optimize time cost for Shrink thread

ge/graph/load/model_manager/davinci_model.cc

@@ -173,6 +173,7 @@ DavinciModel::DavinciModel(int32_t priority, const std::shared_ptr<ModelListener
       runtime_model_id_(0),
       version_(0),
       ge_model_(nullptr),
+      is_shrink_created_(false),
       listener_(listener),
       run_flg_(false),
       priority_(priority),
@@ -765,7 +766,6 @@ Status DavinciModel::Init(void *dev_ptr, size_t mem_size, void *weight_ptr, size
     }
   }
 
-  CREATE_STD_THREAD(shrink_id_, &DavinciModel::Shrink, this);
   return SUCCESS;
 }
 
@@ -3708,6 +3708,10 @@ Status DavinciModel::NnExecute(rtStream_t stream, bool async_mode, const InputDa
   GELOGD("current_data.index=%u", input_data.index);
   GE_IF_BOOL_EXEC(ProfilingManager::Instance().ProfilingModelExecuteOn(), SetProfileTime(MODEL_PRE_PROC_END));
 
+  if (!is_shrink_created_) {
+    CREATE_STD_THREAD(shrink_id_, &DavinciModel::Shrink, this);
+    is_shrink_created_ = true;
+  }
   if (!task_list_.empty()) {
     GELOGD("rtModelExecute do");
     GE_IF_BOOL_EXEC(ProfilingManager::Instance().ProfilingModelExecuteOn(), SetProfileTime(MODEL_INFER_START));

ge/graph/load/model_manager/davinci_model.h

@@ -905,6 +905,7 @@ class DavinciModel {
 
   thread thread_id_;
   thread shrink_id_;
+  bool is_shrink_created_;
 
   shared_ptr<ModelListener> listener_;
 

ge/graph/load/model_manager/model_manager.cc

@@ -1058,7 +1058,6 @@ Status ModelManager::LoadModelOffline(uint32_t &model_id, const ModelData &model
   }
 
   do {
-    GeModelPtr ge_model = model_helper.GetGeModel();
     shared_ptr<DavinciModel> davinci_model = MakeShared<DavinciModel>(model.priority, listener);
     if (davinci_model == nullptr) {
       GELOGE(ACL_ERROR_GE_MEMORY_ALLOCATION, "Make shared failed");
@@ -1066,11 +1065,12 @@ Status ModelManager::LoadModelOffline(uint32_t &model_id, const ModelData &model
     }
     davinci_model->SetProfileTime(MODEL_LOAD_START, (timespec.tv_sec * kTimeSpecNano +
                                                      timespec.tv_nsec));  // 1000 ^ 3 converts second to nanosecond
-    ret = davinci_model->Assign(ge_model);
+    ret = davinci_model->Assign(model_helper.GetGeModel());
     if (ret != SUCCESS) {
       GELOGW("assign model failed.");
       break;
     }
+    model_helper.ResetGeModel();
     davinci_model->SetId(model_id);
 
     int32_t device_id = 0;

inc/framework/common/helper/model_helper.h

@@ -53,6 +53,11 @@ class GE_FUNC_VISIBILITY ModelHelper {
   Status GetBaseNameFromFileName(const std::string &file_name, std::string &base_name);
   Status GetModelNameFromMergedGraphName(const std::string &graph_name, std::string &model_name);
 
+  void ResetGeModel() {
+    root_model_.reset();
+    model_.reset();
+  }
+
  private:
   bool is_assign_model_ = false;
   bool is_offline_ = true;