!10339 [MSLITE][DEVELOP] fix bug of npu benchmark

From: @yangruoqi713
Reviewed-by: @zhang_xue_tong,@hangangqiang
Signed-off-by: @zhang_xue_tong

mindspore/lite/src/runtime/agent/npu/npu_executor.cc

@@ -19,6 +19,11 @@
 #include "src/runtime/agent/npu/npu_manager.h"
 #include "nnacl/pack.h"
 namespace mindspore::lite {
+NPUExecutor::~NPUExecutor() {
+  npu_input_tensors_.clear();
+  npu_output_tensors_.clear();
+}
+
 int NPUExecutor::Prepare(const std::vector<kernel::LiteKernel *> &kernels) {
   this->client_ = mindspore::lite::NPUManager::GetInstance()->GetClient(model_name_);
   if (this->client_ == nullptr) {

mindspore/lite/src/runtime/agent/npu/npu_executor.h

@@ -28,7 +28,7 @@ namespace mindspore::lite {
 class NPUExecutor : public Executor {
  public:
   explicit NPUExecutor(const std::string &model_name) { this->model_name_ = model_name; }
-  ~NPUExecutor() override = default;
+  ~NPUExecutor() override;
   int Prepare(const std::vector<kernel::LiteKernel *> &kernels) override;
 
   int Run(const std::vector<Tensor *> &in_tensors, const std::vector<Tensor *> &out_tensors,

mindspore/lite/src/runtime/agent/npu/npu_manager.cc

@@ -70,6 +70,7 @@ void NPUManager::Reset() {
   }
   models_.clear();
   for (auto client : clients_) {
+    client->UnLoadModel();
     client.reset();
   }
   clients_.clear();
@@ -213,7 +214,6 @@ int NPUManager::LoadModel(const std::shared_ptr<hiai::AiModelMngerClient> &clien
   }
 
   for (const auto &desc : desc_list) {
-    MS_LOG(ERROR) << desc->GetName();
     auto it = models_.find(desc->GetName());
     it->second->is_loaded = true;
     it->second->client_ = client;

mindspore/lite/src/runtime/agent/npu/npu_manager.h

@@ -42,7 +42,7 @@ struct SubGraphModel {
   bool is_loaded = false;
   int index_;
   std::string model_name_;
-  domi::ModelBufferData *model_buffer_data_;
+  std::shared_ptr<domi::ModelBufferData> model_buffer_data_;
   std::shared_ptr<hiai::AiModelMngerClient> client_;
   std::shared_ptr<hiai::AiModelDescription> desc_;
 };

mindspore/lite/src/runtime/agent/npu/optimizer/npu_fusion_pass.cc

@@ -55,7 +55,6 @@ void NPUFusionPass::UpdatePreKernels(kernel::LiteKernel *cur_kernel) {
       }
     }
     cur_kernel->set_in_kernels(cur_in_kernels);
-    kernels->erase(find(kernels->begin(), kernels->end(), in_kernel));
   }
 }
 
@@ -80,7 +79,6 @@ void NPUFusionPass::UpdatePostKernels(kernel::LiteKernel *cur_kernel) {
       }
     }
     cur_kernel->set_out_kernels(cur_out_kernels);
-    kernels->erase(find(kernels->begin(), kernels->end(), out_kernel));
   }
 }
 
@@ -191,10 +189,8 @@ int NPUFusionPass::FormatFusion(kernel::LiteKernel *kernel) {
       post_kernel->set_in_kernels(post_in_kernels);
       pre_out_kernels.push_back(post_kernel);
     }
-    kernels->erase(find(kernels->begin(), kernels->end(), nc2nh));
   }
   pre_kernel->set_out_kernels(pre_out_kernels);
-  kernels->erase(find(kernels->begin(), kernels->end(), kernel));
   return RET_OK;
 }
 

mindspore/lite/src/runtime/agent/npu/subgraph_npu_kernel.cc

@@ -33,6 +33,14 @@ namespace mindspore::kernel {
 using mindspore::lite::RET_ERROR;
 using mindspore::lite::RET_OK;
 
+SubGraphNpuKernel::~SubGraphNpuKernel() {
+  subgraph_input_op_.clear();
+  subgraph_output_op_.clear();
+  if (executor_ != nullptr) {
+    delete executor_;
+  }
+}
+
 domi::ModelBufferData *SubGraphNpuKernel::BuildIRModel() {
   ge::Graph graph("NPUGraph");
 

mindspore/lite/src/runtime/agent/npu/subgraph_npu_kernel.h

@@ -36,7 +36,7 @@ class SubGraphNpuKernel : public SubGraphKernel {
     subgraph_type_ = kNpuSubGraph;
   }
 
-  ~SubGraphNpuKernel() override = default;
+  ~SubGraphNpuKernel() override;
 
   int Init() override;