fix reset8p pynative performance

mindspore/ccsrc/backend/session/ascend_session.cc

@@ -1308,6 +1308,8 @@ void AscendSession::RunOpMemoryAllocNew(const std::vector<tensor::TensorPtr> &in
 
 void AscendSession::RunOpGenKernelEvent(const KernelGraph *graph) const {
   MS_EXCEPTION_IF_NULL(graph);
+  auto kernels = graph->execution_order();
+  device::ascend::AscendStreamAssign::GetInstance().AssignStreamForNonTaskSink(kernels);
   auto runtime_instance = device::KernelRuntimeManager::Instance().GetKernelRuntime(kAscendDevice, device_id_);
   MS_EXCEPTION_IF_NULL(runtime_instance);
   runtime_instance->GenKernelEvents(*graph);

mindspore/ccsrc/runtime/device/ascend/ascend_stream_assign.cc

@@ -214,8 +214,7 @@ void SetNodeStreamIDAttr(const NotNull<KernelGraphPtr> &graph_ptr) {
 }
 }  // namespace
 
-void AscendStreamAssign::AssignStreamForNonTaskSink(const NotNull<KernelGraphPtr> &graph_ptr) {
-  auto &kernels = graph_ptr->execution_order();
+void AscendStreamAssign::AssignStreamForNonTaskSink(const std::vector<CNodePtr> &kernels) {
   if (kernels.empty()) {
     return;
   }
@@ -255,7 +254,8 @@ void AscendStreamAssign::AssignStreamForNonTaskSink(const NotNull<KernelGraphPtr
 
 void AscendStreamAssign::AssignStream(const NotNull<KernelGraphPtr> &graph_ptr) {
   if (!IsTaskSink()) {
-    AssignStreamForNonTaskSink(graph_ptr);
+    auto kernels = graph_ptr->execution_order();
+    AssignStreamForNonTaskSink(kernels);
     return;
   }
   if (!graph_ptr->is_dynamic_shape()) {

mindspore/ccsrc/runtime/device/ascend/ascend_stream_assign.h

@@ -59,6 +59,7 @@ class AscendStreamAssign {
   void AssignStream(const NotNull<KernelGraphPtr> &graph_ptr);
   void GetHcomStreams(std::vector<uint32_t> *streams);
   void GetWaitStreams(vector<uint32_t> *wait_active_stream_list);
+  void AssignStreamForNonTaskSink(const std::vector<CNodePtr> &kernels);
   const std::vector<std::vector<uint32_t>> &get_stream_group() const { return stream_groups_; }
   const std::map<CNodePtr, CNodePtr> &get_event_map() const { return event_map_; }
 
@@ -68,7 +69,6 @@ class AscendStreamAssign {
   void Reset();
   CNodePtr CreateSendApplyKernel(const NotNull<KernelGraphPtr> &graph_ptr, uint32_t event_id, uint32_t stream_id);
   CNodePtr CreateRecvApplyKernel(const NotNull<KernelGraphPtr> &graph_ptr, uint32_t event_id, uint32_t stream_id);
-  void AssignStreamForNonTaskSink(const NotNull<KernelGraphPtr> &graph_ptr);
   void CheckResourceAssign(const NotNull<KernelGraphPtr> &graph_ptr);
   void CheckStreamAssign(const NotNull<KernelGraphPtr> &graph_ptr);
   void CheckEventAssign(const NotNull<KernelGraphPtr> &graph_ptr);

tests/ut/cpp/stub/tasksink/ascend_stream_assign_stub.cc

@@ -30,6 +30,8 @@ void AscendStreamAssign::AssignStream(const NotNull<KernelGraphPtr> &graph_ptr)
 void AscendStreamAssign::GetWaitStreams(vector<uint32_t> *wait_active_stream_list) { return; }
 
 void AscendStreamAssign::GetHcomStreams(std::vector<uint32_t> *streams) { return; }
+
+void AscendStreamAssign::AssignStreamForNonTaskSink(const std::vector<CNodePtr> &kernels) { return; }
 }  // namespace ascend
 
 void KernelAdjust::InsertDeviceLoopCtrl(const std::shared_ptr<session::KernelGraph> &kernel_graph_ptr) { return; }