subgraph optimize multi thread env

ge/graph/manager/graph_manager.cc

@@ -533,11 +533,19 @@ Status GraphManager::CopySubGraphAndMarkFusion(const ComputeGraphPtr &compute_gr
   return SUCCESS;
 }
 
-Status GraphManager::OptimizeSubGraphWithMultiThreads(ComputeGraphPtr compute_graph, 
+Status GraphManager::OptimizeSubGraphWithMultiThreads(ComputeGraphPtr compute_graph,
                                                       Graph2SubGraphInfoList &sub_graph_map, uint64_t session_id) {
   GE_CHECK_NOTNULL(compute_graph);
   // use default 16 multi thread
-  const uint32_t thread_num = 16;
+  uint32_t thread_num = 16;
+
+  const char *env = std::getenv("THREAD_MULTI_NUM");
+  if (env != nullptr) {
+    thread_num = atoi(env);
+    GEEVENT("OptimizeSubGraphWithMultiThreads thread num: %u", thread_num);
+  }
+
+
   ThreadPool executor(thread_num);
   std::vector<std::future<Status>> vector_future;
   const auto &root_subgraph_list = sub_graph_map[compute_graph];
@@ -2111,6 +2119,8 @@ Status GraphManager::OptimizeStage1(ge::ComputeGraphPtr &compute_graph) {
                                                new (std::nothrow) TransOpWithoutReshapeFusionPass))
   GE_CHK_STATUS_RET(after_merge_passes.AddPass("OptimizeStage1_1::TransOpBreadthFusionPass",
                                                new (std::nothrow) TransOpBreadthFusionPass))
+  GE_CHK_STATUS_RET(
+      after_merge_passes.AddPass("OptimizeStage1_1::HcclMemcpyPass", new (std::nothrow) HcclMemcpyPass));
 
   GE_TIMESTAMP_START(after_merge_passes);
   auto ret = after_merge_passes.Run(compute_graph);