f895a9e996
The idea of the refactoring is that XlaExpression is the canonical XLA representation of a symbolic TF value. So in general a computation to compile is a function with type [XlaExpression] -> [XlaExpression], and in a future change we will add a method to XlaCompiler that exposes pretty much exactly that API. The current TF function/graph/op compilation methods are specific ways to build such a function. * Move XlaExpression into its own file. Improve its ergonomics; it is really a kind of sum type. Also move some useful common methods on XlaExpressions into the XlaExpression class. * Add support for passing and returning XlaExpressions via XlaOpKernelContext, since they are the underlying representation. The remaining *Input() and *Output() methods are really just conveniences built on top. * Simplify _Arg and _Retval to just get and set an XlaExpression from an XlaContext. Move logic to flatten return values out of _Retval and move it instead into XlaCompiler so it can be reused when compiling non-graph computations. * Move logic to assign cores to arguments and return values into a common place in XlaCompiler. PiperOrigin-RevId: 221104314
70 lines
2.6 KiB
C++
70 lines
2.6 KiB
C++
/* 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_TF2XLA_XLA_COMPILATION_DEVICE_H_
|
|
#define TENSORFLOW_COMPILER_TF2XLA_XLA_COMPILATION_DEVICE_H_
|
|
|
|
#include <memory>
|
|
|
|
#include "tensorflow/core/common_runtime/local_device.h"
|
|
#include "tensorflow/core/framework/device_base.h"
|
|
#include "tensorflow/core/framework/tensor.h"
|
|
#include "tensorflow/core/lib/core/status.h"
|
|
#include "tensorflow/core/platform/mem.h"
|
|
#include "tensorflow/core/public/session_options.h"
|
|
|
|
namespace tensorflow {
|
|
|
|
// Class is defined in xla_compilation_device.cc, reference
|
|
// included here only so the XlaCompilationDevice allocator_ member can be
|
|
// declared.
|
|
class XlaCompilationAllocator;
|
|
|
|
// This is a 'dummy' TensorFlow device that is only used to execute a
|
|
// subgraph of XLA compilation Ops to construct a compiled version
|
|
// of the subgraph's computation. It has a 'dummy' allocator that
|
|
// backs each Tensor with an XlaExpression. The shape of the Tensor
|
|
// matches the shape of XlaExpression.
|
|
//
|
|
// We deliberately don't register a device factory because we *never*
|
|
// want placement to put Ops on a compilation device. The device is created
|
|
// manually, not using a factory.
|
|
//
|
|
// XLA compilation is not thread-safe. OpKernels registered on the
|
|
// XlaCompilationDevice must not use threads or concurrency.
|
|
class XlaCompilationDevice : public LocalDevice {
|
|
public:
|
|
XlaCompilationDevice(const SessionOptions& options, DeviceType type);
|
|
|
|
~XlaCompilationDevice() override;
|
|
|
|
Allocator* GetAllocator(AllocatorAttributes attr) override;
|
|
|
|
void Compute(OpKernel* op_kernel, OpKernelContext* context) override;
|
|
|
|
Status Sync() override;
|
|
|
|
Status MakeTensorFromProto(const TensorProto& tensor_proto,
|
|
const AllocatorAttributes alloc_attrs,
|
|
Tensor* tensor) override;
|
|
|
|
private:
|
|
std::unique_ptr<XlaCompilationAllocator> allocator_;
|
|
};
|
|
|
|
} // namespace tensorflow
|
|
|
|
#endif // TENSORFLOW_COMPILER_TF2XLA_XLA_COMPILATION_DEVICE_H_
|