Speed up `OpKernel::InputRange()` by ~2.6x.

This reduces the typical amount of time spent converting an argument
name into a range of positional arguments from ~59.5ns to ~22.5ns. To
achieve this, the `NameRangeMap` is converted to borrow the argument
names (as `tensorflow::StringPiece` objects) from the appropriate
`OpDef`, instead of storing a string. This avoids allocating a string
for each lookup. The map data structure is also changed from a
`std::unordered_map` to a `gtl::FlatMap`.

PiperOrigin-RevId: 180588460

1 files changed, 2 insertions, 2 deletions

diff --git a/tensorflow/core/framework/op_kernel.cc b/tensorflow/core/framework/op_kernel.cc
index 4d410809e7..433005c8ab 100644
--- a/tensorflow/core/framework/op_kernel.cc
+++ b/tensorflow/core/framework/op_kernel.cc
@@ -112,7 +112,7 @@ const string& OpKernel::requested_input(int i) const { return def_->input(i); }
 
 Status OpKernel::InputRange(StringPiece input_name, int* start,
                             int* stop) const {
-  const auto result = input_name_map_.find(input_name.ToString());
+  const auto result = input_name_map_.find(input_name);
   if (result == input_name_map_.end()) {
     return errors::InvalidArgument("Unknown input name: ", input_name);
   } else {
@@ -124,7 +124,7 @@ Status OpKernel::InputRange(StringPiece input_name, int* start,
 
 Status OpKernel::OutputRange(StringPiece output_name, int* start,
                              int* stop) const {
-  const auto result = output_name_map_.find(output_name.ToString());
+  const auto result = output_name_map_.find(output_name);
   if (result == output_name_map_.end()) {
     return errors::InvalidArgument("Unknown output name: ", output_name);
   } else {