experiments/STT-tensorflow
1
0
Fork 0
Code Issues Pull Requests Projects Releases Wiki Activity

[XLA:CPU] Remove code from parallel CPU backend outlining that was causing unnecessary copies to be inserted, and which is no longer necessary since we added co-located buffer support for kCall.

Browse Source 
*) All bitcast copy is no longer necessary as CopyInsertion will insert copies
at the root of the computation for a parameter which is live-out.
*) Copy if root does not define buffer no longer necessary because colocated
assignment looks at points-to set of root instruction.

PiperOrigin-RevId: 168412076
This commit is contained in:
A. Unique TensorFlower 2017-09-12 11:48:18 -07:00 committed by TensorFlower Gardener
parent 5da4df92c2
commit 7e023d865d
2 changed files with 0 additions and 43 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

2
tensorflow/compiler/xla/service/cpu/BUILD
View File

@ -519,8 +519,6 @@ cc_library(
"//tensorflow/compiler/xla/service:hlo",
"//tensorflow/compiler/xla/service:hlo_cost_analysis",
"//tensorflow/compiler/xla/service:hlo_pass",
"//tensorflow/compiler/xla/service:logical_buffer",
"//tensorflow/compiler/xla/service:tuple_points_to_analysis",
"//tensorflow/core:lib",
],
)

41
tensorflow/compiler/xla/service/cpu/cpu_parallelization_preparation.cc
View File

@ -21,8 +21,6 @@ limitations under the License.
#include "tensorflow/compiler/xla/service/hlo_computation.h"
#include "tensorflow/compiler/xla/service/hlo_instruction.h"
#include "tensorflow/compiler/xla/service/hlo_opcode.h"
#include "tensorflow/compiler/xla/service/logical_buffer.h"
#include "tensorflow/compiler/xla/service/tuple_points_to_analysis.h"
#include "tensorflow/compiler/xla/types.h"
#include "tensorflow/compiler/xla/util.h"
#include "tensorflow/core/lib/strings/strcat.h"
@ -64,13 +62,11 @@ StatusOr<bool> ParallelizationPreparation::Run(HloModule* module) {
instructions_to_outline.clear();
HloInstruction* outline_candidate = instruction;
instructions_to_outline.push_back(outline_candidate);
bool all_bitcasts = outline_candidate->opcode() == HloOpcode::kBitcast;
// Outline sole users with the current instruction.
while (CanOutlineWithUser(outline_candidate)) {
HloInstruction* prior_candidate = outline_candidate;
outline_candidate = *outline_candidate->users().begin();
all_bitcasts &= outline_candidate->opcode() == HloOpcode::kBitcast;
if (std::any_of(outline_candidate->operands().begin(),
outline_candidate->operands().end(),
[&](const HloInstruction* operand) {
@ -86,24 +82,6 @@ StatusOr<bool> ParallelizationPreparation::Run(HloModule* module) {
}
instructions_to_outline.push_back(outline_candidate);
}
// If all instructions in the outline candidates are a bitcast, then create
// a copy at the head of the bitcasts and include it in the outlined
// instructions. The underlying problem is that a computation which forwards
// a parameter buffer to the output is not properly handled by the backends
// or analysis.
//
// This would be better handled by being smarter about choosing outline
// candidates in the first place.
if (all_bitcasts) {
// 'head' is the first instruction in the chain of bitcasts.
HloInstruction* head = instructions_to_outline[0];
HloInstruction* head_operand = head->mutable_operand(0);
HloInstruction* copy =
entry_computation->AddInstruction(HloInstruction::CreateUnary(
head_operand->shape(), HloOpcode::kCopy, head_operand));
TF_RETURN_IF_ERROR(head->ReplaceOperandWith(0, copy));
instructions_to_outline.insert(instructions_to_outline.begin(), copy);
}
outlined.insert(instructions_to_outline.begin(),
instructions_to_outline.end());
@ -122,25 +100,6 @@ StatusOr<bool> ParallelizationPreparation::Run(HloModule* module) {
changed = true;
}
TF_ASSIGN_OR_RETURN(auto points_to_analysis,
TuplePointsToAnalysis::Run(module));
for (auto& computation : module->computations()) {
if (computation->IsFusionComputation()) {
continue;
}
HloInstruction* root = computation->root_instruction();
// Copy root instruction if it does not define its own top-level buffer.
// TODO(b/32885001) Remove these copies (at least for the unambiguous case).
// TODO(b/32885001) Perform shallow copy if root value is a tuple.
if (!points_to_analysis->InstructionDefinesBufferAtIndex(root,
/*index=*/{})) {
HloInstruction* copy = computation->AddInstruction(
HloInstruction::CreateUnary(root->shape(), HloOpcode::kCopy, root));
computation->set_root_instruction(copy);
changed = true;
}
}
XLA_VLOG_LINES(2, "ParallelizationPreparation EXIT");
XLA_VLOG_LINES(2, module->ToString());
return changed;