f9d6dd6269
Memory space assignment requires ROOT instructions of while bodies to the the latest in the schedule. This pass sinks the ROOTs to the end of the schedule. To make sure dependencies are respected (e.g. the instructions after the ROOT may depend on the ROOT instruction), this pass will insert either tuple(gte(old_root), gte(old_root), ...) or bitcast(old_root) at the end of the computation. Note that Hlo live range, hence copy insertion, always calculate ROOTs' live ranges to be until the end of the computation, so it should be safe to move the ROOTs to the end of the schedule. PiperOrigin-RevId: 307902025 Change-Id: I2bdcc9cc660fcffbbe7611f76034c9924cf29c3d
42 lines
1.7 KiB
C++
42 lines
1.7 KiB
C++
/* 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_XLA_SERVICE_ROOT_INSTRUCTION_SINKER_H_
|
|
#define TENSORFLOW_COMPILER_XLA_SERVICE_ROOT_INSTRUCTION_SINKER_H_
|
|
|
|
#include "tensorflow/compiler/xla/service/hlo_module.h"
|
|
#include "tensorflow/compiler/xla/service/hlo_pass_interface.h"
|
|
|
|
namespace xla {
|
|
|
|
// Given a scheduled HLO module, this pass sinks the ROOT of the instruction to
|
|
// the bottom of the non-fusion computations. To avoid dependency violations of
|
|
// moving the ROOT instruction, it creates a new ROOT instruction that looks
|
|
// like the following:
|
|
// - For tuple ROOT type:
|
|
// new_root = tuple(gte(old_root), gte(old_root), ...)
|
|
// - For non-tuple ROOT type:
|
|
// new_root = bitcast(old_root)
|
|
class RootInstructionSinker : public HloModulePass {
|
|
public:
|
|
~RootInstructionSinker() override = default;
|
|
absl::string_view name() const override { return "root-instruction-sinker"; }
|
|
StatusOr<bool> Run(HloModule* module) override;
|
|
};
|
|
|
|
} // namespace xla
|
|
|
|
#endif // TENSORFLOW_COMPILER_XLA_SERVICE_ROOT_INSTRUCTION_SINKER_H_
|