Performance microtweaks: Pass by reference rather than by value; pre-reserve capacity when total vectoroid size is known.

PiperOrigin-RevId: 201172723

4 files changed, 10 insertions, 6 deletions

diff --git a/tensorflow/compiler/xla/service/gpu/ir_emitter_nested.cc b/tensorflow/compiler/xla/service/gpu/ir_emitter_nested.cc
index bb47a42805..c9574c87a3 100644
--- a/tensorflow/compiler/xla/service/gpu/ir_emitter_nested.cc
+++ b/tensorflow/compiler/xla/service/gpu/ir_emitter_nested.cc
@@ -120,9 +120,10 @@ Status IrEmitterNested::EmitTargetElementLoop(
   // For MOF we give the loop emitter an array for every output it should
   // generate.
   if (hlo.IsMultiOutputFusion()) {
+    const int64 num_elems = ShapeUtil::TupleElementCount(hlo.shape());
     std::vector<llvm_ir::IrArray> target_arrays;
-    for (int64 i = 0, e = ShapeUtil::TupleElementCount(hlo.shape()); i != e;
-         ++i) {
+    target_arrays.reserve(num_elems);
+    for (int64 i = 0; i != num_elems; ++i) {
       target_arrays.push_back(GetIrArray(hlo, hlo, {i}));
     }
     TF_RETURN_IF_ERROR(
@@ -130,6 +131,7 @@ Status IrEmitterNested::EmitTargetElementLoop(
             .EmitLoop());
 
     std::vector<llvm::Value*> tuple_operand_ptrs;
+    tuple_operand_ptrs.reserve(num_elems);
     for (const llvm_ir::IrArray& array : target_arrays) {
       tuple_operand_ptrs.push_back(array.GetBasePointer());
     }
diff --git a/tensorflow/compiler/xla/service/hlo_query.cc b/tensorflow/compiler/xla/service/hlo_query.cc
index d45038f1f4..2418c19f3d 100644
--- a/tensorflow/compiler/xla/service/hlo_query.cc
+++ b/tensorflow/compiler/xla/service/hlo_query.cc
@@ -61,7 +61,7 @@ bool AllOperandsAreConstants(const HloInstruction& instruction) {
 }
 
 HloInstruction* GetMatchingOperand(
-    std::function<bool(const HloInstruction*)> matcher,
+    const std::function<bool(const HloInstruction*)>& matcher,
     HloInstruction* instruction) {
   for (HloInstruction* op : instruction->operands()) {
     if (matcher(op)) {
@@ -72,7 +72,7 @@ HloInstruction* GetMatchingOperand(
 }
 
 bool MatchBinaryInstructionOperand(
-    std::function<bool(const HloInstruction*)> matcher,
+    const std::function<bool(const HloInstruction*)>& matcher,
     HloInstruction* instruction, HloInstruction** matching_operand,
     HloInstruction** other_operand) {
   CHECK_EQ(instruction->operand_count(), 2);
diff --git a/tensorflow/compiler/xla/service/hlo_query.h b/tensorflow/compiler/xla/service/hlo_query.h
index c79347bbf9..c0826a6aee 100644
--- a/tensorflow/compiler/xla/service/hlo_query.h
+++ b/tensorflow/compiler/xla/service/hlo_query.h
@@ -45,7 +45,7 @@ bool IsScalarConstant(const HloInstruction* instruction);
 // multiple matching operands, then the first matching operand is returned. If
 // there are no matching operands then nullptr is returned.
 HloInstruction* GetMatchingOperand(
-    std::function<bool(const HloInstruction*)> matcher,
+    const std::function<bool(const HloInstruction*)>& matcher,
     HloInstruction* instruction);
 
 // Returns whether a binary instruction has a matching operand. Sets
@@ -53,7 +53,7 @@ HloInstruction* GetMatchingOperand(
 // other_operand. Note: in the case where both operands match, the first operand
 // of the instruction is returned.
 bool MatchBinaryInstructionOperand(
-    std::function<bool(const HloInstruction*)> matcher,
+    const std::function<bool(const HloInstruction*)>& matcher,
     HloInstruction* instruction, HloInstruction** matching_operand,
     HloInstruction** other_operand);
 
diff --git a/tensorflow/compiler/xla/service/shape_inference.cc b/tensorflow/compiler/xla/service/shape_inference.cc
index e25f5e67c7..4606d8f202 100644
--- a/tensorflow/compiler/xla/service/shape_inference.cc
+++ b/tensorflow/compiler/xla/service/shape_inference.cc
@@ -939,6 +939,7 @@ ShapeInference::InferDegenerateDimensionBroadcastShape(HloOpcode operation,
     HloOpcode opcode,
     tensorflow::gtl::ArraySlice<const HloInstruction*> operands) {
   std::vector<const Shape*> operand_shapes;
+  operand_shapes.reserve(operands.size());
   for (const HloInstruction* operand : operands) {
     operand_shapes.push_back(&operand->shape());
   }
@@ -954,6 +955,7 @@ ShapeInference::InferDegenerateDimensionBroadcastShape(HloOpcode operation,
   switch (opcode) {
     case HloOpcode::kTuple: {
       Shape result = ShapeUtil::MakeTupleShape({});
+      result.mutable_tuple_shapes()->Reserve(operand_shapes.size());
       for (const Shape* shape : operand_shapes) {
         ShapeUtil::AppendShapeToTuple(*shape, &result);
       }