optimize find input need lock tensor

mindspore/ccsrc/backend/session/ascend_session.cc

@@ -187,6 +187,8 @@ GraphId AscendSession::CompileGraphImpl(NotNull<FuncGraphPtr> func_graph) {
   MemoryAlloc(root_graph.get());
   // generate and load task into device
   Load(root_graph);
+  root_graph->SetInputNodes();
+  root_graph->SetOptimizerFlag();
   DumpAllGraphs(all_graphs);
   // return the root_graph id to backend
   auto graph_id = root_graph->graph_id();

mindspore/ccsrc/backend/session/executor.cc

@@ -271,7 +271,7 @@ void Executor::RunGraphAsync(const SessionPtr &session, const GraphId &graph_id,
   task->session_ = session;
   task->graph_id_ = graph_id;
   task->input_tensors_ = inputs;
-  task->input_need_lock_tensors_ = session->GetNeedLockInputTensors(graph_id, inputs);
+  task->input_need_lock_tensors_ = session->GetInputNeedLockTensors(graph_id, inputs);
   for (auto &tensor : inputs) {
     if (tensor->NeedWait()) {
       if (tensor->IsGraphOutput()) {

mindspore/ccsrc/backend/session/gpu_session.cc

@@ -169,11 +169,7 @@ void GPUSession::LoadInputData(const std::shared_ptr<KernelGraph> &kernel_graph,
                                const std::vector<tensor::TensorPtr> &inputs_const) const {
   std::vector<tensor::TensorPtr> inputs(inputs_const);
   MS_EXCEPTION_IF_NULL(kernel_graph);
-  std::vector<AnfNodePtr> input_nodes;
+  auto &input_nodes = kernel_graph->input_nodes();
-  for (const auto &input_node : kernel_graph->inputs()) {
-    auto params = AnfAlgo::GetAllOutput(input_node);
-    std::copy(params.begin(), params.end(), std::back_inserter(input_nodes));
-  }
   auto ms_context = MsContext::GetInstance();
   MS_EXCEPTION_IF_NULL(ms_context);
   if (inputs.size() != input_nodes.size()) {

mindspore/ccsrc/backend/session/kernel_graph.cc

@@ -1212,6 +1212,25 @@ void KernelGraph::UpdateGraphDynamicAttr() {
   is_dynamic_shape_ = false;
 }
 
+void KernelGraph::SetInputNodes() {
+  input_nodes_.clear();
+  for (const auto &input_node : inputs()) {
+    auto params = AnfAlgo::GetAllOutput(input_node);
+    std::copy(params.begin(), params.end(), std::back_inserter(input_nodes_));
+  }
+}
+
+void KernelGraph::SetOptimizerFlag() {
+  has_optimizer_ = false;
+  for (const auto &cnode : execution_order_) {
+    MS_EXCEPTION_IF_NULL(cnode);
+    if (kOptOperatorSet.find(AnfAlgo::GetCNodeName(cnode)) != kOptOperatorSet.end()) {
+      has_optimizer_ = true;
+      return;
+    }
+  }
+}
+
 std::string KernelGraph::ToString() const { return std::string("kernel_graph_").append(std::to_string(graph_id_)); }
 
 KernelGraph::~KernelGraph() {

mindspore/ccsrc/backend/session/kernel_graph.h

@@ -185,6 +185,10 @@ class KernelGraph : public FuncGraph {
 
   void UpdateGraphDynamicAttr();
   bool is_dynamic_shape() const { return is_dynamic_shape_; }
+  void SetOptimizerFlag();
+  void SetInputNodes();
+  const std::vector<AnfNodePtr> &input_nodes() const { return input_nodes_; }
+  bool has_optimizer() const { return has_optimizer_; }
 
  private:
   // remove value node form graph
@@ -234,9 +238,9 @@ class KernelGraph : public FuncGraph {
   std::unordered_map<AnfNodePtr, std::vector<std::pair<AnfNodePtr, size_t>>> node_output_edges_;
   std::map<std::string, std::pair<AnfNodePtr, int>> summary_nodes_;
   // graph needn't execute
-  bool executable_;
+  bool executable_{false};
   // exist summary node in graph
-  bool summary_node_exist_;
+  bool summary_node_exist_{false};
   // valid inputs
   std::vector<bool> valid_inputs_;
 
@@ -251,7 +255,7 @@ class KernelGraph : public FuncGraph {
 
   CNodePtr start_label_;
   CNodePtr end_goto_;
-  bool null_output_;
+  bool null_output_{false};
   std::unordered_map<AnfNodePtr, AnfNodePtr> front_to_internal_outputs_map_;
   std::unordered_map<AnfNodePtr, std::unordered_map<int, std::pair<AnfNodePtr, bool>>> internal_outputs_to_front_map_;
   std::unordered_map<AnfNodePtr, std::unordered_map<int, tensor::TensorPtr>> internal_outputs_tensor_map_;
@@ -260,7 +264,9 @@ class KernelGraph : public FuncGraph {
   std::set<AnfNodePtr> visited_nodes_;
   std::map<AnfNodePtr, AnfNodePtr> edge_to_;
   std::stack<AnfNodePtr> loop_nodes_;
-  bool is_dynamic_shape_;
+  std::vector<AnfNodePtr> input_nodes_;
+  bool has_optimizer_{false};
+  bool is_dynamic_shape_{false};
 };
 }  // namespace session
 using KernelGraphPtr = std::shared_ptr<session::KernelGraph>;

mindspore/ccsrc/backend/session/session_basic.cc

@@ -852,6 +852,8 @@ KernelGraphPtr SessionBasic::ConstructKernelGraph(const AnfNodePtrList &lst, con
     graph->set_summary_node_exist(true);
   }
   opt::BackendCommonOptimization(graph);
+  graph->SetInputNodes();
+  graph->SetOptimizerFlag();
   return graph;
 }
 
@@ -971,11 +973,8 @@ void SessionBasic::LoadInputData(const std::shared_ptr<KernelGraph> &kernel_grap
   if (kernel_graph->input_ctrl_tensors()) {
     input_ctrl_size = LoadCtrlInputTensor(kernel_graph, &inputs);
   }
-  std::vector<AnfNodePtr> input_nodes;
+  auto &input_nodes = kernel_graph->input_nodes();
-  for (const auto &input_node : kernel_graph->inputs()) {
+
-    auto params = AnfAlgo::GetAllOutput(input_node);
-    std::copy(params.begin(), params.end(), std::back_inserter(input_nodes));
-  }
   if ((inputs.size() + input_ctrl_size) - 3 != input_nodes.size()) {
     MS_LOG(EXCEPTION) << "Tensor input:" << inputs.size() << " is not equal graph inputs:" << input_nodes.size()
                       << ", input_ctrl_size:" << input_ctrl_size;
@@ -1026,19 +1025,11 @@ void SessionBasic::UpdateOutputs(const std::shared_ptr<KernelGraph> &kernel_grap
   }
 }
 
-std::vector<tensor::TensorPtr> SessionBasic::GetNeedLockInputTensors(const GraphId &graph_id,
+std::vector<tensor::TensorPtr> SessionBasic::GetInputNeedLockTensors(const GraphId &graph_id,
                                                                      const std::vector<tensor::TensorPtr> &inputs) {
   auto graph = GetGraph(graph_id);
   MS_EXCEPTION_IF_NULL(graph);
-  bool has_optimizer = false;
+  if (!graph->has_optimizer()) {
-  for (const auto &cnode : graph->execution_order()) {
-    MS_EXCEPTION_IF_NULL(cnode);
-    if (kOptOperatorSet.find(AnfAlgo::GetCNodeName(cnode)) != kOptOperatorSet.end()) {
-      has_optimizer = true;
-      break;
-    }
-  }
-  if (!has_optimizer) {
     return {};
   }
   std::vector<tensor::TensorPtr> result;
@@ -1339,6 +1330,7 @@ std::shared_ptr<KernelGraph> SessionBasic::ConstructSingleOpGraph(const OpRunInf
   graph->set_execution_order(exe_order);
   // set output
   CreateOutputNode(cnode, graph);
+  graph->SetInputNodes();
   return graph;
 }
 

mindspore/ccsrc/backend/session/session_basic.h

@@ -98,7 +98,7 @@ class SessionBasic : public std::enable_shared_from_this<SessionBasic> {
     return true;
   }
   virtual void GetModelInputsInfo(uint32_t graph_id, std::vector<tensor::TensorPtr> *inputs) const {}
-  std::vector<tensor::TensorPtr> GetNeedLockInputTensors(const GraphId &graph_id,
+  std::vector<tensor::TensorPtr> GetInputNeedLockTensors(const GraphId &graph_id,
                                                          const std::vector<tensor::TensorPtr> &inputs);
 #ifdef ENABLE_DEBUGGER
   // set debugger