Wait VarDummyHandle generated

paddle/fluid/framework/details/computation_op_handle.cc

@@ -35,7 +35,7 @@ void ComputationOpHandle::RunImpl() {
 
 bool ComputationOpHandle::NeedWait(VarHandleBase *in_var) {
   bool need_wait =
-      dynamic_cast<VarHandle *>(in_var) && in_var->generated_op_ &&
+      in_var && in_var->generated_op_ &&
       in_var->generated_op_->DeviceContext(place_) != dev_ctxes_[place_];
   return need_wait;
 }

paddle/fluid/framework/details/op_handle_base.cc

@@ -105,7 +105,7 @@ void OpHandleBase::WaitInputVarGenerated(const platform::Place &place) {
 }
 
 bool OpHandleBase::NeedWait(VarHandleBase *in_var) {
-  return dynamic_cast<VarHandle *>(in_var) && in_var->generated_op_;
+  return in_var && in_var->generated_op_;
 }
 
 void OpHandleBase::RunAndRecordEvent(const std::function<void()> &callback) {

paddle/fluid/framework/details/send_op_handle.cc

@@ -26,8 +26,17 @@ SendOpHandle::SendOpHandle(const framework::OpDesc &op_desc,
       place_(place) {}
 
 void SendOpHandle::RunImpl() {
+  // TODO(wuyi): need further analysis whether wait VarDummyHandle.
   // Wait input done
-  WaitInputVarGenerated(place_);
+  for (auto *in : inputs_) {
+    auto &p = static_cast<VarHandle *>(in)->place_;
+    if (in->DebugString() == "dummy") {  // HACK
+      continue;
+    }
+    if (in->generated_op_) {
+      in->generated_op_->RecordWaitEventOnCtx(dev_ctxes_[p]);
+    }
+  }
   auto &tmp_scope = local_scope_->FindVar(kLocalExecScopeName)->Get<Scope *>();
   // FIXME(wuyi): can not use RunAndRecordEvent here, for it will cause dead
   // lock.