aboutsummaryrefslogtreecommitdiffhomepage
path: root/tensorflow/compiler/xla/service/gpu/ir_emitter_unnested.cc
diff options
context:
space:
mode:
authorGravatar Justin Lebar <jlebar@google.com>2018-07-10 13:50:19 -0700
committerGravatar TensorFlower Gardener <gardener@tensorflow.org>2018-07-10 13:58:27 -0700
commite7fb6f5f0e1745f6140e0add84e931e4786e8269 (patch)
tree3dc56b830586332f98bd918f162979b61ab265a9 /tensorflow/compiler/xla/service/gpu/ir_emitter_unnested.cc
parentc4ac9855539dc228881707c69a9ef2fe703dadd4 (diff)
[XLA:GPU] Delete AnnotateBufferLoadStoreInstructionWithMetadata.
This function was trying to annotate loads/stores to shared memory with metadata copied from another IRArray. But it was over-eager and copied more than we wanted. In addition, it CHECKed that the source buffer was invariant, and that's not necessarily true. For now, remove this function, which is just a performance optimization. PiperOrigin-RevId: 204005618
Diffstat (limited to 'tensorflow/compiler/xla/service/gpu/ir_emitter_unnested.cc')
-rw-r--r--tensorflow/compiler/xla/service/gpu/ir_emitter_unnested.cc6
1 files changed, 1 insertions, 5 deletions
diff --git a/tensorflow/compiler/xla/service/gpu/ir_emitter_unnested.cc b/tensorflow/compiler/xla/service/gpu/ir_emitter_unnested.cc
index 186672047b..1002963dc2 100644
--- a/tensorflow/compiler/xla/service/gpu/ir_emitter_unnested.cc
+++ b/tensorflow/compiler/xla/service/gpu/ir_emitter_unnested.cc
@@ -3040,12 +3040,10 @@ LaunchDimensions IrEmitterUnnested::EmitHlo021Tile(
llvm_ir::IrArray& input_in_logical_shape =
param_in_reduced_shape_arrays[id];
llvm::Value* buffer = param_buffers[id];
- llvm::Instruction* store_to_buffer = ir_builder_.CreateStore(
+ ir_builder_.CreateStore(
input_in_logical_shape.EmitReadArrayElement(index, &ir_builder_,
"input_element"),
ir_builder_.CreateGEP(buffer, {index_typed_const(0), y_loc, x}));
- param_arrays[id].AnnotateBufferLoadStoreInstructionWithMetadata(
- store_to_buffer);
}
};
@@ -3073,8 +3071,6 @@ LaunchDimensions IrEmitterUnnested::EmitHlo021Tile(
ir_builder_.CreateGEP(param_buffers[0],
{ir_builder_.getInt64(0), x, y_loc}),
"output_element");
- param_arrays[0].AnnotateBufferLoadStoreInstructionWithMetadata(
- load_from_buffer);
output_in_reduced_shape_arrays[0].EmitWriteArrayElement(
index, load_from_buffer, &ir_builder_);
} else {