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

[Executor tracing] Add num_output_edges to the PropagateOutputs TraceMe.

Browse Source 
This enables someone reading a trace to distinguish between long PropagateOutputs events that are due to contention, and long events that are due to a lot of propagation work going on.

PiperOrigin-RevId: 296072451
Change-Id: Ic7b731f3ccb2b0ba12e0e7d23f31e89cef9b0a97
This commit is contained in:
Derek Murray 2020-02-19 15:32:49 -08:00 committed by TensorFlower Gardener
parent 120b0f57f0
commit 0546460a7d
1 changed files with 3 additions and 1 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

4
tensorflow/core/common_runtime/executor.cc
View File

@ -2161,7 +2161,9 @@ void ExecutorState::PropagateOutputs(const TaggedNode& tagged_node,
profiler::TraceMe activity( profiler::TraceMe activity(
[&]() { [&]() {
return strings::StrCat("ExecutorPropagateOutputs#", "id=", step_id_, return strings::StrCat("ExecutorPropagateOutputs#", "id=", step_id_,
",kernel_name=", item->kernel->name_view(), "#"); ",kernel_name=", item->kernel->name_view(),
",num_output_edges=", item->num_output_edges,
"#");
}, },
profiler::GetTFTraceMeLevel(/*is_expensive=*/false)); profiler::GetTFTraceMeLevel(/*is_expensive=*/false));