From dbc6147e9b5ff599f6f35b9299674fe0d7e7bd82 Mon Sep 17 00:00:00 2001
From: laiyongqiang <laiyongqiang1@huawei.com>
Date: Wed, 28 Oct 2020 09:37:09 +0800
Subject: [PATCH] fix bug when add atomic clean for getnext's output

---
 .../runtime/device/ascend/kernel_build_ascend.cc    | 13 +++++++++----
 mindspore/ccsrc/runtime/device/kernel_adjust.cc     |  6 ++++++
 2 files changed, 15 insertions(+), 4 deletions(-)

diff --git a/mindspore/ccsrc/runtime/device/ascend/kernel_build_ascend.cc b/mindspore/ccsrc/runtime/device/ascend/kernel_build_ascend.cc
index 01b0114909..7cb248bc43 100644
--- a/mindspore/ccsrc/runtime/device/ascend/kernel_build_ascend.cc
+++ b/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()
diff --git a/mindspore/ccsrc/runtime/device/kernel_adjust.cc b/mindspore/ccsrc/runtime/device/kernel_adjust.cc
index a77089314b..64bc184252 100644
--- a/mindspore/ccsrc/runtime/device/kernel_adjust.cc
+++ b/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);