!1679 Add OptimizeAfterStage1

From: @selfws
Reviewed-by: @ji_chen,@wqtshg
Signed-off-by: @ji_chen

ge/graph/manager/graph_manager.cc

@@ -818,6 +818,7 @@ Status GraphManager::PreRunOptimizeOriginalGraph(const GraphNodePtr &graph_node,
     GM_RUN_AND_DUMP_PERF("OptimizeSwitchOp", stages.preparer.SwitchOpOptimize, compute_graph);
   }
   GM_RUN_AND_DUMP_PERF("Optimize1", OptimizeStage1, compute_graph);
+  GM_RUN_AND_DUMP_PERF("OptimizeAfterStage1", stages.optimizer.OptimizeAfterStage1, compute_graph);
   GM_RUN_AND_DUMP_PERF("InferShape2", compute_graph->InferShapeInNeed);
 
   PassManager graph_pass;

ge/graph/optimize/graph_optimize.cc

@@ -144,7 +144,7 @@ Status GraphOptimize::OptimizeOriginalGraph(ComputeGraphPtr &compute_graph) {
   }
 
   auto graph_optimizer = instance_ptr->OpsKernelManagerObj().GetAllGraphOptimizerObjsByPriority();
-  GELOGI("optimize by opskernel in original graph optimize phase. num of graph_optimizer is %lu.",
+  GELOGI("optimize by opskernel in original graph optimize phase. num of graph_optimizer is %zu.",
          graph_optimizer.size());
   string exclude_core_Type = (core_type_ == kVectorCore) ? kAicoreEngine : kVectorEngine;
   GELOGD("[OptimizeOriginalGraph]: engine type will exclude: %s", exclude_core_Type.c_str());
@@ -179,7 +179,7 @@ Status GraphOptimize::OptimizeOriginalGraphJudgeInsert(ComputeGraphPtr &compute_
   }
 
   auto graph_optimizer = instance_ptr->OpsKernelManagerObj().GetAllGraphOptimizerObjsByPriority();
-  GELOGI("optimize by opskernel in original graph optimize phase. num of graph_optimizer is %lu.",
+  GELOGI("optimize by opskernel in judging insert phase. num of graph_optimizer is %zu.",
          graph_optimizer.size());
   string exclude_core_Type = (core_type_ == kVectorCore) ? kAicoreEngine : kVectorEngine;
   if (graph_optimizer.size() != 0) {
@@ -266,6 +266,44 @@ Status GraphOptimize::OptimizeGraphBeforeBuildForRts(ComputeGraphPtr &compute_gr
   return ret;
 }
 
+Status GraphOptimize::OptimizeAfterStage1(ComputeGraphPtr &compute_graph) {
+  GE_CHECK_NOTNULL(compute_graph);
+  GELOGD("OptimizeAfterStage1 in");
+  if (GetContext().GetHostExecFlag()) {
+    // graph exec on host, no need OptimizeAfterStage1
+    return SUCCESS;
+  }
+
+  Status ret = SUCCESS;
+  std::shared_ptr<GELib> instance_ptr = ge::GELib::GetInstance();
+  if (instance_ptr == nullptr || !instance_ptr->InitFlag()) {
+    REPORT_INNER_ERROR("E19999", "Gelib not init before, check invalid");
+    GELOGE(GE_CLI_GE_NOT_INITIALIZED, "OptimizeAfterStage1 failed.");
+    return GE_CLI_GE_NOT_INITIALIZED;
+  }
+
+  auto graph_optimizer = instance_ptr->OpsKernelManagerObj().GetAllGraphOptimizerObjsByPriority();
+  GELOGI("Optimize by ops kernel in after stage1 phase, num of graph_optimizer is %zu.", graph_optimizer.size());
+  string exclude_core_type = (core_type_ == kVectorCore) ? kAicoreEngine : kVectorEngine;
+  if (graph_optimizer.size() != 0) {
+    for (auto iter = graph_optimizer.begin(); iter != graph_optimizer.end(); ++iter) {
+      if (iter->first == exclude_core_type) {
+        GELOGI("[OptimizeAfterStage1]: engine type will exclude:%s.", exclude_core_type.c_str());
+        continue;
+      }
+      GELOGI("Begin to optimize graph after stage1 by engine %s.", iter->first.c_str());
+      ret = (iter->second)->OptimizeAfterStage1(*compute_graph);
+      if (ret != SUCCESS) {
+        REPORT_INNER_ERROR("E19999", "Call OptimizeAfterStage1 failed, ret:%d, engine_name:%s, "
+                           "graph_name:%s.", ret, iter->first.c_str(), compute_graph->GetName().c_str());
+        GELOGE(ret, "[OptimizeAfterStage1]: graph optimize failed, ret:%d.", ret);
+        return ret;
+      }
+    }
+  }
+  return ret;
+}
+
 Status GraphOptimize::SetOptions(const ge::GraphManagerOptions &options) {
   if (options.framework_type >= static_cast<int32_t>(domi::FrameworkType::FRAMEWORK_RESERVED)) {
     GELOGE(GE_GRAPH_OPTIONS_INVALID, "Optimize Type %d invalid.", options.framework_type);

ge/graph/optimize/graph_optimize.h

@@ -58,6 +58,9 @@ class GraphOptimize {
   // for rts optimize before build to add attr and insert memcpy op
   Status OptimizeGraphBeforeBuildForRts(ComputeGraphPtr &compute_graph);
 
+  // optimize whole graph, using after stage1
+  Status OptimizeAfterStage1(ComputeGraphPtr &graph);
+
   // set options
   Status SetOptions(const GraphManagerOptions &options);
 

metadef

@@ -1 +1 @@
-Subproject commit 05b2882f8d0364d295aae6ed4ab818a6dc83ad9a
+Subproject commit 82d8168408306116c0d69227ec33956ff988b316