Reduce Layer.__call__ overhead by ~15%.

Performance improvements to how Variable autocasting is enabled:

- Uses a class-based context manager to enable autocasting rather than
  contextlib.contextmanager wrapper (class-based is faster).
- Skips property accesses
- Fewer reads of expensive thread local objects

PiperOrigin-RevId: 314377779
Change-Id: Ia1cca0b994d3d3745407e1b568cc46e4b6cafd9e

tensorflow/python/framework/ops.py

@@ -5142,9 +5142,8 @@ class Graph(object):
   def _auto_cast_variable_read_dtype(self, dtype):
     self._thread_local._auto_cast_variable_read_dtype = dtype  # pylint: disable=protected-access
 
-  @tf_contextlib.contextmanager
   def _enable_auto_casting_variables(self, dtype):
-    """Context manager to automatically cast AutoCastVariables.
+    """Returns a context manager to automatically cast AutoCastVariables.
 
     If an AutoCastVariable `var` is used under this context manager, it will be
     casted to `dtype` before being used.
@@ -5154,15 +5153,10 @@ class Graph(object):
     Args:
       dtype: The dtype that AutoCastVariables should be casted to.
 
-    Yields:
+    Returns:
-      Nothing.
+      Context manager.
     """
-    prev_read_dtype = self._auto_cast_variable_read_dtype
+    return enable_auto_cast_variables(dtype, graph=self)
-    try:
-      self._auto_cast_variable_read_dtype = dtype
-      yield
-    finally:
-      self._auto_cast_variable_read_dtype = prev_read_dtype
 
   def _mutation_lock(self):
     """Returns a lock to guard code that creates & mutates ops.
@@ -5179,6 +5173,36 @@ class Graph(object):
     return self._group_lock.group(_SESSION_RUN_LOCK_GROUP)
 
 
+class enable_auto_cast_variables(object):
+  """Enables the autocasting of `AutoCastVariable`s.
+
+  Under this context manager, `AutoCastVariable`s will be cast to `dtype` if
+  `dtype` is floating-point. Otherwise, `AutoCastVariable`s will not be cast.
+  """
+
+  def __init__(self, dtype, graph=None):
+    if dtype and not dtype.is_floating:
+      self._dtype = None
+    else:
+      self._dtype = dtype
+    if graph is None:
+      self._graph = get_default_graph()
+    else:
+      self._graph = graph
+
+  def __enter__(self):
+    # For performance, access `_thread_local` attr directly rather than
+    # @property wrappers.
+    graph_thread_local = self._graph._thread_local
+    self._prev_read_dtype = getattr(graph_thread_local,
+                                    "_auto_cast_variable_read_dtype", None)
+    graph_thread_local._auto_cast_variable_read_dtype = self._dtype
+
+  def __exit__(self, type_arg, value_arg, traceback_arg):
+    self._graph._thread_local._auto_cast_variable_read_dtype = (
+        self._prev_read_dtype)
+
+
 # TODO(agarwal): currently device directives in an outer eager scope will not
 # apply to inner graph mode code. Fix that.
 

tensorflow/python/keras/engine/base_layer.py

@@ -937,8 +937,7 @@ class Layer(module.Module, version_utils.LayerVersionSelector):
               call_fn = self.call
 
             try:
-              with base_layer_utils.autocast_context_manager(
+              with ops.enable_auto_cast_variables(self._compute_dtype_object):
-                  self._compute_dtype_object):
                 # Add auto_control_deps in V2 when they are not already added by
                 # a `tf.function`.
                 if (ops.executing_eagerly_outside_functions() and
@@ -999,8 +998,7 @@ class Layer(module.Module, version_utils.LayerVersionSelector):
         with ops.name_scope_v2(self.name):
           self._maybe_build(inputs)
           cast_inputs = self._maybe_cast_inputs(inputs, input_list)
-          with base_layer_utils.autocast_context_manager(
+          with ops.enable_auto_cast_variables(self._compute_dtype_object):
-              self._compute_dtype_object):
             outputs = self.call(cast_inputs, *args, **kwargs)
           self._handle_activity_regularization(inputs, outputs)
           self._set_mask_metadata(inputs, outputs, input_masks, build_graph)
@@ -1288,7 +1286,7 @@ class Layer(module.Module, version_utils.LayerVersionSelector):
       if callable(loss):
         # We run the loss without autocasting, as regularizers are often
         # numerically unstable in float16.
-        with base_layer_utils.autocast_context_manager(None):
+        with ops.enable_auto_cast_variables(None):
           loss = loss()
       if loss is None:
         return None  # Will be filtered out when computing the .losses property

tensorflow/python/keras/engine/base_layer_utils.py

@@ -481,23 +481,6 @@ def training_arg_passed_to_call(argspec, args, kwargs):
   return 'training' in full_args and full_args['training'] is not None
 
 
-def autocast_context_manager(dtype):
-  """Returns a context manager to autocast AutoCastVariables.
-
-  Under this context manager, AutoCastVariables will be casted to `dtype` if
-  `dtype` is floating-point. Otherwise, AutoCastVariables will not be casted.
-
-  Args:
-    dtype: The dtype to cast AutoCastVariables to, or None.
-
-  Returns:
-    A context manager to automatically cast AutoCastVariables.
-  """
-  if dtype and not dtype.is_floating:
-    dtype = None
-  return ops.get_default_graph()._enable_auto_casting_variables(dtype)  # pylint: disable=protected-access
-
-
 def is_subclassed(layer):
   """Returns True if the object is a subclassed layer or subclassed model."""
   return (layer.__module__.find('keras.engine') == -1 and

tensorflow/python/keras/engine/base_layer_v1.py

@@ -767,8 +767,7 @@ class Layer(base_layer.Layer):
 
           if not self.dynamic:
             try:
-              with base_layer_utils.autocast_context_manager(
+              with ops.enable_auto_cast_variables(self._compute_dtype_object):
-                  self._compute_dtype_object):
                 outputs = call_fn(cast_inputs, *args, **kwargs)
 
             except errors.OperatorNotAllowedInGraphError as e:
@@ -812,8 +811,7 @@ class Layer(base_layer.Layer):
         with backend.name_scope(self._name_scope()):
           self._maybe_build(inputs)
           cast_inputs = self._maybe_cast_inputs(inputs)
-          with base_layer_utils.autocast_context_manager(
+          with ops.enable_auto_cast_variables(self._compute_dtype_object):
-              self._compute_dtype_object):
             outputs = self.call(cast_inputs, *args, **kwargs)
           self._handle_activity_regularization(inputs, outputs)
           self._set_mask_metadata(inputs, outputs, input_masks)
@@ -1023,7 +1021,7 @@ class Layer(base_layer.Layer):
       if callable(loss):
         # We run the loss without autocasting, as regularizers are often
         # numerically unstable in float16.
-        with base_layer_utils.autocast_context_manager(None):
+        with ops.enable_auto_cast_variables(None):
           loss = loss()
       if loss is None:
         return None  # Will be filtered out when computing the .losses property

tensorflow/python/keras/saving/saved_model/save_impl.py

@@ -26,6 +26,7 @@ import weakref
 
 from tensorflow.python.eager import def_function
 from tensorflow.python.eager import function
+from tensorflow.python.framework import ops
 from tensorflow.python.framework import tensor_shape
 from tensorflow.python.framework import tensor_spec
 from tensorflow.python.keras import backend as K
@@ -521,8 +522,7 @@ def layer_call_wrapper(call_collection, method):
     with base_layer_utils.call_context().enter(
         layer, inputs=inputs, build_graph=False, training=training,
         saving=True):
-      with base_layer_utils.autocast_context_manager(
+      with ops.enable_auto_cast_variables(layer._compute_dtype_object):  # pylint: disable=protected-access
-          layer._compute_dtype_object):  # pylint: disable=protected-access
         ret = method(*args, **kwargs)
     _restore_layer_losses(original_losses)
     return ret