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

Fill the runner field in FLR options with device private runner.

Browse Source 
When running multi-device eager functions, pass the device-private runner (if available) to launch internal ops. It makes sure that GPU kernel launching ops are executed in GPU private thread pool.

PiperOrigin-RevId: 239103306
This commit is contained in:
Haoyu Zhang 2019-03-18 18:19:37 -07:00 committed by TensorFlower Gardener
parent 6e9cb400d1
commit e4df992d2e
1 changed files with 6 additions and 1 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

7
tensorflow/core/common_runtime/process_function_library_runtime.cc
View File

@ -790,11 +790,16 @@ void ProcessFunctionLibraryRuntime::RunMultiDevice(
opts_copy.args_alloc_attrs = comp_data.arg_alloc_attrs_;
opts_copy.rets_alloc_attrs = comp_data.ret_alloc_attrs_;
opts_copy.remote_execution = false;
FunctionLibraryRuntime* flr = GetFLR(target);
// When target device has private thread pool, use the target device runner
thread::ThreadPool* pool = flr->device()->tensorflow_device_thread_pool();
opts_copy.runner = (pool == nullptr) ? opts_copy.runner : flr->runner();
std::vector<Tensor> comp_args =
GetArgsForIndices(comp_data.arg_indices_, args);
std::vector<Tensor>* comp_rets = new std::vector<Tensor>;
rets->resize(data->num_outputs_);
GetFLR(target)->Run(
flr->Run(
opts_copy, handle, comp_args, comp_rets,
[comp_rets, rets, comp_data, refcounted_done](const Status& status) {
if (!status.ok()) {