fix bug when add atomic clean for getnext's output

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

@@ -234,10 +234,15 @@ std::map<AnfNodePtr, std::vector<size_t>> GetCommunicationOpInputInfo(
       for (size_t i = 0; i < input_num; i++) {
         auto input_node = kernel->input(i + 1);
         auto kernel_input = AnfAlgo::VisitKernelWithReturnType(input_node, 0, true);
-        if (AnfAlgo::IsCommunicationOp(kernel_input.first)) {
-          // no need to add atomic for communication op's output
-          MS_LOG(INFO) << "No need to add atomic clean for communication op "
-                       << kernel_input.first->fullname_with_scope() << "'s output";
+        if (!kernel_input.first->isa<CNode>()) {
+          continue;
+        }
+        auto cnode = kernel_input.first->cast<CNodePtr>();
+        if (AnfAlgo::IsCommunicationOp(cnode) || AnfAlgo::IsIndependentNode(cnode) ||
+            AnfAlgo::GetCNodeName(cnode) == kGetNextOpName) {
+          // no need to add atomic for communication/independent/getnext op 's output
+          MS_LOG(INFO) << "No need to add atomic clean for op " << kernel_input.first->fullname_with_scope()
+                       << "'s output";
           continue;
         }
         MS_LOG(INFO) << "Add atomic clean for single communication op input, comm:" << kernel->fullname_with_scope()

mindspore/ccsrc/runtime/device/kernel_adjust.cc

@@ -280,6 +280,12 @@ void KernelAdjust::InsertSwitchLoop(const std::shared_ptr<session::KernelGraph>
   for (size_t idx = i + 1; idx < orders.size(); idx++) {
     cur_cnode = orders[idx];
     if (AnfAlgo::HasNodeAttr(kAttrLabelForInsertStreamActive, cur_cnode)) {
+      auto pre_node = orders[idx - 1];
+      auto pre_kernel_name = AnfAlgo::GetCNodeName(pre_node);
+      if (pre_kernel_name == kAtomicAddrCleanOpName) {
+        other_list.pop_back();
+        memcpy_list.push_back(pre_node);
+      }
       memcpy_list.emplace_back(cur_cnode);
     } else {
       other_list.emplace_back(cur_cnode);