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

[Executor] Fix measurement error in the microbenchmark.

Browse Source 
Previously, we were measuring the time taken to build the input graph, which
dominated the cost of actually executing the executor code to be benchmarked.

PiperOrigin-RevId: 301168024
Change-Id: I6a9fa8d6baa6a9afd670c97ae5bf34799aebd2d4
This commit is contained in:
Derek Murray 2020-03-16 08:58:23 -07:00 committed by TensorFlower Gardener
parent 14a46de40c
commit c3dc660512
1 changed files with 4 additions and 0 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

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

@ -416,6 +416,7 @@ TEST_F(ExecutorTest, RecvInvalidRefDtype) {
// maximum of 'width' nodes. All nodes are no-ops and all dependencies are
// control dependencies.
static void BM_executor(int iters, int width, int depth) {
testing::StopTiming();
#ifdef PLATFORM_GOOGLE
BenchmarkUseRealTime();
#endif // PLATFORM_GOOGLE
@ -451,6 +452,7 @@ static void BM_executor(int iters, int width, int depth) {
SetBenchmarkLabel(strings::StrCat("Nodes = ", cur));
SetBenchmarkItemsProcessed(cur * static_cast<int64>(iters));
#endif // PLATFORM_GOOGLE
testing::StartTiming();
test::Benchmark("cpu", g).Run(iters);
}
@ -466,6 +468,7 @@ BENCHMARK(BM_executor)->ArgPair(8192, 32);
BENCHMARK(BM_executor)->ArgPair(1024, 1024);
static void BM_FeedInputFetchOutput(int iters) {
testing::StopTiming();
Graph* g = new Graph(OpRegistry::Global());
// z = x + y: x and y are provided as benchmark inputs. z is the
// output of the benchmark. Conceptually, the caller is ALICE, the
@ -484,6 +487,7 @@ static void BM_FeedInputFetchOutput(int iters) {
#ifdef PLATFORM_GOOGLE
SetBenchmarkItemsProcessed(static_cast<int64>(iters));
#endif // PLATFORM_GOOGLE
testing::StartTiming();
test::Benchmark("cpu", g).RunWithRendezvousArgs({{x_key, val}, {y_key, val}},
{z_key}, iters);
}