Relax the stringent memory allocator constraints in AssignOp if a Grappler graph...
authorA. Unique TensorFlower <gardener@tensorflow.org>
Tue, 1 May 2018 20:34:39 +0000 (13:34 -0700)
committerTensorFlower Gardener <gardener@tensorflow.org>
Tue, 1 May 2018 20:37:38 +0000 (13:37 -0700)
commit3b7f22f9180935919bab478adb45037b1f0d38c2
tree535358ced05946650aab972fc0a83b4499ef0fc4
parent9149558a639efe82baf1b5201feccf2411343a8a
Relax the stringent memory allocator constraints in AssignOp if a Grappler graph analysis determines it to be safe. This will allow Assign to reuse the input buffer to initialize the variable in many cases.

PiperOrigin-RevId: 194988134
tensorflow/core/grappler/op_types.cc
tensorflow/core/grappler/op_types.h
tensorflow/core/grappler/optimizers/memory_optimizer.cc
tensorflow/core/grappler/optimizers/memory_optimizer_test.cc
tensorflow/core/grappler/utils.cc
tensorflow/core/grappler/utils.h
tensorflow/core/kernels/assign_op.h
tensorflow/core/kernels/resource_variable_ops.cc