STT-tensorflow/tensorflow/compiler/xla/service/root_instruction_sinker.cc
Berkin Ilbeyi f9d6dd6269 [XLA] Add a RootInstructionSinker pass.
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
2020-04-22 14:46:15 -07:00

74 lines
2.6 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.
==============================================================================*/
#include "tensorflow/compiler/xla/service/root_instruction_sinker.h"
#include "tensorflow/compiler/xla/service/tuple_util.h"
namespace xla {
namespace {
// Sinks the root of the given computation for tuple root types.
void SinkTupleRoot(HloComputation* computation) {
HloInstruction* root = computation->root_instruction();
CHECK(root->shape().IsTuple());
HloInstruction* new_root = TupleUtil::Duplicate(root);
// Add the new instructions to the schedule.
HloInstructionSequence& sequence =
computation->parent()->schedule().GetOrCreateSequence(computation);
for (HloInstruction* operand : new_root->operands()) {
sequence.push_back(operand);
}
sequence.push_back(new_root);
computation->set_root_instruction(new_root);
}
// Sinks the root of the given computation for not-tuple root types.
void SinkNontupleRoot(HloComputation* computation) {
HloInstruction* root = computation->root_instruction();
CHECK(!root->shape().IsTuple());
HloInstruction* new_root = computation->AddInstruction(
HloInstruction::CreateBitcast(root->shape(), root));
HloInstructionSequence& sequence =
computation->parent()->schedule().GetOrCreateSequence(computation);
sequence.push_back(new_root);
computation->set_root_instruction(new_root);
}
} // namespace
StatusOr<bool> RootInstructionSinker::Run(HloModule* module) {
TF_RET_CHECK(module->has_schedule());
bool modified = false;
for (HloComputation* computation : module->MakeNonfusionComputations()) {
HloInstructionSequence& sequence =
module->schedule().GetOrCreateSequence(computation);
if (computation->root_instruction() ==
sequence.instructions().at(sequence.size() - 1)) {
continue;
}
if (computation->root_instruction()->shape().IsTuple()) {
SinkTupleRoot(computation);
} else {
SinkNontupleRoot(computation);
}
modified = true;
}
return modified;
}
} // namespace xla