optimize-backend-anf-ir-proto-dump

mindspore/ccsrc/pre_activate/ascend/ascend_backend_optimization.cc

@@ -89,8 +89,8 @@ void AddAscendBackendOptionalIRFusion(PassManager *ir_fusion_pm) {
   ir_fusion_pm->AddPass(std::make_shared<ClipByNormNoDivSquareSumFusion>());
   ir_fusion_pm->AddPass(std::make_shared<LambUpdateWithLRRuleFusion>());
   ir_fusion_pm->AddPass(std::make_shared<ConfusionSoftmaxGradRule>());
-  ir_fusion_pm->AddPass(std::make_shared<LambNextMVRule>());
   ir_fusion_pm->AddPass(std::make_shared<LambNextMVWithDecayRule>());
+  ir_fusion_pm->AddPass(std::make_shared<LambNextMVRule>());
   ir_fusion_pm->AddPass(std::make_shared<LambNextRightRule>());
   ir_fusion_pm->AddPass(std::make_shared<LambUpdateWithLrV2>());
   ir_fusion_pm->AddPass(std::make_shared<ReshapeTransposeFusion>());
@@ -201,7 +201,7 @@ void AscendBackendIRFusionOptimization(const std::shared_ptr<session::KernelGrap
     std::string file_path = save_graphs_path + "/" + "hwopt_d_ir_fusion_before" + "_graph_" +
                             std::to_string(kernel_graph->graph_id()) + ".ir";
     DumpIR(file_path, kernel_graph);
-    DumpIRProto(kernel_graph, "before_hwopt");
+    DumpIRProto(kernel_graph, "before_hwopt_" + std::to_string(kernel_graph->graph_id()));
   }
   auto optimizer = std::make_shared<GraphOptimizer>();
   auto ir_fusion_pm = std::make_shared<PassManager>("ir_fusion_pm");
@@ -305,7 +305,7 @@ void AscendBackendOptimization(const std::shared_ptr<session::KernelGraph> &kern
     std::string file_path =
       save_graphs_path + "/" + "hwopt_d_end" + "_graph_" + std::to_string(kernel_graph->graph_id()) + ".ir";
     DumpIR(file_path, kernel_graph, true);
-    DumpIRProto(kernel_graph, "after_hwopt");
+    DumpIRProto(kernel_graph, "after_hwopt_" + std::to_string(kernel_graph->graph_id()));
   }
 }
 

mindspore/ccsrc/session/ascend_session.cc

@@ -263,6 +263,7 @@ void AscendSession::BuildGraph(GraphId graph_id) {
   }
   // sync the inital const tensor to device
   SyncInitialTenosrToDevice();
+  ExportChildGraphs(graph_id);
   MS_LOG(INFO) << "end";
 }
 
@@ -558,6 +559,36 @@ void AscendSession::Dump(const std::shared_ptr<KernelGraph> &kernel_graph) const
   MS_LOG(INFO) << "Finish!";
 }
 
+void AscendSession::ExportChildGraphs(const GraphId graph_id) {
+#ifdef ENABLE_DUMP_IR
+  auto context_ptr = MsContext::GetInstance();
+  MS_EXCEPTION_IF_NULL(context_ptr);
+  bool save_graphs = context_ptr->save_graphs_flag();
+  if (!save_graphs) {
+    return;
+  }
+  auto save_graphs_path = context_ptr->save_graphs_path();
+  if (save_graphs_path.empty()) {
+    save_graphs_path = ".";
+  }
+  if (graph_id == final_graph_id_) {
+    auto &graph_order = GetGraphOrder(final_graph_id_);
+    auto &graph_type = GetGraphOrderType(final_graph_id_);
+    for (size_t i = 0; i < graph_order.size(); i++) {
+      if (graph_type[i] == BRANCH_END || graph_type[i] == BRANCH_START) {
+        continue;
+      }
+      auto child_graph = GetGraph(graph_order[i]);
+      MS_LOG(DEBUG) << "Start export child graph " << graph_order[i];
+      std::string file_path = save_graphs_path + "/graph_build_" + std::to_string(child_graph->graph_id()) + ".ir";
+      DumpIR(file_path, child_graph, true);
+      DumpIRProto(child_graph, "vm_build_" + std::to_string(child_graph->graph_id()));
+      MS_LOG(DEBUG) << "End export child graph " << graph_order[i];
+    }
+  }
+#endif
+}
+
 GraphId AscendSession::SetFinalGraphInput(const std::vector<AnfNodePtr> &args) {
   MS_LOG(INFO) << "Start! Args size " << args.size();
   auto final_graph = NewKernelGraph();

mindspore/ccsrc/session/ascend_session.h

@@ -82,6 +82,7 @@ class AscendSession : public SessionBasic {
   void LoadTask(const std::shared_ptr<KernelGraph> &kernel_graph) const;
   void ExecTask(const std::shared_ptr<KernelGraph> &kernel_graph) const;
   void Dump(const std::shared_ptr<KernelGraph> &kernel_graph) const;
+  void ExportChildGraphs(const GraphId graph_id);
   // below functions are used for run op
   void RunOpHardwareOptimize(const std::shared_ptr<session::KernelGraph> &kernel_graph) const;
   void RunOpExecTask(const std::shared_ptr<KernelGraph> &kernel_graph) const;