Merge pull request #14955 from guschmue/win-quantized-ops

add support for quantized ops on windows

tensorflow/contrib/cmake/README.md

@@ -19,23 +19,6 @@ for instructions on how to install a pre-built TensorFlow package on Windows.
 ### Current known limitations
 * It is not possible to load a custom Op library.
 * GCS file system is not supported.
-* The following Ops are not currently implemented:
- - Dequantize
- - QuantizeAndDequantize
- - QuantizedAvgPool
- - QuantizedBatchNomWithGlobalNormalization
- - QuantizedBiasAdd
- - QuantizedConcat
- - QuantizedConv2D
- - QuantizedMatmul
- - QuantizedMaxPoo
- - QuantizeDownAndShrinkRange
- - QuantizedRelu
- - QuantizedRelu6
- - QuantizedReshape
- - QuantizeV2
- - RequantizationRange
- - Requantize
 
 ## Building with CMake
 

tensorflow/contrib/cmake/external/gemmlowp.cmake

@@ -14,8 +14,8 @@
 # ==============================================================================
 include (ExternalProject)
 
-set(gemmlowp_URL https://mirror.bazel.build/github.com/google/gemmlowp/archive/010bb3e71a26ca1d0884a167081d092b43563996.zip)
-set(gemmlowp_HASH SHA256=dd2557072bde12141419cb8320a9c25e6ec41a8ae53c2ac78c076a347bb46d9d)
+set(gemmlowp_URL https://github.com/google/gemmlowp/archive/6a2a90822e8546fc2bfa7044de0faf1c1cb4862f.zip)
+set(gemmlowp_HASH SHA256=3447948d219f3270383766bbe08942888c0eb4e0ca6663c0e0548502ec5bb77d)
 set(gemmlowp_BUILD ${CMAKE_CURRENT_BINARY_DIR}/gemmlowp/src/gemmlowp)
 set(gemmlowp_INCLUDE_DIR ${CMAKE_CURRENT_BINARY_DIR}/gemmlowp/src/gemmlowp)
 

tensorflow/contrib/cmake/tf_core_kernels.cmake

@@ -150,9 +150,6 @@ list(REMOVE_ITEM tf_core_kernels_srcs ${tf_core_kernels_exclude_srcs})
 if(WIN32)
   file(GLOB_RECURSE tf_core_kernels_windows_exclude_srcs
       # not working on windows yet
-      "${tensorflow_source_dir}/tensorflow/core/kernels/meta_support.*"
-      "${tensorflow_source_dir}/tensorflow/core/kernels/*quantiz*.h"
-      "${tensorflow_source_dir}/tensorflow/core/kernels/*quantiz*.cc"
       "${tensorflow_source_dir}/tensorflow/core/kernels/neon/*"
       # not in core - those are loaded dynamically as dll
       "${tensorflow_source_dir}/tensorflow/contrib/nearest_neighbor/kernels/hyperplane_lsh_probes.cc"

tensorflow/contrib/cmake/tf_tests.cmake

@@ -145,6 +145,8 @@ if (tensorflow_BUILD_PYTHON_TESTS)
     "${tensorflow_source_dir}/tensorflow/contrib/estimator/python/estimator/*_test.py"
     "${tensorflow_source_dir}/tensorflow/python/kernel_tests/*.py"
     "${tensorflow_source_dir}/tensorflow/python/meta_graph_transform/*_test.py"
+    "${tensorflow_source_dir}/tensorflow/python/ops/quantized_conv_ops_test.py"
+    "${tensorflow_source_dir}/tensorflow/python/ops/quantized_ops_test.py"
     "${tensorflow_source_dir}/tensorflow/python/platform/build_info_test.py"
     "${tensorflow_source_dir}/tensorflow/python/profiler/*_test.py"
     "${tensorflow_source_dir}/tensorflow/python/profiler/internal/*_test.py"

tensorflow/core/kernels/quantized_conv_ops.cc

@@ -268,6 +268,13 @@ class Im2ColConvFunctor {
     Im2ColBufferResource<T1, chunk_value_count>* im2col_buffer_resource;
     std::function<Status(Im2ColBufferResource<T1, chunk_value_count>**)>
         creator = [](Im2ColBufferResource<T1, chunk_value_count>** resource) {
+#ifdef _MSC_VER
+          // MSVC complains about the capture of chunk_value_count which oddly
+          // works fine in conv_ops_using_gemm.cc for example.
+          // Define chunk_value_count inside the lambda for now.
+          const int64 chunk_value_count =
+              (kMaxChunkSize + (sizeof(T1) - 1)) / sizeof(T1);
+#endif
           *resource = new Im2ColBufferResource<T1, chunk_value_count>();
           return Status::OK();
         };

tensorflow/python/ops/quantized_conv_ops_test.py

@@ -93,7 +93,7 @@ class Conv2DTest(test.TestCase):
     quantized_range = ((quantized_max - quantized_min) * range_adjust)
     range_scale = (quantized_range / number_of_steps)
     lowest_quantized = -(1 << (number_of_bits - 1))
-    result = np.array([(quantized_min + ((x - lowest_quantized) * range_scale))
+    result = np.array([(quantized_min + ((float(x) - lowest_quantized) * range_scale))
                        for x in quantized.flatten()])
     return result
 

tensorflow/python/ops/quantized_ops_test.py

@@ -0,0 +1,57 @@
+# 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.
+# ==============================================================================
+"""Functional tests for quantized operations."""
+
+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.ops import array_ops
+from tensorflow.python.platform import test
+
+
+class QuantizedOpsTest(test.TestCase):
+
+  def __init__(self, method_name="runTest"):
+    super(QuantizedOpsTest, self).__init__(method_name)
+
+  def testQuantizeOp(self):
+    expected_output = [1, 1, 2, 127, 255, 255]
+    with self.test_session(use_gpu=False) as sess:
+      x = constant_op.constant([1.0, 1.25, 1.75, 127.0, 255.0, 500.0], shape=[6], dtype=dtypes.float32)
+      x_min = 0.0
+      x_max = 255.0
+      op = array_ops.quantize(x, x_min, x_max, dtypes.quint8, mode="MIN_FIRST")
+      value = sess.run(op)
+      self.assertArrayNear(expected_output, value.output, 0.1)
+
+  def testDequantizeOp(self):
+    expected_output = [1.0, 2.0, 4.0, 8.0, 16.0, 255.0]
+    inp = np.array([1, 2, 4, 8, 16, 255]).astype(np.uint8)
+    with self.test_session(use_gpu=False) as sess:
+      x = constant_op.constant(inp, shape=[6], dtype=dtypes.quint8)
+      x_min = 0.0
+      x_max = 255.0
+      op = array_ops.dequantize(x, x_min, x_max, mode="MIN_FIRST")
+      value = sess.run(op)
+      self.assertArrayNear(expected_output, value, 0.1)
+
+
+if __name__ == "__main__":
+  test.main()