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

Use a reader lock when reading the TfToPlatformGpuIdMap.

Browse Source 
We perform a look up in this map every time a kernel is executed on a GPU. A contention profile shows non-trivial contention on the current exclusive lock, which can be avoided by switching to a reader lock.

PiperOrigin-RevId: 260820700
This commit is contained in:
Derek Murray 2019-07-30 16:32:09 -07:00 committed by TensorFlower Gardener
parent 890f21cb3f
commit 25a15dda8e
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/gpu/gpu_id_manager.cc
View File

@ -59,7 +59,9 @@ class TfToPlatformGpuIdMap {
bool Find(TfGpuId tf_gpu_id, PlatformGpuId* platform_gpu_id) const
LOCKS_EXCLUDED(mu_) {
mutex_lock lock(mu_);
// TODO(mrry): Consider replacing this with an atomic `is_initialized` bit,
// to avoid writing to a shared cache line in the tf_shared_lock.
tf_shared_lock lock(mu_);
auto result = id_map_.find(tf_gpu_id.value());
if (result == id_map_.end()) return false;
*platform_gpu_id = result->second;