diff options
| author |  Derek Murray <mrry@google.com> | 2018-01-02 13:53:37 -0800 |
|---|---|---|
| committer |  TensorFlower Gardener <gardener@tensorflow.org> | 2018-01-02 13:57:52 -0800 |
| commit | 844e0c1984c626c440246b707f958953e0f6ef49 (patch) | |
| tree | 142f4f882cdb04299f729bc91182bdb9b7009ca4 /tensorflow/c/c_api_function.cc | |
| parent | 4c81792ebebc9c1dbd3bcd13073f85f3c8688ca5 (diff) | |
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
Diffstat (limited to 'tensorflow/c/c_api_function.cc')
| -rw-r--r-- | tensorflow/c/c_api_function.cc | 2 |
1 files changed, 1 insertions, 1 deletions
diff --git a/tensorflow/c/c_api_function.cc b/tensorflow/c/c_api_function.cc index d60d1de315..dd6119e8aa 100644 --- a/tensorflow/c/c_api_function.cc +++ b/tensorflow/c/c_api_function.cc @@ -312,7 +312,7 @@ Status GraphToFunctionDef(const Graph& fn_body, const string& fn_name, TF_RETURN_IF_ERROR( NameRangesForNode(*node, node->op_def(), nullptr, &output_ranges)); for (const auto& output : output_ranges) { - const string& output_name = output.first; + const StringPiece& output_name = output.first; int index_start = output.second.first; int index_end = output.second.second; for (int i = index_start; i < index_end; ++i) { |