Merge pull request #14395 from sneaxiy/fix_num_threads_in_fast_pe

Fix num_threads settings in fast_pe

paddle/fluid/framework/details/fast_threaded_ssa_graph_executor.cc

@@ -30,8 +30,8 @@ FastThreadedSSAGraphExecutor::FastThreadedSSAGraphExecutor(
       local_scopes_(local_scopes),
       places_(places),
       graph_(std::move(graph)),
-      pool_(strategy.num_threads_ +
-            1),  // add one more thread for generate op_deps
+      pool_(strategy.num_threads_),
+      prepare_pool_(1),  // add one more thread for generate op_deps
       fetch_ctxs_(places) {
   for (auto &op : ir::FilterByNodeWrapper<OpHandleBase>(*graph_)) {
     int dep = static_cast<int>(op->NotReadyInputSize());
@@ -160,7 +160,7 @@ void FastThreadedSSAGraphExecutor::RunOpAsync(
   });
 }
 void FastThreadedSSAGraphExecutor::PrepareAtomicOpDeps() {
-  atomic_op_deps_ = pool_.enqueue([&] {
+  atomic_op_deps_ = prepare_pool_.enqueue([&] {
     auto *op_deps = new std::unordered_map<OpHandleBase *, std::atomic<int>>;
     for (auto &pair : op_deps_) {
       (*op_deps)[pair.first] = pair.second;

paddle/fluid/framework/details/fast_threaded_ssa_graph_executor.h

@@ -46,6 +46,7 @@ class FastThreadedSSAGraphExecutor : public SSAGraphExecutor {
   std::vector<OpHandleBase *> bootstrap_ops_;
 
   ::ThreadPool pool_;
+  ::ThreadPool prepare_pool_;
   platform::DeviceContextPool fetch_ctxs_;
   std::atomic<int> remaining_;