LITE_MICRO_EXPECT_EQ(nullptr, registration->init(nullptr, nullptr, 0)); - TF_LITE_MICRO_EXPECT_EQ(kTfLiteOk, registration->prepare(nullptr, nullptr)); - TF_LITE_MICRO_EXPECT_EQ(kTfLiteOk, registration->invoke(nullptr, nullptr)); - - TF_LITE_MICRO_EXPECT_EQ(2, micro_op_resolver.GetRegistrationLength()); - - registration = resolver->FindOp(BuiltinOperator_RELU); + resolver->FindOp(BuiltinOperator_RELU); TF_LITE_MICRO_EXPECT_EQ(nullptr, registration); registration = resolver->FindOp("mock_custom"); @@ -116,12 +103,7 @@ TF_LITE_MICRO_TEST(TestErrorReporting) { tflite::MockPrepare, tflite::MockInvoke}; tflite::MockErrorReporter mock_reporter; - MicroMutableOpResolver<2> micro_op_resolver(&mock_reporter); - TF_LITE_MICRO_EXPECT_EQ(false, mock_reporter.HasBeenCalled()); - mock_reporter.ResetState(); - - TF_LITE_MICRO_EXPECT_EQ( - kTfLiteOk, micro_op_resolver.AddBuiltin(BuiltinOperator_CONV_2D, &r)); + MicroMutableOpResolver<1> micro_op_resolver(&mock_reporter); TF_LITE_MICRO_EXPECT_EQ(false, mock_reporter.HasBeenCalled()); mock_reporter.ResetState(); @@ -132,10 +114,7 @@ TF_LITE_MICRO_TEST(TestErrorReporting) { // Attempting to Add more operators than the class template parameter for // MicroMutableOpResolver should result in errors. - TF_LITE_MICRO_EXPECT_EQ( - kTfLiteError, micro_op_resolver.AddBuiltin(BuiltinOperator_RELU, &r)); - TF_LITE_MICRO_EXPECT_EQ(true, mock_reporter.HasBeenCalled()); - mock_reporter.ResetState(); + TF_LITE_MICRO_EXPECT_EQ(kTfLiteError, micro_op_resolver.AddRelu()); TF_LITE_MICRO_EXPECT_EQ(kTfLiteError, micro_op_resolver.AddCustom("mock_custom_1", &r)); From e74a115bcf5cd27f476b46161a639e9ec599491d Mon Sep 17 00:00:00 2001 From: Peng Wang Date: Mon, 22 Jun 2020 17:47:47 -0700 Subject: [PATCH 510/867] [TF-numpy] Adds __rmatmul__ method to ndarray. PiperOrigin-RevId: 317771125 Change-Id: I719c46d97ae1c68ac59dcd1cf8f65d067ddc7658 --- tensorflow/python/ops/numpy_ops/np_math_ops.py | 3 ++- 1 file changed, 2 insertions(+), 1 deletion(-) diff --git a/tensorflow/python/ops/numpy_ops/np_math_ops.py b/tensorflow/python/ops/numpy_ops/np_math_ops.py index 427aa96e5a4..03329bbdbf1 100644 --- a/tensorflow/python/ops/numpy_ops/np_math_ops.py +++ b/tensorflow/python/ops/numpy_ops/np_math_ops.py @@ -950,11 +950,12 @@ setattr(np_arrays.ndarray, '__sub__', _wrap(subtract)) setattr(np_arrays.ndarray, '__rsub__', _wrap(subtract, True)) setattr(np_arrays.ndarray, '__mul__', _wrap(multiply)) setattr(np_arrays.ndarray, '__rmul__', _wrap(multiply, True)) +setattr(np_arrays.ndarray, '__matmul__', _wrap(matmul)) +setattr(np_arrays.ndarray, '__rmatmul__', _wrap(matmul, True)) setattr(np_arrays.ndarray, '__pow__', _wrap(power)) setattr(np_arrays.ndarray, '__rpow__', _wrap(power, True)) setattr(np_arrays.ndarray, '__truediv__', _wrap(true_divide)) setattr(np_arrays.ndarray, '__rtruediv__', _wrap(true_divide, True)) -setattr(np_arrays.ndarray, '__matmul__', _wrap(matmul)) def _comparison(tf_fun, x1, x2, cast_bool_to_int=False): From 4d13d6416da49b82b375d631082607be3c58b57f Mon Sep 17 00:00:00 2001 From: Rick Chao Date: Mon, 22 Jun 2020 17:49:06 -0700 Subject: [PATCH 511/867] Make cluster_resolver standard property in tf.distribute strategies. PiperOrigin-RevId: 317771299 Change-Id: I71b5c585cef7bd7ef80e66b75e30287fddcf89e2 --- tensorflow/python/distribute/BUILD | 5 +- .../collective_all_reduce_strategy.py | 12 ++++ .../collective_all_reduce_strategy_test.py | 15 ++++- .../python/distribute/distribute_lib.py | 59 +++++++++++++++++++ .../python/distribute/distribute_lib_test.py | 13 ++++ .../python/distribute/strategy_common_test.py | 35 +++++++++++ tensorflow/python/distribute/tpu_strategy.py | 12 ++++ .../python/distribute/tpu_strategy_test.py | 7 +++ ...orflow.distribute.-mirrored-strategy.pbtxt | 4 ++ ...flow.distribute.-one-device-strategy.pbtxt | 4 ++ .../v1/tensorflow.distribute.-strategy.pbtxt | 4 ++ ...perimental.-central-storage-strategy.pbtxt | 4 ++ ...ntal.-multi-worker-mirrored-strategy.pbtxt | 4 ++ ...erimental.-parameter-server-strategy.pbtxt | 4 ++ ...tribute.experimental.-t-p-u-strategy.pbtxt | 4 ++ ...orflow.distribute.-mirrored-strategy.pbtxt | 4 ++ ...flow.distribute.-one-device-strategy.pbtxt | 4 ++ .../v2/tensorflow.distribute.-strategy.pbtxt | 4 ++ ...ensorflow.distribute.-t-p-u-strategy.pbtxt | 4 ++ ...perimental.-central-storage-strategy.pbtxt | 4 ++ ...ntal.-multi-worker-mirrored-strategy.pbtxt | 4 ++ ...erimental.-parameter-server-strategy.pbtxt | 4 ++ ...tribute.experimental.-t-p-u-strategy.pbtxt | 4 ++ 23 files changed, 215 insertions(+), 3 deletions(-) diff --git a/tensorflow/python/distribute/BUILD b/tensorflow/python/distribute/BUILD index 74d80b63e12..85ee8de5635 100644 --- a/tensorflow/python/distribute/BUILD +++ b/tensorflow/python/distribute/BUILD @@ -204,6 +204,7 @@ py_test( "//tensorflow/python:variables", "//tensorflow/python/autograph/core:test_lib", "//tensorflow/python/data/ops:dataset_ops", + "//tensorflow/python/distribute/cluster_resolver:cluster_resolver_lib", "//third_party/py/numpy", ], ) @@ -1847,10 +1848,11 @@ py_test( ], ) -cuda_py_test( +distribute_py_test( name = "strategy_common_test", srcs = ["strategy_common_test.py"], python_version = "PY3", + shard_count = 12, tags = [ "multi_and_single_gpu", # TODO(b/155301154): Enable this test on multi-gpu guitar once multi process @@ -1859,6 +1861,7 @@ cuda_py_test( ], xla_enable_strict_auto_jit = True, deps = [ + ":collective_all_reduce_strategy", ":combinations", ":multi_worker_test_base", ":reduce_util", diff --git a/tensorflow/python/distribute/collective_all_reduce_strategy.py b/tensorflow/python/distribute/collective_all_reduce_strategy.py index e2b039ceb23..e754cc43a41 100644 --- a/tensorflow/python/distribute/collective_all_reduce_strategy.py +++ b/tensorflow/python/distribute/collective_all_reduce_strategy.py @@ -138,6 +138,18 @@ class CollectiveAllReduceStrategy(distribute_lib.Strategy): """ return super(CollectiveAllReduceStrategy, self).scope() + @property + def cluster_resolver(self): + """Returns the cluster resolver associated with this strategy. + + As a multi-worker strategy, + `tf.distribute.experimental.MultiWorkerMirroredStrategy` provides the + associated `tf.distribute.cluster_resolver.ClusterResolver`. If the user + provides one in `__init__`, that instance is returned; if the user does + not, a default `TFConfigClusterResolver` is provided. + """ + return self.extended._cluster_resolver # pylint: disable=protected-access + @tf_export(v1=["distribute.experimental.MultiWorkerMirroredStrategy"]) # pylint: disable=missing-docstring class CollectiveAllReduceStrategyV1(distribute_lib.StrategyV1): diff --git a/tensorflow/python/distribute/collective_all_reduce_strategy_test.py b/tensorflow/python/distribute/collective_all_reduce_strategy_test.py index 87212c85fc4..9d16c6f5a1f 100644 --- a/tensorflow/python/distribute/collective_all_reduce_strategy_test.py +++ b/tensorflow/python/distribute/collective_all_reduce_strategy_test.py @@ -505,8 +505,7 @@ class DistributedCollectiveAllReduceStrategyTest( self.assertEqual(['CollectiveReduce'], new_rewrite_options.scoped_allocator_opts.enable_op) - @combinations.generate(combinations.combine(mode=['eager'])) - def testEnableCollectiveOps(self): + def _get_strategy_with_mocked_methods(self): mock_called = [False] # pylint: disable=dangerous-default-value @@ -525,9 +524,21 @@ class DistributedCollectiveAllReduceStrategyTest( mock_configure_collective_ops): strategy, _, _ = self._get_test_object( task_type='worker', task_id=1, num_gpus=2) + + return strategy, mock_called + + @combinations.generate(combinations.combine(mode=['eager'])) + def testEnableCollectiveOps(self): + strategy, mock_called = self._get_strategy_with_mocked_methods() self.assertTrue(strategy.extended._std_server_started) self.assertTrue(mock_called[0]) + @combinations.generate(combinations.combine(mode=['eager'])) + def testEnableCollectiveOpsAndClusterResolver(self): + strategy, _ = self._get_strategy_with_mocked_methods() + self.assertEqual(strategy.cluster_resolver.task_type, 'worker') + self.assertEqual(strategy.cluster_resolver.task_id, 1) + class DistributedCollectiveAllReduceStrategyTestWithChief( CollectiveAllReduceStrategyTestBase, parameterized.TestCase): diff --git a/tensorflow/python/distribute/distribute_lib.py b/tensorflow/python/distribute/distribute_lib.py index 5abfb6e1c09..f32427b88e0 100644 --- a/tensorflow/python/distribute/distribute_lib.py +++ b/tensorflow/python/distribute/distribute_lib.py @@ -1439,6 +1439,65 @@ class StrategyBase(object): def __copy__(self): raise RuntimeError("Must only deepcopy DistributionStrategy.") + @property + def cluster_resolver(self): + """Returns the cluster resolver associated with this strategy. + + In general, when using a multi-worker `tf.distribute` strategy such as + `tf.distribute.experimental.MultiWorkerMirroredStrategy` or + `tf.distribute.experimental.TPUStrategy()`, there is a + `tf.distribute.cluster_resolver.ClusterResolver` associated with the + strategy used, and such an instance is returned by this property. + + Strategies that intend to have an associated + `tf.distribute.cluster_resolver.ClusterResolver` must set the + relevant attribute, or override this property; otherwise, `None` is returned + by default. Those strategies should also provide information regarding what + is returned by this property. + + Single-worker strategies usually do not have a + `tf.distribute.cluster_resolver.ClusterResolver`, and in those cases this + property will return `None`. + + The `tf.distribute.cluster_resolver.ClusterResolver` may be useful when the + user needs to access information such as the cluster spec, task type or task + id. For example, + + ```python + + os.environ['TF_CONFIG'] = json.dumps({ + 'cluster': { + 'worker': ["localhost:12345", "localhost:23456"], + 'ps': ["localhost:34567"] + }, + 'task': {'type': 'worker', 'index': 0} + }) + + # This implicitly uses TF_CONFIG for the cluster and current task info. + strategy = tf.distribute.experimental.MultiWorkerMirroredStrategy() + + ... + + if strategy.cluster_resolver.task_type == 'worker': + # Perform something that's only applicable on workers. Since we set this + # as a worker above, this block will run on this particular instance. + elif strategy.cluster_resolver.task_type == 'ps': + # Perform something that's only applicable on parameter servers. Since we + # set this as a worker above, this block will not run on this particular + # instance. + ``` + + For more information, please see + `tf.distribute.cluster_resolver.ClusterResolver`'s API docstring. + + Returns: + The cluster resolver associated with this strategy. Returns `None` if a + cluster resolver is not applicable or available in this strategy. + """ + if hasattr(self.extended, "_cluster_resolver"): + return self.extended._cluster_resolver # pylint: disable=protected-access + return None + @tf_export("distribute.Strategy", v1=[]) # pylint: disable=g-missing-docstring class Strategy(StrategyBase): diff --git a/tensorflow/python/distribute/distribute_lib_test.py b/tensorflow/python/distribute/distribute_lib_test.py index 8ea1cac6f02..b5924ec3b67 100644 --- a/tensorflow/python/distribute/distribute_lib_test.py +++ b/tensorflow/python/distribute/distribute_lib_test.py @@ -28,6 +28,7 @@ from tensorflow.python.distribute import distribute_lib from tensorflow.python.distribute import distribution_strategy_context as ds_context from tensorflow.python.distribute import input_lib from tensorflow.python.distribute import reduce_util +from tensorflow.python.distribute.cluster_resolver import SimpleClusterResolver from tensorflow.python.eager import context from tensorflow.python.eager import def_function from tensorflow.python.framework import constant_op @@ -36,6 +37,7 @@ from tensorflow.python.framework import ops from tensorflow.python.ops import variable_scope from tensorflow.python.ops import variables from tensorflow.python.platform import test +from tensorflow.python.training import server_lib from tensorflow.python.util import nest @@ -422,6 +424,17 @@ class TestStrategyTest(test.TestCase): test_fn() + def testClusterResolverDefaultNotImplemented(self): + dist = _TestStrategy() + self.assertIsNone(dist.cluster_resolver) + base_cluster_spec = server_lib.ClusterSpec({ + "ps": ["ps0:2222", "ps1:2222"], + "worker": ["worker0:2222", "worker1:2222", "worker2:2222"] + }) + cluster_resolver = SimpleClusterResolver(base_cluster_spec) + dist.extended._cluster_resolver = cluster_resolver + self.assertIs(dist.cluster_resolver, cluster_resolver) + # _TestStrategy2 is like _TestStrategy, except it doesn't change variable # creation. diff --git a/tensorflow/python/distribute/strategy_common_test.py b/tensorflow/python/distribute/strategy_common_test.py index b823e2de331..d1a72da7e7c 100644 --- a/tensorflow/python/distribute/strategy_common_test.py +++ b/tensorflow/python/distribute/strategy_common_test.py @@ -27,6 +27,8 @@ from tensorflow.python.distribute import multi_worker_test_base from tensorflow.python.distribute import reduce_util from tensorflow.python.distribute import strategy_combinations from tensorflow.python.distribute import strategy_test_lib +from tensorflow.python.distribute.collective_all_reduce_strategy import CollectiveAllReduceStrategy +from tensorflow.python.distribute.tpu_strategy import TPUStrategy from tensorflow.python.eager import def_function from tensorflow.python.framework import dtypes from tensorflow.python.ops import array_ops @@ -184,5 +186,38 @@ class DistributedCollectiveAllReduceStrategyTest( # worker strategy combinations can run on a fixed number of GPUs. +class StrategyClusterResolverTest(test.TestCase, parameterized.TestCase): + + @combinations.generate( + combinations.combine( + strategy=[strategy_combinations.multi_worker_mirrored_two_workers] + + strategy_combinations.all_strategies, + mode=['eager'])) + def testClusterResolverProperty(self, strategy): + # CollectiveAllReduceStrategy and TPUStrategy must have a cluster resolver. + # `None` otherwise. + resolver = strategy.cluster_resolver + if not isinstance(strategy, CollectiveAllReduceStrategy) and not isinstance( + strategy, TPUStrategy): + self.assertIsNone(resolver) + return + + with strategy.scope(): + self.assertIs(strategy.cluster_resolver, resolver) + self.assertTrue(hasattr(resolver, 'cluster_spec')) + self.assertTrue(hasattr(resolver, 'environment')) + self.assertTrue(hasattr(resolver, 'master')) + self.assertTrue(hasattr(resolver, 'num_accelerators')) + self.assertIsNone(resolver.rpc_layer) + if isinstance(strategy, CollectiveAllReduceStrategy): + self.assertGreaterEqual(resolver.task_id, 0) + self.assertLessEqual(resolver.task_id, 1) + self.assertEqual(resolver.task_type, 'worker') + elif isinstance(strategy, TPUStrategy): + # TPUStrategy does not have task_id and task_type applicable. + self.assertIsNone(resolver.task_id) + self.assertIsNone(resolver.task_type) + + if __name__ == '__main__': combinations.main() diff --git a/tensorflow/python/distribute/tpu_strategy.py b/tensorflow/python/distribute/tpu_strategy.py index 4b3c4be0ccd..df393c61dbb 100644 --- a/tensorflow/python/distribute/tpu_strategy.py +++ b/tensorflow/python/distribute/tpu_strategy.py @@ -345,6 +345,18 @@ class TPUStrategy(distribute_lib.Strategy): options = options or distribute_lib.RunOptions() return self.extended.tpu_run(fn, args, kwargs, options) + @property + def cluster_resolver(self): + """Returns the cluster resolver associated with this strategy. + + `tf.distribute.experimental.TPUStrategy` provides the + associated `tf.distribute.cluster_resolver.ClusterResolver`. If the user + provides one in `__init__`, that instance is returned; if the user does + not, a default + `tf.distribute.cluster_resolver.TPUClusterResolver` is provided. + """ + return self.extended._tpu_cluster_resolver # pylint: disable=protected-access + @tf_export(v1=["distribute.experimental.TPUStrategy"]) class TPUStrategyV1(distribute_lib.StrategyV1): diff --git a/tensorflow/python/distribute/tpu_strategy_test.py b/tensorflow/python/distribute/tpu_strategy_test.py index 5e47e750d87..142743a6ec2 100644 --- a/tensorflow/python/distribute/tpu_strategy_test.py +++ b/tensorflow/python/distribute/tpu_strategy_test.py @@ -555,6 +555,13 @@ class TPUStrategyTest(test.TestCase, parameterized.TestCase): update_variable.get_concrete_function() self.assertLen(strategy.extended.worker_devices, trace_count[0]) + def test_cluster_resolver_available(self, enable_packed_var): + resolver = get_tpu_cluster_resolver() + remote.connect_to_cluster(resolver) + tpu_strategy_util.initialize_tpu_system(resolver) + strategy = tpu_lib.TPUStrategy(resolver) + self.assertIs(strategy.cluster_resolver, resolver) + class TPUStrategyDataPrefetchTest(test.TestCase): diff --git a/tensorflow/tools/api/golden/v1/tensorflow.distribute.-mirrored-strategy.pbtxt b/tensorflow/tools/api/golden/v1/tensorflow.distribute.-mirrored-strategy.pbtxt index 36c78c406b7..0c5db602029 100644 --- a/tensorflow/tools/api/golden/v1/tensorflow.distribute.-mirrored-strategy.pbtxt +++ b/tensorflow/tools/api/golden/v1/tensorflow.distribute.-mirrored-strategy.pbtxt @@ -4,6 +4,10 @@ tf_class { is_instance: "" is_instance: "" is_instance: "" + member { + name: "cluster_resolver" + mtype: "" + } member { name: "extended" mtype: "" diff --git a/tensorflow/tools/api/golden/v1/tensorflow.distribute.-one-device-strategy.pbtxt b/tensorflow/tools/api/golden/v1/tensorflow.distribute.-one-device-strategy.pbtxt index 09865ab02ee..ae62acffa44 100644 --- a/tensorflow/tools/api/golden/v1/tensorflow.distribute.-one-device-strategy.pbtxt +++ b/tensorflow/tools/api/golden/v1/tensorflow.distribute.-one-device-strategy.pbtxt @@ -4,6 +4,10 @@ tf_class { is_instance: "" is_instance: "" is_instance: "" + member { + name: "cluster_resolver" + mtype: "" + } member { name: "extended" mtype: "" diff --git a/tensorflow/tools/api/golden/v1/tensorflow.distribute.-strategy.pbtxt b/tensorflow/tools/api/golden/v1/tensorflow.distribute.-strategy.pbtxt index 0e6c10bd533..9285405ea4f 100644 --- a/tensorflow/tools/api/golden/v1/tensorflow.distribute.-strategy.pbtxt +++ b/tensorflow/tools/api/golden/v1/tensorflow.distribute.-strategy.pbtxt @@ -3,6 +3,10 @@ tf_class { is_instance: "" is_instance: "" is_instance: "" + member { + name: "cluster_resolver" + mtype: "" + } member { name: "extended" mtype: "" diff --git a/tensorflow/tools/api/golden/v1/tensorflow.distribute.experimental.-central-storage-strategy.pbtxt b/tensorflow/tools/api/golden/v1/tensorflow.distribute.experimental.-central-storage-strategy.pbtxt index fbc4c107a1a..3c3d785ac7c 100644 --- a/tensorflow/tools/api/golden/v1/tensorflow.distribute.experimental.-central-storage-strategy.pbtxt +++ b/tensorflow/tools/api/golden/v1/tensorflow.distribute.experimental.-central-storage-strategy.pbtxt @@ -4,6 +4,10 @@ tf_class { is_instance: "" is_instance: "" is_instance: "" + member { + name: "cluster_resolver" + mtype: "" + } member { name: "extended" mtype: "" diff --git a/tensorflow/tools/api/golden/v1/tensorflow.distribute.experimental.-multi-worker-mirrored-strategy.pbtxt b/tensorflow/tools/api/golden/v1/tensorflow.distribute.experimental.-multi-worker-mirrored-strategy.pbtxt index cd67e7d27c4..e1f8bea251b 100644 --- a/tensorflow/tools/api/golden/v1/tensorflow.distribute.experimental.-multi-worker-mirrored-strategy.pbtxt +++ b/tensorflow/tools/api/golden/v1/tensorflow.distribute.experimental.-multi-worker-mirrored-strategy.pbtxt @@ -4,6 +4,10 @@ tf_class { is_instance: "" is_instance: "" is_instance: "" + member { + name: "cluster_resolver" + mtype: "" + } member { name: "extended" mtype: "" diff --git a/tensorflow/tools/api/golden/v1/tensorflow.distribute.experimental.-parameter-server-strategy.pbtxt b/tensorflow/tools/api/golden/v1/tensorflow.distribute.experimental.-parameter-server-strategy.pbtxt index 0eff82474ff..6ae83d18589 100644 --- a/tensorflow/tools/api/golden/v1/tensorflow.distribute.experimental.-parameter-server-strategy.pbtxt +++ b/tensorflow/tools/api/golden/v1/tensorflow.distribute.experimental.-parameter-server-strategy.pbtxt @@ -4,6 +4,10 @@ tf_class { is_instance: "" is_instance: "" is_instance: "" + member { + name: "cluster_resolver" + mtype: "" + } member { name: "extended" mtype: "" diff --git a/tensorflow/tools/api/golden/v1/tensorflow.distribute.experimental.-t-p-u-strategy.pbtxt b/tensorflow/tools/api/golden/v1/tensorflow.distribute.experimental.-t-p-u-strategy.pbtxt index 2af9a5ad095..0e548eca9b5 100644 --- a/tensorflow/tools/api/golden/v1/tensorflow.distribute.experimental.-t-p-u-strategy.pbtxt +++ b/tensorflow/tools/api/golden/v1/tensorflow.distribute.experimental.-t-p-u-strategy.pbtxt @@ -4,6 +4,10 @@ tf_class { is_instance: "" is_instance: "" is_instance: "" + member { + name: "cluster_resolver" + mtype: "" + } member { name: "extended" mtype: "" diff --git a/tensorflow/tools/api/golden/v2/tensorflow.distribute.-mirrored-strategy.pbtxt b/tensorflow/tools/api/golden/v2/tensorflow.distribute.-mirrored-strategy.pbtxt index be4c841aed7..8817f16d808 100644 --- a/tensorflow/tools/api/golden/v2/tensorflow.distribute.-mirrored-strategy.pbtxt +++ b/tensorflow/tools/api/golden/v2/tensorflow.distribute.-mirrored-strategy.pbtxt @@ -4,6 +4,10 @@ tf_class { is_instance: "" is_instance: "" is_instance: "" + member { + name: "cluster_resolver" + mtype: "" + } member { name: "extended" mtype: "" diff --git a/tensorflow/tools/api/golden/v2/tensorflow.distribute.-one-device-strategy.pbtxt b/tensorflow/tools/api/golden/v2/tensorflow.distribute.-one-device-strategy.pbtxt index 4557fe1060b..992243ffe8a 100644 --- a/tensorflow/tools/api/golden/v2/tensorflow.distribute.-one-device-strategy.pbtxt +++ b/tensorflow/tools/api/golden/v2/tensorflow.distribute.-one-device-strategy.pbtxt @@ -4,6 +4,10 @@ tf_class { is_instance: "" is_instance: "" is_instance: "" + member { + name: "cluster_resolver" + mtype: "" + } member { name: "extended" mtype: "" diff --git a/tensorflow/tools/api/golden/v2/tensorflow.distribute.-strategy.pbtxt b/tensorflow/tools/api/golden/v2/tensorflow.distribute.-strategy.pbtxt index 9f6a2ac32be..8140088e701 100644 --- a/tensorflow/tools/api/golden/v2/tensorflow.distribute.-strategy.pbtxt +++ b/tensorflow/tools/api/golden/v2/tensorflow.distribute.-strategy.pbtxt @@ -3,6 +3,10 @@ tf_class { is_instance: "" is_instance: "" is_instance: "" + member { + name: "cluster_resolver" + mtype: "" + } member { name: "extended" mtype: "" diff --git a/tensorflow/tools/api/golden/v2/tensorflow.distribute.-t-p-u-strategy.pbtxt b/tensorflow/tools/api/golden/v2/tensorflow.distribute.-t-p-u-strategy.pbtxt index f41a08454e2..29947a1c9c5 100644 --- a/tensorflow/tools/api/golden/v2/tensorflow.distribute.-t-p-u-strategy.pbtxt +++ b/tensorflow/tools/api/golden/v2/tensorflow.distribute.-t-p-u-strategy.pbtxt @@ -4,6 +4,10 @@ tf_class { is_instance: "" is_instance: "" is_instance: "" + member { + name: "cluster_resolver" + mtype: "" + } member { name: "extended" mtype: "" diff --git a/tensorflow/tools/api/golden/v2/tensorflow.distribute.experimental.-central-storage-strategy.pbtxt b/tensorflow/tools/api/golden/v2/tensorflow.distribute.experimental.-central-storage-strategy.pbtxt index dd61960c66f..695fb52358b 100644 --- a/tensorflow/tools/api/golden/v2/tensorflow.distribute.experimental.-central-storage-strategy.pbtxt +++ b/tensorflow/tools/api/golden/v2/tensorflow.distribute.experimental.-central-storage-strategy.pbtxt @@ -4,6 +4,10 @@ tf_class { is_instance: "" is_instance: "" is_instance: "" + member { + name: "cluster_resolver" + mtype: "" + } member { name: "extended" mtype: "" diff --git a/tensorflow/tools/api/golden/v2/tensorflow.distribute.experimental.-multi-worker-mirrored-strategy.pbtxt b/tensorflow/tools/api/golden/v2/tensorflow.distribute.experimental.-multi-worker-mirrored-strategy.pbtxt index 500ae362e5f..43632e17b6d 100644 --- a/tensorflow/tools/api/golden/v2/tensorflow.distribute.experimental.-multi-worker-mirrored-strategy.pbtxt +++ b/tensorflow/tools/api/golden/v2/tensorflow.distribute.experimental.-multi-worker-mirrored-strategy.pbtxt @@ -4,6 +4,10 @@ tf_class { is_instance: "" is_instance: "" is_instance: "" + member { + name: "cluster_resolver" + mtype: "" + } member { name: "extended" mtype: "" diff --git a/tensorflow/tools/api/golden/v2/tensorflow.distribute.experimental.-parameter-server-strategy.pbtxt b/tensorflow/tools/api/golden/v2/tensorflow.distribute.experimental.-parameter-server-strategy.pbtxt index d6dc9627d9a..39181625469 100644 --- a/tensorflow/tools/api/golden/v2/tensorflow.distribute.experimental.-parameter-server-strategy.pbtxt +++ b/tensorflow/tools/api/golden/v2/tensorflow.distribute.experimental.-parameter-server-strategy.pbtxt @@ -4,6 +4,10 @@ tf_class { is_instance: "" is_instance: "" is_instance: "" + member { + name: "cluster_resolver" + mtype: "" + } member { name: "extended" mtype: "" diff --git a/tensorflow/tools/api/golden/v2/tensorflow.distribute.experimental.-t-p-u-strategy.pbtxt b/tensorflow/tools/api/golden/v2/tensorflow.distribute.experimental.-t-p-u-strategy.pbtxt index 82a4362a597..855cdbfb175 100644 --- a/tensorflow/tools/api/golden/v2/tensorflow.distribute.experimental.-t-p-u-strategy.pbtxt +++ b/tensorflow/tools/api/golden/v2/tensorflow.distribute.experimental.-t-p-u-strategy.pbtxt @@ -4,6 +4,10 @@ tf_class { is_instance: "" is_instance: "" is_instance: "" + member { + name: "cluster_resolver" + mtype: "" + } member { name: "extended" mtype: "" From e60cf08994a837d1dc37ec2f85063e44b3987a81 Mon Sep 17 00:00:00 2001 From: Rahul Joshi Date: Mon, 22 Jun 2020 18:11:15 -0700 Subject: [PATCH 512/867] [NFC] Re-instate use of FuncOp::isPublic() PiperOrigin-RevId: 317774876 Change-Id: I6a832236377d403b1ebd24ecbc26025c37dc1c13 --- .../compiler/mlir/tensorflow/ir/tf_saved_model.cc | 13 +++++-------- 1 file changed, 5 insertions(+), 8 deletions(-) diff --git a/tensorflow/compiler/mlir/tensorflow/ir/tf_saved_model.cc b/tensorflow/compiler/mlir/tensorflow/ir/tf_saved_model.cc index ef55761686e..5a7d81d4c0c 100644 --- a/tensorflow/compiler/mlir/tensorflow/ir/tf_saved_model.cc +++ b/tensorflow/compiler/mlir/tensorflow/ir/tf_saved_model.cc @@ -232,22 +232,19 @@ static LogicalResult VerifySavedModelModule( for (auto func : module.getOps()) { const bool is_exported = IsExported(func); - if (is_exported && func.getVisibility() != FuncOp::Visibility::Public) { + if (is_exported && !func.isPublic()) { return func.emitError() << "exported function @" << func.getName() << " should be public"; } - if (!is_exported && func.getVisibility() == FuncOp::Visibility::Public) { + if (!is_exported && func.isPublic()) { return func.emitError() << "non-exported function @" << func.getName() << " should be private"; } - if (HasAnyTfSavedModelArgAttr(func)) { - if (!is_exported) { - return func.emitError() - << "can only apply 'tf_saved_model' argument attributes " - "to exported functions"; - } + if (!is_exported && HasAnyTfSavedModelArgAttr(func)) { + return func.emitError() << "can only apply 'tf_saved_model' argument " + "attributes to exported functions"; } } From 265de52331a10af793f733d21e2152123819a269 Mon Sep 17 00:00:00 2001 From: Scott Zhu Date: Mon, 22 Jun 2020 18:20:43 -0700 Subject: [PATCH 513/867] Fix input mapping issue when model is constructed/tested with dict input tensor. The mapping of the dict input tensors was not correct since it was still using the tensor name, rather than the key of the tensor when build the model. This cause the issue down the stream when the inputs are provided with unknown keys. We had some backup logic, which will probably do correct things, eg just flat the dict to keep the original order, which was correct most of the case, but not very reliable. In this change, we make the behavior change: 1. When model is build with dict input tensors, the key of the tensor, instead of the name, will be used to map the tensor with input data. 2. Unknown keys in the input data will result into a warning. We didn't throw error since user might do it intentionally, eg using part of the model to test with full input data. PiperOrigin-RevId: 317776370 Change-Id: I91983443f2b770cb0b45ddb7726f52708cb91d61 --- tensorflow/python/keras/engine/functional.py | 24 ++++++++-- .../python/keras/engine/functional_test.py | 46 +++++++++++++++++++ 2 files changed, 67 insertions(+), 3 deletions(-) diff --git a/tensorflow/python/keras/engine/functional.py b/tensorflow/python/keras/engine/functional.py index 0612d70044d..fd80e7f8bb4 100644 --- a/tensorflow/python/keras/engine/functional.py +++ b/tensorflow/python/keras/engine/functional.py @@ -22,6 +22,7 @@ from __future__ import print_function import collections import copy import itertools +import warnings from six.moves import zip # pylint: disable=redefined-builtin @@ -131,10 +132,10 @@ class Functional(training_lib.Model): # Models constructed with a single Tensor or list of Tensors can # be called with a dict, where the keys of the dict are the names - # of the `Input` objects. Extra keys are ignored. + # of the `Input` objects. Extra keys are ignored with warning. self._enable_dict_to_input_mapping = ( not nest.is_sequence(self._nested_inputs) or - (isinstance(self._nested_inputs, (list, tuple)) and + (isinstance(self._nested_inputs, (list, tuple, dict)) and not any(nest.is_sequence(t) for t in self._nested_inputs))) if any(not hasattr(tensor, '_keras_history') for tensor in self.outputs): @@ -524,10 +525,27 @@ class Functional(training_lib.Model): ref_inputs = self._nested_inputs if not nest.is_sequence(ref_inputs): ref_inputs = [self._nested_inputs] + if isinstance(ref_inputs, dict): + # In the case that the graph is constructed with dict input tensors, + # We will use the original dict key to map with the keys in the input + # data. Note that the model.inputs is using nest.flatten to process the + # input tensors, which means the dict input tensors are ordered by their + # keys. + ref_input_names = sorted(ref_inputs.keys()) + else: + ref_input_names = [inp._keras_history.layer.name for inp in ref_inputs] + + # Raise an warning if there are more input data comparing to input tensor + if len(tensors) > len(ref_input_names): + warnings.warn( + 'Input dict contained keys {} which did not match any model input. ' + 'They will be ignored by the model.'.format( + [n for n in tensors.keys() if n not in ref_input_names]) + ) try: # Flatten in the order `Input`s were passed during Model construction. - return [tensors[inp._keras_history.layer.name] for inp in ref_inputs] + return [tensors[n] for n in ref_input_names] except KeyError: # TODO(b/151582614) return nest.flatten(tensors) diff --git a/tensorflow/python/keras/engine/functional_test.py b/tensorflow/python/keras/engine/functional_test.py index 3c14411deb9..0e82d95d3de 100644 --- a/tensorflow/python/keras/engine/functional_test.py +++ b/tensorflow/python/keras/engine/functional_test.py @@ -18,8 +18,11 @@ from __future__ import absolute_import from __future__ import division from __future__ import print_function +import warnings + import numpy as np + from tensorflow.python.eager import context from tensorflow.python.eager import def_function from tensorflow.python.framework import constant_op @@ -43,6 +46,7 @@ from tensorflow.python.keras.utils import tf_utils from tensorflow.python.ops import array_ops from tensorflow.python.ops import math_ops from tensorflow.python.ops import state_ops +from tensorflow.python.ops import string_ops from tensorflow.python.ops.ragged import ragged_factory_ops from tensorflow.python.platform import test from tensorflow.python.training.tracking.util import Checkpoint @@ -1565,6 +1569,48 @@ class DefaultShapeInferenceBehaviorTest(keras_parameterized.TestCase): self.assertEqual(config['layers'][2]['inbound_nodes'], [[['in1', 0, 0, {}], ['in2', 0, 0, {}]]]) + @combinations.generate(combinations.combine(mode=['eager'])) + def test_dict_inputs_tensors(self): + # Note that this test is running with v2 eager only, since the v1 + # will behave differently wrt to dict input for training. + inputs = { + 'sentence2': input_layer_lib.Input( + shape=(), name='a', dtype=dtypes.string), + 'sentence1': input_layer_lib.Input( + shape=(), name='b', dtype=dtypes.string), + } + strlen = layers.Lambda(string_ops.string_length_v2) + diff = layers.Subtract()( + [strlen(inputs['sentence1']), strlen(inputs['sentence2'])]) + diff = math_ops.cast(diff, dtypes.float32) + model = training_lib.Model(inputs, diff) + + extra_keys = { + 'sentence1': constant_op.constant(['brown fox', 'lazy dog']), + 'sentence2': constant_op.constant(['owl', 'cheeky cat']), + 'label': constant_op.constant([0, 1]), + } + + with warnings.catch_warnings(record=True) as w: + warnings.simplefilter('always') + model(extra_keys) + self.assertIn('ignored by the model', str(w[-1].message)) + + model.compile('sgd', 'mse') + with warnings.catch_warnings(record=True) as w: + warnings.simplefilter('always') + model.fit(extra_keys, y=constant_op.constant([0, 1]), steps_per_epoch=1) + self.assertIn('ignored by the model', str(w[-1].message)) + + with warnings.catch_warnings(record=True) as w: + warnings.simplefilter('always') + model.evaluate(extra_keys, constant_op.constant([0, 1])) + self.assertIn('ignored by the model', str(w[-1].message)) + + # Make sure the model inputs are sorted with the dict keys. + self.assertEqual(model.inputs[0]._keras_history.layer.name, 'b') + self.assertEqual(model.inputs[1]._keras_history.layer.name, 'a') + class GraphUtilsTest(test.TestCase): From d8b90acf95bdff09e29b1065739e28e8bd7500d9 Mon Sep 17 00:00:00 2001 From: Dan Moldovan Date: Mon, 22 Jun 2020 18:46:12 -0700 Subject: [PATCH 514/867] Enable type annotations for python/ops. PiperOrigin-RevId: 317779660 Change-Id: Ife6b7319ef394b611798f5d01b64ebdb3c0a01cc --- tensorflow/python/ops/logging_ops.py | 18 +++--------------- 1 file changed, 3 insertions(+), 15 deletions(-) diff --git a/tensorflow/python/ops/logging_ops.py b/tensorflow/python/ops/logging_ops.py index 8ca63f55987..02fce277690 100644 --- a/tensorflow/python/ops/logging_ops.py +++ b/tensorflow/python/ops/logging_ops.py @@ -54,11 +54,9 @@ except NameError: # call relies on certain conditionals for its dependencies. Use # control_flow_ops.Assert. -# Assert and Print are special symbols in python, so we must -# have an upper-case version of them. -# -# For users with Python 3 or Python 2.7 -# with `from __future__ import print_function`, we could also allow lowercase. +# Assert and Print are special symbols in Python 2, so we must +# have an upper-case version of them. When support for it is dropped, +# we can allow lowercase. # See https://github.com/tensorflow/tensorflow/issues/18053 @@ -83,11 +81,6 @@ def Print(input_, data, message=None, first_n=None, summarize=None, name=None): with jupyter notebook (printing to the notebook *server's* output, not into the notebook). - Additionally, to use tf.print in python 2.7, users must make sure to import - the following: - - `from __future__ import print_function` - Args: input_: A tensor passed through this op. data: A list of tensors to print out when op is evaluated. @@ -148,11 +141,6 @@ def print_v2(*inputs, **kwargs): Python objects. Printed tensors will recursively show the first and last elements of each dimension to summarize. - @compatibility(python2) - In python 2.7, make sure to import the following: - `from __future__ import print_function` - @end_compatibility - Example: Single-input usage: From 7c669e5f795a75bf6665c9596d730e438ab36511 Mon Sep 17 00:00:00 2001 From: Bixia Zheng Date: Mon, 22 Jun 2020 19:13:19 -0700 Subject: [PATCH 515/867] [TF:TRT] Add #if to logger_registry.h for consistency. Add #if GOOGLE_TENSORRT to logger_registry.h to make it consistent with logger_registry.cc. Rewrite two lines of #if into #if GOOGLE_CUDA && GOOGLE_TENSORRT. PiperOrigin-RevId: 317782636 Change-Id: Icaf7116b5013f57bb37b80511c7c195f8bd7e2d5 --- tensorflow/compiler/tf2tensorrt/common/utils.h | 6 ++---- .../compiler/tf2tensorrt/convert/convert_graph.cc | 6 ++---- .../compiler/tf2tensorrt/convert/convert_graph.h | 6 ++---- .../tf2tensorrt/convert/convert_graph_test.cc | 6 ++---- .../compiler/tf2tensorrt/convert/convert_nodes.cc | 6 ++---- .../compiler/tf2tensorrt/convert/convert_nodes.h | 6 ++---- .../tf2tensorrt/convert/convert_nodes_test.cc | 6 ++---- .../compiler/tf2tensorrt/convert/logger_registry.cc | 6 ++---- .../compiler/tf2tensorrt/convert/logger_registry.h | 5 +++-- .../tf2tensorrt/convert/trt_optimization_pass.cc | 6 ++---- .../tf2tensorrt/convert/trt_optimization_pass.h | 6 ++---- .../tf2tensorrt/kernels/get_calibration_data_op.cc | 6 ++---- .../compiler/tf2tensorrt/kernels/trt_engine_op.cc | 6 ++---- .../tf2tensorrt/kernels/trt_engine_op_test.cc | 6 ++---- .../tf2tensorrt/kernels/trt_engine_resource_ops.cc | 6 ++---- .../kernels/trt_engine_resource_ops_test.cc | 6 ++---- .../tf2tensorrt/ops/get_calibration_data_op.cc | 6 ++---- tensorflow/compiler/tf2tensorrt/ops/trt_engine_op.cc | 6 ++---- .../tf2tensorrt/ops/trt_engine_resource_ops.cc | 6 ++---- .../compiler/tf2tensorrt/plugin/plugin_cast.cu.cc | 6 ++---- tensorflow/compiler/tf2tensorrt/plugin/trt_plugin.cc | 6 ++---- tensorflow/compiler/tf2tensorrt/plugin/trt_plugin.h | 6 ++---- tensorflow/compiler/tf2tensorrt/segment/segment.cc | 6 ++---- tensorflow/compiler/tf2tensorrt/segment/segment.h | 6 ++---- .../compiler/tf2tensorrt/segment/segment_test.cc | 6 ++---- tensorflow/compiler/tf2tensorrt/segment/union_find.h | 6 ++---- tensorflow/compiler/tf2tensorrt/tensorrt_test.cc | 6 ++---- .../compiler/tf2tensorrt/utils/trt_allocator.cc | 12 ++++-------- .../compiler/tf2tensorrt/utils/trt_allocator.h | 12 ++++-------- .../compiler/tf2tensorrt/utils/trt_engine_utils.cc | 6 ++---- .../compiler/tf2tensorrt/utils/trt_engine_utils.h | 6 ++---- .../tf2tensorrt/utils/trt_int8_calibrator.cc | 6 ++---- .../compiler/tf2tensorrt/utils/trt_int8_calibrator.h | 6 ++---- tensorflow/compiler/tf2tensorrt/utils/trt_logger.cc | 6 ++---- tensorflow/compiler/tf2tensorrt/utils/trt_logger.h | 6 ++---- .../compiler/tf2tensorrt/utils/trt_lru_cache.cc | 6 ++---- .../compiler/tf2tensorrt/utils/trt_lru_cache.h | 6 ++---- .../utils/trt_shape_optimization_profiles.h | 6 ++---- .../utils/trt_shape_optimization_profiles_test.cc | 6 ++---- 39 files changed, 83 insertions(+), 162 deletions(-) diff --git a/tensorflow/compiler/tf2tensorrt/common/utils.h b/tensorflow/compiler/tf2tensorrt/common/utils.h index 9ab0145e1ec..b428733ecd4 100644 --- a/tensorflow/compiler/tf2tensorrt/common/utils.h +++ b/tensorflow/compiler/tf2tensorrt/common/utils.h @@ -16,8 +16,7 @@ limitations under the License. #ifndef TENSORFLOW_COMPILER_TF2TENSORRT_COMMON_UTILS_H_ #define TENSORFLOW_COMPILER_TF2TENSORRT_COMMON_UTILS_H_ -#if GOOGLE_CUDA -#if GOOGLE_TENSORRT +#if GOOGLE_CUDA && GOOGLE_TENSORRT #include "tensorflow/core/platform/logging.h" @@ -29,7 +28,6 @@ namespace tensorrt { } // namespace tensorrt } // namespace tensorflow -#endif -#endif +#endif // GOOGLE_CUDA && GOOGLE_TENSORRT #endif // TENSORFLOW_COMPILER_TF2TENSORRT_COMMON_UTILS_H_ diff --git a/tensorflow/compiler/tf2tensorrt/convert/convert_graph.cc b/tensorflow/compiler/tf2tensorrt/convert/convert_graph.cc index 1c51d51f1c9..5429aaf3362 100644 --- a/tensorflow/compiler/tf2tensorrt/convert/convert_graph.cc +++ b/tensorflow/compiler/tf2tensorrt/convert/convert_graph.cc @@ -53,8 +53,7 @@ limitations under the License. #include "tensorflow/core/util/device_name_utils.h" #include "tensorflow/tools/graph_transforms/transform_utils.h" -#if GOOGLE_CUDA -#if GOOGLE_TENSORRT +#if GOOGLE_CUDA && GOOGLE_TENSORRT #include "third_party/gpus/cuda/include/cuda_runtime_api.h" #include "third_party/tensorrt/NvInfer.h" namespace tensorflow { @@ -884,5 +883,4 @@ Status ConvertAfterShapes(const ConversionParams& params) { } // namespace tensorrt } // namespace tensorflow -#endif // GOOGLE_TENSORRT -#endif // GOOGLE_CUDA +#endif // GOOGLE_CUDA && GOOGLE_TENSORRT diff --git a/tensorflow/compiler/tf2tensorrt/convert/convert_graph.h b/tensorflow/compiler/tf2tensorrt/convert/convert_graph.h index 53ab84a6fa9..d3897e864fa 100644 --- a/tensorflow/compiler/tf2tensorrt/convert/convert_graph.h +++ b/tensorflow/compiler/tf2tensorrt/convert/convert_graph.h @@ -24,8 +24,7 @@ limitations under the License. #include "tensorflow/core/lib/core/status.h" #include "tensorflow/core/platform/types.h" -#if GOOGLE_CUDA -#if GOOGLE_TENSORRT +#if GOOGLE_CUDA && GOOGLE_TENSORRT namespace tensorflow { namespace tensorrt { @@ -66,7 +65,6 @@ Status RegisterGraphToFunctionLibrary(const GraphDef& segment_graph_def, } // namespace tensorrt } // namespace tensorflow -#endif // GOOGLE_TENSORRT -#endif // GOOGLE_CUDA +#endif // GOOGLE_CUDA && GOOGLE_TENSORRT #endif // TENSORFLOW_COMPILER_TF2TENSORRT_CONVERT_CONVERT_GRAPH_H_ diff --git a/tensorflow/compiler/tf2tensorrt/convert/convert_graph_test.cc b/tensorflow/compiler/tf2tensorrt/convert/convert_graph_test.cc index a1f523d6bfa..54fb1d56441 100644 --- a/tensorflow/compiler/tf2tensorrt/convert/convert_graph_test.cc +++ b/tensorflow/compiler/tf2tensorrt/convert/convert_graph_test.cc @@ -34,8 +34,7 @@ limitations under the License. #include "tensorflow/core/protobuf/config.pb.h" // NOLINT #include "tensorflow/core/public/session.h" -#if GOOGLE_CUDA -#if GOOGLE_TENSORRT +#if GOOGLE_CUDA && GOOGLE_TENSORRT namespace tensorflow { namespace tensorrt { @@ -231,5 +230,4 @@ TEST_F(ConvertAfterShapesTest, DirectlyConnectedEngines) { } // namespace tensorrt } // namespace tensorflow -#endif // GOOGLE_TENSORRT -#endif // GOOGLE_CUDA +#endif // GOOGLE_CUDA && GOOGLE_TENSORRT diff --git a/tensorflow/compiler/tf2tensorrt/convert/convert_nodes.cc b/tensorflow/compiler/tf2tensorrt/convert/convert_nodes.cc index 96cec556942..2ec616ba621 100644 --- a/tensorflow/compiler/tf2tensorrt/convert/convert_nodes.cc +++ b/tensorflow/compiler/tf2tensorrt/convert/convert_nodes.cc @@ -59,8 +59,7 @@ limitations under the License. #include "tensorflow/core/util/env_var.h" #include "tensorflow/core/util/strided_slice_op.h" -#if GOOGLE_CUDA -#if GOOGLE_TENSORRT +#if GOOGLE_CUDA && GOOGLE_TENSORRT #include "third_party/tensorrt/NvInfer.h" #include "third_party/tensorrt/NvInferPlugin.h" @@ -6258,5 +6257,4 @@ bool OutputEdgeValidator::operator()(const Edge* out_edge) const { } // namespace tensorrt } // namespace tensorflow -#endif // GOOGLE_TENSORRT -#endif // GOOGLE_CUDA +#endif // GOOGLE_CUDA && GOOGLE_TENSORRT diff --git a/tensorflow/compiler/tf2tensorrt/convert/convert_nodes.h b/tensorflow/compiler/tf2tensorrt/convert/convert_nodes.h index 7a1276c645c..a621735fad1 100644 --- a/tensorflow/compiler/tf2tensorrt/convert/convert_nodes.h +++ b/tensorflow/compiler/tf2tensorrt/convert/convert_nodes.h @@ -33,8 +33,7 @@ limitations under the License. #include "tensorflow/core/lib/core/status.h" #include "tensorflow/stream_executor/lib/statusor.h" -#if GOOGLE_CUDA -#if GOOGLE_TENSORRT +#if GOOGLE_CUDA && GOOGLE_TENSORRT #include "third_party/tensorrt/NvInfer.h" namespace tensorflow { @@ -694,7 +693,6 @@ BinaryOperationMap(); } // namespace tensorrt } // namespace tensorflow -#endif // GOOGLE_TENSORRT -#endif // GOOGLE_CUDA +#endif // GOOGLE_CUDA && GOOGLE_TENSORRT #endif // TENSORFLOW_COMPILER_TF2TENSORRT_CONVERT_CONVERT_NODES_H_ diff --git a/tensorflow/compiler/tf2tensorrt/convert/convert_nodes_test.cc b/tensorflow/compiler/tf2tensorrt/convert/convert_nodes_test.cc index c24b169f651..53ec9ee7ada 100644 --- a/tensorflow/compiler/tf2tensorrt/convert/convert_nodes_test.cc +++ b/tensorflow/compiler/tf2tensorrt/convert/convert_nodes_test.cc @@ -21,8 +21,7 @@ limitations under the License. #include #include -#if GOOGLE_CUDA -#if GOOGLE_TENSORRT +#if GOOGLE_CUDA && GOOGLE_TENSORRT #include #include @@ -6636,5 +6635,4 @@ TEST_F(OpConverterTest, ConvertPad) { } // namespace tensorrt } // namespace tensorflow -#endif // GOOGLE_TENSORRT -#endif // GOOGLE_CUDA +#endif // GOOGLE_CUDA && GOOGLE_TENSORRT diff --git a/tensorflow/compiler/tf2tensorrt/convert/logger_registry.cc b/tensorflow/compiler/tf2tensorrt/convert/logger_registry.cc index 82e68cbb28d..07c9c2f1ea0 100644 --- a/tensorflow/compiler/tf2tensorrt/convert/logger_registry.cc +++ b/tensorflow/compiler/tf2tensorrt/convert/logger_registry.cc @@ -12,8 +12,7 @@ WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the License for the specific language governing permissions and limitations under the License. ==============================================================================*/ -#if GOOGLE_CUDA -#if GOOGLE_TENSORRT +#if GOOGLE_CUDA && GOOGLE_TENSORRT #include "tensorflow/compiler/tf2tensorrt/convert/logger_registry.h" @@ -58,5 +57,4 @@ LoggerRegistry* GetLoggerRegistry() { } // namespace tensorrt } // namespace tensorflow -#endif // GOOGLE_TENSORRT -#endif // GOOGLE_CUDA +#endif // GOOGLE_CUDA && GOOGLE_TENSORRT diff --git a/tensorflow/compiler/tf2tensorrt/convert/logger_registry.h b/tensorflow/compiler/tf2tensorrt/convert/logger_registry.h index 45b302742d0..2a265cf7caa 100644 --- a/tensorflow/compiler/tf2tensorrt/convert/logger_registry.h +++ b/tensorflow/compiler/tf2tensorrt/convert/logger_registry.h @@ -19,7 +19,8 @@ limitations under the License. #include "tensorflow/core/platform/macros.h" #include "tensorflow/core/platform/types.h" -#if GOOGLE_CUDA +#if GOOGLE_CUDA && GOOGLE_TENSORRT + #include "third_party/tensorrt/NvInfer.h" namespace tensorflow { @@ -53,5 +54,5 @@ class RegisterLogger { } // namespace tensorrt } // namespace tensorflow -#endif // GOOGLE_CUDA +#endif // GOOGLE_CUDA && GOOGLE_TENSORRT #endif // TENSORFLOW_COMPILER_TF2TENSORRT_CONVERT_LOGGER_REGISTRY_H_ diff --git a/tensorflow/compiler/tf2tensorrt/convert/trt_optimization_pass.cc b/tensorflow/compiler/tf2tensorrt/convert/trt_optimization_pass.cc index 72f4fe5ef9b..1cf98d135cb 100644 --- a/tensorflow/compiler/tf2tensorrt/convert/trt_optimization_pass.cc +++ b/tensorflow/compiler/tf2tensorrt/convert/trt_optimization_pass.cc @@ -28,8 +28,7 @@ limitations under the License. #include "tensorflow/core/platform/logging.h" #include "tensorflow/core/platform/stacktrace.h" -#if GOOGLE_CUDA -#if GOOGLE_TENSORRT +#if GOOGLE_CUDA && GOOGLE_TENSORRT namespace tensorflow { namespace tensorrt { namespace convert { @@ -302,5 +301,4 @@ static VerboseCustomGraphOptimizerRegistrar TRTOptimizationPass_Registrar( } // namespace tensorrt } // namespace tensorflow -#endif -#endif +#endif // GOOGLE_CUDA && GOOGLE_TENSORRT diff --git a/tensorflow/compiler/tf2tensorrt/convert/trt_optimization_pass.h b/tensorflow/compiler/tf2tensorrt/convert/trt_optimization_pass.h index f79048bb5f6..e0aaa5500ab 100644 --- a/tensorflow/compiler/tf2tensorrt/convert/trt_optimization_pass.h +++ b/tensorflow/compiler/tf2tensorrt/convert/trt_optimization_pass.h @@ -23,8 +23,7 @@ limitations under the License. #include "tensorflow/core/grappler/optimizers/custom_graph_optimizer.h" #include "tensorflow/core/platform/logging.h" -#if GOOGLE_CUDA -#if GOOGLE_TENSORRT +#if GOOGLE_CUDA && GOOGLE_TENSORRT namespace tensorflow { namespace tensorrt { @@ -83,6 +82,5 @@ class TRTOptimizationPass : public grappler::CustomGraphOptimizer { } // namespace tensorrt } // namespace tensorflow -#endif // GOOGLE_CUDA -#endif // GOOGLE_TENSORRT +#endif // GOOGLE_CUDA && GOOGLE_TENSORRT #endif // TENSORFLOW_COMPILER_TF2TENSORRT_CONVERT_TRT_OPTIMIZATION_PASS_H_ diff --git a/tensorflow/compiler/tf2tensorrt/kernels/get_calibration_data_op.cc b/tensorflow/compiler/tf2tensorrt/kernels/get_calibration_data_op.cc index 3143b06817e..76fb40b9520 100644 --- a/tensorflow/compiler/tf2tensorrt/kernels/get_calibration_data_op.cc +++ b/tensorflow/compiler/tf2tensorrt/kernels/get_calibration_data_op.cc @@ -22,8 +22,7 @@ limitations under the License. #include "tensorflow/core/framework/resource_mgr.h" #include "tensorflow/core/lib/core/refcount.h" -#if GOOGLE_CUDA -#if GOOGLE_TENSORRT +#if GOOGLE_CUDA && GOOGLE_TENSORRT namespace tensorflow { namespace tensorrt { @@ -67,5 +66,4 @@ REGISTER_KERNEL_BUILDER(Name("GetCalibrationDataOp").Device(DEVICE_GPU), } // namespace tensorrt } // namespace tensorflow -#endif // GOOGLE_TENSORRT -#endif // GOOGLE_CUDA +#endif // GOOGLE_CUDA && GOOGLE_TENSORRT diff --git a/tensorflow/compiler/tf2tensorrt/kernels/trt_engine_op.cc b/tensorflow/compiler/tf2tensorrt/kernels/trt_engine_op.cc index 98d199ca9ab..1094555a622 100644 --- a/tensorflow/compiler/tf2tensorrt/kernels/trt_engine_op.cc +++ b/tensorflow/compiler/tf2tensorrt/kernels/trt_engine_op.cc @@ -48,8 +48,7 @@ limitations under the License. #include "tensorflow/core/util/env_var.h" #include "tensorflow/stream_executor/lib/statusor.h" -#if GOOGLE_CUDA -#if GOOGLE_TENSORRT +#if GOOGLE_CUDA && GOOGLE_TENSORRT #include "third_party/gpus/cuda/include/cuda_runtime_api.h" #include "third_party/tensorrt/NvInfer.h" @@ -1009,5 +1008,4 @@ REGISTER_KERNEL_BUILDER(Name("TRTEngineOp").Device(DEVICE_GPU), TRTEngineOp); } // namespace tensorrt } // namespace tensorflow -#endif // GOOGLE_TENSORRT -#endif // GOOGLE_CUDA +#endif // GOOGLE_CUDA && GOOGLE_TENSORRT diff --git a/tensorflow/compiler/tf2tensorrt/kernels/trt_engine_op_test.cc b/tensorflow/compiler/tf2tensorrt/kernels/trt_engine_op_test.cc index a06010de1c7..71193dc24cf 100644 --- a/tensorflow/compiler/tf2tensorrt/kernels/trt_engine_op_test.cc +++ b/tensorflow/compiler/tf2tensorrt/kernels/trt_engine_op_test.cc @@ -50,8 +50,7 @@ limitations under the License. #include "tensorflow/core/platform/status.h" #include "tensorflow/core/public/version.h" -#if GOOGLE_CUDA -#if GOOGLE_TENSORRT +#if GOOGLE_CUDA && GOOGLE_TENSORRT namespace tensorflow { namespace tensorrt { @@ -306,5 +305,4 @@ TYPED_TEST(TRTEngineOpTest, Basic) { } // namespace tensorrt } // namespace tensorflow -#endif // GOOGLE_TENSORRT -#endif // GOOGLE_CUDA +#endif // GOOGLE_CUDA && GOOGLE_TENSORRT diff --git a/tensorflow/compiler/tf2tensorrt/kernels/trt_engine_resource_ops.cc b/tensorflow/compiler/tf2tensorrt/kernels/trt_engine_resource_ops.cc index 2c5821df6ac..3b6e7e91d3b 100644 --- a/tensorflow/compiler/tf2tensorrt/kernels/trt_engine_resource_ops.cc +++ b/tensorflow/compiler/tf2tensorrt/kernels/trt_engine_resource_ops.cc @@ -33,8 +33,7 @@ limitations under the License. #include "tensorflow/core/platform/mutex.h" #include "tensorflow/core/platform/thread_annotations.h" -#if GOOGLE_CUDA -#if GOOGLE_TENSORRT +#if GOOGLE_CUDA && GOOGLE_TENSORRT #include "third_party/tensorrt/NvInfer.h" namespace tensorflow { @@ -251,5 +250,4 @@ REGISTER_KERNEL_BUILDER(Name("SerializeTRTResource").Device(DEVICE_GPU), } // namespace tensorrt } // namespace tensorflow -#endif // GOOGLE_TENSORRT -#endif // GOOGLE_CUDA +#endif // GOOGLE_CUDA && GOOGLE_TENSORRT diff --git a/tensorflow/compiler/tf2tensorrt/kernels/trt_engine_resource_ops_test.cc b/tensorflow/compiler/tf2tensorrt/kernels/trt_engine_resource_ops_test.cc index 4a24160569d..6a073ee24d0 100644 --- a/tensorflow/compiler/tf2tensorrt/kernels/trt_engine_resource_ops_test.cc +++ b/tensorflow/compiler/tf2tensorrt/kernels/trt_engine_resource_ops_test.cc @@ -48,8 +48,7 @@ limitations under the License. #include "tensorflow/core/platform/tstring.h" #include "tensorflow/core/platform/types.h" -#if GOOGLE_CUDA -#if GOOGLE_TENSORRT +#if GOOGLE_CUDA && GOOGLE_TENSORRT namespace tensorflow { namespace tensorrt { @@ -246,5 +245,4 @@ TEST_F(TRTEngineResourceOpsTest, Basic) { } // namespace tensorrt } // namespace tensorflow -#endif // GOOGLE_TENSORRT -#endif // GOOGLE_CUDA +#endif // GOOGLE_CUDA && GOOGLE_TENSORRT diff --git a/tensorflow/compiler/tf2tensorrt/ops/get_calibration_data_op.cc b/tensorflow/compiler/tf2tensorrt/ops/get_calibration_data_op.cc index 573172b92e6..2af3164c3e2 100644 --- a/tensorflow/compiler/tf2tensorrt/ops/get_calibration_data_op.cc +++ b/tensorflow/compiler/tf2tensorrt/ops/get_calibration_data_op.cc @@ -13,8 +13,7 @@ See the License for the specific language governing permissions and limitations under the License. ==============================================================================*/ -#if GOOGLE_CUDA -#if GOOGLE_TENSORRT +#if GOOGLE_CUDA && GOOGLE_TENSORRT #include "tensorflow/core/framework/common_shape_fns.h" #include "tensorflow/core/framework/op.h" @@ -34,5 +33,4 @@ Returns calibration data for the given resource name } // namespace tensorflow -#endif // GOOGLE_TENSORRT -#endif // GOOGLE_CUDA +#endif // GOOGLE_CUDA && GOOGLE_TENSORRT diff --git a/tensorflow/compiler/tf2tensorrt/ops/trt_engine_op.cc b/tensorflow/compiler/tf2tensorrt/ops/trt_engine_op.cc index bd3c2b299a9..2527fe9b910 100644 --- a/tensorflow/compiler/tf2tensorrt/ops/trt_engine_op.cc +++ b/tensorflow/compiler/tf2tensorrt/ops/trt_engine_op.cc @@ -13,8 +13,7 @@ See the License for the specific language governing permissions and limitations under the License. ==============================================================================*/ -#if GOOGLE_CUDA -#if GOOGLE_TENSORRT +#if GOOGLE_CUDA && GOOGLE_TENSORRT #include "tensorflow/core/framework/common_shape_fns.h" #include "tensorflow/core/framework/op.h" @@ -59,5 +58,4 @@ REGISTER_OP("TRTEngineOp") .Attr("static_engine: bool = true"); } // namespace tensorflow -#endif // GOOGLE_TENSORRT -#endif // GOOGLE_CUDA +#endif // GOOGLE_CUDA && GOOGLE_TENSORRT diff --git a/tensorflow/compiler/tf2tensorrt/ops/trt_engine_resource_ops.cc b/tensorflow/compiler/tf2tensorrt/ops/trt_engine_resource_ops.cc index 01911de66ec..3141092de03 100644 --- a/tensorflow/compiler/tf2tensorrt/ops/trt_engine_resource_ops.cc +++ b/tensorflow/compiler/tf2tensorrt/ops/trt_engine_resource_ops.cc @@ -13,8 +13,7 @@ See the License for the specific language governing permissions and limitations under the License. ==============================================================================*/ -#if GOOGLE_CUDA -#if GOOGLE_TENSORRT +#if GOOGLE_CUDA && GOOGLE_TENSORRT #include "tensorflow/core/framework/common_shape_fns.h" #include "tensorflow/core/framework/op.h" @@ -46,5 +45,4 @@ REGISTER_OP("SerializeTRTResource") } // namespace tensorflow -#endif // GOOGLE_TENSORRT -#endif // GOOGLE_CUDA +#endif // GOOGLE_CUDA && GOOGLE_TENSORRT diff --git a/tensorflow/compiler/tf2tensorrt/plugin/plugin_cast.cu.cc b/tensorflow/compiler/tf2tensorrt/plugin/plugin_cast.cu.cc index 4c0d8b0392a..141a7d1f462 100644 --- a/tensorflow/compiler/tf2tensorrt/plugin/plugin_cast.cu.cc +++ b/tensorflow/compiler/tf2tensorrt/plugin/plugin_cast.cu.cc @@ -17,8 +17,7 @@ limitations under the License. #include "tensorflow/compiler/tf2tensorrt/plugin/trt_plugin.h" #include "tensorflow/core/platform/logging.h" -#if GOOGLE_CUDA -#if GOOGLE_TENSORRT +#if GOOGLE_CUDA && GOOGLE_TENSORRT #define EIGEN_USE_GPU // For definition of Eigen::GpuDevice. #include "third_party/gpus/cuda/include/cuda_runtime_api.h" #include "tensorflow/core/util/gpu_kernel_helper.h" @@ -234,5 +233,4 @@ REGISTER_TFTRT_PLUGIN(CastPluginCreator); } // namespace tensorrt } // namespace tensorflow -#endif // GOOGLE_CUDA -#endif // GOOGLE_TENSORRT +#endif // GOOGLE_CUDA && GOOGLE_TENSORRT diff --git a/tensorflow/compiler/tf2tensorrt/plugin/trt_plugin.cc b/tensorflow/compiler/tf2tensorrt/plugin/trt_plugin.cc index 563ce724f43..83d5f9b5965 100644 --- a/tensorflow/compiler/tf2tensorrt/plugin/trt_plugin.cc +++ b/tensorflow/compiler/tf2tensorrt/plugin/trt_plugin.cc @@ -17,8 +17,7 @@ limitations under the License. #include -#if GOOGLE_CUDA -#if GOOGLE_TENSORRT +#if GOOGLE_CUDA && GOOGLE_TENSORRT namespace tensorflow { namespace tensorrt { @@ -30,5 +29,4 @@ const char* kTfTrtPluginNamespace = "TF"; } // namespace tensorrt } // namespace tensorflow -#endif // GOOGLE_CUDA -#endif // GOOGLE_TENSORRT +#endif // GOOGLE_CUDA && GOOGLE_TENSORRT diff --git a/tensorflow/compiler/tf2tensorrt/plugin/trt_plugin.h b/tensorflow/compiler/tf2tensorrt/plugin/trt_plugin.h index bdb046e6c71..600ac6683da 100644 --- a/tensorflow/compiler/tf2tensorrt/plugin/trt_plugin.h +++ b/tensorflow/compiler/tf2tensorrt/plugin/trt_plugin.h @@ -20,8 +20,7 @@ limitations under the License. #include "tensorflow/core/platform/logging.h" -#if GOOGLE_CUDA -#if GOOGLE_TENSORRT +#if GOOGLE_CUDA && GOOGLE_TENSORRT #include "third_party/tensorrt/NvInfer.h" namespace tensorflow { @@ -90,7 +89,6 @@ class TrtPluginRegistrar { } // namespace tensorrt } // namespace tensorflow -#endif // GOOGLE_TENSORRT -#endif // GOOGLE_CUDA +#endif // GOOGLE_CUDA && GOOGLE_TENSORRT #endif // TENSORFLOW_COMPILER_TF2TENSORRT_PLUGIN_TRT_PLUGIN_H_ diff --git a/tensorflow/compiler/tf2tensorrt/segment/segment.cc b/tensorflow/compiler/tf2tensorrt/segment/segment.cc index 32e30006f58..d9080b6f69a 100644 --- a/tensorflow/compiler/tf2tensorrt/segment/segment.cc +++ b/tensorflow/compiler/tf2tensorrt/segment/segment.cc @@ -35,8 +35,7 @@ limitations under the License. #include "tensorflow/core/platform/types.h" #include "tensorflow/core/util/env_var.h" -#if GOOGLE_CUDA -#if GOOGLE_TENSORRT +#if GOOGLE_CUDA && GOOGLE_TENSORRT namespace tensorflow { namespace tensorrt { @@ -1062,5 +1061,4 @@ Status SegmentGraph(const Graph* tf_graph, } // namespace tensorrt } // namespace tensorflow -#endif // GOOGLE_TENSORRT -#endif // GOOGLE_CUDA +#endif // GOOGLE_CUDA && GOOGLE_TENSORRT diff --git a/tensorflow/compiler/tf2tensorrt/segment/segment.h b/tensorflow/compiler/tf2tensorrt/segment/segment.h index 7295c8f0d9d..3f79983cfd2 100644 --- a/tensorflow/compiler/tf2tensorrt/segment/segment.h +++ b/tensorflow/compiler/tf2tensorrt/segment/segment.h @@ -25,8 +25,7 @@ limitations under the License. #include "tensorflow/core/lib/core/status.h" #include "tensorflow/core/platform/types.h" -#if GOOGLE_CUDA -#if GOOGLE_TENSORRT +#if GOOGLE_CUDA && GOOGLE_TENSORRT namespace tensorflow { namespace tensorrt { @@ -67,7 +66,6 @@ Status SegmentGraph(const Graph* tf_graph, } // namespace tensorrt } // namespace tensorflow -#endif // GOOGLE_TENSORRT -#endif // GOOGLE_CUDA +#endif // GOOGLE_CUDA && GOOGLE_TENSORRT #endif // TENSORFLOW_COMPILER_TF2TENSORRT_SEGMENT_SEGMENT_H_ diff --git a/tensorflow/compiler/tf2tensorrt/segment/segment_test.cc b/tensorflow/compiler/tf2tensorrt/segment/segment_test.cc index 2437481a9c4..f3bc5bfbee6 100644 --- a/tensorflow/compiler/tf2tensorrt/segment/segment_test.cc +++ b/tensorflow/compiler/tf2tensorrt/segment/segment_test.cc @@ -26,8 +26,7 @@ limitations under the License. #include "tensorflow/core/platform/types.h" #include "tensorflow/core/public/session.h" -#if GOOGLE_CUDA -#if GOOGLE_TENSORRT +#if GOOGLE_CUDA && GOOGLE_TENSORRT namespace tensorflow { namespace tensorrt { @@ -522,5 +521,4 @@ TEST_F(SegmentTest, IncompatibleBatchSizes) { } // namespace tensorrt } // namespace tensorflow -#endif // GOOGLE_TENSORRT -#endif // GOOGLE_CUDA +#endif // GOOGLE_CUDA && GOOGLE_TENSORRT diff --git a/tensorflow/compiler/tf2tensorrt/segment/union_find.h b/tensorflow/compiler/tf2tensorrt/segment/union_find.h index 70e83c12fca..b53615ec019 100644 --- a/tensorflow/compiler/tf2tensorrt/segment/union_find.h +++ b/tensorflow/compiler/tf2tensorrt/segment/union_find.h @@ -19,8 +19,7 @@ limitations under the License. #include "absl/strings/str_format.h" #include "absl/types/optional.h" -#if GOOGLE_CUDA -#if GOOGLE_TENSORRT +#if GOOGLE_CUDA && GOOGLE_TENSORRT namespace tensorflow { namespace tensorrt { @@ -217,7 +216,6 @@ UnionFind* UnionFind::FindRoot() { } // namespace tensorrt } // namespace tensorflow -#endif // GOOGLE_TENSORRT -#endif // GOOGLE_CUDA +#endif // GOOGLE_CUDA && GOOGLE_TENSORRT #endif // TENSORFLOW_COMPILER_TF2TENSORRT_SEGMENT_UNION_FIND_H_ diff --git a/tensorflow/compiler/tf2tensorrt/tensorrt_test.cc b/tensorflow/compiler/tf2tensorrt/tensorrt_test.cc index 510591bfe00..e994d20df33 100644 --- a/tensorflow/compiler/tf2tensorrt/tensorrt_test.cc +++ b/tensorflow/compiler/tf2tensorrt/tensorrt_test.cc @@ -18,8 +18,7 @@ limitations under the License. #include "tensorflow/core/platform/stream_executor.h" #include "tensorflow/core/platform/test.h" -#if GOOGLE_CUDA -#if GOOGLE_TENSORRT +#if GOOGLE_CUDA && GOOGLE_TENSORRT #include "third_party/gpus/cuda/include/cuda.h" #include "third_party/gpus/cuda/include/cuda_runtime_api.h" #include "third_party/tensorrt/NvInfer.h" @@ -164,5 +163,4 @@ TEST(TensorrtTest, BasicFunctions) { } // namespace } // namespace tensorflow -#endif // GOOGLE_TENSORRT -#endif // GOOGLE_CUDA +#endif // GOOGLE_CUDA && GOOGLE_TENSORRT diff --git a/tensorflow/compiler/tf2tensorrt/utils/trt_allocator.cc b/tensorflow/compiler/tf2tensorrt/utils/trt_allocator.cc index 617ea7fad5c..d4f3a524577 100644 --- a/tensorflow/compiler/tf2tensorrt/utils/trt_allocator.cc +++ b/tensorflow/compiler/tf2tensorrt/utils/trt_allocator.cc @@ -17,11 +17,9 @@ limitations under the License. #include "tensorflow/core/platform/logging.h" -#if GOOGLE_CUDA -#if GOOGLE_TENSORRT +#if GOOGLE_CUDA && GOOGLE_TENSORRT #include "third_party/gpus/cuda/include/cuda_runtime_api.h" -#endif // GOOGLE_TENSORRT -#endif // GOOGLE_CUDA +#endif // GOOGLE_CUDA && GOOGLE_TENSORRT namespace tensorflow { namespace tensorrt { @@ -52,8 +50,7 @@ void* Align(uint64_t alignment, uint64_t size, void*& ptr, uint64_t& space) { } // namespace tensorrt } // namespace tensorflow -#if GOOGLE_CUDA -#if GOOGLE_TENSORRT +#if GOOGLE_CUDA && GOOGLE_TENSORRT namespace tensorflow { namespace tensorrt { @@ -113,5 +110,4 @@ void TRTDeviceAllocator::free(void* memory) { } // namespace tensorrt } // namespace tensorflow -#endif // GOOGLE_TENSORRT -#endif // GOOGLE_CUDA +#endif // GOOGLE_CUDA && GOOGLE_TENSORRT diff --git a/tensorflow/compiler/tf2tensorrt/utils/trt_allocator.h b/tensorflow/compiler/tf2tensorrt/utils/trt_allocator.h index 4ab8b52f523..d219a8a14e8 100644 --- a/tensorflow/compiler/tf2tensorrt/utils/trt_allocator.h +++ b/tensorflow/compiler/tf2tensorrt/utils/trt_allocator.h @@ -20,11 +20,9 @@ limitations under the License. #include "tensorflow/core/framework/allocator.h" -#if GOOGLE_CUDA -#if GOOGLE_TENSORRT +#if GOOGLE_CUDA && GOOGLE_TENSORRT #include "third_party/tensorrt/NvInfer.h" -#endif // GOOGLE_TENSORRT -#endif // GOOGLE_CUDA +#endif // GOOGLE_CUDA && GOOGLE_TENSORRT namespace tensorflow { namespace tensorrt { @@ -33,8 +31,7 @@ void* Align(uint64_t alignment, uint64_t size, void*& ptr, uint64_t& space); } // namespace tensorrt } // namespace tensorflow -#if GOOGLE_CUDA -#if GOOGLE_TENSORRT +#if GOOGLE_CUDA && GOOGLE_TENSORRT namespace tensorflow { namespace tensorrt { @@ -69,6 +66,5 @@ class TRTDeviceAllocator : public TRTBaseAllocator { } // namespace tensorrt } // namespace tensorflow -#endif // GOOGLE_TENSORRT -#endif // GOOGLE_CUDA +#endif // GOOGLE_CUDA && GOOGLE_TENSORRT #endif // TENSORFLOW_COMPILER_TF2TENSORRT_UTILS_TRT_ALLOCATOR_H_ diff --git a/tensorflow/compiler/tf2tensorrt/utils/trt_engine_utils.cc b/tensorflow/compiler/tf2tensorrt/utils/trt_engine_utils.cc index ed997b267b1..8ccfb8b06f0 100644 --- a/tensorflow/compiler/tf2tensorrt/utils/trt_engine_utils.cc +++ b/tensorflow/compiler/tf2tensorrt/utils/trt_engine_utils.cc @@ -25,8 +25,7 @@ limitations under the License. #include "tensorflow/core/lib/core/status.h" #include "tensorflow/core/platform/errors.h" -#if GOOGLE_CUDA -#if GOOGLE_TENSORRT +#if GOOGLE_CUDA && GOOGLE_TENSORRT #include "third_party/tensorrt/NvInfer.h" namespace tensorflow { @@ -257,5 +256,4 @@ Status TrtEnqueue(nvinfer1::IExecutionContext* execution_context, } // namespace tensorrt } // namespace tensorflow -#endif // GOOGLE_TENSORRT -#endif // GOOGLE_CUDA +#endif // GOOGLE_CUDA && GOOGLE_TENSORRT diff --git a/tensorflow/compiler/tf2tensorrt/utils/trt_engine_utils.h b/tensorflow/compiler/tf2tensorrt/utils/trt_engine_utils.h index a471749877a..1ea4fe28cb4 100644 --- a/tensorflow/compiler/tf2tensorrt/utils/trt_engine_utils.h +++ b/tensorflow/compiler/tf2tensorrt/utils/trt_engine_utils.h @@ -24,8 +24,7 @@ limitations under the License. #include "tensorflow/core/framework/tensor_shape.h" #include "tensorflow/core/lib/core/status.h" -#if GOOGLE_CUDA -#if GOOGLE_TENSORRT +#if GOOGLE_CUDA && GOOGLE_TENSORRT #include "third_party/tensorrt/NvInfer.h" namespace tensorflow { @@ -91,7 +90,6 @@ Status TrtEnqueue(nvinfer1::IExecutionContext* execution_context, } // namespace tensorrt } // namespace tensorflow -#endif // GOOGLE_TENSORRT -#endif // GOOGLE_CUDA +#endif // GOOGLE_CUDA && GOOGLE_TENSORRT #endif // TENSORFLOW_COMPILER_TF2TENSORRT_UTILS_TRT_ENGINE_UTILS_H_ diff --git a/tensorflow/compiler/tf2tensorrt/utils/trt_int8_calibrator.cc b/tensorflow/compiler/tf2tensorrt/utils/trt_int8_calibrator.cc index 554c127fa37..24271e352a7 100644 --- a/tensorflow/compiler/tf2tensorrt/utils/trt_int8_calibrator.cc +++ b/tensorflow/compiler/tf2tensorrt/utils/trt_int8_calibrator.cc @@ -20,8 +20,7 @@ limitations under the License. #include "tensorflow/core/platform/logging.h" -#if GOOGLE_CUDA -#if GOOGLE_TENSORRT +#if GOOGLE_CUDA && GOOGLE_TENSORRT #include "third_party/gpus/cuda/include/cuda_runtime_api.h" namespace tensorflow { @@ -147,5 +146,4 @@ TRTInt8Calibrator::~TRTInt8Calibrator() { } // namespace tensorrt } // namespace tensorflow -#endif -#endif +#endif // GOOGLE_CUDA && GOOGLE_TENSORRT diff --git a/tensorflow/compiler/tf2tensorrt/utils/trt_int8_calibrator.h b/tensorflow/compiler/tf2tensorrt/utils/trt_int8_calibrator.h index 06b39716490..4c670e85f52 100644 --- a/tensorflow/compiler/tf2tensorrt/utils/trt_int8_calibrator.h +++ b/tensorflow/compiler/tf2tensorrt/utils/trt_int8_calibrator.h @@ -22,8 +22,7 @@ limitations under the License. #include #include "tensorflow/core/platform/mutex.h" -#if GOOGLE_CUDA -#if GOOGLE_TENSORRT +#if GOOGLE_CUDA && GOOGLE_TENSORRT #include "third_party/gpus/cuda/include/cuda_runtime_api.h" #include "third_party/tensorrt/NvInfer.h" @@ -101,6 +100,5 @@ struct TRTInt8Calibrator : public nvinfer1::IInt8EntropyCalibrator { } // namespace tensorrt } // namespace tensorflow -#endif -#endif +#endif // GOOGLE_CUDA && GOOGLE_TENSORRT #endif // TENSORFLOW_COMPILER_TF2TENSORRT_UTILS_TRT_INT8_CALIBRATOR_H_ diff --git a/tensorflow/compiler/tf2tensorrt/utils/trt_logger.cc b/tensorflow/compiler/tf2tensorrt/utils/trt_logger.cc index 193687ebc8c..e34bf5e7397 100644 --- a/tensorflow/compiler/tf2tensorrt/utils/trt_logger.cc +++ b/tensorflow/compiler/tf2tensorrt/utils/trt_logger.cc @@ -15,8 +15,7 @@ limitations under the License. #include "tensorflow/compiler/tf2tensorrt/utils/trt_logger.h" -#if GOOGLE_CUDA -#if GOOGLE_TENSORRT +#if GOOGLE_CUDA && GOOGLE_TENSORRT #include "tensorflow/compiler/tf2tensorrt/common/utils.h" #include "tensorflow/compiler/tf2tensorrt/convert/logger_registry.h" #include "tensorflow/core/platform/logging.h" @@ -68,5 +67,4 @@ REGISTER_TENSORRT_LOGGER("DefaultLogger", Logger::GetLogger()); } // namespace tensorrt } // namespace tensorflow -#endif // GOOGLE_CUDA -#endif // GOOGLE_TENSORRT +#endif // GOOGLE_CUDA && GOOGLE_TENSORRT diff --git a/tensorflow/compiler/tf2tensorrt/utils/trt_logger.h b/tensorflow/compiler/tf2tensorrt/utils/trt_logger.h index 2ade1b48f47..ce6552e8fe9 100644 --- a/tensorflow/compiler/tf2tensorrt/utils/trt_logger.h +++ b/tensorflow/compiler/tf2tensorrt/utils/trt_logger.h @@ -18,8 +18,7 @@ limitations under the License. #include "tensorflow/core/platform/types.h" -#if GOOGLE_CUDA -#if GOOGLE_TENSORRT +#if GOOGLE_CUDA && GOOGLE_TENSORRT #include "third_party/tensorrt/NvInfer.h" namespace tensorflow { @@ -40,7 +39,6 @@ class Logger : public nvinfer1::ILogger { } // namespace tensorrt } // namespace tensorflow -#endif // GOOGLE_TENSORRT -#endif // GOOGLE_CUDA +#endif // GOOGLE_CUDA && GOOGLE_TENSORRT #endif // TENSORFLOW_COMPILER_TF2TENSORRT_UTILS_TRT_LOGGER_H_ diff --git a/tensorflow/compiler/tf2tensorrt/utils/trt_lru_cache.cc b/tensorflow/compiler/tf2tensorrt/utils/trt_lru_cache.cc index fbcdaad52c0..ee7e6272372 100644 --- a/tensorflow/compiler/tf2tensorrt/utils/trt_lru_cache.cc +++ b/tensorflow/compiler/tf2tensorrt/utils/trt_lru_cache.cc @@ -23,8 +23,7 @@ limitations under the License. #include "tensorflow/core/framework/tensor_shape.h" #include "tensorflow/core/platform/mutex.h" -#if GOOGLE_CUDA -#if GOOGLE_TENSORRT +#if GOOGLE_CUDA && GOOGLE_TENSORRT #include "third_party/tensorrt/NvInfer.h" namespace tensorflow { @@ -141,5 +140,4 @@ EngineContext* TRTEngineCacheResource::GetEngineContext(const int profile_id) { } // namespace tensorrt } // namespace tensorflow -#endif // GOOGLE_TENSORRT -#endif // GOOGLE_CUDA +#endif // GOOGLE_CUDA && GOOGLE_TENSORRT diff --git a/tensorflow/compiler/tf2tensorrt/utils/trt_lru_cache.h b/tensorflow/compiler/tf2tensorrt/utils/trt_lru_cache.h index 8e345254f75..991b9a949e4 100644 --- a/tensorflow/compiler/tf2tensorrt/utils/trt_lru_cache.h +++ b/tensorflow/compiler/tf2tensorrt/utils/trt_lru_cache.h @@ -115,8 +115,7 @@ class LRUCache { } }; -#if GOOGLE_CUDA -#if GOOGLE_TENSORRT +#if GOOGLE_CUDA && GOOGLE_TENSORRT struct EngineContext { EngineContext() {} // Creates an empty context. @@ -223,8 +222,7 @@ class TRTEngineCacheResource : public ResourceBase { TrtShapeOptimizationProfile profiles_; }; -#endif // GOOGLE_TENSORRT -#endif // GOOGLE_CUDA +#endif // GOOGLE_CUDA && GOOGLE_TENSORRT } // namespace tensorrt } // namespace tensorflow diff --git a/tensorflow/compiler/tf2tensorrt/utils/trt_shape_optimization_profiles.h b/tensorflow/compiler/tf2tensorrt/utils/trt_shape_optimization_profiles.h index 40c7f5dcf31..fc688b14139 100644 --- a/tensorflow/compiler/tf2tensorrt/utils/trt_shape_optimization_profiles.h +++ b/tensorflow/compiler/tf2tensorrt/utils/trt_shape_optimization_profiles.h @@ -29,8 +29,7 @@ limitations under the License. #include "tensorflow/core/lib/strings/str_util.h" #include "tensorflow/core/lib/strings/strcat.h" -#if GOOGLE_CUDA -#if GOOGLE_TENSORRT +#if GOOGLE_CUDA && GOOGLE_TENSORRT #include "third_party/tensorrt/NvInfer.h" @@ -173,6 +172,5 @@ class TrtShapeOptimizationProfile { } // namespace tensorrt } // namespace tensorflow -#endif // GOOGLE_TENSORRT -#endif // GOOGLE_CUDA +#endif // GOOGLE_CUDA && GOOGLE_TENSORRT #endif // TENSORFLOW_COMPILER_TF2TENSORRT_UTILS_TRT_SHAPE_OPTIMIZATION_PROFILES_H_ diff --git a/tensorflow/compiler/tf2tensorrt/utils/trt_shape_optimization_profiles_test.cc b/tensorflow/compiler/tf2tensorrt/utils/trt_shape_optimization_profiles_test.cc index 501810587e0..32c2200fb71 100644 --- a/tensorflow/compiler/tf2tensorrt/utils/trt_shape_optimization_profiles_test.cc +++ b/tensorflow/compiler/tf2tensorrt/utils/trt_shape_optimization_profiles_test.cc @@ -13,8 +13,7 @@ See the License for the specific language governing permissions and limitations under the License. ==============================================================================*/ -#if GOOGLE_CUDA -#if GOOGLE_TENSORRT +#if GOOGLE_CUDA && GOOGLE_TENSORRT #include @@ -214,5 +213,4 @@ TEST_F(TrtShapeOptimizationProfileTest, Dynamic) { } // namespace tensorrt } // namespace tensorflow -#endif // GOOGLE_TENSORRT -#endif // GOOGLE_CUDA +#endif // GOOGLE_CUDA && GOOGLE_TENSORRT From d8e73a2f303ddebb7da3694fe7ae20a752eea01b Mon Sep 17 00:00:00 2001 From: Jiho Choi Date: Mon, 22 Jun 2020 19:16:42 -0700 Subject: [PATCH 516/867] Add element tracing for prefetch. PiperOrigin-RevId: 317782986 Change-Id: Ia6e6331a59559423f5a0eac76bf2e7ffd677f876 --- tensorflow/core/kernels/data/BUILD | 2 ++ .../core/kernels/data/prefetch_dataset_op.cc | 23 +++++++++++++++++-- 2 files changed, 23 insertions(+), 2 deletions(-) diff --git a/tensorflow/core/kernels/data/BUILD b/tensorflow/core/kernels/data/BUILD index 6d0351202df..0972dc83ccf 100644 --- a/tensorflow/core/kernels/data/BUILD +++ b/tensorflow/core/kernels/data/BUILD @@ -674,6 +674,8 @@ tf_kernel_library( "//tensorflow/core:lib", "//tensorflow/core:lib_internal", "//tensorflow/core:protos_all_cc", + "//tensorflow/core/profiler/lib:traceme", + "//tensorflow/core/profiler/lib:traceme_encode", ], ) diff --git a/tensorflow/core/kernels/data/prefetch_dataset_op.cc b/tensorflow/core/kernels/data/prefetch_dataset_op.cc index 0230bcd146d..20b78ba14ad 100644 --- a/tensorflow/core/kernels/data/prefetch_dataset_op.cc +++ b/tensorflow/core/kernels/data/prefetch_dataset_op.cc @@ -28,6 +28,8 @@ limitations under the License. #include "tensorflow/core/lib/strings/str_util.h" #include "tensorflow/core/lib/strings/stringprintf.h" #include "tensorflow/core/platform/stringprintf.h" +#include "tensorflow/core/profiler/lib/traceme.h" +#include "tensorflow/core/profiler/lib/traceme_encode.h" #include "tensorflow/core/protobuf/error_codes.pb.h" namespace tensorflow { @@ -303,6 +305,7 @@ class PrefetchDatasetOp::Dataset : public DatasetBase { // The buffered data element. std::vector value; int64 created_us; + int64 id; }; int64 buffer_limit() const TF_EXCLUSIVE_LOCKS_REQUIRED(*mu_) { @@ -339,6 +342,13 @@ class PrefetchDatasetOp::Dataset : public DatasetBase { // (if we successfully got an element) the output values. Status s = buffer_.front().status; if (s.ok()) { + int64 buffer_element_id = buffer_.front().id; + profiler::TraceMe traceme( + [&] { + return profiler::TraceMeEncode( + "PrefetchConsume", {{"element_id", buffer_element_id}}); + }, + profiler::kInfo); if (dataset()->slack_period_ > 0 && (num_elements() + 1) % dataset()->slack_period_ == 0) { // TODO(rachelim): Consider doing something more sophisticated @@ -423,8 +433,16 @@ class PrefetchDatasetOp::Dataset : public DatasetBase { mutex_lock input_l(input_mu_); bool end_of_sequence; BufferElement buffer_element; - buffer_element.status = input_impl_->GetNext( - ctx.get(), &buffer_element.value, &end_of_sequence); + { + profiler::TraceMe traceme( + [&] { + return profiler::TraceMeEncode("PrefetchProduce", + {{"element_id", num_produced}}); + }, + profiler::kInfo); + buffer_element.status = input_impl_->GetNext( + ctx.get(), &buffer_element.value, &end_of_sequence); + } if (buffer_element.status.ok() && end_of_sequence) { mutex_lock l(*mu_); prefetch_thread_finished_ = true; @@ -437,6 +455,7 @@ class PrefetchDatasetOp::Dataset : public DatasetBase { mutex_lock l(*mu_); RecordBufferEnqueue(ctx.get(), buffer_element.value); buffer_element.created_us = EnvTime::NowMicros(); + buffer_element.id = num_produced; buffer_.push_back(std::move(buffer_element)); cond_var_->notify_all(); } From 90078a19ee1ce5c7bc644bac1ceff7c16d12e4bb Mon Sep 17 00:00:00 2001 From: David Rim Date: Mon, 22 Jun 2020 19:16:48 -0700 Subject: [PATCH 517/867] Update quantize weights to use hybrid per-channel quantization and asymmetric quantization by default PiperOrigin-RevId: 317783008 Change-Id: I3c0f026ded030407c434bca70fe15d5fe723c744 --- tensorflow/lite/kernels/register.cc | 18 ++-- .../lite/tools/optimize/quantize_weights.cc | 92 ++++++++++++++++--- .../tools/optimize/quantize_weights_test.cc | 2 + .../lite/tools/versioning/op_version.cc | 63 ++++++++++++- tensorflow/lite/tools/versioning/op_version.h | 9 ++ .../lite/tools/versioning/op_version_test.cc | 27 ++++++ .../lite/tools/versioning/runtime_version.cc | 12 +++ 7 files changed, 195 insertions(+), 28 deletions(-) diff --git a/tensorflow/lite/kernels/register.cc b/tensorflow/lite/kernels/register.cc index c3a4aaad16d..452ce35ec78 100644 --- a/tensorflow/lite/kernels/register.cc +++ b/tensorflow/lite/kernels/register.cc @@ -54,24 +54,24 @@ BuiltinOpResolver::BuiltinOpResolver() { AddBuiltin(BuiltinOperator_L2_POOL_2D, Register_L2_POOL_2D()); AddBuiltin(BuiltinOperator_CONV_2D, Register_CONV_2D(), /* min_version = */ 1, - /* max_version = */ 4); + /* max_version = */ 5); AddBuiltin(BuiltinOperator_DEPTHWISE_CONV_2D, Register_DEPTHWISE_CONV_2D(), /* min_version = */ 1, - /* max_version = */ 5); + /* max_version = */ 6); AddBuiltin(BuiltinOperator_SVDF, Register_SVDF(), /* min_version = */ 1, - /* max_version = */ 3); + /* max_version = */ 4); AddBuiltin(BuiltinOperator_RNN, Register_RNN(), /* min_version = */ 1, - /* max_version = */ 2); + /* max_version = */ 3); AddBuiltin(BuiltinOperator_BIDIRECTIONAL_SEQUENCE_RNN, Register_BIDIRECTIONAL_SEQUENCE_RNN(), /* min_version = */ 1, - /* max_version = */ 2); + /* max_version = */ 3); AddBuiltin(BuiltinOperator_UNIDIRECTIONAL_SEQUENCE_RNN, Register_UNIDIRECTIONAL_SEQUENCE_RNN(), /* min_version = */ 1, - /* max_version = */ 2); + /* max_version = */ 3); AddBuiltin(BuiltinOperator_EMBEDDING_LOOKUP, Register_EMBEDDING_LOOKUP(), /* min_version = */ 1, /* max_version = */ 3); @@ -79,7 +79,7 @@ BuiltinOpResolver::BuiltinOpResolver() { Register_EMBEDDING_LOOKUP_SPARSE()); AddBuiltin(BuiltinOperator_FULLY_CONNECTED, Register_FULLY_CONNECTED(), /* min_version = */ 1, - /* max_version = */ 8); + /* max_version = */ 9); AddBuiltin(BuiltinOperator_LSH_PROJECTION, Register_LSH_PROJECTION()); AddBuiltin(BuiltinOperator_HASHTABLE_LOOKUP, Register_HASHTABLE_LOOKUP()); AddBuiltin(BuiltinOperator_SOFTMAX, Register_SOFTMAX(), @@ -105,13 +105,13 @@ BuiltinOpResolver::BuiltinOpResolver() { AddBuiltin(BuiltinOperator_LOCAL_RESPONSE_NORMALIZATION, Register_LOCAL_RESPONSE_NORMALIZATION()); AddBuiltin(BuiltinOperator_LSTM, Register_LSTM(), /* min_version = */ 1, - /* max_version = */ 3); + /* max_version = */ 4); AddBuiltin(BuiltinOperator_BIDIRECTIONAL_SEQUENCE_LSTM, Register_BIDIRECTIONAL_SEQUENCE_LSTM(), /* min_version = */ 1, /* max_version = */ 3); AddBuiltin(BuiltinOperator_UNIDIRECTIONAL_SEQUENCE_LSTM, Register_UNIDIRECTIONAL_SEQUENCE_LSTM(), /* min_version = */ 1, - /* max_version = */ 2); + /* max_version = */ 3); AddBuiltin(BuiltinOperator_PAD, Register_PAD(), /* min_version = */ 1, /* max_version = */ 2); AddBuiltin(BuiltinOperator_PADV2, Register_PADV2(), /* min_version = */ 1, diff --git a/tensorflow/lite/tools/optimize/quantize_weights.cc b/tensorflow/lite/tools/optimize/quantize_weights.cc index 7e3853c645c..8bef019a83e 100644 --- a/tensorflow/lite/tools/optimize/quantize_weights.cc +++ b/tensorflow/lite/tools/optimize/quantize_weights.cc @@ -43,6 +43,12 @@ typedef struct { int32_t op_input_idx; } ConsumerOpInfo; +typedef struct { + TensorT* t; + bool is_per_channel; + int channel_dim; +} TensorPerChannel; + // The default minimum number of elements a weights array must have to be // quantized by this transformation. const int kWeightsMinNumElementsDefault = 1024; @@ -138,6 +144,7 @@ bool IsHybridEvaluationOp(const OperatorT* op, const OperatorCodeT* op_code, } } else if (builtin_op_code == BuiltinOperator_FULLY_CONNECTED || builtin_op_code == BuiltinOperator_CONV_2D || + builtin_op_code == BuiltinOperator_DEPTHWISE_CONV_2D || builtin_op_code == BuiltinOperator_SVDF || builtin_op_code == BuiltinOperator_RNN || builtin_op_code == BuiltinOperator_BIDIRECTIONAL_SEQUENCE_LSTM || @@ -181,9 +188,10 @@ bool CheckAllOpInputsQuantized(const SubGraphT* subgraph, const OperatorT* op, // Inserts Tensors for each input tensor of op that should be // quantized into tensor_map. TfLiteStatus InsertQuantizableInputTensorsFromOperator( - const ModelT* model, const OperatorT* op, uint64_t weights_min_num_elements, + const ModelT* model, OperatorT* op, uint64_t weights_min_num_elements, const CustomOpMap& custom_op_map, - absl::flat_hash_map* tensor_map, int subgraph_index) { + absl::flat_hash_map* tensor_map, + int subgraph_index) { SubGraphT* subgraph = model->subgraphs.at(subgraph_index).get(); const OperatorCodeT* op_code = model->operator_codes[op->opcode_index].get(); @@ -222,7 +230,50 @@ TfLiteStatus InsertQuantizableInputTensorsFromOperator( continue; } - tensor_map->insert({tensor_idx, tensor}); + if (op_code->builtin_code == BuiltinOperator_DEPTHWISE_CONV_2D) { + tensor_map->insert( + {tensor_idx, {tensor, /*is_per_channel=*/true, /*dim=*/3}}); + } else if (op_code->builtin_code == BuiltinOperator_CONV_2D) { + tensor_map->insert( + {tensor_idx, {tensor, /*is_per_channel=*/true, /*dim=*/0}}); + } else { + switch (op_code->builtin_code) { + case BuiltinOperator_BIDIRECTIONAL_SEQUENCE_LSTM: + op->builtin_options.AsBidirectionalSequenceLSTMOptions() + ->asymmetric_quantize_inputs = true; + break; + case BuiltinOperator_BIDIRECTIONAL_SEQUENCE_RNN: + op->builtin_options.AsBidirectionalSequenceRNNOptions() + ->asymmetric_quantize_inputs = true; + break; + case BuiltinOperator_FULLY_CONNECTED: + op->builtin_options.AsFullyConnectedOptions() + ->asymmetric_quantize_inputs = true; + break; + case BuiltinOperator_LSTM: + op->builtin_options.AsLSTMOptions()->asymmetric_quantize_inputs = + true; + break; + case BuiltinOperator_RNN: + op->builtin_options.AsRNNOptions()->asymmetric_quantize_inputs = true; + break; + case BuiltinOperator_SVDF: + op->builtin_options.AsSVDFOptions()->asymmetric_quantize_inputs = + true; + break; + case BuiltinOperator_UNIDIRECTIONAL_SEQUENCE_LSTM: + op->builtin_options.AsUnidirectionalSequenceLSTMOptions() + ->asymmetric_quantize_inputs = true; + break; + case BuiltinOperator_UNIDIRECTIONAL_SEQUENCE_RNN: + op->builtin_options.AsSequenceRNNOptions() + ->asymmetric_quantize_inputs = true; + break; + default: + break; + } + tensor_map->insert({tensor_idx, {tensor, /*is_per_channel=*/false}}); + } } return kTfLiteOk; @@ -275,17 +326,22 @@ void MakeTensor(const string& name, const std::vector& shape, void UpdateInt8OperatorVersions(ModelT* model) { for (int i = 0; i < model->operator_codes.size(); ++i) { const BuiltinOperator& op_code = model->operator_codes[i]->builtin_code; - if (op_code == BuiltinOperator_CONV_2D || op_code == BuiltinOperator_SVDF || - op_code == BuiltinOperator_RNN || + if (op_code == BuiltinOperator_BIDIRECTIONAL_SEQUENCE_LSTM || op_code == BuiltinOperator_BIDIRECTIONAL_SEQUENCE_RNN || + op_code == BuiltinOperator_EMBEDDING_LOOKUP || + op_code == BuiltinOperator_RNN || op_code == BuiltinOperator_UNIDIRECTIONAL_SEQUENCE_LSTM || op_code == BuiltinOperator_UNIDIRECTIONAL_SEQUENCE_RNN) { - model->operator_codes[i]->version = 2; - } else if (op_code == BuiltinOperator_FULLY_CONNECTED || - op_code == BuiltinOperator_BIDIRECTIONAL_SEQUENCE_LSTM || - op_code == BuiltinOperator_EMBEDDING_LOOKUP || - op_code == BuiltinOperator_LSTM) { model->operator_codes[i]->version = 3; + } else if (op_code == BuiltinOperator_LSTM || + op_code == BuiltinOperator_SVDF) { + model->operator_codes[i]->version = 4; + } else if (op_code == BuiltinOperator_CONV_2D) { + model->operator_codes[i]->version = 5; + } else if (op_code == BuiltinOperator_DEPTHWISE_CONV_2D) { + model->operator_codes[i]->version = 6; + } else if (op_code == BuiltinOperator_FULLY_CONNECTED) { + model->operator_codes[i]->version = 9; } } } @@ -354,7 +410,7 @@ TfLiteStatus QuantizeWeightsInt8(flatbuffers::FlatBufferBuilder* builder, ++subgraph_index) { SubGraphT* subgraph = model->subgraphs.at(subgraph_index).get(); - absl::flat_hash_map tensor_map; + absl::flat_hash_map tensor_map; for (int i = 0; i < subgraph->operators.size(); ++i) { OperatorT* op = subgraph->operators[i].get(); TF_LITE_ENSURE_STATUS(InsertQuantizableInputTensorsFromOperator( @@ -362,16 +418,22 @@ TfLiteStatus QuantizeWeightsInt8(flatbuffers::FlatBufferBuilder* builder, subgraph_index)); } - for (std::pair tensor_pair : tensor_map) { + for (std::pair tensor_pair : tensor_map) { // Quantize the tensor. - TF_LITE_ENSURE_STATUS( - utils::SymmetricQuantizeTensor(model.get(), tensor_pair.second)); + if (tensor_pair.second.is_per_channel) { + TF_LITE_ENSURE_STATUS(utils::SymmetricQuantizeTensorPerChannel( + model.get(), tensor_pair.second.t, tensor_pair.second.channel_dim, + nullptr)); + } else { + TF_LITE_ENSURE_STATUS( + utils::SymmetricQuantizeTensor(model.get(), tensor_pair.second.t)); + } } // Examine the tensor consumers to determine which require dequantize ops. for (const auto& tensor_pair : tensor_map) { int32_t tensor_idx = tensor_pair.first; - TensorT* tensor = tensor_pair.second; + TensorT* tensor = tensor_pair.second.t; std::vector consumer_op_infos = GetTensorConsumers(model.get(), subgraph, tensor_idx); if (IsQuantizationPassThroughOps(model.get(), consumer_op_infos)) { diff --git a/tensorflow/lite/tools/optimize/quantize_weights_test.cc b/tensorflow/lite/tools/optimize/quantize_weights_test.cc index 76f2815ef0b..2f92a9ad71c 100644 --- a/tensorflow/lite/tools/optimize/quantize_weights_test.cc +++ b/tensorflow/lite/tools/optimize/quantize_weights_test.cc @@ -215,6 +215,8 @@ TEST_F(QuantizeWeightsTest, HybridConv) { } else if (quant_tensor->buffer() != 0) { EXPECT_EQ(quant_tensor->type(), TensorType_INT8) << quant_tensor->name()->str(); + auto shape = GetAsVector(quant_tensor->shape()); + EXPECT_EQ(quant_tensor->quantization()->scale()->size(), shape[0]); } else { EXPECT_EQ(quant_tensor->type(), TensorType_FLOAT32); } diff --git a/tensorflow/lite/tools/versioning/op_version.cc b/tensorflow/lite/tools/versioning/op_version.cc index a97b9da47f1..a339976739b 100644 --- a/tensorflow/lite/tools/versioning/op_version.cc +++ b/tensorflow/lite/tools/versioning/op_version.cc @@ -70,10 +70,13 @@ int GetBuiltinOperatorVersion(const OpSignature& op_sig) { return 3; } // If the op is a signed int8 hybrid operation, we need to return - // version 2. + // version 2 or 5 if per channel. if (op_sig.input_types.at(0) == TensorType_FLOAT32 && op_sig.input_types.at(1) == TensorType_INT8 && op_sig.output_types.at(0) == TensorType_FLOAT32) { + if (op_sig.options.conv_2d.is_per_channel_quantized) { + return 5; + } return 2; } return 1; @@ -87,10 +90,13 @@ int GetBuiltinOperatorVersion(const OpSignature& op_sig) { } // If the op is a signed int8 hybrid operation, we need to return - // version 4. + // version 4 or 6 if per-channel. if (op_sig.input_types.at(0) == TensorType_FLOAT32 && op_sig.input_types.at(1) == TensorType_INT8 && op_sig.output_types.at(0) == TensorType_FLOAT32) { + if (op_sig.options.depthwise_conv_2d.is_per_channel_quantized) { + return 6; + } return 4; } // If the op has signed int8 op_sig.inputs and op_sig.outputs, its @@ -154,6 +160,10 @@ int GetBuiltinOperatorVersion(const OpSignature& op_sig) { if (op_sig.input_types.at(0) == TensorType_FLOAT32 && op_sig.input_types.at(1) == TensorType_INT8 && op_sig.output_types.at(0) == TensorType_FLOAT32) { + if (op_sig.options.fully_connected.asymmetric_quantize_inputs) { + // This is to use the updated quantization scheme. + return 9; + } return 3; } // For float and uint8 fixed point kernels, if the weight is @@ -185,6 +195,10 @@ int GetBuiltinOperatorVersion(const OpSignature& op_sig) { if (op_sig.input_types.at(0) == TensorType_FLOAT32 && op_sig.input_types.at(1) == TensorType_INT8 && op_sig.output_types.at(0) == TensorType_FLOAT32) { + // This is to use the updated quantization scheme + if (op_sig.options.input_quantization.asymmetric_quantize_inputs) { + return 4; + } return 2; } return 1; @@ -251,6 +265,9 @@ int GetBuiltinOperatorVersion(const OpSignature& op_sig) { op_sig.input_types.at(0) == TensorType_FLOAT32 && op_sig.input_types.at(2) == TensorType_INT8 && op_sig.output_types.at(0) == TensorType_FLOAT32) { + if (op_sig.options.lstm.asymmetric_quantize_inputs) { + return 4; + } return 3; } // KERNEL_BASIC was added in version 2. @@ -265,6 +282,9 @@ int GetBuiltinOperatorVersion(const OpSignature& op_sig) { if (op_sig.input_types.at(0) == TensorType_FLOAT32 && op_sig.input_types.at(2) == TensorType_INT8 && op_sig.output_types.at(0) == TensorType_FLOAT32) { + if (op_sig.options.lstm.asymmetric_quantize_inputs) { + return 3; + } return 2; } return 1; @@ -450,7 +470,6 @@ int GetBuiltinOperatorVersion(const OpSignature& op_sig) { return 2; } return 1; - case BuiltinOperator_TANH: case BuiltinOperator_LOGISTIC: if (op_sig.input_types.at(0) == TensorType_INT16 && @@ -500,6 +519,19 @@ int GetBuiltinOperatorVersion(const OpSignature& op_sig) { } return 1; + case BuiltinOperator_RNN: + case BuiltinOperator_UNIDIRECTIONAL_SEQUENCE_RNN: + case BuiltinOperator_BIDIRECTIONAL_SEQUENCE_RNN: + case BuiltinOperator_BIDIRECTIONAL_SEQUENCE_LSTM: + if (op_sig.input_types.at(1) == TensorType_INT8 && + op_sig.output_types.at(0) == TensorType_FLOAT32) { + if (op_sig.options.input_quantization.asymmetric_quantize_inputs) { + return 3; + } else { + return 2; + } + } + return 1; case BuiltinOperator_ADD: case BuiltinOperator_PAD: case BuiltinOperator_PADV2: @@ -566,6 +598,16 @@ OpSignature GetOpSignature(const OperatorCode* op_code, const Operator* op, op_sig.options.depthwise_conv_2d.dilation_h_factor = conv_option->dilation_h_factor(); } + const Tensor* filter_tensor = + subgraph->tensors()->Get(op->inputs()->Get(1)); + const QuantizationParameters* filter_quant = + filter_tensor->quantization(); + int num_channels = filter_tensor->shape()->Get(3); + if (filter_quant && filter_quant->scale() && + filter_quant->scale()->Length() && + filter_quant->scale()->Length() == num_channels) { + op_sig.options.depthwise_conv_2d.is_per_channel_quantized = true; + } } break; case BuiltinOperator_FAKE_QUANT: { @@ -584,6 +626,8 @@ OpSignature GetOpSignature(const OperatorCode* op_code, const Operator* op, fully_connected_option->keep_num_dims(); op_sig.options.fully_connected.weights_format = fully_connected_option->weights_format(); + op_sig.options.fully_connected.asymmetric_quantize_inputs = + fully_connected_option->asymmetric_quantize_inputs(); } const Tensor* weight_tensor = @@ -644,6 +688,18 @@ OpSignature GetOpSignature(const OperatorCode* op_code, const Operator* op, op_sig.options.resize.align_corners = resize_nn_option->align_corners(); } } break; + case BuiltinOperator_CONV_2D: { + const Tensor* filter_tensor = + subgraph->tensors()->Get(op->inputs()->Get(1)); + const QuantizationParameters* filter_quant = + filter_tensor->quantization(); + int num_channels = filter_tensor->shape()->Get(0); + if (filter_quant && filter_quant->scale() && + filter_quant->scale()->Length() && + filter_quant->scale()->Length() == num_channels) { + op_sig.options.conv_2d.is_per_channel_quantized = true; + } + } break; // TODO(b/150176627): Add tests for GetOpSignature. case BuiltinOperator_STRIDED_SLICE: case BuiltinOperator_SPACE_TO_BATCH_ND: @@ -651,7 +707,6 @@ OpSignature GetOpSignature(const OperatorCode* op_code, const Operator* op, case BuiltinOperator_TRANSPOSE: { op_sig.options.single_input_op.num_dims = GetNumDims(subgraph, op, 0); } break; - case BuiltinOperator_SUB: case BuiltinOperator_DIV: case BuiltinOperator_MAXIMUM: diff --git a/tensorflow/lite/tools/versioning/op_version.h b/tensorflow/lite/tools/versioning/op_version.h index df74ffaf6dd..71362001387 100644 --- a/tensorflow/lite/tools/versioning/op_version.h +++ b/tensorflow/lite/tools/versioning/op_version.h @@ -30,6 +30,7 @@ typedef struct { struct { int32_t dilation_w_factor; int32_t dilation_h_factor; + bool is_per_channel_quantized; } depthwise_conv_2d; struct { bool narrow_range; @@ -40,6 +41,7 @@ typedef struct { // TODO(b/156530611): Make this global when more ops support sparse // computation. bool sparse_weight; + bool asymmetric_quantize_inputs; } fully_connected; struct { float input1_scale; @@ -48,6 +50,7 @@ typedef struct { } mul; struct { LSTMKernelType kernel_type; + bool asymmetric_quantize_inputs; } lstm; struct { bool half_pixel_centers; @@ -60,6 +63,12 @@ typedef struct { int32_t num_dims; bool need_broadcast; } broadcast; + struct { + bool is_per_channel_quantized; + } conv_2d; + struct { + bool asymmetric_quantize_inputs; + } input_quantization; } options; } OpSignature; diff --git a/tensorflow/lite/tools/versioning/op_version_test.cc b/tensorflow/lite/tools/versioning/op_version_test.cc index 4017fc3bff0..e9fd857a3f5 100644 --- a/tensorflow/lite/tools/versioning/op_version_test.cc +++ b/tensorflow/lite/tools/versioning/op_version_test.cc @@ -361,6 +361,19 @@ TEST(OpVersionTest, VersioningFullyConnectedTest) { fake_op_sig.options.fully_connected = { false, FullyConnectedOptionsWeightsFormat_DEFAULT, true}; EXPECT_EQ(GetBuiltinOperatorVersion(fake_op_sig), 8); + + fake_op_sig = { + .op = BuiltinOperator_FULLY_CONNECTED, + .input_types = + std::vector{TensorType_FLOAT32, TensorType_INT8, + TensorType_FLOAT32}, + .output_types = std::vector{TensorType_FLOAT32}, + }; + fake_op_sig.options.fully_connected = { + false, FullyConnectedOptionsWeightsFormat_DEFAULT, false, false}; + EXPECT_EQ(GetBuiltinOperatorVersion(fake_op_sig), 3); + fake_op_sig.options.fully_connected.asymmetric_quantize_inputs = true; + EXPECT_EQ(GetBuiltinOperatorVersion(fake_op_sig), 9); } TEST(OpVersionTest, VersioningDequantizeTest) { @@ -412,6 +425,15 @@ TEST(OpVersionTest, VersioningConv2DTest) { .output_types = std::vector{TensorType_FLOAT32}, }; EXPECT_EQ(GetBuiltinOperatorVersion(fake_op_sig), 2); + + fake_op_sig = { + .op = BuiltinOperator_CONV_2D, + .input_types = + std::vector{TensorType_FLOAT32, TensorType_INT8}, + .output_types = std::vector{TensorType_FLOAT32}, + }; + fake_op_sig.options.conv_2d.is_per_channel_quantized = true; + EXPECT_EQ(GetBuiltinOperatorVersion(fake_op_sig), 5); } TEST(OpVersionTest, VersioningFloorDivOperatorTest) { @@ -479,6 +501,8 @@ TEST(OpVersionTest, VersioningSVDFOperatorTest) { .output_types = std::vector{TensorType_FLOAT32}, }; EXPECT_EQ(GetBuiltinOperatorVersion(fake_op_sig), 2); + fake_op_sig.options.input_quantization.asymmetric_quantize_inputs = true; + EXPECT_EQ(GetBuiltinOperatorVersion(fake_op_sig), 4); fake_op_sig = { .op = BuiltinOperator_SVDF, @@ -489,6 +513,7 @@ TEST(OpVersionTest, VersioningSVDFOperatorTest) { }; EXPECT_EQ(GetBuiltinOperatorVersion(fake_op_sig), 3); } + TEST(OpVersionTest, VersioningDepthwiseConv2DTest) { OpSignature fake_op_sig = { .op = BuiltinOperator_DEPTHWISE_CONV_2D, @@ -497,6 +522,8 @@ TEST(OpVersionTest, VersioningDepthwiseConv2DTest) { .output_types = std::vector{TensorType_FLOAT32}, }; EXPECT_EQ(GetBuiltinOperatorVersion(fake_op_sig), 4); + fake_op_sig.options.depthwise_conv_2d.is_per_channel_quantized = true; + EXPECT_EQ(GetBuiltinOperatorVersion(fake_op_sig), 6); fake_op_sig = { .op = BuiltinOperator_DEPTHWISE_CONV_2D, diff --git a/tensorflow/lite/tools/versioning/runtime_version.cc b/tensorflow/lite/tools/versioning/runtime_version.cc index 36976354685..efec5a7da18 100644 --- a/tensorflow/lite/tools/versioning/runtime_version.cc +++ b/tensorflow/lite/tools/versioning/runtime_version.cc @@ -63,11 +63,13 @@ std::string FindMinimumRuntimeVersionForOp(tflite::BuiltinOperator op_code, {{BuiltinOperator_CONV_2D, 2}, "1.14.0"}, {{BuiltinOperator_CONV_2D, 3}, "1.14.0"}, {{BuiltinOperator_CONV_2D, 4}, kPendingReleaseVersion}, + {{BuiltinOperator_CONV_2D, 5}, kPendingReleaseVersion}, {{BuiltinOperator_DEPTHWISE_CONV_2D, 1}, "1.5.0"}, {{BuiltinOperator_DEPTHWISE_CONV_2D, 2}, "1.12.0"}, {{BuiltinOperator_DEPTHWISE_CONV_2D, 3}, "1.14.0"}, {{BuiltinOperator_DEPTHWISE_CONV_2D, 4}, "2.2.0"}, {{BuiltinOperator_DEPTHWISE_CONV_2D, 5}, kPendingReleaseVersion}, + {{BuiltinOperator_DEPTHWISE_CONV_2D, 6}, kPendingReleaseVersion}, {{BuiltinOperator_ADD, 1}, "1.5.0"}, {{BuiltinOperator_ADD, 2}, "1.14.0"}, {{BuiltinOperator_ADD_N, 1}, "1.14.0"}, @@ -102,6 +104,7 @@ std::string FindMinimumRuntimeVersionForOp(tflite::BuiltinOperator op_code, {{BuiltinOperator_FULLY_CONNECTED, 6}, "2.1.0"}, {{BuiltinOperator_FULLY_CONNECTED, 7}, kPendingReleaseVersion}, {{BuiltinOperator_FULLY_CONNECTED, 8}, kPendingReleaseVersion}, + {{BuiltinOperator_FULLY_CONNECTED, 9}, kPendingReleaseVersion}, {{BuiltinOperator_GATHER, 1}, "1.6.0"}, {{BuiltinOperator_GATHER, 2}, "1.14.0"}, {{BuiltinOperator_GATHER, 3}, "1.15.0"}, @@ -111,6 +114,7 @@ std::string FindMinimumRuntimeVersionForOp(tflite::BuiltinOperator op_code, {{BuiltinOperator_SVDF, 1}, "1.5.0"}, {{BuiltinOperator_SVDF, 2}, "1.14.0"}, {{BuiltinOperator_SVDF, 3}, "2.2.0"}, + {{BuiltinOperator_SVDF, 4}, kPendingReleaseVersion}, {{BuiltinOperator_L2_NORMALIZATION, 1}, "1.5.0"}, {{BuiltinOperator_L2_NORMALIZATION, 2}, "1.14.0"}, {{BuiltinOperator_L2_POOL_2D, 1}, "1.5.0"}, @@ -151,13 +155,18 @@ std::string FindMinimumRuntimeVersionForOp(tflite::BuiltinOperator op_code, {{BuiltinOperator_LSTM, 1}, "1.7.0"}, {{BuiltinOperator_LSTM, 2}, "1.10.0"}, {{BuiltinOperator_LSTM, 3}, "1.14.0"}, + {{BuiltinOperator_LSTM, 4}, kPendingReleaseVersion}, {{BuiltinOperator_UNIDIRECTIONAL_SEQUENCE_LSTM, 1}, "1.13.1"}, {{BuiltinOperator_UNIDIRECTIONAL_SEQUENCE_LSTM, 2}, "1.14.0"}, + {{BuiltinOperator_UNIDIRECTIONAL_SEQUENCE_LSTM, 3}, + kPendingReleaseVersion}, {{BuiltinOperator_BIDIRECTIONAL_SEQUENCE_LSTM, 1}, "1.14.0"}, {{BuiltinOperator_BIDIRECTIONAL_SEQUENCE_LSTM, 2}, "1.14.0"}, {{BuiltinOperator_BIDIRECTIONAL_SEQUENCE_LSTM, 3}, "1.14.0"}, {{BuiltinOperator_BIDIRECTIONAL_SEQUENCE_RNN, 1}, "1.14.0"}, {{BuiltinOperator_BIDIRECTIONAL_SEQUENCE_RNN, 2}, "1.14.0"}, + {{BuiltinOperator_BIDIRECTIONAL_SEQUENCE_RNN, 3}, + kPendingReleaseVersion}, {{BuiltinOperator_MEAN, 1}, "1.6.0"}, {{BuiltinOperator_MEAN, 2}, "1.14.0"}, {{BuiltinOperator_SUM, 1}, "1.10.0"}, @@ -179,6 +188,7 @@ std::string FindMinimumRuntimeVersionForOp(tflite::BuiltinOperator op_code, kPendingReleaseVersion}, {{BuiltinOperator_RNN, 1}, "1.5.0"}, {{BuiltinOperator_RNN, 2}, "1.14.0"}, + {{BuiltinOperator_RNN, 3}, kPendingReleaseVersion}, {{BuiltinOperator_SKIP_GRAM, 1}, "1.5.0"}, {{BuiltinOperator_SQUEEZE, 1}, "1.6.0"}, {{BuiltinOperator_SPLIT, 1}, "1.5.0"}, @@ -233,6 +243,8 @@ std::string FindMinimumRuntimeVersionForOp(tflite::BuiltinOperator op_code, {{BuiltinOperator_UNIQUE, 1}, "1.14.0"}, {{BuiltinOperator_UNIDIRECTIONAL_SEQUENCE_RNN, 1}, "1.14.0"}, {{BuiltinOperator_UNIDIRECTIONAL_SEQUENCE_RNN, 2}, "1.14.0"}, + {{BuiltinOperator_UNIDIRECTIONAL_SEQUENCE_RNN, 3}, + kPendingReleaseVersion}, {{BuiltinOperator_WHERE, 1}, "1.14.0"}, {{BuiltinOperator_DEQUANTIZE, 1}, "1.13.1"}, {{BuiltinOperator_DEQUANTIZE, 2}, "1.14.0"}, From 27d1a803165af2cab5025ebf2387aa4785349c5a Mon Sep 17 00:00:00 2001 From: Tare Gaskin Date: Tue, 23 Jun 2020 02:34:18 +0000 Subject: [PATCH 518/867] [-Wsign-compare] warning fixes batch 3 --- indexed_warning_files.json | 1 + tensorflow/core/data/compression_utils.cc | 2 +- .../optimizers/arithmetic_optimizer.cc | 24 +++++++++-------- .../grappler/optimizers/constant_folding.cc | 27 ++++++++++--------- .../generic_layout_optimizer_transposer.cc | 24 ++++++++++------- .../generic_layout_optimizer_transposer.h | 10 ++++--- .../optimizers/graph_optimizer_stage.cc | 2 +- .../optimizers/implementation_selector.cc | 2 +- .../grappler/optimizers/memory_optimizer.cc | 2 +- .../optimizers/scoped_allocator_optimizer.cc | 11 ++++---- .../grappler/optimizers/shape_optimizer.cc | 3 ++- .../kernels/data/single_threaded_executor.cc | 4 +-- .../kernels/initializable_lookup_table.cc | 2 +- .../remote_fused_graph_execute_utils.cc | 15 ++++++----- tensorflow/core/kernels/stack.cc | 3 ++- ...embedding_optimization_parameters_utils.cc | 4 +-- tensorflow/core/util/bcast.h | 4 +-- 17 files changed, 79 insertions(+), 61 deletions(-) create mode 100644 indexed_warning_files.json diff --git a/indexed_warning_files.json b/indexed_warning_files.json new file mode 100644 index 00000000000..cbe0560ba18 --- /dev/null +++ b/indexed_warning_files.json @@ -0,0 +1 @@ +{"0": "tensorflow/lite/arena_planner.cc", "1": "tensorflow/core/platform/protobuf.cc", "2": "tensorflow/core/platform/protobuf.cc", "3": "tensorflow/core/platform/default/logging.cc", "4": "tensorflow/core/platform/default/logging.cc", "5": "tensorflow/core/lib/strings/proto_serialization.cc", "6": "tensorflow/core/lib/strings/proto_serialization.cc", "7": "tensorflow/core/platform/default/stacktrace_handler.cc", "8": "tensorflow/core/platform/default/stacktrace_handler.cc", "9": "tensorflow/core/framework/cpu_allocator_impl.cc", "10": "tensorflow/core/framework/allocator_registry.cc", "11": "tensorflow/core/framework/cpu_allocator_impl.cc", "12": "tensorflow/core/framework/allocator_registry.cc", "13": "tensorflow/lite/experimental/microfrontend/lib/fft.cc", "14": "tensorflow/python/util/tf_stack.cc", "15": "external/com_github_grpc_grpc/src/core/tsi/ssl_transport_security.cc", "16": "tensorflow/core/profiler/internal/parse_annotation.cc", "17": "tensorflow/compiler/mlir/xla/ir/chlo_ops.cc", "18": "tensorflow/core/platform/status.cc", "19": "tensorflow/core/platform/file_system_helper.cc", "20": "tensorflow/core/platform/file_system.cc", "21": "tensorflow/core/platform/env.cc", "22": "tensorflow/core/lib/io/random_inputstream.cc", "23": "tensorflow/core/lib/io/snappy/snappy_inputbuffer.cc", "24": "tensorflow/core/lib/io/inputbuffer.cc", "25": "tensorflow/core/lib/io/zlib_outputbuffer.cc", "26": "tensorflow/core/lib/io/snappy/snappy_outputbuffer.cc", "27": "tensorflow/core/framework/tensor_shape.cc", "28": "tensorflow/compiler/mlir/xla/ir/hlo_ops.cc", "29": "tensorflow/compiler/mlir/lite/quantization/quantization_utils.cc", "30": "tensorflow/compiler/mlir/lite/quantization/quantization_driver.cc", "31": "tensorflow/compiler/mlir/lite/quantization/quantization_config.cc", "32": "tensorflow/core/kernels/data/prefetch_autotuner.cc", "33": "tensorflow/core/kernels/quantization_utils.cc", "34": "tensorflow/core/profiler/utils/derived_timeline.cc", "35": "tensorflow/core/profiler/utils/xplane_utils.cc", "36": "tensorflow/core/profiler/lib/profiler_session.cc", "37": "tensorflow/core/platform/s3/s3_file_system.cc", "38": "tensorflow/lite/toco/model_cmdline_flags.cc", "39": "tensorflow/lite/toco/toco_cmdline_flags.cc", "40": "tensorflow/lite/toco/toco_cmdline_flags.cc", "41": "tensorflow/compiler/xla/window_util.cc", "42": "tensorflow/compiler/mlir/tensorflow/utils/dump_mlir_util.cc", "43": "tensorflow/compiler/mlir/lite/quantization/import_quant_stats_pass.cc", "44": "tensorflow/core/kernels/batch_kernels.cc", "45": "tensorflow/lite/toco/graph_transformations/reorder_elementwise_unary.cc", "46": "tensorflow/core/kernels/range_sampler.cc", "47": "tensorflow/core/kernels/data/experimental/sql/sqlite_query_connection.cc", "48": "tensorflow/core/grappler/utils.cc", "49": "tensorflow/core/grappler/costs/op_level_cost_estimator.cc", "50": "tensorflow/lite/toco/graph_transformations/convert_trivial_tile_to_concat.cc", "51": "tensorflow/core/grappler/utils/topological_sort.cc", "52": "tensorflow/core/grappler/utils/frame.cc", "53": "tensorflow/core/grappler/optimizers/common_subgraph_elimination.cc", "54": "tensorflow/core/grappler/optimizers/model_pruner.cc", "55": "tensorflow/python/grappler/model_analyzer.cc", "56": "tensorflow/core/grappler/optimizers/debug_stripper.cc", "57": "tensorflow/core/grappler/utils/graph_view.cc", "58": "tensorflow/core/grappler/utils/functions.cc", "59": "tensorflow/core/grappler/costs/graph_memory.cc", "60": "tensorflow/core/grappler/optimizers/generic_layout_optimizer.cc", "61": "tensorflow/core/grappler/optimizers/function_optimizer.cc", "62": "tensorflow/core/grappler/optimizers/generic_layout_optimizer_transposer.cc", "63": "tensorflow/core/grappler/costs/virtual_scheduler.cc", "64": "tensorflow/core/grappler/optimizers/implementation_selector.cc", "65": "tensorflow/core/grappler/optimizers/pin_to_host_optimizer.cc", "66": "tensorflow/core/grappler/optimizers/shape_optimizer.cc", "67": "tensorflow/core/grappler/optimizers/graph_optimizer_stage.cc", "68": "tensorflow/core/grappler/optimizers/memory_optimizer.cc", "69": "tensorflow/core/grappler/optimizers/scoped_allocator_optimizer.cc", "70": "tensorflow/core/grappler/costs/graph_properties.cc", "71": "tensorflow/core/grappler/optimizers/constant_folding.cc", "72": "tensorflow/core/grappler/optimizers/loop_optimizer.cc", "73": "tensorflow/core/grappler/optimizers/arithmetic_optimizer.cc", "74": "tensorflow/core/data/service/compression_utils.cc", "75": "tensorflow/core/kernels/stack.cc", "76": "tensorflow/core/kernels/remote_fused_graph_execute_utils.cc", "77": "tensorflow/core/tpu/tpu_embedding_optimization_parameters_utils.cc", "78": "tensorflow/core/kernels/initializable_lookup_table.cc", "79": "tensorflow/core/kernels/data/single_threaded_executor.cc", "80": "tensorflow/core/kernels/lookup_util.cc", "81": "tensorflow/lite/toco/graph_transformations/merge_reshape_into_preceding_transpose.cc", "82": "tensorflow/lite/toco/graph_transformations/identify_nearest_upsample.cc", "83": "tensorflow/lite/toco/graph_transformations/remove_successive_transpose.cc", "84": "tensorflow/lite/toco/graph_transformations/remove_trivial_passthrough.cc", "85": "tensorflow/lite/toco/graph_transformations/propagate_fixed_sizes.cc", "86": "tensorflow/lite/toco/graph_transformations/resolve_batch_normalization.cc", "87": "tensorflow/lite/toco/graph_transformations/resolve_constant_concatenation.cc", "88": "tensorflow/lite/toco/graph_transformations/reorder_reshape_transpose.cc", "89": "tensorflow/lite/toco/graph_transformations/resolve_constant_unary.cc", "90": "tensorflow/lite/toco/graph_transformations/drop_fake_quant.cc", "91": "tensorflow/lite/toco/graph_transformations/unpartition_embedding_lookup.cc", "92": "tensorflow/lite/toco/graph_transformations/dequantize.cc", "93": "tensorflow/lite/toco/graph_transformations/unroll_batch_matmul.cc", "94": "tensorflow/lite/toco/graph_transformations/hardcode_min_max.cc", "95": "tensorflow/lite/toco/graph_transformations/propagate_array_data_types.cc", "96": "tensorflow/lite/toco/graph_transformations/group_bidirectional_sequence_ops.cc", "97": "tensorflow/lite/toco/graph_transformations/propagate_fake_quant_num_bits.cc", "98": "tensorflow/lite/toco/graph_transformations/resolve_constant_slice.cc", "99": "tensorflow/lite/toco/graph_transformations/resolve_constant_transpose.cc", "100": "tensorflow/lite/toco/graph_transformations/resolve_constant_pack.cc", "101": "tensorflow/lite/toco/graph_transformations/fuse_broadcast_into_following_binary.cc", "102": "tensorflow/lite/toco/graph_transformations/ensure_uint8_weights_safe_for_fast_int8_kernels.cc", "103": "tensorflow/lite/toco/graph_transformations/convert_trivial_transpose_to_reshape.cc", "104": "tensorflow/core/profiler/convert/op_metrics_to_record.cc", "105": "tensorflow/core/profiler/utils/event_span.cc", "106": "tensorflow/python/framework/python_op_gen.cc", "107": "tensorflow/python/framework/python_op_gen_internal.cc", "108": "tensorflow/compiler/tf2xla/ops/xla_ops.cc", "109": "tensorflow/compiler/mlir/tensorflow/ir/tf_device.cc", "110": "tensorflow/compiler/mlir/tensorflow/transforms/unroll_batch_matmul.cc", "111": "tensorflow/compiler/mlir/tensorflow/ir/tf_executor.cc", "112": "tensorflow/core/kernels/data/captured_function.cc", "113": "tensorflow/core/profiler/convert/xplane_to_tf_functions.cc", "114": "tensorflow/core/common_runtime/bfc_allocator.cc", "115": "tensorflow/core/util/padding.cc", "116": "tensorflow/core/framework/op_def_util.cc", "117": "tensorflow/core/framework/node_def_util.cc", "118": "tensorflow/core/framework/shape_inference.cc", "119": "tensorflow/core/framework/common_shape_fns.cc", "120": "tensorflow/core/common_runtime/lower_case_op.cc", "121": "tensorflow/core/common_runtime/gradients.cc", "122": "tensorflow/compiler/mlir/tensorflow/ir/tf_ops.cc", "123": "tensorflow/stream_executor/device_description.cc", "124": "tensorflow/compiler/jit/shape_inference.cc", "125": "tensorflow/compiler/mlir/tensorflow/utils/export_utils.cc", "126": "tensorflow/compiler/xla/index_util.cc", "127": "tensorflow/compiler/xla/metric_table_report.cc", "128": "tensorflow/compiler/xla/layout.cc", "129": "tensorflow/stream_executor/stream_executor_pimpl.cc", "130": "tensorflow/compiler/xla/shape_util.cc", "131": "tensorflow/compiler/xla/service/hlo_lexer.cc", "132": "tensorflow/compiler/xla/service/cpu/shape_partition.cc", "133": "tensorflow/compiler/xla/util.cc", "134": "tensorflow/compiler/xla/service/name_uniquer.cc", "135": "tensorflow/compiler/xla/shape_layout.cc", "136": "tensorflow/compiler/xla/client/sharding_builder.cc", "137": "tensorflow/compiler/xla/service/computation_layout.cc", "138": "tensorflow/compiler/mlir/xla/type_to_shape.cc", "139": "tensorflow/compiler/mlir/tensorflow/translate/mlir_roundtrip_flags.cc", "140": "tensorflow/compiler/xla/service/shaped_buffer.cc", "141": "tensorflow/compiler/tf2xla/sharding_util.cc", "142": "tensorflow/stream_executor/stream.cc", "143": "tensorflow/compiler/xla/layout_util.cc", "144": "tensorflow/compiler/xla/shape.cc", "145": "tensorflow/core/profiler/convert/xplane_to_memory_profile.cc", "146": "tensorflow/core/profiler/convert/op_stats_to_overview_page.cc", "147": "tensorflow/compiler/mlir/xla/transforms/legalize_tf.cc", "148": "tensorflow/compiler/xla/client/xla_computation.cc", "149": "tensorflow/compiler/mlir/tensorflow/utils/xla_sharding_util.cc", "150": "tensorflow/compiler/tf2xla/rearrange_function_argument.cc", "151": "tensorflow/compiler/jit/encapsulate_util.cc", "152": "tensorflow/compiler/tf2xla/tf2xla_util.cc", "153": "tensorflow/compiler/tf2xla/functionalize_while.cc", "154": "tensorflow/compiler/tf2xla/functionalize_cond.cc", "155": "tensorflow/core/kernels/boosted_trees/resources.cc", "156": "tensorflow/python/client/session_ref.cc", "157": "tensorflow/core/distributed_runtime/rpc/grpc_state.cc", "158": "tensorflow/core/distributed_runtime/rpc/grpc_remote_master.cc", "159": "tensorflow/core/distributed_runtime/rpc/eager/grpc_eager_client.cc", "160": "tensorflow/core/distributed_runtime/collective_param_resolver_distributed.cc", "161": "tensorflow/core/distributed_runtime/collective_rma_distributed.cc", "162": "tensorflow/core/platform/status.cc", "163": "tensorflow/core/framework/tensor_shape.cc", "164": "tensorflow/core/profiler/internal/tfprof_timeline.cc", "165": "tensorflow/core/platform/file_system.cc", "166": "tensorflow/core/platform/file_system_helper.cc", "167": "tensorflow/core/platform/env.cc", "168": "tensorflow/core/lib/io/snappy/snappy_outputbuffer.cc", "169": "tensorflow/core/lib/io/zlib_outputbuffer.cc", "170": "tensorflow/core/lib/io/snappy/snappy_inputbuffer.cc", "171": "tensorflow/core/lib/io/random_inputstream.cc", "172": "tensorflow/core/lib/io/inputbuffer.cc", "173": "tensorflow/core/util/padding.cc", "174": "tensorflow/core/framework/op_def_util.cc", "175": "tensorflow/core/platform/s3/s3_file_system.cc", "176": "tensorflow/stream_executor/device_description.cc", "177": "tensorflow/core/framework/shape_inference.cc", "178": "tensorflow/core/framework/node_def_util.cc", "179": "tensorflow/core/common_runtime/bfc_allocator.cc", "180": "tensorflow/core/framework/common_shape_fns.cc", "181": "tensorflow/stream_executor/stream_executor_pimpl.cc", "182": "tensorflow/core/profiler/utils/xplane_utils.cc", "183": "tensorflow/core/grappler/utils.cc", "184": "tensorflow/core/grappler/costs/op_level_cost_estimator.cc", "185": "tensorflow/core/grappler/utils/symbolic_shapes.cc", "186": "tensorflow/core/grappler/utils/frame.cc", "187": "tensorflow/core/grappler/utils/topological_sort.cc", "188": "tensorflow/stream_executor/stream.cc", "189": "tensorflow/core/kernels/initializable_lookup_table.cc", "190": "tensorflow/core/grappler/optimizers/common_subgraph_elimination.cc", "191": "tensorflow/core/grappler/utils/graph_view.cc", "192": "tensorflow/core/grappler/optimizers/model_pruner.cc", "193": "tensorflow/core/kernels/lookup_util.cc", "194": "tensorflow/compiler/tf2xla/ops/xla_ops.cc", "195": "tensorflow/python/framework/python_op_gen_internal.cc", "196": "tensorflow/python/framework/python_op_gen.cc", "197": "tensorflow/core/profiler/utils/derived_timeline.cc", "198": "tensorflow/core/grappler/optimizers/debug_stripper.cc", "199": "tensorflow/core/grappler/optimizers/implementation_selector.cc", "200": "tensorflow/core/grappler/utils/functions.cc", "201": "tensorflow/core/grappler/optimizers/graph_optimizer_stage.cc", "202": "tensorflow/core/grappler/optimizers/pin_to_host_optimizer.cc", "203": "tensorflow/core/grappler/optimizers/shape_optimizer.cc", "204": "tensorflow/core/grappler/optimizers/generic_layout_optimizer_transposer.cc", "205": "tensorflow/core/grappler/optimizers/function_optimizer.cc", "206": "tensorflow/core/grappler/optimizers/scoped_allocator_optimizer.cc", "207": "tensorflow/core/grappler/optimizers/generic_layout_optimizer.cc", "208": "tensorflow/core/grappler/costs/virtual_scheduler.cc", "209": "tensorflow/core/grappler/optimizers/loop_optimizer.cc", "210": "tensorflow/core/grappler/optimizers/constant_folding.cc", "211": "tensorflow/core/profiler/lib/profiler_session.cc", "212": "tensorflow/core/grappler/costs/graph_properties.cc", "213": "tensorflow/core/grappler/costs/graph_memory.cc", "214": "tensorflow/core/grappler/optimizers/arithmetic_optimizer.cc", "215": "tensorflow/core/grappler/optimizers/memory_optimizer.cc", "216": "tensorflow/core/common_runtime/gradients.cc", "217": "tensorflow/core/common_runtime/lower_case_op.cc", "218": "tensorflow/core/grappler/utils/symbolic_shapes.cc", "219": "tensorflow/compiler/jit/graphcycles/graphcycles.cc", "220": "tensorflow/compiler/jit/xla_cluster_util.cc", "221": "tensorflow/core/kernels/data/experimental/snapshot_util.cc", "222": "tensorflow/core/grappler/optimizers/data/vectorization/parse_single_example_vectorizer.cc", "223": "tensorflow/core/distributed_runtime/graph_mgr.cc", "224": "tensorflow/core/grappler/optimizers/data/vectorization_utils.cc", "225": "tensorflow/core/grappler/optimizers/data/map_vectorization.cc", "226": "tensorflow/cc/framework/while_gradients.cc", "227": "tensorflow/cc/framework/gradients.cc", "228": "tensorflow/core/grappler/graph_analyzer/subgraph.cc", "229": "tensorflow/core/grappler/graph_analyzer/graph_analyzer.cc", "230": "tensorflow/core/grappler/graph_analyzer/sig_node.cc", "231": "tensorflow/compiler/tf2xla/shape_util.cc", "232": "tensorflow/compiler/xla/service/computation_placer.cc", "233": "tensorflow/compiler/xla/client/executable_build_options.cc", "234": "tensorflow/compiler/xla/service/hlo_module_config.cc", "235": "tensorflow/core/distributed_runtime/master_session.cc", "236": "tensorflow/compiler/mlir/tensorflow/utils/tpu_rewrite_device_util.cc", "237": "tensorflow/compiler/tf2xla/literal_util.cc", "238": "tensorflow/compiler/xla/service/cpu/cpu_options.cc", "239": "tensorflow/compiler/xla/client/client.cc", "240": "tensorflow/compiler/xla/literal_util.cc", "241": "tensorflow/core/distributed_runtime/rpc/grpc_session.cc", "242": "tensorflow/compiler/xla/service/hlo_input_output_alias_config.cc", "243": "tensorflow/compiler/xla/service/call_graph.cc", "244": "tensorflow/compiler/xla/service/tuple_util.cc", "245": "tensorflow/compiler/xla/service/hlo_reachability.cc", "246": "tensorflow/compiler/xla/service/hlo_cost_analysis.cc", "247": "tensorflow/compiler/xla/service/hlo_execution_profile.cc", "248": "tensorflow/compiler/xla/service/hlo_module_group.cc", "249": "tensorflow/compiler/mlir/xla/hlo_utils.cc", "250": "tensorflow/compiler/xla/service/hlo_sharding_metadata.cc", "251": "tensorflow/compiler/xla/service/map_inliner.cc", "252": "tensorflow/compiler/xla/service/flatten_call_graph.cc", "253": "tensorflow/compiler/xla/service/hlo_domain_map.cc", "254": "tensorflow/compiler/xla/service/hlo_cse.cc", "255": "tensorflow/compiler/xla/service/batchnorm_expander.cc", "256": "tensorflow/compiler/xla/service/dynamic_index_splitter.cc", "257": "tensorflow/compiler/xla/service/dfs_hlo_visitor.cc", "258": "tensorflow/compiler/xla/service/hlo_subcomputation_unification.cc", "259": "tensorflow/compiler/xla/service/slice_sinker.cc", "260": "tensorflow/compiler/xla/service/dot_decomposer.cc", "261": "tensorflow/compiler/xla/service/sort_simplifier.cc", "262": "tensorflow/compiler/xla/service/reshape_mover.cc", "263": "tensorflow/compiler/xla/service/gpu/partition_assignment.cc", "264": "tensorflow/compiler/xla/service/fusion_node_indexing_evaluation.cc", "265": "tensorflow/compiler/xla/service/dynamic_parameter_binding.cc", "266": "tensorflow/compiler/xla/literal.cc", "267": "tensorflow/compiler/xla/service/hlo_schedule.cc", "268": "tensorflow/compiler/xla/service/buffer_value.cc", "269": "tensorflow/compiler/xla/service/hlo_dce.cc", "270": "tensorflow/compiler/xla/service/hlo_module.cc", "271": "tensorflow/compiler/xla/service/hlo_sharding.cc", "272": "tensorflow/compiler/xla/service/transpose_folding.cc", "273": "tensorflow/compiler/xla/service/logical_buffer.cc", "274": "tensorflow/compiler/xla/service/call_inliner.cc", "275": "tensorflow/compiler/xla/service/hlo_buffer.cc", "276": "tensorflow/compiler/xla/service/hlo_instructions.cc", "277": "tensorflow/compiler/xla/service/hlo_value.cc", "278": "tensorflow/compiler/xla/service/zero_sized_hlo_elimination.cc", "279": "tensorflow/compiler/xla/service/tuple_simplifier.cc", "280": "tensorflow/compiler/xla/service/hlo_computation.cc", "281": "tensorflow/compiler/xla/service/logical_buffer_analysis.cc", "282": "tensorflow/compiler/xla/service/hlo_phi_graph.cc", "283": "tensorflow/compiler/xla/service/conditional_simplifier.cc", "284": "tensorflow/compiler/xla/service/hlo_query.cc", "285": "tensorflow/compiler/xla/service/channel_tracker.cc", "286": "tensorflow/compiler/xla/client/lib/constants.cc", "287": "tensorflow/compiler/xla/service/collective_ops_utils.cc", "288": "tensorflow/compiler/xla/service/hlo_dataflow_analysis.cc", "289": "tensorflow/compiler/tf2xla/lib/util.cc", "290": "tensorflow/compiler/xla/service/shape_inference.cc", "291": "tensorflow/compiler/tf2xla/lib/broadcast.cc", "292": "tensorflow/compiler/tf2xla/lib/data_format.cc", "293": "tensorflow/compiler/xla/service/hlo_parser.cc", "294": "tensorflow/compiler/xla/service/hlo_ordering.cc", "295": "tensorflow/compiler/xla/service/hlo_instruction.cc", "296": "tensorflow/compiler/xla/service/instruction_fusion.cc", "297": "tensorflow/compiler/xla/service/tuple_points_to_analysis.cc", "298": "tensorflow/compiler/xla/service/hlo_verifier.cc", "299": "tensorflow/compiler/xla/client/lib/comparators.cc", "300": "tensorflow/compiler/xla/client/lib/arithmetic.cc", "301": "tensorflow/compiler/xla/client/lib/sorting.cc", "302": "tensorflow/compiler/mlir/xla/ir/mlir_hlo_builder.cc", "303": "tensorflow/compiler/xla/client/lib/loops.cc", "304": "tensorflow/compiler/tf2xla/lib/scatter.cc", "305": "tensorflow/compiler/xla/client/lib/slicing.cc", "306": "tensorflow/compiler/xla/client/lib/prng.cc", "307": "tensorflow/compiler/xla/client/xla_builder.cc", "308": "tensorflow/compiler/xla/service/hlo_graph_dumper.cc", "309": "tensorflow/compiler/xla/service/hlo_alias_analysis.cc", "310": "tensorflow/compiler/xla/service/hlo_live_range.cc", "311": "tensorflow/compiler/xla/client/lib/pooling.cc", "312": "tensorflow/compiler/xla/client/lib/matrix.cc", "313": "tensorflow/compiler/xla/client/lib/tridiagonal.cc", "314": "tensorflow/compiler/xla/service/conditional_to_select.cc", "315": "tensorflow/compiler/xla/service/batch_dot_simplification.cc", "316": "tensorflow/compiler/xla/service/rng_expander.cc", "317": "tensorflow/compiler/xla/service/cholesky_expander.cc", "318": "tensorflow/compiler/xla/client/lib/svd.cc", "319": "tensorflow/compiler/xla/service/hlo_memory_scheduler.cc", "320": "tensorflow/compiler/xla/client/lib/qr.cc", "321": "tensorflow/compiler/xla/client/lib/self_adjoint_eig.cc", "322": "tensorflow/compiler/xla/service/convolution_group_converter.cc", "323": "tensorflow/compiler/xla/client/lib/math.cc", "324": "tensorflow/compiler/xla/service/hlo_creation_utils.cc", "325": "tensorflow/compiler/xla/service/layout_assignment.cc", "326": "tensorflow/compiler/xla/service/while_loop_constant_sinking.cc", "327": "tensorflow/compiler/xla/service/scatter_expander.cc", "328": "tensorflow/compiler/xla/service/heap_simulator.cc", "329": "tensorflow/compiler/xla/service/while_util.cc", "330": "tensorflow/compiler/xla/service/dynamic_dimension_inference.cc", "331": "tensorflow/compiler/xla/service/cpu/buffer_info_util.cc", "332": "tensorflow/compiler/xla/service/hlo_proto_util.cc", "333": "tensorflow/compiler/xla/service/llvm_ir/buffer_assignment_util.cc", "334": "tensorflow/compiler/xla/service/executable.cc", "335": "tensorflow/compiler/xla/service/hlo_pass_pipeline.cc", "336": "tensorflow/compiler/xla/service/triangular_solve_expander.cc", "337": "tensorflow/compiler/xla/service/compilation_cache.cc", "338": "tensorflow/compiler/xla/service/platform_util.cc", "339": "tensorflow/compiler/xla/service/copy_insertion.cc", "340": "tensorflow/compiler/xla/service/compiler.cc", "341": "tensorflow/compiler/xla/service/hlo_constant_folding.cc", "342": "tensorflow/compiler/xla/service/memory_space_assignment.cc", "343": "tensorflow/compiler/xla/service/transfer_manager.cc", "344": "tensorflow/compiler/xla/service/generic_transfer_manager.cc", "345": "tensorflow/compiler/xla/service/hlo_element_type_converter.cc", "346": "tensorflow/compiler/xla/service/execution_tracker.cc", "347": "tensorflow/compiler/xla/service/buffer_assignment.cc", "348": "tensorflow/compiler/xla/service/backend.cc", "349": "tensorflow/compiler/xla/service/op_expander_pass.cc", "350": "tensorflow/compiler/xla/service/indexed_array_analysis.cc", "351": "tensorflow/compiler/xla/service/hlo_evaluator_typed_visitor_complex128.cc", "352": "tensorflow/compiler/xla/service/tree_reduction_rewriter.cc", "353": "tensorflow/compiler/xla/service/rng_bit_generator_expander.cc", "354": "tensorflow/compiler/xla/service/hlo_evaluator_typed_visitor_complex64.cc", "355": "tensorflow/compiler/xla/service/while_loop_analysis.cc", "356": "tensorflow/compiler/xla/service/hlo_evaluator_typed_visitor_int32.cc", "357": "tensorflow/compiler/xla/service/hlo_evaluator_typed_visitor_uint16.cc", "358": "tensorflow/compiler/xla/service/hlo_evaluator_typed_visitor_uint8.cc", "359": "tensorflow/compiler/xla/service/algebraic_simplifier.cc", "360": "tensorflow/compiler/xla/service/hlo_evaluator_typed_visitor_int8.cc", "361": "tensorflow/compiler/xla/service/hlo_evaluator_typed_visitor_half.cc", "362": "tensorflow/compiler/xla/service/hlo_evaluator_typed_visitor_int64.cc", "363": "tensorflow/compiler/xla/service/allocation_tracker.cc", "364": "tensorflow/compiler/xla/service/hlo_evaluator_typed_visitor_bfloat16.cc", "365": "tensorflow/compiler/xla/service/while_loop_invariant_code_motion.cc", "366": "tensorflow/compiler/xla/service/hlo_evaluator_typed_visitor_float.cc", "367": "tensorflow/compiler/xla/service/hlo_evaluator_typed_visitor_int16.cc", "368": "tensorflow/compiler/xla/service/hlo_get_dimension_size_rewriter.cc", "369": "tensorflow/compiler/xla/service/hlo_evaluator_typed_visitor_uint32.cc", "370": "tensorflow/compiler/xla/service/dynamic_padder.cc", "371": "tensorflow/compiler/xla/service/dump.cc", "372": "tensorflow/compiler/xla/service/while_loop_simplifier.cc", "373": "tensorflow/compiler/xla/service/hlo_evaluator_typed_visitor_uint64.cc", "374": "tensorflow/compiler/xla/service/compile_only_service.cc", "375": "tensorflow/compiler/xla/service/hlo_evaluator_typed_visitor_bool.cc", "376": "tensorflow/compiler/xla/service/local_service.cc", "377": "tensorflow/compiler/xla/client/compile_only_client.cc", "378": "tensorflow/compiler/tf2xla/xla_expression.cc", "379": "tensorflow/compiler/xla/service/hlo_evaluator_typed_visitor_double.cc", "380": "tensorflow/compiler/xla/client/client_library.cc", "381": "tensorflow/compiler/xla/client/local_client.cc", "382": "tensorflow/compiler/jit/xla_tensor.cc", "383": "tensorflow/compiler/xla/service/service.cc", "384": "tensorflow/compiler/tf2xla/const_analysis.cc", "385": "tensorflow/compiler/tf2xla/kernels/conv_op_helpers.cc", "386": "tensorflow/compiler/jit/device_util.cc", "387": "tensorflow/compiler/tf2xla/xla_context.cc", "388": "tensorflow/compiler/tf2xla/kernels/tensor_list_utils.cc", "389": "tensorflow/compiler/tf2xla/xla_op_registry.cc", "390": "tensorflow/compiler/tf2xla/xla_helpers.cc", "391": "tensorflow/compiler/tf2xla/xla_compilation_device.cc", "392": "tensorflow/compiler/tf2xla/graph_compiler.cc", "393": "tensorflow/compiler/tf2xla/xla_resource.cc", "394": "tensorflow/compiler/tf2xla/xla_op_kernel.cc", "395": "tensorflow/compiler/tf2xla/lib/random.cc", "396": "tensorflow/compiler/jit/compilability_check_util.cc", "397": "tensorflow/compiler/xla/service/hlo_evaluator.cc", "398": "tensorflow/compiler/tf2xla/graph_compiler_util.cc", "399": "tensorflow/compiler/tf2xla/kernels/if_while_utils.cc", "400": "tensorflow/compiler/tf2xla/xla_compiler.cc", "401": "tensorflow/compiler/jit/increase_dynamism_for_auto_jit_pass.cc", "402": "tensorflow/compiler/jit/build_xla_ops_pass.cc", "403": "tensorflow/compiler/jit/encapsulate_subgraphs_pass.cc", "404": "tensorflow/compiler/jit/encapsulate_xla_computations_pass.cc", "405": "tensorflow/compiler/jit/extract_outside_compilation_pass.cc", "406": "tensorflow/compiler/aot/aot_only_var_handle_op.cc", "407": "tensorflow/compiler/tf2xla/tf2xla.cc", "408": "tensorflow/compiler/mlir/tensorflow/transforms/tpu_merge_variables_with_execute.cc", "409": "tensorflow/compiler/mlir/xla/transforms/legalize_tf_with_tf2xla.cc", "410": "tensorflow/compiler/mlir/tensorflow/translate/export_graphdef.cc", "411": "tensorflow/compiler/mlir/xla/mlir_hlo_to_hlo.cc", "412": "tensorflow/compiler/mlir/tensorflow/transforms/tpu_cluster_formation.cc", "413": "tensorflow/compiler/mlir/tensorflow/transforms/tpu_sharding_identification_pass.cc", "414": "tensorflow/compiler/mlir/tensorflow/transforms/tpu_variable_runtime_reformatting.cc", "415": "tensorflow/compiler/jit/xla_compilation_cache.cc", "416": "tensorflow/compiler/mlir/tensorflow/utils/compile_mlir_util.cc", "417": "tensorflow/compiler/mlir/tensorflow/translate/import_model.cc", "418": "tensorflow/compiler/jit/xla_device_context.cc", "419": "tensorflow/compiler/jit/xla_launch_util.cc", "420": "tensorflow/compiler/mlir/tensorflow/transforms/resource_op_lifting.cc", "421": "tensorflow/compiler/jit/xla_compile_on_demand_op.cc", "422": "tensorflow/compiler/jit/xla_device_ops.cc", "423": "tensorflow/compiler/mlir/tensorflow/transforms/einsum.cc", "424": "tensorflow/compiler/mlir/tensorflow/transforms/batchmatmul_to_einsum.cc", "425": "tensorflow/compiler/jit/xla_device.cc", "426": "tensorflow/compiler/jit/kernels/xla_ops.cc", "427": "tensorflow/compiler/jit/xla_kernel_creator_util.cc", "428": "tensorflow/compiler/xla/service/llvm_compiler.cc", "429": "tensorflow/core/distributed_runtime/eager/remote_mgr.cc", "430": "tensorflow/core/common_runtime/eager/execute.cc", "431": "tensorflow/compiler/mlir/tensorflow/transforms/promote_resources_to_args.cc", "432": "tensorflow/core/distributed_runtime/eager/eager_service_impl.cc", "433": "tensorflow/core/distributed_runtime/rpc/eager/grpc_eager_service_impl.cc", "434": "tensorflow/core/distributed_runtime/rpc/grpc_server_lib.cc", "435": "tensorflow/compiler/xla/service/cpu/xfeed_manager.cc", "436": "tensorflow/compiler/xla/service/gpu/target_util.cc", "437": "tensorflow/compiler/aot/embedded_protocol_buffers.cc", "438": "tensorflow/compiler/xla/service/llvm_ir/math_ops.cc", "439": "tensorflow/compiler/xla/service/llvm_ir/tuple_ops.cc", "440": "tensorflow/compiler/xla/service/cpu/ir_emission_utils.cc", "441": "tensorflow/compiler/xla/service/llvm_ir/ir_array.cc", "442": "tensorflow/compiler/xla/service/cpu/vector_support_library.cc", "443": "tensorflow/compiler/xla/service/llvm_ir/llvm_util.cc", "444": "tensorflow/compiler/xla/service/llvm_ir/llvm_loop.cc", "445": "tensorflow/compiler/xla/service/cpu/cpu_transfer_manager.cc", "446": "tensorflow/compiler/xla/service/llvm_ir/loop_emitter.cc", "447": "tensorflow/python/tfcompile_wrapper.cc", "448": "tensorflow/compiler/xla/service/cpu/conv_canonicalization.cc", "449": "tensorflow/compiler/xla/service/cpu/cpu_runtime.cc", "450": "tensorflow/compiler/xla/service/llvm_ir/kernel_support_library.cc", "451": "tensorflow/compiler/xla/service/llvm_ir/alias_analysis.cc", "452": "tensorflow/compiler/xla/service/cpu/llvm_ir_runtime.cc", "453": "tensorflow/compiler/xla/service/cpu/ir_function.cc", "454": "tensorflow/compiler/xla/service/cpu/parallel_loop_emitter.cc", "455": "tensorflow/compiler/xla/service/gpu/parallel_loop_emitter.cc", "456": "tensorflow/compiler/xla/service/cpu/tiled_dot_emitter.cc", "457": "tensorflow/compiler/xla/service/cpu/cpu_instruction_fusion.cc", "458": "tensorflow/compiler/xla/service/cpu/compiler_functor.cc", "459": "tensorflow/compiler/xla/service/llvm_ir/fused_ir_emitter.cc", "460": "tensorflow/compiler/xla/service/cpu/dot_op_emitter.cc", "461": "tensorflow/compiler/xla/service/elemental_ir_emitter.cc", "462": "tensorflow/compiler/xla/service/llvm_ir/dynamic_update_slice_util.cc", "463": "tensorflow/compiler/xla/service/cpu/cpu_layout_assignment.cc", "464": "tensorflow/compiler/xla/service/cpu/simple_orc_jit.cc", "465": "tensorflow/compiler/xla/service/cpu/cpu_executable.cc", "466": "tensorflow/compiler/xla/service/cpu/parallel_task_assignment.cc", "467": "tensorflow/compiler/xla/service/cpu/elemental_ir_emitter.cc", "468": "tensorflow/core/profiler/internal/tfprof_op.cc", "469": "tensorflow/compiler/jit/xla_cpu_device.cc", "470": "tensorflow/python/lib/core/ndarray_tensor.cc", "471": "tensorflow/compiler/xla/service/cpu/ir_emitter.cc", "472": "tensorflow/core/profiler/internal/tfprof_code.cc", "473": "tensorflow/compiler/xla/service/cpu/cpu_compiler.cc", "474": "tensorflow/core/profiler/internal/tfprof_stats.cc", "475": "tensorflow/core/profiler/internal/print_model_analysis.cc", "476": "tensorflow/compiler/aot/codegen.cc", "477": "tensorflow/compiler/aot/compile.cc", "478": "tensorflow/compiler/tf2xla/mlir_tf2xla.cc", "479": "tensorflow/python/eager/pywrap_tfe_src.cc", "480": "tensorflow/compiler/mlir/lite/utils/lstm_utils.cc", "481": "tensorflow/compiler/mlir/lite/transforms/prepare_composite_functions_tf.cc", "482": "tensorflow/compiler/mlir/lite/transforms/dilated_conv.cc", "483": "tensorflow/lite/delegates/nnapi/quant_lstm_sup.cc", "484": "tensorflow/compiler/mlir/lite/transforms/dense_to_sparse.cc", "485": "tensorflow/compiler/mlir/lite/transforms/lower_static_tensor_list.cc", "486": "tensorflow/lite/delegates/nnapi/nnapi_delegate.cc", "487": "tensorflow/compiler/mlir/lite/transforms/prepare_tf.cc", "488": "tensorflow/compiler/mlir/lite/transforms/legalize_tf.cc", "489": "tensorflow/compiler/mlir/lite/transforms/optimize.cc", "490": "tensorflow/lite/python/interpreter_wrapper/interpreter_wrapper.cc", "491": "tensorflow/lite/tools/verifier.cc", "492": "tensorflow/lite/tools/optimize/model_utils.cc", "493": "tensorflow/lite/tools/optimize/quantization_utils.cc", "494": "tensorflow/lite/tools/versioning/op_version.cc", "495": "tensorflow/lite/tools/versioning/runtime_version.cc", "496": "tensorflow/lite/tools/optimize/quantize_model.cc", "497": "tensorflow/lite/tools/optimize/quantize_weights.cc", "498": "tensorflow/lite/python/optimize/calibration_wrapper.cc", "499": "tensorflow/lite/toco/tflite/import.cc", "500": "tensorflow/compiler/mlir/lite/flatbuffer_import.cc", "501": "tensorflow/compiler/mlir/lite/ir/tfl_ops.cc", "502": "tensorflow/compiler/mlir/lite/flatbuffer_export.cc", "503": "tensorflow/compiler/mlir/lite/python/saved_model_to_tfl_flatbuffer.cc", "504": "tensorflow/lite/experimental/microfrontend/lib/fft.cc", "505": "tensorflow/core/tpu/tpu_embedding_optimization_parameters_utils.cc", "506": "tensorflow/core/platform/default/logging.cc", "507": "tensorflow/core/platform/default/logging.cc", "508": "tensorflow/core/platform/protobuf.cc", "509": "tensorflow/core/framework/allocator_registry.cc", "510": "tensorflow/core/framework/cpu_allocator_impl.cc", "511": "tensorflow/core/kernels/lookup_util.cc", "512": "tensorflow/lite/delegates/nnapi/nnapi_delegate.cc", "513": "tensorflow/core/framework/allocator_registry.cc", "514": "tensorflow/core/common_runtime/lower_case_op.cc", "515": "tensorflow/core/data/service/compression_utils.cc", "516": "tensorflow/core/kernels/data/prefetch_autotuner.cc", "517": "tensorflow/lite/tools/versioning/runtime_version.cc", "518": "tensorflow/compiler/xla/service/hlo_evaluator.cc", "519": "tensorflow/compiler/tf2xla/xla_resource.cc", "520": "tensorflow/core/kernels/stack.cc", "521": "tensorflow/core/profiler/internal/parse_annotation.cc", "522": "tensorflow/core/framework/cpu_allocator_impl.cc", "523": "tensorflow/lite/experimental/microfrontend/lib/fft.cc", "524": "tensorflow/lite/toco/graph_transformations/ensure_uint8_weights_safe_for_fast_int8_kernels.cc", "525": "tensorflow/compiler/xla/service/while_loop_invariant_code_motion.cc", "526": "tensorflow/core/tpu/tpu_embedding_optimization_parameters_utils.cc"} \ No newline at end of file diff --git a/tensorflow/core/data/compression_utils.cc b/tensorflow/core/data/compression_utils.cc index d132bdca8da..2ab51712580 100644 --- a/tensorflow/core/data/compression_utils.cc +++ b/tensorflow/core/data/compression_utils.cc @@ -116,7 +116,7 @@ Status UncompressElement(const CompressedElement& compressed, compressed_data.data(), compressed_data.size(), &uncompressed_size)) { return errors::Internal("Could not get snappy uncompressed length"); } - if (uncompressed_size != total_size) { + if (uncompressed_size != static_cast(total_size)) { return errors::Internal( "Uncompressed size mismatch. Snappy expects ", uncompressed_size, " whereas the tensor metadata suggests ", total_size); diff --git a/tensorflow/core/grappler/optimizers/arithmetic_optimizer.cc b/tensorflow/core/grappler/optimizers/arithmetic_optimizer.cc index 520346b0166..c8015a6e50c 100644 --- a/tensorflow/core/grappler/optimizers/arithmetic_optimizer.cc +++ b/tensorflow/core/grappler/optimizers/arithmetic_optimizer.cc @@ -598,7 +598,7 @@ class AddOpsRewriteStage : public ArithmeticNodesGroupOptimizerStage { std::deque add_ops; // Prepare leaf AddN nodes for inputs of equal shape - for (int i = 0; i < shapes.size(); ++i) { + for (int i = 0, iter_limit = shapes.size(); i < iter_limit; ++i) { const auto node_name = leaf_node_name(i); const auto& inputs = shape_sig_to_inputs[ShapeSignature(shapes[i])]; add_ops.push_back(AddInputsOfSymbolicallyEqualShape(*group.root_node, @@ -750,7 +750,7 @@ class HoistCommonFactorOutOfAggregation : public ArithmeticOptimizerStage { ctx().node_map->AddOutput(new_add_node->name(), new_outer_node->name()); // Hoist non-shared factors up into the new AddN node. - for (int i = 0; i < unique_factors.size(); ++i) { + for (int i = 0, iter_limit = unique_factors.size(); i < iter_limit; ++i) { const string& unique_factor_i = unique_factors[i]; new_add_node->set_input(i, unique_factor_i); ctx().node_map->AddOutput(unique_factor_i, new_add_node->name()); @@ -1190,7 +1190,7 @@ class RemoveIdentityTranspose : public ArithmeticOptimizerStage { if (a.size() != b.size()) { return false; } - for (int i = 0; i < a.size(); ++i) { + for (int i = 0, iter_limit = a.size(); i < iter_limit; ++i) { if (a[b[i]] != i) { return false; } @@ -1199,7 +1199,7 @@ class RemoveIdentityTranspose : public ArithmeticOptimizerStage { } bool IsIdentityPermutation(const std::vector& perm) { - for (int64 i = 0; i < perm.size(); ++i) { + for (int64 i = 0, iter_limit = perm.size(); i < iter_limit; ++i) { if (i != perm[i]) { return false; } @@ -1500,7 +1500,8 @@ class HoistCWiseUnaryChainsStage : public ArithmeticOptimizerStage { for (int i = start; i < end; ++i) { unique_inputs.insert(node.input(i)); } - return unique_inputs.size() == n; + int unique_input_size = unique_inputs.size(); + return unique_input_size == n; } // Returns the length of the common unary chain of ops that can be @@ -3248,14 +3249,15 @@ class RemoveStackSliceSameAxis : public ArithmeticOptimizerStage { slice_begin_vec.size(), ") and size (", slice_size_vec.size(), ") vectors."); } + int slice_begin_vec_size = slice_begin_vec.size(); if (!pack_output_shape.unknown_rank() && - slice_begin_vec.size() != pack_output_shape.dims()) { + slice_begin_vec_size != pack_output_shape.dims()) { return Status::OK(); } - if (pack_axis >= slice_begin_vec.size()) { + if (pack_axis >= slice_begin_vec_size) { return errors::InvalidArgument( "Input to node ", node->name(), " had pack_axis ", pack_axis, - " but rank was ", slice_begin_vec.size(), "."); + " but rank was ", slice_begin_vec_size, "."); } *slice_start_value = slice_begin_vec[pack_axis]; @@ -3264,7 +3266,7 @@ class RemoveStackSliceSameAxis : public ArithmeticOptimizerStage { return Status::OK(); } - for (size_t i = 0; i < slice_begin_vec.size(); ++i) { + for (int i = 0; i < slice_begin_vec_size; ++i) { if (i != pack_axis) { if (slice_begin_vec[i] != 0 || !(slice_size_vec[i] == -1 || @@ -3352,7 +3354,7 @@ class RemoveStackSliceSameAxis : public ArithmeticOptimizerStage { int begin_index = -1; int64 begin_value = 0; - for (int i = 0; i < slice_begin_vec.size(); ++i) { + for (int i = 0, iter_limit = slice_begin_vec.size(); i < iter_limit; ++i) { const int64 v = slice_begin_vec[i]; if (v != 0) { if (begin_index != -1) { @@ -3366,7 +3368,7 @@ class RemoveStackSliceSameAxis : public ArithmeticOptimizerStage { int end_index = -1; int64 end_value = 0; - for (int i = 0; i < slice_end_vec.size(); ++i) { + for (int i = 0, iter_limit = slice_begin_vec.size(); i < iter_limit; ++i) { const int64 v = slice_end_vec[i]; if (v != pack_output_shape.dim_size(i)) { if (end_index != -1) { diff --git a/tensorflow/core/grappler/optimizers/constant_folding.cc b/tensorflow/core/grappler/optimizers/constant_folding.cc index d912eb7857b..f42340f9d09 100644 --- a/tensorflow/core/grappler/optimizers/constant_folding.cc +++ b/tensorflow/core/grappler/optimizers/constant_folding.cc @@ -479,7 +479,7 @@ Status ConstantFolding::MaterializeShapes(const GraphProperties& properties) { CHECK_EQ(op, "ShapeN"); CHECK_EQ(input.size(), output.size()); const NodeDef* const shape_n_node = node; - for (int port_idx = 0; port_idx < output.size(); ++port_idx) { + for (int port_idx = 0, idx_limit = output.size(); port_idx < idx_limit; ++port_idx) { const DataType type = output[port_idx].dtype(); CHECK(type == DT_INT32 || type == DT_INT64); const PartialTensorShape shape(input[port_idx].shape()); @@ -641,12 +641,12 @@ Status ConstantFolding::MaterializeBroadcastGradientArgs( // These extra dims could be equal to 1, in which case there is no // broadcasting. It could also be greater than 1, in which case there would // be broadcasting. Since we don't know, we'll just punt. - for (int i = common_dims; i < shape1.size(); ++i) { + for (int i = common_dims, iter_limit = shape1.size(); i < iter_limit; ++i) { if (shape1[i] < 0) { return Status::OK(); } } - for (int i = common_dims; i < shape2.size(); ++i) { + for (int i = common_dims, iter_limit = shape2.size(); i < iter_limit; ++i) { if (shape2[i] < 0) { return Status::OK(); } @@ -1165,7 +1165,7 @@ bool IsValidConstShapeForMulConvPushDown( // If the const is a scalar, or it has fewer or same number of dimensions // than the filter and it only has single element, the optimization should // work. - if (mul_const_input_shape.dim_size() <= data_format.size() && + if (mul_const_input_shape.dim_size() <= static_cast(data_format.size()) && TensorShape(mul_const_input_shape).num_elements() == 1) { return true; } @@ -1461,7 +1461,7 @@ Status ConstantFolding::FoldNode(NodeDef* node, GraphDef* output_graph, VLOG(2) << "Folded node: " << SummarizeNodeDef(*node); NodeDef* constant_output = nullptr; - for (int i = 0; i < const_nodes.size(); i++) { + for (int i = 0, iter_limit = const_nodes.size(); i < iter_limit; i++) { NodeDef* const_node = &const_nodes[i]; VLOG(3) << "Generated constant node: " << SummarizeNodeDef(*const_node); if (const_node->name().empty()) { @@ -1549,7 +1549,7 @@ Status ConstantFolding::FoldNode(NodeDef* node, GraphDef* output_graph, constant_output->name()); *output->mutable_input(i) = AsControlDependency(*constant_output); } - } else if (port < const_nodes.size() && + } else if (port < static_cast(const_nodes.size()) && !const_nodes[port].name().empty()) { // Replace alive outputs with the corresponding constant. node_map_->UpdateInput(output->name(), NodeName(output->input(i)), @@ -2068,7 +2068,8 @@ Status ConstantFolding::RemoveShuffleOrTranspose( permutation.push_back(permutation_tensor.vec()(j)); } } - if (permutation.size() != shape.dim_size()) { + int permutation_size = permutation.size(); + if (permutation_size != shape.dim_size()) { // Number of elements in perm should be same as dim_size. Skip if not. return Status::OK(); } @@ -2245,9 +2246,10 @@ Status ConstantFolding::SimplifyStridedSlice(const GraphProperties& properties, // as many as expanded_ellipsis_indices.size() axes during computation. // We need to subtract this number from j. int i = j; + int expanded_ellipsis_indices_size = expanded_ellipsis_indices.size(); if (ellipsis_index != -1 && - j >= ellipsis_index + expanded_ellipsis_indices.size()) { - i = j - expanded_ellipsis_indices.size(); + j >= ellipsis_index + expanded_ellipsis_indices_size) { + i = j - expanded_ellipsis_indices_size; } int b = begin.dtype() == DT_INT32 ? begin.vec()(i) : begin.vec()(i); @@ -3479,15 +3481,16 @@ bool ConstantFolding::PartialAssocOpConstFolding(GraphDef* optimized_graph, } // Promote AccumulateNV2 with all constant inputs to AddN, since it is // a fake node that cannot be constant folded by itself. - if (const_inputs.size() == num_non_control_inputs && + int const_inputs_size = const_inputs.size(); + if (const_inputs_size == num_non_control_inputs && node->op() == "AccumulateNV2") { node->set_op("AddN"); node->mutable_attr()->erase("shape"); return true; } const string new_node_name = OptimizedNodeName( - *node, strings::StrCat("_partial_split_", const_inputs.size())); - if (const_inputs.size() > 1 && const_inputs.size() < num_non_control_inputs && + *node, strings::StrCat("_partial_split_", const_inputs_size)); + if (const_inputs_size > 1 && const_inputs_size < num_non_control_inputs && !node_map_->NodeExists(new_node_name)) { NodeDef* added_node = optimized_graph->add_node(); *added_node = *node; diff --git a/tensorflow/core/grappler/optimizers/generic_layout_optimizer_transposer.cc b/tensorflow/core/grappler/optimizers/generic_layout_optimizer_transposer.cc index ab7d8fcd6cf..63239082134 100644 --- a/tensorflow/core/grappler/optimizers/generic_layout_optimizer_transposer.cc +++ b/tensorflow/core/grappler/optimizers/generic_layout_optimizer_transposer.cc @@ -242,7 +242,7 @@ Status Transposer::CreateConstPermNode(TransposeContext* context, AttrValue attr_tensor; Tensor tensor(DT_INT32, TensorShape({4})); - for (int i = 0; i < permutation.size(); i++) { + for (int i = 0, iter_limit = permutation.size(); i < iter_limit; i++) { tensor.flat()(i) = permutation[i]; } tensor.AsProtoTensorContent(attr_tensor.mutable_tensor()); @@ -538,10 +538,11 @@ bool Transposer::IsFaninPortDimsNIfConst(const utils::MutableNodeView& node, if (!tensor.FromProto(value_attr->tensor())) { return false; } - if (tensor.dims() != dims.size()) { + int dims_size = dims.size(); + if (tensor.dims() != dims_size) { return false; } - for (int i = 0; i < dims.size(); ++i) { + for (int i = 0; i < dims_size; ++i) { if (tensor.dim_size(i) != dims[i]) { return false; } @@ -863,12 +864,13 @@ inline bool IsValidConstPermTransposeNode(const utils::MutableNodeView& node, if (!GetValueAttrFromConstInputNode(node, IsTranspose, 1, &tensor)) { return false; } - if (tensor.NumElements() != permutation.size()) { + int permutation_size = permutation.size(); + if (tensor.NumElements() != permutation_size) { return false; } const auto& tensor_data = tensor.unaligned_flat(); - for (int i = 0; i < permutation.size(); i++) { + for (int i = 0; i < permutation_size; i++) { if (permutation[i] != tensor_data(i)) { return false; } @@ -1229,10 +1231,11 @@ bool ReduceTransposer::KeepDims(const utils::MutableNodeView& node) { bool ReduceTransposer::IsAlongAxis(const Tensor& tensor, absl::Span axis, int rank) { - if (tensor.dims() != 1 || tensor.dim_size(0) != axis.size()) { + int axis_size = axis.size(); + if (tensor.dims() != 1 || tensor.dim_size(0) != axis_size) { return false; } - for (int i = 0; i < axis.size(); ++i) { + for (int i = 0; i < axis_size; ++i) { int local_axis = tensor.flat()(i); if (local_axis < 0) { local_axis += rank; @@ -1444,12 +1447,13 @@ bool SqueezeTransposer::IsAlongAxis(const AttrValue& attr, int rank) const { const auto& list = attr.list(); // If list is empty, Squeeze op will squeeze all dimensions of size 1. + int axis_size = axis.size(); if (list.i_size() == 0) { return true; - } else if (list.i_size() != axis.size()) { + } else if (list.i_size() != axis_size) { return false; } - for (int i = 0; i < axis.size(); ++i) { + for (int i = 0; i < axis_size; ++i) { int local_axis = list.i(i); if (local_axis < 0) { local_axis += rank; @@ -1563,7 +1567,7 @@ Status StridedSliceTransposer::PermuteMask(TransposeContext* context, return errors::InvalidArgument("invalid mask value: ", mask_i); } int result = 0; - for (int i = 0; i < context->src_to_dst.size(); i++) { + for (int i = 0, iter_limit = context->src_to_dst.size(); i < iter_limit; i++) { const int final_pos = context->src_to_dst[i]; const int position_mask = 1 << final_pos; const int bit_i = (mask_i & position_mask) >> final_pos; diff --git a/tensorflow/core/grappler/optimizers/generic_layout_optimizer_transposer.h b/tensorflow/core/grappler/optimizers/generic_layout_optimizer_transposer.h index b518c32d8ec..bb00e965872 100644 --- a/tensorflow/core/grappler/optimizers/generic_layout_optimizer_transposer.h +++ b/tensorflow/core/grappler/optimizers/generic_layout_optimizer_transposer.h @@ -528,11 +528,12 @@ template Status PermuteSingle(absl::string_view location, absl::Span permutation, T* values) { DCHECK(values != nullptr); - if (values->size() != permutation.size()) { + int permutation_size = permutation.size(); + if (values->size() != permutation_size) { return Status(tensorflow::error::Code::INVALID_ARGUMENT, absl::StrCat("Size of values ", values->size(), " does not match size of permutation ", - permutation.size(), " @ ", location)); + permutation_size, " @ ", location)); } typedef typename T::value_type V; std::vector elements(values->begin(), values->end()); @@ -549,11 +550,12 @@ template Status PermuteDouble(absl::string_view location, absl::Span permutation, T* values) { DCHECK(values != nullptr); - if (values->size() != permutation.size() * 2) { + int permutation_size = permutation.size(); + if (values->size() != permutation_size * 2) { return Status(tensorflow::error::Code::INVALID_ARGUMENT, absl::StrCat("Size of values ", values->size(), " does not match twice the size of permutation ", - permutation.size(), " @ ", location)); + permutation_size, " @ ", location)); } typedef typename T::value_type V; std::vector elements(values->begin(), values->end()); diff --git a/tensorflow/core/grappler/optimizers/graph_optimizer_stage.cc b/tensorflow/core/grappler/optimizers/graph_optimizer_stage.cc index 97033a180a6..4e955db2f5a 100644 --- a/tensorflow/core/grappler/optimizers/graph_optimizer_stage.cc +++ b/tensorflow/core/grappler/optimizers/graph_optimizer_stage.cc @@ -58,7 +58,7 @@ Status GetTensorProperties(const GraphOptimizerContext& ctx, const auto& output_properties = ctx.graph_properties->GetOutputProperties(tensor_id.node()); - auto num_outputs = output_properties.size(); + int num_outputs = output_properties.size(); if (num_outputs == 0 || tensor_id.index() > num_outputs - 1) { return errors::InvalidArgument( diff --git a/tensorflow/core/grappler/optimizers/implementation_selector.cc b/tensorflow/core/grappler/optimizers/implementation_selector.cc index 2b0a27aaa2d..51d61cfef2e 100644 --- a/tensorflow/core/grappler/optimizers/implementation_selector.cc +++ b/tensorflow/core/grappler/optimizers/implementation_selector.cc @@ -130,7 +130,7 @@ string FindForwardNode(utils::MutableNodeView* backward_node) { void UpdateForwardIdentityNodeDtype(utils::MutableNodeView* forward_node, const DataTypeVector& dtypes) { const auto& fanouts_vector = forward_node->GetRegularFanouts(); - for (int pos = 0; pos < fanouts_vector.size(); ++pos) { + for (int pos = 0, pos_limit = fanouts_vector.size(); pos < pos_limit; ++pos) { const auto& fanouts_at_pos = fanouts_vector[pos]; for (const auto& fanout : fanouts_at_pos) { if ("Identity" == fanout.node_view()->GetOp()) { diff --git a/tensorflow/core/grappler/optimizers/memory_optimizer.cc b/tensorflow/core/grappler/optimizers/memory_optimizer.cc index 867433dcff5..c6fd2b48ac3 100644 --- a/tensorflow/core/grappler/optimizers/memory_optimizer.cc +++ b/tensorflow/core/grappler/optimizers/memory_optimizer.cc @@ -468,7 +468,7 @@ void RecomputationRewritingPass(RewriterConfig::MemOptType optimization_level, // with "gradients/" or contains "/gradients/". return absl::StartsWith(node.name(), recomputation_targets_name_scope) || - node.name().find("/" + recomputation_targets_name_scope) != -1; + static_cast(node.name().find("/" + recomputation_targets_name_scope)) != -1; }; if (optimization_level == RewriterConfig::RECOMPUTATION_HEURISTICS || diff --git a/tensorflow/core/grappler/optimizers/scoped_allocator_optimizer.cc b/tensorflow/core/grappler/optimizers/scoped_allocator_optimizer.cc index 464a2c17197..7a0079d2b4c 100644 --- a/tensorflow/core/grappler/optimizers/scoped_allocator_optimizer.cc +++ b/tensorflow/core/grappler/optimizers/scoped_allocator_optimizer.cc @@ -73,7 +73,8 @@ bool HasOpName(const string& node_name, const string& op_name) { Status GetOutputDataType( const std::vector& output_props, int output_index, DataType* dtype) { - if (output_index >= output_props.size()) { + int output_props_size = output_props.size(); + if (output_index >= output_props_size) { return errors::Internal("Invalid output index ", output_index, " size of output_props ", output_props.size()); } @@ -520,7 +521,7 @@ class UnaryElementwiseRewriter : public ScopedAllocatorOptimizer::Rewriter { // Add control edges from the ScopedAllocatorOp to all of the // input nodes and mark them for allocation from backing tensor. - for (int i = 0; i < inputs.size(); ++i) { + for (int i = 0, iter_limit = inputs.size(); i < iter_limit; ++i) { auto& nd = inputs[i]; if (IsArg(*nd.from_node_def)) { return errors::Internal( @@ -547,7 +548,7 @@ class UnaryElementwiseRewriter : public ScopedAllocatorOptimizer::Rewriter { std::vector inputs_to_first; LOG_WARNING_AND_RETURN_IF_ERROR(GetDataInputs( graph, sa_opti->node_map(), nd.from_node_def, &inputs_to_first)); - for (int i = 0; i < inputs_to_first.size(); ++i) { + for (int i = 0, iter_limit = inputs_to_first.size(); i < iter_limit; ++i) { if (fanout.find(inputs_to_first[i].from_node_def) != fanout.end()) { VLOG(2) << "Found node " << inputs_to_first[i].from_node_def->name() << " in the fanout of " << sa_name; @@ -587,7 +588,7 @@ class UnaryElementwiseRewriter : public ScopedAllocatorOptimizer::Rewriter { VLOG(2) << "BuildSAConcatNode " << sac_name; // control input: edge name -> source node name absl::flat_hash_map sac_ctl_inputs; - for (int i = 0; i < ops.size(); ++i) { + for (int i = 0, iter_limit = ops.size(); i < iter_limit; ++i) { NodeDef* old_op = ops[i]; for (const string& old_op_input : old_op->input()) { int position = 0; @@ -708,7 +709,7 @@ class UnaryElementwiseRewriter : public ScopedAllocatorOptimizer::Rewriter { const std::set& op_instance_names, const string& op_name, const string& sas_name) { VLOG(2) << "RewireSubgraph"; - for (int op_idx = 0; op_idx < ops.size(); ++op_idx) { + for (int op_idx = 0, idx_limit = ops.size(); op_idx < idx_limit; ++op_idx) { NodeDef* old_op = ops[op_idx]; // Copy the output node set since we'll be modifying the version // maintained by NodeMap in the loop. diff --git a/tensorflow/core/grappler/optimizers/shape_optimizer.cc b/tensorflow/core/grappler/optimizers/shape_optimizer.cc index 69de1cde4ca..656c1a1db1c 100644 --- a/tensorflow/core/grappler/optimizers/shape_optimizer.cc +++ b/tensorflow/core/grappler/optimizers/shape_optimizer.cc @@ -99,7 +99,8 @@ Status ShapeOptimizer::Optimize(Cluster* cluster, const GrapplerItem& item, } const auto& prop = properties.GetOutputProperties(reduce_indices.node->name()); - if (prop.size() <= reduce_indices.port_id) { + int prop_size = prop.size(); + if (prop_size <= reduce_indices.port_id) { continue; } const TensorShapeProto& reduction_indices_shape = diff --git a/tensorflow/core/kernels/data/single_threaded_executor.cc b/tensorflow/core/kernels/data/single_threaded_executor.cc index 3a16f1018dd..eeb1ffd5ad0 100644 --- a/tensorflow/core/kernels/data/single_threaded_executor.cc +++ b/tensorflow/core/kernels/data/single_threaded_executor.cc @@ -51,8 +51,8 @@ class SingleThreadedExecutorImpl : public Executor { std::vector ordered_nodes; ordered_nodes.reserve(graph.num_nodes()); GetReversePostOrder(graph, &ordered_nodes); - - if (ordered_nodes.size() != graph.num_nodes()) { + int ordered_nodes_size = ordered_nodes.size(); + if (ordered_nodes_size != graph.num_nodes()) { return errors::InvalidArgument("Graph had ", graph.num_nodes(), " but reverse post-order had ", ordered_nodes.size()); diff --git a/tensorflow/core/kernels/initializable_lookup_table.cc b/tensorflow/core/kernels/initializable_lookup_table.cc index 196c2fe95a3..48041526022 100644 --- a/tensorflow/core/kernels/initializable_lookup_table.cc +++ b/tensorflow/core/kernels/initializable_lookup_table.cc @@ -74,7 +74,7 @@ Status InitializableLookupTable::Initialize(InitTableIterator& iter) { Status InitializableLookupTable::AreEntriesSame(const InitTableIterator& iter, bool* result) { - *result = iter.total_size() == size(); + *result = static_cast(iter.total_size()) == size(); return Status::OK(); } diff --git a/tensorflow/core/kernels/remote_fused_graph_execute_utils.cc b/tensorflow/core/kernels/remote_fused_graph_execute_utils.cc index b57d6163cac..e3f220b9ff4 100644 --- a/tensorflow/core/kernels/remote_fused_graph_execute_utils.cc +++ b/tensorflow/core/kernels/remote_fused_graph_execute_utils.cc @@ -431,7 +431,8 @@ RemoteFusedGraphExecuteUtils::AddOutputTensorShapeTypeByTensorShapeMap( if (data_types.empty()) { return false; } - CHECK(data_types.size() > port); + int data_types_size = data_types.size(); + CHECK(data_types_size > port); *data_type = data_types.at(port); *shape = shapes.at(port); return true; @@ -788,7 +789,8 @@ RemoteFusedGraphExecuteUtils::BuildRemoteFusedGraphExecuteOpNode( ++input_count; } } - CHECK(input_count == 0 || input_count == node->in_edges().size()) + int node_in_edges_size = node->in_edges().size(); + CHECK(input_count == 0 || input_count == node_in_edges_size) << "Invalid input_count(" << input_count << ", " << node->in_edges().size() << ") " << node_name; @@ -968,10 +970,10 @@ RemoteFusedGraphExecuteUtils::BuildRemoteFusedGraphExecuteOpNode( border_inputs, border_outputs, require_shape_type, &graph, &fused_node)); for (const Node* node : graph.nodes()) { - for (int i = 0; i < node->num_inputs(); ++i) { + for (int i = 0, iter_limit = node->num_inputs(); i < iter_limit; ++i) { const Edge* edge = nullptr; TF_RETURN_IF_ERROR(node->input_edge(i, &edge)); - for (int j = 0; j < border_outputs.size(); ++j) { + for (int j = 0, second_iter_limit = border_outputs.size(); j < second_iter_limit; ++j) { const string& output = border_outputs.at(j); const TensorId tid = ParseTensorName(output); const string output_name(tid.first); @@ -1333,8 +1335,9 @@ RemoteFusedGraphExecuteUtils::FuseRemoteGraphByPlacedArguments( /* static */ Status RemoteFusedGraphExecuteUtils::CopyByteArrayToTensor( const void* src_ptr, const int src_size, Tensor* tensor) { - CHECK(tensor->TotalBytes() >= src_size) - << tensor->TotalBytes() << ", " << src_size; + int tensor_TotalBytes = tensor->TotalBytes(); + CHECK(tensor_TotalBytes >= src_size) + << tensor_TotalBytes << ", " << src_size; void* dst_ptr; switch (tensor->dtype()) { case DT_FLOAT: diff --git a/tensorflow/core/kernels/stack.cc b/tensorflow/core/kernels/stack.cc index a30729c3b2f..dd20902a26d 100644 --- a/tensorflow/core/kernels/stack.cc +++ b/tensorflow/core/kernels/stack.cc @@ -57,7 +57,8 @@ class Stack : public ResourceBase { Status Push(const TensorAndAllocation& value) { mutex_lock l(mu_); TF_RETURN_IF_ERROR(CheckNotClosed()); - if (max_size_ >= 0 && stack_.size() >= max_size_) { + int stack_size = stack_.size(); + if (max_size_ >= 0 && stack_size >= max_size_) { return errors::InvalidArgument("Stack[", stack_name_, "] overflowed ", "its max_size (", max_size_, ")"); } diff --git a/tensorflow/core/tpu/tpu_embedding_optimization_parameters_utils.cc b/tensorflow/core/tpu/tpu_embedding_optimization_parameters_utils.cc index 9daf9a2cef7..c40fea7c61c 100644 --- a/tensorflow/core/tpu/tpu_embedding_optimization_parameters_utils.cc +++ b/tensorflow/core/tpu/tpu_embedding_optimization_parameters_utils.cc @@ -336,7 +336,7 @@ Status LoadOpShapeFunction::operator()( }); std::vector inputs(user_param_count); int input_index = 0; - for (int i = 0; i < state_variable_specs.size(); ++i) { + for (int i = 0, iter_limit = state_variable_specs.size(); i < iter_limit; ++i) { if (state_variable_specs[i].has_user_defined() || is_debug_op_) { std::vector input_temp; TF_RETURN_IF_ERROR(c->input(state_variable_specs[i].name(), &input_temp)); @@ -388,7 +388,7 @@ Status RetrieveOpShapeFunction::operator()( TF_RETURN_IF_ERROR(c->GetAttr("num_shards", &num_shards)); int shard_id; TF_RETURN_IF_ERROR(c->GetAttr("shard_id", &shard_id)); - for (int j = 0; j < state_variable_specs.size(); ++j) { + for (int j = 0, iter_limit = state_variable_specs.size(); j < iter_limit; ++j) { if (state_variable_specs[j].has_user_defined() || is_debug_op_) { auto shape = c->MakeShape( std::vector(2, c->UnknownDim())); diff --git a/tensorflow/core/util/bcast.h b/tensorflow/core/util/bcast.h index 7bb8ea18ad3..a79a3f08622 100644 --- a/tensorflow/core/util/bcast.h +++ b/tensorflow/core/util/bcast.h @@ -139,7 +139,7 @@ BCastList::BCastList(const BCastList::Vec (&x)[N], if (x[i] != x[0]) { all_equal = false; } - if (x[i].size() > largest_rank) { + if ( static_cast(x[i].size()) > largest_rank) { largest_rank = x[i].size(); } } @@ -176,7 +176,7 @@ BCastList::BCastList(const BCastList::Vec (&x)[N], // 1-extend and align all vectors. for (int i = 0; i < N; ++i) { - if (copy[i].size() < largest_rank) { + if (static_cast(copy[i].size()) < largest_rank) { copy[i].resize(largest_rank, 1); } } From 2399d37889c044fb6fb52fb3abc2bb5d41e64a28 Mon Sep 17 00:00:00 2001 From: Mihai Maruseac Date: Mon, 22 Jun 2020 19:18:54 -0700 Subject: [PATCH 519/867] Create `RELEASE.md` stub. Now, every PR and commit can add to the release notes before the branch is cut for a new release. PiperOrigin-RevId: 317783211 Change-Id: If65da4b956f83940c7bd539d40f915cfd9af4db0 --- RELEASE.md | 54 ++++++++++++++++++++++++++++++++++++++++++++++++++++++ 1 file changed, 54 insertions(+) diff --git a/RELEASE.md b/RELEASE.md index f93626cc876..68d9399676a 100644 --- a/RELEASE.md +++ b/RELEASE.md @@ -1,3 +1,57 @@ +# Release 2.4.0 + + + +## Breaking Changes + +* +* + +## Known Caveats + +* + +## Major Features and Improvements + +* +* + +## Bug Fixes and Other Changes + +* +* +* +* TF Core: + * +* `tf.data`: + * +* `tf.distribute`: + * +* `tf.keras`: + * +* `tf.function`/AutoGraph: + * +* `tf.lite`: + * +* `tf.random`: + * +* Math and Linear Algebra: + * +* TPU Enhancements: + * +* XLA Support: + * +* Tracing and Debugging: + * +* Other: + * + +## Thanks to our Contributors + +This release contains contributions from many people at Google, as well as: + +, , , , , + # Release 2.3.0 ## Breaking Changes From 821d1b087be221aa3c58be8e077d4553d6851ef2 Mon Sep 17 00:00:00 2001 From: Jiho Choi Date: Mon, 22 Jun 2020 19:30:34 -0700 Subject: [PATCH 520/867] Add element tracing for parallel_map to track exactly which upstream event produces an element for each parallel_map's GetNext call. PiperOrigin-RevId: 317784362 Change-Id: I981692f8d152ab8b6c6c343d01563800a8f94600 --- tensorflow/core/kernels/data/BUILD | 2 ++ .../kernels/data/parallel_map_dataset_op.cc | 19 ++++++++++++++++++- 2 files changed, 20 insertions(+), 1 deletion(-) diff --git a/tensorflow/core/kernels/data/BUILD b/tensorflow/core/kernels/data/BUILD index 0972dc83ccf..74283b63b67 100644 --- a/tensorflow/core/kernels/data/BUILD +++ b/tensorflow/core/kernels/data/BUILD @@ -470,6 +470,8 @@ tf_kernel_library( "//tensorflow/core:lib", "//tensorflow/core:lib_internal", "//tensorflow/core:protos_all_cc", + "//tensorflow/core/profiler/lib:traceme", + "//tensorflow/core/profiler/lib:traceme_encode", ], ) diff --git a/tensorflow/core/kernels/data/parallel_map_dataset_op.cc b/tensorflow/core/kernels/data/parallel_map_dataset_op.cc index bae90549841..1dc27ce6635 100644 --- a/tensorflow/core/kernels/data/parallel_map_dataset_op.cc +++ b/tensorflow/core/kernels/data/parallel_map_dataset_op.cc @@ -29,6 +29,8 @@ limitations under the License. #include "tensorflow/core/lib/core/errors.h" #include "tensorflow/core/lib/random/random.h" #include "tensorflow/core/platform/stringprintf.h" +#include "tensorflow/core/profiler/lib/traceme.h" +#include "tensorflow/core/profiler/lib/traceme_encode.h" #include "tensorflow/core/protobuf/error_codes.pb.h" namespace tensorflow { @@ -241,6 +243,10 @@ class ParallelMapDatasetOp::Dataset : public DatasetBase { RecordStop(ctx); result->notification.WaitForNotification(); RecordStart(ctx); + profiler::TraceMe traceme([&] { + return profiler::TraceMeEncode("ParallelMapConsume", + {{"element_id", result->id}}); + }); return ProcessResult(ctx, result, out_tensors, end_of_sequence); } @@ -358,10 +364,14 @@ class ParallelMapDatasetOp::Dataset : public DatasetBase { private: struct InvocationResult { + InvocationResult() {} + explicit InvocationResult(int64 id) : id(id) {} + Notification notification; Status status; std::vector return_values; bool end_of_input; + int64 id; }; void CancelThreads(bool wait) TF_LOCKS_EXCLUDED(mu_) { @@ -402,6 +412,10 @@ class ParallelMapDatasetOp::Dataset : public DatasetBase { void CallFunction(const std::shared_ptr& ctx, const std::shared_ptr& result) TF_LOCKS_EXCLUDED(*mu_) { + profiler::TraceMe traceme([&] { + return profiler::TraceMeEncode("ParallelMapProduce", + {{"element_id", result->id}}); + }); // Get the next input element. std::vector input_element; result->status = input_impl_->GetNext(ctx.get(), &input_element, @@ -490,6 +504,8 @@ class ParallelMapDatasetOp::Dataset : public DatasetBase { return num_calls_ >= num_parallel_calls || invocation_results_.size() >= num_parallel_calls; }; + // Counts the total number of calls to use as an id of InvocationResult. + int64 num_total_calls = 0; while (true) { { mutex_lock l(*mu_); @@ -502,7 +518,8 @@ class ParallelMapDatasetOp::Dataset : public DatasetBase { return; } while (!busy()) { - invocation_results_.push_back(std::make_shared()); + invocation_results_.push_back( + std::make_shared(num_total_calls++)); new_calls.push_back(invocation_results_.back()); num_calls_++; } From b16994e2d646df982c2c8614a513742ac06c3cf0 Mon Sep 17 00:00:00 2001 From: "A. Unique TensorFlower" Date: Mon, 22 Jun 2020 20:27:46 -0700 Subject: [PATCH 521/867] Add element tracing for parallel_map to track exactly which upstream event produces an element for each parallel_map's GetNext call. PiperOrigin-RevId: 317789815 Change-Id: Ic8a50a3a311051069ed25973b33cd2b0b9bbece5 --- tensorflow/core/kernels/data/BUILD | 2 -- .../kernels/data/parallel_map_dataset_op.cc | 19 +------------------ 2 files changed, 1 insertion(+), 20 deletions(-) diff --git a/tensorflow/core/kernels/data/BUILD b/tensorflow/core/kernels/data/BUILD index 74283b63b67..0972dc83ccf 100644 --- a/tensorflow/core/kernels/data/BUILD +++ b/tensorflow/core/kernels/data/BUILD @@ -470,8 +470,6 @@ tf_kernel_library( "//tensorflow/core:lib", "//tensorflow/core:lib_internal", "//tensorflow/core:protos_all_cc", - "//tensorflow/core/profiler/lib:traceme", - "//tensorflow/core/profiler/lib:traceme_encode", ], ) diff --git a/tensorflow/core/kernels/data/parallel_map_dataset_op.cc b/tensorflow/core/kernels/data/parallel_map_dataset_op.cc index 1dc27ce6635..bae90549841 100644 --- a/tensorflow/core/kernels/data/parallel_map_dataset_op.cc +++ b/tensorflow/core/kernels/data/parallel_map_dataset_op.cc @@ -29,8 +29,6 @@ limitations under the License. #include "tensorflow/core/lib/core/errors.h" #include "tensorflow/core/lib/random/random.h" #include "tensorflow/core/platform/stringprintf.h" -#include "tensorflow/core/profiler/lib/traceme.h" -#include "tensorflow/core/profiler/lib/traceme_encode.h" #include "tensorflow/core/protobuf/error_codes.pb.h" namespace tensorflow { @@ -243,10 +241,6 @@ class ParallelMapDatasetOp::Dataset : public DatasetBase { RecordStop(ctx); result->notification.WaitForNotification(); RecordStart(ctx); - profiler::TraceMe traceme([&] { - return profiler::TraceMeEncode("ParallelMapConsume", - {{"element_id", result->id}}); - }); return ProcessResult(ctx, result, out_tensors, end_of_sequence); } @@ -364,14 +358,10 @@ class ParallelMapDatasetOp::Dataset : public DatasetBase { private: struct InvocationResult { - InvocationResult() {} - explicit InvocationResult(int64 id) : id(id) {} - Notification notification; Status status; std::vector return_values; bool end_of_input; - int64 id; }; void CancelThreads(bool wait) TF_LOCKS_EXCLUDED(mu_) { @@ -412,10 +402,6 @@ class ParallelMapDatasetOp::Dataset : public DatasetBase { void CallFunction(const std::shared_ptr& ctx, const std::shared_ptr& result) TF_LOCKS_EXCLUDED(*mu_) { - profiler::TraceMe traceme([&] { - return profiler::TraceMeEncode("ParallelMapProduce", - {{"element_id", result->id}}); - }); // Get the next input element. std::vector input_element; result->status = input_impl_->GetNext(ctx.get(), &input_element, @@ -504,8 +490,6 @@ class ParallelMapDatasetOp::Dataset : public DatasetBase { return num_calls_ >= num_parallel_calls || invocation_results_.size() >= num_parallel_calls; }; - // Counts the total number of calls to use as an id of InvocationResult. - int64 num_total_calls = 0; while (true) { { mutex_lock l(*mu_); @@ -518,8 +502,7 @@ class ParallelMapDatasetOp::Dataset : public DatasetBase { return; } while (!busy()) { - invocation_results_.push_back( - std::make_shared(num_total_calls++)); + invocation_results_.push_back(std::make_shared()); new_calls.push_back(invocation_results_.back()); num_calls_++; } From 3ed1e3029e68ca8cb6306c8f31182306741dcf0c Mon Sep 17 00:00:00 2001 From: Nick Kreeger Date: Mon, 22 Jun 2020 20:37:38 -0700 Subject: [PATCH 522/867] Remove static_assert for type checking in FlatBufferVectorToTfLiteTypeArray. It turns out that std::is_same() has dropped the non-string argument in c++17. This breaks internal users that are building against qualcomm. PiperOrigin-RevId: 317790812 Change-Id: If56a61d20426670251b55f370a6b5fa886a49e21 --- tensorflow/lite/micro/micro_allocator.cc | 9 +++------ 1 file changed, 3 insertions(+), 6 deletions(-) diff --git a/tensorflow/lite/micro/micro_allocator.cc b/tensorflow/lite/micro/micro_allocator.cc index bf9e38d1050..239a23335a6 100644 --- a/tensorflow/lite/micro/micro_allocator.cc +++ b/tensorflow/lite/micro/micro_allocator.cc @@ -401,12 +401,9 @@ TfLiteStatus FlatBufferVectorToTfLiteTypeArray( kTfLiteArrayType** result) { TFLITE_DCHECK(error_reporter != nullptr); TFLITE_DCHECK(flatbuffer_array != nullptr); - // Only two conversions are supported - float and int32 - ensure that these - // match at compile time instead of duplicating functions here: - static_assert((std::is_same() && - std::is_same()) || - (std::is_same() && - std::is_same())); + // TODO(b/159668691): Consider adding type assertion or breaking this function + // into multiple functions for each type. std::is_same is c++11 and has a + // special updated constructor in c++17 that requires a string argument. if (FLATBUFFERS_LITTLEENDIAN) { // On little-endian machines, TfLite*Array happens to have the same memory // layout as flatbuffers:Vector, so we can From 0453c02f43b64e54caf1a3f001f0c01def38ba37 Mon Sep 17 00:00:00 2001 From: Andy Ly Date: Mon, 22 Jun 2020 21:36:57 -0700 Subject: [PATCH 523/867] Update tf.DeviceIndex description in TensorFlow MLIR ODS to match description in TensorFlow op registry (NFC). PiperOrigin-RevId: 317796799 Change-Id: Ic4d6ca197cce7c56e491f59814cf913dcd64713c --- tensorflow/compiler/mlir/tensorflow/ir/tf_generated_ops.td | 7 +++++++ 1 file changed, 7 insertions(+) diff --git a/tensorflow/compiler/mlir/tensorflow/ir/tf_generated_ops.td b/tensorflow/compiler/mlir/tensorflow/ir/tf_generated_ops.td index 7f250392cb2..65ca3ea4dbd 100644 --- a/tensorflow/compiler/mlir/tensorflow/ir/tf_generated_ops.td +++ b/tensorflow/compiler/mlir/tensorflow/ir/tf_generated_ops.td @@ -2378,6 +2378,13 @@ horizontal and vertices strides, `strides = [1, stride, stride, 1]`. def TF_DeviceIndexOp : TF_Op<"DeviceIndex", [NoSideEffect]> { let summary = "Return the index of device the op runs."; + let description = [{ +Given a list of device names, this operation returns the index of the device +this op runs. The length of the list is returned in two cases: +(1) Device does not exist in the given device list. +(2) It is in XLA compilation. + }]; + let arguments = (ins StrArrayAttr:$device_names ); From 79fc3fd8fcc4b2936ca630cec6898a83ee92f59b Mon Sep 17 00:00:00 2001 From: Rick Chao Date: Mon, 22 Jun 2020 21:41:24 -0700 Subject: [PATCH 524/867] Skip data loading error in multi_worker_tutorial_test (the test does not aim to cover this). PiperOrigin-RevId: 317797271 Change-Id: I8336d7ffeda0836beef0a2d04e633614a44e7fa4 --- .../python/keras/distribute/multi_worker_tutorial_test.py | 3 +++ 1 file changed, 3 insertions(+) diff --git a/tensorflow/python/keras/distribute/multi_worker_tutorial_test.py b/tensorflow/python/keras/distribute/multi_worker_tutorial_test.py index 3f9ab18f89c..aafa00bcc9f 100644 --- a/tensorflow/python/keras/distribute/multi_worker_tutorial_test.py +++ b/tensorflow/python/keras/distribute/multi_worker_tutorial_test.py @@ -19,6 +19,7 @@ from __future__ import print_function import contextlib import os import re +import zipfile from absl.testing import parameterized import numpy as np from tensorflow.python import keras @@ -43,6 +44,8 @@ class MultiWorkerTutorialTest(parameterized.TestCase, test.TestCase): def skip_fetch_failure_exception(self): try: yield + except zipfile.BadZipfile as e: + self.skipTest('Data loading error: Bad magic number for file header.') except Exception as e: # pylint: disable=broad-except if 'URL fetch failure' in str(e): self.skipTest('URL fetch error not considered failure of the test.') From e3904ac3aef005326e126e8240eaf5bf23cb3b04 Mon Sep 17 00:00:00 2001 From: Henry Tan Date: Mon, 22 Jun 2020 22:01:28 -0700 Subject: [PATCH 525/867] Internal TPU library change. PiperOrigin-RevId: 317799072 Change-Id: Ia9b4fba3aa8292e5b089d5741f359eef9763df09 --- .../core/tpu/kernels/tpu_compile_c_api.h | 2 +- tensorflow/core/tpu/kernels/tpu_op_util.cc | 37 ++++++++++++++----- tensorflow/core/tpu/kernels/tpu_op_util.h | 7 ++++ 3 files changed, 36 insertions(+), 10 deletions(-) diff --git a/tensorflow/core/tpu/kernels/tpu_compile_c_api.h b/tensorflow/core/tpu/kernels/tpu_compile_c_api.h index c101e489d56..eab53fe9da4 100644 --- a/tensorflow/core/tpu/kernels/tpu_compile_c_api.h +++ b/tensorflow/core/tpu/kernels/tpu_compile_c_api.h @@ -35,7 +35,7 @@ struct CompilationCacheKeyProperty { const char* mlir_module; const int32_t* device_ids; size_t device_ids_size; - int32_t guaranteed_constants_size; + size_t guaranteed_constants_size; uint64_t function_library_fingerprint; int32_t num_cores_per_replica; int32_t num_replicas; diff --git a/tensorflow/core/tpu/kernels/tpu_op_util.cc b/tensorflow/core/tpu/kernels/tpu_op_util.cc index e2f717fea8b..31b0cc6c72d 100644 --- a/tensorflow/core/tpu/kernels/tpu_op_util.cc +++ b/tensorflow/core/tpu/kernels/tpu_op_util.cc @@ -24,12 +24,13 @@ namespace tpu { namespace { // Return fingerprint_in_metadata if it's not empty; otherwise read input tensor // data to compute the fingerprint. -std::string GuaranteedConstFingerprint( - const string& fingerprint_in_metadata, - const OpInputList& guaranteed_constants) { +std::string GuaranteedConstFingerprint(const string& fingerprint_in_metadata, + const Tensor* guaranteed_constants, + size_t guaranteed_constants_size) { if (fingerprint_in_metadata.empty()) { uint64_t fingerprint = 0; - for (const auto& constant : guaranteed_constants) { + for (size_t i = 0; i < guaranteed_constants_size; ++i) { + const Tensor& constant = guaranteed_constants[i]; fingerprint = TpuCompile_CreateGuaranteedConstFingerprint( fingerprint, constant.tensor_data().data(), constant.tensor_data().size()); @@ -86,9 +87,12 @@ std::string CreateConfigPrefix(const TPUCompileMetadataProto& metadata) { } } // namespace +// The `guaranteed_constants` must be passed as reference due to the lazy +// evaluation of `guaranteed_const_fingerprint()` callback. TpuCompilationCacheKey CreateCompilationCacheKey( absl::string_view function_name, uint64 function_library_fingerprint, - absl::string_view mlir_module, const OpInputList& guaranteed_constants, + absl::string_view mlir_module, const Tensor* guaranteed_constants, + size_t guaranteed_constants_size, const std::vector& dynamic_shapes, const TPUCompileMetadataProto& metadata, const TpuMeshStateInterface& mesh_state) { @@ -114,7 +118,7 @@ TpuCompilationCacheKey CreateCompilationCacheKey( mlir_module.data(), flattened_device_ids.data(), flattened_device_ids.size(), - guaranteed_constants.size(), + guaranteed_constants_size, function_library_fingerprint, metadata.num_cores_per_replica(), metadata.num_replicas(), @@ -128,7 +132,7 @@ TpuCompilationCacheKey CreateCompilationCacheKey( // Guaranteed constants can be different across sessions. Use session_handle // and guaranteed_const fingerprint to guarantee no collision. - if (guaranteed_constants.size() > 0) { + if (guaranteed_constants != nullptr && guaranteed_constants_size > 0) { key.has_guaranteed_const = true; key.session_handle = metadata.session_handle(); // Both `metadata` and `guaranteed_constants` lifetime are captured by @@ -136,16 +140,31 @@ TpuCompilationCacheKey CreateCompilationCacheKey( // managed through the `TPUCompileOpKernelImpl` that outlives the // lifetime of the compilation cache lookups. string fingerprint; - key.guaranteed_const_fingerprint = [&metadata, &guaranteed_constants, + key.guaranteed_const_fingerprint = [&metadata, guaranteed_constants, + guaranteed_constants_size, fingerprint]() mutable { if (fingerprint.empty()) { fingerprint = GuaranteedConstFingerprint( - metadata.guaranteed_const_fingerprint(), guaranteed_constants); + metadata.guaranteed_const_fingerprint(), guaranteed_constants, + guaranteed_constants_size); } return fingerprint; }; } return key; } + +TpuCompilationCacheKey CreateCompilationCacheKey( + absl::string_view function_name, uint64 function_library_fingerprint, + absl::string_view mlir_module, const OpInputList& guaranteed_constants, + size_t guaranteed_constants_size, + const std::vector& dynamic_shapes, + const TPUCompileMetadataProto& metadata, + const TpuMeshStateInterface& mesh_state) { + return CreateCompilationCacheKey( + function_name, function_library_fingerprint, mlir_module, + (guaranteed_constants.size() > 0 ? &guaranteed_constants[0] : nullptr), + guaranteed_constants.size(), dynamic_shapes, metadata, mesh_state); +} } // namespace tpu } // namespace tensorflow diff --git a/tensorflow/core/tpu/kernels/tpu_op_util.h b/tensorflow/core/tpu/kernels/tpu_op_util.h index 0a9657ca05e..bbaa05682e6 100644 --- a/tensorflow/core/tpu/kernels/tpu_op_util.h +++ b/tensorflow/core/tpu/kernels/tpu_op_util.h @@ -34,6 +34,13 @@ TpuCompilationCacheKey CreateCompilationCacheKey( const std::vector& dynamic_shapes, const TPUCompileMetadataProto& metadata, const TpuMeshStateInterface& mesh_state); +TpuCompilationCacheKey CreateCompilationCacheKey( + absl::string_view function_name, uint64 function_library_fingerprint, + absl::string_view mlir_module, const Tensor* guaranteed_constants, + size_t guaranteed_constants_size, + const std::vector& dynamic_shapes, + const TPUCompileMetadataProto& metadata, + const TpuMeshStateInterface& mesh_state); } // namespace tpu } // namespace tensorflow From ff4490f636c87bf9efacfca06bd1f70a61bcfdc9 Mon Sep 17 00:00:00 2001 From: Taehee Jeong Date: Mon, 22 Jun 2020 22:05:00 -0700 Subject: [PATCH 526/867] Config Core ML delegate provider correctly PiperOrigin-RevId: 317799599 Change-Id: I6eef7b9de910b2401cf24909d1e245dc972103f9 --- tensorflow/lite/tools/delegates/coreml_delegate_provider.cc | 1 + 1 file changed, 1 insertion(+) diff --git a/tensorflow/lite/tools/delegates/coreml_delegate_provider.cc b/tensorflow/lite/tools/delegates/coreml_delegate_provider.cc index c6509618aee..7d88f04c8b4 100644 --- a/tensorflow/lite/tools/delegates/coreml_delegate_provider.cc +++ b/tensorflow/lite/tools/delegates/coreml_delegate_provider.cc @@ -17,6 +17,7 @@ limitations under the License. #include "tensorflow/lite/tools/delegates/delegate_provider.h" #include "tensorflow/lite/tools/evaluation/utils.h" #if defined(__APPLE__) +#include "TargetConditionals.h" #if TARGET_OS_IPHONE && !TARGET_IPHONE_SIMULATOR // Only enable metal delegate when using a real iPhone device. #define REAL_IPHONE_DEVICE From 40e860b7a46cf3768bd177b25b1e35750860f73f Mon Sep 17 00:00:00 2001 From: Mehdi Amini Date: Mon, 22 Jun 2020 22:21:55 -0700 Subject: [PATCH 527/867] Add a doc on the current short-term plan for XLA GPU CodeGen with MLIR PiperOrigin-RevId: 317801702 Change-Id: Idabeb7717b357622e9cbcc2d4352df6b4ce25373 --- .../compiler/mlir/g3doc/xla_gpu_codegen.md | 265 ++++++++++++++++++ 1 file changed, 265 insertions(+) create mode 100644 tensorflow/compiler/mlir/g3doc/xla_gpu_codegen.md diff --git a/tensorflow/compiler/mlir/g3doc/xla_gpu_codegen.md b/tensorflow/compiler/mlir/g3doc/xla_gpu_codegen.md new file mode 100644 index 00000000000..06c55abf1fa --- /dev/null +++ b/tensorflow/compiler/mlir/g3doc/xla_gpu_codegen.md @@ -0,0 +1,265 @@ +# MLIR CodeGen for XLA + + + +XLA operates on `HloInstruction` and performs many optimizations on this +representation, sharing a lot of these between targeted devices. As some point a +linear schedule is computed and the memory buffer is assigned to each value +statically. The device specific codegen operates by traversing this sequence and +calling "emitters" to generate a representation suitable for the device (for +example a single LLVM function per XLA computation on CPU, or a sequence of +"thunks" encapsulating GPU operations and possibly generated PTX when targeting +GPU). + +As a staging step, we're currently in the process of intercepting the process +right after XLA completes the buffer-assignment phase and emit instead an MLIR +module in the `lhlo` dialect. From there we perform the codegen using MLIR +components (Linalg, affine, and GPU dialect mainly) depending on the device. + +Below is the plan of record to incrementally migrate XLA/GPU by using `lhlo` as +the codegen input. + +## Tasks + + | Host | Device +------------- | ------------------------ | ------------------------ +Input format | HloInstruction* (Task 1) | HloInstruction* (Task 1) +Output format | xla::Thunk (Task 2) | LLVM IR (Task 3) + +* **Task 1** changes both host and device input format from HloInstruction* to + LHLO. +* **Task 2** changes output format of host from thunks to "some landing pad + for host" (see below). +* **Task 3** migrates device output from LLVM IR to some form of MLIR. It's + optional to this project, and see the section "Migrating Device LLVM IR" for + details. + +This project prioritizes having end-to-end runnable models with LHLO-emitters +enabled as much as possible. This implies that the following order list of +objectives by priority: + +* Make XLA/GPU runnable with LHLO emitters, with existing Thunks and emitters + unmodified. +* Eliminate the references to HloInstruction\* in LHLO, case by case: + * Switch a legacy emitter to an MLIR-based emitter (e.g. Linalg), or + * Mechanically translate the existing emitter to take MLIR representation + (migrate to Standard with GPU Dialect). + +## Migrating Thunks (Task 2) + +xla::gpu::Thunk is a data structure that: + +* Can be called into from the host (xla::gpu::Thunk::ExecuteOnStream()). +* Carries various data in its subclasses. +* Interacts with BufferAllocation::Slice and StreamExecutor. +* Launches kernels +* Calls into all runtime libraries. + +The cost of that includes: + +* Representing op-specific configuration data (e.g. convolution configs). +* Migrating op shape and operand shapes. +* Representing a tree of thunks (while, condition, etc). + +The migration work is independent from LHLO / emitter migration. Under limited +resources, it's prioritized behind LHLO / emitter migration. + +We have several choices on how to lower the host-side part from LHLO: + +* TFRT + * (Pro) great CUDA and HIP wrappers for use. + * (Pro) easy to implement library calls (cuDNN, cuBLAS, cuFFT, etc), as + TFRT ops are interpreted by C++ code. + * (Con) host side is under development and not tested. + * (Con) the JAX integration isn’t clear from a runtime point of view +* Jitted CPU code + * (Pro) great lower-ability. Create a few loops and conditions and it's + done. + * (Con) GPUDialect doesn't yet model chains/streams/asynchronicity/device + allocation. + * (Con) CUDA / HIP runtime support is minimal (toolkit path, version, + dynamic loading, etc). +* Existing (interpreting) XLA runtime + +Tentative conclusion: Use jitted CPU code during the transition, and optionally +adopt TFRT in the end. + +## Migrating Device LLVM IR (Task 3) + +An elemental emitter generates target op by filling it element by element. Each +output element depends on a set of elements from the operands. All elements are +described by combining the buffer with dynamic indices. It's sufficient to +describe almost all "math" ops, but for performance reasons only a large subset +of "math" ops are implemented directly in (Cpu|Gpu)ElementalIrEmitter. + +ElementalIrEmitter is unique in that: + +* A large portion of the code is shared between XLA/GPU and CPU. +* It represents a large portion of ops seen in models, including all + element-wise ops. +* Most fusions solely depend on ElementalIrEmitter. +* It's structurally simple, as it describes a data dependency DAG between op + elements and operand elements. +* It's mostly portable and high-level (e.g. unlike GPU kReduce and GPU kCopy). +* Dynamic shape support is easy for at least element-wise ops. + +Now, for all ops, elementally-emitted or not, there are several flavors of the +end state of each XLA op: + +1. Device code stays as LLVM IR. +1. Refactor the old emitter to be like LHLO -> MLIR LLVM Dialect: + * (Cost) Will be throw-away work if we want to ultimately migrate to + Standard. + * (Benefit) It is easy and mechanical. Can be done in a short period. + * (Benefit) It doesn't benefit more compared to a). +1. Refactor old emitters to be like LHLO -> MLIR GPU + Standard + Loops: + * (Cost) Lifting existing emitters to Standard introduces some challenges. + Pointers and GEPs need to be converted to MemRefs and SubViews. Ensuring + amdgpu completeness is another one. + * (Cost) XLA/GPU heavily relies on LLVM metadata: + * `range` for block/thread indices. + * `align`, `dereferenceable`, `invariant.load`, `alias.scope`, + `noalias` for load/stores. + * `llvm.loop.unroll.disable`, `llvm.loop.unroll.full`, + `llvm.loop.vectorize.enable` for sequential loops. + * (Benefit) Can be long-term. More portable. +1. Refactor old emitters to be LHLO -> Linalg, and write new Linalg emitters + * (Cost) This is case by case. Compared to previous options, a new + implementation that matches XLA's performance needs to go through the + benchmark <-> optimize workflow, which can be a significant cost for + some ops. + * (Benefit) unified stack; community support; portability; more + optimization potentials. + +## Prioritization + +While all three tasks mentioned above are parallelizable, under limited +resources they have to be serialized. The prioritization focuses on visible +results for completion of each task. + +The prioritization is: Task1 (LHLO for legacy emitters) > Task 2 (Thunks) > Task +3 (MLIR emitters). + +By the end of Task 1, users of XLA can generate an LHLO (e.g. kernel generator) +and execute them. The compilation format will not be serializable MLIR. + +By the end of Task 2, LHLO lowers to proper, serializable MLIR. This enables +offline compilation. + +By the end of Task 3, all XLA emitters are MLIR-based in its implementation. + +## Detailed Design + +### Step 1: (Task 1) Complete LHLO and Make Legacy Emitters Take LHLO + +This step makes all existing XLA/GPU emitters interact with MLIR ops. This step +is pure refactoring and NFC. + +This step is mostly mechanical, but it's worth noticing the following +discrepancies between an unnested HloComputation and LHLO: + +* Each HloInstruction has direct access to its operands (a data-flow DAG). On + contrary, each LHLO op only has access to its operand buffers (a bipartite + between ops and buffers). LHLO ops have to go through use-def chains to + access their operand ops. +* Unnested legacy emitters empirically almost never access their operands. The + only exception is kReduce. +* Unnested legacy emitters access BufferAssignment only for getting slices, + not for accessing aux data structures like dataflow\_analysis() or + alias\_analysis(). llvm\_ir builds its own alias\_analysis() based on slice + information. + +The conclusion is that LHLO should fit right-in without major hassle. + +### Step 2: (Optional) Profiling Support + +**This step is only needed if we start to discard some of the XLA Thunk logic +(see the next step).** + +Before actually turning on any MLIR-based emitters, we need profiling for +MLIR-based emitters. + +Currently XLA performs its own profiling by calling into StreamExecutor's timer. +The timer under the hood inserts two events before and after a kernel launch, +and measures the sync time between these two events. + +There are roughly three approaches to support profiling in MLIR: + +* Run a profiler end-to-end +* Add a profile op for each op in LHLO, using an injected profiler. + +The "end-to-end" approach is transparent to MLIR, but suffers the same problem +that makes XLA not use it in the first place: library calls collected by a +profiler (nvprof/...) can't easily relate to HLO ops. For example, cuDNN +launches multiple kernels for each HLO, and it's hard to tell which kernels +correspond to which HLO. + +The "injected profiler" approach requires: + +* LHLO to take a profiler as a parameter. +* inserting profile.start / profile.end before and after each op. +* a pass from that lowers profile.{start,end} to a C++ implementation. + +The exact profiling can't be easily done for MLIR-generated ops, since: + +* MLIR doesn't have a timer, nor it depends on TFRT / StreamExecutor. +* MLIR doesn't easily call into C functions with complicated parameters. + +### Step 3: (Task 2) Migrating Thunks + +This step migrates all host ops and library calls. This step will eliminate most +of the thunks and produce serializable MLIR instead. + +There are roughly three kinds of thunks: + +* KernelThunk, which launches a kernel. +* Control flow thunks, which has host control flow logic (conditional, while, + for, sequence) and launch body kernels. +* Library thunks: cuDNN, cuBLAS, cuFFT, NCCL, etc. + +The **bottom line** is to: + +* Create a Thunk dialect that provides (de)serialize logic for all existing + C++-based Thunks. +* Change emitters to emit a graph of Thunk dialect. + +**Optionally**, we can relieve some thunks from C++ implementation. KernelThunk +can lower to the GPU LaunchKernelOp. Control flow thunks can leverage the CFG +Dialect for loops and conditions, combined with LaunchKernelOp. This optional +step requires profiling and stream support. + +### Step 4: (Task 3) Migrated ElementalIrEmitter + +Once profiling is ready, we can complete and tune all ElementalIrEmitter-based +emitters in MLIR. Then we turn them on by default, assuming that all of these +MLIR-based emitters use a single stream. + +Notice that it's beneficial to migrate XLA/CPU's ElementalIrEmitter as well, +since they share a large portion of the code. + +With all benchmarking and performance hunting done (TODO: define performance +parity), we turn on the new MLIR-based elemental emitter, and delete the legacy +ElementalIrEmitter. + +This step also provides easy fusion transitions (nested ops) for the later +migration. + +### Step 5: Multi-Stream Support or Drop + +We can't delete +[some of the emitters](https://github.com/tensorflow/tensorflow/blob/master/tensorflow/compiler/xla/service/gpu/stream_assignment.cc#L140) +until we support it in MLIR, or we drop the feature. It's a relatively large +amount of work in MLIR and a small amount of gain for XLA. We should investigate +current users of multi-stream XLA/GPU users, and try to delete this feature if +reasonable. + +### Step 6: (Task 3) Migrated Device Ops + +This step migrates all unnested ops, then we can delete all unnested emitters. + +This calls on a rewrite/refactor for kCopy and kReduce. kReduce is already +worked on for plenty, so the actual amount of work that needs to be done remains +to be seen. From 934df8dcea0c176314a52d0062dcca08638bb52d Mon Sep 17 00:00:00 2001 From: Advait Jain Date: Mon, 22 Jun 2020 22:29:15 -0700 Subject: [PATCH 528/867] Remove portable_optimized depthwise_conv implementation. PR #34999 surfaced inconsistencies between the reference and portable_optimized implementations and we are also slowly transitioning to CMSIS-NN for the optimized implementations for ARM Cortex-M targets. There may still be an issue with the reference depthwise conv + dilation > 1 but that would be a separate bug. PiperOrigin-RevId: 317802639 Change-Id: Ic8c9acffb060bb3ec5f802a2045ac6ac8b9f2233 --- tensorflow/lite/micro/BUILD | 24 - tensorflow/lite/micro/kernels/BUILD | 81 --- .../portable_optimized/depthwise_conv.cc | 515 ------------------ .../lite/micro/tools/ci_build/test_mbed.sh | 2 +- .../make/targets/apollo3evb_makefile.inc | 3 - 5 files changed, 1 insertion(+), 624 deletions(-) delete mode 100644 tensorflow/lite/micro/kernels/portable_optimized/depthwise_conv.cc diff --git a/tensorflow/lite/micro/BUILD b/tensorflow/lite/micro/BUILD index f63d9778634..bdfa0c909db 100644 --- a/tensorflow/lite/micro/BUILD +++ b/tensorflow/lite/micro/BUILD @@ -102,30 +102,6 @@ cc_library( ], ) -# TODO(b/144176795): This target should really be handled differently so that we -# do not have a fork in the build graph. The bug has some initial ideas. -cc_library( - name = "portable_optimized_op_resolver", - srcs = [ - "all_ops_resolver.cc", - "micro_mutable_op_resolver.h", - "micro_op_resolver.h", - ], - hdrs = [ - "all_ops_resolver.h", - ], - copts = micro_copts(), - deps = [ - ":micro_compatibility", - "//tensorflow/lite/c:common", - "//tensorflow/lite/core/api", - "//tensorflow/lite/kernels:op_macros", - "//tensorflow/lite/kernels/internal:compatibility", - "//tensorflow/lite/micro/kernels:portable_optimized_micro_ops", - "//tensorflow/lite/schema:schema_fbs", - ], -) - cc_library( name = "debug_log", srcs = [ diff --git a/tensorflow/lite/micro/kernels/BUILD b/tensorflow/lite/micro/kernels/BUILD index c7fa19b8cea..0fd0be4e3a4 100644 --- a/tensorflow/lite/micro/kernels/BUILD +++ b/tensorflow/lite/micro/kernels/BUILD @@ -20,7 +20,6 @@ package_group( packages = ["//tensorflow/lite/micro"], ) -# LINT.IfChange(micro_ops) cc_library( name = "micro_ops", srcs = [ @@ -106,73 +105,6 @@ cc_library( ], }), ) -# LINT.ThenChange(//tensorflow/lite/micro/kernels/BUILD:portable_optimized_micro_ops) - -# LINT.IfChange(portable_optimized_micro_ops) -cc_library( - name = "portable_optimized_micro_ops", - srcs = [ - "activations.cc", - "add.cc", - "arg_min_max.cc", - "ceil.cc", - "circular_buffer.cc", - "comparisons.cc", - "concatenation.cc", - "conv.cc", - "dequantize.cc", - "elementwise.cc", - "ethosu.cc", - "floor.cc", - "fully_connected.cc", - "l2norm.cc", - "logical.cc", - "logistic.cc", - "maximum_minimum.cc", - "mul.cc", - "neg.cc", - "pack.cc", - "pad.cc", - "pooling.cc", - "portable_optimized/depthwise_conv.cc", - "prelu.cc", - "quantize.cc", - "reduce.cc", - "reshape.cc", - "resize_nearest_neighbor.cc", - "round.cc", - "softmax.cc", - "split.cc", - "strided_slice.cc", - "sub.cc", - "svdf.cc", - "tanh.cc", - "unpack.cc", - ], - hdrs = ["micro_ops.h"], - copts = micro_copts(), - visibility = [ - # Needed for micro:portable_optimized_ops_resolver but visibility can not be - # finer-grained than a package. - ":micro_top_level", - ], - deps = [ - ":activation_utils", - ":micro_utils", - "//tensorflow/lite/c:common", - "//tensorflow/lite/kernels:kernel_util", - "//tensorflow/lite/kernels:op_macros", - "//tensorflow/lite/kernels:padding", - "//tensorflow/lite/kernels/internal:common", - "//tensorflow/lite/kernels/internal:compatibility", - "//tensorflow/lite/kernels/internal:quantization_util", - "//tensorflow/lite/kernels/internal:reference_base", - "//tensorflow/lite/kernels/internal:tensor", - "//tensorflow/lite/kernels/internal:types", - "//tensorflow/lite/micro:micro_utils", - ], -) -# LINT.ThenChange(//tensorflow/lite/micro/kernels/BUILD:micro_ops) test_suite( name = "all_tests", @@ -214,19 +146,6 @@ tflite_micro_cc_test( ], ) -tflite_micro_cc_test( - name = "portable_optimized_depthwise_conv_test", - srcs = [ - "depthwise_conv_test.cc", - ], - deps = [ - "//tensorflow/lite/c:common", - "//tensorflow/lite/kernels/internal:tensor", - "//tensorflow/lite/micro:portable_optimized_op_resolver", - "//tensorflow/lite/micro/testing:micro_test", - ], -) - tflite_micro_cc_test( name = "fully_connected_test", srcs = [ diff --git a/tensorflow/lite/micro/kernels/portable_optimized/depthwise_conv.cc b/tensorflow/lite/micro/kernels/portable_optimized/depthwise_conv.cc deleted file mode 100644 index 9fb8f2e32cc..00000000000 --- a/tensorflow/lite/micro/kernels/portable_optimized/depthwise_conv.cc +++ /dev/null @@ -1,515 +0,0 @@ -/* Copyright 2017 The TensorFlow Authors. All Rights Reserved. - -Licensed under the Apache License, Version 2.0 (the "License"); -you may not use this file except in compliance with the License. -You may obtain a copy of the License at - - http://www.apache.org/licenses/LICENSE-2.0 - -Unless required by applicable law or agreed to in writing, software -distributed under the License is distributed on an "AS IS" BASIS, -WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -See the License for the specific language governing permissions and -limitations under the License. -==============================================================================*/ - -#include "tensorflow/lite/kernels/internal/reference/integer_ops/depthwise_conv.h" - -#include "tensorflow/lite/c/builtin_op_data.h" -#include "tensorflow/lite/c/common.h" -#include "tensorflow/lite/kernels/internal/common.h" -#include "tensorflow/lite/kernels/internal/quantization_util.h" -#include "tensorflow/lite/kernels/internal/reference/depthwiseconv_float.h" -#include "tensorflow/lite/kernels/internal/reference/depthwiseconv_uint8.h" -#include "tensorflow/lite/kernels/internal/tensor_ctypes.h" -#include "tensorflow/lite/kernels/kernel_util.h" -#include "tensorflow/lite/kernels/padding.h" - -namespace tflite { -namespace ops { -namespace micro { -namespace depthwise_conv { -namespace { - -constexpr int kInputTensor = 0; -constexpr int kFilterTensor = 1; -constexpr int kBiasTensor = 2; -constexpr int kOutputTensor = 0; -constexpr int kMaxChannels = 256; - -// Depthwise conv is quantized along dimension 3: -// https://www.tensorflow.org/lite/performance/quantization_spec -constexpr int kDepthwiseConvQuantizedDimension = 3; - -// Size of the cached buffer we'll be using to hold reordered weights. -constexpr int kReshapedFilterDataSize = 1 * 1024; - -struct OpData { - TfLitePaddingValues padding; - // The scaling factor from input to output (aka the 'real multiplier') can - // be represented as a fixed point multiplier plus a left shift. - int32_t output_multiplier; - int output_shift; - - // Per channel output multiplier and shift. - int32_t per_channel_output_multiplier[kMaxChannels]; - int32_t per_channel_output_shift[kMaxChannels]; - - // The range of the fused activation layer. For example for kNone and - // uint8_t these would be 0 and 255. - int32_t output_activation_min; - int32_t output_activation_max; -}; - -TfLiteStatus CalculateOpData(TfLiteContext* context, TfLiteNode* node, - TfLiteDepthwiseConvParams* params, int width, - int height, int filter_width, int filter_height, - int out_width, int out_height, - const TfLiteType data_type, OpData* data) { - bool has_bias = node->inputs->size == 3; - // Check number of inputs/outputs - TF_LITE_ENSURE(context, has_bias || node->inputs->size == 2); - TF_LITE_ENSURE_EQ(context, node->outputs->size, 1); - - // Matching GetWindowedOutputSize in TensorFlow. - auto padding = params->padding; - data->padding = ComputePaddingHeightWidth( - params->stride_height, params->stride_width, - params->dilation_height_factor, params->dilation_width_factor, height, - width, filter_height, filter_width, padding, &out_height, &out_width); - - // Note that quantized inference requires that all tensors have their - // parameters set. This is usually done during quantized training. - if (data_type != kTfLiteFloat32) { - const TfLiteTensor* input = GetInput(context, node, kInputTensor); - const TfLiteTensor* filter = GetInput(context, node, kFilterTensor); - const TfLiteTensor* bias = - GetOptionalInputTensor(context, node, kBiasTensor); - TfLiteTensor* output = GetOutput(context, node, kOutputTensor); - int num_channels = filter->dims->data[kDepthwiseConvQuantizedDimension]; - - TF_LITE_ENSURE_STATUS(tflite::PopulateConvolutionQuantizationParams( - context, input, filter, bias, output, params->activation, - &data->output_multiplier, &data->output_shift, - &data->output_activation_min, &data->output_activation_max, - data->per_channel_output_multiplier, - reinterpret_cast(data->per_channel_output_shift), num_channels)); - } - return kTfLiteOk; -} - -// Specialized implementation of the depthwise convolution operation designed to -// work with the particular filter width of eight used by the default micro -// speech sample code. It uses 1KB of RAM to hold reordered weight parameters, -// converted from TFLite's NHWC format to NCHW format, and expressed as signed -// eight bit integers, rather than unsigned. Care must be taken when calling -// this not to use it for more than one node since there's only a single static -// buffer holding the weights. You should use this implementation if depthwise -// convolutions are a performance bottleneck, you have a layer that meets the -// parameter requirements, and the extra RAM usage and additional code size are -// not an issue. -static inline void DepthwiseConvOptimizedForFilterWidthEight( - TfLiteContext* context, const DepthwiseParams& params, - const RuntimeShape& input_shape, const uint8* input_data, - const RuntimeShape& filter_shape, const uint8* filter_data, - const RuntimeShape& bias_shape, const int32* bias_data, - const RuntimeShape& output_shape, uint8* output_data) { - const int stride_width = params.stride_width; - const int stride_height = params.stride_height; - const int pad_width = params.padding_values.width; - const int pad_height = params.padding_values.height; - const int depth_multiplier = params.depth_multiplier; - const int32 output_activation_min = params.quantized_activation_min; - const int32 output_activation_max = params.quantized_activation_max; - const int32 input_offset = params.input_offset; - const int32 filter_offset = params.weights_offset; - const int32 output_offset = params.output_offset; - const int32 output_multiplier = params.output_multiplier; - const int output_shift = params.output_shift; - TFLITE_DCHECK_EQ(input_shape.DimensionsCount(), 4); - TFLITE_DCHECK_EQ(filter_shape.DimensionsCount(), 4); - TFLITE_DCHECK_EQ(output_shape.DimensionsCount(), 4); - - TFLITE_DCHECK_LE(output_activation_min, output_activation_max); - const int batches = MatchingDim(input_shape, 0, output_shape, 0); - const int output_depth = MatchingDim(filter_shape, 3, output_shape, 3); - const int input_height = input_shape.Dims(1); - const int input_width = input_shape.Dims(2); - const int input_depth = input_shape.Dims(3); - const int filter_height = filter_shape.Dims(1); - const int filter_width = filter_shape.Dims(2); - const int output_height = output_shape.Dims(1); - const int output_width = output_shape.Dims(2); - TFLITE_DCHECK_EQ(output_depth, input_depth * depth_multiplier); - TFLITE_DCHECK_EQ(bias_shape.FlatSize(), output_depth); - - static int16_t reshaped_filter_data[kReshapedFilterDataSize]; - const int needed_size = - output_depth * filter_width * filter_height * input_depth; - if (needed_size > kReshapedFilterDataSize) { - TF_LITE_KERNEL_LOG( - context, - "Size too large for reshaped weight buffer (%d needed, %d available)", - needed_size, kReshapedFilterDataSize); - return; - } - - RuntimeShape reshaped_filter_shape; - reshaped_filter_shape.BuildFrom( - {1, output_depth, filter_height, filter_width}); - - // If this is the first time through, repack the weights into a cached buffer - // so that they can be accessed sequentially. - static bool is_reshaped_filter_initialized = false; - if (!is_reshaped_filter_initialized) { - for (int filter_y = 0; filter_y < filter_height; ++filter_y) { - for (int filter_x = 0; filter_x < filter_width; ++filter_x) { - for (int oc = 0; oc < output_depth; ++oc) { - const uint8* current_filter = - filter_data + Offset(filter_shape, 0, filter_y, filter_x, oc); - int16_t* reshaped_filter = - reshaped_filter_data + - Offset(reshaped_filter_shape, 0, oc, filter_y, filter_x); - *reshaped_filter = - static_cast(*current_filter) + filter_offset; - } - } - } - is_reshaped_filter_initialized = true; - } - - for (int b = 0; b < batches; ++b) { - for (int out_y = 0; out_y < output_height; ++out_y) { - for (int out_x = 0; out_x < output_width; ++out_x) { - for (int ic = 0; ic < input_depth; ++ic) { - for (int m = 0; m < depth_multiplier; m++) { - const int oc = m + ic * depth_multiplier; - const int in_x_origin = (out_x * stride_width) - pad_width; - const int in_y_origin = (out_y * stride_height) - pad_height; - int32 acc = 0; - int in_y_start = in_y_origin; - int filter_y_start = 0; - if (in_y_origin < 0) { - in_y_start = 0; - filter_y_start = 0 - in_y_origin; - } - int filter_y_end = filter_height; - if ((in_y_origin + filter_height) >= input_height) { - filter_y_end -= (in_y_origin + filter_height) - input_height; - } - int in_y = in_y_start; - int in_x_start = in_x_origin; - int filter_x_start = 0; - bool is_out_of_x_bounds = false; - if (in_x_origin < 0) { - in_x_start = 0; - filter_x_start = 0 - in_x_origin; - is_out_of_x_bounds = true; - } - int filter_x_end = filter_width; - if ((in_x_origin + filter_width) >= input_width) { - filter_x_end -= (in_x_origin + filter_width) - input_width; - is_out_of_x_bounds = true; - } - for (int filter_y = filter_y_start; filter_y < filter_y_end; - ++filter_y, ++in_y) { - const uint8* current_input = - input_data + Offset(input_shape, b, in_y, in_x_start, ic); - if ((filter_width == 8) && !is_out_of_x_bounds) { - int16* current_filter = - reshaped_filter_data + Offset(reshaped_filter_shape, 0, oc, - filter_y, filter_x_start); - const uint32_t input_vals0 = - *reinterpret_cast(current_input); - current_input += 4; - const int32_t filter_vals0 = - *reinterpret_cast(current_filter); - current_filter += 2; - const uint8 input_val0 = input_vals0 & 0xff; - const int16 filter_val0 = filter_vals0 & 0xffff; - acc += filter_val0 * input_val0; - const uint8 input_val1 = (input_vals0 >> 8) & 0xff; - const int16 filter_val1 = (filter_vals0 >> 16) & 0xffff; - acc += filter_val1 * input_val1; - - const int32_t filter_vals1 = - *reinterpret_cast(current_filter); - current_filter += 2; - const uint8 input_val2 = (input_vals0 >> 16) & 0xff; - const int16 filter_val2 = filter_vals1 & 0xffff; - acc += filter_val2 * input_val2; - const uint8 input_val3 = (input_vals0 >> 24) & 0xff; - const int16 filter_val3 = (filter_vals1 >> 16) & 0xffff; - acc += filter_val3 * input_val3; - - const uint32_t input_vals1 = - *reinterpret_cast(current_input); - const int32_t filter_vals2 = - *reinterpret_cast(current_filter); - current_filter += 2; - const uint8 input_val4 = input_vals1 & 0xff; - const int16 filter_val4 = filter_vals2 & 0xffff; - acc += filter_val4 * input_val4; - const uint8 input_val5 = (input_vals1 >> 8) & 0xff; - const int16 filter_val5 = (filter_vals2 >> 16) & 0xffff; - acc += filter_val5 * input_val5; - - const int32_t filter_vals3 = - *reinterpret_cast(current_filter); - const uint8 input_val6 = (input_vals1 >> 16) & 0xff; - const int16 filter_val6 = filter_vals3 & 0xffff; - acc += filter_val6 * input_val6; - const uint8 input_val7 = (input_vals1 >> 24) & 0xff; - const int16 filter_val7 = (filter_vals3 >> 16) & 0xffff; - acc += filter_val7 * input_val7; - } else { - const uint8* current_filter = - filter_data + - Offset(filter_shape, 0, filter_y, filter_x_start, oc); - for (int filter_x = filter_x_start; filter_x < filter_x_end; - ++filter_x) { - int32 input_val = *current_input; - current_input += input_depth; - int32 filter_val = *current_filter; - current_filter += output_depth; - acc += - (filter_val + filter_offset) * (input_val + input_offset); - } - } - } - if (bias_data) { - acc += bias_data[oc]; - } - acc = reference_ops::depthwise_conv::DepthwiseConvRound< - DepthwiseConvOutputRounding::kAwayFromZero>( - acc, output_multiplier, output_shift); - acc += output_offset; - acc = std::max(acc, output_activation_min); - acc = std::min(acc, output_activation_max); - output_data[Offset(output_shape, b, out_y, out_x, oc)] = - static_cast(acc); - } - } - } - } - } -} // namespace - -} // namespace - -void EvalFloat(TfLiteContext* context, TfLiteNode* node, - TfLiteDepthwiseConvParams* params, OpData* data, - const TfLiteTensor* input, const TfLiteTensor* filter, - const TfLiteTensor* bias, TfLiteTensor* output) { - float output_activation_min, output_activation_max; - CalculateActivationRange(params->activation, &output_activation_min, - &output_activation_max); - - tflite::DepthwiseParams op_params; - // Padding type is ignored, but still set. - op_params.padding_type = PaddingType::kSame; - op_params.padding_values.width = data->padding.width; - op_params.padding_values.height = data->padding.height; - op_params.stride_width = params->stride_width; - op_params.stride_height = params->stride_height; - op_params.dilation_width_factor = 1; - op_params.dilation_height_factor = 1; - op_params.depth_multiplier = params->depth_multiplier; - op_params.float_activation_min = output_activation_min; - op_params.float_activation_max = output_activation_max; - - tflite::reference_ops::DepthwiseConv( - op_params, GetTensorShape(input), GetTensorData(input), - GetTensorShape(filter), GetTensorData(filter), - GetTensorShape(bias), GetTensorData(bias), GetTensorShape(output), - GetTensorData(output)); -} - -// TODO(njeff): Optimize for int8 like we do for uint8. - -void EvalQuantizedPerChannel(TfLiteContext* context, TfLiteNode* node, - TfLiteDepthwiseConvParams* params, OpData* data, - const TfLiteTensor* input, - const TfLiteTensor* filter, - const TfLiteTensor* bias, TfLiteTensor* output) { - DepthwiseParams op_params; - op_params.padding_type = PaddingType::kSame; - op_params.padding_values.width = data->padding.width; - op_params.padding_values.height = data->padding.height; - op_params.stride_width = params->stride_width; - op_params.stride_height = params->stride_height; - op_params.dilation_width_factor = params->dilation_width_factor; - op_params.dilation_height_factor = params->dilation_height_factor; - op_params.depth_multiplier = params->depth_multiplier; - op_params.input_offset = -input->params.zero_point; - op_params.weights_offset = 0; - op_params.output_offset = output->params.zero_point; - // TODO(b/130439627): Use calculated value for clamping. - op_params.quantized_activation_min = std::numeric_limits::min(); - op_params.quantized_activation_max = std::numeric_limits::max(); - - reference_integer_ops::DepthwiseConvPerChannel( - op_params, data->per_channel_output_multiplier, - data->per_channel_output_shift, GetTensorShape(input), - GetTensorData(input), GetTensorShape(filter), - GetTensorData(filter), GetTensorShape(bias), - GetTensorData(bias), GetTensorShape(output), - GetTensorData(output)); -} - -void EvalQuantized(TfLiteContext* context, TfLiteNode* node, - TfLiteDepthwiseConvParams* params, OpData* data, - const TfLiteTensor* input, const TfLiteTensor* filter, - const TfLiteTensor* bias, TfLiteTensor* output) { - const int32_t input_offset = -input->params.zero_point; - const int32_t filter_offset = -filter->params.zero_point; - const int32_t output_offset = output->params.zero_point; - - tflite::DepthwiseParams op_params; - // Padding type is ignored, but still set. - op_params.padding_type = PaddingType::kSame; - op_params.padding_values.width = data->padding.width; - op_params.padding_values.height = data->padding.height; - op_params.stride_width = params->stride_width; - op_params.stride_height = params->stride_height; - op_params.dilation_width_factor = 1; - op_params.dilation_height_factor = 1; - op_params.depth_multiplier = params->depth_multiplier; - op_params.quantized_activation_min = data->output_activation_min; - op_params.quantized_activation_max = data->output_activation_max; - op_params.input_offset = input_offset; - op_params.weights_offset = filter_offset; - op_params.output_offset = output_offset; - op_params.output_multiplier = data->output_multiplier; - // Legacy ops used mixed left and right shifts. Now all are +ve-means-left. - op_params.output_shift = -data->output_shift; - - // Figure out if we can use the optimized path for this set of parameters. - const int filter_width = GetTensorShape(filter).Dims(2); - const int input_depth = GetTensorShape(input).Dims(3); - const int output_depth = GetTensorShape(filter).Dims(3); - const int filter_height = GetTensorShape(filter).Dims(1); - const int needed_size = - output_depth * filter_width * filter_height * input_depth; - bool use_optimized_path = false; - if ((filter_width == 8) && (input_offset == 0) && (input_depth == 1) && - (needed_size <= kReshapedFilterDataSize)) { - // FIXME(petewarden) - We need a more robust way of handling this, ideally - // with an allocation mechanism available through the context API. - // Use the address of the node as a proxy for its identity, since we need - // to ensure the weight values are consistent between calls, and there's - // no easy way to do that quickly other than relying on the identity of - // the owning node. - static TfLiteNode* initialized_node_address = node; - if (initialized_node_address == node) { - use_optimized_path = true; - } else { - static bool has_warned = false; - if (!has_warned) { - TF_LITE_KERNEL_LOG( - context, - "Multiple depthwise conv ops match optimization parameters, but " - "only the first will use the fast path, because there's only one " - "RAM cache available"); - has_warned = true; - } - } - } - if (use_optimized_path) { - DepthwiseConvOptimizedForFilterWidthEight( - context, op_params, GetTensorShape(input), - GetTensorData(input), GetTensorShape(filter), - GetTensorData(filter), GetTensorShape(bias), - GetTensorData(bias), GetTensorShape(output), - GetTensorData(output)); - } else { - tflite::reference_ops::DepthwiseConv( - op_params, GetTensorShape(input), GetTensorData(input), - GetTensorShape(filter), GetTensorData(filter), - GetTensorShape(bias), GetTensorData(bias), - GetTensorShape(output), GetTensorData(output)); - } -} - -TfLiteStatus Eval(TfLiteContext* context, TfLiteNode* node) { - auto* params = - reinterpret_cast(node->builtin_data); - - TfLiteTensor* output = GetOutput(context, node, kOutputTensor); - const TfLiteTensor* input = GetInput(context, node, kInputTensor); - const TfLiteTensor* filter = GetInput(context, node, kFilterTensor); - const TfLiteTensor* bias = - (NumInputs(node) == 3) ? GetInput(context, node, kBiasTensor) : nullptr; - - const TfLiteType data_type = input->type; - int width = SizeOfDimension(input, 2); - int height = SizeOfDimension(input, 1); - int filter_width = SizeOfDimension(filter, 2); - int filter_height = SizeOfDimension(filter, 1); - int out_width = ComputeOutSize(params->padding, width, filter_width, - params->stride_width); - int out_height = ComputeOutSize(params->padding, height, filter_height, - params->stride_height); - OpData data; - - // All per-channel quantized tensors need valid zero point and scale arrays. - if (input->type == kTfLiteInt8) { - TF_LITE_ENSURE_EQ(context, filter->quantization.type, - kTfLiteAffineQuantization); - - const auto* affine_quantization = - reinterpret_cast( - filter->quantization.params); - TF_LITE_ENSURE(context, affine_quantization); - TF_LITE_ENSURE(context, affine_quantization->scale); - TF_LITE_ENSURE(context, affine_quantization->zero_point); - TF_LITE_ENSURE( - context, affine_quantization->scale->size == 1 || - affine_quantization->scale->size == - filter->dims->data[kDepthwiseConvQuantizedDimension]); - TF_LITE_ENSURE_EQ(context, affine_quantization->scale->size, - affine_quantization->zero_point->size); - } - - TF_LITE_ENSURE_STATUS(CalculateOpData(context, node, params, width, height, - filter_width, filter_height, out_width, - out_height, data_type, &data)); - - // TODO(aselle): Consider whether float conv and quantized conv should be - // separate ops to avoid dispatch overhead here. - switch (input->type) { // Already know in/out types are same. - case kTfLiteFloat32: - EvalFloat(context, node, params, &data, input, filter, bias, output); - break; - case kTfLiteInt8: - EvalQuantizedPerChannel(context, node, params, &data, input, filter, bias, - output); - break; - case kTfLiteUInt8: - EvalQuantized(context, node, params, &data, input, filter, bias, output); - break; - default: - TF_LITE_KERNEL_LOG(context, "Type %s (%d) not supported.", - TfLiteTypeGetName(input->type), input->type); - return kTfLiteError; - } - return kTfLiteOk; -} - -} // namespace depthwise_conv - -TfLiteRegistration* Register_DEPTHWISE_CONV_2D() { - static TfLiteRegistration r = {/*init=*/nullptr, - /*free=*/nullptr, - /*prepare=*/nullptr, - /*invoke=*/depthwise_conv::Eval, - /*profiling_string=*/nullptr, - /*builtin_code=*/0, - /*custom_name=*/nullptr, - /*version=*/0}; - return &r; -} - -} // namespace micro -} // namespace ops -} // namespace tflite diff --git a/tensorflow/lite/micro/tools/ci_build/test_mbed.sh b/tensorflow/lite/micro/tools/ci_build/test_mbed.sh index a4d47009c93..fa4506fa6b8 100755 --- a/tensorflow/lite/micro/tools/ci_build/test_mbed.sh +++ b/tensorflow/lite/micro/tools/ci_build/test_mbed.sh @@ -49,7 +49,7 @@ fi make -f tensorflow/lite/micro/tools/make/Makefile \ TARGET=${TARGET} \ - TAGS="portable_optimized disco_f746ng" \ + TAGS="disco_f746ng" \ ${PROJECTS} readable_run tensorflow/lite/micro/tools/ci_build/install_mbed_cli.sh diff --git a/tensorflow/lite/micro/tools/make/targets/apollo3evb_makefile.inc b/tensorflow/lite/micro/tools/make/targets/apollo3evb_makefile.inc index 7d2a0e65b97..dc7a689daed 100644 --- a/tensorflow/lite/micro/tools/make/targets/apollo3evb_makefile.inc +++ b/tensorflow/lite/micro/tools/make/targets/apollo3evb_makefile.inc @@ -24,9 +24,6 @@ ifeq ($(TARGET),$(filter $(TARGET),\ $(MAKEFILE_DIR)/downloads/$(AM_SDK_DEST)/$(SF_BSPS_DEST): $(MAKEFILE_DIR)/downloads/$(AM_SDK_DEST) endif - # Use the faster depthwise conv implementation. - ALL_TAGS += portable_optimized - PLATFORM_FLAGS = \ -DPART_apollo3 \ -DAM_PACKAGE_BGA \ From 088d4376945c0ee6f3def04284ff1a835ea57097 Mon Sep 17 00:00:00 2001 From: Hye Soo Yang Date: Mon, 22 Jun 2020 22:38:05 -0700 Subject: [PATCH 529/867] Fix BMP header bytes. PiperOrigin-RevId: 317803857 Change-Id: Iafa8e4e44400ec3c119af8897ba3fa693138501f --- tensorflow/python/kernel_tests/decode_bmp_op_test.py | 9 ++++++--- 1 file changed, 6 insertions(+), 3 deletions(-) diff --git a/tensorflow/python/kernel_tests/decode_bmp_op_test.py b/tensorflow/python/kernel_tests/decode_bmp_op_test.py index 5e7991382ed..b0a8f9ffbbb 100644 --- a/tensorflow/python/kernel_tests/decode_bmp_op_test.py +++ b/tensorflow/python/kernel_tests/decode_bmp_op_test.py @@ -25,14 +25,14 @@ from tensorflow.python.ops import image_ops from tensorflow.python.platform import test - class DecodeBmpOpTest(test.TestCase): def testex1(self): img_bytes = [[[0, 0, 255], [0, 255, 0]], [[255, 0, 0], [255, 255, 255]]] # Encoded BMP bytes from Wikipedia + # BMP header bytes: https://en.wikipedia.org/wiki/List_of_file_signatures encoded_bytes = [ - 0x42, 0x40, + 0x42, 0x4d, 0x46, 0, 0, 0, 0, 0, 0, 0, @@ -66,9 +66,10 @@ class DecodeBmpOpTest(test.TestCase): def testGrayscale(self): img_bytes = [[[255], [0]], [[255], [0]]] + # BMP header bytes: https://en.wikipedia.org/wiki/List_of_file_signatures encoded_bytes = [ 0x42, - 0x40, + 0x4d, 0x3d, 0, 0, @@ -133,6 +134,8 @@ class DecodeBmpOpTest(test.TestCase): byte_string = bytes(bytearray(encoded_bytes)) img_in = constant_op.constant(byte_string, dtype=dtypes.string) + # TODO(b/159600494): Currently, `decode_bmp` op does not validate input + # magic bytes. decode = image_ops.decode_bmp(img_in) with self.cached_session(): From dce31a431df80aa55f2cfd87d60952c413d6f690 Mon Sep 17 00:00:00 2001 From: chuanqiw Date: Tue, 23 Jun 2020 14:08:27 +0800 Subject: [PATCH 530/867] Update sqlite version to 3.32.1 to fix the CVE-2020-13630 and CVE-2020-11656 --- tensorflow/workspace.bzl | 8 ++++---- 1 file changed, 4 insertions(+), 4 deletions(-) diff --git a/tensorflow/workspace.bzl b/tensorflow/workspace.bzl index 0f591ba8b90..29c6ef99397 100755 --- a/tensorflow/workspace.bzl +++ b/tensorflow/workspace.bzl @@ -409,12 +409,12 @@ def tf_repositories(path_prefix = "", tf_repo_name = ""): tf_http_archive( name = "org_sqlite", build_file = clean_dep("//third_party:sqlite.BUILD"), - sha256 = "f3c79bc9f4162d0b06fa9fe09ee6ccd23bb99ce310b792c5145f87fbcc30efca", - strip_prefix = "sqlite-amalgamation-3310100", + sha256 = "8d46ef69b96628bedb781bd8309210f2a1f4a353792097302f6b754044e6540f", + strip_prefix = "sqlite-amalgamation-3320100", system_build_file = clean_dep("//third_party/systemlibs:sqlite.BUILD"), urls = [ - "https://storage.googleapis.com/mirror.tensorflow.org/www.sqlite.org/2020/sqlite-amalgamation-3310100.zip", - "https://www.sqlite.org/2020/sqlite-amalgamation-3310100.zip", + "https://storage.googleapis.com/mirror.tensorflow.org/www.sqlite.org/2020/sqlite-amalgamation-3320100.zip", + "https://www.sqlite.org/2020/sqlite-amalgamation-3320100.zip", ], ) From 51ee63247d1d4e4c6db089875b605963f31de2d9 Mon Sep 17 00:00:00 2001 From: Gaurav Jain Date: Mon, 22 Jun 2020 23:04:46 -0700 Subject: [PATCH 531/867] Rollforward: Add uint32 & uint64 to TF_CALL_INTEGRAL_TYPES PiperOrigin-RevId: 317806755 Change-Id: I3752198bf98d5659bf163455a5dfe1fbe39ffd1c --- tensorflow/core/framework/register_types.h | 21 +++++--------- tensorflow/core/framework/types.cc | 5 ---- tensorflow/core/kernels/BUILD | 2 ++ tensorflow/core/kernels/concat_lib_cpu.cc | 2 -- tensorflow/core/kernels/concat_op.cc | 2 -- tensorflow/core/kernels/constant_op.cc | 1 - tensorflow/core/kernels/control_flow_ops.cc | 5 ---- .../core/kernels/data/dataset_test_base.cc | 2 -- tensorflow/core/kernels/dense_update_ops.cc | 1 - .../core/kernels/dynamic_partition_op.cc | 2 -- tensorflow/core/kernels/fill_functor.cc | 5 +++- tensorflow/core/kernels/gather_op.cc | 2 -- tensorflow/core/kernels/identity_op.cc | 1 - tensorflow/core/kernels/ragged_gather_op.cc | 2 -- .../kernels/ragged_tensor_from_variant_op.cc | 2 -- .../kernels/ragged_tensor_to_tensor_op.cc | 2 -- .../kernels/ragged_tensor_to_variant_op.cc | 2 -- .../core/kernels/resource_variable_ops.cc | 1 - tensorflow/core/kernels/split_lib_cpu.cc | 1 - tensorflow/core/kernels/split_op.cc | 1 - tensorflow/core/kernels/strided_slice_op.cc | 2 -- .../core/kernels/strided_slice_op_impl.h | 2 -- tensorflow/core/kernels/topk_op.cc | 2 -- .../core/kernels/topk_op_gpu_uint32.cu.cc | 28 +++++++++++++++++++ .../core/kernels/topk_op_gpu_uint64.cu.cc | 28 +++++++++++++++++++ tensorflow/core/util/batch_util.cc | 8 ------ .../core/util/saved_tensor_slice_util.h | 2 ++ 27 files changed, 71 insertions(+), 63 deletions(-) create mode 100644 tensorflow/core/kernels/topk_op_gpu_uint32.cu.cc create mode 100644 tensorflow/core/kernels/topk_op_gpu_uint64.cu.cc diff --git a/tensorflow/core/framework/register_types.h b/tensorflow/core/framework/register_types.h index bc3e5e1743b..0cf6536e8c2 100644 --- a/tensorflow/core/framework/register_types.h +++ b/tensorflow/core/framework/register_types.h @@ -153,16 +153,9 @@ limitations under the License. #endif // defined(IS_MOBILE_PLATFORM) - end of TF_CALL_type defines // Defines for sets of types. - -// TODO(b/111604096): Add uint32 and uint64 to TF_CALL_INTEGRAL_TYPES. -// -// The uint32 and uint64 types were introduced in 10/2017 to be used via XLA and -// thus were not included in TF_CALL_INTEGRAL_TYPES. Including them in -// TF_CALL_INTEGRAL_TYPES should only happen after evaluating the effect on the -// TF binary size and performance. -#define TF_CALL_INTEGRAL_TYPES(m) \ - TF_CALL_int64(m) TF_CALL_int32(m) TF_CALL_uint16(m) TF_CALL_int16(m) \ - TF_CALL_uint8(m) TF_CALL_int8(m) +#define TF_CALL_INTEGRAL_TYPES(m) \ + TF_CALL_uint64(m) TF_CALL_int64(m) TF_CALL_uint32(m) TF_CALL_int32(m) \ + TF_CALL_uint16(m) TF_CALL_int16(m) TF_CALL_uint8(m) TF_CALL_int8(m) #define TF_CALL_FLOAT_TYPES(m) \ TF_CALL_half(m) TF_CALL_bfloat16(m) TF_CALL_float(m) TF_CALL_double(m) @@ -174,10 +167,10 @@ limitations under the License. #define TF_CALL_REAL_NUMBER_TYPES_NO_BFLOAT16(m) \ TF_CALL_INTEGRAL_TYPES(m) TF_CALL_half(m) TF_CALL_float(m) TF_CALL_double(m) -#define TF_CALL_REAL_NUMBER_TYPES_NO_INT32(m) \ - TF_CALL_half(m) TF_CALL_bfloat16(m) TF_CALL_float(m) TF_CALL_double(m) \ - TF_CALL_int64(m) TF_CALL_uint16(m) TF_CALL_int16(m) TF_CALL_uint8(m) \ - TF_CALL_int8(m) +#define TF_CALL_REAL_NUMBER_TYPES_NO_INT32(m) \ + TF_CALL_half(m) TF_CALL_bfloat16(m) TF_CALL_float(m) TF_CALL_double(m) \ + TF_CALL_uint64(m) TF_CALL_int64(m) TF_CALL_uint32(m) TF_CALL_uint16(m) \ + TF_CALL_int16(m) TF_CALL_uint8(m) TF_CALL_int8(m) #define TF_CALL_COMPLEX_TYPES(m) TF_CALL_complex64(m) TF_CALL_complex128(m) diff --git a/tensorflow/core/framework/types.cc b/tensorflow/core/framework/types.cc index 97eaec98ffe..d6455e012d0 100644 --- a/tensorflow/core/framework/types.cc +++ b/tensorflow/core/framework/types.cc @@ -238,11 +238,6 @@ int DataTypeSize(DataType dt) { TF_CALL_qint16(CASE); TF_CALL_quint16(CASE); - // uint32 and uint64 aren't included in TF_CALL_POD_TYPES because we - // don't want to define kernels for them at this stage to avoid binary - // bloat. - TF_CALL_uint32(CASE); - TF_CALL_uint64(CASE); default: return 0; } diff --git a/tensorflow/core/kernels/BUILD b/tensorflow/core/kernels/BUILD index 7da864a6027..e2ff5aed283 100644 --- a/tensorflow/core/kernels/BUILD +++ b/tensorflow/core/kernels/BUILD @@ -4919,7 +4919,9 @@ tf_kernel_library( "topk_op_gpu_double.cu.cc", "topk_op_gpu_float.cu.cc", "topk_op_gpu_half.cu.cc", + "topk_op_gpu_uint64.cu.cc", "topk_op_gpu_int64.cu.cc", + "topk_op_gpu_uint32.cu.cc", "topk_op_gpu_int32.cu.cc", "topk_op_gpu_int16.cu.cc", "topk_op_gpu_uint16.cu.cc", diff --git a/tensorflow/core/kernels/concat_lib_cpu.cc b/tensorflow/core/kernels/concat_lib_cpu.cc index da73d3d2c56..1dec589d3ff 100644 --- a/tensorflow/core/kernels/concat_lib_cpu.cc +++ b/tensorflow/core/kernels/concat_lib_cpu.cc @@ -116,8 +116,6 @@ REGISTER(qint8) REGISTER(quint16) REGISTER(qint16) REGISTER(qint32) -REGISTER(uint32) -REGISTER(uint64) #if defined(IS_MOBILE_PLATFORM) && !defined(SUPPORT_SELECTIVE_REGISTRATION) && \ !defined(__ANDROID_TYPES_FULL__) diff --git a/tensorflow/core/kernels/concat_op.cc b/tensorflow/core/kernels/concat_op.cc index be3e9a67c5f..d3f3a04f33b 100644 --- a/tensorflow/core/kernels/concat_op.cc +++ b/tensorflow/core/kernels/concat_op.cc @@ -208,8 +208,6 @@ REGISTER_CONCAT(qint8); REGISTER_CONCAT(quint16); REGISTER_CONCAT(qint16); REGISTER_CONCAT(qint32); -REGISTER_CONCAT(uint32); -REGISTER_CONCAT(uint64); #undef REGISTER_CONCAT diff --git a/tensorflow/core/kernels/constant_op.cc b/tensorflow/core/kernels/constant_op.cc index 4bcbc076446..dc178d17d49 100644 --- a/tensorflow/core/kernels/constant_op.cc +++ b/tensorflow/core/kernels/constant_op.cc @@ -211,7 +211,6 @@ TF_CALL_ALL_TYPES(REGISTER_CPU_KERNEL); // the conversion from uint8 to quint8. REGISTER_KERNEL(CPU, quint8); REGISTER_KERNEL(CPU, quint16); -REGISTER_KERNEL(CPU, uint32); #undef REGISTER_CPU_KERNEL #ifdef TENSORFLOW_USE_SYCL diff --git a/tensorflow/core/kernels/control_flow_ops.cc b/tensorflow/core/kernels/control_flow_ops.cc index 435de3c5954..1a0082c6a3b 100644 --- a/tensorflow/core/kernels/control_flow_ops.cc +++ b/tensorflow/core/kernels/control_flow_ops.cc @@ -101,16 +101,12 @@ TF_CALL_ALL_TYPES(REGISTER_CPU_SWITCH); TF_CALL_ALL_TYPES(REGISTER_CPU_REF_SWITCH); TF_CALL_QUANTIZED_TYPES(REGISTER_CPU_SWITCH); TF_CALL_QUANTIZED_TYPES(REGISTER_CPU_REF_SWITCH); -REGISTER_CPU_SWITCH(uint64); TF_CALL_NUMBER_TYPES_NO_INT32(REGISTER_GPU_SWITCH); TF_CALL_QUANTIZED_TYPES(REGISTER_GPU_SWITCH); TF_CALL_NUMBER_TYPES_NO_INT32(REGISTER_GPU_REF_SWITCH); TF_CALL_QUANTIZED_TYPES(REGISTER_GPU_REF_SWITCH); -REGISTER_GPU_SWITCH(uint64); TF_CALL_variant(REGISTER_GPU_SWITCH); -TF_CALL_uint32(REGISTER_GPU_SWITCH); -TF_CALL_uint32(REGISTER_GPU_REF_SWITCH); TF_CALL_bool(REGISTER_GPU_SWITCH); TF_CALL_bool(REGISTER_GPU_REF_SWITCH); @@ -311,7 +307,6 @@ TF_CALL_QUANTIZED_TYPES(REGISTER_GPU_KERNEL); TF_CALL_QUANTIZED_TYPES(REGISTER_GPU_REF_KERNEL); REGISTER_GPU_KERNEL(bool); REGISTER_GPU_REF_KERNEL(bool); -REGISTER_GPU_KERNEL(uint64); TF_CALL_variant(REGISTER_GPU_KERNEL); #undef REGISTER_GPU_KERNEL diff --git a/tensorflow/core/kernels/data/dataset_test_base.cc b/tensorflow/core/kernels/data/dataset_test_base.cc index b91ab9b733c..e41e35be1e9 100644 --- a/tensorflow/core/kernels/data/dataset_test_base.cc +++ b/tensorflow/core/kernels/data/dataset_test_base.cc @@ -220,8 +220,6 @@ Status DatasetOpsTestBase::ExpectEqual(const Tensor& a, const Tensor& b) { break; TF_CALL_NUMBER_TYPES(CASE); TF_CALL_tstring(CASE); - TF_CALL_uint32(CASE); - TF_CALL_uint64(CASE); // TODO(feihugis): figure out how to support variant tensors. #undef CASE default: diff --git a/tensorflow/core/kernels/dense_update_ops.cc b/tensorflow/core/kernels/dense_update_ops.cc index 55e4cd7606a..71235fca143 100644 --- a/tensorflow/core/kernels/dense_update_ops.cc +++ b/tensorflow/core/kernels/dense_update_ops.cc @@ -98,7 +98,6 @@ typedef Eigen::SyclDevice SYCLDevice; TF_CALL_ALL_TYPES(REGISTER_KERNELS); // uint32 not included in ALL_TYPES -TF_CALL_uint32(REGISTER_KERNELS); TF_CALL_QUANTIZED_TYPES(REGISTER_KERNELS); // quint16 not included in QUANTIZIED_TYPES TF_CALL_quint16(REGISTER_KERNELS); diff --git a/tensorflow/core/kernels/dynamic_partition_op.cc b/tensorflow/core/kernels/dynamic_partition_op.cc index 90ed71dccce..95af19c4c48 100644 --- a/tensorflow/core/kernels/dynamic_partition_op.cc +++ b/tensorflow/core/kernels/dynamic_partition_op.cc @@ -164,8 +164,6 @@ class DynamicPartitionOp : public DynamicPartitionOp_Shared { DynamicPartitionOp) TF_CALL_ALL_TYPES(REGISTER_DYNAMIC_PARTITION); -// For partitioning fingerprints. -TF_CALL_uint64(REGISTER_DYNAMIC_PARTITION); #undef REGISTER_DYNAMIC_PARTITION } // namespace tensorflow diff --git a/tensorflow/core/kernels/fill_functor.cc b/tensorflow/core/kernels/fill_functor.cc index 10dd3df1915..174a4e45a79 100644 --- a/tensorflow/core/kernels/fill_functor.cc +++ b/tensorflow/core/kernels/fill_functor.cc @@ -45,6 +45,8 @@ DEFINE_SETZERO_CPU(Eigen::half); DEFINE_SETZERO_CPU(bfloat16); DEFINE_SETZERO_CPU(float); DEFINE_SETZERO_CPU(double); +DEFINE_SETZERO_CPU(uint32); +DEFINE_SETZERO_CPU(uint64); DEFINE_SETZERO_CPU(uint8); DEFINE_SETZERO_CPU(int8); DEFINE_SETZERO_CPU(uint16); @@ -96,6 +98,8 @@ DEFINE_SETONE_CPU(Eigen::half); DEFINE_SETONE_CPU(bfloat16); DEFINE_SETONE_CPU(float); DEFINE_SETONE_CPU(double); +DEFINE_SETONE_CPU(uint32); +DEFINE_SETONE_CPU(uint64); DEFINE_SETONE_CPU(uint8); DEFINE_SETONE_CPU(int8); DEFINE_SETONE_CPU(uint16); @@ -137,7 +141,6 @@ struct FillFunctor { TF_CALL_ALL_TYPES(DEFINE_FILL_CPU); DEFINE_FILL_CPU(quint8); DEFINE_FILL_CPU(quint16); -DEFINE_FILL_CPU(uint32); #undef DEFINE_FILL_CPU #ifdef TENSORFLOW_USE_SYCL diff --git a/tensorflow/core/kernels/gather_op.cc b/tensorflow/core/kernels/gather_op.cc index 6d493a5f2ea..948567e019a 100644 --- a/tensorflow/core/kernels/gather_op.cc +++ b/tensorflow/core/kernels/gather_op.cc @@ -211,8 +211,6 @@ TF_CALL_ALL_TYPES(REGISTER_GATHER_CPU); TF_CALL_QUANTIZED_TYPES(REGISTER_GATHER_CPU); TF_CALL_quint16(REGISTER_GATHER_CPU); TF_CALL_qint16(REGISTER_GATHER_CPU); -TF_CALL_uint32(REGISTER_GATHER_CPU); -TF_CALL_uint64(REGISTER_GATHER_CPU); #undef REGISTER_GATHER_CPU diff --git a/tensorflow/core/kernels/identity_op.cc b/tensorflow/core/kernels/identity_op.cc index d15b64597f5..4b226dd72d4 100644 --- a/tensorflow/core/kernels/identity_op.cc +++ b/tensorflow/core/kernels/identity_op.cc @@ -122,7 +122,6 @@ REGISTER_SYCL_HOST_KERNEL(bool); TF_CALL_NUMBER_TYPES_NO_INT32(REGISTER_GPU_KERNEL); REGISTER_GPU_KERNEL(Variant); -TF_CALL_uint32(REGISTER_GPU_KERNEL); REGISTER_GPU_KERNEL(bool); #undef REGISTER_GPU_KERNEL diff --git a/tensorflow/core/kernels/ragged_gather_op.cc b/tensorflow/core/kernels/ragged_gather_op.cc index 88c0d1ebd69..3bf82cba050 100644 --- a/tensorflow/core/kernels/ragged_gather_op.cc +++ b/tensorflow/core/kernels/ragged_gather_op.cc @@ -296,8 +296,6 @@ TF_CALL_tstring(REGISTER_CPU_KERNEL); TF_CALL_QUANTIZED_TYPES(REGISTER_CPU_KERNEL); TF_CALL_quint16(REGISTER_CPU_KERNEL); TF_CALL_qint16(REGISTER_CPU_KERNEL); -TF_CALL_uint32(REGISTER_CPU_KERNEL); -TF_CALL_uint64(REGISTER_CPU_KERNEL); #undef REGISTER_CPU_KERNEL #undef REGISTER_CPU_KERNEL_WITH_INDEX_TYPE diff --git a/tensorflow/core/kernels/ragged_tensor_from_variant_op.cc b/tensorflow/core/kernels/ragged_tensor_from_variant_op.cc index f83bcb38c6c..ad0712e6fd0 100644 --- a/tensorflow/core/kernels/ragged_tensor_from_variant_op.cc +++ b/tensorflow/core/kernels/ragged_tensor_from_variant_op.cc @@ -308,8 +308,6 @@ TF_CALL_tstring(REGISTER_KERNELS); TF_CALL_QUANTIZED_TYPES(REGISTER_KERNELS); TF_CALL_quint16(REGISTER_KERNELS); TF_CALL_qint16(REGISTER_KERNELS); -TF_CALL_uint32(REGISTER_KERNELS); -TF_CALL_uint64(REGISTER_KERNELS); #undef REGISTER_KERNELS #undef REGISTER_KERNELS_WITH_SPLIT_TYPE } // namespace tensorflow diff --git a/tensorflow/core/kernels/ragged_tensor_to_tensor_op.cc b/tensorflow/core/kernels/ragged_tensor_to_tensor_op.cc index d729c43f25a..9ae5d7ffbdc 100644 --- a/tensorflow/core/kernels/ragged_tensor_to_tensor_op.cc +++ b/tensorflow/core/kernels/ragged_tensor_to_tensor_op.cc @@ -561,8 +561,6 @@ TF_CALL_string(REGISTER_CPU_KERNEL); TF_CALL_QUANTIZED_TYPES(REGISTER_CPU_KERNEL); TF_CALL_quint16(REGISTER_CPU_KERNEL); TF_CALL_qint16(REGISTER_CPU_KERNEL); -TF_CALL_uint32(REGISTER_CPU_KERNEL); -TF_CALL_uint64(REGISTER_CPU_KERNEL); #undef REGISTER_CPU_KERNEL diff --git a/tensorflow/core/kernels/ragged_tensor_to_variant_op.cc b/tensorflow/core/kernels/ragged_tensor_to_variant_op.cc index 7a5ae1c6240..64c372b005e 100644 --- a/tensorflow/core/kernels/ragged_tensor_to_variant_op.cc +++ b/tensorflow/core/kernels/ragged_tensor_to_variant_op.cc @@ -213,8 +213,6 @@ TF_CALL_tstring(REGISTER_KERNELS); TF_CALL_QUANTIZED_TYPES(REGISTER_KERNELS); TF_CALL_quint16(REGISTER_KERNELS); TF_CALL_qint16(REGISTER_KERNELS); -TF_CALL_uint32(REGISTER_KERNELS); -TF_CALL_uint64(REGISTER_KERNELS); #undef REGISTER_KERNELS #undef REGISTER_KERNELS_WITH_SPLIT_TYPE } // namespace tensorflow diff --git a/tensorflow/core/kernels/resource_variable_ops.cc b/tensorflow/core/kernels/resource_variable_ops.cc index b9c883c7e2f..510e95ca606 100644 --- a/tensorflow/core/kernels/resource_variable_ops.cc +++ b/tensorflow/core/kernels/resource_variable_ops.cc @@ -508,7 +508,6 @@ class AssignVariableOp : public OpKernel { TF_CALL_ALL_TYPES(REGISTER_KERNELS); TF_CALL_QUANTIZED_TYPES(REGISTER_KERNELS); -TF_CALL_uint32(REGISTER_KERNELS); #undef REGISTER_KERNELS #if GOOGLE_CUDA || TENSORFLOW_USE_ROCM diff --git a/tensorflow/core/kernels/split_lib_cpu.cc b/tensorflow/core/kernels/split_lib_cpu.cc index 0cb0a94d498..a3060e4e90d 100644 --- a/tensorflow/core/kernels/split_lib_cpu.cc +++ b/tensorflow/core/kernels/split_lib_cpu.cc @@ -43,7 +43,6 @@ void Split::operator()( TF_CALL_ALL_TYPES(DEFINE_CPU_KERNELS) DEFINE_CPU_KERNELS(quint8) -DEFINE_CPU_KERNELS(uint64) #ifdef TENSORFLOW_USE_SYCL template diff --git a/tensorflow/core/kernels/split_op.cc b/tensorflow/core/kernels/split_op.cc index f09740c6198..08575f01f67 100644 --- a/tensorflow/core/kernels/split_op.cc +++ b/tensorflow/core/kernels/split_op.cc @@ -404,7 +404,6 @@ class SplitOpSYCL : public SplitOpBase { TF_CALL_ALL_TYPES(REGISTER_SPLIT); REGISTER_SPLIT(quint8); -REGISTER_SPLIT(uint64); #undef REGISTER_SPLIT diff --git a/tensorflow/core/kernels/strided_slice_op.cc b/tensorflow/core/kernels/strided_slice_op.cc index ccc1984bb98..b4099213303 100644 --- a/tensorflow/core/kernels/strided_slice_op.cc +++ b/tensorflow/core/kernels/strided_slice_op.cc @@ -440,8 +440,6 @@ class StridedSliceAssignOp : public OpKernel { StridedSliceAssignOp) TF_CALL_ALL_TYPES(REGISTER_STRIDED_SLICE); -TF_CALL_uint32(REGISTER_STRIDED_SLICE); -TF_CALL_uint64(REGISTER_STRIDED_SLICE); #undef REGISTER_STRIDED_SLICE diff --git a/tensorflow/core/kernels/strided_slice_op_impl.h b/tensorflow/core/kernels/strided_slice_op_impl.h index 1ae959b7b3f..5ce1d773e33 100644 --- a/tensorflow/core/kernels/strided_slice_op_impl.h +++ b/tensorflow/core/kernels/strided_slice_op_impl.h @@ -287,8 +287,6 @@ TF_CALL_GPU_ALL_TYPES(DECLARE_FOR_N_GPU); #endif // END GOOGLE_CUDA || TENSORFLOW_USE_ROCM TF_CALL_ALL_TYPES(DECLARE_FOR_N_CPU); -TF_CALL_uint32(DECLARE_FOR_N_CPU); -TF_CALL_uint64(DECLARE_FOR_N_CPU); #ifdef TENSORFLOW_USE_SYCL #define PREVENT_FOR_N_SYCL(T) \ diff --git a/tensorflow/core/kernels/topk_op.cc b/tensorflow/core/kernels/topk_op.cc index c555b42f005..50325b7bcfe 100644 --- a/tensorflow/core/kernels/topk_op.cc +++ b/tensorflow/core/kernels/topk_op.cc @@ -258,7 +258,6 @@ namespace functor { TF_CALL_GPU_NUMBER_TYPES(DECLARE_GPU_SPEC); TF_CALL_INTEGRAL_TYPES(DECLARE_GPU_SPEC); -TF_CALL_uint32(DECLARE_GPU_SPEC); #undef DECLARE_GPU_SPEC @@ -276,7 +275,6 @@ TF_CALL_uint32(DECLARE_GPU_SPEC); TF_CALL_GPU_NUMBER_TYPES(REGISTER_KERNELS); TF_CALL_INTEGRAL_TYPES(REGISTER_KERNELS); -TF_CALL_uint32(REGISTER_KERNELS) #undef REGISTER_KERNELS #endif // end GOOGLE_CUDA diff --git a/tensorflow/core/kernels/topk_op_gpu_uint32.cu.cc b/tensorflow/core/kernels/topk_op_gpu_uint32.cu.cc new file mode 100644 index 00000000000..16e2e0e9420 --- /dev/null +++ b/tensorflow/core/kernels/topk_op_gpu_uint32.cu.cc @@ -0,0 +1,28 @@ +/* Copyright 2020 The TensorFlow Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. +==============================================================================*/ + +#if GOOGLE_CUDA +#define EIGEN_USE_GPU + +#include "tensorflow/core/kernels/topk_op.h" +#include "tensorflow/core/kernels/topk_op_gpu.h" + +namespace tensorflow { +using Eigen::GpuDevice; + +template struct functor::TopKFunctor; +} // namespace tensorflow + +#endif // GOOGLE_CUDA diff --git a/tensorflow/core/kernels/topk_op_gpu_uint64.cu.cc b/tensorflow/core/kernels/topk_op_gpu_uint64.cu.cc new file mode 100644 index 00000000000..895247a63a2 --- /dev/null +++ b/tensorflow/core/kernels/topk_op_gpu_uint64.cu.cc @@ -0,0 +1,28 @@ +/* Copyright 2020 The TensorFlow Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. +==============================================================================*/ + +#if GOOGLE_CUDA +#define EIGEN_USE_GPU + +#include "tensorflow/core/kernels/topk_op.h" +#include "tensorflow/core/kernels/topk_op_gpu.h" + +namespace tensorflow { +using Eigen::GpuDevice; + +template struct functor::TopKFunctor; +} // namespace tensorflow + +#endif // GOOGLE_CUDA diff --git a/tensorflow/core/util/batch_util.cc b/tensorflow/core/util/batch_util.cc index b88c365ced0..e03188b04da 100644 --- a/tensorflow/core/util/batch_util.cc +++ b/tensorflow/core/util/batch_util.cc @@ -182,8 +182,6 @@ Status CopyElementToSlice(Tensor element, Tensor* parent, int64 index) { switch (element.dtype()) { TF_CALL_ALL_TYPES(HANDLE_TYPE); TF_CALL_QUANTIZED_TYPES(HANDLE_TYPE); - TF_CALL_uint32(HANDLE_TYPE); - TF_CALL_uint64(HANDLE_TYPE); #undef HANDLE_TYPE default: return errors::Unimplemented("CopyElementToSlice Unhandled data type: ", @@ -207,8 +205,6 @@ Status CopySliceToElement(const Tensor& parent, Tensor* element, int64 index) { switch (parent.dtype()) { TF_CALL_ALL_TYPES(HANDLE_TYPE); TF_CALL_QUANTIZED_TYPES(HANDLE_TYPE); - TF_CALL_uint32(HANDLE_TYPE); - TF_CALL_uint64(HANDLE_TYPE); #undef HANDLE_TYPE default: return errors::Unimplemented("CopySliceToElement Unhandled data type: ", @@ -280,8 +276,6 @@ Status CopyContiguousSlices(const Tensor& src, int64 src_offset, switch (src.dtype()) { TF_CALL_ALL_TYPES(HANDLE_TYPE); TF_CALL_QUANTIZED_TYPES(HANDLE_TYPE); - TF_CALL_uint32(HANDLE_TYPE); - TF_CALL_uint64(HANDLE_TYPE); #undef HANDLE_TYPE default: return errors::Unimplemented("CopyContiguousSlices unhandled data type: ", @@ -308,8 +302,6 @@ Status MaybeMoveSliceToElement(Tensor* parent, Tensor* element, int64 index) { switch (parent->dtype()) { TF_CALL_ALL_TYPES(HANDLE_TYPE); TF_CALL_QUANTIZED_TYPES(HANDLE_TYPE); - TF_CALL_uint32(HANDLE_TYPE); - TF_CALL_uint64(HANDLE_TYPE); #undef HANDLE_TYPE default: return errors::Unimplemented( diff --git a/tensorflow/core/util/saved_tensor_slice_util.h b/tensorflow/core/util/saved_tensor_slice_util.h index 09b9235b711..1f9768f5163 100644 --- a/tensorflow/core/util/saved_tensor_slice_util.h +++ b/tensorflow/core/util/saved_tensor_slice_util.h @@ -116,7 +116,9 @@ TENSOR_PROTO_EXTRACT_TYPE(double, double, double); TENSOR_PROTO_EXTRACT_TYPE_COMPLEX(complex64, scomplex, float); TENSOR_PROTO_EXTRACT_TYPE_COMPLEX(complex128, dcomplex, double); TENSOR_PROTO_EXTRACT_TYPE(int32, int, int32); +TENSOR_PROTO_EXTRACT_TYPE(uint32, uint32, uint32); TENSOR_PROTO_EXTRACT_TYPE(int64, int64, protobuf_int64); +TENSOR_PROTO_EXTRACT_TYPE(uint64, uint64, protobuf_uint64); TENSOR_PROTO_EXTRACT_TYPE(uint16, int, int32); TENSOR_PROTO_EXTRACT_TYPE(uint8, int, int32); TENSOR_PROTO_EXTRACT_TYPE(int8, int, int32); From e5023a1738cce7efcdf9d87863b85c80ab2f8c9e Mon Sep 17 00:00:00 2001 From: Karim Nosir Date: Mon, 22 Jun 2020 23:10:17 -0700 Subject: [PATCH 532/867] - Fix type checking in elementwise.cc - Update error messages for some Abs Cast Ceil Cos Sin Not Square Sqrt RSqrt Log PiperOrigin-RevId: 317807251 Change-Id: I2a4f359f04346551eda5a382b25f34bab2c73dc7 --- tensorflow/lite/kernels/cast.cc | 24 ++++++++------- tensorflow/lite/kernels/ceil.cc | 4 +++ tensorflow/lite/kernels/elementwise.cc | 42 ++++++++++++++++++-------- tensorflow/lite/kernels/op_macros.h | 9 ++++++ 4 files changed, 55 insertions(+), 24 deletions(-) diff --git a/tensorflow/lite/kernels/cast.cc b/tensorflow/lite/kernels/cast.cc index 415f1270328..ab95afa979f 100644 --- a/tensorflow/lite/kernels/cast.cc +++ b/tensorflow/lite/kernels/cast.cc @@ -20,6 +20,7 @@ limitations under the License. #include "tensorflow/lite/kernels/internal/tensor.h" #include "tensorflow/lite/kernels/internal/tensor_ctypes.h" #include "tensorflow/lite/kernels/kernel_util.h" +#include "tensorflow/lite/kernels/op_macros.h" namespace tflite { namespace ops { @@ -67,8 +68,8 @@ void copyCast(const std::complex* in, std::complex* out, } template -TfLiteStatus copyToTensor(const FromT* in, TfLiteTensor* out, - int num_elements) { +TfLiteStatus copyToTensor(TfLiteContext* context, const FromT* in, + TfLiteTensor* out, int num_elements) { switch (out->type) { case kTfLiteInt64: copyCast(in, out->data.i64, num_elements); @@ -91,7 +92,7 @@ TfLiteStatus copyToTensor(const FromT* in, TfLiteTensor* out, break; default: // Unsupported type. - return kTfLiteError; + TF_LITE_UNSUPPORTED_TYPE(context, out->type, "Cast"); } return kTfLiteOk; } @@ -103,22 +104,23 @@ TfLiteStatus Eval(TfLiteContext* context, TfLiteNode* node) { TF_LITE_ENSURE_EQ(context, num_elements, NumElements(output)); switch (input->type) { case kTfLiteInt64: - return copyToTensor(input->data.i64, output, num_elements); + return copyToTensor(context, input->data.i64, output, num_elements); case kTfLiteInt32: - return copyToTensor(input->data.i32, output, num_elements); + return copyToTensor(context, input->data.i32, output, num_elements); case kTfLiteUInt8: - return copyToTensor(input->data.uint8, output, num_elements); + return copyToTensor(context, input->data.uint8, output, num_elements); case kTfLiteFloat32: - return copyToTensor(GetTensorData(input), output, num_elements); + return copyToTensor(context, GetTensorData(input), output, + num_elements); case kTfLiteBool: - return copyToTensor(input->data.b, output, num_elements); + return copyToTensor(context, input->data.b, output, num_elements); case kTfLiteComplex64: return copyToTensor( - reinterpret_cast*>(input->data.c64), output, - num_elements); + context, reinterpret_cast*>(input->data.c64), + output, num_elements); default: // Unsupported type. - return kTfLiteError; + TF_LITE_UNSUPPORTED_TYPE(context, input->type, "Cast"); } return kTfLiteOk; } diff --git a/tensorflow/lite/kernels/ceil.cc b/tensorflow/lite/kernels/ceil.cc index d8c6eaad7a4..95c660f3376 100644 --- a/tensorflow/lite/kernels/ceil.cc +++ b/tensorflow/lite/kernels/ceil.cc @@ -18,6 +18,7 @@ limitations under the License. #include "tensorflow/lite/kernels/internal/tensor.h" #include "tensorflow/lite/kernels/internal/tensor_ctypes.h" #include "tensorflow/lite/kernels/kernel_util.h" +#include "tensorflow/lite/kernels/op_macros.h" namespace tflite { namespace ops { @@ -41,6 +42,9 @@ TfLiteStatus Prepare(TfLiteContext* context, TfLiteNode* node) { TfLiteStatus Eval(TfLiteContext* context, TfLiteNode* node) { const TfLiteTensor* input = GetInput(context, node, kInputTensor); TfLiteTensor* output = GetOutput(context, node, kOutputTensor); + if (input->type != kTfLiteFloat32) { + TF_LITE_UNSUPPORTED_TYPE(context, input->type, "Ceil"); + } optimized_ops::Ceil(GetTensorShape(input), GetTensorData(input), GetTensorShape(output), GetTensorData(output)); diff --git a/tensorflow/lite/kernels/elementwise.cc b/tensorflow/lite/kernels/elementwise.cc index 1b91244af33..61c6aeaa811 100644 --- a/tensorflow/lite/kernels/elementwise.cc +++ b/tensorflow/lite/kernels/elementwise.cc @@ -23,6 +23,7 @@ limitations under the License. #include "tensorflow/lite/kernels/internal/tensor.h" #include "tensorflow/lite/kernels/internal/tensor_ctypes.h" #include "tensorflow/lite/kernels/kernel_util.h" +#include "tensorflow/lite/kernels/op_macros.h" namespace tflite { namespace ops { @@ -39,17 +40,15 @@ bool IsLogicalSupportedType(const TfLiteType type) { } typedef bool (*IsSupportedType)(TfLiteType); -template +template TfLiteStatus GenericPrepare(TfLiteContext* context, TfLiteNode* node) { TF_LITE_ENSURE_EQ(context, NumInputs(node), 1); TF_LITE_ENSURE_EQ(context, NumOutputs(node), 1); const TfLiteTensor* input = GetInput(context, node, 0); TfLiteTensor* output = GetOutput(context, node, 0); TF_LITE_ENSURE_TYPES_EQ(context, input->type, output->type); - if (!IsSupportedType(input->type)) { - context->ReportError(context, "Current data type %d is not supported.", - input->type); - return kTfLiteError; + if (!is_supported_type(input->type)) { + TF_LITE_UNSUPPORTED_TYPE(context, input->type, op_name); } return context->ResizeTensor(context, output, TfLiteIntArrayCopy(input->dims)); @@ -112,13 +111,23 @@ TfLiteStatus LogicalNotEval(TfLiteContext* context, TfLiteNode* node) { return EvalLogical(context, node, [](bool v) { return !v; }); } +constexpr char kAbsName[] = "Abs"; +constexpr char kSinName[] = "Sin"; +constexpr char kCosName[] = "Cos"; +constexpr char kLogName[] = "Log"; +constexpr char kSqrtName[] = "Sqrt"; +constexpr char kRsqrtName[] = "Rsqrt"; +constexpr char kSquareName[] = "Square"; +constexpr char kNotName[] = "Not"; + } // namespace } // namespace elementwise TfLiteRegistration* Register_ABS() { static TfLiteRegistration r = { /*init=*/nullptr, /*free=*/nullptr, - elementwise::GenericPrepare, + elementwise::GenericPrepare, elementwise::AbsEval}; return &r; } @@ -126,7 +135,8 @@ TfLiteRegistration* Register_ABS() { TfLiteRegistration* Register_SIN() { static TfLiteRegistration r = { /*init=*/nullptr, /*free=*/nullptr, - elementwise::GenericPrepare, + elementwise::GenericPrepare, elementwise::SinEval}; return &r; } @@ -134,7 +144,8 @@ TfLiteRegistration* Register_SIN() { TfLiteRegistration* Register_COS() { static TfLiteRegistration r = { /*init=*/nullptr, /*free=*/nullptr, - elementwise::GenericPrepare, + elementwise::GenericPrepare, elementwise::CosEval}; return &r; } @@ -142,7 +153,8 @@ TfLiteRegistration* Register_COS() { TfLiteRegistration* Register_LOG() { static TfLiteRegistration r = { /*init=*/nullptr, /*free=*/nullptr, - elementwise::GenericPrepare, + elementwise::GenericPrepare, elementwise::LogEval}; return &r; } @@ -150,7 +162,8 @@ TfLiteRegistration* Register_LOG() { TfLiteRegistration* Register_SQRT() { static TfLiteRegistration r = { /*init=*/nullptr, /*free=*/nullptr, - elementwise::GenericPrepare, + elementwise::GenericPrepare, elementwise::SqrtEval}; return &r; } @@ -158,7 +171,8 @@ TfLiteRegistration* Register_SQRT() { TfLiteRegistration* Register_RSQRT() { static TfLiteRegistration r = { /*init=*/nullptr, /*free=*/nullptr, - elementwise::GenericPrepare, + elementwise::GenericPrepare, elementwise::RsqrtEval}; return &r; } @@ -166,7 +180,8 @@ TfLiteRegistration* Register_RSQRT() { TfLiteRegistration* Register_SQUARE() { static TfLiteRegistration r = { /*init=*/nullptr, /*free=*/nullptr, - elementwise::GenericPrepare, + elementwise::GenericPrepare, elementwise::SquareEval}; return &r; } @@ -174,7 +189,8 @@ TfLiteRegistration* Register_SQUARE() { TfLiteRegistration* Register_LOGICAL_NOT() { static TfLiteRegistration r = { /*init=*/nullptr, /*free=*/nullptr, - elementwise::GenericPrepare, + elementwise::GenericPrepare, elementwise::LogicalNotEval}; return &r; } diff --git a/tensorflow/lite/kernels/op_macros.h b/tensorflow/lite/kernels/op_macros.h index 5c190f1c595..5786756f408 100644 --- a/tensorflow/lite/kernels/op_macros.h +++ b/tensorflow/lite/kernels/op_macros.h @@ -44,6 +44,15 @@ inline void InfiniteLoop() { fprintf(stderr, "%s", (x)); \ } while (0) +// Report Error for unsupported type by op 'op_name' and returns kTfLiteError. +#define TF_LITE_UNSUPPORTED_TYPE(context, type, op_name) \ + do { \ + TF_LITE_KERNEL_LOG((context), "%s:%d Type %s is unsupported by op %s.", \ + __FILE__, __LINE__, TfLiteTypeGetName(type), \ + (op_name)); \ + return kTfLiteError; \ + } while (0) + #define TFLITE_ABORT abort() #endif // TF_LITE_MCU_DEBUG_LOG From daa3c52aa1c301fc698574a2943f2a6eed2de08f Mon Sep 17 00:00:00 2001 From: =?UTF-8?q?M=C3=A5ns=20Nilsson?= Date: Fri, 13 Mar 2020 17:05:54 +0100 Subject: [PATCH 533/867] TFL: Port HARD_SWISH operator from TFLite to TFLu Change-Id: I18092fa30dee33df4577a4abaf5321e7ba96a04c --- tensorflow/lite/kernels/activations.cc | 16 - tensorflow/lite/kernels/internal/common.h | 47 +++ .../internal/reference/reference_ops.h | 19 - tensorflow/lite/micro/all_ops_resolver.cc | 1 + tensorflow/lite/micro/kernels/activations.cc | 193 +++++++++- .../lite/micro/kernels/activations_test.cc | 354 ++++++++++++++++++ tensorflow/lite/micro/kernels/micro_ops.h | 1 + .../lite/micro/micro_mutable_op_resolver.h | 8 + tensorflow/lite/micro/micro_utils.h | 10 + 9 files changed, 612 insertions(+), 37 deletions(-) diff --git a/tensorflow/lite/kernels/activations.cc b/tensorflow/lite/kernels/activations.cc index 7ad33973b38..b6181973c7d 100644 --- a/tensorflow/lite/kernels/activations.cc +++ b/tensorflow/lite/kernels/activations.cc @@ -298,22 +298,6 @@ void HardSwishFree(TfLiteContext* context, void* buffer) { delete static_cast(buffer); } -void DownScaleInt32ToInt16Multiplier(int32_t multiplier_int32, - int16_t* multiplier_int16) { - TFLITE_DCHECK_GE(multiplier_int32, 0); - static constexpr int32_t kRoundingOffset = 1 << 15; - if (multiplier_int32 >= - std::numeric_limits::max() - kRoundingOffset) { - *multiplier_int16 = std::numeric_limits::max(); - return; - } - const int32_t result = (multiplier_int32 + kRoundingOffset) >> 16; - TFLITE_DCHECK_LE(result << 16, multiplier_int32 + kRoundingOffset); - TFLITE_DCHECK_GT(result << 16, multiplier_int32 - kRoundingOffset); - *multiplier_int16 = result; - TFLITE_DCHECK_EQ(*multiplier_int16, result); -} - TfLiteStatus HardSwishPrepare(TfLiteContext* context, TfLiteNode* node) { TF_LITE_ENSURE_STATUS(GenericPrepare(context, node)); TfLiteTensor* output = GetOutput(context, node, 0); diff --git a/tensorflow/lite/kernels/internal/common.h b/tensorflow/lite/kernels/internal/common.h index c1db3587415..9bc65029f34 100644 --- a/tensorflow/lite/kernels/internal/common.h +++ b/tensorflow/lite/kernels/internal/common.h @@ -932,6 +932,53 @@ void optimized_ops_prefetch_write_l1_keep(const T* ptr) { #endif } +// Similar to ARM instruction SQDMULH. +// Similar to gemmlowp::SaturatingRoundingDoublingHighMul except +// rounding to zero instead of to nearest (SQRDMULH). +inline std::int16_t SaturatingDoublingHighMul(std::int16_t a, std::int16_t b) { + bool overflow = a == b && a == std::numeric_limits::min(); + std::int32_t a_32(a); + std::int32_t b_32(b); + std::int32_t ab_32 = a_32 * b_32; + std::int16_t ab_x2_high16 = static_cast((ab_32) / (1 << 15)); + return overflow ? std::numeric_limits::max() : ab_x2_high16; +} + +// Similar to gemmlowp::SaturatingRoundingDoublingHighMul +inline std::int16_t SaturatingRoundingDoublingHighMul(std::int16_t a, std::int16_t b) { + bool overflow = a == b && a == std::numeric_limits::min(); + std::int32_t a_32(a); + std::int32_t b_32(b); + std::int32_t ab_32 = a_32 * b_32; + std::int16_t nudge = ab_32 >= 0 ? (1 << 14) : (1 - (1 << 14)); + std::int16_t ab_x2_high16 = + static_cast((ab_32 + nudge) / (1 << 15)); + return overflow ? std::numeric_limits::max() : ab_x2_high16; +} + +inline int16_t SaturatingLeftShift(int16_t value, int amount) { + int32_t result = static_cast(value) * (1 << amount); + result = std::min(result, std::numeric_limits::max()); + result = std::max(result, std::numeric_limits::min()); + return result; +} + +inline void DownScaleInt32ToInt16Multiplier(int32_t multiplier_int32, + int16_t* multiplier_int16) { + TFLITE_DCHECK_GE(multiplier_int32, 0); + static constexpr int32_t kRoundingOffset = 1 << 15; + if (multiplier_int32 >= + std::numeric_limits::max() - kRoundingOffset) { + *multiplier_int16 = std::numeric_limits::max(); + return; + } + const int32_t result = (multiplier_int32 + kRoundingOffset) >> 16; + TFLITE_DCHECK_LE(result << 16, multiplier_int32 + kRoundingOffset); + TFLITE_DCHECK_GT(result << 16, multiplier_int32 - kRoundingOffset); + *multiplier_int16 = result; + TFLITE_DCHECK_EQ(*multiplier_int16, result); +} + } // namespace tflite #endif // TENSORFLOW_LITE_KERNELS_INTERNAL_COMMON_H_ diff --git a/tensorflow/lite/kernels/internal/reference/reference_ops.h b/tensorflow/lite/kernels/internal/reference/reference_ops.h index 5208b21eb4d..b94c2060b2d 100644 --- a/tensorflow/lite/kernels/internal/reference/reference_ops.h +++ b/tensorflow/lite/kernels/internal/reference/reference_ops.h @@ -2615,25 +2615,6 @@ inline void HardSwish(const RuntimeShape& input_shape, const T* input_data, } } -inline int16_t SaturatingLeftShift(int16_t value, int amount) { - int32_t result = static_cast(value) * (1 << amount); - result = std::min(result, std::numeric_limits::max()); - result = std::max(result, std::numeric_limits::min()); - return result; -} - -// Similar to ARM instruction SQDMULH. -// Similar to gemmlowp::SaturatingRoundingDoublingHighMul except -// rounding to zero instead of to nearest (SQRDMULH). -inline std::int16_t SaturatingDoublingHighMul(std::int16_t a, std::int16_t b) { - bool overflow = a == b && a == std::numeric_limits::min(); - std::int32_t a_32(a); - std::int32_t b_32(b); - std::int32_t ab_32 = a_32 * b_32; - std::int16_t ab_x2_high16 = static_cast((ab_32) / (1 << 15)); - return overflow ? std::numeric_limits::max() : ab_x2_high16; -} - template inline void HardSwish(const HardSwishParams& params, const RuntimeShape& input_shape, const T* input_data, diff --git a/tensorflow/lite/micro/all_ops_resolver.cc b/tensorflow/lite/micro/all_ops_resolver.cc index e728a95360a..ff461cb947e 100644 --- a/tensorflow/lite/micro/all_ops_resolver.cc +++ b/tensorflow/lite/micro/all_ops_resolver.cc @@ -42,6 +42,7 @@ AllOpsResolver::AllOpsResolver() { AddFullyConnected(); AddGreater(); AddGreaterEqual(); + AddHardSwish(); AddL2Normalization(); AddLess(); AddLessEqual(); diff --git a/tensorflow/lite/micro/kernels/activations.cc b/tensorflow/lite/micro/kernels/activations.cc index 4a9b8ce5d8e..75b0339002d 100644 --- a/tensorflow/lite/micro/kernels/activations.cc +++ b/tensorflow/lite/micro/kernels/activations.cc @@ -18,6 +18,7 @@ limitations under the License. #include "tensorflow/lite/kernels/internal/common.h" #include "tensorflow/lite/kernels/internal/quantization_util.h" #include "tensorflow/lite/kernels/internal/tensor_ctypes.h" +#include "tensorflow/lite/kernels/internal/types.h" #include "tensorflow/lite/kernels/kernel_util.h" #include "tensorflow/lite/kernels/op_macros.h" #include "tensorflow/lite/micro/micro_utils.h" @@ -77,6 +78,157 @@ inline void Relu6Quantized(Q lower, Q upper, const RuntimeShape& input_shape, } } +inline std::int32_t RoundingDivideByPOT(std::int32_t numerator, int exponent) { + std::int32_t sign = numerator >= 0 ? 1 : -1; + std::int32_t abs_numerator = std::abs(numerator); + std::int32_t mask = (1LL << exponent) - 1; + std::int32_t remainder = abs_numerator & mask; + std::int32_t threshold = mask >> 1; + std::int32_t abs_result = + (abs_numerator >> exponent) + (remainder > threshold ? 1 : 0); + return sign * abs_result; +} + +inline void HardSwishFloatOp(const RuntimeShape& input_shape, const float* input_data, + const RuntimeShape& output_shape, float* output_data) { + auto matching_size = MatchingFlatSize(input_shape, output_shape); + const float* in_end = input_data + matching_size; + for (; input_data < in_end; input_data++, output_data++) { + const float in = *input_data; + *output_data = + in * std::min(static_cast(6), std::max(static_cast(0), in + 3)) / + 6; + } +} + +template +void HardSwishOp(HardSwishParams& params, + const RuntimeShape& input_shape, const T* input_data, + const RuntimeShape& output_shape, T* output_data) { + const int flat_size = MatchingFlatSize(input_shape, output_shape); + + for (int i = 0; i < flat_size; i++) { + const int16_t input_value = input_data[i] - params.input_zero_point; + // Left-shift as much as we can without overflow/saturation to put + // significant bits in the high bits of our 16-bit fixedpoint values, so + // that fixed-point approximate computations below are as accurate as + // possible. + const int16_t input_value_on_hires_input_scale = input_value << 7; + // Compute the input value on essentially the output scale, just not + // right-shifted yet. This is the value that we'll use in the (x >= +3) + // case, and that in the general case we'll multiply against the "relu-ish" + // fixed-point multiplier in [0, 1]. + const int16_t input_value_on_preshift_output_scale = + SaturatingRoundingDoublingHighMul(input_value_on_hires_input_scale, + params.output_multiplier_fixedpoint_int16); + // Now compute the "relu-ish multiplier". In the (-3 <= x <= +3) case, that + // is just an affine rescaling of x from [-3, 3] to [0, 1]. In the general + // case, it is just that plus saturation at the boundaries of [-3, 3]. + // First, we rescale from [-3, 3] to [-1, 1], saturating. + // That is done by rescaling the input value with a fixed-point multiplier + // (reluish_multiplier_fixedpoint) and bit-shift such that we represent + // that input value on the scale where the real value 3.0f is represented + // by the quantized value 32768. (+32768 is actually not representable as + // int16, so this saturates at +32767, and that is seen empirically to be + // a negligible contribution to numerical error/bias). + // + // This code is careful to correctly implement any magnitude of multiplier, + // involving either a right shift or a left shift, with correct saturation + // behavior in the left-shift case. This forces this code to be more + // complicated, but is necessary for real applications: a partially + // trained quantized MobileNet v3-small model that motivated this code + // exhibits some large [min, max] range boundaries, of the order of + // magnitude of 10 or 100 depending on layers. + // + // The next few lines are basically just an ordinary + // MultiplyByQuantizedMultiplier, except that we are more careful here + // about the fine details of saturation when left-shifting, because here + // overflow in left-shift is a common case, not an anomaly as + // MultiplyByQuantizedMultiplier assumes. + int16_t reluish_value = input_value_on_hires_input_scale; + // Shift left, saturating, as much as we can while ensuring that this + // saturation will not contribute to the result. That is, left shift amount + // reduced by 1. + if (params.reluish_multiplier_exponent > 0) { + reluish_value = SaturatingLeftShift( + reluish_value, params.reluish_multiplier_exponent - 1); + } + // Apply the fixed-point multiplier, dividing the value by a divisor + // ranging in [1, 2]. + reluish_value = SaturatingRoundingDoublingHighMul(reluish_value, params.reluish_multiplier_fixedpoint_int16); + // Apply the last bit of left-shift. Thus, in the left-shifting case, if + // any saturation affects the result, it is happening here --- any + // saturation having occurred above is overwritten here, not affecting the + // result. + if (params.reluish_multiplier_exponent > 0) { + reluish_value = SaturatingLeftShift(reluish_value, 1); + } + // Shift right, in the right-shifting case. + if (params.reluish_multiplier_exponent < 0) { + reluish_value = RoundingDivideByPOT( + reluish_value, -params.reluish_multiplier_exponent); + } + // At this point we have rescaled the value into a 16bit fixedpoint + // reluish_value in [-1, 1]. + // We now convert that to a 16bit fixedpoint value in [0, 1]. + reluish_value = (reluish_value + (1 << 15)) >> 1; + // Use of SaturatingDoublingHighMul here is important to cancel the biases + // from the above SaturatingRoundingDoublingHighMul. + // + const int16_t preshift_output_value = SaturatingDoublingHighMul( + reluish_value, input_value_on_preshift_output_scale); + // We were so far operating on the pre-shift output scale. Now we finally + // apply that output shift, arriving at the final output scale. + int16_t output_value = RoundingDivideByPOT( + preshift_output_value, -params.output_multiplier_exponent); + output_value += params.output_zero_point; + output_value = + std::min(output_value, std::numeric_limits::max()); + output_value = + std::max(output_value, std::numeric_limits::min()); + output_data[i] = output_value; + } +} + + +template +TfLiteStatus HardSwishQuantized(TfLiteContext* context, TfLiteNode* node) { + const TfLiteTensor* input = GetInput(context, node, kInputTensor); + TfLiteTensor* output = GetOutput(context, node, kOutputTensor); + + HardSwishParams params; + + params.input_zero_point = input->params.zero_point; + params.output_zero_point = output->params.zero_point; + + const float input_scale = input->params.scale; + const float hires_input_scale = (1.0f / 128.0f) * input_scale; + const float reluish_scale = 3.0f / 32768.0f; + const float output_scale = output->params.scale; + + const double output_multiplier = static_cast(hires_input_scale / output_scale); + int32_t output_multiplier_fixedpoint_int32; + QuantizeMultiplier(output_multiplier, &output_multiplier_fixedpoint_int32, + ¶ms.output_multiplier_exponent); + DownScaleInt32ToInt16Multiplier( + output_multiplier_fixedpoint_int32, + ¶ms.output_multiplier_fixedpoint_int16); + + TF_LITE_ENSURE(context, params.output_multiplier_exponent <= 0); + + const double reluish_multiplier = static_cast(hires_input_scale / reluish_scale); + int32_t reluish_multiplier_fixedpoint_int32; + QuantizeMultiplier(reluish_multiplier, &reluish_multiplier_fixedpoint_int32, + ¶ms.reluish_multiplier_exponent); + DownScaleInt32ToInt16Multiplier( + reluish_multiplier_fixedpoint_int32, + ¶ms.reluish_multiplier_fixedpoint_int16); + + HardSwishOp(params, GetTensorShape(input), + GetTensorData(input), GetTensorShape(output), GetTensorData(output)); + return kTfLiteOk; +} + TfLiteStatus ReluPrepare(TfLiteContext* context, TfLiteNode* node) { return kTfLiteOk; } @@ -107,7 +259,7 @@ TfLiteStatus ReluEval(TfLiteContext* context, TfLiteNode* node) { return kTfLiteOk; } default: { - TF_LITE_KERNEL_LOG(context, "Only float32 is supported currently, got %s", + TF_LITE_KERNEL_LOG(context, "Only float32/int8/uint8 are supported currently, got %s", TfLiteTypeGetName(input->type)); return kTfLiteError; } @@ -148,13 +300,43 @@ TfLiteStatus Relu6Eval(TfLiteContext* context, TfLiteNode* node) { return kTfLiteOk; } default: { - TF_LITE_KERNEL_LOG(context, "Only float32 is supported currently, got %s", + TF_LITE_KERNEL_LOG(context, "Only float32/int8/uint8 are supported currently, got %s", TfLiteTypeGetName(input->type)); return kTfLiteError; } } } +TfLiteStatus HardSwishPrepare(TfLiteContext* context, TfLiteNode* node) { + return kTfLiteOk; +} + +TfLiteStatus HardSwishEval(TfLiteContext* context, TfLiteNode* node) { + const TfLiteTensor* input = GetInput(context, node, kInputTensor); + TfLiteTensor* output = GetOutput(context, node, kOutputTensor); + + switch (input->type) { + case kTfLiteFloat32: { + HardSwishFloatOp( + GetTensorShape(input), + GetTensorData(input), + GetTensorShape(output), + GetTensorData(output)); + return kTfLiteOk; + } break; + case kTfLiteUInt8: { + return HardSwishQuantized(context, node); + } break; + case kTfLiteInt8: { + return HardSwishQuantized(context, node); + } break; + default: + TF_LITE_KERNEL_LOG(context, "Only float32/int8/uint8 are supported currently, got %s", + TfLiteTypeGetName(input->type)); + return kTfLiteError; + } +} + } // namespace activations TfLiteRegistration* Register_RELU() { @@ -181,6 +363,13 @@ TfLiteRegistration* Register_RELU6() { return &r; } +TfLiteRegistration* Register_HARD_SWISH() { + static TfLiteRegistration r = {}; + r.prepare = activations::HardSwishPrepare; + r.invoke = activations::HardSwishEval; + return &r; +} + } // namespace micro } // namespace ops } // namespace tflite diff --git a/tensorflow/lite/micro/kernels/activations_test.cc b/tensorflow/lite/micro/kernels/activations_test.cc index 221f8f66d58..0a7db1e6589 100644 --- a/tensorflow/lite/micro/kernels/activations_test.cc +++ b/tensorflow/lite/micro/kernels/activations_test.cc @@ -13,6 +13,8 @@ See the License for the specific language governing permissions and limitations under the License. ==============================================================================*/ +#include + #include "tensorflow/lite/c/builtin_op_data.h" #include "tensorflow/lite/c/common.h" #include "tensorflow/lite/micro/all_ops_resolver.h" @@ -135,6 +137,318 @@ void TestRelu6Float(const int* input_dims_data, const float* input_data, } } +void GenerateUniformRandomVector(int size, float min, float max, + std::minstd_rand* random_engine, + std::vector* result) { + // Never use std::uniform_*_distribution in tests, it's + // implementation-defined. Likewise, don't use std::default_random_engine, + // implementation-defined. Implementation-defined is bad because it means that + // any toolchain update or new platform may run into test failures. + // std::minstd_rand is a standard instantiation of + // std::linear_congruential_engine, the cheapest generator in c++11 stdlib, + // it's good enough here. + result->resize(size); + for (int i = 0; i < size; i++) { + // We don't care whether the `max` value may ever be produced exactly. + // It may actually be thanks to rounding, as std::minstd_rand::modulus + // is 2^31 - 1 is greater than the inverse float epsilon. + float random_value_scaled_0_1 = + (*random_engine)() * + (1.0f / static_cast(std::minstd_rand::modulus)); + (*result)[i] = min + (max - min) * random_value_scaled_0_1; + } +} + +void EvalTestReferenceHardSwish(int size, const std::vector& input, + std::vector* result) { + result->resize(size); + for (int i = 0; i < size; i++) { + const float in = input[i]; + (*result)[i] = in * std::min(6.0f, std::max(0.0f, in + 3)) * (1.0f / 6.0f); + } +} + +template +void TestHardSwishQuantized(int size, float input_min, + float input_max, float output_min, + float output_max, std::minstd_rand* random_engine) { + T output_data[size]; + T input_data_quantized[size]; + const int input_dims_data[] = {2, 1, size}; + const int output_dims_data[] = {2, 1, size}; + const float input_scale = ScaleFromMinMax(input_min, input_max); + const int input_zero_point = ZeroPointFromMinMax(input_min, input_max); + const float output_scale = ScaleFromMinMax(output_min, output_max); + const int output_zero_point = ZeroPointFromMinMax(output_min, output_max); + + // The numerical error for any 8bit quantized function is at least one half + // times the quantization step: 0.5 * (kOutMax - kOutMin) / 256. + // To that we add again the quantization step (kOutMax - kOutMin) / 256 + // to allow for an off-by-one rounding error. + const float kTolerance = std::max(input_max - input_min, output_max - output_min) * (1.5f / 256.f); + + TfLiteIntArray* input_dims = IntArrayFromInts(input_dims_data); + TfLiteIntArray* output_dims = IntArrayFromInts(output_dims_data); + const int output_elements_count = ElementCount(*output_dims); + + TF_LITE_MICRO_EXPECT_EQ(output_elements_count, size); + + float dequantized_output[output_elements_count]; + + std::vector float_input_values; + std::vector float_ref_output_values; + GenerateUniformRandomVector(size, input_min, input_max, random_engine, + &float_input_values); + EvalTestReferenceHardSwish(size, float_input_values, + &float_ref_output_values); + for (float& val : float_ref_output_values) { + val = std::min(output_max, std::max(output_min, val)); + } + + constexpr int inputs_size = 1; + constexpr int outputs_size = 1; + constexpr int tensors_size = inputs_size + outputs_size; + TfLiteTensor tensors[tensors_size] = { + CreateQuantizedTensor(float_input_values.data(), input_data_quantized, input_dims, + input_scale, input_zero_point, "input_tensor"), + CreateQuantizedTensor(output_data, output_dims, output_scale, + output_zero_point, "output_tensor"), + }; + + TfLiteContext context; + PopulateContext(tensors, tensors_size, micro_test::reporter, &context); + + ::tflite::AllOpsResolver resolver; + const TfLiteRegistration* registration = + resolver.FindOp(tflite::BuiltinOperator_HARD_SWISH); + TF_LITE_MICRO_EXPECT_NE(nullptr, registration); + + const char* init_data = nullptr; + size_t init_data_size = 0; + void* user_data = nullptr; + if (registration->init) { + user_data = registration->init(&context, init_data, init_data_size); + } + int inputs_array_data[] = {1, 0}; + TfLiteIntArray* inputs_array = IntArrayFromInts(inputs_array_data); + int outputs_array_data[] = {1, 1}; + TfLiteIntArray* outputs_array = IntArrayFromInts(outputs_array_data); + + TfLiteNode node; + node.inputs = inputs_array; + node.outputs = outputs_array; + node.temporaries = nullptr; + node.user_data = user_data; + node.builtin_data = nullptr; + node.custom_initial_data = nullptr; + node.custom_initial_data_size = 0; + node.delegate = nullptr; + if (registration->prepare) { + TF_LITE_MICRO_EXPECT_EQ(kTfLiteOk, registration->prepare(&context, &node)); + } + + TF_LITE_MICRO_EXPECT_NE(nullptr, registration->invoke); + TF_LITE_MICRO_EXPECT_EQ(kTfLiteOk, registration->invoke(&context, &node)); + + if (registration->free) { + registration->free(&context, user_data); + } + + AsymmetricDequantize(output_data, output_elements_count, output_scale, output_zero_point, dequantized_output); + + for (int i = 0; i < output_elements_count; ++i) { + TF_LITE_MICRO_EXPECT_NEAR(float_ref_output_values.data()[i], dequantized_output[i], kTolerance); + } +} + +template +void TestHardSwishQuantizedBias(float input_min, float input_max, float output_min, + float output_max, float tolerated_bias) { + const float quantized_type_range = + static_cast(std::numeric_limits::max()) - + static_cast(std::numeric_limits::min()); + + + const float input_scale = ScaleFromMinMax(input_min, input_max); + const float output_scale = ScaleFromMinMax(output_min, output_max); + + const int input_zero_point = ZeroPointFromMinMax(input_min, input_max); + const int output_zero_point = ZeroPointFromMinMax(output_min, output_max); + + const float max_scale = std::max(output_scale, input_scale); + + // In this bias-focused test case, no need for randomly generated input + // values. + TF_LITE_MICRO_EXPECT_LE(input_min, -3.0f); + TF_LITE_MICRO_EXPECT_GE(input_max, 3.0f); + const int quantized_input_negative_three = + std::round(std::numeric_limits::min() + + (-3.0f - input_min) / input_scale); + const int quantized_input_positive_three = + std::round(std::numeric_limits::min() + + (3.0f - input_min) / input_scale); + std::vector float_input_values; + for (int i = quantized_input_negative_three; + i <= quantized_input_positive_three; i++) { + float_input_values.push_back( + input_min + + (i - std::numeric_limits::min()) * input_scale); + } + const int size = float_input_values.size(); + std::vector float_ref_output_values; + EvalTestReferenceHardSwish(size, float_input_values, + &float_ref_output_values); + for (float& val : float_ref_output_values) { + val = std::min(output_max, std::max(output_min, val)); + } + + T output_data[size]; + T input_data_quantized[size]; + const int input_dims_data[] = {2, 1, size}; + const int output_dims_data[] = {2, 1, size}; + + // The numerical error for any 8bit quantized function is at least one half + // times the quantization step: 0.5 * (kOutMax - kOutMin) / 256. + // To that we add again the quantization step (kOutMax - kOutMin) / 256 + // to allow for an off-by-one rounding error. + const float kTolerance = std::max(input_max - input_min, output_max - output_min) * (1.5f / 256.f); + + TfLiteIntArray* input_dims = IntArrayFromInts(input_dims_data); + TfLiteIntArray* output_dims = IntArrayFromInts(output_dims_data); + const int output_elements_count = ElementCount(*output_dims); + + TF_LITE_MICRO_EXPECT_EQ(output_elements_count, size); + + float dequantized_output[output_elements_count]; + + constexpr int inputs_size = 1; + constexpr int outputs_size = 1; + constexpr int tensors_size = inputs_size + outputs_size; + TfLiteTensor tensors[tensors_size] = { + CreateQuantizedTensor(float_input_values.data(), input_data_quantized, input_dims, + input_scale, input_zero_point, "input_tensor"), + CreateQuantizedTensor(output_data, output_dims, output_scale, + output_zero_point, "output_tensor"), + }; + + TfLiteContext context; + PopulateContext(tensors, tensors_size, micro_test::reporter, &context); + + ::tflite::AllOpsResolver resolver; + const TfLiteRegistration* registration = + resolver.FindOp(tflite::BuiltinOperator_HARD_SWISH); + TF_LITE_MICRO_EXPECT_NE(nullptr, registration); + + const char* init_data = nullptr; + size_t init_data_size = 0; + void* user_data = nullptr; + if (registration->init) { + user_data = registration->init(&context, init_data, init_data_size); + } + int inputs_array_data[] = {1, 0}; + TfLiteIntArray* inputs_array = IntArrayFromInts(inputs_array_data); + int outputs_array_data[] = {1, 1}; + TfLiteIntArray* outputs_array = IntArrayFromInts(outputs_array_data); + + TfLiteNode node; + node.inputs = inputs_array; + node.outputs = outputs_array; + node.temporaries = nullptr; + node.user_data = user_data; + node.builtin_data = nullptr; + node.custom_initial_data = nullptr; + node.custom_initial_data_size = 0; + node.delegate = nullptr; + if (registration->prepare) { + TF_LITE_MICRO_EXPECT_EQ(kTfLiteOk, registration->prepare(&context, &node)); + } + + TF_LITE_MICRO_EXPECT_NE(nullptr, registration->invoke); + TF_LITE_MICRO_EXPECT_EQ(kTfLiteOk, registration->invoke(&context, &node)); + + if (registration->free) { + registration->free(&context, user_data); + } + + AsymmetricDequantize(output_data, output_elements_count, output_scale, output_zero_point, dequantized_output); + + float sum_diff = 0; + for (int i = 0; i < size; i++) { + sum_diff += dequantized_output[i] - float_ref_output_values[i]; + } + const float bias = sum_diff / (size * max_scale); + TF_LITE_MICRO_EXPECT_LE(std::abs(bias), tolerated_bias); +} + +void TestHardSwishFloat(int size, std::minstd_rand* random_engine) { + std::vector float_input_values; + const float kMin = -10.0f; + const float kMax = 10.0f; + GenerateUniformRandomVector(size, kMin, kMax, random_engine, + &float_input_values); + std::vector float_ref_output_values; + EvalTestReferenceHardSwish(size, float_input_values, + &float_ref_output_values); + + float output_data[size]; + const int input_dims_data[] = {1, size}; + const int output_dims_data[] = {1, size}; + + TfLiteIntArray* input_dims = IntArrayFromInts(input_dims_data); + TfLiteIntArray* output_dims = IntArrayFromInts(output_dims_data); + const int output_elements_count = ElementCount(*output_dims); + + TF_LITE_MICRO_EXPECT_EQ(output_elements_count, size); + + constexpr int inputs_size = 1; + constexpr int outputs_size = 1; + constexpr int tensors_size = inputs_size + outputs_size; + TfLiteTensor tensors[tensors_size] = { + CreateFloatTensor(float_input_values.data(), input_dims, "input_tensor"), + CreateFloatTensor(output_data, output_dims, "output_tensor"), + }; + TfLiteContext context; + PopulateContext(tensors, tensors_size, micro_test::reporter, &context); + + ::tflite::AllOpsResolver resolver; + const TfLiteRegistration* registration = + resolver.FindOp(tflite::BuiltinOperator_HARD_SWISH); + TF_LITE_MICRO_EXPECT_NE(nullptr, registration); + + const char* init_data = nullptr; + size_t init_data_size = 0; + void* user_data = nullptr; + if (registration->init) { + user_data = registration->init(&context, init_data, init_data_size); + } + int inputs_array_data[] = {1, 0}; + TfLiteIntArray* inputs_array = IntArrayFromInts(inputs_array_data); + int outputs_array_data[] = {1, 1}; + TfLiteIntArray* outputs_array = IntArrayFromInts(outputs_array_data); + TfLiteNode node; + node.inputs = inputs_array; + node.outputs = outputs_array; + node.temporaries = nullptr; + node.user_data = user_data; + node.builtin_data = nullptr; + node.custom_initial_data = nullptr; + node.custom_initial_data_size = 0; + node.delegate = nullptr; + if (registration->prepare) { + TF_LITE_MICRO_EXPECT_EQ(kTfLiteOk, registration->prepare(&context, &node)); + } + TF_LITE_MICRO_EXPECT_NE(nullptr, registration->invoke); + TF_LITE_MICRO_EXPECT_EQ(kTfLiteOk, registration->invoke(&context, &node)); + + if (registration->free) { + registration->free(&context, user_data); + } + + for (int i = 0; i < output_elements_count; ++i) { + TF_LITE_MICRO_EXPECT_NEAR(float_ref_output_values.data()[i], output_data[i], 1e-5f); + } +} + void TestReluUint8(const int* input_dims_data, const float* input_data, uint8_t* input_data_quantized, const float input_scale, const int input_zero_point, const float* golden, @@ -431,6 +745,46 @@ TF_LITE_MICRO_TEST(SimpleRelu6TestFloat) { output_data); } +TF_LITE_MICRO_TEST(SimpleHardSwishTestFloat) { + std::minstd_rand random_engine; + for (int size : {1, 2, 3, 4, 10, 20, 30, 40, 100}) { + tflite::testing::TestHardSwishFloat(size, &random_engine); + } +} + +TF_LITE_MICRO_TEST(SimpleHardSwishTestQuantized) { + std::minstd_rand random_engine; + std::vector> minmax_pairs{ + {0.f, 1.f}, {-2.f, 1.f}, {-5.f, 10.f}, {-40.f, 60.f}}; + for (const auto& input_minmax : minmax_pairs) { + for (const auto& output_minmax : minmax_pairs) { + float input_min = input_minmax.first; + float input_max = input_minmax.second; + float output_min = output_minmax.first; + float output_max = output_minmax.second; + for (int size : {1, 3, 10, 100}) { + tflite::testing::TestHardSwishQuantized(size, input_min, input_max, + output_min, output_max, + &random_engine); + tflite::testing::TestHardSwishQuantized(size, input_min, input_max, + output_min, output_max, + &random_engine); + } + } + } +} + +// See the comment in the reference implementation of quantized HardSwish: +// A numerical issue significantly affecting ImageNet classification accuracy +// with MobileNet v3 is only observable at the scale of HardSwish unit tests +// if we monitor specifically bias. This testcase is extracted from one of the +// HardSwish nodes in that MobileNet v3 that exhibited this issue. +TF_LITE_MICRO_TEST(SimpleHardSwishTestQuantizedBias) { + tflite::testing::TestHardSwishQuantizedBias(-11.654928f, 25.036512f, + -0.3905796f, 24.50887f, 0.035); +} + + TF_LITE_MICRO_TEST(SimpleReluTestUint8) { const int elements_count = 10; diff --git a/tensorflow/lite/micro/kernels/micro_ops.h b/tensorflow/lite/micro/kernels/micro_ops.h index 24180aab8c5..25f9bd41d63 100644 --- a/tensorflow/lite/micro/kernels/micro_ops.h +++ b/tensorflow/lite/micro/kernels/micro_ops.h @@ -46,6 +46,7 @@ TfLiteRegistration* Register_FLOOR(); TfLiteRegistration* Register_FULLY_CONNECTED(); TfLiteRegistration* Register_GREATER(); TfLiteRegistration* Register_GREATER_EQUAL(); +TfLiteRegistration* Register_HARD_SWISH(); TfLiteRegistration* Register_LESS(); TfLiteRegistration* Register_LESS_EQUAL(); TfLiteRegistration* Register_LOG(); diff --git a/tensorflow/lite/micro/micro_mutable_op_resolver.h b/tensorflow/lite/micro/micro_mutable_op_resolver.h index 1b76f440a61..d77dc9577fd 100644 --- a/tensorflow/lite/micro/micro_mutable_op_resolver.h +++ b/tensorflow/lite/micro/micro_mutable_op_resolver.h @@ -217,6 +217,14 @@ class MicroMutableOpResolver : public MicroOpResolver { ParseOpData); } + TfLiteStatus AddHardSwish() { + // TODO(b/149408647): Replace ParseOpData with the operator specific parse + // function. + return AddBuiltin(BuiltinOperator_HARD_SWISH, + *tflite::ops::micro::Register_HARD_SWISH(), + ParseOpData); + } + TfLiteStatus AddL2Normalization() { // TODO(b/149408647): Replace ParseOpData with the operator specific parse // function. diff --git a/tensorflow/lite/micro/micro_utils.h b/tensorflow/lite/micro/micro_utils.h index 4f8689b943e..0a4ab1ef09c 100644 --- a/tensorflow/lite/micro/micro_utils.h +++ b/tensorflow/lite/micro/micro_utils.h @@ -94,6 +94,16 @@ void SymmetricDequantize(const int8_t* values, const int size, const float dequantization_scale, float* dequantized_values); +template +void AsymmetricDequantize(const T* values, const int size, + const float dequantization_scale, + int dequantization_zero_point, + float* dequantized_values) { + for (int i = 0; i < size; ++i) { + dequantized_values[i] = (values[i] - dequantization_zero_point) * dequantization_scale; + } +} + } // namespace tflite #endif // TENSORFLOW_LITE_MICRO_MICRO_UTILS_H_ From 820257026fa125ee102948986f0e6826b6bee883 Mon Sep 17 00:00:00 2001 From: =?UTF-8?q?M=C3=A5ns=20Nilsson?= Date: Tue, 2 Jun 2020 09:35:02 +0200 Subject: [PATCH 534/867] TFLu: Fix variable length array error in hard_swish unit test --- .../lite/micro/kernels/activations_test.cc | 87 ++++++++++++------- 1 file changed, 57 insertions(+), 30 deletions(-) diff --git a/tensorflow/lite/micro/kernels/activations_test.cc b/tensorflow/lite/micro/kernels/activations_test.cc index 0a7db1e6589..03a6b698a86 100644 --- a/tensorflow/lite/micro/kernels/activations_test.cc +++ b/tensorflow/lite/micro/kernels/activations_test.cc @@ -169,11 +169,9 @@ void EvalTestReferenceHardSwish(int size, const std::vector& input, } template -void TestHardSwishQuantized(int size, float input_min, - float input_max, float output_min, +void TestHardSwishQuantized(int size, const T* output_data, T* input_data_quantized, float* dequantized_output, + float input_min, float input_max, float output_min, float output_max, std::minstd_rand* random_engine) { - T output_data[size]; - T input_data_quantized[size]; const int input_dims_data[] = {2, 1, size}; const int output_dims_data[] = {2, 1, size}; const float input_scale = ScaleFromMinMax(input_min, input_max); @@ -193,8 +191,6 @@ void TestHardSwishQuantized(int size, float input_min, TF_LITE_MICRO_EXPECT_EQ(output_elements_count, size); - float dequantized_output[output_elements_count]; - std::vector float_input_values; std::vector float_ref_output_values; GenerateUniformRandomVector(size, input_min, input_max, random_engine, @@ -262,8 +258,9 @@ void TestHardSwishQuantized(int size, float input_min, } template -void TestHardSwishQuantizedBias(float input_min, float input_max, float output_min, - float output_max, float tolerated_bias) { +void TestHardSwishQuantizedBias(const int size, const T* output_data, T* input_data_quantized, + float* dequantized_output, float input_min, float input_max, + float output_min, float output_max, float tolerated_bias) { const float quantized_type_range = static_cast(std::numeric_limits::max()) - static_cast(std::numeric_limits::min()); @@ -294,7 +291,8 @@ void TestHardSwishQuantizedBias(float input_min, float input_max, float output_m input_min + (i - std::numeric_limits::min()) * input_scale); } - const int size = float_input_values.size(); + TF_LITE_MICRO_EXPECT_EQ(float_input_values.size(), size); + std::vector float_ref_output_values; EvalTestReferenceHardSwish(size, float_input_values, &float_ref_output_values); @@ -302,8 +300,6 @@ void TestHardSwishQuantizedBias(float input_min, float input_max, float output_m val = std::min(output_max, std::max(output_min, val)); } - T output_data[size]; - T input_data_quantized[size]; const int input_dims_data[] = {2, 1, size}; const int output_dims_data[] = {2, 1, size}; @@ -319,8 +315,6 @@ void TestHardSwishQuantizedBias(float input_min, float input_max, float output_m TF_LITE_MICRO_EXPECT_EQ(output_elements_count, size); - float dequantized_output[output_elements_count]; - constexpr int inputs_size = 1; constexpr int outputs_size = 1; constexpr int tensors_size = inputs_size + outputs_size; @@ -370,7 +364,8 @@ void TestHardSwishQuantizedBias(float input_min, float input_max, float output_m registration->free(&context, user_data); } - AsymmetricDequantize(output_data, output_elements_count, output_scale, output_zero_point, dequantized_output); + AsymmetricDequantize(output_data, output_elements_count, output_scale, + output_zero_point, dequantized_output); float sum_diff = 0; for (int i = 0; i < size; i++) { @@ -380,7 +375,7 @@ void TestHardSwishQuantizedBias(float input_min, float input_max, float output_m TF_LITE_MICRO_EXPECT_LE(std::abs(bias), tolerated_bias); } -void TestHardSwishFloat(int size, std::minstd_rand* random_engine) { +void TestHardSwishFloat(const int size, float* output_data, std::minstd_rand* random_engine) { std::vector float_input_values; const float kMin = -10.0f; const float kMax = 10.0f; @@ -390,7 +385,6 @@ void TestHardSwishFloat(int size, std::minstd_rand* random_engine) { EvalTestReferenceHardSwish(size, float_input_values, &float_ref_output_values); - float output_data[size]; const int input_dims_data[] = {1, size}; const int output_dims_data[] = {1, size}; @@ -747,29 +741,57 @@ TF_LITE_MICRO_TEST(SimpleRelu6TestFloat) { TF_LITE_MICRO_TEST(SimpleHardSwishTestFloat) { std::minstd_rand random_engine; - for (int size : {1, 2, 3, 4, 10, 20, 30, 40, 100}) { - tflite::testing::TestHardSwishFloat(size, &random_engine); - } + constexpr int size = 100; + float output_data[size] = {0.f}; + + tflite::testing::TestHardSwishFloat(size, output_data, &random_engine); } -TF_LITE_MICRO_TEST(SimpleHardSwishTestQuantized) { + +TF_LITE_MICRO_TEST(SimpleHardSwishTestInt8) { std::minstd_rand random_engine; std::vector> minmax_pairs{ {0.f, 1.f}, {-2.f, 1.f}, {-5.f, 10.f}, {-40.f, 60.f}}; + constexpr int size = 101; + constexpr int8_t output_data[size] = {0}; + int8_t input_data_quantized[size] = {0}; + float dequantized_output[size] = {0.f}; + for (const auto& input_minmax : minmax_pairs) { for (const auto& output_minmax : minmax_pairs) { float input_min = input_minmax.first; float input_max = input_minmax.second; float output_min = output_minmax.first; float output_max = output_minmax.second; - for (int size : {1, 3, 10, 100}) { - tflite::testing::TestHardSwishQuantized(size, input_min, input_max, - output_min, output_max, - &random_engine); - tflite::testing::TestHardSwishQuantized(size, input_min, input_max, - output_min, output_max, - &random_engine); - } + + tflite::testing::TestHardSwishQuantized(size, output_data, input_data_quantized, dequantized_output, + input_min, input_max, output_min, output_max, + &random_engine); + + } + } +} + +TF_LITE_MICRO_TEST(SimpleHardSwishTestUint8) { + std::minstd_rand random_engine; + std::vector> minmax_pairs{ + {0.f, 1.f}, {-2.f, 1.f}, {-5.f, 10.f}, {-40.f, 60.f}}; + constexpr int size = 99; + constexpr uint8_t output_data[size] = {0}; + uint8_t input_data_quantized[size] = {0}; + float dequantized_output[size] = {0.f}; + + for (const auto& input_minmax : minmax_pairs) { + for (const auto& output_minmax : minmax_pairs) { + float input_min = input_minmax.first; + float input_max = input_minmax.second; + float output_min = output_minmax.first; + float output_max = output_minmax.second; + + tflite::testing::TestHardSwishQuantized(size, output_data, input_data_quantized, dequantized_output, + input_min, input_max, output_min, output_max, + &random_engine); + } } } @@ -780,8 +802,13 @@ TF_LITE_MICRO_TEST(SimpleHardSwishTestQuantized) { // if we monitor specifically bias. This testcase is extracted from one of the // HardSwish nodes in that MobileNet v3 that exhibited this issue. TF_LITE_MICRO_TEST(SimpleHardSwishTestQuantizedBias) { - tflite::testing::TestHardSwishQuantizedBias(-11.654928f, 25.036512f, - -0.3905796f, 24.50887f, 0.035); + constexpr int size = 43; + constexpr uint8_t output_data[size] = {0}; + uint8_t input_data_quantized[size] = {0}; + float dequantized_output[size] = {0.f}; + + tflite::testing::TestHardSwishQuantizedBias(size, output_data, input_data_quantized, dequantized_output, + -11.654928f, 25.036512f, -0.3905796f, 24.50887f, 0.035); } From c2fc5dccfa021fa8f7ec48da874c0d91a2d211c4 Mon Sep 17 00:00:00 2001 From: =?UTF-8?q?M=C3=A5ns=20Nilsson?= Date: Wed, 10 Jun 2020 11:14:27 +0200 Subject: [PATCH 535/867] TFLu: Update hard swish op as per review comments --- tensorflow/lite/kernels/internal/BUILD | 2 + .../kernels/internal/reference/activations.h | 54 +++++++ .../internal/reference/reference_ops.h | 15 +- tensorflow/lite/micro/kernels/activations.cc | 150 +++++++++--------- .../lite/micro/kernels/activations_test.cc | 121 +++++++------- 5 files changed, 196 insertions(+), 146 deletions(-) create mode 100644 tensorflow/lite/kernels/internal/reference/activations.h diff --git a/tensorflow/lite/kernels/internal/BUILD b/tensorflow/lite/kernels/internal/BUILD index a02a5bf3981..31a6a74eb5e 100644 --- a/tensorflow/lite/kernels/internal/BUILD +++ b/tensorflow/lite/kernels/internal/BUILD @@ -437,6 +437,7 @@ cc_library( name = "reference_base", srcs = [], hdrs = [ + "reference/activations.h", "reference/add.h", "reference/arg_min_max.h", "reference/batch_matmul.h", @@ -525,6 +526,7 @@ cc_library( name = "legacy_reference_base", srcs = [], hdrs = [ + "reference/activations.h", "reference/add.h", "reference/arg_min_max.h", "reference/binary_function.h", diff --git a/tensorflow/lite/kernels/internal/reference/activations.h b/tensorflow/lite/kernels/internal/reference/activations.h new file mode 100644 index 00000000000..2110f5e3cbc --- /dev/null +++ b/tensorflow/lite/kernels/internal/reference/activations.h @@ -0,0 +1,54 @@ +/* Copyright 2020 The TensorFlow Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. +==============================================================================*/ +#ifndef TENSORFLOW_LITE_KERNELS_INTERNAL_REFERENCE_ACTIVATIONS_H_ +#define TENSORFLOW_LITE_KERNELS_INTERNAL_REFERENCE_ACTIVATIONS_H_ + +#include "tensorflow/lite/kernels/internal/types.h" +#include "tensorflow/lite/kernels/internal/common.h" + + +namespace tflite { +namespace reference_ops { + +inline std::int32_t RoundingDivideByPOT(std::int32_t numerator, int exponent) { + std::int32_t sign = numerator >= 0 ? 1 : -1; + std::int32_t abs_numerator = std::abs(numerator); + std::int32_t mask = (1LL << exponent) - 1; + std::int32_t remainder = abs_numerator & mask; + std::int32_t threshold = mask >> 1; + std::int32_t abs_result = + (abs_numerator >> exponent) + (remainder > threshold ? 1 : 0); + return sign * abs_result; +} + +template +inline void HardSwish(const RuntimeShape& input_shape, const T* input_data, + const RuntimeShape& output_shape, T* output_data) { + ruy::profiler::ScopeLabel label("ReferenceHardSwish/Float"); + auto matching_size = MatchingFlatSize(input_shape, output_shape); + const T* in_end = input_data + matching_size; + for (; input_data < in_end; input_data++, output_data++) { + const float in = *input_data; + *output_data = + in * std::min(static_cast(6), std::max(static_cast(0), in + 3)) / + 6; + } +} + +} // namespace reference_ops +} // namespace tflite + + +#endif // TENSORFLOW_LITE_KERNELS_INTERNAL_REFERENCE_CONV_H_ diff --git a/tensorflow/lite/kernels/internal/reference/reference_ops.h b/tensorflow/lite/kernels/internal/reference/reference_ops.h index b94c2060b2d..3e9ebff69ad 100644 --- a/tensorflow/lite/kernels/internal/reference/reference_ops.h +++ b/tensorflow/lite/kernels/internal/reference/reference_ops.h @@ -32,6 +32,7 @@ limitations under the License. #include "tensorflow/lite/c/common.h" #include "tensorflow/lite/kernels/internal/common.h" #include "tensorflow/lite/kernels/internal/quantization_util.h" +#include "tensorflow/lite/kernels/internal/reference/activations.h" #include "tensorflow/lite/kernels/internal/reference/add.h" #include "tensorflow/lite/kernels/internal/reference/arg_min_max.h" #include "tensorflow/lite/kernels/internal/reference/binary_function.h" @@ -2601,20 +2602,6 @@ void ReverseSequence(const TS* seq_lengths, const int seq_dim, } } -template -inline void HardSwish(const RuntimeShape& input_shape, const T* input_data, - const RuntimeShape& output_shape, T* output_data) { - ruy::profiler::ScopeLabel label("ReferenceHardSwish/Float"); - auto matching_size = MatchingFlatSize(input_shape, output_shape); - const T* in_end = input_data + matching_size; - for (; input_data < in_end; input_data++, output_data++) { - const float in = *input_data; - *output_data = - in * std::min(static_cast(6), std::max(static_cast(0), in + 3)) / - 6; - } -} - template inline void HardSwish(const HardSwishParams& params, const RuntimeShape& input_shape, const T* input_data, diff --git a/tensorflow/lite/micro/kernels/activations.cc b/tensorflow/lite/micro/kernels/activations.cc index 75b0339002d..934f029ac6a 100644 --- a/tensorflow/lite/micro/kernels/activations.cc +++ b/tensorflow/lite/micro/kernels/activations.cc @@ -19,6 +19,7 @@ limitations under the License. #include "tensorflow/lite/kernels/internal/quantization_util.h" #include "tensorflow/lite/kernels/internal/tensor_ctypes.h" #include "tensorflow/lite/kernels/internal/types.h" +#include "tensorflow/lite/kernels/internal/reference/reference_ops.h" #include "tensorflow/lite/kernels/kernel_util.h" #include "tensorflow/lite/kernels/op_macros.h" #include "tensorflow/lite/micro/micro_utils.h" @@ -78,33 +79,10 @@ inline void Relu6Quantized(Q lower, Q upper, const RuntimeShape& input_shape, } } -inline std::int32_t RoundingDivideByPOT(std::int32_t numerator, int exponent) { - std::int32_t sign = numerator >= 0 ? 1 : -1; - std::int32_t abs_numerator = std::abs(numerator); - std::int32_t mask = (1LL << exponent) - 1; - std::int32_t remainder = abs_numerator & mask; - std::int32_t threshold = mask >> 1; - std::int32_t abs_result = - (abs_numerator >> exponent) + (remainder > threshold ? 1 : 0); - return sign * abs_result; -} - -inline void HardSwishFloatOp(const RuntimeShape& input_shape, const float* input_data, - const RuntimeShape& output_shape, float* output_data) { - auto matching_size = MatchingFlatSize(input_shape, output_shape); - const float* in_end = input_data + matching_size; - for (; input_data < in_end; input_data++, output_data++) { - const float in = *input_data; - *output_data = - in * std::min(static_cast(6), std::max(static_cast(0), in + 3)) / - 6; - } -} - template void HardSwishOp(HardSwishParams& params, - const RuntimeShape& input_shape, const T* input_data, - const RuntimeShape& output_shape, T* output_data) { + const RuntimeShape& input_shape, const T* input_data, + const RuntimeShape& output_shape, T* output_data) { const int flat_size = MatchingFlatSize(input_shape, output_shape); for (int i = 0; i < flat_size; i++) { @@ -165,7 +143,7 @@ void HardSwishOp(HardSwishParams& params, } // Shift right, in the right-shifting case. if (params.reluish_multiplier_exponent < 0) { - reluish_value = RoundingDivideByPOT( + reluish_value = tflite::reference_ops::RoundingDivideByPOT( reluish_value, -params.reluish_multiplier_exponent); } // At this point we have rescaled the value into a 16bit fixedpoint @@ -179,7 +157,7 @@ void HardSwishOp(HardSwishParams& params, reluish_value, input_value_on_preshift_output_scale); // We were so far operating on the pre-shift output scale. Now we finally // apply that output shift, arriving at the final output scale. - int16_t output_value = RoundingDivideByPOT( + int16_t output_value = tflite::reference_ops::RoundingDivideByPOT( preshift_output_value, -params.output_multiplier_exponent); output_value += params.output_zero_point; output_value = @@ -190,46 +168,11 @@ void HardSwishOp(HardSwishParams& params, } } - -template -TfLiteStatus HardSwishQuantized(TfLiteContext* context, TfLiteNode* node) { - const TfLiteTensor* input = GetInput(context, node, kInputTensor); - TfLiteTensor* output = GetOutput(context, node, kOutputTensor); - - HardSwishParams params; - - params.input_zero_point = input->params.zero_point; - params.output_zero_point = output->params.zero_point; - - const float input_scale = input->params.scale; - const float hires_input_scale = (1.0f / 128.0f) * input_scale; - const float reluish_scale = 3.0f / 32768.0f; - const float output_scale = output->params.scale; - - const double output_multiplier = static_cast(hires_input_scale / output_scale); - int32_t output_multiplier_fixedpoint_int32; - QuantizeMultiplier(output_multiplier, &output_multiplier_fixedpoint_int32, - ¶ms.output_multiplier_exponent); - DownScaleInt32ToInt16Multiplier( - output_multiplier_fixedpoint_int32, - ¶ms.output_multiplier_fixedpoint_int16); - - TF_LITE_ENSURE(context, params.output_multiplier_exponent <= 0); - - const double reluish_multiplier = static_cast(hires_input_scale / reluish_scale); - int32_t reluish_multiplier_fixedpoint_int32; - QuantizeMultiplier(reluish_multiplier, &reluish_multiplier_fixedpoint_int32, - ¶ms.reluish_multiplier_exponent); - DownScaleInt32ToInt16Multiplier( - reluish_multiplier_fixedpoint_int32, - ¶ms.reluish_multiplier_fixedpoint_int16); - - HardSwishOp(params, GetTensorShape(input), - GetTensorData(input), GetTensorShape(output), GetTensorData(output)); - return kTfLiteOk; -} - TfLiteStatus ReluPrepare(TfLiteContext* context, TfLiteNode* node) { + // Validate number of inputs and outputs + TF_LITE_ENSURE_EQ(context, NumInputs(node), 1); + TF_LITE_ENSURE_EQ(context, NumOutputs(node), 1); + return kTfLiteOk; } @@ -267,6 +210,10 @@ TfLiteStatus ReluEval(TfLiteContext* context, TfLiteNode* node) { } TfLiteStatus Relu6Prepare(TfLiteContext* context, TfLiteNode* node) { + // Validate number of inputs and outputs + TF_LITE_ENSURE_EQ(context, NumInputs(node), 1); + TF_LITE_ENSURE_EQ(context, NumOutputs(node), 1); + return kTfLiteOk; } @@ -307,34 +254,84 @@ TfLiteStatus Relu6Eval(TfLiteContext* context, TfLiteNode* node) { } } +void* HardSwishInit(TfLiteContext* context, const char* buffer, size_t length) { + void* data = nullptr; + if (context->AllocatePersistentBuffer(context, sizeof(HardSwishParams), &data) == + kTfLiteError) { + return nullptr; + } + return data; +} + TfLiteStatus HardSwishPrepare(TfLiteContext* context, TfLiteNode* node) { + TFLITE_DCHECK(node->user_data != nullptr); + TF_LITE_ENSURE_EQ(context, NumInputs(node), 1); + TF_LITE_ENSURE_EQ(context, NumOutputs(node), 1); + + const TfLiteTensor* input = GetInput(context, node, kInputTensor); + TfLiteTensor* output = GetOutput(context, node, kOutputTensor); + + if (input->type == kTfLiteUInt8 || input->type == kTfLiteInt8) { + HardSwishParams* params = static_cast(node->user_data); + + params->input_zero_point = input->params.zero_point; + params->output_zero_point = output->params.zero_point; + + const float input_scale = input->params.scale; + const float hires_input_scale = (1.0f / 128.0f) * input_scale; + const float reluish_scale = 3.0f / 32768.0f; + const float output_scale = output->params.scale; + + const double output_multiplier = static_cast(hires_input_scale / output_scale); + int32_t output_multiplier_fixedpoint_int32; + QuantizeMultiplier(output_multiplier, &output_multiplier_fixedpoint_int32, + ¶ms->output_multiplier_exponent); + DownScaleInt32ToInt16Multiplier( + output_multiplier_fixedpoint_int32, + ¶ms->output_multiplier_fixedpoint_int16); + + TF_LITE_ENSURE(context, params->output_multiplier_exponent <= 0); + + const double reluish_multiplier = static_cast(hires_input_scale / reluish_scale); + int32_t reluish_multiplier_fixedpoint_int32; + QuantizeMultiplier(reluish_multiplier, &reluish_multiplier_fixedpoint_int32, + ¶ms->reluish_multiplier_exponent); + DownScaleInt32ToInt16Multiplier( + reluish_multiplier_fixedpoint_int32, + ¶ms->reluish_multiplier_fixedpoint_int16); + } + return kTfLiteOk; } TfLiteStatus HardSwishEval(TfLiteContext* context, TfLiteNode* node) { const TfLiteTensor* input = GetInput(context, node, kInputTensor); TfLiteTensor* output = GetOutput(context, node, kOutputTensor); + HardSwishParams* params = static_cast(node->user_data); switch (input->type) { case kTfLiteFloat32: { - HardSwishFloatOp( + tflite::reference_ops::HardSwish( GetTensorShape(input), GetTensorData(input), GetTensorShape(output), GetTensorData(output)); - return kTfLiteOk; } break; case kTfLiteUInt8: { - return HardSwishQuantized(context, node); + HardSwishOp(*params, GetTensorShape(input), + GetTensorData(input), GetTensorShape(output), GetTensorData(output)); } break; case kTfLiteInt8: { - return HardSwishQuantized(context, node); + HardSwishOp(*params, GetTensorShape(input), + GetTensorData(input), GetTensorShape(output), GetTensorData(output)); } break; - default: + default: { TF_LITE_KERNEL_LOG(context, "Only float32/int8/uint8 are supported currently, got %s", TfLiteTypeGetName(input->type)); return kTfLiteError; + } } + return kTfLiteOk; } } // namespace activations @@ -364,9 +361,14 @@ TfLiteRegistration* Register_RELU6() { } TfLiteRegistration* Register_HARD_SWISH() { - static TfLiteRegistration r = {}; - r.prepare = activations::HardSwishPrepare; - r.invoke = activations::HardSwishEval; + static TfLiteRegistration r = {/*init=*/activations::HardSwishInit, + /*free=*/nullptr, + /*prepare=*/activations::HardSwishPrepare, + /*invoke=*/activations::HardSwishEval, + /*profiling_string=*/nullptr, + /*builtin_code=*/0, + /*custom_name=*/nullptr, + /*version=*/0}; return &r; } diff --git a/tensorflow/lite/micro/kernels/activations_test.cc b/tensorflow/lite/micro/kernels/activations_test.cc index 03a6b698a86..b1d8e429025 100644 --- a/tensorflow/lite/micro/kernels/activations_test.cc +++ b/tensorflow/lite/micro/kernels/activations_test.cc @@ -139,7 +139,7 @@ void TestRelu6Float(const int* input_dims_data, const float* input_data, void GenerateUniformRandomVector(int size, float min, float max, std::minstd_rand* random_engine, - std::vector* result) { + float* result) { // Never use std::uniform_*_distribution in tests, it's // implementation-defined. Likewise, don't use std::default_random_engine, // implementation-defined. Implementation-defined is bad because it means that @@ -147,7 +147,6 @@ void GenerateUniformRandomVector(int size, float min, float max, // std::minstd_rand is a standard instantiation of // std::linear_congruential_engine, the cheapest generator in c++11 stdlib, // it's good enough here. - result->resize(size); for (int i = 0; i < size; i++) { // We don't care whether the `max` value may ever be produced exactly. // It may actually be thanks to rounding, as std::minstd_rand::modulus @@ -155,23 +154,23 @@ void GenerateUniformRandomVector(int size, float min, float max, float random_value_scaled_0_1 = (*random_engine)() * (1.0f / static_cast(std::minstd_rand::modulus)); - (*result)[i] = min + (max - min) * random_value_scaled_0_1; + result[i] = min + (max - min) * random_value_scaled_0_1; } } -void EvalTestReferenceHardSwish(int size, const std::vector& input, - std::vector* result) { - result->resize(size); +void EvalTestReferenceHardSwish(int size, float* input, + float* result) { for (int i = 0; i < size; i++) { const float in = input[i]; - (*result)[i] = in * std::min(6.0f, std::max(0.0f, in + 3)) * (1.0f / 6.0f); + result[i] = in * std::min(6.0f, std::max(0.0f, in + 3)) * (1.0f / 6.0f); } } template void TestHardSwishQuantized(int size, const T* output_data, T* input_data_quantized, float* dequantized_output, float input_min, float input_max, float output_min, - float output_max, std::minstd_rand* random_engine) { + float output_max, std::minstd_rand* random_engine, + float *float_input_values, float* float_ref_output_values) { const int input_dims_data[] = {2, 1, size}; const int output_dims_data[] = {2, 1, size}; const float input_scale = ScaleFromMinMax(input_min, input_max); @@ -191,21 +190,20 @@ void TestHardSwishQuantized(int size, const T* output_data, T* input_data_quanti TF_LITE_MICRO_EXPECT_EQ(output_elements_count, size); - std::vector float_input_values; - std::vector float_ref_output_values; GenerateUniformRandomVector(size, input_min, input_max, random_engine, - &float_input_values); + float_input_values); EvalTestReferenceHardSwish(size, float_input_values, - &float_ref_output_values); - for (float& val : float_ref_output_values) { - val = std::min(output_max, std::max(output_min, val)); + float_ref_output_values); + for (int i = 0; i < size; i++) { + float val = float_ref_output_values[i]; + float_ref_output_values[i] = std::min(output_max, std::max(output_min, val)); } constexpr int inputs_size = 1; constexpr int outputs_size = 1; constexpr int tensors_size = inputs_size + outputs_size; TfLiteTensor tensors[tensors_size] = { - CreateQuantizedTensor(float_input_values.data(), input_data_quantized, input_dims, + CreateQuantizedTensor(float_input_values, input_data_quantized, input_dims, input_scale, input_zero_point, "input_tensor"), CreateQuantizedTensor(output_data, output_dims, output_scale, output_zero_point, "output_tensor"), @@ -253,19 +251,19 @@ void TestHardSwishQuantized(int size, const T* output_data, T* input_data_quanti AsymmetricDequantize(output_data, output_elements_count, output_scale, output_zero_point, dequantized_output); for (int i = 0; i < output_elements_count; ++i) { - TF_LITE_MICRO_EXPECT_NEAR(float_ref_output_values.data()[i], dequantized_output[i], kTolerance); + TF_LITE_MICRO_EXPECT_NEAR(float_ref_output_values[i], dequantized_output[i], kTolerance); } } template void TestHardSwishQuantizedBias(const int size, const T* output_data, T* input_data_quantized, float* dequantized_output, float input_min, float input_max, - float output_min, float output_max, float tolerated_bias) { + float output_min, float output_max, float tolerated_bias, + float* float_input_values, float* float_ref_output_values) { const float quantized_type_range = static_cast(std::numeric_limits::max()) - static_cast(std::numeric_limits::min()); - const float input_scale = ScaleFromMinMax(input_min, input_max); const float output_scale = ScaleFromMinMax(output_min, output_max); @@ -284,20 +282,18 @@ void TestHardSwishQuantizedBias(const int size, const T* output_data, T* input_d const int quantized_input_positive_three = std::round(std::numeric_limits::min() + (3.0f - input_min) / input_scale); - std::vector float_input_values; + for (int i = quantized_input_negative_three; i <= quantized_input_positive_three; i++) { - float_input_values.push_back( - input_min + - (i - std::numeric_limits::min()) * input_scale); + float_input_values[i] = input_min + + (i - std::numeric_limits::min()) * input_scale; } - TF_LITE_MICRO_EXPECT_EQ(float_input_values.size(), size); - std::vector float_ref_output_values; EvalTestReferenceHardSwish(size, float_input_values, - &float_ref_output_values); - for (float& val : float_ref_output_values) { - val = std::min(output_max, std::max(output_min, val)); + float_ref_output_values); + for (int i = 0; i < size; i++) { + float val = float_ref_output_values[i]; + float_ref_output_values[i] = std::min(output_max, std::max(output_min, val)); } const int input_dims_data[] = {2, 1, size}; @@ -319,7 +315,7 @@ void TestHardSwishQuantizedBias(const int size, const T* output_data, T* input_d constexpr int outputs_size = 1; constexpr int tensors_size = inputs_size + outputs_size; TfLiteTensor tensors[tensors_size] = { - CreateQuantizedTensor(float_input_values.data(), input_data_quantized, input_dims, + CreateQuantizedTensor(float_input_values, input_data_quantized, input_dims, input_scale, input_zero_point, "input_tensor"), CreateQuantizedTensor(output_data, output_dims, output_scale, output_zero_point, "output_tensor"), @@ -375,15 +371,15 @@ void TestHardSwishQuantizedBias(const int size, const T* output_data, T* input_d TF_LITE_MICRO_EXPECT_LE(std::abs(bias), tolerated_bias); } -void TestHardSwishFloat(const int size, float* output_data, std::minstd_rand* random_engine) { - std::vector float_input_values; +void TestHardSwishFloat(const int size, float* output_data, std::minstd_rand* random_engine, + float* float_input_values, float* float_ref_output_values) { const float kMin = -10.0f; const float kMax = 10.0f; GenerateUniformRandomVector(size, kMin, kMax, random_engine, - &float_input_values); - std::vector float_ref_output_values; + float_input_values); + EvalTestReferenceHardSwish(size, float_input_values, - &float_ref_output_values); + float_ref_output_values); const int input_dims_data[] = {1, size}; const int output_dims_data[] = {1, size}; @@ -398,7 +394,7 @@ void TestHardSwishFloat(const int size, float* output_data, std::minstd_rand* ra constexpr int outputs_size = 1; constexpr int tensors_size = inputs_size + outputs_size; TfLiteTensor tensors[tensors_size] = { - CreateFloatTensor(float_input_values.data(), input_dims, "input_tensor"), + CreateFloatTensor(float_input_values, input_dims, "input_tensor"), CreateFloatTensor(output_data, output_dims, "output_tensor"), }; TfLiteContext context; @@ -439,7 +435,7 @@ void TestHardSwishFloat(const int size, float* output_data, std::minstd_rand* ra } for (int i = 0; i < output_elements_count; ++i) { - TF_LITE_MICRO_EXPECT_NEAR(float_ref_output_values.data()[i], output_data[i], 1e-5f); + TF_LITE_MICRO_EXPECT_NEAR(float_ref_output_values[i], output_data[i], 1e-5f); } } @@ -743,55 +739,62 @@ TF_LITE_MICRO_TEST(SimpleHardSwishTestFloat) { std::minstd_rand random_engine; constexpr int size = 100; float output_data[size] = {0.f}; + float input_values[size] = {0.f}; + float output_values[size] = {0.f}; - tflite::testing::TestHardSwishFloat(size, output_data, &random_engine); + tflite::testing::TestHardSwishFloat(size, output_data, &random_engine, + input_values, output_values); } TF_LITE_MICRO_TEST(SimpleHardSwishTestInt8) { std::minstd_rand random_engine; - std::vector> minmax_pairs{ - {0.f, 1.f}, {-2.f, 1.f}, {-5.f, 10.f}, {-40.f, 60.f}}; + constexpr int pairs = 4, one_pair = 2; constexpr int size = 101; + constexpr float minmax_pairs[pairs][one_pair] = { + {0.f, 1.f}, {-2.f, 1.f}, {-5.f, 10.f}, {-40.f, 60.f}}; constexpr int8_t output_data[size] = {0}; int8_t input_data_quantized[size] = {0}; float dequantized_output[size] = {0.f}; + float input_values[size] = {0.f}; + float output_values[size] = {0.f}; - for (const auto& input_minmax : minmax_pairs) { - for (const auto& output_minmax : minmax_pairs) { - float input_min = input_minmax.first; - float input_max = input_minmax.second; - float output_min = output_minmax.first; - float output_max = output_minmax.second; + for (int x = 0; x < pairs; x++) { + for (int y = 0; y < pairs; y++) { + float input_min = minmax_pairs[x][0]; + float input_max = minmax_pairs[x][1]; + float output_min = minmax_pairs[y][0]; + float output_max = minmax_pairs[y][1]; tflite::testing::TestHardSwishQuantized(size, output_data, input_data_quantized, dequantized_output, input_min, input_max, output_min, output_max, - &random_engine); - + &random_engine, input_values, output_values); } } } TF_LITE_MICRO_TEST(SimpleHardSwishTestUint8) { std::minstd_rand random_engine; - std::vector> minmax_pairs{ - {0.f, 1.f}, {-2.f, 1.f}, {-5.f, 10.f}, {-40.f, 60.f}}; constexpr int size = 99; + constexpr int pairs = 4, one_pair = 2; + constexpr float minmax_pairs[pairs][one_pair] = { + {0.f, 1.f}, {-2.f, 1.f}, {-5.f, 10.f}, {-40.f, 60.f}}; constexpr uint8_t output_data[size] = {0}; uint8_t input_data_quantized[size] = {0}; float dequantized_output[size] = {0.f}; + float input_values[size] = {0.f}; + float output_values[size] = {0.f}; - for (const auto& input_minmax : minmax_pairs) { - for (const auto& output_minmax : minmax_pairs) { - float input_min = input_minmax.first; - float input_max = input_minmax.second; - float output_min = output_minmax.first; - float output_max = output_minmax.second; + for (int x = 0; x < pairs; x++) { + for (int y = 0; y < pairs; y++) { + float input_min = minmax_pairs[x][0]; + float input_max = minmax_pairs[x][1]; + float output_min = minmax_pairs[y][0]; + float output_max = minmax_pairs[y][1]; tflite::testing::TestHardSwishQuantized(size, output_data, input_data_quantized, dequantized_output, input_min, input_max, output_min, output_max, - &random_engine); - + &random_engine, input_values, output_values); } } } @@ -806,12 +809,14 @@ TF_LITE_MICRO_TEST(SimpleHardSwishTestQuantizedBias) { constexpr uint8_t output_data[size] = {0}; uint8_t input_data_quantized[size] = {0}; float dequantized_output[size] = {0.f}; + float input_values[size] = {0.f}; + float output_values[size] = {0.f}; tflite::testing::TestHardSwishQuantizedBias(size, output_data, input_data_quantized, dequantized_output, - -11.654928f, 25.036512f, -0.3905796f, 24.50887f, 0.035); + -11.654928f, 25.036512f, -0.3905796f, 24.50887f, 0.035, + input_values, output_values); } - TF_LITE_MICRO_TEST(SimpleReluTestUint8) { const int elements_count = 10; From 801779e9231ec58497b6ce4d921abe459045a87d Mon Sep 17 00:00:00 2001 From: =?UTF-8?q?M=C3=A5ns=20Nilsson?= Date: Thu, 11 Jun 2020 15:56:58 +0200 Subject: [PATCH 536/867] TFLu: do not directly include reference_ops.h --- tensorflow/lite/kernels/internal/reference/activations.h | 2 +- tensorflow/lite/micro/kernels/activations.cc | 2 +- 2 files changed, 2 insertions(+), 2 deletions(-) diff --git a/tensorflow/lite/kernels/internal/reference/activations.h b/tensorflow/lite/kernels/internal/reference/activations.h index 2110f5e3cbc..c043f1e0845 100644 --- a/tensorflow/lite/kernels/internal/reference/activations.h +++ b/tensorflow/lite/kernels/internal/reference/activations.h @@ -15,10 +15,10 @@ limitations under the License. #ifndef TENSORFLOW_LITE_KERNELS_INTERNAL_REFERENCE_ACTIVATIONS_H_ #define TENSORFLOW_LITE_KERNELS_INTERNAL_REFERENCE_ACTIVATIONS_H_ +#include "ruy/profiler/instrumentation.h" // from @ruy #include "tensorflow/lite/kernels/internal/types.h" #include "tensorflow/lite/kernels/internal/common.h" - namespace tflite { namespace reference_ops { diff --git a/tensorflow/lite/micro/kernels/activations.cc b/tensorflow/lite/micro/kernels/activations.cc index 934f029ac6a..a536878e4cb 100644 --- a/tensorflow/lite/micro/kernels/activations.cc +++ b/tensorflow/lite/micro/kernels/activations.cc @@ -19,7 +19,7 @@ limitations under the License. #include "tensorflow/lite/kernels/internal/quantization_util.h" #include "tensorflow/lite/kernels/internal/tensor_ctypes.h" #include "tensorflow/lite/kernels/internal/types.h" -#include "tensorflow/lite/kernels/internal/reference/reference_ops.h" +#include "tensorflow/lite/kernels/internal/reference/activations.h" #include "tensorflow/lite/kernels/kernel_util.h" #include "tensorflow/lite/kernels/op_macros.h" #include "tensorflow/lite/micro/micro_utils.h" From 7473be170522cf994f541de07f867648b25c450e Mon Sep 17 00:00:00 2001 From: =?UTF-8?q?M=C3=A5ns=20Nilsson?= Date: Tue, 23 Jun 2020 10:09:18 +0200 Subject: [PATCH 537/867] TFLu: remove parameter to CreateTensor calls as it was removed --- tensorflow/lite/micro/kernels/activations_test.cc | 12 ++++++------ 1 file changed, 6 insertions(+), 6 deletions(-) diff --git a/tensorflow/lite/micro/kernels/activations_test.cc b/tensorflow/lite/micro/kernels/activations_test.cc index b1d8e429025..d8fa1c56a5d 100644 --- a/tensorflow/lite/micro/kernels/activations_test.cc +++ b/tensorflow/lite/micro/kernels/activations_test.cc @@ -204,9 +204,9 @@ void TestHardSwishQuantized(int size, const T* output_data, T* input_data_quanti constexpr int tensors_size = inputs_size + outputs_size; TfLiteTensor tensors[tensors_size] = { CreateQuantizedTensor(float_input_values, input_data_quantized, input_dims, - input_scale, input_zero_point, "input_tensor"), + input_scale, input_zero_point), CreateQuantizedTensor(output_data, output_dims, output_scale, - output_zero_point, "output_tensor"), + output_zero_point), }; TfLiteContext context; @@ -316,9 +316,9 @@ void TestHardSwishQuantizedBias(const int size, const T* output_data, T* input_d constexpr int tensors_size = inputs_size + outputs_size; TfLiteTensor tensors[tensors_size] = { CreateQuantizedTensor(float_input_values, input_data_quantized, input_dims, - input_scale, input_zero_point, "input_tensor"), + input_scale, input_zero_point), CreateQuantizedTensor(output_data, output_dims, output_scale, - output_zero_point, "output_tensor"), + output_zero_point), }; TfLiteContext context; @@ -394,8 +394,8 @@ void TestHardSwishFloat(const int size, float* output_data, std::minstd_rand* ra constexpr int outputs_size = 1; constexpr int tensors_size = inputs_size + outputs_size; TfLiteTensor tensors[tensors_size] = { - CreateFloatTensor(float_input_values, input_dims, "input_tensor"), - CreateFloatTensor(output_data, output_dims, "output_tensor"), + CreateFloatTensor(float_input_values, input_dims), + CreateFloatTensor(output_data, output_dims), }; TfLiteContext context; PopulateContext(tensors, tensors_size, micro_test::reporter, &context); From f5e1a27293823b799a080c384d854aff8f3ef133 Mon Sep 17 00:00:00 2001 From: Tres Popp Date: Tue, 23 Jun 2020 02:01:33 -0700 Subject: [PATCH 538/867] Integrate LLVM at https://github.com/llvm/llvm-project/commit/f570d5810485 PiperOrigin-RevId: 317824506 Change-Id: I5da78511de1e20d9b0a62fd3a8092b791a35b9da --- .../compiler/mlir/xla/transforms/lhlo_legalize_to_llvm.cc | 2 +- tensorflow/workspace.bzl | 4 ++-- third_party/mlir/BUILD | 1 + 3 files changed, 4 insertions(+), 3 deletions(-) diff --git a/tensorflow/compiler/mlir/xla/transforms/lhlo_legalize_to_llvm.cc b/tensorflow/compiler/mlir/xla/transforms/lhlo_legalize_to_llvm.cc index 99d2c08aa98..78a77dc3b4d 100644 --- a/tensorflow/compiler/mlir/xla/transforms/lhlo_legalize_to_llvm.cc +++ b/tensorflow/compiler/mlir/xla/transforms/lhlo_legalize_to_llvm.cc @@ -129,7 +129,7 @@ struct DynamicMemRefCastOpConverter void PopulateLhloToLLVMConversionPatterns(LLVMTypeConverter *converter, OwningRewritePatternList *patterns) { patterns->insert( - *converter); + *converter, LowerToLLVMOptions()); } } // namespace xla_lhlo diff --git a/tensorflow/workspace.bzl b/tensorflow/workspace.bzl index 0f591ba8b90..a55dcd17b07 100755 --- a/tensorflow/workspace.bzl +++ b/tensorflow/workspace.bzl @@ -710,8 +710,8 @@ def tf_repositories(path_prefix = "", tf_repo_name = ""): ) # Check out LLVM and MLIR from llvm-project. - LLVM_COMMIT = "7e825abd5704ce28b166f9463d4bd304348fd2a9" - LLVM_SHA256 = "a21b752ee1866e195f3f72c7931c79f8c4ecc0f14861488284bdc2bdf14d6fe9" + LLVM_COMMIT = "f570d5810485fa6fb2e1009f795a899d79bd429f" + LLVM_SHA256 = "e154a1a97c3b6bead73a32bcb6fc37aac2d80628abee2961315304a4964f04fe" LLVM_URLS = [ "https://storage.googleapis.com/mirror.tensorflow.org/github.com/llvm/llvm-project/archive/{commit}.tar.gz".format(commit = LLVM_COMMIT), "https://github.com/llvm/llvm-project/archive/{commit}.tar.gz".format(commit = LLVM_COMMIT), diff --git a/third_party/mlir/BUILD b/third_party/mlir/BUILD index 2be59fc44b1..8fd0a94bf64 100644 --- a/third_party/mlir/BUILD +++ b/third_party/mlir/BUILD @@ -1973,6 +1973,7 @@ cc_library( ":StandardOps", ":Support", ":Transforms", + ":VectorOps", "@llvm-project//llvm:Support", ], ) From 7a124ecab8070e7affc2fabf019b5b57588d85ee Mon Sep 17 00:00:00 2001 From: "A. Unique TensorFlower" Date: Tue, 23 Jun 2020 02:01:48 -0700 Subject: [PATCH 539/867] Update GraphDef version to 441. PiperOrigin-RevId: 317824531 Change-Id: I7fbf95f3a662d2d8d823656e9ddf5ae2dfd487f6 --- tensorflow/core/public/version.h | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/tensorflow/core/public/version.h b/tensorflow/core/public/version.h index e63ca48863f..52a926c8d8b 100644 --- a/tensorflow/core/public/version.h +++ b/tensorflow/core/public/version.h @@ -108,7 +108,7 @@ limitations under the License. #define TF_GRAPH_DEF_VERSION_MIN_PRODUCER 0 #define TF_GRAPH_DEF_VERSION_MIN_CONSUMER 0 -#define TF_GRAPH_DEF_VERSION 440 // Updated: 2020/6/22 +#define TF_GRAPH_DEF_VERSION 441 // Updated: 2020/6/23 // Checkpoint compatibility versions (the versions field in SavedSliceMeta). // From 5aa8be781eebe1f3544d4c19f83ca0fa6351b4a3 Mon Sep 17 00:00:00 2001 From: "A. Unique TensorFlower" Date: Tue, 23 Jun 2020 02:01:50 -0700 Subject: [PATCH 540/867] compat: Update forward compatibility horizon to 2020-06-23 PiperOrigin-RevId: 317824535 Change-Id: I12a1d7a71a420b996f67cf3c536a9c5fd3365154 --- tensorflow/python/compat/compat.py | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/tensorflow/python/compat/compat.py b/tensorflow/python/compat/compat.py index 301f4608aba..521d7eaf30f 100644 --- a/tensorflow/python/compat/compat.py +++ b/tensorflow/python/compat/compat.py @@ -33,7 +33,7 @@ from tensorflow.python.util.tf_export import tf_export # This value changes every day with an automatic CL. It can be modified in code # via `forward_compatibility_horizon()` or with the environment variable # TF_FORWARD_COMPATIBILITY_DELTA_DAYS, which is added to the compatibility date. -_FORWARD_COMPATIBILITY_HORIZON = datetime.date(2020, 6, 22) +_FORWARD_COMPATIBILITY_HORIZON = datetime.date(2020, 6, 23) _FORWARD_COMPATIBILITY_DELTA_DAYS_VAR_NAME = "TF_FORWARD_COMPATIBILITY_DELTA_DAYS" _FORWARD_COMPATIBILITY_DATE_NUMBER = None From 775b9cb5acffaf86c1ce09d4d1460f40b1392c2a Mon Sep 17 00:00:00 2001 From: Alexander Belyaev Date: Tue, 23 Jun 2020 03:19:37 -0700 Subject: [PATCH 541/867] [XLA][MLIR] Use callback-based builders for (indexed_)generic in (L)HLO->Linalg. PiperOrigin-RevId: 317833249 Change-Id: Ifb60bd3a4797ce360c2a2685e3e96a19dcd6b161 --- .../xla/transforms/xla_legalize_to_linalg.cc | 124 +++++++----------- 1 file changed, 44 insertions(+), 80 deletions(-) diff --git a/tensorflow/compiler/mlir/xla/transforms/xla_legalize_to_linalg.cc b/tensorflow/compiler/mlir/xla/transforms/xla_legalize_to_linalg.cc index 8a2f8ce7d04..e7bb5df8233 100644 --- a/tensorflow/compiler/mlir/xla/transforms/xla_legalize_to_linalg.cc +++ b/tensorflow/compiler/mlir/xla/transforms/xla_legalize_to_linalg.cc @@ -125,32 +125,19 @@ class PointwiseToLinalgConverter : public OpConversionPattern { opResultTypes.push_back(shapedType); } + int64_t args_count = bodyArgTypes.size(); + int64_t results_count = bodyResultTypes.size(); auto linalgOp = rewriter.create( - loc, opResultTypes, args, - /*inputCount=*/bodyArgTypes.size(), - /*outputCount=*/bodyResultTypes.size(), indexing_maps, - GetNParallelLoopsAttrs(nloops)); - - // Add a block to the region. - auto* region = &linalgOp.region(); - auto* block = rewriter.createBlock(region, region->end()); - block->addArguments(bodyArgTypes); - if (isLHLO) block->addArguments(bodyResultTypes); - - SmallVector bodyArgs; - for (int i = 0, e = bodyArgTypes.size(); i < e; ++i) { - bodyArgs.push_back(block->getArgument(i)); - } - - rewriter.setInsertionPointToEnd(block); - // TODO(ravishankarm) : For now use the method in xla_lhlo namespace. That - // method needs to be moved out of there. - Value opResult = xla_lhlo::XlaOpToStdScalarOp::map( - op, bodyResultTypes, bodyArgs, &rewriter); - if (!opResult) { - return failure(); - } - rewriter.create(loc, opResult); + loc, opResultTypes, args, args_count, results_count, indexing_maps, + GetNParallelLoopsAttrs(nloops), + [&](OpBuilder& nestedBuilder, Location nestedLoc, ValueRange args) { + // TODO(ravishankarm) : For now use the method in xla_lhlo namespace. + // That method needs to be moved out of there. + Value opResult = xla_lhlo::XlaOpToStdScalarOp::map( + op, bodyResultTypes, + llvm::to_vector<2>(args.take_front(args_count)), &rewriter); + nestedBuilder.create(loc, opResult); + }); rewriter.replaceOp(op, linalgOp.getOperation()->getResults()); return success(); } @@ -301,27 +288,20 @@ class DataMovementOpConverter : public OpConversionPattern { OpTy op, ArrayRef args, ConversionPatternRewriter& rewriter) const final { if (!verifyXLAOpBufferOrTensorSemantics(op)) return failure(); - auto operandType = op.operand().getType().template cast(); auto resultType = getXLAOpResultType(op); SmallVector indexing_maps = Derived::getIndexingMaps(op, &rewriter); if (indexing_maps.empty()) return failure(); - OpBuilder::InsertionGuard linalgOpGuard(rewriter); auto nloops = resultType.getRank(); auto loc = op.getLoc(); auto linalgOp = rewriter.create( loc, isLHLO ? ArrayRef{} : resultType, args, /*inputCount=*/1, - /*outputCount=*/1, indexing_maps, GetNParallelLoopsAttrs(nloops)); - - auto* region = &linalgOp.region(); - auto* block = rewriter.createBlock(region, region->end()); - block->addArguments(operandType.getElementType()); - if (isLHLO) block->addArgument(resultType.getElementType()); - - rewriter.setInsertionPointToEnd(block); - rewriter.create(loc, block->getArgument(0)); + /*outputCount=*/1, indexing_maps, GetNParallelLoopsAttrs(nloops), + [&](OpBuilder& nestedBuilder, Location nestedLoc, ValueRange args) { + nestedBuilder.create(loc, *args.begin()); + }); rewriter.replaceOp(op, linalgOp.getOperation()->getResults()); return success(); @@ -437,36 +417,26 @@ class LhloBroadcastInDimConverter Value zero = rewriter.create(loc, 0); Value val = rewriter.create(loc, operand, llvm::makeArrayRef({zero})); - auto linalgOp = rewriter.create( + rewriter.create( loc, llvm::None, llvm::makeArrayRef(operand_adaptor.output()), /*inputCount=*/0, /*outputCount=*/1, llvm::makeArrayRef(rewriter.getMultiDimIdentityMap(nloops)), - GetNParallelLoopsAttrs(nloops)); + GetNParallelLoopsAttrs(nloops), + [&](OpBuilder& nestedBuilder, Location nestedLoc, ValueRange args) { + nestedBuilder.create(loc, val); + }); - auto* region = &linalgOp.region(); - auto* block = rewriter.createBlock(region, region->end()); - block->addArgument(result_type.getElementType()); - - rewriter.setInsertionPointToEnd(block); - rewriter.create(loc, val); } else { auto indexing_maps = getIndexingMaps(op, broadcast_dims, result_shape, operand_type, &rewriter); - - OpBuilder::InsertionGuard linalgOpGuard(rewriter); - auto linalgOp = rewriter.create( + rewriter.create( loc, llvm::None, llvm::makeArrayRef({operand, operand_adaptor.output()}), /*inputCount=*/1, /*outputCount=*/1, indexing_maps, - GetNParallelLoopsAttrs(nloops)); - - auto* region = &linalgOp.region(); - auto* block = rewriter.createBlock(region, region->end()); - block->addArguments(operand_type.getElementType()); - block->addArgument(result_type.getElementType()); - - rewriter.setInsertionPointToEnd(block); - rewriter.create(loc, block->getArgument(0)); + GetNParallelLoopsAttrs(nloops), + [&](OpBuilder& nestedBuilder, Location nestedLoc, ValueRange args) { + nestedBuilder.create(loc, *args.begin()); + }); } rewriter.replaceOp(op, llvm::None); return success(); @@ -686,32 +656,26 @@ class IotaConverter : public OpConversionPattern { // Construct the indexing maps needed for linalg.generic ops. unsigned nloops = resultMemrefType.getRank(); - auto loc = iotaOp.getLoc(); - auto linalgOp = rewriter.create( - loc, ArrayRef{}, args, + rewriter.create( + iotaOp.getLoc(), ArrayRef{}, args, 0, // args_in 1, // args_out llvm::makeArrayRef(rewriter.getMultiDimIdentityMap(nloops)), - GetNParallelLoopsAttrs(nloops)); + GetNParallelLoopsAttrs(nloops), + [&](OpBuilder& nestedBuilder, Location nestedLoc, ValueRange ivs, + ValueRange args) { + Value castOp = nestedBuilder.create( + nestedLoc, ivs[iotaOp.iota_dimension().getZExtValue()], + nestedBuilder.getIntegerType( + resultElementType.getIntOrFloatBitWidth())); + if (resultElementType.isa()) { + castOp = nestedBuilder.create(nestedLoc, castOp, + resultElementType); + } + nestedBuilder.create(nestedLoc, castOp); + }); - // Add a block to the region. - auto* region = &linalgOp.region(); - auto* block = rewriter.createBlock(region, region->end()); - for (unsigned i = 0; i < nloops; ++i) { - block->addArgument(rewriter.getIndexType()); - } - block->addArguments(llvm::makeArrayRef(resultElementType)); - - rewriter.setInsertionPointToEnd(block); - Operation* castOp = rewriter.create( - loc, block->getArgument(iotaOp.iota_dimension().getZExtValue()), - rewriter.getIntegerType(resultElementType.getIntOrFloatBitWidth())); - if (resultElementType.isa()) { - castOp = rewriter.create(loc, castOp->getResult(0), - resultElementType); - } - rewriter.create(loc, castOp->getResult(0)); - rewriter.eraseOp(iotaOp); + rewriter.replaceOp(iotaOp, llvm::None); return success(); } }; @@ -867,7 +831,7 @@ struct LhloLegalizeToLinalg auto func = getFunction(); populateLHLOToLinalgConversionPattern(func.getContext(), &patterns); - if (failed(applyPartialConversion(func, target, patterns))) { + if (failed(applyPartialConversion(func, target, patterns, nullptr))) { signalPassFailure(); } } @@ -882,7 +846,7 @@ struct HloLegalizeToLinalg auto func = getFunction(); xla_hlo::populateHLOToLinalgConversionPattern(func.getContext(), &patterns); - if (failed(applyPartialConversion(func, target, patterns))) { + if (failed(applyPartialConversion(func, target, patterns, nullptr))) { signalPassFailure(); } } From 8eea0658d48c4110076b478ac507445e2b2e47b3 Mon Sep 17 00:00:00 2001 From: Adrian Kuegel Date: Tue, 23 Jun 2020 04:51:06 -0700 Subject: [PATCH 542/867] Replace tensorflow command line flags with LLVM command line flags. This avoids a dependency on the framework_internal library, and also LLVM directly supports comma-separated value lists. PiperOrigin-RevId: 317842748 Change-Id: I9207375c753fbab33f6211736d57248c8c3f558a --- .../compiler/mlir/tools/kernel_gen/BUILD | 3 +- .../mlir/tools/kernel_gen/tf_to_cubin.cc | 88 +++++-------------- 2 files changed, 26 insertions(+), 65 deletions(-) diff --git a/tensorflow/compiler/mlir/tools/kernel_gen/BUILD b/tensorflow/compiler/mlir/tools/kernel_gen/BUILD index cebfa7cd9d4..80b597d962d 100644 --- a/tensorflow/compiler/mlir/tools/kernel_gen/BUILD +++ b/tensorflow/compiler/mlir/tools/kernel_gen/BUILD @@ -44,8 +44,9 @@ tf_cc_binary( visibility = ["//tensorflow/core/kernels/cubin_headers:__pkg__"], deps = [ ":cubin_creator", - "//tensorflow/core:framework_internal", + "//tensorflow/compiler/mlir:init_mlir", "//tensorflow/core:lib", "@com_google_absl//absl/strings", + "@llvm-project//llvm:Support", ], ) diff --git a/tensorflow/compiler/mlir/tools/kernel_gen/tf_to_cubin.cc b/tensorflow/compiler/mlir/tools/kernel_gen/tf_to_cubin.cc index 66fcabde0ac..96831689600 100644 --- a/tensorflow/compiler/mlir/tools/kernel_gen/tf_to_cubin.cc +++ b/tensorflow/compiler/mlir/tools/kernel_gen/tf_to_cubin.cc @@ -21,77 +21,37 @@ #include #include -#include "absl/strings/numbers.h" -#include "absl/strings/str_split.h" #include "absl/strings/string_view.h" +#include "llvm/Support/CommandLine.h" +#include "tensorflow/compiler/mlir/init_mlir.h" #include "tensorflow/compiler/mlir/tools/kernel_gen/cubin_creator.h" #include "tensorflow/core/platform/env.h" -#include "tensorflow/core/platform/init_main.h" #include "tensorflow/core/platform/logging.h" -#include "tensorflow/core/util/command_line_flags.h" - -namespace { -bool ParseStringList(std::string string_list, std::vector* result) { - result->clear(); - uint32_t item; - auto items = absl::StrSplit(string_list, ','); - for (const auto& item_str : items) { - if (!absl::SimpleAtoi(item_str, &item)) { - LOG(ERROR) << "Expected token " << item_str << " to be an integer"; - return false; - } - result->push_back(item); - } - return true; -} -} // namespace int main(int argc, char** argv) { - std::string input_file = "foo.mlir"; - std::string output_file = "foo.bin"; - int32_t architecture = 50; - std::vector tile_sizes; - std::vector unroll_factors; - std::vector same_shape; + llvm::cl::opt input_file("input", llvm::cl::desc("input file"), + llvm::cl::value_desc("filename"), + llvm::cl::init("foo.mlir")); + llvm::cl::opt output_file( + "output", llvm::cl::desc("output file"), llvm::cl::value_desc("filename"), + llvm::cl::init("foo.bin")); + llvm::cl::opt architecture( + "arch", llvm::cl::desc("target architecture (e.g. 50 for sm_50)"), + llvm::cl::init(50)); + llvm::cl::list tile_sizes( + "tile_sizes", llvm::cl::desc("tile sizes to use"), llvm::cl::ZeroOrMore, + llvm::cl::CommaSeparated); + llvm::cl::list unroll_factors( + "unroll_factors", + llvm::cl::desc("factors to unroll by, separated by commas"), + llvm::cl::ZeroOrMore, llvm::cl::CommaSeparated); + llvm::cl::list same_shape( + "same_shape", + llvm::cl::desc("arguments with same shape, separated by commas"), + llvm::cl::ZeroOrMore, llvm::cl::CommaSeparated); - auto parse_tile_sizes = [&tile_sizes](std::string tile_sizes_str) { - if (!ParseStringList(tile_sizes_str, &tile_sizes)) { - return false; - } - // Initialize with the default. - if (tile_sizes.empty()) { - tile_sizes.push_back(16); - tile_sizes.push_back(64); - } - return true; - }; - - auto parse_unroll_factors = - [&unroll_factors](std::string unroll_factors_str) { - return ParseStringList(unroll_factors_str, &unroll_factors); - }; - - auto parse_same_shape = [&same_shape](std::string same_shape_str) { - return ParseStringList(same_shape_str, &same_shape); - }; - - std::vector flag_list = { - tensorflow::Flag("input", &input_file, "input file"), - tensorflow::Flag("output", &output_file, "output file"), - tensorflow::Flag("arch", &architecture, - "target architecture (e.g. 50 for sm_50)"), - tensorflow::Flag("tile_sizes", parse_tile_sizes, "16,64", - "tile sizes to use"), - tensorflow::Flag("unroll_factors", parse_unroll_factors, "", - "factors to unroll by, separated by commas"), - tensorflow::Flag("same_shape", parse_same_shape, "", - "arguments with same shape, separated by commas"), - }; - bool parse_ok = tensorflow::Flags::Parse(&argc, argv, flag_list); - tensorflow::port::InitMain("usage", &argc, &argv); - if (!parse_ok) { - return 1; - } + tensorflow::InitMlir y(&argc, &argv); + llvm::cl::ParseCommandLineOptions(argc, argv, "TF op GPU kernel generator\n"); std::pair compute_capability(architecture / 10, architecture % 10); From af94e801cf90fe8fb70d930d8658c20913d09ca5 Mon Sep 17 00:00:00 2001 From: Stephan Herhut Date: Tue, 23 Jun 2020 04:56:42 -0700 Subject: [PATCH 543/867] Use ForLoopSpecializationPass to help with vectorization of kernels. PiperOrigin-RevId: 317843378 Change-Id: I4ead02c24f957269888af5491934567cd3e311fb --- tensorflow/compiler/xla/service/mlir_gpu/kernel_lowering.cc | 5 +++++ 1 file changed, 5 insertions(+) diff --git a/tensorflow/compiler/xla/service/mlir_gpu/kernel_lowering.cc b/tensorflow/compiler/xla/service/mlir_gpu/kernel_lowering.cc index 3f99d40c717..196ea218ef3 100644 --- a/tensorflow/compiler/xla/service/mlir_gpu/kernel_lowering.cc +++ b/tensorflow/compiler/xla/service/mlir_gpu/kernel_lowering.cc @@ -505,6 +505,11 @@ Status LowerLHLOToGPU(mlir::ModuleOp module, LowerLHLOToGPUOptions options) { // Some basic cleanup. pm.addNestedPass<::mlir::FuncOp>(::mlir::createCanonicalizerPass()); pm.addNestedPass<::mlir::FuncOp>(::mlir::createCSEPass()); + // Make loops with min bounds into a conditional plus static bounds. + // Only do this if we unrolled in the first place. + if (!options.unroll_factors.empty()) { + pm.addNestedPass<::mlir::FuncOp>(mlir::createForLoopSpecializationPass()); + } // Approximate of requested. if (options.use_approximations) { pm.addNestedPass<::mlir::FuncOp>( From 7198070f4d032493eeae9741ba9fd868f3d3fb7c Mon Sep 17 00:00:00 2001 From: Tres Popp Date: Tue, 23 Jun 2020 04:57:39 -0700 Subject: [PATCH 544/867] Integrate LLVM at https://github.com/llvm/llvm-project/commit/7a55d984971c PiperOrigin-RevId: 317843475 Change-Id: I2824384ac6a2535fd779ee40301d83004e55b5f3 --- tensorflow/workspace.bzl | 4 ++-- 1 file changed, 2 insertions(+), 2 deletions(-) diff --git a/tensorflow/workspace.bzl b/tensorflow/workspace.bzl index a55dcd17b07..7c47628818b 100755 --- a/tensorflow/workspace.bzl +++ b/tensorflow/workspace.bzl @@ -710,8 +710,8 @@ def tf_repositories(path_prefix = "", tf_repo_name = ""): ) # Check out LLVM and MLIR from llvm-project. - LLVM_COMMIT = "f570d5810485fa6fb2e1009f795a899d79bd429f" - LLVM_SHA256 = "e154a1a97c3b6bead73a32bcb6fc37aac2d80628abee2961315304a4964f04fe" + LLVM_COMMIT = "7a55d984971c11daa55e9423934f98bdc9c04f2f" + LLVM_SHA256 = "6ef6d1f92f51936ed3027433ea26875ac4d2ac8eed88b1fbce472018c4fb7720" LLVM_URLS = [ "https://storage.googleapis.com/mirror.tensorflow.org/github.com/llvm/llvm-project/archive/{commit}.tar.gz".format(commit = LLVM_COMMIT), "https://github.com/llvm/llvm-project/archive/{commit}.tar.gz".format(commit = LLVM_COMMIT), From e071e66f03eb6ae234212400f34061f25e79a699 Mon Sep 17 00:00:00 2001 From: Lev Proleev Date: Tue, 23 Jun 2020 05:29:56 -0700 Subject: [PATCH 545/867] Add support for TENSOR_QUANT8_ASYMM_SIGNED in NNAPI delegate PiperOrigin-RevId: 317846923 Change-Id: I1c61f53e89228cd2482435e9255e390864bd83e3 --- tensorflow/lite/delegates/nnapi/BUILD | 26 + .../delegates/nnapi/acceleration_test_list.cc | 14 +- .../lite/delegates/nnapi/nnapi_delegate.cc | 246 +++-- .../delegates/nnapi/nnapi_delegate_kernel.h | 3 + .../nnapi/nnapi_delegate_mock_test.h | 2 + ...nnapi_delegate_signed_quantization_test.cc | 920 ++++++++++++++++++ tensorflow/lite/nnapi/NeuralNetworksTypes.h | 1 + tensorflow/lite/nnapi/nnapi_handler.h | 22 + 8 files changed, 1154 insertions(+), 80 deletions(-) create mode 100644 tensorflow/lite/delegates/nnapi/nnapi_delegate_signed_quantization_test.cc diff --git a/tensorflow/lite/delegates/nnapi/BUILD b/tensorflow/lite/delegates/nnapi/BUILD index ec9f6907f21..beeaff1b99d 100644 --- a/tensorflow/lite/delegates/nnapi/BUILD +++ b/tensorflow/lite/delegates/nnapi/BUILD @@ -190,6 +190,32 @@ cc_test( ], ) +cc_test( + name = "nnapi_delegate_signed_quantization_test", + size = "small", + srcs = [ + "nnapi_delegate_signed_quantization_test.cc", + ], + tags = [ + "no_mac", + "no_windows", + "tflite_not_portable_ios", + ], + deps = [ + ":nnapi_delegate", + ":nnapi_delegate_mock_test", + "//tensorflow/lite:framework", + "//tensorflow/lite:kernel_api", + "//tensorflow/lite:minimal_logging", + "//tensorflow/lite/c:common", + "//tensorflow/lite/kernels:builtin_ops", + "//tensorflow/lite/kernels:test_util", + "//tensorflow/lite/nnapi:nnapi_implementation", + "//tensorflow/lite/nnapi:nnapi_lib", + "@com_google_googletest//:gtest", + ], +) + cc_test( name = "quant_lstm_sup_test", size = "small", diff --git a/tensorflow/lite/delegates/nnapi/acceleration_test_list.cc b/tensorflow/lite/delegates/nnapi/acceleration_test_list.cc index b20628016f0..31bdc5f8b99 100644 --- a/tensorflow/lite/delegates/nnapi/acceleration_test_list.cc +++ b/tensorflow/lite/delegates/nnapi/acceleration_test_list.cc @@ -60,6 +60,10 @@ FloatActivationsOpTest/Elu,30 FloatActivationsOpTest/HardSwish QuantizedActivationsOpTest/HardSwish QuantizedActivationsOpTest/HardSwishBias +QuantizedActivationsOpTest/Relu* +QuantizedActivationsOpTest/PRelu,29 +QuantizedActivationsOpTest/PReluSameShapes,29 +QuantizedActivationsOpTest/PReluInt8.+,30 # add_test FloatAddOpModel/.+ @@ -145,6 +149,7 @@ ConvolutionOpTest/ConvolutionOpTest/.+/\d+ # dequantize_test DequantizeOpTest/Uint8 +DequantizeOpTest/Int8,30 # depth_to_space_test DepthToSpaceOpModel/Float32 @@ -190,6 +195,7 @@ QuantizedFullyConnectedOpTest/SimpleTestQuantizedOutputMultiplierGreaterThan1Uin QuantizedFullyConnectedOpTest/SimpleTestQuantizedOutputMultiplierGreaterThan1Int8/\d+,29 HybridFullyConnectedOpTest/SimpleTestQuantizedUint8,29 HybridFullyConnectedOpTest/SimpleTestQuantizedInt8,29 +HybridAsymmetricInputFullyConnectedOpTest.SimpleTestQuantizedUint8,29 FloatFullyConnectedOpTest/FloatFullyConnectedOpTest/SimpleTest4DInput/\d+ QuantizedFullyConnectedOpTest/QuantizedFullyConnectedOpTest/SimpleTest4dInputQuantizedUint8/\d+ QuantizedFullyConnectedOpTest/QuantizedFullyConnectedOpTest/SimpleTest4dInputQuantizedOutputMultiplierGreaterThan1Uint8/\d+,29 @@ -207,6 +213,7 @@ FloatGatherOpTest/LastAxis,29 TypesGatherOpTest/Float32Int32,29 TypesGatherOpTest/Int32Int32,29 TypesGatherOpTest/Uint8Int32,29 +TypesGatherOpTest/Int8Int32,29 # hashtable_lookup_test # All test excepted the string one should be accelerated @@ -286,13 +293,18 @@ QuantizedLstmTest/BasicQuantizedLstmTest/29 # quantize_test QuantizeOpTest/UINT8,29 +QuantizeOpTest/INT8,30 + +# rank # reduce_test -Dynamic.+(Mean|Sum|Prod|Max|Min)OpTest/.+ -ConstUint8(Mean|Sum)OpTest/.+ +-ConstInt8MeanOpTest.NonSpecialAxisNonSameScale +-ConstInt8MeanOpTest.QuantizedDifferentScale ConstUint8(Max|Min)OpTest/.+,29 ConstUint8(Mean)OpTest/.+ -Constint8(Mean|Max|Min)OpTest/.+ +ConstInt8(Mean|Max|Min)OpTest/.+,29 ConstFloat(Sum|Prod|Max|Min)OpTest/NotKeepDims,29 ConstFloat(Sum|Prod|Max|Min)OpTest/KeepDims,29 ConstFloat(Mean|Any)OpTest/NotKeepDims diff --git a/tensorflow/lite/delegates/nnapi/nnapi_delegate.cc b/tensorflow/lite/delegates/nnapi/nnapi_delegate.cc index 1c35ee370c2..58ab13ab657 100644 --- a/tensorflow/lite/delegates/nnapi/nnapi_delegate.cc +++ b/tensorflow/lite/delegates/nnapi/nnapi_delegate.cc @@ -201,6 +201,7 @@ bool NeedInt8Conversion(const TfLiteContext* context, int builtin_code, case kTfLiteBuiltinConcatenation: case kTfLiteBuiltinEqual: case kTfLiteBuiltinExpandDims: + case kTfLiteBuiltinGather: case kTfLiteBuiltinGreater: case kTfLiteBuiltinGreaterEqual: case kTfLiteBuiltinHardSwish: @@ -377,6 +378,7 @@ bool HasZeroes(TfLiteIntArrayView array) { enum { NN_TENSOR_FLAG_SCALAR_AS_TENSOR = 1U << 0, NN_TENSOR_FLAG_INT8_CONVERSION = 1U << 1, + NN_TENSOR_FLAG_USE_INT8_ASYMM_SIGNED = 1U << 2, }; // Returns the SDK level to target when delegating to the given devices. @@ -1065,6 +1067,8 @@ class NNAPIOpBuilder { tensor_flags & NN_TENSOR_FLAG_SCALAR_AS_TENSOR; const bool need_int8_conversion = tensor_flags & NN_TENSOR_FLAG_INT8_CONVERSION; + const bool use_int8_asymm_signed = + tensor_flags & NN_TENSOR_FLAG_USE_INT8_ASYMM_SIGNED; int ann_tensor_index = operand_mapping_->lite_index_to_ann(tensor_index); if (ann_tensor_index != -1) { indices->push_back(ann_tensor_index); @@ -1095,12 +1099,25 @@ class NNAPIOpBuilder { nn_type = ANEURALNETWORKS_TENSOR_FLOAT32; break; case kTfLiteUInt8: + nn_type = ANEURALNETWORKS_TENSOR_QUANT8_ASYMM; + scale = tensor->params.scale; + zeroPoint = tensor->params.zero_point; + if (scale == 0) { + // ANEURALNETWORKS_TENSOR_QUANT8_ASYMM with zero scale is not valid in + // NNAPI. + scale = 1; + } + break; case kTfLiteInt8: // If explicit int8 conversion is needed, we still need // ANEURALNETWORKS_TENSOR_QUANT8_ASYMM type. - nn_type = (tensor_type == kTfLiteUInt8 || need_int8_conversion) - ? ANEURALNETWORKS_TENSOR_QUANT8_ASYMM - : ANEURALNETWORKS_TENSOR_QUANT8_SYMM; + if (use_int8_asymm_signed) { + nn_type = ANEURALNETWORKS_TENSOR_QUANT8_ASYMM_SIGNED; + } else if (need_int8_conversion) { + nn_type = ANEURALNETWORKS_TENSOR_QUANT8_ASYMM; + } else { + nn_type = ANEURALNETWORKS_TENSOR_QUANT8_SYMM; + } scale = tensor->params.scale; zeroPoint = tensor->params.zero_point; if (tensor->quantization.type == kTfLiteAffineQuantization) { @@ -1130,8 +1147,7 @@ class NNAPIOpBuilder { operand_mapping_->add_type_conversion(tensor_index, kTfLiteUInt8); } if (scale == 0) { - // TENSOR_QUANT8_ASYMM and ANEURALNETWORKS_TENSOR_QUANT8_ASYMM - // with zero scale are not valid in NNAPI. + // QUANT8 tensors with zero scale are not valid in NNAPI. scale = 1; } } @@ -1248,7 +1264,6 @@ class NNAPIOpBuilder { "setting new operand value", nnapi_errno_); } } - indices->push_back(ann_tensor_index); return kTfLiteOk; } @@ -1437,7 +1452,6 @@ bool NNAPIDelegateKernel::Validate( bool is_accelerator_specified, std::vector* map_failures) { OpValidationContext val_ctx{true, map_failures}; - switch (builtin_code) { case kTfLiteBuiltinAdd: { ExpectMaxOpVersion(version, 2, &val_ctx); @@ -1789,18 +1803,21 @@ bool NNAPIDelegateKernel::Validate( "Supported op versions are 1 and 2 only", &val_ctx); const auto& input = context->tensors[node->inputs->data[0]]; - Expect(input.type != kTfLiteFloat16, - NNAPIValidationFailureType::kUnsupportedInputType, - "kTfLiteFloat16 not supported as input", &val_ctx); + if (android_sdk_version < kMinSdkVersionForNNAPI12) { + EXPECT_INPUT_TYPE_IN(input.type, kTfLiteUInt8); + } else { + EXPECT_INPUT_TYPE_IN(input.type, kTfLiteUInt8, kTfLiteInt8); - const auto zero_point = input.params.zero_point; - Expect(input.type != kTfLiteInt8 || - (zero_point == 0 && - android_sdk_version >= kMinSdkVersionForNNAPI12), - NNAPIValidationFailureType::kUnsupportedInputType, - "NN API supports int8 type since version 1.2 but only for " - "symmetric quantization.", - &val_ctx); + if (android_sdk_version == kMinSdkVersionForNNAPI12 && + input.type == kTfLiteInt8) { + const auto zero_point = input.params.zero_point; + Expect(zero_point == 0, + NNAPIValidationFailureType::kUnsupportedInputType, + "NN API supports int8 type since version 1.2 but only for " + "symmetric quantization.", + &val_ctx); + } + } } break; case kTfLiteBuiltinFloor: { ExpectOpVersion(version, 1, &val_ctx); @@ -2150,21 +2167,38 @@ bool NNAPIDelegateKernel::Validate( &val_ctx); const TfLiteType input_type = context->tensors[node->inputs->data[0]].type; - EXPECT_INPUT_TYPE_IN(input_type, kTfLiteFloat32, kTfLiteInt32, - kTfLiteUInt8); const TfLiteType output_type = context->tensors[node->outputs->data[0]].type; - ExpectTypeIn(output_type, {kTfLiteFloat32, kTfLiteInt32, kTfLiteUInt8}, - NNAPIValidationFailureType::kUnsupportedOutputType, - "Output type should be one of kTfLiteFloat32, kTfLiteInt32, " - "kTfLiteUInt8.", - &val_ctx); + if (android_sdk_version >= kMinSdkVersionForNNAPI13) { + EXPECT_INPUT_TYPE_IN(input_type, kTfLiteFloat32, kTfLiteInt32, + kTfLiteUInt8, kTfLiteInt8); + + ExpectTypeIn( + output_type, + {kTfLiteFloat32, kTfLiteInt32, kTfLiteUInt8, kTfLiteInt8}, + NNAPIValidationFailureType::kUnsupportedOutputType, + "Output type should be one of kTfLiteFloat32, kTfLiteInt32, " + "kTfLiteUInt8, kTfLiteInt8.", + &val_ctx); + } else { + EXPECT_INPUT_TYPE_IN(input_type, kTfLiteFloat32, kTfLiteInt32, + kTfLiteUInt8); + + ExpectTypeIn( + output_type, {kTfLiteFloat32, kTfLiteInt32, kTfLiteUInt8}, + NNAPIValidationFailureType::kUnsupportedOutputType, + "Output type should be one of kTfLiteFloat32, kTfLiteInt32, " + "kTfLiteUInt8.", + &val_ctx); + } } break; case kTfLiteBuiltinPrelu: { ExpectOpVersion(version, 1, &val_ctx); ExpectMinAndroidSdkVersion(android_sdk_version, kMinSdkVersionForNNAPI12, &val_ctx); - ExpectIsFloatOrUint8Operator(context, node, &val_ctx); + const auto input_type = context->tensors[node->inputs->data[0]].type; + EXPECT_INPUT_TYPE_IN(input_type, kTfLiteFloat32, kTfLiteUInt8, + kTfLiteInt8); } break; case kTfLiteBuiltinTile: { ExpectOpVersion(version, 1, &val_ctx); @@ -2240,19 +2274,18 @@ bool NNAPIDelegateKernel::Validate( &val_ctx); } break; case kTfLiteBuiltinGather: { - ExpectOpVersion(version, 1, &val_ctx); + ExpectOpVersion(version, 2, &val_ctx); ExpectMinAndroidSdkVersion(android_sdk_version, kMinSdkVersionForNNAPI12, &val_ctx); const auto input_type = context->tensors[node->inputs->data[0]].type; const auto& positions = context->tensors[node->inputs->data[1]]; + EXPECT_INPUT_TYPE_IN(input_type, kTfLiteFloat32, kTfLiteFloat16, - kTfLiteInt32, kTfLiteUInt8); - ExpectTypeIn(positions.type, - {kTfLiteFloat32, kTfLiteFloat16, kTfLiteInt32, kTfLiteUInt8}, - NNAPIValidationFailureType::kUnsupportedInputType, - "Positions type should be one of kTfLiteFloat32, " - "kTfLiteFloat16, kTfLiteInt32, kTfLiteUInt8", - &val_ctx); + kTfLiteInt32, kTfLiteUInt8, kTfLiteInt8); + + Expect(positions.type == kTfLiteInt32, + NNAPIValidationFailureType::kUnsupportedInputType, + "Positions type should be one of kTfLiteInt32", &val_ctx); Expect(positions.dims->size != 0, NNAPIValidationFailureType::kUnsupportedOperandRank, "0-dimension args are not supported by NNAPI.", &val_ctx); @@ -2283,8 +2316,13 @@ bool NNAPIDelegateKernel::Validate( &val_ctx); // Tensor indices: split_dim: 0, value: 1 const TfLiteTensor& input = context->tensors[node->inputs->data[1]]; - EXPECT_INPUT_TYPE_IN(input.type, kTfLiteFloat32, kTfLiteUInt8, - kTfLiteInt32); + if (android_sdk_version >= kMinSdkVersionForNNAPI13) { + EXPECT_INPUT_TYPE_IN(input.type, kTfLiteFloat32, kTfLiteUInt8, + kTfLiteInt8, kTfLiteInt32); + } else { + EXPECT_INPUT_TYPE_IN(input.type, kTfLiteFloat32, kTfLiteUInt8, + kTfLiteInt32); + } const TfLiteTensor& axis = context->tensors[node->inputs->data[0]]; Expect(axis.type == kTfLiteInt32 && axis.allocation_type == kTfLiteMmapRo, NNAPIValidationFailureType::kUnsupportedInputType, @@ -2308,30 +2346,41 @@ bool NNAPIDelegateKernel::Validate( NNAPIValidationFailureType::kUnsupportedInputType, "Value should be Float32.", &val_ctx); const auto output_type = context->tensors[node->outputs->data[0]].type; - Expect(output_type == kTfLiteUInt8, - NNAPIValidationFailureType::kUnsupportedOutputType, - "Output should be kTfLiteUInt8.", &val_ctx); + if (android_sdk_version < kMinSdkVersionForNNAPI13) { + Expect(output_type == kTfLiteUInt8, + NNAPIValidationFailureType::kUnsupportedOutputType, + "Output should be kTfLiteUInt8.", &val_ctx); + } else { + ExpectTypeIn(output_type, {kTfLiteUInt8, kTfLiteInt8}, + NNAPIValidationFailureType::kUnsupportedOutputType, + "Output should be kTfLiteUInt8.", &val_ctx); + } const auto quantization_params = context->tensors[node->outputs->data[0]].params; Expect(quantization_params.scale > 0.f, NNAPIValidationFailureType::kUnsupportedQuantizationParameters, "Quantization scale should be > 0.", &val_ctx); } break; - case kTfLiteBuiltinReduceAny: - case kTfLiteBuiltinReduceMin: - case kTfLiteBuiltinReduceMax: { - ExpectOpVersion(version, 1, &val_ctx); + case kTfLiteBuiltinReduceAny: { + ExpectOpVersion(version, 2, &val_ctx); ExpectMinAndroidSdkVersion(android_sdk_version, kMinSdkVersionForNNAPI12, &val_ctx); Expect(context->tensors[node->outputs->data[0]].dims->size != 0, NNAPIValidationFailureType::kUnsupportedOutputType, "NNAPI does not support generating a scalar as output.", &val_ctx); - if (builtin_code == kTfLiteBuiltinReduceProd) { - const auto input_type = context->tensors[node->inputs->data[0]].type; - Expect(input_type == kTfLiteFloat32, - NNAPIValidationFailureType::kUnsupportedInputType, - "NNAPI only supports floating point REDUCE_PROD.", &val_ctx); - } + } break; + case kTfLiteBuiltinReduceMin: + case kTfLiteBuiltinReduceMax: { + ExpectMaxOpVersion(version, 2, &val_ctx); + ExpectMinAndroidSdkVersion(android_sdk_version, kMinSdkVersionForNNAPI12, + &val_ctx); + const auto input_tensor = context->tensors[node->inputs->data[0]]; + const auto input_type = input_tensor.type; + EXPECT_INPUT_TYPE_IN(input_type, kTfLiteFloat32, kTfLiteUInt8, + kTfLiteInt8); + Expect(input_tensor.dims->size != 0, + NNAPIValidationFailureType::kUnsupportedOutputType, + "NNAPI does not support generating a scalar as output.", &val_ctx); } break; case kTfLiteBuiltinDepthToSpace: { const TfLiteType input_type = @@ -3093,16 +3142,10 @@ TfLiteStatus NNAPIDelegateKernel::Map( case kTfLiteBuiltinGather: { auto builtin = reinterpret_cast( mapping_args.node->builtin_data); - mapping_args.builder->AddTensorInput(mapping_args.node->inputs->data[0], - /* hybrid_op */ false, - /* scalar_as_tensor */ false); - mapping_args.builder->AddScalarInt32Operand(builtin->axis); - mapping_args.builder->AddTensorInput(mapping_args.node->inputs->data[1], /* hybrid_op */ false, - /* scalar_as_tensor */ false); - + /* tensor_flags */ 0); *nn_op_type = ANEURALNETWORKS_GATHER; } break; case kTfLiteBuiltinBidirectionalSequenceLstm: { @@ -3430,6 +3473,9 @@ TfLiteStatus NNAPIDelegateKernel::Invoke(TfLiteContext* context, // absolute indices but NN api indices inputs by relative indices. int relative_input_index = 0; + const bool use_int8_asymm_signed = + target_sdk_version_ >= kMinSdkVersionForNNAPI13; + size_t input_offset = 0; for (auto absolute_input_index : TfLiteIntArrayView(node->inputs)) { if (absolute_input_index == kTfLiteOptionalTensor) { @@ -3472,9 +3518,16 @@ TfLiteStatus NNAPIDelegateKernel::Invoke(TfLiteContext* context, } } else if (tensor->type == kTfLiteInt8 && ann_type_equivalent == kTfLiteInt32) { - for (int i = 0; i < num_elements; ++i) { - reinterpret_cast(input_ptr)[i] = - static_cast(tensor->data.int8[i]) + 128; + if (use_int8_asymm_signed) { + for (int i = 0; i < num_elements; ++i) { + reinterpret_cast(input_ptr)[i] = + static_cast(tensor->data.int8[i]); + } + } else { + for (int i = 0; i < num_elements; ++i) { + reinterpret_cast(input_ptr)[i] = + static_cast(tensor->data.int8[i]) + 128; + } } } else { context->ReportError( @@ -3685,6 +3738,15 @@ TfLiteStatus NNAPIDelegateKernel::AddOpsAndTensors(TfLiteContext* context, &dequantize_mapping, &allocation_memory_mapping_, &nnapi_to_tflite_op_mapping_, nn_model_.get(), nnapi_errno); + + // If we have target accelerators the target SDK version might be + // different than the current android version. + target_sdk_version_ = nnapi_->android_sdk_version; + if (!nnapi_devices_.empty()) { + TF_LITE_ENSURE_STATUS(GetTargetSdkVersion( + context, nnapi_, nnapi_devices_, &target_sdk_version_, nnapi_errno)); + } + // Add Tensors. for (auto node_index : nodes_) { // Obtain the op and registration. @@ -3696,11 +3758,18 @@ TfLiteStatus NNAPIDelegateKernel::AddOpsAndTensors(TfLiteContext* context, const bool hybrid_op = IsHybridOperator(context, reg->builtin_code, node); const bool scalar_as_tensor = IsScalarInputSupported(reg->builtin_code); const bool need_int8_conversion = + target_sdk_version_ < kMinSdkVersionForNNAPI13 && NeedInt8Conversion(context, reg->builtin_code, node); + const bool use_int8_asymm_signed = + target_sdk_version_ >= kMinSdkVersionForNNAPI13 && !hybrid_op; + int input_tensor_flags = 0; if (scalar_as_tensor) { input_tensor_flags |= NN_TENSOR_FLAG_SCALAR_AS_TENSOR; } + if (use_int8_asymm_signed) { + input_tensor_flags |= NN_TENSOR_FLAG_USE_INT8_ASYMM_SIGNED; + } // On SDK level less than 30, h_swish will be lowered into supported NNAPI // operations. Since SDK level 30, h_swish is supported as a single @@ -3807,8 +3876,12 @@ TfLiteStatus NNAPIDelegateKernel::AddOpsAndTensors(TfLiteContext* context, break; case kTfLiteInt8: if (constant_value.allocation_type == kTfLiteMmapRo) { - builder.AddScalarInt32Operand( - static_cast(*constant_value.data.int8) + 128); + if (need_int8_conversion) { + builder.AddScalarInt32Operand( + static_cast(*constant_value.data.int8) + 128); + } else { + builder.AddScalarInt32Operand(*constant_value.data.int8); + } } else { builder.AddSingleValueTensorAsScalarOperand( constant_value_id, ANEURALNETWORKS_INT32); @@ -3836,7 +3909,8 @@ TfLiteStatus NNAPIDelegateKernel::AddOpsAndTensors(TfLiteContext* context, // specifying the output height and width, is not added and // instead the height and width will be added individually as // scalars by the mapping function returned by Map(). - TF_LITE_ENSURE_STATUS(builder.AddTensorInput(input_index, hybrid_op)); + TF_LITE_ENSURE_STATUS(builder.AddTensorInput(input_index, hybrid_op, + input_tensor_flags)); } } else if (reg->builtin_code == kTfLiteBuiltinTopkV2 && input_pos > 0) { // The K parameter tensor is not handled here but by the functor @@ -3844,8 +3918,12 @@ TfLiteStatus NNAPIDelegateKernel::AddOpsAndTensors(TfLiteContext* context, // the else clause below continue; } else if (reg->builtin_code == kTfLiteBuiltinGather) { - // Everything is added during Map since input tensors + // Everything else is added during Map since input tensors // have different order. + if (input_pos == 0) { + TF_LITE_ENSURE_STATUS(builder.AddTensorInput(input_index, hybrid_op, + input_tensor_flags)); + } continue; } else if (reg->builtin_code == kTfLiteBuiltinExpandDims && input_pos == 1) { @@ -3862,7 +3940,8 @@ TfLiteStatus NNAPIDelegateKernel::AddOpsAndTensors(TfLiteContext* context, // the axis, needs to be converted to a scalar since TFLite uses a // tensor but NNAPI uses a scalar as the axis. if (input_pos == 0) { - TF_LITE_ENSURE_STATUS(builder.AddTensorInput(input_index, hybrid_op)); + TF_LITE_ENSURE_STATUS(builder.AddTensorInput(input_index, hybrid_op, + input_tensor_flags)); } else { const int axis_id = node->inputs->data[1]; const TfLiteTensor& axis_tensor = context->tensors[axis_id]; @@ -3908,12 +3987,26 @@ TfLiteStatus NNAPIDelegateKernel::AddOpsAndTensors(TfLiteContext* context, std::vector(1, operand_tensor.data.uint8[0]), operand_tensor.params, &tensor_index)); break; - case kTfLiteInt8: - TF_LITE_ENSURE_STATUS(builder.AddNewInputConstantTensor( - ANEURALNETWORKS_TENSOR_QUANT8_SYMM, operand_tensor.type, {1}, - std::vector(1, operand_tensor.data.int8[0]), - operand_tensor.params, &tensor_index)); - break; + case kTfLiteInt8: { + auto params = operand_tensor.params; + if (params.scale == 0.0) { + params.scale = 1.0; + } + + if (use_int8_asymm_signed) { + TF_LITE_ENSURE_STATUS(builder.AddNewInputConstantTensor( + ANEURALNETWORKS_TENSOR_QUANT8_ASYMM_SIGNED, + operand_tensor.type, {1}, + std::vector(1, operand_tensor.data.int8[0]), params, + &tensor_index)); + } else { + TF_LITE_ENSURE_STATUS(builder.AddNewInputConstantTensor( + ANEURALNETWORKS_TENSOR_QUANT8_ASYMM, operand_tensor.type, + {1}, + std::vector(1, operand_tensor.data.int8[0] + 128), + params, &tensor_index)); + } + } break; case kTfLiteInt32: TF_LITE_ENSURE_STATUS(builder.AddNewInputConstantTensor( ANEURALNETWORKS_TENSOR_INT32, operand_tensor.type, {1}, @@ -3995,19 +4088,11 @@ TfLiteStatus NNAPIDelegateKernel::AddOpsAndTensors(TfLiteContext* context, } } - // If we have target accelerators the target SDK version might be - // different than the current android version. - int target_sdk_version = nnapi_->android_sdk_version; - if (!nnapi_devices_.empty()) { - TF_LITE_ENSURE_STATUS(GetTargetSdkVersion( - context, nnapi_, nnapi_devices_, &target_sdk_version, nnapi_errno)); - } - // Get op type and operands // Fails if the Validate function failed int nn_op_type; TF_LITE_ENSURE_STATUS( - Map(context, reg->builtin_code, reg->version, target_sdk_version, + Map(context, reg->builtin_code, reg->version, target_sdk_version_, {context, &builder, node, &model_state_outputs_, &model_state_tfl_inputs_, &feedback_loops_, nnapi_errno}, &nn_op_type)); @@ -4017,6 +4102,9 @@ TfLiteStatus NNAPIDelegateKernel::AddOpsAndTensors(TfLiteContext* context, if (need_int8_conversion) { output_tensor_flags |= NN_TENSOR_FLAG_INT8_CONVERSION; } + if (use_int8_asymm_signed) { + output_tensor_flags |= NN_TENSOR_FLAG_USE_INT8_ASYMM_SIGNED; + } for (int output_pos = 0; output_pos < node->outputs->size; ++output_pos) { const auto output_index = node->outputs->data[output_pos]; diff --git a/tensorflow/lite/delegates/nnapi/nnapi_delegate_kernel.h b/tensorflow/lite/delegates/nnapi/nnapi_delegate_kernel.h index 26822c011e3..9aa0f303cc2 100644 --- a/tensorflow/lite/delegates/nnapi/nnapi_delegate_kernel.h +++ b/tensorflow/lite/delegates/nnapi/nnapi_delegate_kernel.h @@ -341,6 +341,9 @@ class NNAPIDelegateKernel { std::vector nnapi_to_tflite_op_mapping_; + // Fully initialized in NNAPIDelegateKernel::AddOpsAndTensors + int target_sdk_version_ = 27; // kMinSdkVersionForNNAPI13 + void AddDequantizeOperatorsWhereNeeded( const TfLiteContext* context, int builtin_code, const TfLiteNode* node, int tflite_node_index, NNAPIOpBuilder* builder, int* nnapi_errno); diff --git a/tensorflow/lite/delegates/nnapi/nnapi_delegate_mock_test.h b/tensorflow/lite/delegates/nnapi/nnapi_delegate_mock_test.h index fa7ff9dd1f1..5dbe4110131 100644 --- a/tensorflow/lite/delegates/nnapi/nnapi_delegate_mock_test.h +++ b/tensorflow/lite/delegates/nnapi/nnapi_delegate_mock_test.h @@ -71,6 +71,8 @@ class NnApiMock : public ::tflite::nnapi::NnApiHandler { ExecutionComputeReturns(); ExecutionStartComputeReturns(); EventWaitReturns(); + SetPriorityReturns(); + SetOperandSymmPerChannelQuantParamsReturns(); SetNnapiSupportedDevice("test-device", android_sdk_version); } diff --git a/tensorflow/lite/delegates/nnapi/nnapi_delegate_signed_quantization_test.cc b/tensorflow/lite/delegates/nnapi/nnapi_delegate_signed_quantization_test.cc new file mode 100644 index 00000000000..b9d702015c2 --- /dev/null +++ b/tensorflow/lite/delegates/nnapi/nnapi_delegate_signed_quantization_test.cc @@ -0,0 +1,920 @@ +/* Copyright 2020 The TensorFlow Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. +==============================================================================*/ +#include +#include "tensorflow/lite/builtin_ops.h" +#include "tensorflow/lite/c/common.h" +#include "tensorflow/lite/delegates/nnapi/nnapi_delegate.h" +#include "tensorflow/lite/delegates/nnapi/nnapi_delegate_mock_test.h" +#include "tensorflow/lite/interpreter.h" +#include "tensorflow/lite/kernels/fully_connected.h" +#include "tensorflow/lite/kernels/test_util.h" +#include "tensorflow/lite/minimal_logging.h" +#include "tensorflow/lite/model.h" +#include "tensorflow/lite/nnapi/NeuralNetworksTypes.h" +#include "tensorflow/lite/nnapi/nnapi_implementation.h" + +namespace tflite { + +namespace ops { +namespace builtin { + +TfLiteRegistration* Register_CONVOLUTION_REF(); +TfLiteRegistration* Register_DEQUANTIZE(); + +} // namespace builtin +} // namespace ops + +namespace { + +class SingleOpModelWithNNAPI : public SingleOpModel { + public: + SingleOpModelWithNNAPI() = default; + void Init(const NnApi* nnapi) { + stateful_delegate_.reset(new StatefulNnApiDelegate(nnapi)); + SetDelegate(stateful_delegate_.get()); + } + + StatefulNnApiDelegate* GetDelegate() { return stateful_delegate_.get(); } + + void SetBufferHandle(int index, TfLiteBufferHandle handle) { + interpreter_->SetBufferHandle(index, handle, stateful_delegate_.get()); + } + TfLiteStatus GetCompilationStatus() { return compilation_status_; } + + protected: + std::unique_ptr stateful_delegate_; + TfLiteStatus compilation_status_; +}; + +class HybridFullyConnectedOpModel : public SingleOpModelWithNNAPI { + public: + HybridFullyConnectedOpModel(const NnApi* nnapi, int units, int batches, + const TensorData& input, + const TensorData& weights, + const TensorData& output = {TensorType_FLOAT32}, + bool asymmetric_inputs = false) + : batches_(batches), units_(units) { + SingleOpModelWithNNAPI::Init(nnapi); + int total_input_size = 1; + for (size_t i = 0; i < input.shape.size(); ++i) { + total_input_size *= input.shape[i]; + } + input_size_ = total_input_size / batches_; + + input_ = AddInput(input); + weights_ = AddInput(weights); + + TensorData bias{TensorType_FLOAT32, {units_}}; + bias_ = AddInput(bias); + + output_ = AddOutput(output); + + auto options = CreateFullyConnectedOptions( + builder_, ActivationFunctionType_RELU, + tflite::FullyConnectedOptionsWeightsFormat_DEFAULT, + false, asymmetric_inputs) + .Union(); + SetBuiltinOp(BuiltinOperator_FULLY_CONNECTED, + BuiltinOptions_FullyConnectedOptions, options); + resolver_ = absl::make_unique( + BuiltinOperator_FULLY_CONNECTED, + ops::builtin::Register_FULLY_CONNECTED_PIE()); + BuildInterpreter({GetShape(input_), GetShape(weights_), GetShape(bias_)}, + /*num_threads=*/-1, + /* allow_fp32_relax_to_fp16 */ false, + /*apply_delegate=*/false); + compilation_status_ = ApplyDelegate(); + } + void SetBias(const std::vector& f) { PopulateTensor(bias_, f); } + void SetWeights(const std::vector& data) { + SymmetricQuantizeAndPopulate(weights_, data); + } + void SetSignedWeights(std::initializer_list f) { + SignedSymmetricQuantizeAndPopulate(weights_, f); + } + + void SetInput(const std::vector& f) { PopulateTensor(input_, f); } + std::vector GetOutput() { return ExtractVector(output_); } + std::vector GetOutputShape() { return GetTensorShape(output_); } + + int input_size() { return input_size_; } + int num_units() { return units_; } + int num_batches() { return batches_; } + + protected: + int input_; + int weights_; + int bias_; + int output_; + + int batches_; + int units_; + int input_size_; +}; + +struct NnApiSignedQuantizationTest + : ::tflite::delegate::nnapi::NnApiDelegateMockTest { + static void SetUpTestSuite() { tensors_count = new std::map(); } + void SetUp() override { + ::tflite::delegate::nnapi::NnApiDelegateMockTest::SetUp(); + nnapi_mock_->StubAddOperandWith( + [](ANeuralNetworksModel* model, + const ANeuralNetworksOperandType* type) -> int { + const auto nn_tensor_type = type->type; + if (tensors_count->find(nn_tensor_type) == tensors_count->end()) { + tensors_count->insert({nn_tensor_type, 0}); + } + tensors_count->at(nn_tensor_type)++; + return ANEURALNETWORKS_NO_ERROR; + }); + } + void TearDown() override { tensors_count->clear(); } + static void TearDownTestSuite() { + delete tensors_count; + tensors_count = nullptr; + } + static std::map* tensors_count; +}; +std::map* NnApiSignedQuantizationTest::tensors_count = nullptr; + +TEST_F(NnApiSignedQuantizationTest, + HybridFullyConnectedMapsToSignedSymmOnSdk29) { + nnapi_mock_->SetAndroidSdkVersion(29); + + HybridFullyConnectedOpModel m( + nnapi_mock_->GetNnApi(), /*units=*/3, /*batches=*/2, + /*input=*/{TensorType_FLOAT32, {2, 10}}, + /*weights=*/{TensorType_INT8, {3, 10}, 0, 0, 10.0 / 127.0, 0}); + m.SetSignedWeights({ + 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, // u = 0 + 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, // u = 1 + 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, // u = 2 + }); + m.SetBias({1, 2, 3}); + m.SetInput({ + 1, 2, 3, 4, 5, 6, 7, 8, -9, -10, // b = 0 + 1, 2, 3, 4, 5, 6, 7, -8, 9, -10, // b = 1 + }); + + m.Invoke(); + EXPECT_EQ(m.GetCompilationStatus(), kTfLiteOk); + + ASSERT_EQ(tensors_count->size(), 3); + ASSERT_NE(tensors_count->find(ANEURALNETWORKS_TENSOR_FLOAT32), + tensors_count->end()); + ASSERT_NE(tensors_count->find(ANEURALNETWORKS_INT32), tensors_count->end()); + ASSERT_NE(tensors_count->find(ANEURALNETWORKS_TENSOR_QUANT8_SYMM), + tensors_count->end()); + + EXPECT_EQ(tensors_count->at(ANEURALNETWORKS_TENSOR_FLOAT32), + 4); // fc_input, fc_weights, fc_bias, fc_output + EXPECT_EQ(tensors_count->at(ANEURALNETWORKS_INT32), 1); // activation + EXPECT_EQ(tensors_count->at(ANEURALNETWORKS_TENSOR_QUANT8_SYMM), + 1); // dequantize_weights_input +} + +TEST_F(NnApiSignedQuantizationTest, + HybridFullyConnectedMapsToSignedSymmOnSdk30) { + nnapi_mock_->SetAndroidSdkVersion(30); + + HybridFullyConnectedOpModel m( + nnapi_mock_->GetNnApi(), /*units=*/3, /*batches=*/2, + /*input=*/{TensorType_FLOAT32, {2, 10}}, + /*weights=*/{TensorType_INT8, {3, 10}, 0, 0, 10.0 / 127.0, 0}); + m.SetSignedWeights({ + 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, // u = 0 + 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, // u = 1 + 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, // u = 2 + }); + m.SetBias({1, 2, 3}); + m.SetInput({ + 1, 2, 3, 4, 5, 6, 7, 8, -9, -10, // b = 0 + 1, 2, 3, 4, 5, 6, 7, -8, 9, -10, // b = 1 + }); + + m.Invoke(); + EXPECT_EQ(m.GetCompilationStatus(), kTfLiteOk); + + ASSERT_EQ(tensors_count->size(), 3); + ASSERT_NE(tensors_count->find(ANEURALNETWORKS_TENSOR_FLOAT32), + tensors_count->end()); + ASSERT_NE(tensors_count->find(ANEURALNETWORKS_INT32), tensors_count->end()); + ASSERT_NE(tensors_count->find(ANEURALNETWORKS_TENSOR_QUANT8_SYMM), + tensors_count->end()); + + EXPECT_EQ(tensors_count->at(ANEURALNETWORKS_TENSOR_FLOAT32), + 4); // fc_input, fc_weights, fc_bias, fc_output + EXPECT_EQ(tensors_count->at(ANEURALNETWORKS_INT32), 1); // activation + EXPECT_EQ(tensors_count->at(ANEURALNETWORKS_TENSOR_QUANT8_SYMM), + 1); // dequantize_weights_input +} + +template +class BaseConvolutionOpModel : public SingleOpModelWithNNAPI { + public: + BaseConvolutionOpModel( + const NnApi* nnapi, TfLiteRegistration* registration, + const TensorData& input, const TensorData& filter, + const TensorData& output, int stride_width = 2, int stride_height = 2, + enum Padding padding = Padding_VALID, + enum ActivationFunctionType activation = ActivationFunctionType_NONE, + int dilation_width_factor = 1, int dilation_height_factor = 1, + std::initializer_list filter_data = {}) { + SingleOpModelWithNNAPI::Init(nnapi); + + input_ = AddInput(input); + + if (filter_data.size()) { + filter_ = AddConstInput(filter, filter_data); + } else { + filter_ = AddInput(filter); + } + + int bias_size = GetShape(filter_)[0]; + if (input.type == TensorType_FLOAT32) { + bias_ = AddInput({TensorType_FLOAT32, {bias_size}}); + } else { + // This is a quantized version. The scale of 'bias' depends on the scales + // of input and filter. Supposedly this is correctly set during quantized + // training. + if (filter.per_channel_quantization) { + // per channel quantization. + std::vector bias_scale( + filter.per_channel_quantization_scales.size()); + std::vector bias_zero_points( + filter.per_channel_quantization_scales.size()); + for (size_t i = 0; i < filter.per_channel_quantization_scales.size(); + ++i) { + bias_scale[i] = + input.scale * filter.per_channel_quantization_scales[i]; + bias_zero_points[i] = 0; + } + tflite::TensorType bias_type = TensorType_INT32; + if (input.type == TensorType_INT16) { + // In case of 16-bit, the bias type is set to be int 64. + bias_type = TensorType_INT64; + } + TensorData bias{bias_type, + {bias_size}, + /*min=*/0, + /*max=*/0, + /*scale=*/0, + /*zero_point=*/0, + true, + /*per_channel_quantization_scales=*/bias_scale, + /*per_channel_quantization_offsets=*/bias_zero_points, + /*channel_index==*/0}; + bias_ = AddInput(bias); + } else { + // per tensor quantization. + auto bias_scale = GetScale(input_) * GetScale(filter_); + TensorData bias{TensorType_INT32, {bias_size}, 0, 0, bias_scale}; + bias_ = AddInput(bias); + } + } + + output_ = AddOutput(output); + + SetBuiltinOp(BuiltinOperator_CONV_2D, BuiltinOptions_Conv2DOptions, + CreateConv2DOptions( + builder_, padding, stride_width, stride_height, activation, + dilation_width_factor, dilation_height_factor) + .Union()); + + resolver_ = absl::make_unique(BuiltinOperator_CONV_2D, + registration); + BuildInterpreter({GetShape(input_), GetShape(filter_), GetShape(bias_)}, + /*num_threads=*/-1, + /* allow_fp32_relax_to_fp16 */ false, + /*apply_delegate=*/false); + compilation_status_ = ApplyDelegate(); + } + + protected: + int input_; + int filter_; + int bias_; + int output_; +}; + +class QuantizedConvolutionOpModel : public BaseConvolutionOpModel { + public: + using BaseConvolutionOpModel::BaseConvolutionOpModel; + + void SetInput(std::initializer_list data) { + QuantizeAndPopulate(input_, data); + } + + void SetFilter(std::initializer_list data) { + QuantizeAndPopulate(filter_, data); + } + + void SetBias(std::initializer_list data) { + QuantizeAndPopulate(bias_, data); + } + + std::vector GetOutput() { return ExtractVector(output_); } + std::vector GetDequantizedOutput() { + return Dequantize(ExtractVector(output_), + GetScale(output_), GetZeroPoint(output_)); + } +}; + +TEST_F(NnApiSignedQuantizationTest, + Conv2DUnsignedPerTensorMapsToUnsignedOnSdk29) { + QuantizedConvolutionOpModel m(nnapi_mock_->GetNnApi(), + ops::builtin::Register_CONVOLUTION_REF(), + {TensorType_UINT8, {2, 2, 4, 1}, -63.5, 64}, + {TensorType_UINT8, {3, 2, 2, 1}, -63.5, 64}, + {TensorType_UINT8, {}, -127, 128}); + m.SetInput({ + // First batch + 1, 1, 1, 1, // row = 1 + 2, 2, 2, 2, // row = 2 + // Second batch + 1, 2, 3, 4, // row = 1 + 1, 2, 3, 4, // row = 2 + }); + m.SetFilter({ + 1, 2, 3, 4, // first 2x2 filter + -1, 1, -1, 1, // second 2x2 filter + -1, -1, 1, 1, // third 2x2 filter + }); + m.SetBias({1, 2, 3}); + + m.Invoke(); + EXPECT_EQ(m.GetCompilationStatus(), kTfLiteOk); + + ASSERT_EQ(tensors_count->size(), 3); + ASSERT_NE(tensors_count->find(ANEURALNETWORKS_TENSOR_QUANT8_ASYMM), + tensors_count->end()); + ASSERT_NE(tensors_count->find(ANEURALNETWORKS_TENSOR_INT32), + tensors_count->end()); + ASSERT_NE(tensors_count->find(ANEURALNETWORKS_INT32), tensors_count->end()); + + EXPECT_EQ(tensors_count->at(ANEURALNETWORKS_TENSOR_QUANT8_ASYMM), + 3); // input, filter, output + EXPECT_EQ(tensors_count->at(ANEURALNETWORKS_TENSOR_INT32), 1); // bias + EXPECT_EQ(tensors_count->at(ANEURALNETWORKS_INT32), + 4); // padding, stride_width, stride_height, activation +} + +TEST_F(NnApiSignedQuantizationTest, + Conv2dUnsignedPerTensorMapsToUnsignedOnSdk30) { + nnapi_mock_->SetAndroidSdkVersion(30); + QuantizedConvolutionOpModel m(nnapi_mock_->GetNnApi(), + ops::builtin::Register_CONVOLUTION_REF(), + {TensorType_UINT8, {2, 2, 4, 1}, -63.5, 64}, + {TensorType_UINT8, {3, 2, 2, 1}, -63.5, 64}, + {TensorType_UINT8, {}, -127, 128}); + m.SetInput({ + // First batch + 1, 1, 1, 1, // row = 1 + 2, 2, 2, 2, // row = 2 + // Second batch + 1, 2, 3, 4, // row = 1 + 1, 2, 3, 4, // row = 2 + }); + m.SetFilter({ + 1, 2, 3, 4, // first 2x2 filter + -1, 1, -1, 1, // second 2x2 filter + -1, -1, 1, 1, // third 2x2 filter + }); + m.SetBias({1, 2, 3}); + + m.Invoke(); + EXPECT_EQ(m.GetCompilationStatus(), kTfLiteOk); + + ASSERT_EQ(tensors_count->size(), 3); + ASSERT_NE(tensors_count->find(ANEURALNETWORKS_TENSOR_QUANT8_ASYMM), + tensors_count->end()); + ASSERT_NE(tensors_count->find(ANEURALNETWORKS_TENSOR_INT32), + tensors_count->end()); + ASSERT_NE(tensors_count->find(ANEURALNETWORKS_INT32), tensors_count->end()); + + EXPECT_EQ(tensors_count->at(ANEURALNETWORKS_TENSOR_QUANT8_ASYMM), + 3); // input, filter, output + EXPECT_EQ(tensors_count->at(ANEURALNETWORKS_TENSOR_INT32), 1); // bias + EXPECT_EQ(tensors_count->at(ANEURALNETWORKS_INT32), + 4); // padding, stride_width, stride_height, activation +} + +class PerChannelQuantizedConvolutionOpModel + : public BaseConvolutionOpModel { + public: + using BaseConvolutionOpModel::BaseConvolutionOpModel; + + void SetInput(std::initializer_list data) { + QuantizeAndPopulate(input_, data); + } + + void SetFilter(std::initializer_list data) { + PerChannelSymmetricQuantizeAndPopulate(filter_, data); + } + + void SetBias(std::initializer_list data) { + PerChannelQuantizeBias(bias_, data); + } + + std::vector GetOutput() { return ExtractVector(output_); } + std::vector GetDequantizedOutput() { + return Dequantize(ExtractVector(output_), GetScale(output_), + GetZeroPoint(output_)); + } +}; + +TEST_F(NnApiSignedQuantizationTest, + Conv2dSignedPerTensorMapsToUnsignedOnSdk29) { + nnapi_mock_->SetAndroidSdkVersion(29); + PerChannelQuantizedConvolutionOpModel m( + nnapi_mock_->GetNnApi(), ops::builtin::Register_CONVOLUTION_REF(), + {TensorType_INT8, {1, 2, 3, 2}, -63.5, 64, 0.5, -1}, + {TensorType_INT8, + // [2 * 2 * 2 * 2] as [output_channel, y, x, input_channel] + {2, 2, 2, 2}, + 0, + 0, + 0, + 0, + /*per_channel_quantization=*/true, + /*per_channel_quantization_scales=*/{1}, + /*per_channel_quantization_offsets=*/{0}, + /*channel_index=*/0}, + {TensorType_INT8, {}, -63.5, 64, 0.5, -1}, + /*stride_width=*/1, /*stride_height=*/1); + m.SetInput({ + // [1 * 2 * 3 * 2] as [batch, y, x, input_channel] + 3, 2, // batch = 0, y = 0, x = 0 + 1, -1, // batch = 0, y = 0, x = 1 + -2, -3, // batch = 0, y = 0, x = 2 + 4, 3, // batch = 0, y = 1, x = 0 + 2, -2, // batch = 0, y = 1, x = 1 + -3, -4, // batch = 0, y = 1, x = 2 + }); + m.SetFilter( + // [2 * 2 * 2 * 2] as [output_channel, y, x, input_channel] + { + 1, 2, // out channel = 0, y = 0, x = 0 + 3, 4, // out channel = 0, y = 0, x = 1 + 3, 4, // out channel = 0, y = 1, x = 0 + 5, 6, // out channel = 0, y = 1, x = 1 + 7, 8, // out channel = 1, y = 0, x = 0 + 5, 6, // out channel = 1, y = 0, x = 1 + 3, 4, // out channel = 1, y = 1, x = 0 + 1, 2, // out channel = 1, y = 1, x = 1 + }); + m.SetBias({3, -2}); + + // Invoke and verify output. + // output has dimension [1 * 1 * 2 * 2] as [batch, y, x, output_channel] + m.Invoke(); + EXPECT_EQ(m.GetCompilationStatus(), kTfLiteOk); + + ASSERT_EQ(tensors_count->size(), 3); + ASSERT_NE(tensors_count->find(ANEURALNETWORKS_TENSOR_QUANT8_ASYMM), + tensors_count->end()); + ASSERT_NE(tensors_count->find(ANEURALNETWORKS_TENSOR_INT32), + tensors_count->end()); + ASSERT_NE(tensors_count->find(ANEURALNETWORKS_INT32), tensors_count->end()); + + EXPECT_EQ(tensors_count->at(ANEURALNETWORKS_TENSOR_QUANT8_ASYMM), + 3); // input, filter, output + EXPECT_EQ(tensors_count->at(ANEURALNETWORKS_TENSOR_INT32), 1); // bias + EXPECT_EQ(tensors_count->at(ANEURALNETWORKS_INT32), + 4); // padding, stride_width, stride_height, activation +} + +TEST_F(NnApiSignedQuantizationTest, + Conv2dSignedPerTensorMapsToUnsignedOnSdk30) { + nnapi_mock_->SetAndroidSdkVersion(30); + PerChannelQuantizedConvolutionOpModel m( + nnapi_mock_->GetNnApi(), ops::builtin::Register_CONVOLUTION_REF(), + {TensorType_INT8, {1, 2, 3, 2}, -63.5, 64, 0.5, -1}, + {TensorType_INT8, + // [2 * 2 * 2 * 2] as [output_channel, y, x, input_channel] + {2, 2, 2, 2}, + 0, + 0, + 0, + 0, + /*per_channel_quantization=*/true, + /*per_channel_quantization_scales=*/{1}, + /*per_channel_quantization_offsets=*/{0}, + /*channel_index=*/0}, + {TensorType_INT8, {}, -63.5, 64, 0.5, -1}, + /*stride_width=*/1, /*stride_height=*/1); + m.SetInput({ + // [1 * 2 * 3 * 2] as [batch, y, x, input_channel] + 3, 2, // batch = 0, y = 0, x = 0 + 1, -1, // batch = 0, y = 0, x = 1 + -2, -3, // batch = 0, y = 0, x = 2 + 4, 3, // batch = 0, y = 1, x = 0 + 2, -2, // batch = 0, y = 1, x = 1 + -3, -4, // batch = 0, y = 1, x = 2 + }); + m.SetFilter( + // [2 * 2 * 2 * 2] as [output_channel, y, x, input_channel] + { + 1, 2, // out channel = 0, y = 0, x = 0 + 3, 4, // out channel = 0, y = 0, x = 1 + 3, 4, // out channel = 0, y = 1, x = 0 + 5, 6, // out channel = 0, y = 1, x = 1 + 7, 8, // out channel = 1, y = 0, x = 0 + 5, 6, // out channel = 1, y = 0, x = 1 + 3, 4, // out channel = 1, y = 1, x = 0 + 1, 2, // out channel = 1, y = 1, x = 1 + }); + m.SetBias({3, -2}); + + // Invoke and verify output. + // output has dimension [1 * 1 * 2 * 2] as [batch, y, x, output_channel] + m.Invoke(); + EXPECT_EQ(m.GetCompilationStatus(), kTfLiteOk); + + ASSERT_EQ(tensors_count->size(), 3); + ASSERT_NE(tensors_count->find(ANEURALNETWORKS_TENSOR_QUANT8_ASYMM_SIGNED), + tensors_count->end()); + ASSERT_NE(tensors_count->find(ANEURALNETWORKS_TENSOR_INT32), + tensors_count->end()); + ASSERT_NE(tensors_count->find(ANEURALNETWORKS_INT32), tensors_count->end()); + + EXPECT_EQ(tensors_count->at(ANEURALNETWORKS_TENSOR_QUANT8_ASYMM_SIGNED), + 3); // input, filter, output + EXPECT_EQ(tensors_count->at(ANEURALNETWORKS_TENSOR_INT32), 1); // bias + EXPECT_EQ(tensors_count->at(ANEURALNETWORKS_INT32), + 4); // padding, stride_width, stride_height, activation +} + +TEST_F(NnApiSignedQuantizationTest, + Conv2dSignedPerChannelMapsToUnsignedOnSdk29) { + PerChannelQuantizedConvolutionOpModel m( + nnapi_mock_->GetNnApi(), ops::builtin::Register_CONVOLUTION_REF(), + {TensorType_INT8, {1, 2, 3, 2}, -63.5, 64, 0.5, -1}, + {TensorType_INT8, + // [2 * 2 * 2 * 2] as [output_channel, y, x, input_channel] + {2, 2, 2, 2}, + 0, + 0, + 0, + 0, + /*per_channel_quantization=*/true, + /*per_channel_quantization_scales=*/{1, 2}, + /*per_channel_quantization_offsets=*/{0, 0}, + /*channel_index=*/0}, + {TensorType_INT8, {}, -63.5, 64, 0.5, -1}, + /*stride_width=*/1, /*stride_height=*/1); + m.SetInput({ + // [1 * 2 * 3 * 2] as [batch, y, x, input_channel] + 3, 2, // batch = 0, y = 0, x = 0 + 1, -1, // batch = 0, y = 0, x = 1 + -2, -3, // batch = 0, y = 0, x = 2 + 4, 3, // batch = 0, y = 1, x = 0 + 2, -2, // batch = 0, y = 1, x = 1 + -3, -4, // batch = 0, y = 1, x = 2 + }); + m.SetFilter( + // [2 * 2 * 2 * 2] as [output_channel, y, x, input_channel] + { + 1, 2, // out channel = 0, y = 0, x = 0 + 3, 4, // out channel = 0, y = 0, x = 1 + 3, 4, // out channel = 0, y = 1, x = 0 + 5, 6, // out channel = 0, y = 1, x = 1 + 7, 8, // out channel = 1, y = 0, x = 0 + 5, 6, // out channel = 1, y = 0, x = 1 + 3, 4, // out channel = 1, y = 1, x = 0 + 1, 2, // out channel = 1, y = 1, x = 1 + }); + m.SetBias({3, -2}); + + // Invoke and verify output. + // output has dimension [1 * 1 * 2 * 2] as [batch, y, x, output_channel] + m.Invoke(); + EXPECT_EQ(m.GetCompilationStatus(), kTfLiteOk); + + ASSERT_EQ(tensors_count->size(), 4); + ASSERT_NE(tensors_count->find(ANEURALNETWORKS_TENSOR_QUANT8_ASYMM), + tensors_count->end()); + ASSERT_NE(tensors_count->find(ANEURALNETWORKS_TENSOR_QUANT8_SYMM_PER_CHANNEL), + tensors_count->end()); + ASSERT_NE(tensors_count->find(ANEURALNETWORKS_TENSOR_INT32), + tensors_count->end()); + ASSERT_NE(tensors_count->find(ANEURALNETWORKS_INT32), tensors_count->end()); + + EXPECT_EQ(tensors_count->at(ANEURALNETWORKS_TENSOR_QUANT8_ASYMM), + 2); // input, output + EXPECT_EQ(tensors_count->at(ANEURALNETWORKS_TENSOR_QUANT8_SYMM_PER_CHANNEL), + 1); // filter + EXPECT_EQ(tensors_count->at(ANEURALNETWORKS_TENSOR_INT32), 1); // bias + EXPECT_EQ(tensors_count->at(ANEURALNETWORKS_INT32), + 4); // padding, stride_width, stride_height, activation +} + +TEST_F(NnApiSignedQuantizationTest, Conv2dSignedPerChannelMapsToSignedOnSdk30) { + nnapi_mock_->SetAndroidSdkVersion(30); + PerChannelQuantizedConvolutionOpModel m( + nnapi_mock_->GetNnApi(), ops::builtin::Register_CONVOLUTION_REF(), + {TensorType_INT8, {1, 2, 3, 2}, -63.5, 64, 0.5, -1}, + {TensorType_INT8, + // [2 * 2 * 2 * 2] as [output_channel, y, x, input_channel] + {2, 2, 2, 2}, + 0, + 0, + 0, + 0, + /*per_channel_quantization=*/true, + /*per_channel_quantization_scales=*/{1, 2}, + /*per_channel_quantization_offsets=*/{0, 0}, + /*channel_index=*/0}, + {TensorType_INT8, {}, -63.5, 64, 0.5, -1}, + /*stride_width=*/1, /*stride_height=*/1); + m.SetInput({ + // [1 * 2 * 3 * 2] as [batch, y, x, input_channel] + 3, 2, // batch = 0, y = 0, x = 0 + 1, -1, // batch = 0, y = 0, x = 1 + -2, -3, // batch = 0, y = 0, x = 2 + 4, 3, // batch = 0, y = 1, x = 0 + 2, -2, // batch = 0, y = 1, x = 1 + -3, -4, // batch = 0, y = 1, x = 2 + }); + m.SetFilter( + // [2 * 2 * 2 * 2] as [output_channel, y, x, input_channel] + { + 1, 2, // out channel = 0, y = 0, x = 0 + 3, 4, // out channel = 0, y = 0, x = 1 + 3, 4, // out channel = 0, y = 1, x = 0 + 5, 6, // out channel = 0, y = 1, x = 1 + 7, 8, // out channel = 1, y = 0, x = 0 + 5, 6, // out channel = 1, y = 0, x = 1 + 3, 4, // out channel = 1, y = 1, x = 0 + 1, 2, // out channel = 1, y = 1, x = 1 + }); + m.SetBias({3, -2}); + + // Invoke and verify output. + // output has dimension [1 * 1 * 2 * 2] as [batch, y, x, output_channel] + m.Invoke(); + EXPECT_EQ(m.GetCompilationStatus(), kTfLiteOk); + + ASSERT_EQ(tensors_count->size(), 4); + ASSERT_NE(tensors_count->find(ANEURALNETWORKS_TENSOR_QUANT8_ASYMM_SIGNED), + tensors_count->end()); + ASSERT_NE(tensors_count->find(ANEURALNETWORKS_TENSOR_QUANT8_SYMM_PER_CHANNEL), + tensors_count->end()); + ASSERT_NE(tensors_count->find(ANEURALNETWORKS_TENSOR_INT32), + tensors_count->end()); + ASSERT_NE(tensors_count->find(ANEURALNETWORKS_INT32), tensors_count->end()); + + EXPECT_EQ(tensors_count->at(ANEURALNETWORKS_TENSOR_QUANT8_ASYMM_SIGNED), + 2); // input, output + EXPECT_EQ(tensors_count->at(ANEURALNETWORKS_TENSOR_QUANT8_SYMM_PER_CHANNEL), + 1); // filter + EXPECT_EQ(tensors_count->at(ANEURALNETWORKS_TENSOR_INT32), 1); // bias + EXPECT_EQ(tensors_count->at(ANEURALNETWORKS_INT32), + 4); // padding, stride_width, stride_height, activation +} + +class QuantizeOpModel : public SingleOpModelWithNNAPI { + public: + QuantizeOpModel(const NnApi* nnapi, const TensorData& input, + const TensorData& output) { + SingleOpModelWithNNAPI::Init(nnapi); + input_ = AddInput(input); + output_ = AddOutput(output); + SetBuiltinOp(BuiltinOperator_QUANTIZE, BuiltinOptions_QuantizeOptions, + CreateQuantizeOptions(builder_).Union()); + + BuildInterpreter({GetShape(input_)}, /*num_threads=*/-1, + /* allow_fp32_relax_to_fp16 */ false, + /*apply_delegate=*/false); + compilation_status_ = ApplyDelegate(); + } + + void SetInput(std::initializer_list data) { + PopulateTensor(input_, data); + } + + template + void SetInputAndQuantize(std::initializer_list data) { + QuantizeAndPopulate(input_, data); + } + + template + std::vector GetOutput() { + return ExtractVector(output_); + } + + private: + int input_; + int output_; +}; + +TEST_F(NnApiSignedQuantizationTest, QuantizeUint8MapsToUint8OnSdk29) { + // [-63.5, 64] -> scale=0.5 zero_point=127 for UINT8 + QuantizeOpModel m(nnapi_mock_->GetNnApi(), {TensorType_FLOAT32, {2, 5}}, + {TensorType_UINT8, {2, 5}, 0, 0, 0.5, 127}); + + m.SetInput({-63.5, -63, -62.5, -62, -61.5, 62, 62.5, 63, 63.5, 64}); + m.Invoke(); + EXPECT_EQ(m.GetCompilationStatus(), kTfLiteOk); + + ASSERT_EQ(tensors_count->size(), 2); + ASSERT_NE(tensors_count->find(ANEURALNETWORKS_TENSOR_FLOAT32), + tensors_count->end()); + ASSERT_NE(tensors_count->find(ANEURALNETWORKS_TENSOR_QUANT8_ASYMM), + tensors_count->end()); + + EXPECT_EQ(tensors_count->at(ANEURALNETWORKS_TENSOR_FLOAT32), + 1); // input + EXPECT_EQ(tensors_count->at(ANEURALNETWORKS_TENSOR_QUANT8_ASYMM), + 1); // output +} + +TEST_F(NnApiSignedQuantizationTest, QuantizeUint8MapsToUint8OnSdk30) { + nnapi_mock_->SetAndroidSdkVersion(30); + // [-63.5, 64] -> scale=0.5 zero_point=127 for UINT8 + QuantizeOpModel m(nnapi_mock_->GetNnApi(), {TensorType_FLOAT32, {2, 5}}, + {TensorType_UINT8, {2, 5}, 0, 0, 0.5, 127}); + + m.SetInput({-63.5, -63, -62.5, -62, -61.5, 62, 62.5, 63, 63.5, 64}); + m.Invoke(); + EXPECT_EQ(m.GetCompilationStatus(), kTfLiteOk); + + ASSERT_EQ(tensors_count->size(), 2); + ASSERT_NE(tensors_count->find(ANEURALNETWORKS_TENSOR_FLOAT32), + tensors_count->end()); + ASSERT_NE(tensors_count->find(ANEURALNETWORKS_TENSOR_QUANT8_ASYMM), + tensors_count->end()); + + EXPECT_EQ(tensors_count->at(ANEURALNETWORKS_TENSOR_FLOAT32), + 1); // input + EXPECT_EQ(tensors_count->at(ANEURALNETWORKS_TENSOR_QUANT8_ASYMM), + 1); // output +} + +// Quantize with Int8 output is only supported since SDK level 30. +TEST_F(NnApiSignedQuantizationTest, QuantizeInt8MapsToInt8OnSdk30) { + nnapi_mock_->SetAndroidSdkVersion(30); + // [-63.5, 64] -> scale=0.5 zero_point=1 for INT8 + QuantizeOpModel m(nnapi_mock_->GetNnApi(), {TensorType_FLOAT32, {2, 5}}, + {TensorType_INT8, {2, 5}, 0, 0, 0.5, -1}); + + m.SetInput({-63.5, -63, -62.5, -62, -61.5, 62, 62.5, 63, 63.5, 64}); + m.Invoke(); + EXPECT_EQ(m.GetCompilationStatus(), kTfLiteOk); + + ASSERT_EQ(tensors_count->size(), 2); + ASSERT_NE(tensors_count->find(ANEURALNETWORKS_TENSOR_FLOAT32), + tensors_count->end()); + ASSERT_NE(tensors_count->find(ANEURALNETWORKS_TENSOR_QUANT8_ASYMM_SIGNED), + tensors_count->end()); + + EXPECT_EQ(tensors_count->at(ANEURALNETWORKS_TENSOR_FLOAT32), + 1); // input + EXPECT_EQ(tensors_count->at(ANEURALNETWORKS_TENSOR_QUANT8_ASYMM_SIGNED), + 1); // output +} + +class DequantizeOpModel : public SingleOpModelWithNNAPI { + public: + DequantizeOpModel(const NnApi* nnapi, TensorType type, + std::initializer_list shape, float scale, + int32_t zero_point, int version) { + SingleOpModelWithNNAPI::Init(nnapi); + const TensorData input_tensor_data = {type, shape, 0, 0, scale, zero_point}; + input_ = AddInput(input_tensor_data); + output_ = AddOutput({TensorType_FLOAT32, shape}); + SetBuiltinOp(BuiltinOperator_DEQUANTIZE, BuiltinOptions_DequantizeOptions, + CreateDequantizeOptions(builder_).Union()); + + resolver_ = absl::make_unique( + BuiltinOperator_DEQUANTIZE, ops::builtin::Register_DEQUANTIZE(), + version); + + BuildInterpreter({GetShape(input_)}, /*num_threads=*/-1, + /* allow_fp32_relax_to_fp16 */ false, + /*apply_delegate=*/false); + compilation_status_ = ApplyDelegate(); + } + + template + void SetInput(std::initializer_list data) { + PopulateTensor(input_, data); + } + + std::vector GetOutput() { return ExtractVector(output_); } + + private: + int input_; + int output_; +}; + +TEST_F(NnApiSignedQuantizationTest, DequantizeUint8MapsToUint8OnSdk29) { + // [-63.5, 64] -> scale=0.5 zero_point=127 for UINT8 + DequantizeOpModel m(nnapi_mock_->GetNnApi(), TensorType_UINT8, {2, 5}, 0.5, + 127, 1); + + m.SetInput({0, 1, 2, 3, 4, 251, 252, 253, 254, 255}); + m.Invoke(); + EXPECT_EQ(m.GetCompilationStatus(), kTfLiteOk); + + ASSERT_EQ(tensors_count->size(), 2); + ASSERT_NE(tensors_count->find(ANEURALNETWORKS_TENSOR_QUANT8_ASYMM), + tensors_count->end()); + ASSERT_NE(tensors_count->find(ANEURALNETWORKS_TENSOR_FLOAT32), + tensors_count->end()); + + EXPECT_EQ(tensors_count->at(ANEURALNETWORKS_TENSOR_QUANT8_ASYMM), + 1); // input + EXPECT_EQ(tensors_count->at(ANEURALNETWORKS_TENSOR_FLOAT32), + 1); // output +} + +TEST_F(NnApiSignedQuantizationTest, DequantizeUint8MapsToUint8OnSdk30) { + nnapi_mock_->SetAndroidSdkVersion(30); + // [-63.5, 64] -> scale=0.5 zero_point=127 for UINT8 + DequantizeOpModel m(nnapi_mock_->GetNnApi(), TensorType_UINT8, {2, 5}, 0.5, + 127, 1); + + m.SetInput({0, 1, 2, 3, 4, 251, 252, 253, 254, 255}); + m.Invoke(); + EXPECT_EQ(m.GetCompilationStatus(), kTfLiteOk); + + ASSERT_EQ(tensors_count->size(), 2); + ASSERT_NE(tensors_count->find(ANEURALNETWORKS_TENSOR_QUANT8_ASYMM), + tensors_count->end()); + ASSERT_NE(tensors_count->find(ANEURALNETWORKS_TENSOR_FLOAT32), + tensors_count->end()); + + EXPECT_EQ(tensors_count->at(ANEURALNETWORKS_TENSOR_QUANT8_ASYMM), + 1); // input + EXPECT_EQ(tensors_count->at(ANEURALNETWORKS_TENSOR_FLOAT32), + 1); // output +} + +// Dequantize with Int8 input is only supported for symmetric quantization on +// SDK level 29 +TEST_F(NnApiSignedQuantizationTest, + DequantizeTestInt8SymmMapsToInt8SymmOnSdk29) { + // [-63.5, 64] -> scale=0.5, zero_point=0 for INT8 + DequantizeOpModel m(nnapi_mock_->GetNnApi(), TensorType_INT8, {2, 5}, 0.5, 0, + 2); + + m.SetInput({-128, -127, -126, -125, -124, 123, 124, 125, 126, 127}); + m.Invoke(); + EXPECT_EQ(m.GetCompilationStatus(), kTfLiteOk); + + ASSERT_EQ(tensors_count->size(), 2); + ASSERT_NE(tensors_count->find(ANEURALNETWORKS_TENSOR_QUANT8_SYMM), + tensors_count->end()); + ASSERT_NE(tensors_count->find(ANEURALNETWORKS_TENSOR_FLOAT32), + tensors_count->end()); + + EXPECT_EQ(tensors_count->at(ANEURALNETWORKS_TENSOR_QUANT8_SYMM), + 1); // input + EXPECT_EQ(tensors_count->at(ANEURALNETWORKS_TENSOR_FLOAT32), + 1); // output +} + +// Dequantize with Int8 input is only supported since SDK level 30. +TEST_F(NnApiSignedQuantizationTest, DequantizeTestInt8MapsToInt8OnSdk30) { + nnapi_mock_->SetAndroidSdkVersion(30); + // [-63.5, 64] -> scale=0.5, zero_point=1 for INT8 + DequantizeOpModel m(nnapi_mock_->GetNnApi(), TensorType_INT8, {2, 5}, 0.5, -1, + 2); + + m.SetInput({-128, -127, -126, -125, -124, 123, 124, 125, 126, 127}); + m.Invoke(); + EXPECT_EQ(m.GetCompilationStatus(), kTfLiteOk); + + ASSERT_EQ(tensors_count->size(), 2); + ASSERT_NE(tensors_count->find(ANEURALNETWORKS_TENSOR_QUANT8_ASYMM_SIGNED), + tensors_count->end()); + ASSERT_NE(tensors_count->find(ANEURALNETWORKS_TENSOR_FLOAT32), + tensors_count->end()); + + EXPECT_EQ(tensors_count->at(ANEURALNETWORKS_TENSOR_QUANT8_ASYMM_SIGNED), + 1); // input + EXPECT_EQ(tensors_count->at(ANEURALNETWORKS_TENSOR_FLOAT32), + 1); // output +} + +} // namespace +} // namespace tflite + +int main(int argc, char** argv) { + ::tflite::LogToStderr(); + ::testing::InitGoogleTest(&argc, argv); + return RUN_ALL_TESTS(); +} diff --git a/tensorflow/lite/nnapi/NeuralNetworksTypes.h b/tensorflow/lite/nnapi/NeuralNetworksTypes.h index 3c30a0479fa..8415df58b8b 100644 --- a/tensorflow/lite/nnapi/NeuralNetworksTypes.h +++ b/tensorflow/lite/nnapi/NeuralNetworksTypes.h @@ -46,6 +46,7 @@ enum { ANEURALNETWORKS_TENSOR_QUANT16_SYMM = 7, ANEURALNETWORKS_TENSOR_QUANT8_SYMM_PER_CHANNEL = 11, ANEURALNETWORKS_TENSOR_QUANT8_SYMM = 13, + ANEURALNETWORKS_TENSOR_QUANT8_ASYMM_SIGNED = 14, }; /** diff --git a/tensorflow/lite/nnapi/nnapi_handler.h b/tensorflow/lite/nnapi/nnapi_handler.h index 00c0b23e3cf..16e1e9fea10 100644 --- a/tensorflow/lite/nnapi/nnapi_handler.h +++ b/tensorflow/lite/nnapi/nnapi_handler.h @@ -118,6 +118,11 @@ class NnApiHandler { const ANeuralNetworksOperandType* type) { return Value; }; } + void StubAddOperandWith(int(stub)(ANeuralNetworksModel* model, + const ANeuralNetworksOperandType* type)) { + nnapi_->ANeuralNetworksModel_addOperand = stub; + } + template void SetOperandValueReturns() { nnapi_->ANeuralNetworksModel_setOperandValue = @@ -268,6 +273,23 @@ class NnApiHandler { }; } + template + void SetPriorityReturns() { + nnapi_->ANeuralNetworksCompilation_setPriority = + [](ANeuralNetworksCompilation* compilation, int priority) -> int { + return Value; + }; + } + + template + void SetOperandSymmPerChannelQuantParamsReturns() { + nnapi_->ANeuralNetworksModel_setOperandSymmPerChannelQuantParams = + [](ANeuralNetworksModel* model, int32_t index, + const ANeuralNetworksSymmPerChannelQuantParams* channelQuant) { + return Value; + }; + } + /* * Sets the SDK Version in the nnapi structure. * If set_unsupported_ops_to_null is set to true, all the functions not From 4fc5c50e15d48ae50e3d7c192a6efad19fafdb3d Mon Sep 17 00:00:00 2001 From: Lukas Geiger Date: Tue, 23 Jun 2020 16:00:35 +0200 Subject: [PATCH 546/867] Prefer generator expressions over list comprehensions --- tensorflow/python/eager/function.py | 4 ++-- tensorflow/python/keras/engine/training_v1.py | 4 ++-- .../keras/layers/preprocessing/category_crossing.py | 8 ++++---- .../python/keras/layers/preprocessing/hashing.py | 12 ++++++------ 4 files changed, 14 insertions(+), 14 deletions(-) diff --git a/tensorflow/python/eager/function.py b/tensorflow/python/eager/function.py index e2f5d86fbd2..beb58d8ff91 100644 --- a/tensorflow/python/eager/function.py +++ b/tensorflow/python/eager/function.py @@ -1192,7 +1192,7 @@ class _TapeGradientFunctions(object): def _wrap_backward_function(self, forward_graph, backward, outputs): """Create a backward function given `outputs` from the forward function.""" capture_mapping = dict( - zip([ops.tensor_id(t) for t in forward_graph.outputs], outputs)) + zip((ops.tensor_id(t) for t in forward_graph.outputs), outputs)) remapped_captures = [ capture_mapping.get(ops.tensor_id(capture), capture) for capture in backward.captured_inputs @@ -1491,7 +1491,7 @@ class ConcreteFunction(object): self._ndarrays_list = ( isinstance(structured_outputs, (list, tuple)) and structured_outputs and - all([isinstance(o, np_arrays.ndarray) for o in structured_outputs])) + all(isinstance(o, np_arrays.ndarray) for o in structured_outputs)) self._ndarray_singleton = isinstance(structured_outputs, np_arrays.ndarray) # function_spec defines the structured signature. diff --git a/tensorflow/python/keras/engine/training_v1.py b/tensorflow/python/keras/engine/training_v1.py index c137c6e517a..c901b73c892 100644 --- a/tensorflow/python/keras/engine/training_v1.py +++ b/tensorflow/python/keras/engine/training_v1.py @@ -1492,8 +1492,8 @@ class Model(training_lib.Model): def _recompile_weights_loss_and_weighted_metrics(self): if not self._is_compiled: return False - recompile = any([e.sample_weights_mismatch() - for e in self._training_endpoints]) + recompile = any(e.sample_weights_mismatch() + for e in self._training_endpoints) if recompile: self._compile_weights_loss_and_weighted_metrics() diff --git a/tensorflow/python/keras/layers/preprocessing/category_crossing.py b/tensorflow/python/keras/layers/preprocessing/category_crossing.py index 594b9741946..e949bd4c87c 100644 --- a/tensorflow/python/keras/layers/preprocessing/category_crossing.py +++ b/tensorflow/python/keras/layers/preprocessing/category_crossing.py @@ -188,15 +188,15 @@ class CategoryCrossing(Layer): def compute_output_signature(self, input_spec): input_shapes = [x.shape for x in input_spec] output_shape = self.compute_output_shape(input_shapes) - if any([ + if any( isinstance(inp_spec, ragged_tensor.RaggedTensorSpec) for inp_spec in input_spec - ]): + ): return tensor_spec.TensorSpec(shape=output_shape, dtype=dtypes.string) - elif any([ + elif any( isinstance(inp_spec, sparse_tensor.SparseTensorSpec) for inp_spec in input_spec - ]): + ): return sparse_tensor.SparseTensorSpec( shape=output_shape, dtype=dtypes.string) return tensor_spec.TensorSpec(shape=output_shape, dtype=dtypes.string) diff --git a/tensorflow/python/keras/layers/preprocessing/hashing.py b/tensorflow/python/keras/layers/preprocessing/hashing.py index faeeec63a86..89c3042ae24 100644 --- a/tensorflow/python/keras/layers/preprocessing/hashing.py +++ b/tensorflow/python/keras/layers/preprocessing/hashing.py @@ -158,10 +158,10 @@ class Hashing(Layer): def _preprocess_inputs(self, inputs): if isinstance(inputs, (tuple, list)): # If any of them is tensor or ndarray, then treat as list - if any([ + if any( tensor_util.is_tensor(inp) or isinstance(inp, np.ndarray) for inp in inputs - ]): + ): return [self._preprocess_single_input(inp) for inp in inputs] return self._preprocess_single_input(inputs) @@ -261,15 +261,15 @@ class Hashing(Layer): return tensor_spec.TensorSpec(shape=output_shape, dtype=output_dtype) input_shapes = [x.shape for x in input_spec] output_shape = self.compute_output_shape(input_shapes) - if any([ + if any( isinstance(inp_spec, ragged_tensor.RaggedTensorSpec) for inp_spec in input_spec - ]): + ): return tensor_spec.TensorSpec(shape=output_shape, dtype=dtypes.int64) - elif any([ + elif any( isinstance(inp_spec, sparse_tensor.SparseTensorSpec) for inp_spec in input_spec - ]): + ): return sparse_tensor.SparseTensorSpec( shape=output_shape, dtype=dtypes.int64) return tensor_spec.TensorSpec(shape=output_shape, dtype=dtypes.int64) From ba915231715348096e18775581027d7e625d8485 Mon Sep 17 00:00:00 2001 From: Stephan Herhut Date: Tue, 23 Jun 2020 06:18:20 -0700 Subject: [PATCH 547/867] Add call-back to further configure the llvm::TargetMachine that was selected for PTX generation. This is used in kernel generation to enable FMA operation fusion. PiperOrigin-RevId: 317851871 Change-Id: Ifaaa1ba99e5aecb90f380d085583bb8f56805d9f --- .../compiler/mlir/tools/kernel_gen/cubin_creator.cc | 12 +++++++++--- .../service/gpu/llvm_gpu_backend/gpu_backend_lib.cc | 12 +++++++++--- .../service/gpu/llvm_gpu_backend/gpu_backend_lib.h | 8 +++++--- 3 files changed, 23 insertions(+), 9 deletions(-) diff --git a/tensorflow/compiler/mlir/tools/kernel_gen/cubin_creator.cc b/tensorflow/compiler/mlir/tools/kernel_gen/cubin_creator.cc index 79969a22572..85a53e042e1 100644 --- a/tensorflow/compiler/mlir/tools/kernel_gen/cubin_creator.cc +++ b/tensorflow/compiler/mlir/tools/kernel_gen/cubin_creator.cc @@ -268,6 +268,7 @@ StatusOr> tensorflow::kernel_gen::GenerateCubinForTfCode( options.tile_sizes = tile_sizes; options.unroll_factors = unroll_factors; options.collapse_parallel_loops = false; + options.use_approximations = true; TF_RETURN_IF_ERROR(xla::mlir_gpu::LowerLHLOToGPU(module.get(), options)); } TF_RETURN_IF_ERROR(xla::mlir_gpu::LowerKernelBodiesToNVVM(module.get())); @@ -287,10 +288,15 @@ StatusOr> tensorflow::kernel_gen::GenerateCubinForTfCode( xla::HloModuleConfig config; config.set_debug_options(xla::GetDebugOptionsFromFlags()); + auto enable_fusion = [](llvm::TargetMachine* target) { + target->Options.AllowFPOpFusion = llvm::FPOpFusion::FPOpFusionMode::Fast; + }; + TF_ASSIGN_OR_RETURN(std::string libdevice_dir, GetLibdeviceDir(config)); - TF_ASSIGN_OR_RETURN(std::string ptx, xla::gpu::nvptx::CompileToPtx( - llvmModule.get(), compute_capability, - config, libdevice_dir)); + TF_ASSIGN_OR_RETURN( + std::string ptx, + xla::gpu::nvptx::CompileToPtx(llvmModule.get(), compute_capability, + config, libdevice_dir, enable_fusion)); VLOG(1) << ptx; #if GOOGLE_CUDA diff --git a/tensorflow/compiler/xla/service/gpu/llvm_gpu_backend/gpu_backend_lib.cc b/tensorflow/compiler/xla/service/gpu/llvm_gpu_backend/gpu_backend_lib.cc index 497dcda4361..d2126a8d17d 100644 --- a/tensorflow/compiler/xla/service/gpu/llvm_gpu_backend/gpu_backend_lib.cc +++ b/tensorflow/compiler/xla/service/gpu/llvm_gpu_backend/gpu_backend_lib.cc @@ -492,9 +492,10 @@ void NVPTXBackendInit(const HloModuleConfig& hlo_module_config) { namespace nvptx { -StatusOr CompileToPtx(llvm::Module* module, GpuVersion gpu_version, - const HloModuleConfig& hlo_module_config, - const string& libdevice_dir_path) { +StatusOr CompileToPtx( + llvm::Module* module, GpuVersion gpu_version, + const HloModuleConfig& hlo_module_config, const string& libdevice_dir_path, + std::function configure_target) { static absl::once_flag backend_init_flag; absl::call_once(backend_init_flag, NVPTXBackendInit, hlo_module_config); @@ -525,6 +526,11 @@ StatusOr CompileToPtx(llvm::Module* module, GpuVersion gpu_version, std::unique_ptr target_machine = NVPTXGetTargetMachine( default_target_triple, *compute_capability, hlo_module_config); + // Apply target machine configuration from call-back if available. + if (configure_target) { + configure_target(target_machine.get()); + } + // Link with libdevice, and optimize the LLVM module. TF_RETURN_IF_ERROR(LinkAndOptimizeModule( module, gpu_version, hlo_module_config, libdevice_dir_path, diff --git a/tensorflow/compiler/xla/service/gpu/llvm_gpu_backend/gpu_backend_lib.h b/tensorflow/compiler/xla/service/gpu/llvm_gpu_backend/gpu_backend_lib.h index 526621de7a5..33ef9280c7a 100644 --- a/tensorflow/compiler/xla/service/gpu/llvm_gpu_backend/gpu_backend_lib.h +++ b/tensorflow/compiler/xla/service/gpu/llvm_gpu_backend/gpu_backend_lib.h @@ -22,6 +22,7 @@ limitations under the License. #include "absl/strings/string_view.h" #include "llvm/IR/Module.h" +#include "llvm/Target/TargetMachine.h" #include "tensorflow/compiler/xla/service/gpu/gpu_types.h" #include "tensorflow/compiler/xla/service/hlo_module_config.h" #include "tensorflow/compiler/xla/statusor.h" @@ -38,9 +39,10 @@ namespace nvptx { // The Compile.* interfaces each create their own llvm::LLVMContext objects for // thread safety, but note that LLVM's multithreaded support is very // preliminary; multithreaded use is not recommended at this time. -StatusOr CompileToPtx(llvm::Module* module, GpuVersion gpu_version, - const HloModuleConfig& hlo_module_config, - const string& libdevice_dir_path); +StatusOr CompileToPtx( + llvm::Module* module, GpuVersion gpu_version, + const HloModuleConfig& hlo_module_config, const string& libdevice_dir_path, + std::function configure_target = nullptr); } // namespace nvptx namespace amdgpu { From c67719c6c34207e0e8a1f5e6b429d20130cd71f9 Mon Sep 17 00:00:00 2001 From: Tres Popp Date: Tue, 23 Jun 2020 06:58:18 -0700 Subject: [PATCH 548/867] Integrate LLVM at https://github.com/llvm/llvm-project/commit/f1c671925b1c PiperOrigin-RevId: 317856949 Change-Id: I192ce19253524e7957d7fc0a83abaac4adb31772 --- tensorflow/workspace.bzl | 4 ++-- 1 file changed, 2 insertions(+), 2 deletions(-) diff --git a/tensorflow/workspace.bzl b/tensorflow/workspace.bzl index 7c47628818b..d34e7d973d3 100755 --- a/tensorflow/workspace.bzl +++ b/tensorflow/workspace.bzl @@ -710,8 +710,8 @@ def tf_repositories(path_prefix = "", tf_repo_name = ""): ) # Check out LLVM and MLIR from llvm-project. - LLVM_COMMIT = "7a55d984971c11daa55e9423934f98bdc9c04f2f" - LLVM_SHA256 = "6ef6d1f92f51936ed3027433ea26875ac4d2ac8eed88b1fbce472018c4fb7720" + LLVM_COMMIT = "f1c671925b1c60ded3e4e7b3c6b1ec984b2d9b93" + LLVM_SHA256 = "57fc8f0ab46bdfdff52b03c2196d658c094bc4179cd1cf9495becf6a8466123a" LLVM_URLS = [ "https://storage.googleapis.com/mirror.tensorflow.org/github.com/llvm/llvm-project/archive/{commit}.tar.gz".format(commit = LLVM_COMMIT), "https://github.com/llvm/llvm-project/archive/{commit}.tar.gz".format(commit = LLVM_COMMIT), From 18dc2474023de36cb686f9900d5cac85763cbf6f Mon Sep 17 00:00:00 2001 From: Yash Katariya Date: Tue, 23 Jun 2020 08:30:44 -0700 Subject: [PATCH 549/867] Save output in lite notebooks to stop them from being tested. PiperOrigin-RevId: 317870420 Change-Id: Ie69635ac083b7788e00acabf53ed3975f8b019f2 --- .../post_training_float16_quant.ipynb | 236 ++++++++++++---- .../post_training_integer_quant.ipynb | 178 +++++++++--- .../performance/post_training_quant.ipynb | 263 ++++++++++++++---- 3 files changed, 534 insertions(+), 143 deletions(-) diff --git a/tensorflow/lite/g3doc/performance/post_training_float16_quant.ipynb b/tensorflow/lite/g3doc/performance/post_training_float16_quant.ipynb index ef08902865e..6015d3e1a65 100644 --- a/tensorflow/lite/g3doc/performance/post_training_float16_quant.ipynb +++ b/tensorflow/lite/g3doc/performance/post_training_float16_quant.ipynb @@ -12,7 +12,7 @@ }, { "cell_type": "code", - "execution_count": 0, + "execution_count": 1, "metadata": { "cellView": "form", "colab": {}, @@ -105,7 +105,7 @@ }, { "cell_type": "code", - "execution_count": 0, + "execution_count": 2, "metadata": { "colab": {}, "colab_type": "code", @@ -124,13 +124,29 @@ }, { "cell_type": "code", - "execution_count": 0, + "execution_count": 3, "metadata": { - "colab": {}, + "colab": { + "height": 34 + }, "colab_type": "code", - "id": "c6nb7OPlXs_3" + "id": "c6nb7OPlXs_3", + "outputId": "be7e4e14-cd67-4554-e928-ad803f36dad9" }, - "outputs": [], + "outputs": [ + { + "data": { + "text/plain": [ + "tf.float16" + ] + }, + "execution_count": 3, + "metadata": { + "tags": [] + }, + "output_type": "execute_result" + } + ], "source": [ "tf.float16" ] @@ -147,13 +163,39 @@ }, { "cell_type": "code", - "execution_count": 0, + "execution_count": 4, "metadata": { - "colab": {}, + "colab": { + "height": 102 + }, "colab_type": "code", - "id": "hWSAjQWagIHl" + "id": "hWSAjQWagIHl", + "outputId": "9bf2b530-5a05-415f-f856-cab3642256e9" }, - "outputs": [], + "outputs": [ + { + "name": "stdout", + "output_type": "stream", + "text": [ + "Downloading data from https://storage.googleapis.com/tensorflow/tf-keras-datasets/mnist.npz\n", + "11493376/11490434 [==============================] - 0s 0us/step\n", + "11501568/11490434 [==============================] - 0s 0us/step\n", + "1875/1875 [==============================] - 12s 6ms/step - loss: 0.2864 - accuracy: 0.9207 - val_loss: 0.1467 - val_accuracy: 0.9560\n" + ] + }, + { + "data": { + "text/plain": [ + "\u003ctensorflow.python.keras.callbacks.History at 0x7fcd75df46a0\u003e" + ] + }, + "execution_count": 4, + "metadata": { + "tags": [] + }, + "output_type": "execute_result" + } + ], "source": [ "# Load MNIST dataset\n", "mnist = keras.datasets.mnist\n", @@ -211,7 +253,7 @@ }, { "cell_type": "code", - "execution_count": 0, + "execution_count": 5, "metadata": { "colab": {}, "colab_type": "code", @@ -235,7 +277,7 @@ }, { "cell_type": "code", - "execution_count": 0, + "execution_count": 6, "metadata": { "colab": {}, "colab_type": "code", @@ -249,13 +291,29 @@ }, { "cell_type": "code", - "execution_count": 0, + "execution_count": 7, "metadata": { - "colab": {}, + "colab": { + "height": 34 + }, "colab_type": "code", - "id": "Ie9pQaQrn5ue" + "id": "Ie9pQaQrn5ue", + "outputId": "5df7381a-78ee-4f3e-e1a9-0f3a028384cf" }, - "outputs": [], + "outputs": [ + { + "data": { + "text/plain": [ + "84452" + ] + }, + "execution_count": 7, + "metadata": { + "tags": [] + }, + "output_type": "execute_result" + } + ], "source": [ "tflite_model_file = tflite_models_dir/\"mnist_model.tflite\"\n", "tflite_model_file.write_bytes(tflite_model)" @@ -273,7 +331,7 @@ }, { "cell_type": "code", - "execution_count": 0, + "execution_count": 8, "metadata": { "colab": {}, "colab_type": "code", @@ -297,13 +355,29 @@ }, { "cell_type": "code", - "execution_count": 0, + "execution_count": 9, "metadata": { - "colab": {}, + "colab": { + "height": 34 + }, "colab_type": "code", - "id": "yuNfl3CoHNK3" + "id": "yuNfl3CoHNK3", + "outputId": "839f02cd-0a8c-4551-aaa3-0c05c845ad2e" }, - "outputs": [], + "outputs": [ + { + "data": { + "text/plain": [ + "44272" + ] + }, + "execution_count": 9, + "metadata": { + "tags": [] + }, + "output_type": "execute_result" + } + ], "source": [ "tflite_fp16_model = converter.convert()\n", "tflite_model_fp16_file = tflite_models_dir/\"mnist_model_quant_f16.tflite\"\n", @@ -322,13 +396,26 @@ }, { "cell_type": "code", - "execution_count": 0, + "execution_count": 10, "metadata": { - "colab": {}, + "colab": { + "height": 68 + }, "colab_type": "code", - "id": "JExfcfLDscu4" + "id": "JExfcfLDscu4", + "outputId": "6ca316c2-cb0e-40e9-ffb1-a8bcf267e101" }, - "outputs": [], + "outputs": [ + { + "name": "stdout", + "output_type": "stream", + "text": [ + "total 128K\n", + "-rw-rw-r-- 1 colaboratory-playground 50844828 44K Jun 23 06:04 mnist_model_quant_f16.tflite\n", + "-rw-rw-r-- 1 colaboratory-playground 50844828 83K Jun 23 06:04 mnist_model.tflite\n" + ] + } + ], "source": [ "!ls -lh {tflite_models_dir}" ] @@ -365,7 +452,7 @@ }, { "cell_type": "code", - "execution_count": 0, + "execution_count": 11, "metadata": { "colab": {}, "colab_type": "code", @@ -379,7 +466,7 @@ }, { "cell_type": "code", - "execution_count": 0, + "execution_count": 12, "metadata": { "colab": {}, "colab_type": "code", @@ -403,7 +490,7 @@ }, { "cell_type": "code", - "execution_count": 0, + "execution_count": 13, "metadata": { "colab": {}, "colab_type": "code", @@ -423,13 +510,29 @@ }, { "cell_type": "code", - "execution_count": 0, + "execution_count": 14, "metadata": { - "colab": {}, + "colab": { + "height": 281 + }, "colab_type": "code", - "id": "XZClM2vo3_bm" + "id": "XZClM2vo3_bm", + "outputId": "fec12377-9f68-45a7-b4a6-ad902d8db171" }, - "outputs": [], + "outputs": [ + { + "data": { + "image/png": "iVBORw0KGgoAAAANSUhEUgAAAP8AAAEICAYAAACQ6CLfAAAABHNCSVQICAgIfAhkiAAAAAlwSFlz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\nxfATSYrhJ5LU/wOdAGX9nfSgHgAAAABJRU5ErkJggg==\n", + "text/plain": [ + "\u003cFigure size 600x400 with 1 Axes\u003e" + ] + }, + "metadata": { + "tags": [] + }, + "output_type": "display_data" + } + ], "source": [ "import matplotlib.pylab as plt\n", "\n", @@ -442,7 +545,7 @@ }, { "cell_type": "code", - "execution_count": 0, + "execution_count": 15, "metadata": { "colab": {}, "colab_type": "code", @@ -462,13 +565,29 @@ }, { "cell_type": "code", - "execution_count": 0, + "execution_count": 16, "metadata": { - "colab": {}, + "colab": { + "height": 281 + }, "colab_type": "code", - "id": "CIH7G_MwbY2x" + "id": "CIH7G_MwbY2x", + "outputId": "6a65e499-6618-4b3e-94f6-1d12af8fb251" }, - "outputs": [], + "outputs": [ + { + "data": { + "image/png": "iVBORw0KGgoAAAANSUhEUgAAAP8AAAEICAYAAACQ6CLfAAAABHNCSVQICAgIfAhkiAAAAAlwSFlz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\nxfATSYrhJ5LU/wOdAGX9nfSgHgAAAABJRU5ErkJggg==\n", + "text/plain": [ + "\u003cFigure size 600x400 with 1 Axes\u003e" + ] + }, + "metadata": { + "tags": [] + }, + "output_type": "display_data" + } + ], "source": [ "plt.imshow(test_images[0])\n", "template = \"True:{true}, predicted:{predict}\"\n", @@ -489,7 +608,7 @@ }, { "cell_type": "code", - "execution_count": 0, + "execution_count": 17, "metadata": { "colab": {}, "colab_type": "code", @@ -531,13 +650,24 @@ }, { "cell_type": "code", - "execution_count": 0, + "execution_count": 18, "metadata": { - "colab": {}, + "colab": { + "height": 34 + }, "colab_type": "code", - "id": "T5mWkSbMcU5z" + "id": "T5mWkSbMcU5z", + "outputId": "818e9142-70cf-420b-8e64-38c2ca11a370" }, - "outputs": [], + "outputs": [ + { + "name": "stdout", + "output_type": "stream", + "text": [ + "0.956\n" + ] + } + ], "source": [ "print(evaluate_model(interpreter))" ] @@ -554,13 +684,24 @@ }, { "cell_type": "code", - "execution_count": 0, + "execution_count": 19, "metadata": { - "colab": {}, + "colab": { + "height": 34 + }, "colab_type": "code", - "id": "-9cnwiPp6EGm" + "id": "-9cnwiPp6EGm", + "outputId": "53e00eac-51af-4030-be1a-3df986640f8d" }, - "outputs": [], + "outputs": [ + { + "name": "stdout", + "output_type": "stream", + "text": [ + "0.956\n" + ] + } + ], "source": [ "# NOTE: Colab runs on server CPUs. At the time of writing this, TensorFlow Lite\n", "# doesn't have super optimized server CPU kernels. For this reason this may be\n", @@ -599,8 +740,7 @@ "metadata": { "colab": { "collapsed_sections": [], - "name": "post_training-float16-quant.ipynb", - "private_outputs": true, + "name": "post_training_float16_quant.ipynb", "provenance": [], "toc_visible": true }, diff --git a/tensorflow/lite/g3doc/performance/post_training_integer_quant.ipynb b/tensorflow/lite/g3doc/performance/post_training_integer_quant.ipynb index ad461f56d6f..8397dbfa69f 100644 --- a/tensorflow/lite/g3doc/performance/post_training_integer_quant.ipynb +++ b/tensorflow/lite/g3doc/performance/post_training_integer_quant.ipynb @@ -12,7 +12,7 @@ }, { "cell_type": "code", - "execution_count": 0, + "execution_count": 17, "metadata": { "cellView": "form", "colab": {}, @@ -106,7 +106,7 @@ }, { "cell_type": "code", - "execution_count": 0, + "execution_count": 18, "metadata": { "colab": {}, "colab_type": "code", @@ -135,13 +135,36 @@ }, { "cell_type": "code", - "execution_count": 0, + "execution_count": 19, "metadata": { - "colab": {}, + "colab": { + "height": 51 + }, "colab_type": "code", - "id": "eMsw_6HujaqM" + "id": "eMsw_6HujaqM", + "outputId": "5662a5f3-fc64-458f-958a-98f9c6348143" }, - "outputs": [], + "outputs": [ + { + "name": "stdout", + "output_type": "stream", + "text": [ + "1875/1875 [==============================] - 2s 1ms/step - loss: 0.2782 - accuracy: 0.9221 - val_loss: 0.1230 - val_accuracy: 0.9664\n" + ] + }, + { + "data": { + "text/plain": [ + "\u003ctensorflow.python.keras.callbacks.History at 0x7f33f1817588\u003e" + ] + }, + "execution_count": 19, + "metadata": { + "tags": [] + }, + "output_type": "execute_result" + } + ], "source": [ "# Load MNIST dataset\n", "mnist = keras.datasets.mnist\n", @@ -199,7 +222,7 @@ }, { "cell_type": "code", - "execution_count": 0, + "execution_count": 20, "metadata": { "colab": {}, "colab_type": "code", @@ -223,7 +246,7 @@ }, { "cell_type": "code", - "execution_count": 0, + "execution_count": 21, "metadata": { "colab": {}, "colab_type": "code", @@ -237,13 +260,29 @@ }, { "cell_type": "code", - "execution_count": 0, + "execution_count": 22, "metadata": { - "colab": {}, + "colab": { + "height": 34 + }, "colab_type": "code", - "id": "Ie9pQaQrn5ue" + "id": "Ie9pQaQrn5ue", + "outputId": "8580b835-61f0-42b3-a21e-b8d476042c11" }, - "outputs": [], + "outputs": [ + { + "data": { + "text/plain": [ + "84528" + ] + }, + "execution_count": 22, + "metadata": { + "tags": [] + }, + "output_type": "execute_result" + } + ], "source": [ "tflite_model_file = tflite_models_dir/\"mnist_model.tflite\"\n", "tflite_model_file.write_bytes(tflite_model)" @@ -266,7 +305,7 @@ }, { "cell_type": "code", - "execution_count": 0, + "execution_count": 23, "metadata": { "colab": {}, "colab_type": "code", @@ -289,7 +328,7 @@ }, { "cell_type": "code", - "execution_count": 0, + "execution_count": 24, "metadata": { "colab": {}, "colab_type": "code", @@ -319,13 +358,29 @@ }, { "cell_type": "code", - "execution_count": 0, + "execution_count": 25, "metadata": { - "colab": {}, + "colab": { + "height": 34 + }, "colab_type": "code", - "id": "yuNfl3CoHNK3" + "id": "yuNfl3CoHNK3", + "outputId": "79a19679-87a2-4dc6-eee4-b33f3e5c1c5d" }, - "outputs": [], + "outputs": [ + { + "data": { + "text/plain": [ + "24720" + ] + }, + "execution_count": 25, + "metadata": { + "tags": [] + }, + "output_type": "execute_result" + } + ], "source": [ "tflite_model_quant = converter.convert()\n", "tflite_model_quant_file = tflite_models_dir/\"mnist_model_quant.tflite\"\n", @@ -344,13 +399,27 @@ }, { "cell_type": "code", - "execution_count": 0, + "execution_count": 26, "metadata": { - "colab": {}, + "colab": { + "height": 85 + }, "colab_type": "code", - "id": "JExfcfLDscu4" + "id": "JExfcfLDscu4", + "outputId": "58238f92-01b0-4faa-e293-35451d08dd7c" }, - "outputs": [], + "outputs": [ + { + "name": "stdout", + "output_type": "stream", + "text": [ + "total 140K\n", + "-rw-rw-r-- 1 yashkatariya 10086651 25K Jun 23 06:06 mnist_model_quant_io.tflite\n", + "-rw-rw-r-- 1 yashkatariya 10086651 25K Jun 23 06:07 mnist_model_quant.tflite\n", + "-rw-rw-r-- 1 yashkatariya 10086651 83K Jun 23 06:06 mnist_model.tflite\n" + ] + } + ], "source": [ "!ls -lh {tflite_models_dir}" ] @@ -369,13 +438,29 @@ }, { "cell_type": "code", - "execution_count": 0, + "execution_count": 27, "metadata": { - "colab": {}, + "colab": { + "height": 34 + }, "colab_type": "code", - "id": "kzjEjcDs3BHa" + "id": "kzjEjcDs3BHa", + "outputId": "8d7370ec-3f3f-41a2-8afb-4ecdd40e9efc" }, - "outputs": [], + "outputs": [ + { + "data": { + "text/plain": [ + "24784" + ] + }, + "execution_count": 27, + "metadata": { + "tags": [] + }, + "output_type": "execute_result" + } + ], "source": [ "converter.target_spec.supported_ops = [tf.lite.OpsSet.TFLITE_BUILTINS_INT8]\n", "converter.inference_input_type = tf.uint8\n", @@ -423,7 +508,7 @@ }, { "cell_type": "code", - "execution_count": 0, + "execution_count": 28, "metadata": { "colab": {}, "colab_type": "code", @@ -437,7 +522,7 @@ }, { "cell_type": "code", - "execution_count": 0, + "execution_count": 29, "metadata": { "colab": {}, "colab_type": "code", @@ -465,7 +550,7 @@ }, { "cell_type": "code", - "execution_count": 0, + "execution_count": 30, "metadata": { "colab": {}, "colab_type": "code", @@ -484,13 +569,29 @@ }, { "cell_type": "code", - "execution_count": 0, + "execution_count": 31, "metadata": { - "colab": {}, + "colab": { + "height": 281 + }, "colab_type": "code", - "id": "XZClM2vo3_bm" + "id": "XZClM2vo3_bm", + "outputId": "3af2e31c-44c6-41f2-c51f-da9d7b71bdfb" }, - "outputs": [], + "outputs": [ + { + "data": { + "image/png": "iVBORw0KGgoAAAANSUhEUgAAAP8AAAEICAYAAACQ6CLfAAAABHNCSVQICAgIfAhkiAAAAAlwSFlz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\nxfATSYrhJ5LU/wOdAGX9nfSgHgAAAABJRU5ErkJggg==\n", + "text/plain": [ + "\u003cFigure size 600x400 with 1 Axes\u003e" + ] + }, + "metadata": { + "tags": [] + }, + "output_type": "display_data" + } + ], "source": [ "import matplotlib.pylab as plt\n", "\n", @@ -513,7 +614,7 @@ }, { "cell_type": "code", - "execution_count": 0, + "execution_count": 33, "metadata": { "colab": {}, "colab_type": "code", @@ -530,7 +631,7 @@ }, { "cell_type": "code", - "execution_count": 0, + "execution_count": null, "metadata": { "colab": {}, "colab_type": "code", @@ -557,7 +658,7 @@ }, { "cell_type": "code", - "execution_count": 0, + "execution_count": null, "metadata": { "colab": {}, "colab_type": "code", @@ -599,7 +700,7 @@ }, { "cell_type": "code", - "execution_count": 0, + "execution_count": null, "metadata": { "colab": {}, "colab_type": "code", @@ -622,7 +723,7 @@ }, { "cell_type": "code", - "execution_count": 0, + "execution_count": null, "metadata": { "colab": {}, "colab_type": "code", @@ -653,7 +754,6 @@ "colab": { "collapsed_sections": [], "name": "post_training_integer_quant.ipynb", - "private_outputs": true, "provenance": [], "toc_visible": true }, diff --git a/tensorflow/lite/g3doc/performance/post_training_quant.ipynb b/tensorflow/lite/g3doc/performance/post_training_quant.ipynb index 201ccf5bdc3..5341fe5e4fb 100644 --- a/tensorflow/lite/g3doc/performance/post_training_quant.ipynb +++ b/tensorflow/lite/g3doc/performance/post_training_quant.ipynb @@ -12,7 +12,7 @@ }, { "cell_type": "code", - "execution_count": 0, + "execution_count": 1, "metadata": { "cellView": "form", "colab": {}, @@ -126,7 +126,7 @@ }, { "cell_type": "code", - "execution_count": 0, + "execution_count": 2, "metadata": { "colab": {}, "colab_type": "code", @@ -155,13 +155,36 @@ }, { "cell_type": "code", - "execution_count": 0, + "execution_count": 3, "metadata": { - "colab": {}, + "colab": { + "height": 51 + }, "colab_type": "code", - "id": "hWSAjQWagIHl" + "id": "hWSAjQWagIHl", + "outputId": "961899f8-1597-4417-b21d-cae94a330ecc" }, - "outputs": [], + "outputs": [ + { + "name": "stdout", + "output_type": "stream", + "text": [ + "1875/1875 [==============================] - 10s 5ms/step - loss: 0.2787 - accuracy: 0.9203 - val_loss: 0.1323 - val_accuracy: 0.9624\n" + ] + }, + { + "data": { + "text/plain": [ + "\u003ctensorflow.python.keras.callbacks.History at 0x7f6443480e80\u003e" + ] + }, + "execution_count": 3, + "metadata": { + "tags": [] + }, + "output_type": "execute_result" + } + ], "source": [ "# Load MNIST dataset\n", "mnist = keras.datasets.mnist\n", @@ -200,8 +223,7 @@ "id": "5NMaNZQCkW9X" }, "source": [ - "For the example, since you trained the model for just a single epoch, so it only trains to ~96% accuracy.\n", - "\n" + "For the example, since you trained the model for just a single epoch, so it only trains to ~96% accuracy.\n" ] }, { @@ -220,7 +242,7 @@ }, { "cell_type": "code", - "execution_count": 0, + "execution_count": 4, "metadata": { "colab": {}, "colab_type": "code", @@ -244,7 +266,7 @@ }, { "cell_type": "code", - "execution_count": 0, + "execution_count": 5, "metadata": { "colab": {}, "colab_type": "code", @@ -258,13 +280,29 @@ }, { "cell_type": "code", - "execution_count": 0, + "execution_count": 6, "metadata": { - "colab": {}, + "colab": { + "height": 34 + }, "colab_type": "code", - "id": "Ie9pQaQrn5ue" + "id": "Ie9pQaQrn5ue", + "outputId": "046db0bc-1745-4e94-9f21-f7e91bdaebda" }, - "outputs": [], + "outputs": [ + { + "data": { + "text/plain": [ + "84452" + ] + }, + "execution_count": 6, + "metadata": { + "tags": [] + }, + "output_type": "execute_result" + } + ], "source": [ "tflite_model_file = tflite_models_dir/\"mnist_model.tflite\"\n", "tflite_model_file.write_bytes(tflite_model)" @@ -282,13 +320,29 @@ }, { "cell_type": "code", - "execution_count": 0, + "execution_count": 7, "metadata": { - "colab": {}, + "colab": { + "height": 34 + }, "colab_type": "code", - "id": "g8PUvLWDlmmz" + "id": "g8PUvLWDlmmz", + "outputId": "d79b45d3-babf-4890-8036-de2f497da88a" }, - "outputs": [], + "outputs": [ + { + "data": { + "text/plain": [ + "23840" + ] + }, + "execution_count": 7, + "metadata": { + "tags": [] + }, + "output_type": "execute_result" + } + ], "source": [ "converter.optimizations = [tf.lite.Optimize.DEFAULT]\n", "tflite_quant_model = converter.convert()\n", @@ -308,13 +362,29 @@ }, { "cell_type": "code", - "execution_count": 0, + "execution_count": 8, "metadata": { - "colab": {}, + "colab": { + "height": 119 + }, "colab_type": "code", - "id": "JExfcfLDscu4" + "id": "JExfcfLDscu4", + "outputId": "d1fda4c2-343e-40fb-f90f-b6bde00c523e" }, - "outputs": [], + "outputs": [ + { + "name": "stdout", + "output_type": "stream", + "text": [ + "total 214M\n", + "-rw-rw-r-- 1 colaboratory-playground 50844828 44K Jun 23 06:04 mnist_model_quant_f16.tflite\n", + "-rw-rw-r-- 1 colaboratory-playground 50844828 24K Jun 23 06:12 mnist_model_quant.tflite\n", + "-rw-rw-r-- 1 colaboratory-playground 50844828 83K Jun 23 06:12 mnist_model.tflite\n", + "-rw-rw-r-- 1 colaboratory-playground 50844828 44M Jun 23 06:10 resnet_v2_101_quantized.tflite\n", + "-rw-rw-r-- 1 colaboratory-playground 50844828 171M Jun 23 06:09 resnet_v2_101.tflite\n" + ] + } + ], "source": [ "!ls -lh {tflite_models_dir}" ] @@ -329,8 +399,7 @@ "## Run the TFLite models\n", "\n", "Run the TensorFlow Lite model using the Python TensorFlow Lite\n", - "Interpreter.\n", - "\n" + "Interpreter.\n" ] }, { @@ -345,7 +414,7 @@ }, { "cell_type": "code", - "execution_count": 0, + "execution_count": 9, "metadata": { "colab": {}, "colab_type": "code", @@ -359,7 +428,7 @@ }, { "cell_type": "code", - "execution_count": 0, + "execution_count": 10, "metadata": { "colab": {}, "colab_type": "code", @@ -383,7 +452,7 @@ }, { "cell_type": "code", - "execution_count": 0, + "execution_count": 11, "metadata": { "colab": {}, "colab_type": "code", @@ -403,13 +472,29 @@ }, { "cell_type": "code", - "execution_count": 0, + "execution_count": 12, "metadata": { - "colab": {}, + "colab": { + "height": 281 + }, "colab_type": "code", - "id": "XZClM2vo3_bm" + "id": "XZClM2vo3_bm", + "outputId": "0fa4155b-01f8-4fea-f586-d9044d73572e" }, - "outputs": [], + "outputs": [ + { + "data": { + "image/png": "iVBORw0KGgoAAAANSUhEUgAAAP8AAAEICAYAAACQ6CLfAAAABHNCSVQICAgIfAhkiAAAAAlwSFlz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\nxfATSYrhJ5LU/wOdAGX9nfSgHgAAAABJRU5ErkJggg==\n", + "text/plain": [ + "\u003cFigure size 600x400 with 1 Axes\u003e" + ] + }, + "metadata": { + "tags": [] + }, + "output_type": "display_data" + } + ], "source": [ "import matplotlib.pylab as plt\n", "\n", @@ -432,7 +517,7 @@ }, { "cell_type": "code", - "execution_count": 0, + "execution_count": 13, "metadata": { "colab": {}, "colab_type": "code", @@ -474,13 +559,24 @@ }, { "cell_type": "code", - "execution_count": 0, + "execution_count": 14, "metadata": { - "colab": {}, + "colab": { + "height": 34 + }, "colab_type": "code", - "id": "DqXBnDfJ7qxL" + "id": "DqXBnDfJ7qxL", + "outputId": "78f393f8-c4a5-41e0-abe4-ab6a5c394e51" }, - "outputs": [], + "outputs": [ + { + "name": "stdout", + "output_type": "stream", + "text": [ + "0.9624\n" + ] + } + ], "source": [ "print(evaluate_model(interpreter))" ] @@ -497,13 +593,24 @@ }, { "cell_type": "code", - "execution_count": 0, + "execution_count": 15, "metadata": { - "colab": {}, + "colab": { + "height": 34 + }, "colab_type": "code", - "id": "-9cnwiPp6EGm" + "id": "-9cnwiPp6EGm", + "outputId": "d82552d7-8a2c-49dc-a19a-56010a013102" }, - "outputs": [], + "outputs": [ + { + "name": "stdout", + "output_type": "stream", + "text": [ + "0.9626\n" + ] + } + ], "source": [ "print(evaluate_model(interpreter_quant))" ] @@ -515,7 +622,6 @@ "id": "L7lfxkor8pgv" }, "source": [ - "\n", "In this example, the compressed model has no difference in the accuracy." ] }, @@ -537,7 +643,7 @@ }, { "cell_type": "code", - "execution_count": 0, + "execution_count": 16, "metadata": { "colab": {}, "colab_type": "code", @@ -557,13 +663,29 @@ }, { "cell_type": "code", - "execution_count": 0, + "execution_count": 17, "metadata": { - "colab": {}, + "colab": { + "height": 34 + }, "colab_type": "code", - "id": "LwnV4KxwVEoG" + "id": "LwnV4KxwVEoG", + "outputId": "7d50f90d-6104-43a3-863c-28db9465d483" }, - "outputs": [], + "outputs": [ + { + "data": { + "text/plain": [ + "178509092" + ] + }, + "execution_count": 17, + "metadata": { + "tags": [] + }, + "output_type": "execute_result" + } + ], "source": [ "# Convert to TF Lite without quantization\n", "resnet_tflite_file = tflite_models_dir/\"resnet_v2_101.tflite\"\n", @@ -572,13 +694,29 @@ }, { "cell_type": "code", - "execution_count": 0, + "execution_count": 18, "metadata": { - "colab": {}, + "colab": { + "height": 34 + }, "colab_type": "code", - "id": "2qkZD0VoVExe" + "id": "2qkZD0VoVExe", + "outputId": "76a47590-fa91-49b9-f568-4e00b46c9537" }, - "outputs": [], + "outputs": [ + { + "data": { + "text/plain": [ + "45182656" + ] + }, + "execution_count": 18, + "metadata": { + "tags": [] + }, + "output_type": "execute_result" + } + ], "source": [ "# Convert to TF Lite with quantization\n", "converter.optimizations = [tf.lite.Optimize.DEFAULT]\n", @@ -588,13 +726,28 @@ }, { "cell_type": "code", - "execution_count": 0, + "execution_count": 19, "metadata": { - "colab": {}, + "colab": { + "height": 102 + }, "colab_type": "code", - "id": "vhOjeg1x9Knp" + "id": "vhOjeg1x9Knp", + "outputId": "c643a660-f815-49f0-ac4b-ac48af3c1203" }, - "outputs": [], + "outputs": [ + { + "name": "stdout", + "output_type": "stream", + "text": [ + "-rw-rw-r-- 1 colaboratory-playground 50844828 44K Jun 23 06:04 /tmp/mnist_tflite_models/mnist_model_quant_f16.tflite\n", + "-rw-rw-r-- 1 colaboratory-playground 50844828 24K Jun 23 06:12 /tmp/mnist_tflite_models/mnist_model_quant.tflite\n", + "-rw-rw-r-- 1 colaboratory-playground 50844828 83K Jun 23 06:12 /tmp/mnist_tflite_models/mnist_model.tflite\n", + "-rw-rw-r-- 1 colaboratory-playground 50844828 44M Jun 23 06:13 /tmp/mnist_tflite_models/resnet_v2_101_quantized.tflite\n", + "-rw-rw-r-- 1 colaboratory-playground 50844828 171M Jun 23 06:12 /tmp/mnist_tflite_models/resnet_v2_101.tflite\n" + ] + } + ], "source": [ "!ls -lh {tflite_models_dir}/*.tflite" ] @@ -606,7 +759,6 @@ "id": "qqHLaqFMCjRZ" }, "source": [ - "\n", "The model size reduces from 171 MB to 43 MB.\n", "The accuracy of this model on imagenet can be evaluated using the scripts provided for [TFLite accuracy measurement](https://github.com/tensorflow/tensorflow/tree/master/tensorflow/lite/tools/accuracy/ilsvrc).\n", "\n", @@ -618,7 +770,6 @@ "colab": { "collapsed_sections": [], "name": "post_training_quant.ipynb", - "private_outputs": true, "provenance": [], "toc_visible": true }, From 89851a6a7725d92735e9817ee6a1f551dd7492ab Mon Sep 17 00:00:00 2001 From: "A. Unique TensorFlower" Date: Tue, 23 Jun 2020 08:42:05 -0700 Subject: [PATCH 550/867] Replace deprecated thread annotations macros. PiperOrigin-RevId: 317872415 Change-Id: I2e3fcd5cebc7a2b4d63c1322a2386484b336479c --- tensorflow/compiler/xla/pjrt/distributed/service.h | 10 +++++----- tensorflow/compiler/xla/pjrt/nvidia_gpu_device.cc | 2 +- 2 files changed, 6 insertions(+), 6 deletions(-) diff --git a/tensorflow/compiler/xla/pjrt/distributed/service.h b/tensorflow/compiler/xla/pjrt/distributed/service.h index 725a76791ce..9ecbdb3cc7c 100644 --- a/tensorflow/compiler/xla/pjrt/distributed/service.h +++ b/tensorflow/compiler/xla/pjrt/distributed/service.h @@ -54,15 +54,15 @@ class DistributedRuntimeServiceImpl final absl::Mutex mu_; enum class State { kInitializing, kRunning }; - State state_ GUARDED_BY(mu_) = State::kInitializing; + State state_ ABSL_GUARDED_BY(mu_) = State::kInitializing; - std::vector local_topologies_ GUARDED_BY(mu_); - GlobalTopologyProto topology_ GUARDED_BY(mu_); + std::vector local_topologies_ ABSL_GUARDED_BY(mu_); + GlobalTopologyProto topology_ ABSL_GUARDED_BY(mu_); struct Node { bool present = false; }; - int num_nodes_present_ GUARDED_BY(mu_) = 0; - std::vector nodes_ GUARDED_BY(mu_); + int num_nodes_present_ ABSL_GUARDED_BY(mu_) = 0; + std::vector nodes_ ABSL_GUARDED_BY(mu_); KeyValueStore key_value_store_; }; diff --git a/tensorflow/compiler/xla/pjrt/nvidia_gpu_device.cc b/tensorflow/compiler/xla/pjrt/nvidia_gpu_device.cc index de760af8fd9..edffaf6c877 100644 --- a/tensorflow/compiler/xla/pjrt/nvidia_gpu_device.cc +++ b/tensorflow/compiler/xla/pjrt/nvidia_gpu_device.cc @@ -169,7 +169,7 @@ class NcclIdStore { const std::shared_ptr client_; absl::Mutex mu_; - absl::flat_hash_map cache_ GUARDED_BY(mu_); + absl::flat_hash_map cache_ ABSL_GUARDED_BY(mu_); }; StatusOr NcclIdStore::GetNcclUniqueId(const NcclCliqueKey& key) { From f5a9d24c847ffcc7ae09e850aad39e2cb55ae4f3 Mon Sep 17 00:00:00 2001 From: "A. Unique TensorFlower" Date: Tue, 23 Jun 2020 08:57:59 -0700 Subject: [PATCH 551/867] Added shape inference for tf_device.LaunchOp. PiperOrigin-RevId: 317875384 Change-Id: Idb070c9e92d07ee19cd8ed26c1beec3de86f43df --- .../mlir/tensorflow/tests/shape_inference.mlir | 13 +++++++++++++ .../mlir/tensorflow/transforms/shape_inference.cc | 4 ++++ 2 files changed, 17 insertions(+) diff --git a/tensorflow/compiler/mlir/tensorflow/tests/shape_inference.mlir b/tensorflow/compiler/mlir/tensorflow/tests/shape_inference.mlir index 1599d53ed15..1af4ba6b3dc 100644 --- a/tensorflow/compiler/mlir/tensorflow/tests/shape_inference.mlir +++ b/tensorflow/compiler/mlir/tensorflow/tests/shape_inference.mlir @@ -433,4 +433,17 @@ func @multiple_blocks_one_return(%arg0: tensor) -> tensor<*xf32> { // CHECK: return %[[CAST_RESULT_0]], %[[CAST_RESULT_1]], %[[ADDI]] return %27, %28, %2 : tensor<*xui8>, tensor<*xi8>, tensor<*xi8> } + + // CHECK-LABEL: infer_device_launch + func @infer_device_launch(%arg0: tensor<1x8x2xi32>) -> (tensor<*xf32>, tensor<*xf32>) { + %0 = "tf.Const"() {value = dense<-1> : tensor} : () -> tensor + %1 = "tf_device.launch"() ({ + %2 = "tf.Cast"(%arg0) {Truncate = false} : (tensor<1x8x2xi32>) -> tensor<1x8x2xf32> + tf_device.return %2 : tensor<1x8x2xf32> + // CHECK: () -> tensor<1x8x2xf32> + }) {device = "/device:CPU:0"} : () -> tensor<*xf32> + // CHECK: (tensor, tensor<1x8x2xf32>) -> (tensor<1x8x1xf32>, tensor<1x8x1xf32>) + %3:2 = "tf.Split"(%0, %1) {device = ""} : (tensor, tensor<*xf32>) -> (tensor<*xf32>, tensor<*xf32>) + return %3#0, %3#1 : tensor<*xf32>, tensor<*xf32> + } } diff --git a/tensorflow/compiler/mlir/tensorflow/transforms/shape_inference.cc b/tensorflow/compiler/mlir/tensorflow/transforms/shape_inference.cc index 7e4baadc397..33ccf5caff2 100644 --- a/tensorflow/compiler/mlir/tensorflow/transforms/shape_inference.cc +++ b/tensorflow/compiler/mlir/tensorflow/transforms/shape_inference.cc @@ -215,6 +215,10 @@ bool InferShapeForNonTFDialectOperation(Operation* op, Dialect* tf_dialect) { return InferShapeForPassThroughOps( tensor_cast.getOperation()->getOperands(), op, tf_dialect); } + if (auto launch_op = dyn_cast(op)) { + return InferShapeForPassThroughOps( + launch_op.GetBody().getTerminator()->getOperands(), op, tf_dialect); + } return false; } From a9e2bceddf4588f1ef0bae6fb4054111ff3d2225 Mon Sep 17 00:00:00 2001 From: Raman Sarokin Date: Tue, 23 Jun 2020 09:51:29 -0700 Subject: [PATCH 552/867] ConvConstants converted to new style. Added memory types for buffer. PiperOrigin-RevId: 317885492 Change-Id: I12dfede368bdf157f746f7823d7d47fd3807ffa1 --- tensorflow/lite/delegates/gpu/cl/arguments.cc | 16 +- tensorflow/lite/delegates/gpu/cl/buffer.cc | 14 ++ tensorflow/lite/delegates/gpu/cl/buffer.h | 3 + tensorflow/lite/delegates/gpu/cl/gpu_object.h | 3 + .../gpu/cl/kernels/conv_constants.cc | 143 ++++++++++-------- .../delegates/gpu/cl/kernels/conv_constants.h | 30 ++-- .../lite/delegates/gpu/cl/linear_storage.cc | 1 + .../lite/delegates/gpu/cl/linear_storage.h | 1 + 8 files changed, 132 insertions(+), 79 deletions(-) diff --git a/tensorflow/lite/delegates/gpu/cl/arguments.cc b/tensorflow/lite/delegates/gpu/cl/arguments.cc index d26c869957c..1fd58ef2454 100644 --- a/tensorflow/lite/delegates/gpu/cl/arguments.cc +++ b/tensorflow/lite/delegates/gpu/cl/arguments.cc @@ -457,8 +457,20 @@ std::string Arguments::GetListOfArgs() { for (auto& t : buffers_) { const std::string type_name = t.second.data_type == DataType::FLOAT32 ? "float" : "half"; - AppendArgument(absl::StrCat("__global ", type_name, t.second.element_size, - "* ", t.first), + std::string memory_type; + switch (t.second.memory_type) { + case MemoryType::GLOBAL: + memory_type = "__global"; + break; + case MemoryType::CONSTANT: + memory_type = "__constant"; + break; + case MemoryType::LOCAL: + memory_type = "__local"; + break; + } + AppendArgument(absl::StrCat(memory_type, " ", type_name, + t.second.element_size, "* ", t.first), &result); } for (auto& t : image_buffers_) { diff --git a/tensorflow/lite/delegates/gpu/cl/buffer.cc b/tensorflow/lite/delegates/gpu/cl/buffer.cc index 68e85593e5d..436d8751e18 100644 --- a/tensorflow/lite/delegates/gpu/cl/buffer.cc +++ b/tensorflow/lite/delegates/gpu/cl/buffer.cc @@ -50,6 +50,7 @@ GPUResources BufferDescriptor::GetGPUResources(AccessType access_type) const { desc.data_type = element_type; desc.access_type = access_type; desc.element_size = element_size; + desc.memory_type = memory_type; resources.buffers.push_back({"buffer", desc}); return resources; } @@ -59,6 +60,8 @@ absl::Status BufferDescriptor::PerformSelector( const std::vector& template_args, std::string* result) const { if (selector == "Read") { return PerformReadSelector(args, result); + } else if (selector == "GetPtr") { + return PerformGetPtrSelector(args, result); } else { return absl::NotFoundError(absl::StrCat( "BufferDescriptor don't have selector with name - ", selector)); @@ -76,6 +79,17 @@ absl::Status BufferDescriptor::PerformReadSelector( return absl::OkStatus(); } +absl::Status BufferDescriptor::PerformGetPtrSelector( + const std::vector& args, std::string* result) const { + if (!args.empty()) { + return absl::NotFoundError( + absl::StrCat("BufferDescriptor GetPtr require zero arguments, but ", + args.size(), " was passed")); + } + *result = "buffer"; + return absl::OkStatus(); +} + Buffer::Buffer(cl_mem buffer, size_t size_in_bytes) : buffer_(buffer), size_(size_in_bytes) {} diff --git a/tensorflow/lite/delegates/gpu/cl/buffer.h b/tensorflow/lite/delegates/gpu/cl/buffer.h index 771aae9e002..0d1072040c1 100644 --- a/tensorflow/lite/delegates/gpu/cl/buffer.h +++ b/tensorflow/lite/delegates/gpu/cl/buffer.h @@ -32,6 +32,7 @@ namespace cl { struct BufferDescriptor : public GPUObjectDescriptor { DataType element_type; // FLOAT32 or FLOAT16 int element_size; + MemoryType memory_type = MemoryType::GLOBAL; absl::Status PerformSelector(const std::string& selector, const std::vector& args, @@ -41,6 +42,8 @@ struct BufferDescriptor : public GPUObjectDescriptor { GPUResources GetGPUResources(AccessType access_type) const override; absl::Status PerformReadSelector(const std::vector& args, std::string* result) const; + absl::Status PerformGetPtrSelector(const std::vector& args, + std::string* result) const; }; // Buffer represent linear GPU data storage with arbitrary data format. diff --git a/tensorflow/lite/delegates/gpu/cl/gpu_object.h b/tensorflow/lite/delegates/gpu/cl/gpu_object.h index faf18b539e2..711c4726bc2 100644 --- a/tensorflow/lite/delegates/gpu/cl/gpu_object.h +++ b/tensorflow/lite/delegates/gpu/cl/gpu_object.h @@ -54,10 +54,13 @@ struct GPUImageBufferDescriptor { cl_mem memory; }; +enum class MemoryType { GLOBAL, CONSTANT, LOCAL }; + struct GPUBufferDescriptor { DataType data_type; AccessType access_type; int element_size; + MemoryType memory_type = MemoryType::GLOBAL; cl_mem memory; }; diff --git a/tensorflow/lite/delegates/gpu/cl/kernels/conv_constants.cc b/tensorflow/lite/delegates/gpu/cl/kernels/conv_constants.cc index d4dc206ffce..e6fc5da36a2 100644 --- a/tensorflow/lite/delegates/gpu/cl/kernels/conv_constants.cc +++ b/tensorflow/lite/delegates/gpu/cl/kernels/conv_constants.cc @@ -27,16 +27,29 @@ namespace gpu { namespace cl { namespace { -std::string GenerateConvolutionConstantCode( - const OperationDef& op_def, const int2& kernel_size, int src_channels, - int dst_channels, bool stride_correction, const CLDevice& device, - const std::vector& linked_operations) { - TensorCodeGenerator src_tensor( - "src_data", WHSPoint{"src_size.x", "src_size.y", "src_size.z"}, - op_def.src_tensors[0]); - TensorCodeGenerator dst_tensor( - "dst_data", WHSPoint{"dst_size.x", "dst_size.y", "dst_size.z"}, - op_def.dst_tensors[0]); +std::string GenerateConvolutionConstantCode(const OperationDef& op_def, + const int2& kernel_size, + int src_channels, int dst_channels, + bool stride_correction, + const CLDevice& device, + Arguments* args) { + auto src_desc = absl::make_unique(op_def.src_tensors[0]); + src_desc->SetTextureAddressMode(GetFastestZeroMode(device)); + if (op_def.IsBatchSupported()) { + src_desc->SetStateVar("BatchedWidth", "true"); + } + args->AddObjectRef("src_tensor", AccessType::READ, std::move(src_desc)); + auto dst_desc = absl::make_unique(op_def.dst_tensors[0]); + if (op_def.IsBatchSupported()) { + dst_desc->SetStateVar("BatchedWidth", "true"); + } + args->AddObjectRef("dst_tensor", AccessType::WRITE, std::move(dst_desc)); + args->AddInt("stride_x"); + args->AddInt("stride_y"); + args->AddInt("padding_x"); + args->AddInt("padding_y"); + args->AddInt("dilation_x"); + args->AddInt("dilation_y"); std::string c = GetCommonDefines(op_def.precision); @@ -89,33 +102,24 @@ std::string GenerateConvolutionConstantCode( const std::string postfixes[] = {".x", ".xy", ".xyz", ""}; c += "__kernel void main_function(\n"; - c += src_tensor.GetDeclaration(AccessType::READ) + ",\n"; - c += " __constant FLT4* filters, \n"; - c += " __constant FLT4* biases"; - c += GetArgsDeclaration(linked_operations); - c += dst_tensor.GetDeclaration(AccessType::WRITE) + ",\n"; - c += " int2 stride, \n"; - c += " int2 padding, \n"; - c += " int2 dilation, \n"; - c += " int4 src_size, \n"; - c += " int4 dst_size \n"; - c += ") {\n"; + c += "$0) {\n"; c += " int X = get_global_id(0);\n"; c += " int Y = get_global_id(1);\n"; - c += " if (X >= dst_size.x || Y >= dst_size.y) return;\n"; + c += " if (X >= args.dst_tensor.Width() || Y >= args.dst_tensor.Height()) " + "return;\n"; if (stride_correction) { c += " int start_x = " + - GetXStrideCorrected("X", "src_size.w", "stride.x", "padding.x") + + GetXStrideCorrected("X", "args.src_tensor.Batch()", "args.stride_x", + "args.padding_x") + ";\n"; } else { - c += " int start_x = X * stride.x + padding.x;\n"; + c += " int start_x = X * args.stride_x + args.padding_x;\n"; } - c += " int start_y = Y * stride.y + padding.y;\n"; + c += " int start_y = Y * args.stride_y + args.padding_y;\n"; c += " ACCUM_FLT4 r[" + kOutZ + "];\n"; c += " for (int i = 0; i < " + kOutZ + "; ++i) {\n"; c += " r[i] = (ACCUM_FLT4)(0.0f, 0.0f, 0.0f, 0.0f);\n"; c += " }\n"; - const auto address_mode = GetFastestZeroMode(device); int filters_counter = 0; for (int s = 0; s < src_depth; ++s) { const int ch_count = std::min(4, src_channels - s * 4); @@ -124,27 +128,29 @@ std::string GenerateConvolutionConstantCode( const std::string s_type = absl::StrCat("FLT", s_count); const std::string s_postfix = postfixes[ch_count - 1]; for (int ky = 0; ky < kernel_size.y; ++ky) { - std::string s_y = absl::StrCat("(start_y + ", ky, " * dilation.y)"); + std::string s_y = absl::StrCat("(start_y + ", ky, " * args.dilation_y)"); if (manual_clamp) { c += " {\n"; - c += " bool y_out = " + s_y + " < 0 || " + s_y + " >= src_size.y;\n"; + c += " bool y_out = " + s_y + " < 0 || " + s_y + + " >= args.src_tensor.Height();\n"; } for (int kx = 0; kx < kernel_size.x; ++kx) { c += " {\n"; - std::string s_x = absl::StrCat("(start_x + ", kx, " * dilation.x)"); + std::string s_x = + absl::StrCat("(start_x + ", kx, " * args.dilation_x)"); if (manual_clamp) { - c += " bool x_out = " + s_x + "< 0 || " + s_x + ">= src_size.x;\n"; + c += " bool x_out = " + s_x + "< 0 || " + s_x + + ">= args.src_tensor.Width();\n"; c += " " + s_type + " src = x_out || y_out ?"; - c += "(" + s_type + ")(0.0) : "; - c += src_tensor.ReadWHS(s_x, s_y, std::to_string(s)) + s_postfix + - ";\n"; + c += "(" + s_type + ")(0.0) : args.src_tensor.Read(" + s_x + ", " + + s_y + ", " + std::to_string(s) + ")" + s_postfix + ";\n"; } else { - c += " " + s_type + " src = " + - src_tensor.ReadWHS(s_x, s_y, std::to_string(s), address_mode) + - s_postfix + ";\n"; + c += " " + s_type + " src = args.src_tensor.Read(" + s_x + ", " + + s_y + ", " + std::to_string(s) + ")" + s_postfix + ";\n"; } for (int d = 0; d < out_z; ++d) { - c += " " + s_conv + "(r[" + std::to_string(d) + "], src, filters,"; + c += " " + s_conv + "(r[" + std::to_string(d) + + "], src, args.weigths.GetPtr(),"; c += " " + std::to_string(filters_counter) + ");\n"; filters_counter += ch_count; } @@ -158,10 +164,9 @@ std::string GenerateConvolutionConstantCode( for (int i = 0; i < out_z; ++i) { std::string s_i = std::to_string(i); c += " {\n"; - c += " FLT4 res = TO_FLT4(r[" + s_i + "]) + biases[" + s_i + "];\n"; - const LinkingContext context{"res", "X", "Y", s_i}; - c += PostProcess(linked_operations, context); - c += " " + dst_tensor.WriteWHS("res", "X", "Y", s_i); + c += " FLT4 res = TO_FLT4(r[" + s_i + "]) + args.biases.Read(" + s_i + + ");\n"; + c += " args.dst_tensor.Write(res, X, Y, " + s_i + ");\n"; c += " }\n"; } c += "}\n"; @@ -191,8 +196,6 @@ int GetOptimalMaxConstantSize(const DeviceInfo& info) { ConvConstants::ConvConstants(ConvConstants&& kernel) : GPUOperation(std::move(kernel)), - weights_(std::move(kernel.weights_)), - biases_(std::move(kernel.biases_)), kernel_size_(kernel.kernel_size_), stride_(kernel.stride_), padding_(kernel.padding_), @@ -204,8 +207,6 @@ ConvConstants::ConvConstants(ConvConstants&& kernel) ConvConstants& ConvConstants::operator=(ConvConstants&& kernel) { if (this != &kernel) { - weights_ = std::move(kernel.weights_); - biases_ = std::move(kernel.biases_); std::swap(kernel_size_, kernel.kernel_size_); std::swap(stride_, kernel.stride_); std::swap(padding_, kernel.padding_); @@ -222,9 +223,15 @@ ConvConstants& ConvConstants::operator=(ConvConstants&& kernel) { absl::Status ConvConstants::Compile(const CreationContext& creation_context) { const bool stride_correction = definition_.IsBatchSupported() && stride_.x != 1; - const auto code = GenerateConvolutionConstantCode( + std::string code = GenerateConvolutionConstantCode( definition_, kernel_size_, src_channels_, dst_channels_, - stride_correction, *creation_context.device, linked_operations_); + stride_correction, *creation_context.device, &args_); + std::string element_wise_code; + RETURN_IF_ERROR( + MergeOperations(linked_operations_, &args_, &element_wise_code)); + RETURN_IF_ERROR(args_.TransformToCLCode(creation_context.device->GetInfo(), + {{"dst_tensor", element_wise_code}}, + &code)); std::vector options; if (definition_.precision == CalculationsPrecision::F16 && creation_context.device->IsAdreno3xx()) { @@ -241,20 +248,16 @@ absl::Status ConvConstants::Compile(const CreationContext& creation_context) { } absl::Status ConvConstants::BindArguments() { - kernel_.ResetBindingCounter(); - RETURN_IF_ERROR(kernel_.SetMemoryAuto(src_[0]->GetMemoryPtr())); - RETURN_IF_ERROR(kernel_.SetMemoryAuto(weights_.GetMemoryPtr())); - RETURN_IF_ERROR(kernel_.SetMemoryAuto(biases_.GetMemoryPtr())); - RETURN_IF_ERROR(BindArgs(&kernel_, linked_operations_)); - RETURN_IF_ERROR(kernel_.SetMemoryAuto(dst_[0]->GetMemoryPtrForWriting())); - RETURN_IF_ERROR(kernel_.SetBytesAuto(stride_)); - RETURN_IF_ERROR( - kernel_.SetBytesAuto(int2(padding_.x * src_[0]->Batch(), padding_.y))); - RETURN_IF_ERROR( - kernel_.SetBytesAuto(int2(dilation_.x * src_[0]->Batch(), dilation_.y))); - RETURN_IF_ERROR(kernel_.SetBytesAuto(src_[0]->GetWBatchedHSB())); - RETURN_IF_ERROR(kernel_.SetBytesAuto(dst_[0]->GetWBatchedHSB())); - return absl::OkStatus(); + RETURN_IF_ERROR(args_.SetObjectRef("src_tensor", src_[0])); + RETURN_IF_ERROR(args_.SetObjectRef("dst_tensor", dst_[0])); + RETURN_IF_ERROR(args_.SetInt("stride_x", stride_.x)); + RETURN_IF_ERROR(args_.SetInt("stride_y", stride_.y)); + RETURN_IF_ERROR(args_.SetInt("padding_x", padding_.x * src_[0]->Batch())); + RETURN_IF_ERROR(args_.SetInt("padding_y", padding_.y)); + RETURN_IF_ERROR(args_.SetInt("dilation_x", dilation_.x * src_[0]->Batch())); + RETURN_IF_ERROR(args_.SetInt("dilation_y", dilation_.y)); + RETURN_IF_ERROR(SetArguments(linked_operations_, &args_)); + return args_.Bind(kernel_.kernel()); } int3 ConvConstants::GetGridSize() const { @@ -304,12 +307,18 @@ absl::Status CreateConvConstants(const CreationContext& creation_context, *result = ConvConstants(definition, attr); RETURN_IF_ERROR( result->UploadWeights(attr.weights, creation_context.context)); - LinearStorageCreateInfo create_info; - create_info.storage_type = LinearStorageType::BUFFER; - create_info.data_type = definition.GetDataType(); - create_info.aligned_size = attr.weights.shape.o; - RETURN_IF_ERROR(CreateLinearStorage( - create_info, attr.bias, creation_context.context, &result->biases_)); + + TensorLinearDescriptor desc; + desc.storage_type = LinearStorageType::BUFFER; + desc.element_type = definition.GetDataType(); + desc.memory_type = MemoryType::CONSTANT; + + LinearStorage lt; + RETURN_IF_ERROR( + CreateLinearStorage(desc, attr.bias, creation_context.context, <)); + result->args_.AddObject("biases", AccessType::READ, + absl::make_unique(std::move(lt)), + absl::make_unique(desc)); return absl::OkStatus(); } diff --git a/tensorflow/lite/delegates/gpu/cl/kernels/conv_constants.h b/tensorflow/lite/delegates/gpu/cl/kernels/conv_constants.h index 8d80d48314d..b9cc52f7e94 100644 --- a/tensorflow/lite/delegates/gpu/cl/kernels/conv_constants.h +++ b/tensorflow/lite/delegates/gpu/cl/kernels/conv_constants.h @@ -71,9 +71,6 @@ class ConvConstants : public GPUOperation { absl::Status BindArguments(); int3 GetGridSize() const; - Buffer weights_; - LinearStorage biases_; - int2 kernel_size_; int2 stride_; int2 padding_; @@ -92,21 +89,34 @@ absl::Status ConvConstants::UploadWeights( const int kernel_x = weights.shape.w; const int kernel_y = weights.shape.h; - const int float_size = - definition_.precision == CalculationsPrecision::F32 ? 4 : 2; + const bool f32_weights = definition_.precision == CalculationsPrecision::F32; + + BufferDescriptor desc; + desc.element_type = f32_weights ? DataType::FLOAT32 : DataType::FLOAT16; + desc.element_size = 4; + desc.memory_type = MemoryType::CONSTANT; + + const int float_size = f32_weights ? 4 : 2; const int float_count = src_channels_ * dst_depth * 4 * kernel_x * kernel_y; - if (definition_.GetDataType() == DataType::FLOAT32) { + Buffer weights_buffer; + if (f32_weights) { std::vector gpu_data(float_count / 4); RearrangeWeightsData(weights, absl::MakeSpan(gpu_data)); - return CreateReadOnlyBuffer(float_size * float_count, gpu_data.data(), - context, &weights_); + RETURN_IF_ERROR(CreateReadOnlyBuffer( + float_size * float_count, gpu_data.data(), context, &weights_buffer)); } else { std::vector gpu_data(float_count / 4); RearrangeWeightsData(weights, absl::MakeSpan(gpu_data)); - return CreateReadOnlyBuffer(float_size * float_count, gpu_data.data(), - context, &weights_); + RETURN_IF_ERROR(CreateReadOnlyBuffer( + float_size * float_count, gpu_data.data(), context, &weights_buffer)); } + + args_.AddObject("weigths", AccessType::READ, + absl::make_unique(std::move(weights_buffer)), + absl::make_unique(desc)); + + return absl::OkStatus(); } template diff --git a/tensorflow/lite/delegates/gpu/cl/linear_storage.cc b/tensorflow/lite/delegates/gpu/cl/linear_storage.cc index 47504a34c2b..ee0ea3efbec 100644 --- a/tensorflow/lite/delegates/gpu/cl/linear_storage.cc +++ b/tensorflow/lite/delegates/gpu/cl/linear_storage.cc @@ -32,6 +32,7 @@ GPUResources TensorLinearDescriptor::GetGPUResources( desc.data_type = element_type; desc.access_type = access_type; desc.element_size = 4; + desc.memory_type = memory_type; resources.buffers.push_back({"buffer", desc}); } else { GPUImage2DDescriptor desc; diff --git a/tensorflow/lite/delegates/gpu/cl/linear_storage.h b/tensorflow/lite/delegates/gpu/cl/linear_storage.h index 2c0770ef3dc..14c8460bf80 100644 --- a/tensorflow/lite/delegates/gpu/cl/linear_storage.h +++ b/tensorflow/lite/delegates/gpu/cl/linear_storage.h @@ -40,6 +40,7 @@ enum class LinearStorageType { BUFFER, TEXTURE_2D }; struct TensorLinearDescriptor : public GPUObjectDescriptor { LinearStorageType storage_type; DataType element_type; // FLOAT32 or FLOAT16 + MemoryType memory_type = MemoryType::GLOBAL; // applicable for BUFFER absl::Status PerformSelector(const std::string& selector, const std::vector& args, From 6262e61826764c5236c1e74596583973c7248f3c Mon Sep 17 00:00:00 2001 From: Akshay Modi Date: Tue, 23 Jun 2020 09:51:51 -0700 Subject: [PATCH 553/867] Add more np array interop tests. PiperOrigin-RevId: 317885566 Change-Id: I26dc3643547f0bd834b8e913e3c6eda7847f1aea --- .../python/ops/numpy_ops/np_interop_test.py | 120 ++++++++++++++++-- 1 file changed, 106 insertions(+), 14 deletions(-) diff --git a/tensorflow/python/ops/numpy_ops/np_interop_test.py b/tensorflow/python/ops/numpy_ops/np_interop_test.py index 17a3bf81a01..33abb58f260 100644 --- a/tensorflow/python/ops/numpy_ops/np_interop_test.py +++ b/tensorflow/python/ops/numpy_ops/np_interop_test.py @@ -93,9 +93,8 @@ class InteropTest(tf.test.TestCase): dx, dy = t.gradient([xx, yy], [x, y]) - # # TODO(nareshmodi): Figure out a way to rewrap ndarray as tensors. - # self.assertIsInstance(dx, np.ndarray) - # self.assertIsInstance(dy, np.ndarray) + self.assertIsInstance(dx, np.ndarray) + self.assertIsInstance(dy, np.ndarray) self.assertAllClose(dx, 2.0) self.assertAllClose(dy, 3.0) @@ -181,19 +180,17 @@ class InteropTest(tf.test.TestCase): multiplier = np.asarray(5.) - with strategy.scope(): + @tf.function + def run(): + ctx = tf.distribute.get_replica_context() + val = np.asarray(ctx.replica_id_in_sync_group) + return val * multiplier - @tf.function - def run(): - ctx = tf.distribute.get_replica_context() - val = np.asarray(ctx.replica_id_in_sync_group) - return val * multiplier + distributed_values = strategy.run(run) + reduced = strategy.reduce( + tf.distribute.ReduceOp.SUM, distributed_values, axis=None) - distributed_values = strategy.run(run) - reduced = strategy.reduce( - tf.distribute.ReduceOp.SUM, distributed_values, axis=None) - - values = distributed_values.values + values = strategy.experimental_local_results(distributed_values) # Note that this should match the number of virtual CPUs. self.assertLen(values, 3) @@ -208,6 +205,101 @@ class InteropTest(tf.test.TestCase): # self.assertIsInstance(reduced, np.ndarray) self.assertAllClose(reduced, 15) + def testPyFuncInterop(self): + def py_func_fn(a, b): + return a + b + + @tf.function + def fn(a, b): + result = tf.py_function(py_func_fn, [a, b], a.dtype) + return np.asarray(result) + + a = np.asarray(1.) + b = np.asarray(2.) + + result = fn(a, b) + self.assertIsInstance(result, np.ndarray) + self.assertAllClose(result, 3.) + + def testDatasetInterop(self): + values = [1, 2, 3, 4, 5, 6] + values_as_array = np.asarray(values) + + # Tensor dataset + dataset = tf.data.Dataset.from_tensors(values_as_array) + + for value, value_from_dataset in zip([values_as_array], dataset): + self.assertIsInstance(value_from_dataset, np.ndarray) + self.assertAllEqual(value_from_dataset, value) + + # Tensor slice dataset + dataset = tf.data.Dataset.from_tensor_slices(values_as_array) + + for value, value_from_dataset in zip(values, dataset): + self.assertIsInstance(value_from_dataset, np.ndarray) + self.assertAllEqual(value_from_dataset, value) + + # # TODO(nareshmodi): as_numpy_iterator() doesn't work. + # items = list(dataset.as_numpy_iterator()) + + # Map over a dataset. + dataset = dataset.map(lambda x: np.add(x, 1)) + + for value, value_from_dataset in zip(values, dataset): + self.assertIsInstance(value_from_dataset, np.ndarray) + self.assertAllEqual(value_from_dataset, value + 1) + + # Batch a dataset. + dataset = tf.data.Dataset.from_tensor_slices(values_as_array).batch(2) + + for value, value_from_dataset in zip([[1, 2], [3, 4], [5, 6]], dataset): + self.assertIsInstance(value_from_dataset, np.ndarray) + self.assertAllEqual(value_from_dataset, value) + + def testKerasInterop(self): + # Return an ndarray from the model. + inputs = tf.keras.layers.Input(shape=(10,)) + output_layer = tf.keras.layers.Lambda(np.square)(inputs) + model = tf.keras.Model([inputs], output_layer) + + values = onp.arange(10, dtype=onp.float32) + values_as_array = np.asarray(values) + + result = model(values) + self.assertIsInstance(result, np.ndarray) + self.assertAllClose(result, onp.square(values)) + + result = model(values_as_array) + self.assertIsInstance(result, np.ndarray) + self.assertAllClose(result, onp.square(values)) + + def testPForInterop(self): + def outer_product(a): + return np.tensordot(a, a, 0) + + batch_size = 100 + a = np.ones((batch_size, 32, 32)) + c = tf.vectorized_map(outer_product, a) + + # # TODO(nareshmodi): vectorized_map doesn't rewrap tensors in ndarray. + # self.assertIsInstance(c, np.ndarray) + self.assertEqual(c.shape, (batch_size, 32, 32, 32, 32)) + + def testJacobian(self): + with tf.GradientTape() as g: + x = np.asarray([1., 2.]) + y = np.asarray([3., 4.]) + g.watch(x) + g.watch(y) + z = x * x * y + + jacobian = g.jacobian(z, [x, y]) + answer = [tf.linalg.diag(2 * x * y), tf.linalg.diag(x * x)] + + self.assertIsInstance(jacobian[0], np.ndarray) + self.assertIsInstance(jacobian[1], np.ndarray) + self.assertAllClose(jacobian, answer) + class FunctionTest(InteropTest): From d34bb0c10657b29322b231b08a818738d4eb010f Mon Sep 17 00:00:00 2001 From: Raman Sarokin Date: Tue, 23 Jun 2020 09:52:04 -0700 Subject: [PATCH 554/867] DepthwiseConvolution converted to new style. DepthwiseConvolution3D merged into DepthwiseConvolution. PiperOrigin-RevId: 317885608 Change-Id: Ifead6b0998c75d70f2be18ba2d32936e9e8ecdf0 --- .../lite/delegates/gpu/cl/kernels/BUILD | 23 -- .../gpu/cl/kernels/depthwise_conv.cc | 328 ++++++++++------- .../delegates/gpu/cl/kernels/depthwise_conv.h | 152 +++++++- .../gpu/cl/kernels/depthwise_conv_3d.cc | 338 ------------------ .../gpu/cl/kernels/depthwise_conv_3d.h | 170 --------- 5 files changed, 337 insertions(+), 674 deletions(-) delete mode 100644 tensorflow/lite/delegates/gpu/cl/kernels/depthwise_conv_3d.cc delete mode 100644 tensorflow/lite/delegates/gpu/cl/kernels/depthwise_conv_3d.h diff --git a/tensorflow/lite/delegates/gpu/cl/kernels/BUILD b/tensorflow/lite/delegates/gpu/cl/kernels/BUILD index 24a9a962296..21fb65e8909 100644 --- a/tensorflow/lite/delegates/gpu/cl/kernels/BUILD +++ b/tensorflow/lite/delegates/gpu/cl/kernels/BUILD @@ -594,29 +594,6 @@ cc_library( ], ) -cc_library( - name = "depthwise_conv_3d", - srcs = ["depthwise_conv_3d.cc"], - hdrs = ["depthwise_conv_3d.h"], - deps = [ - ":gpu_operation", - ":util", - ":work_group_picking", - "//tensorflow/lite/delegates/gpu/cl:buffer", - "//tensorflow/lite/delegates/gpu/cl:cl_device", - "//tensorflow/lite/delegates/gpu/cl:linear_storage", - "//tensorflow/lite/delegates/gpu/cl:tensor", - "//tensorflow/lite/delegates/gpu/cl:texture2d", - "//tensorflow/lite/delegates/gpu/cl:util", - "//tensorflow/lite/delegates/gpu/common:data_type", - "//tensorflow/lite/delegates/gpu/common:operations", - "//tensorflow/lite/delegates/gpu/common:shape", - "//tensorflow/lite/delegates/gpu/common:status", - "//tensorflow/lite/delegates/gpu/common:tensor", - "//tensorflow/lite/delegates/gpu/common:types", - ], -) - cc_test( name = "depthwise_conv_test", srcs = ["depthwise_conv_test.cc"], diff --git a/tensorflow/lite/delegates/gpu/cl/kernels/depthwise_conv.cc b/tensorflow/lite/delegates/gpu/cl/kernels/depthwise_conv.cc index 4c5e20abde3..de1a04befa8 100644 --- a/tensorflow/lite/delegates/gpu/cl/kernels/depthwise_conv.cc +++ b/tensorflow/lite/delegates/gpu/cl/kernels/depthwise_conv.cc @@ -34,57 +34,71 @@ bool IsSpecializedCase(int channel_multiplier) { channel_multiplier == 4; } -std::string GetSrcValue(const TensorCodeGenerator& src_tensor, - int channel_multiplier, - TextureAddressMode address_mode) { +std::string GetSrcValue(int channel_multiplier, const std::string coords) { std::string c; if (channel_multiplier == 1) { - c += " FLT4 src_final =" + - src_tensor.ReadWHS("x_c", "y_c", "Z", address_mode) + ";\n"; + c += " FLT4 src_final = args.src_tensor.Read(" + coords + ", S);\n"; } else if (channel_multiplier == 2) { - c += " int z_layer = Z / 2;\n"; - c += " FLT4 src =" + - src_tensor.ReadWHS("x_c", "y_c", "z_layer", address_mode) + ";\n"; - c += " FLT2 t0 = Z % 2 == 0 ? src.xy : src.zw;\n"; + c += " int s_layer = S / 2;\n"; + c += " FLT4 src = args.src_tensor.Read(" + coords + ", s_layer);\n"; + c += " FLT2 t0 = S % 2 == 0 ? src.xy : src.zw;\n"; c += " FLT4 src_final = (FLT4)(t0.x, t0.x, t0.y, t0.y);\n"; } else if (channel_multiplier == 4) { - c += " int z_layer = Z / 4;\n"; - c += " FLT4 src =" + - src_tensor.ReadWHS("x_c", "y_c", "z_layer", address_mode) + ";\n"; + c += " int s_layer = S / 4;\n"; + c += " FLT4 src = args.src_tensor.Read(" + coords + ", s_layer);\n"; c += " FLT t0 = src.x;\n"; - c += " int reminder = Z % 4;\n"; + c += " int reminder = S % 4;\n"; c += " if (reminder == 1) t0 = src.y;\n"; c += " if (reminder == 2) t0 = src.z;\n"; c += " if (reminder == 3) t0 = src.w;\n"; c += " FLT4 src_final = (FLT4)(t0, t0, t0, t0);\n"; } else { - c += " int z_layer = Z / channel_multiplier;\n"; - c += " FLT4 src =" + - src_tensor.ReadWHS("x_c", "y_c", "z_layer", address_mode) + ";\n"; - c += " int z_offset = (Z % channel_multiplier) * 4;\n"; + c += " int s_layer = S / args.ch_multiplier;\n"; + c += " FLT4 src = args.src_tensor.Read(" + coords + ", s_layer);\n"; + c += " int s_offset = (S % args.ch_multiplier) * 4;\n"; c += " FLT4 src_final;\n"; c += " FLT temp_arr[4] = {src.x, src.y, src.z, src.w};\n"; - c += " src_final.x = temp_arr[(z_offset + 0) / channel_multiplier];\n"; - c += " src_final.y = temp_arr[(z_offset + 1) / channel_multiplier];\n"; - c += " src_final.z = temp_arr[(z_offset + 2) / channel_multiplier];\n"; - c += " src_final.w = temp_arr[(z_offset + 3) / channel_multiplier];\n"; + c += " src_final.x = temp_arr[(s_offset + 0) / args.ch_multiplier];\n"; + c += " src_final.y = temp_arr[(s_offset + 1) / args.ch_multiplier];\n"; + c += " src_final.z = temp_arr[(s_offset + 2) / args.ch_multiplier];\n"; + c += " src_final.w = temp_arr[(s_offset + 3) / args.ch_multiplier];\n"; } return c; } std::string GenerateDepthwiseConvolutionCode( - const OperationDef& op_def, bool stride_correction, - const LinearStorage& biases, int channel_multiplier, - bool weights_are_buffer, - const std::vector& linked_operations, - const CLDevice& device) { - TensorCodeGenerator src_tensor( - "src_data", WHSPoint{"src_size.x", "src_size.y", "src_size.z"}, - op_def.src_tensors[0]); - TensorCodeGenerator dst_tensor( - "dst_data", WHSPoint{"dst_size.x", "dst_size.y", "dst_size.z"}, - op_def.dst_tensors[0]); + const OperationDef& op_def, bool stride_correction, int channel_multiplier, + bool weights_are_buffer, const CLDevice& device, Arguments* args) { + auto src_desc = absl::make_unique(op_def.src_tensors[0]); + src_desc->SetTextureAddressMode(GetFastestZeroMode(device)); + if (op_def.IsBatchSupported()) { + src_desc->SetStateVar("BatchedWidth", "true"); + } + args->AddObjectRef("src_tensor", AccessType::READ, std::move(src_desc)); + auto dst_desc = absl::make_unique(op_def.dst_tensors[0]); + if (op_def.IsBatchSupported()) { + dst_desc->SetStateVar("BatchedWidth", "true"); + } + args->AddObjectRef("dst_tensor", AccessType::WRITE, std::move(dst_desc)); + args->AddInt("kernel_size_x"); + args->AddInt("stride_x"); + args->AddInt("padding_x"); + args->AddInt("dilation_x"); + args->AddInt("kernel_size_y"); + args->AddInt("stride_y"); + args->AddInt("padding_y"); + args->AddInt("dilation_y"); + if (op_def.dst_tensors[0].HasAxis(Axis::DEPTH)) { + args->AddInt("kernel_size_z"); + args->AddInt("stride_z"); + args->AddInt("padding_z"); + args->AddInt("dilation_z"); + } + if (!IsSpecializedCase(channel_multiplier)) { + args->AddInt("ch_multiplier"); + } + const auto src_tensor_type = op_def.src_tensors[0].storage_type; std::string c = GetCommonDefines(op_def.precision); @@ -93,86 +107,110 @@ std::string GenerateDepthwiseConvolutionCode( src_tensor_type == TensorStorageType::IMAGE_BUFFER; c += "__kernel void main_function(\n"; - c += src_tensor.GetDeclaration(AccessType::READ) + ",\n"; - if (weights_are_buffer) { - c += " __global FLT4* filters, \n"; - } else { - c += " __read_only image2d_t filters, \n"; - } - c += biases.GetDeclaration(); - c += GetArgsDeclaration(linked_operations); - c += dst_tensor.GetDeclaration(AccessType::WRITE) + ",\n"; - c += " int2 kernel_size, \n"; - c += " int2 stride, \n"; - c += " int2 padding, \n"; - c += " int2 dilation, \n"; - if (!IsSpecializedCase(channel_multiplier)) { - c += " int channel_multiplier, \n"; - } - c += " int4 src_size, \n"; - c += " int4 dst_size \n"; - c += ") {\n"; + c += "$0) {\n"; c += " int X = get_global_id(0);\n"; c += " int Y = get_global_id(1);\n"; - c += " int Z = get_global_id(2);\n"; - c += " if (X >= dst_size.x || Y >= dst_size.y || Z >= dst_size.z) return;\n"; + if (op_def.dst_tensors[0].HasAxis(Axis::DEPTH)) { + c += " int linear_id_2 = get_global_id(2);\n"; + c += " int S = linear_id_2 / args.dst_tensor.Depth();\n"; + c += " int Z = linear_id_2 % args.dst_tensor.Depth();\n"; + } else { + c += " int S = get_global_id(2);\n"; + } + c += " if (X >= args.dst_tensor.Width() || Y >= args.dst_tensor.Height() || " + "S >= args.dst_tensor.Slices()) { \n"; + c += " return; \n"; + c += " } \n"; c += " ACCUM_FLT4 r = (ACCUM_FLT4)(0.0f, 0.0f, 0.0f, 0.0f);\n"; if (stride_correction) { c += " int x_offseted = " + - GetXStrideCorrected("X", "src_size.w", "stride.x", "padding.x") + + GetXStrideCorrected("X", "args.src_tensor.Batch()", "args.stride_x", + "args.padding_x") + ";\n"; } else { - c += " int x_offseted = X * stride.x + padding.x;\n"; + c += " int x_offseted = X * args.stride_x + args.padding_x;\n"; + } + c += " int y_offseted = Y * args.stride_y + args.padding_y;\n"; + std::string weights_offset = "args.kernel_size_x * args.kernel_size_y"; + if (op_def.dst_tensors[0].HasAxis(Axis::DEPTH)) { + c += " int z_offseted = Z * args.stride_z + args.padding_z;\n"; + weights_offset += " * args.kernel_size_z"; } - c += " int y_offseted = Y * stride.y + padding.y;\n"; if (weights_are_buffer) { - c += " int fx_c = Z * kernel_size.x * kernel_size.y;\n"; + c += " int fx_c = S * " + weights_offset + ";\n"; } else { c += " int fx_c = 0;\n"; } + std::string flat_coords = "x_c, y_c"; if (manual_clamp) { - c += " for (int ky = 0; ky < kernel_size.y; ++ky) {\n"; - c += " int y_c = y_offseted + ky * dilation.y;\n"; - c += " bool outside_y = y_c < 0 || y_c >= src_size.y;\n"; - c += " for (int kx = 0; kx < kernel_size.x; ++kx) {\n"; - c += " int x_c = x_offseted + kx * dilation.x;\n"; - c += " bool outside_x = x_c < 0 || x_c >= src_size.x;\n"; - c += " if (!outside_x && !outside_y) {\n"; - if (weights_are_buffer) { - c += " FLT4 f = filters[fx_c];\n"; - } else { - c += " FLT4 f = READ_IMAGE(filters, smp_none, (int2)(fx_c, Z));\n"; + std::string check = "!outside_x && !outside_y"; + if (op_def.dst_tensors[0].HasAxis(Axis::DEPTH)) { + check += " && !outside_z"; + flat_coords += ", z_c"; + c += " for (int kz = 0; kz < args.kernel_size_z; ++kz) {\n"; + c += " int z_c = z_offseted + kz * args.dilation_z;\n"; + c += " bool outside_z = z_c < 0 || z_c >= args.src_tensor.Depth();\n"; } - c += GetSrcValue(src_tensor, channel_multiplier, - TextureAddressMode::DONT_CARE); + c += " for (int ky = 0; ky < args.kernel_size_y; ++ky) {\n"; + c += " int y_c = y_offseted + ky * args.dilation_y;\n"; + c += " bool outside_y = y_c < 0 || y_c >= args.src_tensor.Height();\n"; + c += " for (int kx = 0; kx < args.kernel_size_x; ++kx) {\n"; + c += " int x_c = x_offseted + kx * args.dilation_x;\n"; + c += " bool outside_x = x_c < 0 || x_c >= args.src_tensor.Width();\n"; + c += " if (" + check + ") {\n"; + if (weights_are_buffer) { + c += " FLT4 f = args.weights.Read(fx_c);\n"; + } else { + c += " FLT4 f = args.weights.Read(fx_c, S);\n"; + } + c += GetSrcValue(channel_multiplier, flat_coords); c += " r += TO_ACCUM_TYPE(src_final * f);\n"; c += " };\n"; c += " fx_c++;\n"; c += " }\n"; c += " }\n"; + if (op_def.dst_tensors[0].HasAxis(Axis::DEPTH)) { + c += " }\n"; + } } else { // Texture types with ZERO clamping - c += " for (int ky = 0; ky < kernel_size.y; ++ky) {\n"; - c += " int y_c = y_offseted + ky * dilation.y;\n"; - c += " for (int kx = 0; kx < kernel_size.x; ++kx) {\n"; - c += " int x_c = x_offseted + kx * dilation.x;\n"; - const auto access_mode = GetFastestZeroMode(device); - c += GetSrcValue(src_tensor, channel_multiplier, access_mode); + if (op_def.dst_tensors[0].HasAxis(Axis::DEPTH)) { + flat_coords += ", z_c"; + c += " for (int kz = 0; kz < args.kernel_size_z; ++kz) {\n"; + c += " int z_c = z_offseted + kz * args.dilation_z;\n"; + if (src_tensor_type != + TensorStorageType::TEXTURE_3D) { // Only TEXTURE_3D supports clamping + // in DEPTH dimension + c += " if (z_c < 0 || z_c >= args.src_tensor.Depth()) {\n"; + c += " fx_c += args.kernel_size_y * args.kernel_size_x;\n"; + c += " continue;\n"; + c += " }\n"; + } + } + c += " for (int ky = 0; ky < args.kernel_size_y; ++ky) {\n"; + c += " int y_c = y_offseted + ky * args.dilation_y;\n"; + c += " for (int kx = 0; kx < args.kernel_size_x; ++kx) {\n"; + c += " int x_c = x_offseted + kx * args.dilation_x;\n"; + c += GetSrcValue(channel_multiplier, flat_coords); if (weights_are_buffer) { - c += " FLT4 f = filters[fx_c];\n"; + c += " FLT4 f = args.weights.Read(fx_c);\n"; } else { - c += " FLT4 f = READ_IMAGE(filters, smp_none, (int2)(fx_c, Z));\n"; + c += " FLT4 f = args.weights.Read(fx_c, S);\n"; } c += " fx_c++;\n"; c += " r += TO_ACCUM_TYPE(src_final * f);\n"; c += " }\n"; c += " }\n"; + if (op_def.dst_tensors[0].HasAxis(Axis::DEPTH)) { + c += " }\n"; + } + } + c += " FLT4 res0 = TO_FLT4(r) + args.biases.Read(S);\n"; + if (op_def.dst_tensors[0].HasAxis(Axis::DEPTH)) { + c += " args.dst_tensor.Write(res0, X, Y, Z, S);\n"; + } else { + c += " args.dst_tensor.Write(res0, X, Y, S);\n"; } - c += " FLT4 bias_val = " + biases.ReadLinearFLT4("Z") + ";\n"; - c += " FLT4 res0 = TO_FLT4(r) + bias_val;\n"; - const LinkingContext context{"res0", "X", "Y", "Z"}; - c += PostProcess(linked_operations, context); - c += " " + dst_tensor.WriteWHS("res0", "X", "Y", "Z") + "\n"; c += "}\n"; return c; @@ -184,20 +222,30 @@ DepthwiseConvolution::DepthwiseConvolution( const DepthwiseConvolution2DAttributes& attr, bool weights_are_buffer) : GPUOperation(definition), weights_are_buffer_(weights_are_buffer), - kernel_size_(attr.weights.shape.w, attr.weights.shape.h), - stride_(attr.strides.w, attr.strides.h), - padding_(-attr.padding.prepended.w, -attr.padding.prepended.h), - dilation_(attr.dilations.w, attr.dilations.h), + kernel_size_(attr.weights.shape.w, attr.weights.shape.h, 0, 0), + stride_(attr.strides.w, attr.strides.h, 0, 0), + padding_(-attr.padding.prepended.w, -attr.padding.prepended.h, 0, 0), + dilation_(attr.dilations.w, attr.dilations.h, 0, 0), + channel_multiplier_(attr.weights.shape.o), + work_group_size_(8, 8, 1) {} + +DepthwiseConvolution::DepthwiseConvolution( + const OperationDef& definition, + const DepthwiseConvolution3DAttributes& attr, bool weights_are_buffer) + : GPUOperation(definition), + weights_are_buffer_(weights_are_buffer), + kernel_size_(attr.weights.shape.w, attr.weights.shape.h, + attr.weights.shape.d, 0), + stride_(attr.strides.w, attr.strides.h, attr.strides.d, 0), + padding_(-attr.padding.prepended.w, -attr.padding.prepended.h, + -attr.padding.prepended.d, 0), + dilation_(attr.dilations.w, attr.dilations.h, attr.dilations.d, 0), channel_multiplier_(attr.weights.shape.o), work_group_size_(8, 8, 1) {} DepthwiseConvolution::DepthwiseConvolution(DepthwiseConvolution&& operation) : GPUOperation(std::move(operation)), weights_are_buffer_(operation.weights_are_buffer_), - weights_tex2d_(std::move(operation.weights_tex2d_)), - weights_buf_(std::move(operation.weights_buf_)), - weights_(operation.weights_), - biases_(std::move(operation.biases_)), kernel_size_(operation.kernel_size_), stride_(operation.stride_), padding_(operation.padding_), @@ -210,10 +258,6 @@ DepthwiseConvolution& DepthwiseConvolution::operator=( DepthwiseConvolution&& operation) { if (this != &operation) { std::swap(weights_are_buffer_, operation.weights_are_buffer_); - weights_tex2d_ = std::move(operation.weights_tex2d_); - weights_buf_ = std::move(operation.weights_buf_); - std::swap(weights_, operation.weights_); - biases_ = std::move(operation.biases_); std::swap(kernel_size_, operation.kernel_size_); std::swap(stride_, operation.stride_); std::swap(padding_, operation.padding_); @@ -230,39 +274,48 @@ absl::Status DepthwiseConvolution::Compile( const CreationContext& creation_context) { const bool stride_correction = definition_.IsBatchSupported() && stride_.x != 1; - const auto code = GenerateDepthwiseConvolutionCode( - definition_, stride_correction, biases_, channel_multiplier_, - weights_are_buffer_, linked_operations_, *creation_context.device); + std::string code = GenerateDepthwiseConvolutionCode( + definition_, stride_correction, channel_multiplier_, weights_are_buffer_, + *creation_context.device, &args_); + std::string element_wise_code; + RETURN_IF_ERROR( + MergeOperations(linked_operations_, &args_, &element_wise_code)); + RETURN_IF_ERROR(args_.TransformToCLCode(creation_context.device->GetInfo(), + {{"dst_tensor", element_wise_code}}, + &code)); return creation_context.cache->GetOrCreateCLKernel( code, "main_function", *creation_context.context, *creation_context.device, &kernel_); } absl::Status DepthwiseConvolution::BindArguments() { - kernel_.ResetBindingCounter(); - RETURN_IF_ERROR(kernel_.SetMemoryAuto(src_[0]->GetMemoryPtr())); - RETURN_IF_ERROR(kernel_.SetMemoryAuto(weights_)); - RETURN_IF_ERROR(kernel_.SetMemoryAuto(biases_.GetMemoryPtr())); - RETURN_IF_ERROR(BindArgs(&kernel_, linked_operations_)); - RETURN_IF_ERROR(kernel_.SetMemoryAuto(dst_[0]->GetMemoryPtrForWriting())); - RETURN_IF_ERROR(kernel_.SetBytesAuto(kernel_size_)); - RETURN_IF_ERROR(kernel_.SetBytesAuto(stride_)); - RETURN_IF_ERROR( - kernel_.SetBytesAuto(int2(padding_.x * src_[0]->Batch(), padding_.y))); - RETURN_IF_ERROR( - kernel_.SetBytesAuto(int2(dilation_.x * src_[0]->Batch(), dilation_.y))); - if (!IsSpecializedCase(channel_multiplier_)) { - RETURN_IF_ERROR(kernel_.SetBytesAuto(int32_t(channel_multiplier_))); + RETURN_IF_ERROR(args_.SetObjectRef("src_tensor", src_[0])); + RETURN_IF_ERROR(args_.SetObjectRef("dst_tensor", dst_[0])); + RETURN_IF_ERROR(args_.SetInt("kernel_size_x", kernel_size_.x)); + RETURN_IF_ERROR(args_.SetInt("stride_x", stride_.x)); + RETURN_IF_ERROR(args_.SetInt("padding_x", padding_.x * src_[0]->Batch())); + RETURN_IF_ERROR(args_.SetInt("dilation_x", dilation_.x * src_[0]->Batch())); + RETURN_IF_ERROR(args_.SetInt("kernel_size_y", kernel_size_.y)); + RETURN_IF_ERROR(args_.SetInt("stride_y", stride_.y)); + RETURN_IF_ERROR(args_.SetInt("padding_y", padding_.y)); + RETURN_IF_ERROR(args_.SetInt("dilation_y", dilation_.y)); + if (definition_.dst_tensors[0].HasAxis(Axis::DEPTH)) { + RETURN_IF_ERROR(args_.SetInt("kernel_size_z", kernel_size_.z)); + RETURN_IF_ERROR(args_.SetInt("stride_z", stride_.z)); + RETURN_IF_ERROR(args_.SetInt("padding_z", padding_.z)); + RETURN_IF_ERROR(args_.SetInt("dilation_z", dilation_.z)); } - RETURN_IF_ERROR(kernel_.SetBytesAuto(src_[0]->GetWBatchedHSB())); - RETURN_IF_ERROR(kernel_.SetBytesAuto(dst_[0]->GetWBatchedHSB())); - return absl::OkStatus(); + if (!IsSpecializedCase(channel_multiplier_)) { + RETURN_IF_ERROR(args_.SetInt("ch_multiplier", channel_multiplier_)); + } + RETURN_IF_ERROR(SetArguments(linked_operations_, &args_)); + return args_.Bind(kernel_.kernel()); } int3 DepthwiseConvolution::GetGridSize() const { const int grid_x = dst_[0]->Width() * dst_[0]->Batch(); const int grid_y = dst_[0]->Height(); - const int grid_z = dst_[0]->Slices(); + const int grid_z = dst_[0]->Slices() * dst_[0]->Depth(); return int3(grid_x, grid_y, grid_z); } @@ -284,14 +337,41 @@ absl::Status CreateDepthwiseConvolution( *result = DepthwiseConvolution(definition, attr, weights_are_buffer); RETURN_IF_ERROR( result->UploadWeights(attr.weights, creation_context.context)); - LinearStorageCreateInfo create_info; - create_info.storage_type = weights_are_buffer ? LinearStorageType::BUFFER - : LinearStorageType::TEXTURE_2D; - create_info.data_type = definition.GetDataType(); - create_info.name = "biases"; - create_info.aligned_size = attr.weights.shape.o * attr.weights.shape.i; - RETURN_IF_ERROR(CreateLinearStorage( - create_info, attr.bias, creation_context.context, &result->biases_)); + + TensorLinearDescriptor desc; + desc.storage_type = weights_are_buffer ? LinearStorageType::BUFFER + : LinearStorageType::TEXTURE_2D; + desc.element_type = definition.GetDataType(); + + LinearStorage lt; + RETURN_IF_ERROR( + CreateLinearStorage(desc, attr.bias, creation_context.context, <)); + result->args_.AddObject("biases", AccessType::READ, + absl::make_unique(std::move(lt)), + absl::make_unique(desc)); + return absl::OkStatus(); +} + +absl::Status CreateDepthwiseConvolution( + const CreationContext& creation_context, const OperationDef& definition, + const DepthwiseConvolution3DAttributes& attr, + DepthwiseConvolution* result) { + bool weights_are_buffer = creation_context.device->IsMali(); + *result = DepthwiseConvolution(definition, attr, weights_are_buffer); + RETURN_IF_ERROR( + result->UploadWeights(attr.weights, creation_context.context)); + + TensorLinearDescriptor desc; + desc.storage_type = weights_are_buffer ? LinearStorageType::BUFFER + : LinearStorageType::TEXTURE_2D; + desc.element_type = definition.GetDataType(); + + LinearStorage lt; + RETURN_IF_ERROR( + CreateLinearStorage(desc, attr.bias, creation_context.context, <)); + result->args_.AddObject("biases", AccessType::READ, + absl::make_unique(std::move(lt)), + absl::make_unique(desc)); return absl::OkStatus(); } diff --git a/tensorflow/lite/delegates/gpu/cl/kernels/depthwise_conv.h b/tensorflow/lite/delegates/gpu/cl/kernels/depthwise_conv.h index 7655f2abae0..30cd3d06a5a 100644 --- a/tensorflow/lite/delegates/gpu/cl/kernels/depthwise_conv.h +++ b/tensorflow/lite/delegates/gpu/cl/kernels/depthwise_conv.h @@ -54,9 +54,17 @@ class DepthwiseConvolution : public GPUOperation { const CreationContext& creation_context, const OperationDef& definition, const DepthwiseConvolution2DAttributes& attr, DepthwiseConvolution* result); + friend absl::Status CreateDepthwiseConvolution( + const CreationContext& creation_context, const OperationDef& definition, + const DepthwiseConvolution3DAttributes& attr, + DepthwiseConvolution* result); DepthwiseConvolution(const OperationDef& definition, const DepthwiseConvolution2DAttributes& attr, bool weights_are_buffer); + DepthwiseConvolution(const OperationDef& definition, + const DepthwiseConvolution3DAttributes& attr, + bool weights_are_buffer); + template absl::Status UploadWeights(const tflite::gpu::Tensor& weights, CLContext* context); @@ -65,20 +73,23 @@ class DepthwiseConvolution : public GPUOperation { void RearrangeWeightsData(const tflite::gpu::Tensor& weights, absl::Span dst); + template + absl::Status UploadWeights(const tflite::gpu::Tensor& weights, + CLContext* context); + + template + void RearrangeWeightsData(const tflite::gpu::Tensor& weights, + absl::Span dst); + absl::Status BindArguments(); int3 GetGridSize() const; bool weights_are_buffer_; - Texture2D weights_tex2d_; - Buffer weights_buf_; - cl_mem weights_; - LinearStorage biases_; - - int2 kernel_size_; - int2 stride_; - int2 padding_; - int2 dilation_; + int4 kernel_size_; + int4 stride_; + int4 padding_; + int4 dilation_; int channel_multiplier_; CLKernel kernel_; @@ -89,26 +100,28 @@ template absl::Status DepthwiseConvolution::UploadWeights( const tflite::gpu::Tensor& weights, CLContext* context) { const int dst_channels = weights.shape.i * weights.shape.o; - const int dst_depth = DivideRoundUp(dst_channels, 4); + const int dst_slices = DivideRoundUp(dst_channels, 4); const int kernel_x = weights.shape.w; const int kernel_y = weights.shape.h; - const int elements_count = kernel_x * kernel_y * dst_depth; + const int elements_count = kernel_x * kernel_y * dst_slices; const bool fp32_weights = definition_.precision == CalculationsPrecision::F32; const int float4_size = fp32_weights ? 16 : 8; + Texture2D weights_tex2d; + Buffer weights_buf; if (fp32_weights) { std::vector gpu_data(elements_count); RearrangeWeightsData(weights, absl::MakeSpan(gpu_data)); if (weights_are_buffer_) { RETURN_IF_ERROR(CreateReadOnlyBuffer(float4_size * elements_count, gpu_data.data(), context, - &weights_buf_)); + &weights_buf)); } else { RETURN_IF_ERROR(CreateTexture2DRGBA( - definition_.GetDataType(), kernel_x * kernel_y, dst_depth, - gpu_data.data(), context, &weights_tex2d_)); + definition_.GetDataType(), kernel_x * kernel_y, dst_slices, + gpu_data.data(), context, &weights_tex2d)); } } else { std::vector gpu_data(elements_count); @@ -116,18 +129,27 @@ absl::Status DepthwiseConvolution::UploadWeights( if (weights_are_buffer_) { RETURN_IF_ERROR(CreateReadOnlyBuffer(float4_size * elements_count, gpu_data.data(), context, - &weights_buf_)); + &weights_buf)); } else { RETURN_IF_ERROR(CreateTexture2DRGBA( - definition_.GetDataType(), kernel_x * kernel_y, dst_depth, - gpu_data.data(), context, &weights_tex2d_)); + definition_.GetDataType(), kernel_x * kernel_y, dst_slices, + gpu_data.data(), context, &weights_tex2d)); } } if (weights_are_buffer_) { - weights_ = weights_buf_.GetMemoryPtr(); + BufferDescriptor desc; + desc.element_type = fp32_weights ? DataType::FLOAT32 : DataType::FLOAT16; + desc.element_size = 4; + args_.AddObject("weights", AccessType::READ, + absl::make_unique(std::move(weights_buf)), + absl::make_unique(desc)); } else { - weights_ = weights_tex2d_.GetMemoryPtr(); + Texture2DDescriptor desc; + desc.element_type = fp32_weights ? DataType::FLOAT32 : DataType::FLOAT16; + args_.AddObject("weights", AccessType::READ, + absl::make_unique(std::move(weights_tex2d)), + absl::make_unique(desc)); } return absl::OkStatus(); @@ -162,6 +184,98 @@ void DepthwiseConvolution::RearrangeWeightsData( } } +template +absl::Status DepthwiseConvolution::UploadWeights( + const tflite::gpu::Tensor& weights, CLContext* context) { + const int dst_channels = weights.shape.i * weights.shape.o; + const int dst_slices = DivideRoundUp(dst_channels, 4); + const int kernel_x = weights.shape.w; + const int kernel_y = weights.shape.h; + const int kernel_z = weights.shape.d; + + const int elements_count = kernel_x * kernel_y * kernel_z * dst_slices; + + const bool fp32_weights = definition_.precision == CalculationsPrecision::F32; + const int float4_size = fp32_weights ? 16 : 8; + + Texture2D weights_tex2d; + Buffer weights_buf; + if (fp32_weights) { + std::vector gpu_data(elements_count); + RearrangeWeightsData(weights, absl::MakeSpan(gpu_data)); + if (weights_are_buffer_) { + RETURN_IF_ERROR(CreateReadOnlyBuffer(float4_size * elements_count, + gpu_data.data(), context, + &weights_buf)); + } else { + RETURN_IF_ERROR(CreateTexture2DRGBA( + definition_.GetDataType(), kernel_x * kernel_y * kernel_z, dst_slices, + gpu_data.data(), context, &weights_tex2d)); + } + } else { + std::vector gpu_data(elements_count); + RearrangeWeightsData(weights, absl::MakeSpan(gpu_data)); + if (weights_are_buffer_) { + RETURN_IF_ERROR(CreateReadOnlyBuffer(float4_size * elements_count, + gpu_data.data(), context, + &weights_buf)); + } else { + RETURN_IF_ERROR(CreateTexture2DRGBA( + definition_.GetDataType(), kernel_x * kernel_y * kernel_z, dst_slices, + gpu_data.data(), context, &weights_tex2d)); + } + } + + if (weights_are_buffer_) { + BufferDescriptor desc; + desc.element_type = fp32_weights ? DataType::FLOAT32 : DataType::FLOAT16; + desc.element_size = 4; + args_.AddObject("weights", AccessType::READ, + absl::make_unique(std::move(weights_buf)), + absl::make_unique(desc)); + } else { + Texture2DDescriptor desc; + desc.element_type = fp32_weights ? DataType::FLOAT32 : DataType::FLOAT16; + args_.AddObject("weights", AccessType::READ, + absl::make_unique(std::move(weights_tex2d)), + absl::make_unique(desc)); + } + + return absl::OkStatus(); +} + +template +void DepthwiseConvolution::RearrangeWeightsData( + const tflite::gpu::Tensor& weights, absl::Span dst) { + const int dst_channels = weights.shape.i * weights.shape.o; + const int dst_slices = DivideRoundUp(dst_channels, 4); + const int kernel_x = weights.shape.w; + const int kernel_y = weights.shape.h; + const int kernel_z = weights.shape.d; + + int counter = 0; + for (int d = 0; d < dst_slices; ++d) { + for (int z = 0; z < kernel_z; ++z) { + for (int y = 0; y < kernel_y; ++y) { + for (int x = 0; x < kernel_x; ++x) { + T filter_val; + for (int i = 0; i < 4; ++i) { + const int d_ch = d * 4 + i; + if (d_ch < dst_channels) { + const int f_index = weights.shape.LinearIndex( + {d_ch % weights.shape.o, y, x, z, d_ch / weights.shape.o}); + filter_val[i] = weights.data[f_index]; + } else { + filter_val[i] = 0.0f; + } + } + dst[counter++] = filter_val; + } + } + } + } +} + absl::Status CreateDepthwiseConvolution( const CreationContext& creation_context, const OperationDef& definition, const DepthwiseConvolution2DAttributes& attr, DepthwiseConvolution* result); diff --git a/tensorflow/lite/delegates/gpu/cl/kernels/depthwise_conv_3d.cc b/tensorflow/lite/delegates/gpu/cl/kernels/depthwise_conv_3d.cc deleted file mode 100644 index f9926a9f466..00000000000 --- a/tensorflow/lite/delegates/gpu/cl/kernels/depthwise_conv_3d.cc +++ /dev/null @@ -1,338 +0,0 @@ -/* Copyright 2019 The TensorFlow Authors. All Rights Reserved. - -Licensed under the Apache License, Version 2.0 (the "License"); -you may not use this file except in compliance with the License. -You may obtain a copy of the License at - - http://www.apache.org/licenses/LICENSE-2.0 - -Unless required by applicable law or agreed to in writing, software -distributed under the License is distributed on an "AS IS" BASIS, -WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -See the License for the specific language governing permissions and -limitations under the License. -==============================================================================*/ - -#include "tensorflow/lite/delegates/gpu/cl/kernels/depthwise_conv_3d.h" - -#include -#include -#include - -#include "tensorflow/lite/delegates/gpu/cl/cl_device.h" -#include "tensorflow/lite/delegates/gpu/cl/kernels/util.h" -#include "tensorflow/lite/delegates/gpu/cl/kernels/work_group_picking.h" - -namespace tflite { -namespace gpu { -namespace cl { -namespace { - -bool IsSpecializedCase(int channel_multiplier) { - return channel_multiplier == 1 || channel_multiplier == 2 || - channel_multiplier == 4; -} - -std::string GetSrcValue(const TensorCodeGenerator& src_tensor, - int channel_multiplier, - TextureAddressMode address_mode) { - std::string c; - if (channel_multiplier == 1) { - c += " FLT4 src_final =" + - src_tensor.ReadWHDS("x_c", "y_c", "z_c", "S", address_mode) + ";\n"; - } else if (channel_multiplier == 2) { - c += " int z_layer = S / 2;\n"; - c += " FLT4 src =" + - src_tensor.ReadWHDS("x_c", "y_c", "z_c", "z_layer", address_mode) + - ";\n"; - c += " FLT2 t0 = S % 2 == 0 ? src.xy : src.zw;\n"; - c += " FLT4 src_final = (FLT4)(t0.x, t0.x, t0.y, t0.y);\n"; - } else if (channel_multiplier == 4) { - c += " int z_layer = S / 4;\n"; - c += " FLT4 src =" + - src_tensor.ReadWHDS("x_c", "y_c", "z_c", "z_layer", address_mode) + - ";\n"; - c += " FLT t0 = src.x;\n"; - c += " int reminder = S % 4;\n"; - c += " if (reminder == 1) t0 = src.y;\n"; - c += " if (reminder == 2) t0 = src.z;\n"; - c += " if (reminder == 3) t0 = src.w;\n"; - c += " FLT4 src_final = (FLT4)(t0, t0, t0, t0);\n"; - } else { - c += " int z_layer = S / channel_multiplier;\n"; - c += " FLT4 src =" + - src_tensor.ReadWHDS("x_c", "y_c", "z_c", "z_layer", address_mode) + - ";\n"; - c += " int z_offset = (S % channel_multiplier) * 4;\n"; - c += " FLT4 src_final;\n"; - c += " FLT temp_arr[4] = {src.x, src.y, src.z, src.w};\n"; - c += " src_final.x = temp_arr[(z_offset + 0) / " - "channel_multiplier];\n"; - c += " src_final.y = temp_arr[(z_offset + 1) / " - "channel_multiplier];\n"; - c += " src_final.z = temp_arr[(z_offset + 2) / " - "channel_multiplier];\n"; - c += " src_final.w = temp_arr[(z_offset + 3) / " - "channel_multiplier];\n"; - } - - return c; -} - -std::string GenerateDepthwiseConvolution3DCode( - const OperationDef& op_def, bool stride_correction, - const LinearStorage& biases, int channel_multiplier, - bool weights_are_buffer, - const std::vector& linked_operations, - const CLDevice& device) { - TensorCodeGenerator src_tensor( - "src_data", - WHDSPoint{"src_size.x", "src_size.y", "src_size.z", "src_size.w"}, - op_def.src_tensors[0]); - TensorCodeGenerator dst_tensor( - "dst_data", - WHDSPoint{"dst_size.x", "dst_size.y", "dst_size.z", "dst_size.w"}, - op_def.dst_tensors[0]); - const auto src_tensor_type = op_def.src_tensors[0].storage_type; - - std::string c = GetCommonDefines(op_def.precision); - - const bool manual_clamp = src_tensor_type == TensorStorageType::BUFFER || - src_tensor_type == TensorStorageType::IMAGE_BUFFER; - - c += "__kernel void main_function(\n"; - c += src_tensor.GetDeclaration(AccessType::READ) + ",\n"; - if (weights_are_buffer) { - c += " __global FLT4* filters, \n"; - } else { - c += " __read_only image2d_t filters, \n"; - } - c += biases.GetDeclaration(); - c += GetArgsDeclaration(linked_operations); - c += dst_tensor.GetDeclaration(AccessType::WRITE) + ",\n"; - c += " int4 kernel_size, \n"; - c += " int4 stride, \n"; - c += " int4 padding, \n"; - c += " int4 dilation, \n"; - if (!IsSpecializedCase(channel_multiplier)) { - c += " int channel_multiplier, \n"; - } - if (op_def.IsBatchSupported()) { - c += " int batch_size, \n"; - } - c += " int4 src_size, \n"; - c += " int4 dst_size \n"; - c += ") {\n"; - c += " int X = get_global_id(0);\n"; - c += " int Y = get_global_id(1);\n"; - c += " int linear_id_z = get_global_id(2);\n"; - c += " int S = linear_id_z % dst_size.w;\n"; - c += " int Z = linear_id_z / dst_size.w;\n"; - c += " if (X >= dst_size.x || Y >= dst_size.y || Z >= dst_size.z) return;\n"; - c += " ACCUM_FLT4 r = (ACCUM_FLT4)(0.0f, 0.0f, 0.0f, 0.0f);\n"; - if (stride_correction) { - c += " int x_offseted = " + - GetXStrideCorrected("X", "batch_size", "stride.x", "padding.x") + - ";\n"; - } else { - c += " int x_offseted = X * stride.x + padding.x;\n"; - } - c += " int y_offseted = Y * stride.y + padding.y;\n"; - c += " int z_offseted = Z * stride.z + padding.z;\n"; - if (weights_are_buffer) { - c += " int fx_c = S * kernel_size.x * kernel_size.y * kernel_size.z;\n"; - } else { - c += " int fx_c = 0;\n"; - } - - if (manual_clamp) { - c += " for (int kz = 0; kz < kernel_size.z; ++kz) {\n"; - c += " int z_c = z_offseted + kz * dilation.z;\n"; - c += " bool outside_z = z_c < 0 || z_c >= src_size.z;\n"; - c += " for (int ky = 0; ky < kernel_size.y; ++ky) {\n"; - c += " int y_c = y_offseted + ky * dilation.y;\n"; - c += " bool outside_y = y_c < 0 || y_c >= src_size.y;\n"; - c += " for (int kx = 0; kx < kernel_size.x; ++kx) {\n"; - c += " int x_c = x_offseted + kx * dilation.x;\n"; - c += " bool outside_x = x_c < 0 || x_c >= src_size.x;\n"; - c += " if (!outside_x && !outside_y && !outside_z) {\n"; - if (weights_are_buffer) { - c += " FLT4 f = filters[fx_c];\n"; - } else { - c += " FLT4 f = READ_IMAGE(filters, smp_none, (int2)(fx_c, " - "S));\n"; - } - c += GetSrcValue(src_tensor, channel_multiplier, - TextureAddressMode::DONT_CARE); - c += " r += TO_ACCUM_TYPE(src_final * f);\n"; - c += " };\n"; - c += " fx_c++;\n"; - c += " }\n"; - c += " }\n"; - c += " }\n"; - } else { // Texture types with ZERO clamping - c += " for (int kz = 0; kz < kernel_size.z; ++kz) {\n"; - c += " int z_c = z_offseted + kz * dilation.z;\n"; - if (src_tensor_type != - TensorStorageType::TEXTURE_3D) { // Only TEXTURE_3D supports clamping - // in DEPTH dimension - c += " if (z_c < 0 || z_c >= src_size.z) {\n"; - c += " fx_c += kernel_size.y * kernel_size.x;\n"; - c += " continue;\n"; - c += " }\n"; - } - c += " for (int ky = 0; ky < kernel_size.y; ++ky) {\n"; - c += " int y_c = y_offseted + ky * dilation.y;\n"; - c += " for (int kx = 0; kx < kernel_size.x; ++kx) {\n"; - c += " int x_c = x_offseted + kx * dilation.x;\n"; - const auto access_mode = GetFastestZeroMode(device); - c += GetSrcValue(src_tensor, channel_multiplier, access_mode); - if (weights_are_buffer) { - c += " FLT4 f = filters[fx_c];\n"; - } else { - c += " FLT4 f = READ_IMAGE(filters, smp_none, (int2)(fx_c, S));\n"; - } - c += " fx_c++;\n"; - c += " r += TO_ACCUM_TYPE(src_final * f);\n"; - c += " }\n"; - c += " }\n"; - c += " }\n"; - } - c += " FLT4 bias_val = " + biases.ReadLinearFLT4("S") + ";\n"; - c += " FLT4 res0 = TO_FLT4(r) + bias_val;\n"; - const LinkingContext context{"res0", "X", "Y", "S"}; - c += PostProcess(linked_operations, context); - c += " " + dst_tensor.WriteWHDS("res0", "X", "Y", "Z", "S") + "\n"; - c += "}\n"; - return c; -} -} // namespace - -DepthwiseConvolution3D::DepthwiseConvolution3D( - const OperationDef& definition, - const DepthwiseConvolution3DAttributes& attr, const CLDevice& device) - : GPUOperation(definition), - weights_are_buffer_(device.IsMali()), - kernel_size_(attr.weights.shape.w, attr.weights.shape.h, - attr.weights.shape.d), - stride_(attr.strides.w, attr.strides.h, attr.strides.d), - padding_(-attr.padding.prepended.w, -attr.padding.prepended.h, - -attr.padding.prepended.d), - dilation_(attr.dilations.w, attr.dilations.h, attr.dilations.d), - channel_multiplier_(attr.weights.shape.o), - work_group_size_(8, 8, 1) {} - -DepthwiseConvolution3D::DepthwiseConvolution3D( - DepthwiseConvolution3D&& operation) - : GPUOperation(std::move(operation)), - weights_tex2d_(std::move(operation.weights_tex2d_)), - weights_buf_(std::move(operation.weights_buf_)), - weights_are_buffer_(operation.weights_are_buffer_), - biases_(std::move(operation.biases_)), - kernel_size_(operation.kernel_size_), - stride_(operation.stride_), - padding_(operation.padding_), - dilation_(operation.dilation_), - channel_multiplier_(operation.channel_multiplier_), - kernel_(std::move(operation.kernel_)), - work_group_size_(operation.work_group_size_) {} - -DepthwiseConvolution3D& DepthwiseConvolution3D::operator=( - DepthwiseConvolution3D&& operation) { - if (this != &operation) { - weights_tex2d_ = std::move(operation.weights_tex2d_); - weights_buf_ = std::move(operation.weights_buf_); - std::swap(weights_are_buffer_, operation.weights_are_buffer_); - biases_ = std::move(operation.biases_); - std::swap(kernel_size_, operation.kernel_size_); - std::swap(stride_, operation.stride_); - std::swap(padding_, operation.padding_); - std::swap(dilation_, operation.dilation_); - std::swap(channel_multiplier_, operation.channel_multiplier_); - kernel_ = std::move(operation.kernel_); - std::swap(work_group_size_, operation.work_group_size_); - GPUOperation::operator=(std::move(operation)); - } - return *this; -} - -absl::Status DepthwiseConvolution3D::Compile( - const CreationContext& creation_context) { - const bool stride_correction = - definition_.IsBatchSupported() && stride_.x != 1; - const auto code = GenerateDepthwiseConvolution3DCode( - definition_, stride_correction, biases_, channel_multiplier_, - weights_are_buffer_, linked_operations_, *creation_context.device); - return creation_context.cache->GetOrCreateCLKernel( - code, "main_function", *creation_context.context, - *creation_context.device, &kernel_); -} - -absl::Status DepthwiseConvolution3D::BindArguments() { - kernel_.ResetBindingCounter(); - RETURN_IF_ERROR(kernel_.SetMemoryAuto(src_[0]->GetMemoryPtr())); - if (weights_are_buffer_) { - RETURN_IF_ERROR(kernel_.SetMemoryAuto(weights_buf_.GetMemoryPtr())); - } else { - RETURN_IF_ERROR(kernel_.SetMemoryAuto(weights_tex2d_.GetMemoryPtr())); - } - RETURN_IF_ERROR(kernel_.SetMemoryAuto(biases_.GetMemoryPtr())); - RETURN_IF_ERROR(BindArgs(&kernel_, linked_operations_)); - RETURN_IF_ERROR(kernel_.SetMemoryAuto(dst_[0]->GetMemoryPtrForWriting())); - RETURN_IF_ERROR(kernel_.SetBytesAuto( - int4(kernel_size_.x, kernel_size_.y, kernel_size_.z, 1))); - RETURN_IF_ERROR( - kernel_.SetBytesAuto(int4(stride_.x, stride_.y, stride_.z, 1))); - RETURN_IF_ERROR(kernel_.SetBytesAuto( - int4(padding_.x * src_[0]->Batch(), padding_.y, padding_.z, 1))); - RETURN_IF_ERROR(kernel_.SetBytesAuto( - int4(dilation_.x * src_[0]->Batch(), dilation_.y, dilation_.z, 1))); - if (!IsSpecializedCase(channel_multiplier_)) { - RETURN_IF_ERROR(kernel_.SetBytesAuto(int32_t(channel_multiplier_))); - } - if (definition_.IsBatchSupported()) { - RETURN_IF_ERROR(kernel_.SetBytesAuto(src_[0]->Batch())); - } - RETURN_IF_ERROR(kernel_.SetBytesAuto(src_[0]->GetWBatchedHDS())); - RETURN_IF_ERROR(kernel_.SetBytesAuto(dst_[0]->GetWBatchedHDS())); - return absl::OkStatus(); -} - -int3 DepthwiseConvolution3D::GetGridSize() const { - const int grid_x = dst_[0]->Width() * dst_[0]->Batch(); - const int grid_y = dst_[0]->Height(); - const int grid_z = dst_[0]->Slices() * dst_[0]->Depth(); - return int3(grid_x, grid_y, grid_z); -} - -absl::Status DepthwiseConvolution3D::Tune(const TuningParameters& params) { - RETURN_IF_ERROR(BindArguments()); - return GetBestWorkGroup(params, kernel_, GetGridSize(), &work_group_size_); -} - -absl::Status DepthwiseConvolution3D::AddToQueue(CLCommandQueue* queue) { - RETURN_IF_ERROR(BindArguments()); - return queue->DispatchImplicit(kernel_, GetGridSize(), work_group_size_); -} - -absl::Status CreateDepthwiseConvolution3D( - const CreationContext& creation_context, const OperationDef& definition, - const DepthwiseConvolution3DAttributes& attr, - DepthwiseConvolution3D* result) { - *result = DepthwiseConvolution3D(definition, attr, *creation_context.device); - RETURN_IF_ERROR( - result->UploadWeights(attr.weights, creation_context.context)); - LinearStorageCreateInfo create_info; - create_info.storage_type = - DeduceLinearStorageType(definition.GetPrimaryStorageType()); - create_info.data_type = definition.GetDataType(); - create_info.name = "biases"; - create_info.aligned_size = attr.weights.shape.o * attr.weights.shape.i; - RETURN_IF_ERROR(CreateLinearStorage( - create_info, attr.bias, creation_context.context, &result->biases_)); - return absl::OkStatus(); -} - -} // namespace cl -} // namespace gpu -} // namespace tflite diff --git a/tensorflow/lite/delegates/gpu/cl/kernels/depthwise_conv_3d.h b/tensorflow/lite/delegates/gpu/cl/kernels/depthwise_conv_3d.h deleted file mode 100644 index 3c87ba5832c..00000000000 --- a/tensorflow/lite/delegates/gpu/cl/kernels/depthwise_conv_3d.h +++ /dev/null @@ -1,170 +0,0 @@ -/* Copyright 2019 The TensorFlow Authors. All Rights Reserved. - -Licensed under the Apache License, Version 2.0 (the "License"); -you may not use this file except in compliance with the License. -You may obtain a copy of the License at - - http://www.apache.org/licenses/LICENSE-2.0 - -Unless required by applicable law or agreed to in writing, software -distributed under the License is distributed on an "AS IS" BASIS, -WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -See the License for the specific language governing permissions and -limitations under the License. -==============================================================================*/ - -#ifndef TENSORFLOW_LITE_DELEGATES_GPU_CL_KERNELS_DEPTHWISE_CONV_3D_H_ -#define TENSORFLOW_LITE_DELEGATES_GPU_CL_KERNELS_DEPTHWISE_CONV_3D_H_ - -#include - -#include "tensorflow/lite/delegates/gpu/cl/buffer.h" -#include "tensorflow/lite/delegates/gpu/cl/kernels/gpu_operation.h" -#include "tensorflow/lite/delegates/gpu/cl/linear_storage.h" -#include "tensorflow/lite/delegates/gpu/cl/tensor.h" -#include "tensorflow/lite/delegates/gpu/cl/texture2d.h" -#include "tensorflow/lite/delegates/gpu/cl/util.h" -#include "tensorflow/lite/delegates/gpu/common/data_type.h" -#include "tensorflow/lite/delegates/gpu/common/operations.h" -#include "tensorflow/lite/delegates/gpu/common/shape.h" -#include "tensorflow/lite/delegates/gpu/common/status.h" -#include "tensorflow/lite/delegates/gpu/common/tensor.h" -#include "tensorflow/lite/delegates/gpu/common/types.h" - -namespace tflite { -namespace gpu { -namespace cl { - -class DepthwiseConvolution3D : public GPUOperation { - public: - DepthwiseConvolution3D() = default; - absl::Status AddToQueue(CLCommandQueue* queue) override; - absl::Status Tune(const TuningParameters& params) override; - - absl::Status Compile(const CreationContext& creation_context) override; - - // Move only - DepthwiseConvolution3D(DepthwiseConvolution3D&& operation); - DepthwiseConvolution3D& operator=(DepthwiseConvolution3D&& operation); - DepthwiseConvolution3D(const DepthwiseConvolution3D&) = delete; - DepthwiseConvolution3D& operator=(const DepthwiseConvolution3D&) = delete; - - private: - friend absl::Status CreateDepthwiseConvolution3D( - const CreationContext& creation_context, const OperationDef& definition, - const DepthwiseConvolution3DAttributes& attr, - DepthwiseConvolution3D* result); - DepthwiseConvolution3D(const OperationDef& definition, - const DepthwiseConvolution3DAttributes& attr, - const CLDevice& device); - template - absl::Status UploadWeights(const tflite::gpu::Tensor& weights, - CLContext* context); - - template - void RearrangeWeightsData(const tflite::gpu::Tensor& weights, - absl::Span dst); - - absl::Status BindArguments(); - int3 GetGridSize() const; - - Texture2D weights_tex2d_; - Buffer weights_buf_; - bool weights_are_buffer_; - - LinearStorage biases_; - - int3 kernel_size_; - int3 stride_; - int3 padding_; - int3 dilation_; - int channel_multiplier_; - - CLKernel kernel_; - int3 work_group_size_; -}; - -template -absl::Status DepthwiseConvolution3D::UploadWeights( - const tflite::gpu::Tensor& weights, CLContext* context) { - const int dst_channels = weights.shape.i * weights.shape.o; - const int dst_slices = DivideRoundUp(dst_channels, 4); - const int kernel_x = weights.shape.w; - const int kernel_y = weights.shape.h; - const int kernel_z = weights.shape.d; - - const int elements_count = kernel_x * kernel_y * kernel_z * dst_slices; - const bool f32_weights = definition_.precision == CalculationsPrecision::F32; - - const int float4_size = f32_weights ? 16 : 8; - - if (f32_weights) { - std::vector gpu_data(elements_count); - RearrangeWeightsData(weights, absl::MakeSpan(gpu_data)); - if (weights_are_buffer_) { - RETURN_IF_ERROR(CreateReadOnlyBuffer(float4_size * elements_count, - gpu_data.data(), context, - &weights_buf_)); - } else { - RETURN_IF_ERROR(CreateTexture2DRGBA( - definition_.GetDataType(), kernel_x * kernel_y * kernel_z, dst_slices, - gpu_data.data(), context, &weights_tex2d_)); - } - } else { - std::vector gpu_data(elements_count); - RearrangeWeightsData(weights, absl::MakeSpan(gpu_data)); - if (weights_are_buffer_) { - RETURN_IF_ERROR(CreateReadOnlyBuffer(float4_size * elements_count, - gpu_data.data(), context, - &weights_buf_)); - } else { - RETURN_IF_ERROR(CreateTexture2DRGBA( - definition_.GetDataType(), kernel_x * kernel_y * kernel_z, dst_slices, - gpu_data.data(), context, &weights_tex2d_)); - } - } - return absl::OkStatus(); -} - -template -void DepthwiseConvolution3D::RearrangeWeightsData( - const tflite::gpu::Tensor& weights, absl::Span dst) { - const int dst_channels = weights.shape.i * weights.shape.o; - const int dst_slices = DivideRoundUp(dst_channels, 4); - const int kernel_x = weights.shape.w; - const int kernel_y = weights.shape.h; - const int kernel_z = weights.shape.d; - - int counter = 0; - for (int d = 0; d < dst_slices; ++d) { - for (int z = 0; z < kernel_z; ++z) { - for (int y = 0; y < kernel_y; ++y) { - for (int x = 0; x < kernel_x; ++x) { - T filter_val; - for (int i = 0; i < 4; ++i) { - const int d_ch = d * 4 + i; - if (d_ch < dst_channels) { - const int f_index = weights.shape.LinearIndex( - {d_ch % weights.shape.o, y, x, z, d_ch / weights.shape.o}); - filter_val[i] = weights.data[f_index]; - } else { - filter_val[i] = 0.0f; - } - } - dst[counter++] = filter_val; - } - } - } - } -} - -absl::Status CreateDepthwiseConvolution3D( - const CreationContext& creation_context, const OperationDef& definition, - const DepthwiseConvolution3DAttributes& attr, - DepthwiseConvolution3D* result); - -} // namespace cl -} // namespace gpu -} // namespace tflite - -#endif // TENSORFLOW_LITE_DELEGATES_GPU_CL_KERNELS_DEPTHWISE_CONV_3D_H_ From bd6f60bf01875b4151ebf7bb37a7172ffa9ec426 Mon Sep 17 00:00:00 2001 From: Robert David Date: Tue, 23 Jun 2020 09:55:28 -0700 Subject: [PATCH 555/867] Rename few LSTM symbols to a name that is consistent among different versions. Also remove an unused std::vector variable. PiperOrigin-RevId: 317886335 Change-Id: Ib1e131c42d86b983630a4f8130d9950a45119b4a --- tensorflow/lite/kernels/lstm_eval.cc | 26 +++++++++++++------------- 1 file changed, 13 insertions(+), 13 deletions(-) diff --git a/tensorflow/lite/kernels/lstm_eval.cc b/tensorflow/lite/kernels/lstm_eval.cc index 5b4e8a8d479..ca8344d863b 100644 --- a/tensorflow/lite/kernels/lstm_eval.cc +++ b/tensorflow/lite/kernels/lstm_eval.cc @@ -490,7 +490,7 @@ inline void LstmStepFloat( // Temporary pre-allocated storage for quantized values: // quantized_input_ptr (same size as input_ptr) // quantized_output_state_ptr (same size as output_state_ptr) -// quantized_cell_state_ptr (same size as cell_state_ptr) +// quantized_output_scratch (same size as cell_state_ptr) // Temporary pre-allocated storage for recovered values: // recovered_cell_weights (same size as cell_to_*_weights) // @@ -540,7 +540,7 @@ inline void LstmStepHybrid( float* scratch2, float* scratch3, float* scaling_factors, float* scaling_factors_scratch, float* recovered_cell_weights, int8_t* quantized_input_ptr, int8_t* quantized_aux_input_ptr, - int8_t* quantized_output_state_ptr, int8_t* quantized_cell_state_ptr, + int8_t* quantized_output_state_ptr, int8_t* quantized_output_scratch, float* output_state_ptr, float* cell_state_ptr, int32_t* accum_scratch_ptr, float* output_ptr, int32_t* zero_points, int32_t* row_sums, int row_sums_size, bool* compute_row_sums, bool asymmetric_quantize_inputs, @@ -882,10 +882,10 @@ inline void LstmStepHybrid( if (!tensor_utils::IsZeroVector(output_gate_scratch, n_batch * n_cell)) { // Save quantization and matmul computation for all zero input. tensor_utils::BatchQuantizeFloats( - output_gate_scratch, n_batch, n_cell, quantized_cell_state_ptr, + output_gate_scratch, n_batch, n_cell, quantized_output_scratch, scaling_factors, zero_points, asymmetric_quantize_inputs); tensor_utils::MatrixBatchVectorMultiplyAccumulate( - projection_weights_ptr, n_output, n_cell, quantized_cell_state_ptr, + projection_weights_ptr, n_output, n_cell, quantized_output_scratch, projection_weights_scale, scaling_factors, n_batch, output_state_ptr, /*per_channel_scale=*/nullptr, asymmetric_quantize_inputs ? zero_points : nullptr, accum_scratch_ptr, @@ -996,7 +996,7 @@ inline void LstmStepHybrid( // output_state_ptr - size 'n_batch * n_output' // cell_state_ptr - size 'n_batch * n_cell' // output_ptr - size 'n_batch * n_output' -inline void LstmStepInteger( +inline void LstmStepInteger8x8_16( const int8_t* input_ptr, const int8_t* input_to_input_weight_ptr, int32_t effective_input_to_input_scale_a, int32_t effective_input_to_input_scale_b, @@ -1060,7 +1060,7 @@ inline void LstmStepInteger( int32_t output_state_zp, int16_t* cell_state_ptr, int8_t* output_ptr, int16_t* scratch0, int16_t* scratch1, int16_t* scratch2, int16_t* scratch3, int8_t* scratch4, int32_t* scratch5, CpuBackendContext* context) { - ruy::profiler::ScopeLabel label("LstmStepInteger"); + ruy::profiler::ScopeLabel label("LstmStepInteger8x8_16"); // Make named scratch buffers for the different gates. int16_t* input_gate_scratch = scratch0; int16_t* forget_gate_scratch = scratch1; @@ -1336,7 +1336,7 @@ inline void LstmStepInteger( // cell_state_ptr - size 'n_batch * n_cell' // output_ptr - size 'n_batch * n_output' // TODO(b/148688698): Move zero point calculation into Prepare(). -void LstmStepInteger( +inline void LstmStepInteger8x8_8( const int8_t* input_ptr, int32_t input_zp, const int8_t* input_to_input_weight_ptr, int32_t effective_input_to_input_scale_a, @@ -1391,6 +1391,7 @@ void LstmStepInteger( int8_t* scratch0, int8_t* scratch1, int16_t* scratch2, int16_t* scratch3, int16_t* scratch4, int16_t* scratch5, int16_t* scratch6, int16_t* scratch7) { + ruy::profiler::ScopeLabel label("LstmStepInteger8x8_8"); // Make named scratch buffers for the different gates. int16_t* input_gate_scratch = scratch5; int16_t* forget_gate_scratch = scratch2; @@ -1426,7 +1427,7 @@ void LstmStepInteger( tensor_utils::ApplySigmoidFloat(forget_gate_scratch, n_batch, n_cell, forget_gate_scratch); - // Update gate. + // Cell gate. std::fill_n(scratch0, n_batch * n_cell, 0); std::fill_n(scratch1, n_batch * n_cell, 0); tensor_utils::MatrixBatchVectorMultiply( @@ -1444,13 +1445,13 @@ void LstmStepInteger( intermediate_scale_a[4], intermediate_scale_b[4], intermediate_scale_a[5], intermediate_scale_b[5], n_batch, n_cell, cell_gate_scratch); - // Update gate layer norm. + // Cell gate layer norm. tensor_utils::ApplyLayerNormFloat( cell_gate_scratch, layer_norm_cell_weight_ptr, layer_norm_cell_scale_a, layer_norm_cell_scale_b, cell_gate_bias_ptr, n_batch, n_cell, cell_gate_scratch); - // Update gate tanh. + // Cell gate tanh. tensor_utils::ApplyTanhFloat(cell_gate_scratch, n_batch, n_cell, -12, cell_gate_scratch); @@ -1505,7 +1506,6 @@ void LstmStepInteger( tensor_utils::ApplyTanhFloat(cell_state_ptr, n_batch, n_cell, -15, forget_gate_scratch); - std::vector hidden(n_batch * n_cell); tensor_utils::CwiseMul(output_gate_scratch, forget_gate_scratch, n_batch, n_cell, 15 + 15 - 15, cell_gate_scratch); @@ -2004,7 +2004,7 @@ TfLiteStatus EvalInteger8x8_16( const int t_rel = t; int8_t* output_ptr = GetTensorData(output) + t_rel * output_step; const int8_t* input_ptr = GetTensorData(input) + t_rel * input_step; - LstmStepInteger( + LstmStepInteger8x8_16( input_ptr, GetTensorData(input_to_input_weights), integer_lstm_param->effective_input_to_input_scale_a, integer_lstm_param->effective_input_to_input_scale_b, @@ -2140,7 +2140,7 @@ TfLiteStatus EvalInteger8x8_8( int8_t* output_ptr = GetTensorData(output) + t_rel * output_step; // Input can be int8 asymmetric or int16 symmetric. const int8_t* input_ptr = GetTensorData(input) + t_rel * input_step; - lstm_eval::LstmStepInteger( + lstm_eval::LstmStepInteger8x8_8( input_ptr, input_zp, GetTensorData(input_to_input_weights), From b07691301fa26602c35531aa6349e95136c9ef3d Mon Sep 17 00:00:00 2001 From: Smit Hinsu Date: Tue, 23 Jun 2020 10:05:49 -0700 Subject: [PATCH 556/867] MultiProcessRunner: Add UnexpectedSubprocessExitError to be raised if the exit code from a subprocess in unexpected. This results in subprocess having segfault failing the test, which would not have before this change. PiperOrigin-RevId: 317888643 Change-Id: Id1c372ec695e8478bbb9c0888516ca9af4bb9799 --- .../python/distribute/multi_process_runner.py | 38 ++----------------- .../distribute/multi_process_runner_test.py | 35 ----------------- 2 files changed, 3 insertions(+), 70 deletions(-) diff --git a/tensorflow/python/distribute/multi_process_runner.py b/tensorflow/python/distribute/multi_process_runner.py index cb460c8fff5..db31b9c4dd4 100644 --- a/tensorflow/python/distribute/multi_process_runner.py +++ b/tensorflow/python/distribute/multi_process_runner.py @@ -475,28 +475,14 @@ class MultiProcessRunner(object): process_statuses = self._queue_to_list(self._process_status_queue) if not self._all_forced_terminated and len( process_statuses) != self._outstanding_subprocess_count: - raise UnexpectedSubprocessExitError( - 'Missing status(es) from %d subprocess(es). See logs for details.' % - (self._outstanding_subprocess_count - len(process_statuses)), - self._get_mpr_result(process_statuses)) + raise RuntimeError( + 'missing statuses from %d subproceses.' % + (self._outstanding_subprocess_count - len(process_statuses))) for process_status in process_statuses: assert isinstance(process_status, _ProcessStatusInfo) if not process_status.is_successful: six.reraise(*process_status.exc_info) - # Checking all the processes that are expected to exit properly. - for (task_type, task_id), p in self._processes.items(): - if self._dependence_on_chief and task_type != 'chief': - # If _dependence_on_chief, other processes may have been - # forced-terminated, which is expected. - continue - # Successfully exiting process has exit code 0. - if p.exitcode > 0: - raise UnexpectedSubprocessExitError( - 'Subprocess %s-%d exited with exit code %d. See logs for details.' % - (task_type, task_id, p.exitcode), - self._get_mpr_result(process_statuses)) - logging.info('Joining log reading threads.') for thread in self._reading_threads: thread.join() @@ -532,8 +518,6 @@ class MultiProcessRunner(object): for (task_type, task_id), p in self._processes.items(): try: os.kill(p.pid, sig) - logging.info('%s-%d terminated with signal %r.', task_type, task_id, - sig) except ProcessLookupError: logging.info('Attempting to kill %s-%d but it does not exist.', task_type, task_id) @@ -686,9 +670,6 @@ class _ProcFunc(object): self._resources.process_status_queue.put(info) self._close_streaming() - # Exit with code 0 as it's considered successful exit at this point. - sys.exit(0) - class SubprocessTimeoutError(RuntimeError): """An error that indicates there is at least one subprocess timing out. @@ -703,19 +684,6 @@ class SubprocessTimeoutError(RuntimeError): self.mpr_result = mpr_result -class UnexpectedSubprocessExitError(RuntimeError): - """An error indicating there is at least one subprocess with unexpected exit. - - When this is raised, a `MultiProcessRunnerResult` object can be retrieved by - `UnexpectedSubprocessExitError`'s mpr_result attribute. See - `MultiProcessRunner.join()` for more information. - """ - - def __init__(self, msg, mpr_result): - super(UnexpectedSubprocessExitError, self).__init__(msg) - self.mpr_result = mpr_result - - def _set_tf_config(task_type, task_id, cluster_spec, rpc_layer=None): """Set TF_CONFIG environment variable.""" tf_config_dict = { diff --git a/tensorflow/python/distribute/multi_process_runner_test.py b/tensorflow/python/distribute/multi_process_runner_test.py index 529d7fd91a5..d6e04010e34 100644 --- a/tensorflow/python/distribute/multi_process_runner_test.py +++ b/tensorflow/python/distribute/multi_process_runner_test.py @@ -18,7 +18,6 @@ from __future__ import absolute_import from __future__ import division from __future__ import print_function -import ctypes import json import os import threading @@ -299,40 +298,6 @@ class MultiProcessRunnerTest(test.TestCase): self.assertTrue( any('something printed' in line for line in list_to_assert)) - def test_seg_fault_raises_error(self): - - def proc_func_expected_to_seg_fault(): - ctypes.string_at(0) # Intentionally made seg fault. - - with self.assertRaises( - multi_process_runner.UnexpectedSubprocessExitError) as cm: - multi_process_runner.run( - proc_func_expected_to_seg_fault, - multi_worker_test_base.create_cluster_spec(num_workers=1), - list_stdout=True) - self.assertIn('Missing status(es) from 1 subprocess(es).', - str(cm.exception)) - list_to_assert = cm.exception.mpr_result.stdout - self.assertTrue(any('SIGSEGV' in line for line in list_to_assert)) - - def test_seg_fault_in_chief_raises_error(self): - - def proc_func_expected_to_seg_fault(): - if multi_worker_test_base.get_task_type() == 'worker': - time.sleep(10000) - ctypes.string_at(0) # Intentionally made seg fault. - - with self.assertRaises( - multi_process_runner.UnexpectedSubprocessExitError) as cm: - multi_process_runner.run( - proc_func_expected_to_seg_fault, - multi_worker_test_base.create_cluster_spec( - has_chief=True, num_workers=1), - list_stdout=True) - self.assertIn('Subprocess chief-0 exited with exit code', - str(cm.exception)) - list_to_assert = cm.exception.mpr_result.stdout - self.assertTrue(any('SIGSEGV' in line for line in list_to_assert)) if __name__ == '__main__': multi_process_runner.test_main() From e4d6335bcb7a73cd8967c2c12339380aa1ae284f Mon Sep 17 00:00:00 2001 From: Hye Soo Yang Date: Tue, 23 Jun 2020 10:21:31 -0700 Subject: [PATCH 557/867] Refactor DecodeImageOp for the purpose of removing redundant data parsing and format checks from python wrapper and having them take place only in kernels. Remove security concerns. This change: - Creates new op kernel (`DecodeImageV2Op`) that can decode all four image formats (jpg, png, gif, bmp). `DecodeImage` is the op name. `DecodeBmpOp` is moved into `DecodeImageV2Op`. (Now we have `gen_image_ops.decode_image` as opposed to previous `decode_image` which was a pure python implementation.) - Updates GIF decoder to take in `expand_animation` flag for decoding just one frame. - Removes data parsing and format checking logic from python layer entirely. - Updates magic bytes for detecting image formats. - Replicates portions of `convert_image_dtype` functionality in kernel (for optionally converting uint8/uint16 -> float32). PiperOrigin-RevId: 317891936 Change-Id: I84f18e053f6dad845d9f2a61e1119f4de131c85d --- .../base_api/api_def_DecodeImage.pbtxt | 51 ++ .../python_api/api_def_DecodeImage.pbtxt | 4 + tensorflow/core/kernels/decode_image_op.cc | 459 +++++++++++++++++- tensorflow/core/lib/gif/gif_io.cc | 13 +- tensorflow/core/lib/gif/gif_io.h | 2 +- tensorflow/core/ops/image_ops.cc | 45 ++ tensorflow/core/ops/image_ops_test.cc | 33 +- tensorflow/python/ops/image_ops_impl.py | 17 + tensorflow/python/ops/image_ops_test.py | 201 ++++---- .../api/golden/v1/tensorflow.raw_ops.pbtxt | 4 + .../api/golden/v2/tensorflow.raw_ops.pbtxt | 4 + 11 files changed, 738 insertions(+), 95 deletions(-) create mode 100644 tensorflow/core/api_def/base_api/api_def_DecodeImage.pbtxt create mode 100644 tensorflow/core/api_def/python_api/api_def_DecodeImage.pbtxt diff --git a/tensorflow/core/api_def/base_api/api_def_DecodeImage.pbtxt b/tensorflow/core/api_def/base_api/api_def_DecodeImage.pbtxt new file mode 100644 index 00000000000..c534425eb24 --- /dev/null +++ b/tensorflow/core/api_def/base_api/api_def_DecodeImage.pbtxt @@ -0,0 +1,51 @@ +op { + graph_op_name: "DecodeImage" + in_arg { + name: "contents" + description: <