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

Correctly handle S32 and U32 in the AR-CRS combiner

Browse Source 
Some of the transforms we are doing are only valid on floating point types
so we have to condition them on the element type of the operations.

PiperOrigin-RevId: 227836047
This commit is contained in:
A. Unique TensorFlower 2019-01-04 04:10:23 -08:00 committed by TensorFlower Gardener
parent 6965d80c13
commit 8e4ca33c8b
2 changed files with 20 additions and 5 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

1
tensorflow/compiler/xla/service/BUILD
View File

@ -3678,6 +3678,7 @@ cc_library(
":pattern_matcher",
"//tensorflow/compiler/xla:literal",
"//tensorflow/compiler/xla:literal_util",
"//tensorflow/compiler/xla:shape_util",
"//tensorflow/compiler/xla:status_macros",
"//tensorflow/compiler/xla:statusor",
"//tensorflow/compiler/xla:types",

24
tensorflow/compiler/xla/service/ar_crs_combiner.cc
View File

@ -26,6 +26,7 @@ limitations under the License.
#include "tensorflow/compiler/xla/service/hlo_instruction.h"
#include "tensorflow/compiler/xla/service/hlo_opcode.h"
#include "tensorflow/compiler/xla/service/pattern_matcher.h"
#include "tensorflow/compiler/xla/shape_util.h"
#include "tensorflow/compiler/xla/status_macros.h"
#include "tensorflow/compiler/xla/types.h"
#include "tensorflow/compiler/xla/xla_data.pb.h"
@ -44,11 +45,24 @@ bool MatchesArCrsPattern(HloInstruction* instruction) {
if (instruction->user_count() != 1) {
return false;
}
auto opcode = instruction->opcode();
return opcode == HloOpcode::kBitcast || opcode == HloOpcode::kTranspose ||
opcode == HloOpcode::kReshape || opcode == HloOpcode::kConvert ||
opcode == HloOpcode::kAdd || opcode == HloOpcode::kSubtract ||
opcode == HloOpcode::kMultiply;
switch (instruction->opcode()) {
case HloOpcode::kBitcast:
case HloOpcode::kTranspose:
case HloOpcode::kReshape:
return true;
case HloOpcode::kConvert:
// Can be moved across if both input and output is either float or
// integer (e.g. S32<->U32 or F32<->BF16)
return ShapeUtil::ElementIsFloating(instruction->shape()) ==
ShapeUtil::ElementIsFloating(instruction->operand(0)->shape());
case HloOpcode::kAdd:
case HloOpcode::kSubtract:
case HloOpcode::kMultiply:
// Only supported for floating point operands.
return ShapeUtil::ElementIsFloating(instruction->shape());
default:
return false;
}
};
auto computation_is_addition = [](HloComputation* c) {