Merge pull request #10126 from typhoonzero/fix_para_exe_dist_train

typhoonzero · web-flow · commit 3863c6a9a8b0 · 2018-04-23T15:27:07.000+08:00
Fix para exe dist train
diff --git a/paddle/fluid/framework/details/send_op_handle.cc b/paddle/fluid/framework/details/send_op_handle.cc
@@ -34,7 +34,10 @@ void SendOpHandle::RunImpl() {
     }
     in->generated_op_->Wait(dev_ctxes_[p]);
   }
-  this->RunAndRecordEvent([&] { op_->Run(*local_scope_, place_); });
+  auto &tmp_scope = local_scope_->FindVar(kLocalExecScopeName)->Get<Scope *>();
+  // FIXME(wuyi): can not use RunAndRecordEvent here, for it will cause dead
+  // lock.
+  op_->Run(*tmp_scope, place_);
 }
 
 std::string SendOpHandle::Name() const { return "send"; }

Original file line number	Diff line number	Diff line change
`@@ -34,7 +34,10 @@ void SendOpHandle::RunImpl() {`
`34`	`34`	`}`
`35`	`35`	`in->generated_op_->Wait(dev_ctxes_[p]);`
`36`	`36`	`}`
`37`		`- this->RunAndRecordEvent([&] { op_->Run(*local_scope_, place_); });`
	`37`	`+ auto &tmp_scope = local_scope_->FindVar(kLocalExecScopeName)->Get<Scope *>();`
	`38`	`+ // FIXME(wuyi): can not use RunAndRecordEvent here, for it will cause dead`
	`39`	`+ // lock.`
	`40`	`+ op_->Run(*tmp_scope, place_);`
`38`	`41`	`}`
`39`	`42`
`40`	`43`	`std::string SendOpHandle::Name() const { return "send"; }`