!19940 [MSLITE][DEVELOP] modify lite for new api

Merge pull request !19940 from yangruoqi713/master

cmake/package_lite.cmake

@@ -216,8 +216,6 @@ if(PLATFORM_ARM64)
     endif()
     install(FILES ${TOP_DIR}/mindspore/core/ir/dtype/type_id.h DESTINATION ${RUNTIME_INC_DIR}/ir/dtype
             COMPONENT ${RUNTIME_COMPONENT_NAME})
-    install(FILES ${TOP_DIR}/mindspore/core/ir/format.h DESTINATION ${RUNTIME_INC_DIR}/ir
-            COMPONENT ${RUNTIME_COMPONENT_NAME})
     install(DIRECTORY ${TOP_DIR}/include/api/ DESTINATION ${RUNTIME_INC_DIR}/api
             COMPONENT ${RUNTIME_COMPONENT_NAME} FILES_MATCHING PATTERN "*.h" PATTERN "ops*" EXCLUDE)
     __install_micro_wrapper()
@@ -269,8 +267,6 @@ elseif(PLATFORM_ARM32)
     endif()
     install(FILES ${TOP_DIR}/mindspore/core/ir/dtype/type_id.h DESTINATION ${RUNTIME_INC_DIR}/ir/dtype
             COMPONENT ${RUNTIME_COMPONENT_NAME})
-    install(FILES ${TOP_DIR}/mindspore/core/ir/format.h DESTINATION ${RUNTIME_INC_DIR}/ir
-            COMPONENT ${RUNTIME_COMPONENT_NAME})
     install(DIRECTORY ${TOP_DIR}/include/api/ DESTINATION ${RUNTIME_INC_DIR}/api
             COMPONENT ${RUNTIME_COMPONENT_NAME} FILES_MATCHING PATTERN "*.h" PATTERN "ops*" EXCLUDE)
     __install_micro_wrapper()
@@ -306,6 +302,8 @@ elseif(WIN32)
                 DESTINATION ${CONVERTER_ROOT_DIR}/lib COMPONENT ${RUNTIME_COMPONENT_NAME})
         install(FILES ${glog_LIBPATH}/../bin/libglog.dll DESTINATION ${CONVERTER_ROOT_DIR}/lib
                 COMPONENT ${RUNTIME_COMPONENT_NAME})
+        install(DIRECTORY ${TOP_DIR}/include/api/ DESTINATION ${CONVERTER_ROOT_DIR}/include/api
+                COMPONENT ${RUNTIME_COMPONENT_NAME} FILES_MATCHING PATTERN "*.h" PATTERN "ops*" EXCLUDE)
         install(DIRECTORY ${TOP_DIR}/mindspore/core/abstract/ DESTINATION ${CONVERTER_ROOT_DIR}/include/core/abstract
                 COMPONENT ${RUNTIME_COMPONENT_NAME} FILES_MATCHING PATTERN "*.h")
         install(DIRECTORY ${TOP_DIR}/mindspore/core/base/ DESTINATION ${CONVERTER_ROOT_DIR}/include/core/base
@@ -372,8 +370,6 @@ elseif(WIN32)
             COMPONENT ${RUNTIME_COMPONENT_NAME})
     install(FILES ${TOP_DIR}/mindspore/core/ir/dtype/type_id.h DESTINATION ${RUNTIME_INC_DIR}/ir/dtype
             COMPONENT ${RUNTIME_COMPONENT_NAME})
-    install(FILES ${TOP_DIR}/mindspore/core/ir/format.h DESTINATION ${RUNTIME_INC_DIR}/ir
-            COMPONENT ${RUNTIME_COMPONENT_NAME})
     install(DIRECTORY ${TOP_DIR}/include/api/ DESTINATION ${RUNTIME_INC_DIR}/api
             COMPONENT ${RUNTIME_COMPONENT_NAME} FILES_MATCHING PATTERN "*.h" PATTERN "ops*" EXCLUDE)
     install(FILES ${TOP_DIR}/build/mindspore/src/${MINDSPORE_LITE_LIB_NAME}.a DESTINATION ${RUNTIME_LIB_DIR}
@@ -404,8 +400,6 @@ else()
             COMPONENT ${RUNTIME_COMPONENT_NAME})
     install(FILES ${TOP_DIR}/mindspore/core/ir/dtype/type_id.h DESTINATION ${RUNTIME_INC_DIR}/ir/dtype
             COMPONENT ${RUNTIME_COMPONENT_NAME})
-    install(FILES ${TOP_DIR}/mindspore/core/ir/format.h DESTINATION ${RUNTIME_INC_DIR}/ir
-            COMPONENT ${RUNTIME_COMPONENT_NAME})
     install(DIRECTORY ${TOP_DIR}/include/api/ DESTINATION ${RUNTIME_INC_DIR}/api
             COMPONENT ${RUNTIME_COMPONENT_NAME} FILES_MATCHING PATTERN "*.h" PATTERN "ops*" EXCLUDE)
     install(FILES ${TOP_DIR}/mindspore/lite/build/src/${MINDSPORE_LITE_LIB_NAME}.so DESTINATION ${RUNTIME_LIB_DIR}
@@ -423,6 +417,8 @@ else()
         install(DIRECTORY ${TOP_DIR}/mindspore/lite/include/ DESTINATION ${CONVERTER_ROOT_DIR}/include
                 COMPONENT ${RUNTIME_COMPONENT_NAME} FILES_MATCHING PATTERN "*.h"
                 PATTERN "train*" EXCLUDE PATTERN "delegate.h" EXCLUDE PATTERN "lite_session.h" EXCLUDE)
+        install(DIRECTORY ${TOP_DIR}/include/api/ DESTINATION ${CONVERTER_ROOT_DIR}/include/api
+                COMPONENT ${RUNTIME_COMPONENT_NAME} FILES_MATCHING PATTERN "*.h" PATTERN "ops*" EXCLUDE)
         install(DIRECTORY ${TOP_DIR}/mindspore/core/abstract/ DESTINATION ${CONVERTER_ROOT_DIR}/include/core/abstract
                 COMPONENT ${RUNTIME_COMPONENT_NAME} FILES_MATCHING PATTERN "*.h")
         install(DIRECTORY ${TOP_DIR}/mindspore/core/base/ DESTINATION ${CONVERTER_ROOT_DIR}/include/core/base

mindspore/lite/include/allocator.h → include/api/allocator.h

@@ -14,11 +14,11 @@
  * limitations under the License.
  */
 
-#ifndef MINDSPORE_LITE_INCLUDE_ALLOCATOR_H_
-#define MINDSPORE_LITE_INCLUDE_ALLOCATOR_H_
+#ifndef MINDSPORE_INCLUDE_API_ALLOCATOR_H
+#define MINDSPORE_INCLUDE_API_ALLOCATOR_H
 
 #include <memory>
-#include "include/lite_utils.h"
+#include "include/api/types.h"
 
 namespace mindspore {
 /// \brief Allocator defined a memory pool for malloc memory and free memory dynamically.
@@ -85,4 +85,4 @@ class MS_API Allocator {
   size_t aligned_size_ = 32;
 };
 }  // namespace mindspore
-#endif  // MINDSPORE_LITE_INCLUDE_ALLOCATOR_H_
+#endif  // MINDSPORE_INCLUDE_API_ALLOCATOR_H

include/api/context.h

@@ -36,6 +36,7 @@ enum DeviceType {
 };
 
 class Allocator;
+class Delegate;
 class DeviceInfoContext;
 
 class MS_API Context {
@@ -57,6 +58,9 @@ class MS_API Context {
   void SetEnableParallel(bool is_parallel);
   bool GetEnableParallel() const;
 
+  void SetDelegate(const std::shared_ptr<Delegate> &delegate);
+  std::shared_ptr<Delegate> GetDelegate() const;
+
   std::vector<std::shared_ptr<DeviceInfoContext>> &MutableDeviceInfo();
 
  private:

include/api/data_type.h

@@ -23,6 +23,7 @@ enum class DataType : int {
   kObjectTypeList = 13,
   kObjectTypeTuple = 14,
   kObjectTypeTensorType = 17,
+  kNumberTypeBegin = 29,
   kNumberTypeBool = 30,
   kNumberTypeInt8 = 32,
   kNumberTypeInt16 = 33,

mindspore/lite/include/delegate.h → include/api/delegate.h

@@ -14,15 +14,14 @@
  * limitations under the License.
  */
 
-#ifndef MINDSPORE_LITE_DELEGATE_DELEGATE_H_
-#define MINDSPORE_LITE_DELEGATE_DELEGATE_H_
+#ifndef MINDSPORE_INCLUDE_API_DELEGATE_H
+#define MINDSPORE_INCLUDE_API_DELEGATE_H
 
 #include <map>
 #include <vector>
 #include <memory>
-#include "include/ms_tensor.h"
-#include "include/context.h"
-#include "include/kernel.h"
+#include "schema/model_generated.h"
+#include "include/api/kernel.h"
 
 namespace mindspore {
 typedef enum {
@@ -35,8 +34,8 @@ using KernelIter = std::vector<kernel::Kernel *>::iterator;
 class MS_API DelegateModel {
  public:
   /// \brief Constructor of MindSpore Lite DelegateModel.
-  DelegateModel(std::vector<kernel::Kernel *> *kernels, const std::vector<tensor::MSTensor *> &inputs,
-                const std::vector<tensor::MSTensor *> &outputs,
+  DelegateModel(std::vector<kernel::Kernel *> *kernels, const std::vector<MSTensor> &inputs,
+                const std::vector<MSTensor> &outputs,
                 const std::map<kernel::Kernel *, const schema::Primitive *> &primitives, SchemaVersion version)
       : kernels_(kernels), inputs_(inputs), outputs_(outputs), primitives_(primitives), version_(version) {}
 
@@ -71,12 +70,12 @@ class MS_API DelegateModel {
   /// \brief Get the input tensors of DelegateModel.
   ///
   /// \return The input tensor vector of DelegateModel.
-  const std::vector<mindspore::tensor::MSTensor *> &inputs() { return this->inputs_; }
+  const std::vector<mindspore::MSTensor> &inputs() { return this->inputs_; }
 
   /// \brief Get the output tensors of DelegateModel.
   ///
   /// \return The ioutput tensor vector of DelegateModel.
-  const std::vector<mindspore::tensor::MSTensor *> &outputs() { return this->outputs_; }
+  const std::vector<mindspore::MSTensor> &outputs() { return this->outputs_; }
 
   /// \brief Get the ms model version.
   ///
@@ -85,14 +84,12 @@ class MS_API DelegateModel {
 
  protected:
   std::vector<kernel::Kernel *> *kernels_;
-  const std::vector<mindspore::tensor::MSTensor *> &inputs_;
-  const std::vector<mindspore::tensor::MSTensor *> &outputs_;
+  const std::vector<mindspore::MSTensor> &inputs_;
+  const std::vector<mindspore::MSTensor> &outputs_;
   const std::map<kernel::Kernel *, const schema::Primitive *> &primitives_;
   SchemaVersion version_;
 };
 
-typedef void (*DelegateHook)(std::shared_ptr<Delegate> delegate);
-static void HookNullFuc(std::shared_ptr<Delegate> delegate) {}
 class MS_API Delegate {
  public:
   /// \brief Constructor of MindSpore Lite Delegate.
@@ -112,10 +109,6 @@ class MS_API Delegate {
   ///
   /// \param[in] model Define the delegate model to be built.
   virtual int Build(DelegateModel *model) = 0;
-
-  DelegateHook init_hook_ = HookNullFuc;
-  DelegateHook build_hook_ = HookNullFuc;
-  DelegateHook run_hook_ = HookNullFuc;
 };
 }  // namespace mindspore
-#endif  // MINDSPORE_LITE_DELEGATE_DELEGATE_H_
+#endif  // MINDSPORE_INCLUDE_API_DELEGATE_H

mindspore/core/ir/format.h → include/api/format.h

@@ -16,8 +16,8 @@
  * limitations under the License.
  */
 
-#ifndef MINDSPORE_CORE_IR_FORMAT_H_
-#define MINDSPORE_CORE_IR_FORMAT_H_
+#ifndef MINDSPORE_INCLUDE_API_FORMAT_H
+#define MINDSPORE_INCLUDE_API_FORMAT_H
 
 #include <cstdint>
 
@@ -43,4 +43,4 @@ enum Format : int64_t {
   NCW = 17
 };
 }  // namespace mindspore
-#endif  // MINDSPORE_CORE_IR_FORMAT_H_
+#endif  // MINDSPORE_INCLUDE_API_FORMAT_H

mindspore/lite/include/kernel.h → include/api/kernel.h

@@ -14,22 +14,22 @@
  * limitations under the License.
  */
 
-#ifndef MINDSPORE_LITE_SRC_KERNEL_H_
-#define MINDSPORE_LITE_SRC_KERNEL_H_
+#ifndef MINDSPORE_INCLUDE_API_KERNEL_H
+#define MINDSPORE_INCLUDE_API_KERNEL_H
 #include <vector>
 #include <string>
 #include <utility>
 #include "schema/model_generated.h"
-#include "include/lite_utils.h"
-#include "include/context.h"
+#include "include/api/types.h"
+#include "include/api/context.h"
 
 namespace mindspore::kernel {
 class Kernel {
  public:
   Kernel() = default;
 
-  Kernel(const std::vector<tensor::MSTensor *> &inputs, const std::vector<tensor::MSTensor *> &outputs,
-         const schema::Primitive *primitive, const lite::Context *ctx)
+  Kernel(const std::vector<mindspore::MSTensor> &inputs, const std::vector<mindspore::MSTensor> &outputs,
+         const schema::Primitive *primitive, const mindspore::Context *ctx)
       : inputs_(std::move(inputs)), outputs_(std::move(outputs)), primitive_(primitive), context_(ctx) {
     if (primitive != nullptr) {
       type_ = primitive->value_type();
@@ -46,33 +46,34 @@ class Kernel {
 
   virtual schema::PrimitiveType type() const { return type_; }
 
-  virtual void set_inputs(const std::vector<mindspore::tensor::MSTensor *> &in_tensors) { this->inputs_ = in_tensors; }
-  virtual void set_input(mindspore::tensor::MSTensor *in_tensor, int index) { this->inputs_[index] = in_tensor; }
+  virtual void set_inputs(const std::vector<mindspore::MSTensor> &in_tensors) { this->inputs_ = in_tensors; }
 
-  virtual void set_outputs(const std::vector<mindspore::tensor::MSTensor *> &out_tensors) {
-    this->outputs_ = out_tensors;
-  }
+  virtual void set_input(mindspore::MSTensor in_tensor, int index) { this->inputs_[index] = in_tensor; }
+
+  virtual void set_outputs(const std::vector<mindspore::MSTensor> &out_tensors) { this->outputs_ = out_tensors; }
 
-  virtual void set_output(mindspore::tensor::MSTensor *out_tensor, int index) { this->outputs_[index] = out_tensor; }
+  virtual void set_output(mindspore::MSTensor out_tensor, int index) { this->outputs_[index] = out_tensor; }
 
-  virtual const std::vector<mindspore::tensor::MSTensor *> &inputs() { return this->inputs_; }
+  virtual const std::vector<mindspore::MSTensor> &inputs() { return this->inputs_; }
 
-  virtual const std::vector<mindspore::tensor::MSTensor *> &outputs() { return this->outputs_; }
+  virtual const std::vector<mindspore::MSTensor> &outputs() { return this->outputs_; }
 
   std::string name() const { return this->name_; }
 
   void set_name(const std::string &name) { this->name_ = name; }
-  const lite::Context *context() const { return this->context_; }
+
+  const mindspore::Context *context() const { return this->context_; }
+
   const schema::Primitive *primitive() const { return this->primitive_; }
 
  protected:
-  std::vector<mindspore::tensor::MSTensor *> inputs_;
-  std::vector<mindspore::tensor::MSTensor *> outputs_;
+  std::vector<mindspore::MSTensor> inputs_;
+  std::vector<mindspore::MSTensor> outputs_;
   schema::PrimitiveType type_ = schema::PrimitiveType_NONE;
   std::string name_;
   const schema::Primitive *primitive_ = nullptr;
-  const lite::Context *context_ = nullptr;
+  const mindspore::Context *context_ = nullptr;
 };
 }  // namespace mindspore::kernel
 
-#endif  // MINDSPORE_LITE_SRC_KERNEL_H_
+#endif  // MINDSPORE_INCLUDE_API_KERNEL_H

include/api/types.h

@@ -23,6 +23,7 @@
 #include <functional>
 #include "include/api/data_type.h"
 #include "include/api/dual_abi_helper.h"
+#include "include/api/format.h"
 
 #ifdef _WIN32
 #define MS_API __declspec(dllexport)
@@ -56,6 +57,7 @@ enum OptimizationLevel : uint32_t {
   kOptimizationType = 0xFFFFFFFF
 };
 
+class Allocator;
 class MS_API MSTensor {
  public:
   class Impl;
@@ -91,6 +93,17 @@ class MS_API MSTensor {
   MSTensor *Clone() const;
   bool operator==(std::nullptr_t) const;
   bool operator!=(std::nullptr_t) const;
+  bool operator==(const MSTensor &tensor) const;
+
+  void SetShape(const std::vector<int64_t> &shape);
+  void SetDataType(enum DataType data_type);
+  void SetTensorName(const std::string &name);
+  void SetAllocator(std::shared_ptr<Allocator> allocator);
+  std::shared_ptr<Allocator> allocator() const;
+  void SetFormat(mindspore::Format format);
+  mindspore::Format format() const;
+  void SetData(void *data);
+  const std::shared_ptr<Impl> impl() const { return impl_; }
 
  private:
   // api without std::string

mindspore/core/utils/check_convert_utils.h

@@ -27,7 +27,7 @@
 #include "base/base.h"
 #include "ir/anf.h"
 #include "ir/dtype/type_id.h"
-#include "ir/format.h"
+#include "include/api/format.h"
 #include "utils/log_adapter.h"
 namespace mindspore {
 typedef std::pair<std::map<std::string, int64_t>, std::map<int64_t, std::string>> AttrConverterPair;

mindspore/lite/include/registry/kernel_interface.h

@@ -22,7 +22,7 @@
 #include <vector>
 #include <memory>
 #include "include/model.h"
-#include "include/ms_tensor.h"
+#include "include/api/types.h"
 #include "schema/model_generated.h"
 
 namespace mindspore {
@@ -46,7 +46,7 @@ class MS_API KernelInterface {
   /// \param[in] primitive Define the attributes of op.
   ///
   /// \return  STATUS as an error code of inferring, STATUS is defined in errorcode.h..
-  virtual int Infer(const std::vector<tensor::MSTensor *> &inputs, const std::vector<tensor::MSTensor *> &outputs,
+  virtual int Infer(std::vector<mindspore::MSTensor> *inputs, std::vector<mindspore::MSTensor> *outputs,
                     const schema::Primitive *primitive) {
     return 0;
   }
@@ -58,7 +58,7 @@ class MS_API KernelInterface {
   /// \param[in] param Define the contr of performance.
   ///
   /// \return STATUS as an error code of inferring, STATUS is defined in errorcode.h.
-  virtual int GetCapability(const std::vector<tensor::MSTensor *> &tensor_in, const schema::Primitive *primitive,
+  virtual int GetCapability(const std::vector<mindspore::MSTensor> &tensor_in, const schema::Primitive *primitive,
                             CapabilityParam *param) {
     return 0;
   }

mindspore/lite/include/registry/register_kernel.h

@@ -22,9 +22,10 @@
 #include <vector>
 #include <memory>
 #include "schema/model_generated.h"
-#include "include/context.h"
-#include "include/ms_tensor.h"
-#include "include/kernel.h"
+#include "include/api/context.h"
+#include "include/api/types.h"
+#include "include/api/kernel.h"
+#include "ir/dtype/type_id.h"
 
 namespace mindspore {
 namespace kernel {
@@ -57,8 +58,8 @@ struct MS_API KernelDesc {
 ///
 /// \return Smart Pointer of kernel.
 using CreateKernel = std::function<std::shared_ptr<kernel::Kernel>(
-  const std::vector<tensor::MSTensor *> &inputs, const std::vector<tensor::MSTensor *> &outputs,
-  const schema::Primitive *primitive, const lite::Context *ctx)>;
+  const std::vector<MSTensor> &inputs, const std::vector<MSTensor> &outputs, const schema::Primitive *primitive,
+  const mindspore::Context *ctx)>;
 
 /// \brief RegisterKernel Defined registration of kernel.
 class MS_API RegisterKernel {

mindspore/lite/micro/cmake/file_list.cmake

@@ -127,9 +127,13 @@ set(CODER_OPCODERS_SRC
         )
 
 set(LITE_SRC
+        ${LITE_DIR}/src/cxx_api/tensor_utils.cc
+        ${LITE_DIR}/src/cxx_api/types.cc
+        ${LITE_DIR}/src/cxx_api/tensor/tensor_impl.cc
         ${LITE_DIR}/src/common/file_utils.cc
         ${LITE_DIR}/src/common/graph_util.cc
         ${LITE_DIR}/src/common/prim_util.cc
+        ${LITE_DIR}/src/common/string_util.cc
         ${LITE_DIR}/src/common/tensor_util.cc
         ${LITE_DIR}/src/runtime/infer_manager.cc
         ${LITE_DIR}/src/registry/kernel_interface.cc
@@ -137,12 +141,14 @@ set(LITE_SRC
         ${LITE_DIR}/src/registry/register_kernel.cc
         ${LITE_DIR}/src/registry/register_kernel_impl.cc
         ${LITE_DIR}/src/lite_model.cc
+        ${LITE_DIR}/src/ms_tensor.cc
         ${LITE_DIR}/src/tensorlist.cc
         ${LITE_DIR}/src/tensor.cc
         ${LITE_DIR}/src/weight_decoder.cc
         ${LITE_DIR}/src/huffman_decode.cc
         ${LITE_DIR}/src/common/log_adapter.cc
         ${LITE_DIR}/src/common/utils.cc
+        ${LITE_DIR}/../core/utils/status.cc
         ### tools
         ${LITE_DIR}/tools/common/flag_parser.cc
         )

mindspore/lite/micro/coder/generator/component/const_blocks/mtensor.cc

@@ -38,7 +38,7 @@ const char tensor_header[] = R"RAW(
 #define MINDSPORE_LITE_MICRO_LIBRARY_SOURCE_TENSOR_H_
 
 #include "include/ms_tensor.h"
-#include "include/ir/format.h"
+#include "include/api/format.h"
 
 namespace mindspore {
 namespace lite {

mindspore/lite/src/CMakeLists.txt

@@ -62,6 +62,7 @@ endif()
 
 set(LITE_SRC
         ${API_SRC}
+        ${CMAKE_CURRENT_SOURCE_DIR}/common/context_util.cc
         ${CMAKE_CURRENT_SOURCE_DIR}/common/file_utils.cc
         ${CMAKE_CURRENT_SOURCE_DIR}/common/utils.cc
         ${CMAKE_CURRENT_SOURCE_DIR}/common/graph_util.cc

mindspore/lite/src/common/context_util.cc

@@ -0,0 +1,120 @@
+/**
+ * Copyright 2021 Huawei Technologies Co., Ltd
+ *
+ * Licensed under the Apache License, Version 2.0 (the "License");
+ * you may not use this file except in compliance with the License.
+ * You may obtain a copy of the License at
+ *
+ * http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS,
+ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License.
+ */
+
+#include "src/common/context_util.h"
+#include <set>
+#include <map>
+#include <memory>
+#include <string>
+#include "src/common/log_adapter.h"
+
+namespace mindspore {
+namespace lite {
+namespace {
+template <class T>
+void PassBasicProperties(std::shared_ptr<T> device_info, const lite::DeviceContext &device_context) {
+  device_info->SetProvider(device_context.provider_);
+  device_info->SetProviderDevice(device_context.provider_device_);
+  device_info->SetAllocator(device_context.allocator_);
+}
+
+std::shared_ptr<mindspore::CPUDeviceInfo> CPUDeviceInfoFromCPUDeviceContext(const lite::DeviceContext &cpu_context) {
+  if (cpu_context.device_type_ != DT_CPU) {
+    MS_LOG(ERROR) << "function input parameter is not cpu context.";
+    return nullptr;
+  }
+  auto cpu_info = std::make_shared<mindspore::CPUDeviceInfo>();
+  cpu_info->SetEnableFP16(cpu_context.device_info_.cpu_device_info_.enable_float16_);
+  PassBasicProperties(cpu_info, cpu_context);
+  return cpu_info;
+}
+
+std::shared_ptr<mindspore::MaliGPUDeviceInfo> GPUDeviceInfoFromGPUDeviceContext(
+  const lite::DeviceContext &gpu_context) {
+  if (gpu_context.device_type_ != DT_GPU) {
+    MS_LOG(ERROR) << "function input parameter is not gpu context.";
+    return nullptr;
+  }
+  auto gpu_info = std::make_shared<mindspore::MaliGPUDeviceInfo>();
+  gpu_info->SetEnableFP16(gpu_context.device_info_.gpu_device_info_.enable_float16_);
+  PassBasicProperties(gpu_info, gpu_context);
+  return gpu_info;
+}
+
+std::shared_ptr<mindspore::KirinNPUDeviceInfo> NPUDeviceInfoFromNPUDeviceContext(
+  const lite::DeviceContext &npu_context) {
+  if (npu_context.device_type_ != DT_NPU) {
+    MS_LOG(ERROR) << "function input parameter is not npu context.";
+    return nullptr;
+  }
+  auto npu_info = std::make_shared<mindspore::KirinNPUDeviceInfo>();
+  npu_info->SetFrequency(npu_context.device_info_.npu_device_info_.frequency_);
+  PassBasicProperties(npu_info, npu_context);
+  return npu_info;
+}
+}  // namespace
+
+mindspore::Context *MSContextFromContext(const lite::Context *context) {
+  if (context == nullptr) {
+    MS_LOG(ERROR) << "context is nullptr";
+    return nullptr;
+  }
+  auto ms_context = new (std::nothrow) mindspore::Context();
+  if (ms_context == nullptr) {
+    MS_LOG(ERROR) << "New Context failed";
+    return nullptr;
+  }
+  ms_context->SetThreadNum(context->thread_num_);
+  ms_context->SetThreadAffinity(context->affinity_core_list_);
+  ms_context->SetEnableParallel(context->enable_parallel_);
+  ms_context->SetDelegate(context->delegate);
+  auto &device_infos = ms_context->MutableDeviceInfo();
+  std::map<DeviceType, std::function<std::shared_ptr<mindspore::DeviceInfoContext>(const lite::DeviceContext &)>>
+    transfer_funcs = {{DT_CPU, CPUDeviceInfoFromCPUDeviceContext},
+                      {DT_GPU, GPUDeviceInfoFromGPUDeviceContext},
+                      {DT_NPU, NPUDeviceInfoFromNPUDeviceContext}};
+  for (auto &device_context : context->device_list_) {
+    auto device_type = device_context.device_type_;
+    if (transfer_funcs.find(device_type) == transfer_funcs.end()) {
+      MS_LOG(ERROR) << "device type is invalid.";
+      return nullptr;
+    }
+    auto device_info = transfer_funcs[device_type](device_context);
+    if (device_info == nullptr) {
+      MS_LOG(ERROR) << "transfer device context to device info failed.";
+      return nullptr;
+    }
+    if (device_type == DT_CPU) {
+      ms_context->SetThreadAffinity(device_context.device_info_.cpu_device_info_.cpu_bind_mode_);
+    }
+    device_infos.push_back(device_info);
+  }
+  return ms_context;
+}
+
+std::set<std::string> ProvidersFromMSContext(const mindspore::Context *context) {
+  std::set<std::string> providers;
+  if (context == nullptr) {
+    return providers;
+  }
+  auto &device_infos = const_cast<mindspore::Context *>(context)->MutableDeviceInfo();
+  for (auto &device_info : device_infos) {
+    providers.emplace(device_info->GetProvider());
+  }
+  return providers;
+}
+}  // namespace lite
+}  // namespace mindspore

mindspore/lite/src/common/context_util.h

@@ -0,0 +1,31 @@
+/**
+ * Copyright 2021 Huawei Technologies Co., Ltd
+ *
+ * Licensed under the Apache License, Version 2.0 (the "License");
+ * you may not use this file except in compliance with the License.
+ * You may obtain a copy of the License at
+ *
+ * http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS,
+ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License.
+ */
+
+#ifndef MINDSPORE_LITE_SRC_COMMON_CONTEXT_UTIL_H_
+#define MINDSPORE_LITE_SRC_COMMON_CONTEXT_UTIL_H_
+
+#include <set>
+#include <string>
+#include "include/context.h"
+#include "include/api/context.h"
+
+namespace mindspore {
+namespace lite {
+mindspore::Context *MSContextFromContext(const lite::Context *context);
+std::set<std::string> ProvidersFromMSContext(const mindspore::Context *context);
+}  // namespace lite
+}  // namespace mindspore
+#endif  // MINDSPORE_LITE_SRC_COMMON_CONTEXT_UTIL_H_

mindspore/lite/src/common/string_util.h

@@ -20,7 +20,7 @@
 #include <vector>
 #include <string>
 #include <utility>
-#include "mindspore/lite/src/tensor.h"
+#include "src/tensor.h"
 #include "src/common/log_adapter.h"
 #include "tools/common/option.h"
 #include "include/errorcode.h"

mindspore/lite/src/common/tensor_util.cc

@@ -270,5 +270,15 @@ int CheckTensorsInvalid(const std::vector<Tensor *> &tensors) {
   }
   return RET_OK;
 }
+
+std::vector<mindspore::MSTensor> LiteTensorsToMSTensors(const std::vector<lite::Tensor *> &lite_tensors) {
+  std::vector<mindspore::MSTensor> tensors;
+  std::transform(lite_tensors.begin(), lite_tensors.end(), std::back_inserter(tensors), [](lite::Tensor *tensor) {
+    return mindspore::MSTensor(std::make_shared<mindspore::MSTensor::Impl>(tensor));
+  });
+
+  return tensors;
+}
+
 }  // namespace lite
 }  // namespace mindspore

mindspore/lite/src/common/tensor_util.h

@@ -17,10 +17,13 @@
 #ifndef MINDSPORE_LITE_SRC_COMMON_TENSOR_UTIL_H_
 #define MINDSPORE_LITE_SRC_COMMON_TENSOR_UTIL_H_
 #include <vector>
+
+#include <memory>
 #include "src/tensor.h"
 #include "src/tensorlist.h"
 #include "nnacl/tensor_c.h"
 #include "nnacl/infer/common_infer.h"
+#include "src/cxx_api/tensor/tensor_impl.h"
 
 namespace mindspore {
 namespace lite {
@@ -40,6 +43,9 @@ int GenerateOutTensorC(const OpParameter *const parameter, const std::vector<lit
                        const std::vector<lite::Tensor *> &outputs, std::vector<TensorC *> *out_tensor_c);
 
 int CheckTensorsInvalid(const std::vector<Tensor *> &tensors);
+
+std::vector<mindspore::MSTensor> LiteTensorsToMSTensors(const std::vector<lite::Tensor *> &lite_tensors);
+
 }  // namespace lite
 }  // namespace mindspore
 

mindspore/lite/src/cxx_api/context.cc

@@ -40,6 +40,7 @@ struct Context::Data {
   bool enable_parallel_ = false;
   std::vector<int32_t> affinity_core_list_;
   int affinity_mode_ = 2;
+  std::shared_ptr<Delegate> delegate = nullptr;
 };
 
 struct DeviceInfoContext::Data {
@@ -100,6 +101,7 @@ bool Context::GetEnableParallel() const {
     MS_LOG(ERROR) << "Invalid context.";
     return false;
   }
+
   return data_->enable_parallel_;
 }
 
@@ -137,6 +139,22 @@ std::vector<int32_t> Context::GetThreadAffinityCoreList() const {
   return data_->affinity_core_list_;
 }
 
+void Context::SetDelegate(const std::shared_ptr<Delegate> &delegate) {
+  if (data_ == nullptr) {
+    MS_LOG(ERROR) << "Invalid context.";
+    return;
+  }
+  data_->delegate = delegate;
+}
+
+std::shared_ptr<Delegate> Context::GetDelegate() const {
+  if (data_ == nullptr) {
+    MS_LOG(ERROR) << "Invalid context.";
+    return nullptr;
+  }
+  return data_->delegate;
+}
+
 std::vector<std::shared_ptr<DeviceInfoContext>> &Context::MutableDeviceInfo() {
   static std::vector<std::shared_ptr<DeviceInfoContext>> empty;
   if (data_ == nullptr) {

mindspore/lite/src/cxx_api/converters.cc

@@ -82,7 +82,7 @@ Status A2L_ConvertContext(Context *a_context, lite::Context *l_context) {
       return kLiteInputParamInvalid;
     }
   }
-
+  l_context->delegate = a_context->GetDelegate();
   return kSuccess;
 }
 }  // namespace mindspore

mindspore/lite/src/cxx_api/model/model_impl.cc

@@ -132,6 +132,7 @@ Status ModelImpl::RunGraph(const MSKernelCallBack &before, const MSKernelCallBac
     mscall_param.node_type_ = call_param.node_type;
     return before(inputs, outputs, mscall_param);
   };
+
   auto after_call_back = [&](const std::vector<mindspore::tensor::MSTensor *> &before_inputs,
                              const std::vector<mindspore::tensor::MSTensor *> &before_outputs,
                              const CallBackParam &call_param) {

mindspore/lite/src/cxx_api/model/model_impl.h

@@ -100,7 +100,6 @@ class ModelImpl {
   void SetGraph(const std::shared_ptr<Graph> &graph) { graph_ = graph; }
   void SetContext(const std::shared_ptr<Context> &context) { context_ = context; }
   void SetConfig(const std::shared_ptr<TrainCfg> cfg) { cfg_ = cfg; }
-  lite::CpuBindMode GetCpuBindMode();
   Status RunGraph(const MSKernelCallBack &before, const MSKernelCallBack &after);
 };
 }  // namespace mindspore

mindspore/lite/src/cxx_api/tensor/tensor_impl.cc

@@ -14,6 +14,7 @@
  * limitations under the License.
  */
 
+#include "src/cxx_api/tensor/tensor_impl.h"
 #include <cstddef>
 #include <numeric>
 #include <memory>
@@ -21,15 +22,8 @@
 #include <string>
 #include <vector>
 #include <functional>
-#include "src/cxx_api/tensor/tensor_impl.h"
 #include "src/cxx_api/tensor_utils.h"
-#include "include/api/types.h"
-#include "include/api/status.h"
-#include "include/ms_tensor.h"
-#include "src/common/string_util.h"
 #include "src/tensor.h"
-#include "src/common/log_adapter.h"
-#include "ir/dtype/type_id.h"
 
 namespace mindspore {
 using mindspore::lite::RET_OK;
@@ -37,7 +31,12 @@ using mindspore::lite::RET_OK;
 std::shared_ptr<MSTensor::Impl> MSTensor::Impl::CreateTensorImpl(const std::string &name, enum DataType type,
                                                                  const std::vector<int64_t> &shape, const void *data,
                                                                  size_t data_len) {
-  std::vector<int32_t> truncated_shape = TruncateShape(shape, static_cast<enum TypeId>(type), data_len, true);
+  std::vector<int32_t> truncated_shape;
+  if (data_len == 0) {
+    truncated_shape = TruncateShape(shape, static_cast<enum TypeId>(type), data_len, false);
+  } else {
+    truncated_shape = TruncateShape(shape, static_cast<enum TypeId>(type), data_len, true);
+  }
   if (truncated_shape.empty() && !(shape.empty())) {
     MS_LOG(ERROR) << "Invalid shape for creating tensor.";
     return nullptr;
@@ -80,4 +79,14 @@ std::shared_ptr<MSTensor::Impl> MSTensor::Impl::StringsToTensorImpl(const std::s
   impl->set_from_session(false);
   return impl;
 }
+
+std::vector<std::string> MSTensor::Impl::TensorImplToStrings(const std::shared_ptr<Impl> &impl) {
+  std::vector<std::string> empty;
+  auto lite_tensor = impl->lite_tensor();
+  if (lite_tensor == nullptr) {
+    MS_LOG(ERROR) << "Invalid tensor impl.";
+    return empty;
+  }
+  return lite::MSTensorToStrings(lite_tensor);
+}
 }  // namespace mindspore

mindspore/lite/src/cxx_api/tensor/tensor_impl.h

@@ -26,9 +26,9 @@
 #include <functional>
 #include "include/api/types.h"
 #include "include/api/status.h"
+#include "include/errorcode.h"
 #include "include/lite_utils.h"
 #include "include/ms_tensor.h"
-#include "src/tensor.h"
 #include "src/common/log_adapter.h"
 
 namespace mindspore {
@@ -38,7 +38,7 @@ class MSTensor::Impl {
  public:
   Impl() {}
 
-  virtual ~Impl() {
+  ~Impl() {
     if (lite_tensor_ == nullptr) {
       return;
     }
@@ -57,22 +57,15 @@ class MSTensor::Impl {
     }
   }
 
-  static std::shared_ptr<Impl> CreateTensorImpl(const std::string &name, enum DataType type,
-                                                const std::vector<int64_t> &shape, const void *data, size_t data_len);
+  static std::shared_ptr<Impl> MS_API CreateTensorImpl(const std::string &name, enum DataType type,
+                                                       const std::vector<int64_t> &shape, const void *data,
+                                                       size_t data_len);
 
-  static std::shared_ptr<Impl> StringsToTensorImpl(const std::string &name, const std::vector<std::string> &str);
+  static std::shared_ptr<Impl> MS_API StringsToTensorImpl(const std::string &name, const std::vector<std::string> &str);
 
-  static std::vector<std::string> TensorImplToStrings(const std::shared_ptr<Impl> &impl) {
-    std::vector<std::string> empty;
-    auto lite_tensor = impl->lite_tensor();
-    if (lite_tensor == nullptr) {
-      MS_LOG(ERROR) << "Invalid tensor impl.";
-      return empty;
-    }
-    return lite::MSTensorToStrings(lite_tensor);
-  }
+  static std::vector<std::string> MS_API TensorImplToStrings(const std::shared_ptr<Impl> &impl);
 
-  virtual const std::string &Name() const {
+  const std::string &Name() const {
     static std::string empty = "";
     if (lite_tensor_ == nullptr) {
       MS_LOG(ERROR) << "Invalid tensor.";
@@ -81,7 +74,15 @@ class MSTensor::Impl {
     return tensor_name_;
   }
 
-  virtual enum DataType DataType() const {
+  void SetName(const std::string &name) {
+    if (lite_tensor_ == nullptr) {
+      MS_LOG(ERROR) << "Invalid tensor.";
+      return;
+    }
+    lite_tensor_->set_tensor_name(name);
+  }
+
+  enum DataType DataType() const {
     if (lite_tensor_ == nullptr) {
       MS_LOG(ERROR) << "Invalid tensor.";
       return DataType::kTypeUnknown;
@@ -89,6 +90,14 @@ class MSTensor::Impl {
     return static_cast<enum DataType>(lite_tensor_->data_type());
   }
 
+  void SetDataType(enum DataType data_type) {
+    if (lite_tensor_ == nullptr) {
+      MS_LOG(ERROR) << "Invalid tensor.";
+      return;
+    }
+    lite_tensor_->set_data_type(static_cast<enum TypeId>(data_type));
+  }
+
   int64_t ElementNum() const {
     if (lite_tensor_ == nullptr) {
       MS_LOG(ERROR) << "Invalid tensor.";
@@ -97,7 +106,7 @@ class MSTensor::Impl {
     return static_cast<int64_t>(lite_tensor_->ElementsNum());
   }
 
-  virtual const std::vector<int64_t> &Shape() {
+  const std::vector<int64_t> &Shape() {
     static std::vector<int64_t> empty;
     if (lite_tensor_ == nullptr) {
       MS_LOG(ERROR) << "Invalid tensor.";
@@ -109,7 +118,50 @@ class MSTensor::Impl {
     return shape_;
   }
 
-  virtual std::shared_ptr<const void> Data() const {
+  void SetShape(const std::vector<int64_t> &shape) {
+    if (lite_tensor_ == nullptr) {
+      MS_LOG(ERROR) << "Invalid tensor.";
+      return;
+    }
+    std::vector<int> tensor_shape;
+    tensor_shape.resize(shape.size());
+    std::transform(shape.begin(), shape.end(), tensor_shape.begin(), [](int64_t c) { return static_cast<int>(c); });
+    lite_tensor_->set_shape(tensor_shape);
+  }
+
+  std::shared_ptr<Allocator> allocator() const {
+    if (lite_tensor_ == nullptr) {
+      MS_LOG(ERROR) << "Invalid tensor.";
+      return nullptr;
+    }
+    return lite_tensor_->allocator();
+  }
+
+  void SetAllocator(std::shared_ptr<Allocator> allocator) {
+    if (lite_tensor_ == nullptr) {
+      MS_LOG(ERROR) << "Invalid tensor.";
+      return;
+    }
+    lite_tensor_->set_allocator(allocator);
+  }
+
+  mindspore::Format format() {
+    if (lite_tensor_ == nullptr) {
+      MS_LOG(ERROR) << "Invalid tensor.";
+      return mindspore::Format::NHWC;
+    }
+    return lite_tensor_->format();
+  }
+
+  void SetFormat(mindspore::Format format) {
+    if (lite_tensor_ == nullptr) {
+      MS_LOG(ERROR) << "Invalid tensor.";
+      return;
+    }
+    lite_tensor_->set_format(format);
+  }
+
+  std::shared_ptr<const void> Data() const {
     if (lite_tensor_ == nullptr) {
       MS_LOG(ERROR) << "Invalid tensor.";
       return nullptr;
@@ -123,14 +175,15 @@ class MSTensor::Impl {
     return std::shared_ptr<const void>(lite_tensor_->data(), [](const void *) {});
   }
 
-  virtual void *MutableData() {
+  void *MutableData() {
     if (lite_tensor_ == nullptr) {
       MS_LOG(ERROR) << "Invalid tensor.";
       return nullptr;
     }
     return lite_tensor_->MutableData();
   }
-  virtual size_t DataSize() const {
+
+  size_t DataSize() const {
     if (lite_tensor_ == nullptr) {
       MS_LOG(ERROR) << "Invalid tensor.";
       return 0;
@@ -138,7 +191,15 @@ class MSTensor::Impl {
     return lite_tensor_->Size();
   }
 
-  virtual bool IsDevice() const { return false; }
+  void SetData(void *data) {
+    if (lite_tensor_ == nullptr) {
+      MS_LOG(ERROR) << "Invalid tensor.";
+      return;
+    }
+    lite_tensor_->set_data(data);
+  }
+
+  bool IsDevice() const { return false; }
 
   tensor::MSTensor *lite_tensor() const { return lite_tensor_; }
 

mindspore/lite/src/cxx_api/tensor_utils.cc

@@ -16,6 +16,7 @@
 
 #include "src/cxx_api/tensor_utils.h"
 #include "src/common/log_adapter.h"
+#include "src/tensor.h"
 
 namespace mindspore {
 std::vector<int32_t> TruncateShape(const std::vector<int64_t> &shape, enum TypeId type, size_t data_len,

mindspore/lite/src/cxx_api/tensor_utils.h

@@ -26,11 +26,12 @@
 #include "src/cxx_api/tensor/tensor_impl.h"
 
 namespace mindspore {
-std::vector<int32_t> TruncateShape(const std::vector<int64_t> &shape, enum TypeId type, size_t data_len,
-                                   bool verify_size);
-Status LiteTensorToMSTensor(tensor::MSTensor *srcTensor, MSTensor *dstTensor);
+std::vector<int32_t> MS_API TruncateShape(const std::vector<int64_t> &shape, enum TypeId type, size_t data_len,
+                                          bool verify_size);
 
-std::vector<MSTensor> LiteTensorsToMSTensors(const std::vector<mindspore::tensor::MSTensor *> &srcTensors);
+Status MS_API LiteTensorToMSTensor(tensor::MSTensor *srcTensor, MSTensor *dstTensor);
+
+std::vector<MSTensor> MS_API LiteTensorsToMSTensors(const std::vector<mindspore::tensor::MSTensor *> &srcTensors);
 
 }  // namespace mindspore
 

mindspore/lite/src/cxx_api/types.cc

@@ -21,11 +21,12 @@
 #include "include/api/status.h"
 #include "include/api/dual_abi_helper.h"
 #include "src/cxx_api/tensor/tensor_impl.h"
-#include "src/common/string_util.h"
-#include "src/tensor.h"
 #include "src/common/log_adapter.h"
 
 namespace mindspore {
+namespace {
+constexpr int64_t MAX_MALLOC_SIZE = static_cast<size_t>(2000) * 1024 * 1024;
+}
 class Buffer::Impl {
  public:
   Impl() : data_() { MS_LOG(ERROR) << "Unsupported feature."; }
@@ -71,28 +72,37 @@ bool MSTensor::operator==(std::nullptr_t) const { return impl_ == nullptr; }
 
 bool MSTensor::operator!=(std::nullptr_t) const { return impl_ != nullptr; }
 
+bool MSTensor::operator==(const MSTensor &tensor) const { return impl_->lite_tensor() == tensor.impl_->lite_tensor(); }
+
 MSTensor *MSTensor::CreateTensor(const std::vector<char> &name, enum DataType type, const std::vector<int64_t> &shape,
                                  const void *data, size_t data_len) noexcept {
   if (data_len < 0 || data_len > MAX_MALLOC_SIZE) {
     MS_LOG(ERROR) << "data_len is error.";
     return nullptr;
   }
-  auto new_data = malloc(data_len);
-  if (new_data == nullptr) {
-    MS_LOG(ERROR) << "Allocate data failed.";
-    return nullptr;
+  void *new_data = nullptr;
+  if (data != nullptr) {
+    new_data = malloc(data_len);
+    if (new_data == nullptr) {
+      MS_LOG(ERROR) << "Allocate data failed.";
+      return nullptr;
+    }
+    ::memcpy(new_data, data, data_len);
   }
-  ::memcpy(new_data, data, data_len);
   auto impl = Impl::CreateTensorImpl(CharToString(name), type, shape, new_data, data_len);
   if (impl == nullptr) {
     MS_LOG(ERROR) << "Allocate tensor impl failed.";
-    free(new_data);
+    if (new_data != nullptr) {
+      free(new_data);
+    }
     return nullptr;
   }
   auto ms_tensor = new (std::nothrow) MSTensor(impl);
   if (ms_tensor == nullptr) {
     MS_LOG(ERROR) << "Allocate tensor impl failed.";
-    free(new_data);
+    if (new_data != nullptr) {
+      free(new_data);
+    }
     return nullptr;
   }
   impl->set_own_data(true);
@@ -172,7 +182,7 @@ MSTensor *MSTensor::Clone() const {
 
 std::vector<char> MSTensor::CharName() const {
   if (impl_ == nullptr) {
-    MS_LOG(ERROR) << "Invalid tensor inpmlement.";
+    MS_LOG(ERROR) << "Invalid tensor implement.";
     return std::vector<char>();
   }
   return StringToChar(impl_->Name());
@@ -180,7 +190,7 @@ std::vector<char> MSTensor::CharName() const {
 
 int64_t MSTensor::ElementNum() const {
   if (impl_ == nullptr) {
-    MS_LOG(ERROR) << "Invalid tensor inpmlement.";
+    MS_LOG(ERROR) << "Invalid tensor implement.";
     return -1;
   }
   return impl_->ElementNum();
@@ -188,7 +198,7 @@ int64_t MSTensor::ElementNum() const {
 
 enum DataType MSTensor::DataType() const {
   if (impl_ == nullptr) {
-    MS_LOG(ERROR) << "Invalid tensor inpmlement.";
+    MS_LOG(ERROR) << "Invalid tensor implement.";
     return DataType::kTypeUnknown;
   }
   return impl_->DataType();
@@ -197,7 +207,7 @@ enum DataType MSTensor::DataType() const {
 const std::vector<int64_t> &MSTensor::Shape() const {
   static std::vector<int64_t> empty;
   if (impl_ == nullptr) {
-    MS_LOG(ERROR) << "Invalid tensor inpmlement.";
+    MS_LOG(ERROR) << "Invalid tensor implement.";
     return empty;
   }
   return impl_->Shape();
@@ -205,7 +215,7 @@ const std::vector<int64_t> &MSTensor::Shape() const {
 
 std::shared_ptr<const void> MSTensor::Data() const {
   if (impl_ == nullptr) {
-    MS_LOG(ERROR) << "Invalid tensor inpmlement.";
+    MS_LOG(ERROR) << "Invalid tensor implement.";
     return nullptr;
   }
   return impl_->Data();
@@ -213,7 +223,7 @@ std::shared_ptr<const void> MSTensor::Data() const {
 
 void *MSTensor::MutableData() {
   if (impl_ == nullptr) {
-    MS_LOG(ERROR) << "Invalid tensor inpmlement.";
+    MS_LOG(ERROR) << "Invalid tensor implement.";
     return nullptr;
   }
   return impl_->MutableData();
@@ -221,7 +231,7 @@ void *MSTensor::MutableData() {
 
 size_t MSTensor::DataSize() const {
   if (impl_ == nullptr) {
-    MS_LOG(ERROR) << "Invalid tensor inpmlement.";
+    MS_LOG(ERROR) << "Invalid tensor implement.";
     return 0;
   }
   return impl_->DataSize();
@@ -238,6 +248,70 @@ void MSTensor::DestroyTensorPtr(MSTensor *tensor) noexcept {
   }
 }
 
+void MSTensor::SetShape(const std::vector<int64_t> &shape) {
+  if (impl_ == nullptr) {
+    MS_LOG(ERROR) << "Invalid tensor implement.";
+    return;
+  }
+  impl_->SetShape(shape);
+}
+
+void MSTensor::SetDataType(enum DataType data_type) {
+  if (impl_ == nullptr) {
+    MS_LOG(ERROR) << "Invalid tensor implement.";
+    return;
+  }
+  impl_->SetDataType(data_type);
+}
+
+void MSTensor::SetTensorName(const std::string &name) {
+  if (impl_ == nullptr) {
+    MS_LOG(ERROR) << "Invalid tensor implement.";
+    return;
+  }
+  impl_->SetName(name);
+}
+
+void MSTensor::SetAllocator(std::shared_ptr<Allocator> allocator) {
+  if (impl_ == nullptr) {
+    MS_LOG(ERROR) << "Invalid tensor implement.";
+    return;
+  }
+  return impl_->SetAllocator(allocator);
+}
+
+std::shared_ptr<Allocator> MSTensor::allocator() const {
+  if (impl_ == nullptr) {
+    MS_LOG(ERROR) << "Invalid tensor implement.";
+    return nullptr;
+  }
+  return impl_->allocator();
+}
+
+void MSTensor::SetFormat(mindspore::Format format) {
+  if (impl_ == nullptr) {
+    MS_LOG(ERROR) << "Invalid tensor implement.";
+    return;
+  }
+  return impl_->SetFormat(format);
+}
+
+mindspore::Format MSTensor::format() const {
+  if (impl_ == nullptr) {
+    MS_LOG(ERROR) << "Invalid tensor implement.";
+    return mindspore::Format::NHWC;
+  }
+  return impl_->format();
+}
+
+void MSTensor::SetData(void *data) {
+  if (impl_ == nullptr) {
+    MS_LOG(ERROR) << "Invalid tensor implement.";
+    return;
+  }
+  return impl_->SetData(data);
+}
+
 Buffer::Buffer() : impl_(nullptr) { MS_LOG(ERROR) << "Unsupported feature."; }
 Buffer::Buffer(const void *data, size_t data_len) : impl_(nullptr) { MS_LOG(ERROR) << "Unsupported feature."; }
 Buffer::~Buffer() = default;

mindspore/lite/src/delegate/delegate.cc

@@ -14,7 +14,7 @@
  * limitations under the License.
  */
 
-#include "include/delegate.h"
+#include "include/api/delegate.h"
 namespace mindspore {
 const schema::Primitive *DelegateModel::GetPrimitive(kernel::Kernel *kernel) const {
   if (primitives_.find(kernel) != primitives_.end()) {

mindspore/lite/src/delegate/delegate_utils.cc

@@ -16,7 +16,7 @@
 
 #include "src/delegate/delegate_utils.h"
 namespace mindspore::lite {
-bool IsSubGraphInputTensor(const std::vector<mindspore::tensor::MSTensor *> &inputs, tensor::MSTensor *input) {
+bool IsSubGraphInputTensor(const std::vector<mindspore::MSTensor> &inputs, mindspore::MSTensor input) {
   if (find(inputs.begin(), inputs.end(), input) != inputs.end()) {
     return true;
   }

mindspore/lite/src/delegate/delegate_utils.h

@@ -17,17 +17,17 @@
 #define MINDSPORE_LITE_SRC_DELEGATE_DELEGATE_UTILS
 #include <vector>
 #include "include/ms_tensor.h"
-#include "include/delegate.h"
+#include "include/api/delegate.h"
 #include "src/common/log_adapter.h"
 #include "src/delegate/tensorrt/op/tensorrt_op.h"
 
 namespace mindspore::lite {
-bool IsSubGraphInputTensor(const std::vector<mindspore::tensor::MSTensor *> &inputs, tensor::MSTensor *input);
+bool IsSubGraphInputTensor(const std::vector<mindspore::MSTensor> &inputs, mindspore::MSTensor input);
 
 template <typename T>
-std::vector<mindspore::tensor::MSTensor *> GetGraphInTensors(std::vector<T *> ops) {
-  std::vector<mindspore::tensor::MSTensor *> inputs;
-  auto is_op_output = [&](tensor::MSTensor *tensor) -> bool {
+std::vector<mindspore::MSTensor> GetGraphInTensors(std::vector<T *> ops) {
+  std::vector<mindspore::MSTensor> inputs;
+  auto is_op_output = [&](mindspore::MSTensor tensor) -> bool {
     for (auto op : ops) {
       auto out_tensors = op->outputs();
       if (find(out_tensors.begin(), out_tensors.end(), tensor) != out_tensors.end()) {
@@ -39,7 +39,7 @@ std::vector<mindspore::tensor::MSTensor *> GetGraphInTensors(std::vector<T *> op
 
   for (auto op : ops) {
     for (auto in_tensor : op->inputs()) {
-      if (in_tensor->data() == nullptr && !is_op_output(in_tensor)) {
+      if (in_tensor.Data() == nullptr && !is_op_output(in_tensor)) {
         inputs.push_back(in_tensor);
       }
     }
@@ -48,9 +48,9 @@ std::vector<mindspore::tensor::MSTensor *> GetGraphInTensors(std::vector<T *> op
 }
 
 template <typename T>
-std::vector<mindspore::tensor::MSTensor *> GetGraphOutTensors(const std::vector<T *> &ops) {
-  std::vector<mindspore::tensor::MSTensor *> outputs;
-  auto is_op_input = [&](const tensor::MSTensor *tensor) -> bool {
+std::vector<mindspore::MSTensor> GetGraphOutTensors(const std::vector<T *> &ops) {
+  std::vector<mindspore::MSTensor> outputs;
+  auto is_op_input = [&](const mindspore::MSTensor tensor) -> bool {
     for (auto op : ops) {
       auto in_tensors = op->inputs();
       if (find(in_tensors.begin(), in_tensors.end(), tensor) != in_tensors.end()) {
@@ -86,13 +86,13 @@ std::vector<mindspore::tensor::MSTensor *> GetGraphOutTensors(const std::vector<
 }
 
 template <typename T>
-std::vector<tensor::MSTensor *> GraphInTensors(const std::vector<T *> &ops, DelegateModel *model, KernelIter from,
-                                               KernelIter end) {
+std::vector<mindspore::MSTensor> GraphInTensors(const std::vector<T *> &ops, DelegateModel *model, KernelIter from,
+                                                KernelIter end) {
   auto in_tensors = GetGraphInTensors(ops);
-  std::vector<tensor::MSTensor *> all_in_tensors;
+  std::vector<mindspore::MSTensor> all_in_tensors;
   for (auto op : ops) {
     for (auto in_tensor : op->inputs()) {
-      if (in_tensor->data() != nullptr && find(in_tensors.begin(), in_tensors.end(), in_tensor) == in_tensors.end()) {
+      if (in_tensor.Data() != nullptr && find(in_tensors.begin(), in_tensors.end(), in_tensor) == in_tensors.end()) {
         all_in_tensors.push_back(in_tensor);
       }
     }
@@ -113,10 +113,10 @@ std::vector<tensor::MSTensor *> GraphInTensors(const std::vector<T *> &ops, Dele
 }
 
 template <typename T>
-std::vector<tensor::MSTensor *> GraphOutTensors(const std::vector<T *> &ops, DelegateModel *model, KernelIter from,
-                                                KernelIter end) {
+std::vector<mindspore::MSTensor> GraphOutTensors(const std::vector<T *> &ops, DelegateModel *model, KernelIter from,
+                                                 KernelIter end) {
   auto out_tensors = GetGraphOutTensors(ops);
-  std::vector<tensor::MSTensor *> all_out_tensors;
+  std::vector<mindspore::MSTensor> all_out_tensors;
   for (auto op : ops) {
     for (auto out_tensor : op->outputs()) {
       if (find(out_tensors.begin(), out_tensors.end(), out_tensor) == out_tensors.end()) {
@@ -176,9 +176,8 @@ void FindPreNextOps(std::vector<T *> all_ops) {
 }
 
 template <typename T>
-int GetGraphInOutOps(const std::vector<mindspore::tensor::MSTensor *> &inputs,
-                     const std::vector<mindspore::tensor::MSTensor *> &outputs, std::vector<T *> *in_ops,
-                     std::vector<T *> *out_ops, const std::vector<T *> &all_ops) {
+int GetGraphInOutOps(const std::vector<mindspore::MSTensor> &inputs, const std::vector<mindspore::MSTensor> &outputs,
+                     std::vector<T *> *in_ops, std::vector<T *> *out_ops, const std::vector<T *> &all_ops) {
   for (auto in_tensor : inputs) {
     for (auto op : all_ops) {
       if (find(op->inputs().begin(), op->inputs().end(), in_tensor) != op->inputs().end() &&

mindspore/lite/src/delegate/npu/npu_converter_utils.cc

@@ -15,7 +15,6 @@
  */
 
 #include "src/delegate/npu/npu_converter_utils.h"
-#include <arm_neon.h>
 #include "src/common/log_adapter.h"
 namespace mindspore {
 #define C8NUM 8
@@ -54,7 +53,7 @@ void Float16ToFloat32(const float16_t *__restrict input, float *__restrict outpu
 }
 #endif
 
-ge::Shape ConverterToNPUShape(const std::vector<int> &src_shape) {
+ge::Shape ConverterToNPUShape(const std::vector<int64_t> &src_shape) {
   vector<int64_t> shapes;
   shapes.reserve(src_shape.size());
   for (int i = 0; i < src_shape.size(); i++) {
@@ -82,27 +81,26 @@ ge::Format ConverterToNPUFormat(schema::Format format) {
   return ge_format;
 }
 
-ge::DataType ConverterToNPUDataType(TypeId type_id) {
+ge::DataType ConverterToNPUDataType(DataType type_id) {
   ge::DataType data_type;
   switch (type_id) {
-    case kNumberTypeFloat:
-    case kNumberTypeFloat32:
-    case kNumberTypeFloat16:
+    case DataType::kNumberTypeFloat32:
+    case DataType::kNumberTypeFloat16:
       data_type = ge::DT_FLOAT;
       break;
-    case kNumberTypeInt8:
+    case DataType::kNumberTypeInt8:
       data_type = ge::DT_INT8;
       break;
-    case kNumberTypeUInt8:
+    case DataType::kNumberTypeUInt8:
       data_type = ge::DT_UINT8;
       break;
-    case kNumberTypeInt16:
+    case DataType::kNumberTypeInt16:
       data_type = ge::DT_INT16;
       break;
-    case kNumberTypeInt32:
+    case DataType::kNumberTypeInt32:
       data_type = ge::DT_INT32;
       break;
-    case kNumberTypeUInt32:
+    case DataType::kNumberTypeUInt32:
       data_type = ge::DT_UINT32;
       break;
     default:
@@ -112,43 +110,41 @@ ge::DataType ConverterToNPUDataType(TypeId type_id) {
   return data_type;
 }
 
-hiai::op::Data *ConverterToNPUData(tensor::MSTensor *src, const std::string &name) {
+hiai::op::Data *ConverterToNPUData(mindspore::MSTensor src, const std::string &name) {
   auto data = new (std::nothrow) hiai::op::Data(name);
   if (data == nullptr) {
     MS_LOG(ERROR) << "new data failed.";
     return data;
   }
-  ge::TensorDesc tensor_desc(ConverterToNPUShape(src->shape()), ge::FORMAT_NCHW,
-                             ConverterToNPUDataType(src->data_type()));
+  ge::TensorDesc tensor_desc(ConverterToNPUShape(src.Shape()), ge::FORMAT_NCHW, ConverterToNPUDataType(src.DataType()));
   data->update_input_desc_x(tensor_desc);
   return data;
 }
 
-std::shared_ptr<ge::Tensor> ConverterToNPUTensor(tensor::MSTensor *src) {
+std::shared_ptr<ge::Tensor> ConverterToNPUTensor(mindspore::MSTensor src) {
   std::shared_ptr<ge::Tensor> ge_tensor = std::shared_ptr<ge::Tensor>(new (std::nothrow) ge::Tensor());
   if (ge_tensor == nullptr) {
     MS_LOG(ERROR) << "new ge_tensor failed.";
     return nullptr;
   }
-  ge::TensorDesc tensor_desc(ConverterToNPUShape(src->shape()), ge::FORMAT_NCHW,
-                             ConverterToNPUDataType(src->data_type()));
+  ge::TensorDesc tensor_desc(ConverterToNPUShape(src.Shape()), ge::FORMAT_NCHW, ConverterToNPUDataType(src.DataType()));
 
   ge_tensor->SetTensorDesc(tensor_desc);
 
-  if (src->data() != nullptr) {
-    if (src->data_type() == kNumberTypeFloat16) {
+  if (src.Data() != nullptr) {
+    if (src.DataType() == DataType::kNumberTypeFloat16) {
 #ifdef ENABLE_ARM64
-      auto fp32_data = malloc(src->ElementsNum() * sizeof(float));
-      Float16ToFloat32(reinterpret_cast<float16_t *>(src->data()), reinterpret_cast<float *>(fp32_data),
-                       src->ElementsNum());
-      ge_tensor->SetData(reinterpret_cast<const uint8_t *>(fp32_data), src->ElementsNum() * sizeof(float));
+      auto fp32_data = malloc(src.ElementNum() * sizeof(float));
+      Float16ToFloat32(reinterpret_cast<float16_t *>(src.MutableData()), reinterpret_cast<float *>(fp32_data),
+                       src.ElementNum());
+      ge_tensor->SetData(reinterpret_cast<const uint8_t *>(fp32_data), src.ElementNum() * sizeof(float));
       free(fp32_data);
 #else
       MS_LOG(ERROR) << "This platform does not support fp16.";
       return nullptr;
 #endif
     } else {
-      ge_tensor->SetData(reinterpret_cast<const uint8_t *>(src->data()), src->Size());
+      ge_tensor->SetData(reinterpret_cast<const uint8_t *>(src.MutableData()), src.DataSize());
     }
   }
   return ge_tensor;
@@ -189,7 +185,7 @@ int TransFormAxis(int axis) {
   }
 }
 
-bool IsContainMSTensor(const std::vector<tensor::MSTensor *> &tensor_vec, const tensor::MSTensor *tensor) {
+bool IsContainMSTensor(const std::vector<mindspore::MSTensor> &tensor_vec, const mindspore::MSTensor tensor) {
   return find(tensor_vec.begin(), tensor_vec.end(), tensor) != tensor_vec.end();
 }
 }  // namespace mindspore

mindspore/lite/src/delegate/npu/npu_converter_utils.h

@@ -19,29 +19,36 @@
 #include <string>
 #include <memory>
 #include <vector>
+#ifdef ENABLE_ARM64
+#include <arm_neon.h>
+#endif
 #include "schema/ops_generated.h"
 #include "include/graph/tensor.h"
 #include "include/graph/op/array_defs.h"
-#include "include/ms_tensor.h"
+#include "include/api/types.h"
+#include "include/api/data_type.h"
 
 namespace mindspore {
+#ifdef ENABLE_ARM64
+void Float32ToFloat16(const float *__restrict input, float16_t *__restrict output, int number);
 
-std::shared_ptr<ge::Tensor> ConverterToNPUTensor(tensor::MSTensor *src);
+void Float16ToFloat32(const float16_t *__restrict input, float *__restrict output, int number);
+#endif
 
-hiai::op::Data *ConverterToNPUData(tensor::MSTensor *src, const std::string &name);
+std::shared_ptr<ge::Tensor> ConverterToNPUTensor(mindspore::MSTensor src);
 
-ge::Format ConverterToNPUFormat(schema::Format format);
+hiai::op::Data *ConverterToNPUData(mindspore::MSTensor src, const std::string &name);
 
-ge::DataType ConverterToNPUDataType(TypeId type_id);
+ge::Format ConverterToNPUFormat(schema::Format format);
 
-ge::Shape ConverterToNPUShape(const std::vector<int> &src_shape);
+ge::DataType ConverterToNPUDataType(DataType type_id);
 
-int ConverterToNPUActMode(schema::ActivationType type);
+ge::Shape ConverterToNPUShape(const std::vector<int64_t> &src_shape);
 
 int ConverterToNPUEltwiseMode(schema::EltwiseMode mode);
 
 int TransFormAxis(int axis);
 
-bool IsContainMSTensor(const std::vector<tensor::MSTensor *> &tensor_vec, const tensor::MSTensor *tensor);
+bool IsContainMSTensor(const std::vector<mindspore::MSTensor> &tensor_vec, const mindspore::MSTensor tensor);
 }  // namespace mindspore
 #endif  // MINDSPORE_LITE_SRC_RUNTIME_DELEGATE_NPU_NPU_CONVERTER_UITLS_H_

mindspore/lite/src/delegate/npu/npu_delegate.cc

@@ -16,6 +16,7 @@
 
 #include "src/delegate/npu/npu_delegate.h"
 #include <queue>
+#include "include/errorcode.h"
 #include "src/delegate/npu/op/npu_op.h"
 #include "src/delegate/npu/op/activation_npu.h"
 #include "src/delegate/npu/op/argmax_npu.h"
@@ -54,6 +55,9 @@
 #include "src/delegate/npu/pass/npu_insert_transform_pass.h"
 #include "src/delegate/npu/pass/npu_fusion_pass.h"
 
+using mindspore::lite::RET_ERROR;
+using mindspore::lite::RET_OK;
+
 namespace mindspore {
 NPUDelegate::~NPUDelegate() {
   if (npu_manager_ != nullptr) {
@@ -202,42 +206,42 @@ int NPUDelegate::Build(DelegateModel *model) {
 }
 
 NPUOp *NPUDelegate::GetOP(kernel::Kernel *kernel, const schema::Primitive *primitive) {
-  auto in_tensors = kernel->inputs();
-  auto out_tensors = kernel->outputs();
   auto name = kernel->name();
   NPUOp *npu_op = nullptr;
   auto node_type = primitive->value_type();
   if (node_type == schema::PrimitiveType_Conv2DFusion) {
-    npu_op = GetNPUConvOp(primitive, in_tensors, out_tensors, name);
+    npu_op = GetNPUConvOp(primitive, kernel->inputs(), kernel->outputs(), name);
   } else {
     if (op_func_lists_.find(node_type) != op_func_lists_.end()) {
-      npu_op = op_func_lists_[node_type](primitive, in_tensors, out_tensors, name);
+      npu_op = op_func_lists_[node_type](primitive, kernel->inputs(), kernel->outputs(), name);
     } else {
       MS_LOG(DEBUG) << "Unsupported op type for NPU.";
       return nullptr;
     }
   }
 
-  for (auto tensor : in_tensors) {
-    if (tensor->data_type() == kNumberTypeFloat16 && tensor->data() == nullptr) {
-      tensor->set_data_type(kNumberTypeFloat32);
+  for (int i = 0; i < kernel->inputs().size(); i++) {
+    mindspore::MSTensor tensor = kernel->inputs()[i];
+    if (tensor.DataType() == DataType::kNumberTypeFloat16 && tensor.Data() == nullptr) {
+      tensor.SetDataType(DataType::kNumberTypeFloat32);
     }
   }
-  for (auto tensor : out_tensors) {
-    if (tensor->data_type() == kNumberTypeFloat16) {
-      tensor->set_data_type(kNumberTypeFloat32);
+  for (int i = 0; i < kernel->outputs().size(); i++) {
+    mindspore::MSTensor tensor = kernel->outputs()[i];
+    if (tensor.DataType() == DataType::kNumberTypeFloat16) {
+      tensor.SetDataType(DataType::kNumberTypeFloat32);
     }
   }
   return npu_op;
 }
 
-std::vector<tensor::MSTensor *> GraphInTensors(const std::vector<NPUOp *> &ops, DelegateModel *model, KernelIter from,
-                                               KernelIter end) {
+std::vector<mindspore::MSTensor> GraphInTensors(const std::vector<NPUOp *> &ops, DelegateModel *model, KernelIter from,
+                                                KernelIter end) {
   auto in_tensors = NPUGraphUtils::GetGraphInTensors(ops);
-  std::vector<tensor::MSTensor *> all_in_tensors;
+  std::vector<mindspore::MSTensor> all_in_tensors;
   for (auto op : ops) {
     for (auto in_tensor : op->inputs()) {
-      if (in_tensor->data() != nullptr && find(in_tensors.begin(), in_tensors.end(), in_tensor) == in_tensors.end()) {
+      if (in_tensor.Data() != nullptr && find(in_tensors.begin(), in_tensors.end(), in_tensor) == in_tensors.end()) {
         all_in_tensors.push_back(in_tensor);
       }
     }
@@ -257,10 +261,10 @@ std::vector<tensor::MSTensor *> GraphInTensors(const std::vector<NPUOp *> &ops,
   return in_tensors;
 }
 
-std::vector<tensor::MSTensor *> GraphOutTensors(const std::vector<NPUOp *> &ops, DelegateModel *model, KernelIter from,
-                                                KernelIter end) {
+std::vector<mindspore::MSTensor> GraphOutTensors(const std::vector<NPUOp *> &ops, DelegateModel *model, KernelIter from,
+                                                 KernelIter end) {
   auto out_tensors = NPUGraphUtils::GetGraphOutTensors(ops);
-  std::vector<tensor::MSTensor *> all_out_tensors;
+  std::vector<mindspore::MSTensor> all_out_tensors;
   for (auto op : ops) {
     for (auto out_tensor : op->outputs()) {
       if (find(out_tensors.begin(), out_tensors.end(), out_tensor) == out_tensors.end()) {

mindspore/lite/src/delegate/npu/npu_delegate.h

@@ -19,16 +19,11 @@
 
 #include <vector>
 #include <map>
-#include "include/delegate.h"
+#include "include/api/delegate.h"
+#include "include/context.h"
 #include "src/delegate/npu/npu_manager.h"
 #include "src/delegate/npu/pass/npu_pass_manager.h"
-#include "src/delegate/npu/op//npu_op.h"
-#include "include/context.h"
-#include "include/errorcode.h"
-#include "src/common/log_adapter.h"
-
-using mindspore::lite::RET_ERROR;
-using mindspore::lite::RET_OK;
+#include "src/delegate/npu/op/npu_op.h"
 
 namespace mindspore {
 class NPUDelegate : public Delegate {

mindspore/lite/src/delegate/npu/npu_executor.cc

@@ -47,8 +47,8 @@ int NPUExecutor::Prepare() {
   return RET_OK;
 }
 
-std::vector<int> GetNpuTensorShape(int dim, std::shared_ptr<hiai::AiTensor> npu_tensor) {
-  std::vector<int> npu_shape;
+std::vector<int64_t> GetNpuTensorShape(int dim, std::shared_ptr<hiai::AiTensor> npu_tensor) {
+  std::vector<int64_t> npu_shape;
   if (dim > 0) {
     npu_shape.push_back(npu_tensor->GetTensorDimension().GetNumber());
   }
@@ -75,40 +75,40 @@ std::vector<int> ExpandShapeTo4d(const std::vector<int> &shape) {
   return ret;
 }
 
-bool IsSameShapeTensor(tensor::MSTensor *tensor, std::shared_ptr<hiai::AiTensor> npu_tensor) {
-  if (tensor->shape().size() > 4) {
+bool IsSameShapeTensor(mindspore::MSTensor tensor, std::shared_ptr<hiai::AiTensor> npu_tensor) {
+  if (tensor.Shape().size() > 4) {
     MS_LOG(ERROR) << "Npu does not support output tensor dims greater than 4";
     return false;
   }
-  return GetNpuTensorShape(tensor->shape().size(), npu_tensor) == tensor->shape();
+  return GetNpuTensorShape(tensor.Shape().size(), npu_tensor) == tensor.Shape();
 }
 
-int NPUExecutor::Run(const std::vector<tensor::MSTensor *> &in_tensors,
-                     const std::vector<tensor::MSTensor *> &out_tensors, const std::vector<NPUOp *> &in_ops) {
+int NPUExecutor::Run(const std::vector<mindspore::MSTensor> &in_tensors,
+                     const std::vector<mindspore::MSTensor> &out_tensors, const std::vector<NPUOp *> &in_ops) {
   hiai::AiContext context;
-  std::unordered_map<tensor::MSTensor *, int> tensor_uses;
+  std::unordered_map<std::string, int> tensor_uses;
   for (const auto op : in_ops) {
     for (const auto op_input : op->inputs()) {
-      if (tensor_uses.find(op_input) == tensor_uses.end()) {
-        tensor_uses.insert({op_input, 1});
+      if (tensor_uses.find(op_input.Name()) == tensor_uses.end()) {
+        tensor_uses.insert({op_input.Name(), 1});
       } else {
-        tensor_uses[op_input]++;
+        tensor_uses[op_input.Name()]++;
       }
     }
   }
   for (int i = 0; i < npu_input_tensors_.size(); ++i) {
     int index = 0;
     for (; index < in_tensors.size(); index++) {
-      if (tensor_uses[in_tensors[index]] > 0 && IsSameShapeTensor(in_tensors[index], npu_input_tensors_[i])) {
-        void *data = in_tensors[index]->data();
+      if (tensor_uses[in_tensors[index].Name()] > 0 && IsSameShapeTensor(in_tensors[index], npu_input_tensors_[i])) {
+        auto data = in_tensors[index].Data();
         if (data == nullptr) {
-          MS_LOG(ERROR) << "For " << model_name_ << ", the input tensor " << in_tensors[index]->tensor_name()
+          MS_LOG(ERROR) << "For " << model_name_ << ", the input tensor " << in_tensors[index].Name()
                         << " data is nullptr";
           return RET_ERROR;
         }
 
-        memcpy(npu_input_tensors_[i]->GetBuffer(), data, in_tensors[index]->Size());
-        tensor_uses[in_tensors[index]]--;
+        memcpy(npu_input_tensors_[i]->GetBuffer(), data.get(), in_tensors[index].DataSize());
+        tensor_uses[in_tensors[index].Name()]--;
         break;
       }
     }
@@ -135,9 +135,10 @@ int NPUExecutor::Run(const std::vector<tensor::MSTensor *> &in_tensors,
     int index = 0;
     for (; index < out_tensors.size(); index++) {
       if (!outputs_visited[index] && IsSameShapeTensor(out_tensors[index], npu_output_tensors_[i])) {
-        void *data = out_tensors[index]->data();
+        mindspore::MSTensor out_tensor = out_tensors[index];
+        auto data = out_tensor.MutableData();
         if (data == nullptr) {
-          MS_LOG(ERROR) << "For " << model_name_ << ", the output tensor " << in_tensors[index]->tensor_name()
+          MS_LOG(ERROR) << "For " << model_name_ << ", the output tensor " << out_tensors[index].Name()
                         << " data is nullptr";
           return RET_ERROR;
         }

mindspore/lite/src/delegate/npu/npu_executor.h

@@ -33,7 +33,7 @@ class NPUExecutor {
   ~NPUExecutor();
   int Prepare();
 
-  int Run(const std::vector<tensor::MSTensor *> &in_tensors, const std::vector<tensor::MSTensor *> &out_tensors,
+  int Run(const std::vector<mindspore::MSTensor> &in_tensors, const std::vector<mindspore::MSTensor> &out_tensors,
           const std::vector<NPUOp *> &in_ops);
 
  private:

mindspore/lite/src/delegate/npu/npu_graph.cc

@@ -28,12 +28,12 @@ NPUGraph::~NPUGraph() {
   for (auto *op : npu_ops_) {
     delete op;
   }
-  for (auto *tensor : insert_tensors_) {
+  for (auto tensor : insert_tensors_) {
     delete tensor;
   }
 }
 
-void NPUGraph::set_input(tensor::MSTensor *in_tensor, int index) {
+void NPUGraph::set_input(mindspore::MSTensor in_tensor, int index) {
   MS_ASSERT(index < inputs_.size());
   auto origin_tensor = this->inputs_[index];
   for (auto kernel : all_kernels_) {
@@ -46,7 +46,7 @@ void NPUGraph::set_input(tensor::MSTensor *in_tensor, int index) {
   this->inputs_[index] = in_tensor;
 }
 
-void NPUGraph::set_output(tensor::MSTensor *out_tensor, int index) {
+void NPUGraph::set_output(mindspore::MSTensor out_tensor, int index) {
   MS_ASSERT(index < outputs_.size());
   auto origin_tensor = this->outputs_[index];
   for (auto kernel : all_kernels_) {
@@ -199,7 +199,7 @@ int NPUGraph::Prepare() {
     }
     for (auto output : all_kernels_[i]->outputs()) {
       if (find(outputs_.begin(), outputs_.end(), output) == outputs_.end()) {
-        output->MutableData();
+        output.MutableData();
       }
     }
   }
@@ -211,7 +211,7 @@ int NPUGraph::Execute() {
     // 1. malloc graph output data
     for (auto output : all_kernels_[i]->outputs()) {
       if (find(outputs_.begin(), outputs_.end(), output) != outputs_.end()) {
-        output->MutableData();
+        output.MutableData();
       }
     }
     // 2. execute

mindspore/lite/src/delegate/npu/npu_graph.h

@@ -20,15 +20,15 @@
 #include <vector>
 #include <map>
 #include <utility>
-#include "include/kernel.h"
+#include "include/api/kernel.h"
 #include "src/delegate/npu/op/npu_op.h"
 #include "src/delegate/npu/npu_executor.h"
 
 namespace mindspore {
 class NPUGraph : public kernel::Kernel {
  public:
-  NPUGraph(std::vector<NPUOp *> npu_ops, NPUManager *npu_manager, const std::vector<tensor::MSTensor *> &inputs,
-           const std::vector<tensor::MSTensor *> &outputs)
+  NPUGraph(std::vector<NPUOp *> npu_ops, NPUManager *npu_manager, const std::vector<mindspore::MSTensor> &inputs,
+           const std::vector<mindspore::MSTensor> &outputs)
       : kernel::Kernel(inputs, outputs, nullptr, nullptr), npu_ops_(std::move(npu_ops)), npu_manager_(npu_manager) {}
 
   ~NPUGraph() override;
@@ -44,15 +44,15 @@ class NPUGraph : public kernel::Kernel {
     return lite::RET_ERROR;
   }
 
-  void set_input(tensor::MSTensor *in_tensor, int index) override;
+  void set_input(mindspore::MSTensor in_tensor, int index) override;
 
-  void set_output(tensor::MSTensor *out_tensor, int index) override;
+  void set_output(mindspore::MSTensor out_tensor, int index) override;
 
   int FindPreNextOps();
 
   std::vector<NPUOp *> *GetOps() { return &npu_ops_; }
 
-  std::vector<tensor::MSTensor *> *GetInsertTensors() { return &insert_tensors_; }
+  std::vector<mindspore::MSTensor *> *GetInsertTensors() { return &insert_tensors_; }
 
  protected:
   std::vector<NPUOp *> FindPreOps(NPUOp *cur_op);
@@ -69,7 +69,7 @@ class NPUGraph : public kernel::Kernel {
 
   std::vector<kernel::Kernel *> all_kernels_{};
 
-  std::vector<tensor::MSTensor *> insert_tensors_;
+  std::vector<mindspore::MSTensor *> insert_tensors_;
 
   NPUManager *npu_manager_ = nullptr;
 };

mindspore/lite/src/delegate/npu/npu_graph_utils.cc

@@ -16,9 +16,9 @@
 
 #include "src/delegate/npu/npu_graph_utils.h"
 namespace mindspore {
-std::vector<mindspore::tensor::MSTensor *> NPUGraphUtils::GetGraphInTensors(std::vector<NPUOp *> ops) {
-  std::vector<mindspore::tensor::MSTensor *> inputs;
-  auto is_op_output = [&](tensor::MSTensor *tensor) -> bool {
+std::vector<mindspore::MSTensor> NPUGraphUtils::GetGraphInTensors(std::vector<NPUOp *> ops) {
+  std::vector<mindspore::MSTensor> inputs;
+  auto is_op_output = [&](mindspore::MSTensor tensor) -> bool {
     for (auto op : ops) {
       auto out_tensors = op->outputs();
       if (find(out_tensors.begin(), out_tensors.end(), tensor) != out_tensors.end()) {
@@ -30,7 +30,7 @@ std::vector<mindspore::tensor::MSTensor *> NPUGraphUtils::GetGraphInTensors(std:
 
   for (auto op : ops) {
     for (auto in_tensor : op->inputs()) {
-      if (in_tensor->data() == nullptr && !is_op_output(in_tensor)) {
+      if (in_tensor.Data() == nullptr && !is_op_output(in_tensor)) {
         inputs.push_back(in_tensor);
       }
     }
@@ -38,9 +38,9 @@ std::vector<mindspore::tensor::MSTensor *> NPUGraphUtils::GetGraphInTensors(std:
   return inputs;
 }
 
-std::vector<mindspore::tensor::MSTensor *> NPUGraphUtils::GetGraphOutTensors(std::vector<NPUOp *> ops) {
-  std::vector<mindspore::tensor::MSTensor *> outputs;
-  auto is_op_input = [&](const tensor::MSTensor *tensor) -> bool {
+std::vector<mindspore::MSTensor> NPUGraphUtils::GetGraphOutTensors(std::vector<NPUOp *> ops) {
+  std::vector<mindspore::MSTensor> outputs;
+  auto is_op_input = [&](const mindspore::MSTensor tensor) -> bool {
     for (auto op : ops) {
       auto in_tensors = op->inputs();
       if (find(in_tensors.begin(), in_tensors.end(), tensor) != in_tensors.end()) {

mindspore/lite/src/delegate/npu/npu_graph_utils.h

@@ -23,9 +23,9 @@
 namespace mindspore {
 class NPUGraphUtils {
  public:
-  static std::vector<mindspore::tensor::MSTensor *> GetGraphInTensors(std::vector<NPUOp *> ops);
+  static std::vector<mindspore::MSTensor> GetGraphInTensors(std::vector<NPUOp *> ops);
 
-  static std::vector<mindspore::tensor::MSTensor *> GetGraphOutTensors(std::vector<NPUOp *> ops);
+  static std::vector<mindspore::MSTensor> GetGraphOutTensors(std::vector<NPUOp *> ops);
 };
 }  // namespace mindspore
 

mindspore/lite/src/delegate/npu/npu_subgraph.cc

@@ -48,7 +48,7 @@ NPUSubGraph::~NPUSubGraph() {
   op_buffer_.clear();
 }
 
-void NPUSubGraph::set_input(tensor::MSTensor *in_tensor, int index) {
+void NPUSubGraph::set_input(mindspore::MSTensor in_tensor, int index) {
   MS_ASSERT(index < inputs_.size());
   auto origin_tensor = inputs_[index];
   // only in_ops_ input tensors list used in execute function
@@ -62,7 +62,7 @@ void NPUSubGraph::set_input(tensor::MSTensor *in_tensor, int index) {
   this->inputs_[index] = in_tensor;
 }
 
-void NPUSubGraph::set_output(tensor::MSTensor *out_tensor, int index) {
+void NPUSubGraph::set_output(mindspore::MSTensor out_tensor, int index) {
   MS_ASSERT(index < out_tensor_sorted_.size());
   auto origin_tensor = outputs_[index];
   for (size_t i = 0; i < out_tensor_sorted_.size(); i++) {
@@ -217,7 +217,7 @@ int NPUSubGraph::BuildNPUInputOp() {
   return RET_OK;
 }
 
-bool NPUSubGraph::IsSubGraphInputTensor(tensor::MSTensor *input) {
+bool NPUSubGraph::IsSubGraphInputTensor(mindspore::MSTensor input) {
   if (find(this->inputs().begin(), this->inputs().end(), input) != this->inputs().end()) {
     return true;
   }

mindspore/lite/src/delegate/npu/npu_subgraph.h

@@ -20,7 +20,7 @@
 #include <memory>
 #include <vector>
 #include <string>
-#include "include/kernel.h"
+#include "include/api/kernel.h"
 #include "src/delegate/npu/npu_executor.h"
 
 namespace mindspore {
@@ -42,9 +42,9 @@ class NPUSubGraph : public kernel::Kernel {
     return lite::RET_ERROR;
   }
 
-  void set_input(tensor::MSTensor *in_tensor, int index) override;
+  void set_input(mindspore::MSTensor in_tensor, int index) override;
 
-  void set_output(tensor::MSTensor *out_tensor, int index) override;
+  void set_output(mindspore::MSTensor out_tensor, int index) override;
 
   int GetGraphInOutOps();
 
@@ -59,7 +59,7 @@ class NPUSubGraph : public kernel::Kernel {
 
   int GetNPUOperators(const std::vector<NPUOp *> &ops);
 
-  bool IsSubGraphInputTensor(tensor::MSTensor *input);
+  bool IsSubGraphInputTensor(mindspore::MSTensor input);
 
   std::string GetOMModelName();
 
@@ -69,7 +69,7 @@ class NPUSubGraph : public kernel::Kernel {
 
   std::vector<ge::Operator> subgraph_output_ops_;
 
-  std::vector<tensor::MSTensor *> out_tensor_sorted_;
+  std::vector<mindspore::MSTensor> out_tensor_sorted_;
 
   std::vector<ge::Operator *> op_buffer_;
 

mindspore/lite/src/delegate/npu/op/activation_npu.cc

@@ -16,8 +16,8 @@
 
 #include "src/delegate/npu/op/activation_npu.h"
 namespace mindspore {
-int ActivationNPUOp::IsSupport(const schema::Primitive *primitive, const std::vector<tensor::MSTensor *> &in_tensors,
-                               const std::vector<tensor::MSTensor *> &out_tensors) {
+int ActivationNPUOp::IsSupport(const schema::Primitive *primitive, const std::vector<mindspore::MSTensor> &in_tensors,
+                               const std::vector<mindspore::MSTensor> &out_tensors) {
   auto act_prim = primitive->value_as_Activation();
   if (act_prim == nullptr) {
     MS_LOG(ERROR) << "Get null primitive value for op ." << name_;
@@ -33,8 +33,8 @@ int ActivationNPUOp::IsSupport(const schema::Primitive *primitive, const std::ve
   return RET_OK;
 }
 
-int ActivationNPUOp::Init(const schema::Primitive *primitive, const std::vector<tensor::MSTensor *> &in_tensors,
-                          const std::vector<tensor::MSTensor *> &out_tensors) {
+int ActivationNPUOp::Init(const schema::Primitive *primitive, const std::vector<mindspore::MSTensor> &in_tensors,
+                          const std::vector<mindspore::MSTensor> &out_tensors) {
   act_ = new (std::nothrow) hiai::op::Activation(name_);
   if (act_ == nullptr) {
     MS_LOG(ERROR) << "New activation npu operator for activation op " << name_ << " failed.";
@@ -72,8 +72,8 @@ int ActivationNPUOp::Init(const schema::Primitive *primitive, const std::vector<
   return RET_OK;
 }
 
-int ActivationNPUOp::SetNPUInputs(const std::vector<tensor::MSTensor *> &in_tensors,
-                                  const std::vector<tensor::MSTensor *> &out_tensors,
+int ActivationNPUOp::SetNPUInputs(const std::vector<mindspore::MSTensor> &in_tensors,
+                                  const std::vector<mindspore::MSTensor> &out_tensors,
                                   const std::vector<ge::Operator *> &npu_inputs) {
   act_->set_input_x(*npu_inputs[0]);
   return RET_OK;

mindspore/lite/src/delegate/npu/op/activation_npu.h

@@ -18,26 +18,25 @@
 
 #include <vector>
 #include <string>
-#include "include/graph/op/all_ops.h"
 #include "include/graph/compatible/all_ops.h"
 #include "src/delegate/npu/op/npu_op.h"
 namespace mindspore {
 class ActivationNPUOp : public NPUOp {
  public:
-  ActivationNPUOp(const schema::Primitive *primitive, const std::vector<tensor::MSTensor *> &in_tensors,
-                  const std::vector<tensor::MSTensor *> &out_tensors, std::string name)
+  ActivationNPUOp(const schema::Primitive *primitive, const std::vector<mindspore::MSTensor> &in_tensors,
+                  const std::vector<mindspore::MSTensor> &out_tensors, std::string name)
       : NPUOp(primitive, in_tensors, out_tensors, name) {}
 
   ~ActivationNPUOp() override;
 
-  int IsSupport(const schema::Primitive *primitive, const std::vector<tensor::MSTensor *> &in_tensors,
-                const std::vector<tensor::MSTensor *> &out_tensors) override;
+  int IsSupport(const schema::Primitive *primitive, const std::vector<mindspore::MSTensor> &in_tensors,
+                const std::vector<mindspore::MSTensor> &out_tensors) override;
 
-  int Init(const schema::Primitive *primitive, const std::vector<tensor::MSTensor *> &in_tensors,
-           const std::vector<tensor::MSTensor *> &out_tensors) override;
+  int Init(const schema::Primitive *primitive, const std::vector<mindspore::MSTensor> &in_tensors,
+           const std::vector<mindspore::MSTensor> &out_tensors) override;
 
-  int SetNPUInputs(const std::vector<tensor::MSTensor *> &in_tensors,
-                   const std::vector<tensor::MSTensor *> &out_tensors,
+  int SetNPUInputs(const std::vector<mindspore::MSTensor> &in_tensors,
+                   const std::vector<mindspore::MSTensor> &out_tensors,
                    const std::vector<ge::Operator *> &npu_inputs) override;
 
   ge::Operator *GetNPUOp() override;

mindspore/lite/src/delegate/npu/op/argmax_npu.cc

@@ -16,11 +16,10 @@
 
 #include "src/delegate/npu/op/argmax_npu.h"
 #include <memory>
-#include "src/delegate/npu/npu_converter_utils.h"
 
 namespace mindspore {
-int ArgmaxNPUOp::Init(const schema::Primitive *primitive, const std::vector<tensor::MSTensor *> &in_tensors,
-                      const std::vector<tensor::MSTensor *> &out_tensors) {
+int ArgmaxNPUOp::Init(const schema::Primitive *primitive, const std::vector<mindspore::MSTensor> &in_tensors,
+                      const std::vector<mindspore::MSTensor> &out_tensors) {
   argmax_ = new (std::nothrow) hiai::op::ArgMaxExt2(name_);
   if (argmax_ == nullptr) {
     MS_LOG(ERROR) << "New argmax npu operator for " << name_ << " failed.";
@@ -54,8 +53,8 @@ int ArgmaxNPUOp::Init(const schema::Primitive *primitive, const std::vector<tens
   return RET_OK;
 }
 
-int ArgmaxNPUOp::SetNPUInputs(const std::vector<tensor::MSTensor *> &in_tensors,
-                              const std::vector<tensor::MSTensor *> &out_tensors,
+int ArgmaxNPUOp::SetNPUInputs(const std::vector<mindspore::MSTensor> &in_tensors,
+                              const std::vector<mindspore::MSTensor> &out_tensors,
                               const std::vector<ge::Operator *> &npu_inputs) {
   argmax_->set_input_x(*npu_inputs[0]);
   return RET_OK;

mindspore/lite/src/delegate/npu/op/argmax_npu.h

@@ -27,22 +27,22 @@ namespace mindspore {
 
 class ArgmaxNPUOp : public NPUOp {
  public:
-  ArgmaxNPUOp(const schema::Primitive *primitive, const std::vector<tensor::MSTensor *> &in_tensors,
-              const std::vector<tensor::MSTensor *> &out_tensors, std::string name)
+  ArgmaxNPUOp(const schema::Primitive *primitive, const std::vector<mindspore::MSTensor> &in_tensors,
+              const std::vector<mindspore::MSTensor> &out_tensors, std::string name)
       : NPUOp(primitive, in_tensors, out_tensors, name) {}
 
   ~ArgmaxNPUOp() override;
 
-  int IsSupport(const schema::Primitive *primitive, const std::vector<tensor::MSTensor *> &in_tensors,
-                const std::vector<tensor::MSTensor *> &out_tensors) override {
+  int IsSupport(const schema::Primitive *primitive, const std::vector<mindspore::MSTensor> &in_tensors,
+                const std::vector<mindspore::MSTensor> &out_tensors) override {
     return RET_OK;
   }
 
-  int Init(const schema::Primitive *primitive, const std::vector<tensor::MSTensor *> &in_tensors,
-           const std::vector<tensor::MSTensor *> &out_tensors) override;
+  int Init(const schema::Primitive *primitive, const std::vector<mindspore::MSTensor> &in_tensors,
+           const std::vector<mindspore::MSTensor> &out_tensors) override;
 
-  int SetNPUInputs(const std::vector<tensor::MSTensor *> &in_tensors,
-                   const std::vector<tensor::MSTensor *> &out_tensors,
+  int SetNPUInputs(const std::vector<mindspore::MSTensor> &in_tensors,
+                   const std::vector<mindspore::MSTensor> &out_tensors,
                    const std::vector<ge::Operator *> &npu_inputs) override;
 
   ge::Operator *GetNPUOp() override;

mindspore/lite/src/delegate/npu/op/arithmetic_npu.cc

@@ -15,23 +15,22 @@
  */
 
 #include "src/delegate/npu/op/arithmetic_npu.h"
-#include "include/graph/op/all_ops.h"
 namespace mindspore {
 constexpr int RELU_MODE = 1;
 constexpr int RELU6_MODE = 14;
-int ArithmeticNPUOp::IsSupport(const schema::Primitive *primitive, const std::vector<tensor::MSTensor *> &in_tensors,
-                               const std::vector<tensor::MSTensor *> &out_tensors) {
-  if (in_tensors[0]->shape() != in_tensors[1]->shape()) {
+int ArithmeticNPUOp::IsSupport(const schema::Primitive *primitive, const std::vector<mindspore::MSTensor> &in_tensors,
+                               const std::vector<mindspore::MSTensor> &out_tensors) {
+  if (in_tensors[0].Shape() != in_tensors[1].Shape()) {
     MS_LOG(WARNING) << name_ << " for the two inputs, the corresponding dimensions must have the same value."
-                    << " shape 1 is:" << in_tensors[0]->shape() << " shape 2 is:" << in_tensors[1]->shape();
+                    << " shape 1 is:" << in_tensors[0].Shape() << " shape 2 is:" << in_tensors[1].Shape();
     return RET_NOT_SUPPORT;
   }
   auto type = primitive->value_type();
-  if (type == mindspore::schema::PrimitiveType_Less && in_tensors[0]->shape().size() == 1) {
+  if (type == mindspore::schema::PrimitiveType_Less && in_tensors[0].Shape().size() == 1) {
     MS_LOG(WARNING) << name_ << " not support input 1d";
     return RET_NOT_SUPPORT;
   }
-  if (type == mindspore::schema::PrimitiveType_Equal && in_tensors[0]->shape().size() == 2) {
+  if (type == mindspore::schema::PrimitiveType_Equal && in_tensors[0].Shape().size() == 2) {
     MS_LOG(WARNING) << name_ << " not support input 2d";
     return RET_NOT_SUPPORT;
   }
@@ -48,8 +47,8 @@ ge::Operator *CreateOperator(const std::string &name) {
   return op;
 }
 
-int ArithmeticNPUOp::Init(const schema::Primitive *primitive, const std::vector<tensor::MSTensor *> &in_tensors,
-                          const std::vector<tensor::MSTensor *> &out_tensors) {
+int ArithmeticNPUOp::Init(const schema::Primitive *primitive, const std::vector<mindspore::MSTensor> &in_tensors,
+                          const std::vector<mindspore::MSTensor> &out_tensors) {
   switch (type_) {
     case schema::PrimitiveType_MulFusion:
       op_ = CreateOperator<hiai::op::Mul>(name_);
@@ -143,8 +142,8 @@ void SetInputs(const std::vector<ge::Operator *> &npu_inputs, ge::Operator *op)
   return;
 }
 
-int ArithmeticNPUOp::SetNPUInputs(const std::vector<tensor::MSTensor *> &in_tensors,
-                                  const std::vector<tensor::MSTensor *> &out_tensors,
+int ArithmeticNPUOp::SetNPUInputs(const std::vector<mindspore::MSTensor> &in_tensors,
+                                  const std::vector<mindspore::MSTensor> &out_tensors,
                                   const std::vector<ge::Operator *> &npu_inputs) {
   switch (type_) {
     case schema::PrimitiveType_MulFusion:
@@ -203,7 +202,7 @@ int ArithmeticNPUOp::SetNPUInputs(const std::vector<tensor::MSTensor *> &in_tens
 }
 
 int ArithmeticNPUOp::SetNPUInputs(
-  const std::vector<tensor::MSTensor *> &in_tensors, const std::vector<tensor::MSTensor *> &out_tensors,
+  const std::vector<mindspore::MSTensor> &in_tensors, const std::vector<mindspore::MSTensor> &out_tensors,
   const std::vector<ge::Operator *> &npu_inputs,
   const std::unordered_map<int, std::pair<ge::Operator *, int>> &index2_multi_out_index) {
   auto ret = SetNPUInputs(in_tensors, out_tensors, npu_inputs);

mindspore/lite/src/delegate/npu/op/arithmetic_npu.h

@@ -25,24 +25,24 @@
 namespace mindspore {
 class ArithmeticNPUOp : public NPUOp {
  public:
-  ArithmeticNPUOp(const schema::Primitive *primitive, const std::vector<tensor::MSTensor *> &in_tensors,
-                  const std::vector<tensor::MSTensor *> &out_tensors, std::string name)
+  ArithmeticNPUOp(const schema::Primitive *primitive, const std::vector<mindspore::MSTensor> &in_tensors,
+                  const std::vector<mindspore::MSTensor> &out_tensors, std::string name)
       : NPUOp(primitive, in_tensors, out_tensors, name) {}
 
   ~ArithmeticNPUOp() override;
 
-  int IsSupport(const schema::Primitive *primitive, const std::vector<tensor::MSTensor *> &in_tensors,
-                const std::vector<tensor::MSTensor *> &out_tensors) override;
+  int IsSupport(const schema::Primitive *primitive, const std::vector<mindspore::MSTensor> &in_tensors,
+                const std::vector<mindspore::MSTensor> &out_tensors) override;
 
-  int Init(const schema::Primitive *primitive, const std::vector<tensor::MSTensor *> &in_tensors,
-           const std::vector<tensor::MSTensor *> &out_tensors) override;
+  int Init(const schema::Primitive *primitive, const std::vector<mindspore::MSTensor> &in_tensors,
+           const std::vector<mindspore::MSTensor> &out_tensors) override;
 
-  int SetNPUInputs(const std::vector<tensor::MSTensor *> &in_tensors,
-                   const std::vector<tensor::MSTensor *> &out_tensors,
+  int SetNPUInputs(const std::vector<mindspore::MSTensor> &in_tensors,
+                   const std::vector<mindspore::MSTensor> &out_tensors,
                    const std::vector<ge::Operator *> &npu_inputs) override;
 
-  int SetNPUInputs(const std::vector<tensor::MSTensor *> &in_tensors,
-                   const std::vector<tensor::MSTensor *> &out_tensors, const std::vector<ge::Operator *> &npu_inputs,
+  int SetNPUInputs(const std::vector<mindspore::MSTensor> &in_tensors,
+                   const std::vector<mindspore::MSTensor> &out_tensors, const std::vector<ge::Operator *> &npu_inputs,
                    const std::unordered_map<int, std::pair<ge::Operator *, int>> &index2_multi_out_index) override;
 
   ge::Operator *GetNPUOp() override;

mindspore/lite/src/delegate/npu/op/arithmetic_self_npu.cc

@@ -16,7 +16,6 @@
 
 #include "src/delegate/npu/op/arithmetic_self_npu.h"
 #include <string>
-#include "include/graph/op/all_ops.h"
 
 namespace mindspore {
 template <typename T>
@@ -29,8 +28,8 @@ ge::Operator *CreateOperator(const std::string &name) {
   return op;
 }
 
-int ArithmeticSelfNPUOp::Init(const schema::Primitive *primitive, const std::vector<tensor::MSTensor *> &in_tensors,
-                              const std::vector<tensor::MSTensor *> &out_tensors) {
+int ArithmeticSelfNPUOp::Init(const schema::Primitive *primitive, const std::vector<mindspore::MSTensor> &in_tensors,
+                              const std::vector<mindspore::MSTensor> &out_tensors) {
   switch (type_) {
     case schema::PrimitiveType_Cos:
       op_ = CreateOperator<hiai::op::Cos>(name_);
@@ -86,8 +85,8 @@ void SetInputs(const std::vector<ge::Operator *> &npu_inputs, ge::Operator *op)
   return;
 }
 
-int ArithmeticSelfNPUOp::SetNPUInputs(const std::vector<tensor::MSTensor *> &in_tensors,
-                                      const std::vector<tensor::MSTensor *> &out_tensors,
+int ArithmeticSelfNPUOp::SetNPUInputs(const std::vector<mindspore::MSTensor> &in_tensors,
+                                      const std::vector<mindspore::MSTensor> &out_tensors,
                                       const std::vector<ge::Operator *> &npu_inputs) {
   switch (type_) {
     case schema::PrimitiveType_Cos:

mindspore/lite/src/delegate/npu/op/arithmetic_self_npu.h

@@ -24,22 +24,22 @@
 namespace mindspore {
 class ArithmeticSelfNPUOp : public NPUOp {
  public:
-  ArithmeticSelfNPUOp(const schema::Primitive *primitive, const std::vector<tensor::MSTensor *> &in_tensors,
-                      const std::vector<tensor::MSTensor *> &out_tensors, std::string name)
+  ArithmeticSelfNPUOp(const schema::Primitive *primitive, const std::vector<mindspore::MSTensor> &in_tensors,
+                      const std::vector<mindspore::MSTensor> &out_tensors, std::string name)
       : NPUOp(primitive, in_tensors, out_tensors, name) {}
 
   ~ArithmeticSelfNPUOp() override;
 
-  int IsSupport(const schema::Primitive *primitive, const std::vector<tensor::MSTensor *> &in_tensors,
-                const std::vector<tensor::MSTensor *> &out_tensors) override {
+  int IsSupport(const schema::Primitive *primitive, const std::vector<mindspore::MSTensor> &in_tensors,
+                const std::vector<mindspore::MSTensor> &out_tensors) override {
     return RET_OK;
   }
 
-  int Init(const schema::Primitive *primitive, const std::vector<tensor::MSTensor *> &in_tensors,
-           const std::vector<tensor::MSTensor *> &out_tensors) override;
+  int Init(const schema::Primitive *primitive, const std::vector<mindspore::MSTensor> &in_tensors,
+           const std::vector<mindspore::MSTensor> &out_tensors) override;
 
-  int SetNPUInputs(const std::vector<tensor::MSTensor *> &in_tensors,
-                   const std::vector<tensor::MSTensor *> &out_tensors,
+  int SetNPUInputs(const std::vector<mindspore::MSTensor> &in_tensors,
+                   const std::vector<mindspore::MSTensor> &out_tensors,
                    const std::vector<ge::Operator *> &npu_inputs) override;
 
   ge::Operator *GetNPUOp() override;

mindspore/lite/src/delegate/npu/op/avg_pooling_npu.cc

@@ -16,8 +16,8 @@
 
 #include "src/delegate/npu/op/avg_pooling_npu.h"
 namespace mindspore {
-int AvgPoolingNPUOp::IsSupport(const schema::Primitive *primitive, const std::vector<tensor::MSTensor *> &in_tensors,
-                               const std::vector<tensor::MSTensor *> &out_tensors) {
+int AvgPoolingNPUOp::IsSupport(const schema::Primitive *primitive, const std::vector<mindspore::MSTensor> &in_tensors,
+                               const std::vector<mindspore::MSTensor> &out_tensors) {
   auto pooling_prim = primitive->value_as_AvgPoolFusion();
   if (pooling_prim == nullptr) {
     MS_LOG(ERROR) << "Get null primitive value for op ." << name_;
@@ -71,8 +71,8 @@ int AvgPoolingNPUOp::SetPoolingParam(const schema::AvgPoolFusion *pooling_prim)
   return RET_OK;
 }
 
-int AvgPoolingNPUOp::Init(const schema::Primitive *primitive, const std::vector<tensor::MSTensor *> &in_tensors,
-                          const std::vector<tensor::MSTensor *> &out_tensors) {
+int AvgPoolingNPUOp::Init(const schema::Primitive *primitive, const std::vector<mindspore::MSTensor> &in_tensors,
+                          const std::vector<mindspore::MSTensor> &out_tensors) {
   pooling_ = new (std::nothrow) hiai::op::PoolingD(name_ + "_pooling");
   if (pooling_ == nullptr) {
     MS_LOG(ERROR) << "New pooling npu operator for op " << name_ << " failed.";
@@ -99,8 +99,8 @@ int AvgPoolingNPUOp::Init(const schema::Primitive *primitive, const std::vector<
   return RET_OK;
 }
 
-int AvgPoolingNPUOp::SetNPUInputs(const std::vector<tensor::MSTensor *> &in_tensors,
-                                  const std::vector<tensor::MSTensor *> &out_tensors,
+int AvgPoolingNPUOp::SetNPUInputs(const std::vector<mindspore::MSTensor> &in_tensors,
+                                  const std::vector<mindspore::MSTensor> &out_tensors,
                                   const std::vector<ge::Operator *> &npu_inputs) {
   pooling_->set_input_x(*npu_inputs[0]);
   return RET_OK;

mindspore/lite/src/delegate/npu/op/avg_pooling_npu.h

@@ -23,20 +23,20 @@
 namespace mindspore {
 class AvgPoolingNPUOp : public ConvolutionBaseNPUOp {
  public:
-  AvgPoolingNPUOp(const schema::Primitive *primitive, const std::vector<tensor::MSTensor *> &in_tensors,
-                  const std::vector<tensor::MSTensor *> &out_tensors, std::string name)
+  AvgPoolingNPUOp(const schema::Primitive *primitive, const std::vector<mindspore::MSTensor> &in_tensors,
+                  const std::vector<mindspore::MSTensor> &out_tensors, std::string name)
       : ConvolutionBaseNPUOp(primitive, in_tensors, out_tensors, name) {}
 
   ~AvgPoolingNPUOp() override;
 
-  int IsSupport(const schema::Primitive *primitive, const std::vector<tensor::MSTensor *> &in_tensors,
-                const std::vector<tensor::MSTensor *> &out_tensors) override;
+  int IsSupport(const schema::Primitive *primitive, const std::vector<mindspore::MSTensor> &in_tensors,
+                const std::vector<mindspore::MSTensor> &out_tensors) override;
 
-  int Init(const schema::Primitive *primitive, const std::vector<tensor::MSTensor *> &in_tensors,
-           const std::vector<tensor::MSTensor *> &out_tensors) override;
+  int Init(const schema::Primitive *primitive, const std::vector<mindspore::MSTensor> &in_tensors,
+           const std::vector<mindspore::MSTensor> &out_tensors) override;
 
-  int SetNPUInputs(const std::vector<tensor::MSTensor *> &in_tensors,
-                   const std::vector<tensor::MSTensor *> &out_tensors,
+  int SetNPUInputs(const std::vector<mindspore::MSTensor> &in_tensors,
+                   const std::vector<mindspore::MSTensor> &out_tensors,
                    const std::vector<ge::Operator *> &npu_inputs) override;
 
   ge::Operator *GetNPUOp() override;

mindspore/lite/src/delegate/npu/op/batchnorm_npu.cc

@@ -15,11 +15,12 @@
  */
 
 #include "src/delegate/npu/op/batchnorm_npu.h"
+#include "include/graph/op/all_ops.h"
 #include "src/delegate/npu/npu_converter_utils.h"
 
 namespace mindspore {
-int BatchnormNPUOp::Init(const schema::Primitive *primitive, const std::vector<tensor::MSTensor *> &in_tensors,
-                         const std::vector<tensor::MSTensor *> &out_tensors) {
+int BatchnormNPUOp::Init(const schema::Primitive *primitive, const std::vector<mindspore::MSTensor> &in_tensors,
+                         const std::vector<mindspore::MSTensor> &out_tensors) {
   batchnorm_ = new (std::nothrow) ge::op::BatchNormExt2(name_);
   if (batchnorm_ == nullptr) {
     MS_LOG(ERROR) << "New batchnorm npu operator for batchnorm op " << name_ << " failed.";
@@ -36,8 +37,8 @@ int BatchnormNPUOp::Init(const schema::Primitive *primitive, const std::vector<t
   return RET_OK;
 }
 
-int BatchnormNPUOp::SetNPUInputs(const std::vector<tensor::MSTensor *> &in_tensors,
-                                 const std::vector<tensor::MSTensor *> &out_tensors,
+int BatchnormNPUOp::SetNPUInputs(const std::vector<mindspore::MSTensor> &in_tensors,
+                                 const std::vector<mindspore::MSTensor> &out_tensors,
                                  const std::vector<ge::Operator *> &npu_inputs) {
   batchnorm_->set_input_x(*npu_inputs[0]);
   auto scale = new (std::nothrow) hiai::op::Const(name_ + "_scale");

mindspore/lite/src/delegate/npu/op/batchnorm_npu.h

@@ -18,29 +18,28 @@
 
 #include <vector>
 #include <string>
-#include "include/graph/op/all_ops.h"
 #include "include/graph/compatible/all_ops.h"
 #include "src/delegate/npu/op/npu_op.h"
 
 namespace mindspore {
 class BatchnormNPUOp : public NPUOp {
  public:
-  BatchnormNPUOp(const schema::Primitive *primitive, const std::vector<tensor::MSTensor *> &in_tensors,
-                 const std::vector<tensor::MSTensor *> &out_tensors, std::string name)
+  BatchnormNPUOp(const schema::Primitive *primitive, const std::vector<mindspore::MSTensor> &in_tensors,
+                 const std::vector<mindspore::MSTensor> &out_tensors, std::string name)
       : NPUOp(primitive, in_tensors, out_tensors, name) {}
 
   ~BatchnormNPUOp() override;
 
-  int IsSupport(const schema::Primitive *primitive, const std::vector<tensor::MSTensor *> &in_tensors,
-                const std::vector<tensor::MSTensor *> &out_tensors) override {
+  int IsSupport(const schema::Primitive *primitive, const std::vector<mindspore::MSTensor> &in_tensors,
+                const std::vector<mindspore::MSTensor> &out_tensors) override {
     return RET_OK;
   }
 
-  int Init(const schema::Primitive *primitive, const std::vector<tensor::MSTensor *> &in_tensors,
-           const std::vector<tensor::MSTensor *> &out_tensors) override;
+  int Init(const schema::Primitive *primitive, const std::vector<mindspore::MSTensor> &in_tensors,
+           const std::vector<mindspore::MSTensor> &out_tensors) override;
 
-  int SetNPUInputs(const std::vector<tensor::MSTensor *> &in_tensors,
-                   const std::vector<tensor::MSTensor *> &out_tensors,
+  int SetNPUInputs(const std::vector<mindspore::MSTensor> &in_tensors,
+                   const std::vector<mindspore::MSTensor> &out_tensors,
                    const std::vector<ge::Operator *> &npu_inputs) override;
 
   ge::Operator *GetNPUOp() override;

mindspore/lite/src/delegate/npu/op/cast_npu.cc

@@ -18,10 +18,10 @@
 #include "src/delegate/npu/npu_converter_utils.h"
 
 namespace mindspore {
-int CastNPUOp::IsSupport(const schema::Primitive *primitive, const std::vector<tensor::MSTensor *> &in_tensors,
-                         const std::vector<tensor::MSTensor *> &out_tensors) {
-  if (in_tensors.size() >= 2 && in_tensors[1]->ElementsNum() == 1) {
-    dst_type_ = static_cast<int *>(in_tensors[1]->data())[0];
+int CastNPUOp::IsSupport(const schema::Primitive *primitive, const std::vector<mindspore::MSTensor> &in_tensors,
+                         const std::vector<mindspore::MSTensor> &out_tensors) {
+  if (in_tensors.size() >= 2 && in_tensors[1].ElementNum() == 1) {
+    dst_type_ = reinterpret_cast<const int *>(in_tensors[1].Data().get())[0];
   } else {
     MS_LOG(WARNING) << "NPU dst dtype is attribute.";
     return RET_NOT_SUPPORT;
@@ -29,20 +29,20 @@ int CastNPUOp::IsSupport(const schema::Primitive *primitive, const std::vector<t
   return RET_OK;
 }
 
-int CastNPUOp::Init(const schema::Primitive *primitive, const std::vector<tensor::MSTensor *> &in_tensors,
-                    const std::vector<tensor::MSTensor *> &out_tensors) {
+int CastNPUOp::Init(const schema::Primitive *primitive, const std::vector<mindspore::MSTensor> &in_tensors,
+                    const std::vector<mindspore::MSTensor> &out_tensors) {
   cast_ = new (std::nothrow) hiai::op::CastT(name_);
   if (cast_ == nullptr) {
     MS_LOG(ERROR) << name_ << " op is nullptr";
     return RET_ERROR;
   }
-  cast_->set_attr_dst_dtype(ConverterToNPUDataType(static_cast<TypeId>(dst_type_)));
-  cast_->set_attr_src_dtype(ConverterToNPUDataType(static_cast<TypeId>(in_tensors[0]->data_type())));
+  cast_->set_attr_dst_dtype(ConverterToNPUDataType(static_cast<DataType>(dst_type_)));
+  cast_->set_attr_src_dtype(ConverterToNPUDataType(static_cast<DataType>(in_tensors[0].DataType())));
   return RET_OK;
 }
 
-int CastNPUOp::SetNPUInputs(const std::vector<tensor::MSTensor *> &in_tensors,
-                            const std::vector<tensor::MSTensor *> &out_tensors,
+int CastNPUOp::SetNPUInputs(const std::vector<mindspore::MSTensor> &in_tensors,
+                            const std::vector<mindspore::MSTensor> &out_tensors,
                             const std::vector<ge::Operator *> &npu_inputs) {
   cast_->set_input_x(*npu_inputs[0]);
   return RET_OK;

mindspore/lite/src/delegate/npu/op/cast_npu.h

@@ -24,20 +24,20 @@
 namespace mindspore {
 class CastNPUOp : public NPUOp {
  public:
-  CastNPUOp(const schema::Primitive *primitive, const std::vector<tensor::MSTensor *> &in_tensors,
-            const std::vector<tensor::MSTensor *> &out_tensors, std::string name)
+  CastNPUOp(const schema::Primitive *primitive, const std::vector<mindspore::MSTensor> &in_tensors,
+            const std::vector<mindspore::MSTensor> &out_tensors, std::string name)
       : NPUOp(primitive, in_tensors, out_tensors, name) {}
 
   ~CastNPUOp() override;
 
-  int IsSupport(const schema::Primitive *primitive, const std::vector<tensor::MSTensor *> &in_tensors,
-                const std::vector<tensor::MSTensor *> &out_tensors) override;
+  int IsSupport(const schema::Primitive *primitive, const std::vector<mindspore::MSTensor> &in_tensors,
+                const std::vector<mindspore::MSTensor> &out_tensors) override;
 
-  int Init(const schema::Primitive *primitive, const std::vector<tensor::MSTensor *> &in_tensors,
-           const std::vector<tensor::MSTensor *> &out_tensors) override;
+  int Init(const schema::Primitive *primitive, const std::vector<mindspore::MSTensor> &in_tensors,
+           const std::vector<mindspore::MSTensor> &out_tensors) override;
 
-  int SetNPUInputs(const std::vector<tensor::MSTensor *> &in_tensors,
-                   const std::vector<tensor::MSTensor *> &out_tensors,
+  int SetNPUInputs(const std::vector<mindspore::MSTensor> &in_tensors,
+                   const std::vector<mindspore::MSTensor> &out_tensors,
                    const std::vector<ge::Operator *> &npu_inputs) override;
 
   ge::Operator *GetNPUOp() override;

mindspore/lite/src/delegate/npu/op/concat_npu.cc

@@ -18,8 +18,8 @@
 #include "src/delegate/npu/npu_converter_utils.h"
 
 namespace mindspore {
-int ConcatNPUOp::Init(const schema::Primitive *primitive, const std::vector<tensor::MSTensor *> &in_tensors,
-                      const std::vector<tensor::MSTensor *> &out_tensors) {
+int ConcatNPUOp::Init(const schema::Primitive *primitive, const std::vector<mindspore::MSTensor> &in_tensors,
+                      const std::vector<mindspore::MSTensor> &out_tensors) {
   concat_ = new (std::nothrow) hiai::op::ConcatD(name_);
   if (concat_ == nullptr) {
     MS_LOG(ERROR) << name_ << " op is nullptr";
@@ -34,8 +34,8 @@ int ConcatNPUOp::Init(const schema::Primitive *primitive, const std::vector<tens
   return RET_OK;
 }
 
-int ConcatNPUOp::SetNPUInputs(const std::vector<tensor::MSTensor *> &in_tensors,
-                              const std::vector<tensor::MSTensor *> &out_tensors,
+int ConcatNPUOp::SetNPUInputs(const std::vector<mindspore::MSTensor> &in_tensors,
+                              const std::vector<mindspore::MSTensor> &out_tensors,
                               const std::vector<ge::Operator *> &npu_inputs) {
   concat_->set_attr_concat_dim(axis_);
   concat_->set_attr_N(npu_inputs.size());

mindspore/lite/src/delegate/npu/op/concat_npu.h

@@ -23,22 +23,22 @@
 namespace mindspore {
 class ConcatNPUOp : public NPUOp {
  public:
-  ConcatNPUOp(const schema::Primitive *primitive, const std::vector<tensor::MSTensor *> &in_tensors,
-              const std::vector<tensor::MSTensor *> &out_tensors, std::string name)
+  ConcatNPUOp(const schema::Primitive *primitive, const std::vector<mindspore::MSTensor> &in_tensors,
+              const std::vector<mindspore::MSTensor> &out_tensors, std::string name)
       : NPUOp(primitive, in_tensors, out_tensors, name) {}
 
   ~ConcatNPUOp() override;
 
-  int IsSupport(const schema::Primitive *primitive, const std::vector<tensor::MSTensor *> &in_tensors,
-                const std::vector<tensor::MSTensor *> &out_tensors) override {
+  int IsSupport(const schema::Primitive *primitive, const std::vector<mindspore::MSTensor> &in_tensors,
+                const std::vector<mindspore::MSTensor> &out_tensors) override {
     return RET_OK;
   }
 
-  int Init(const schema::Primitive *primitive, const std::vector<tensor::MSTensor *> &in_tensors,
-           const std::vector<tensor::MSTensor *> &out_tensors) override;
+  int Init(const schema::Primitive *primitive, const std::vector<mindspore::MSTensor> &in_tensors,
+           const std::vector<mindspore::MSTensor> &out_tensors) override;
 
-  int SetNPUInputs(const std::vector<tensor::MSTensor *> &in_tensors,
-                   const std::vector<tensor::MSTensor *> &out_tensors,
+  int SetNPUInputs(const std::vector<mindspore::MSTensor> &in_tensors,
+                   const std::vector<mindspore::MSTensor> &out_tensors,
                    const std::vector<ge::Operator *> &npu_inputs) override;
 
   ge::Operator *GetNPUOp() override;

mindspore/lite/src/delegate/npu/op/convolution_base_npu.cc

@@ -17,7 +17,6 @@
 #include "src/delegate/npu/op/convolution_base_npu.h"
 #include "src/delegate/npu/npu_converter_utils.h"
 #include "src/delegate/npu/transpose_kernel.h"
-#include "nnacl/fp16/cast_fp16.h"
 
 namespace mindspore {
 ConvolutionBaseNPUOp::~ConvolutionBaseNPUOp() {
@@ -35,27 +34,39 @@ ConvolutionBaseNPUOp::~ConvolutionBaseNPUOp() {
   }
 }
 
-int ConvolutionBaseNPUOp::InitWeightConst(const std::vector<tensor::MSTensor *> &inputs) {
+int ConvolutionBaseNPUOp::InitWeightConst(const std::vector<mindspore::MSTensor> &inputs) {
   weight_ = new (std::nothrow) hiai::op::Const(name_ + "_w");
   if (weight_ == nullptr) {
     MS_LOG(ERROR) << "New weight const failed.";
     return RET_ERROR;
   }
-  auto w_shape = inputs[1]->shape();
-  auto origin_data = inputs[1]->data();
-  auto fp32_data = origin_data;
-  if (inputs[1]->data_type() == kNumberTypeFloat16) {
-    fp32_data = reinterpret_cast<float *>(malloc(inputs[1]->ElementsNum() * sizeof(float)));
+  auto w_shape = inputs[1].Shape();
+  auto origin_data = inputs[1].Data().get();
+  float *fp32_data = nullptr;
+  if (inputs[1].DataType() == DataType::kNumberTypeFloat16) {
+#ifdef ENABLE_ARM64
+    fp32_data = reinterpret_cast<float *>(malloc(inputs[1].ElementNum() * sizeof(float)));
     // fp16->fp32
-    Float16ToFloat32(reinterpret_cast<float16_t *>(origin_data), reinterpret_cast<float *>(fp32_data),
-                     inputs[1]->ElementsNum());
+    Float16ToFloat32(reinterpret_cast<const float16_t *>(origin_data), reinterpret_cast<float *>(fp32_data),
+                     inputs[1].ElementNum());
+#else
+    MS_LOG(ERROR) << "This platform does not support fp16.";
+    return RET_ERROR;
+#endif
   }
-  auto nchw_data = reinterpret_cast<float *>(malloc(inputs[1]->ElementsNum() * sizeof(float)));
+  auto nchw_data = reinterpret_cast<float *>(malloc(inputs[1].ElementNum() * sizeof(float)));
   if (nchw_data == nullptr) {
     MS_LOG(ERROR) << "Malloc buffer failed.";
     return RET_ERROR;
   }
-  PackNHWCToNCHWFp32(fp32_data, nchw_data, w_shape[0], w_shape[1] * w_shape[2], w_shape[3]);
+  if (inputs[1].DataType() == DataType::kNumberTypeFloat16) {
+    PackNHWCToNCHWFp32(fp32_data, nchw_data, w_shape[0], w_shape[1] * w_shape[2], w_shape[3]);
+  } else if (inputs[1].DataType() == DataType::kNumberTypeFloat32) {
+    PackNHWCToNCHWFp32(origin_data, nchw_data, w_shape[0], w_shape[1] * w_shape[2], w_shape[3]);
+  } else {
+    MS_LOG(ERROR) << "Unsupported data type of weight tensor for npu convolution.";
+    return RET_ERROR;
+  }
 
   std::shared_ptr<ge::Tensor> weight_tensor = std::shared_ptr<ge::Tensor>(new (std::nothrow) ge::Tensor());
   if (weight_tensor == nullptr) {
@@ -63,16 +74,16 @@ int ConvolutionBaseNPUOp::InitWeightConst(const std::vector<tensor::MSTensor *>
     return RET_ERROR;
   }
   ge::TensorDesc tensor_desc(ConverterToNPUShape({w_shape[0], w_shape[3], w_shape[1], w_shape[2]}), ge::FORMAT_NCHW,
-                             ConverterToNPUDataType(inputs[1]->data_type()));
+                             ConverterToNPUDataType(inputs[1].DataType()));
   weight_tensor->SetTensorDesc(tensor_desc);
-  weight_tensor->SetData(reinterpret_cast<const uint8_t *>(nchw_data), inputs[1]->ElementsNum() * sizeof(float));
+  weight_tensor->SetData(reinterpret_cast<const uint8_t *>(nchw_data), inputs[1].ElementNum() * sizeof(float));
 
   weight_->set_attr_value(weight_tensor);
   free(nchw_data);
   return RET_OK;
 }
 
-int ConvolutionBaseNPUOp::InitBiasConst(const std::vector<tensor::MSTensor *> &inputs) {
+int ConvolutionBaseNPUOp::InitBiasConst(const std::vector<mindspore::MSTensor> &inputs) {
   if (inputs.size() >= 3) {
     bias_ = new (std::nothrow) hiai::op::Const(name_ + "_b");
     if (bias_ == nullptr) {

mindspore/lite/src/delegate/npu/op/convolution_base_npu.h

@@ -24,15 +24,15 @@
 namespace mindspore {
 class ConvolutionBaseNPUOp : public NPUOp {
  public:
-  ConvolutionBaseNPUOp(const schema::Primitive *primitive, const std::vector<tensor::MSTensor *> &in_tensors,
-                       const std::vector<tensor::MSTensor *> &out_tensors, std::string name)
+  ConvolutionBaseNPUOp(const schema::Primitive *primitive, const std::vector<mindspore::MSTensor> &in_tensors,
+                       const std::vector<mindspore::MSTensor> &out_tensors, std::string name)
       : NPUOp(primitive, in_tensors, out_tensors, name) {}
 
   ~ConvolutionBaseNPUOp() override;
 
  protected:
-  int InitWeightConst(const std::vector<tensor::MSTensor *> &inputs);
-  int InitBiasConst(const std::vector<tensor::MSTensor *> &inputs);
+  int InitWeightConst(const std::vector<mindspore::MSTensor> &inputs);
+  int InitBiasConst(const std::vector<mindspore::MSTensor> &inputs);
   int SetActivation(const ge::Operator *input, schema::ActivationType act_type);
   hiai::op::Activation *act_ = nullptr;
   hiai::op::Const *weight_ = nullptr;

mindspore/lite/src/delegate/npu/op/convolution_depthwise_npu.cc

@@ -42,8 +42,8 @@ int ConvolutionDepthwiseNPUOp::SetConvDwParam(const schema::Conv2DFusion *conv_p
 }
 
 int ConvolutionDepthwiseNPUOp::Init(const schema::Primitive *primitive,
-                                    const std::vector<tensor::MSTensor *> &in_tensors,
-                                    const std::vector<tensor::MSTensor *> &out_tensors) {
+                                    const std::vector<mindspore::MSTensor> &in_tensors,
+                                    const std::vector<mindspore::MSTensor> &out_tensors) {
   conv_dw_ = new (std::nothrow) hiai::op::ConvolutionDepthwise(name_ + "_conv_depthwise");
   if (conv_dw_ == nullptr) {
     MS_LOG(ERROR) << "New convolution depthwise operator for op " << name_ << " failed.";
@@ -70,8 +70,8 @@ int ConvolutionDepthwiseNPUOp::Init(const schema::Primitive *primitive,
   return RET_OK;
 }
 
-int ConvolutionDepthwiseNPUOp::SetNPUInputs(const std::vector<tensor::MSTensor *> &in_tensors,
-                                            const std::vector<tensor::MSTensor *> &out_tensors,
+int ConvolutionDepthwiseNPUOp::SetNPUInputs(const std::vector<mindspore::MSTensor> &in_tensors,
+                                            const std::vector<mindspore::MSTensor> &out_tensors,
                                             const std::vector<ge::Operator *> &npu_inputs) {
   auto ret = InitWeightConst(in_tensors);
   if (ret != RET_OK) {

mindspore/lite/src/delegate/npu/op/convolution_depthwise_npu.h

@@ -18,28 +18,27 @@
 #define MINDSPORE_LITE_SRC_RUNTIME_DELEGATE_NPU_OP_CONVOLUTION_DEPTHWISE_NPU_H_
 #include <vector>
 #include <string>
-#include "include/graph/op/all_ops.h"
 #include "include/graph/compatible/all_ops.h"
 #include "src/delegate/npu/op/convolution_base_npu.h"
 namespace mindspore {
 class ConvolutionDepthwiseNPUOp : public ConvolutionBaseNPUOp {
  public:
-  ConvolutionDepthwiseNPUOp(const schema::Primitive *primitive, const std::vector<tensor::MSTensor *> &in_tensors,
-                            const std::vector<tensor::MSTensor *> &out_tensors, std::string name)
+  ConvolutionDepthwiseNPUOp(const schema::Primitive *primitive, const std::vector<mindspore::MSTensor> &in_tensors,
+                            const std::vector<mindspore::MSTensor> &out_tensors, std::string name)
       : ConvolutionBaseNPUOp(primitive, in_tensors, out_tensors, name) {}
 
   ~ConvolutionDepthwiseNPUOp() override;
 
-  int IsSupport(const schema::Primitive *primitive, const std::vector<tensor::MSTensor *> &in_tensors,
-                const std::vector<tensor::MSTensor *> &out_tensors) override {
+  int IsSupport(const schema::Primitive *primitive, const std::vector<mindspore::MSTensor> &in_tensors,
+                const std::vector<mindspore::MSTensor> &out_tensors) override {
     return RET_OK;
   }
 
-  int Init(const schema::Primitive *primitive, const std::vector<tensor::MSTensor *> &in_tensors,
-           const std::vector<tensor::MSTensor *> &out_tensors) override;
+  int Init(const schema::Primitive *primitive, const std::vector<mindspore::MSTensor> &in_tensors,
+           const std::vector<mindspore::MSTensor> &out_tensors) override;
 
-  int SetNPUInputs(const std::vector<tensor::MSTensor *> &in_tensors,
-                   const std::vector<tensor::MSTensor *> &out_tensors,
+  int SetNPUInputs(const std::vector<mindspore::MSTensor> &in_tensors,
+                   const std::vector<mindspore::MSTensor> &out_tensors,
                    const std::vector<ge::Operator *> &npu_inputs) override;
 
   ge::Operator *GetNPUOp() override;

mindspore/lite/src/delegate/npu/op/convolution_npu.cc

@@ -17,8 +17,8 @@
 #include "src/delegate/npu/op/convolution_npu.h"
 #include "src/delegate/npu/op/convolution_depthwise_npu.h"
 namespace mindspore {
-int ConvolutionNPUOp::IsSupport(const schema::Primitive *primitive, const std::vector<tensor::MSTensor *> &in_tensors,
-                                const std::vector<tensor::MSTensor *> &out_tensors) {
+int ConvolutionNPUOp::IsSupport(const schema::Primitive *primitive, const std::vector<mindspore::MSTensor> &in_tensors,
+                                const std::vector<mindspore::MSTensor> &out_tensors) {
   auto conv_prim = primitive->value_as_Conv2DFusion();
   if (conv_prim == nullptr) {
     MS_LOG(ERROR) << "Get null primitive value for op ." << name_;
@@ -26,7 +26,7 @@ int ConvolutionNPUOp::IsSupport(const schema::Primitive *primitive, const std::v
   }
   auto stride_h = static_cast<int>(*(conv_prim->stride()->begin()));
   auto stride_w = static_cast<int>(*(conv_prim->stride()->begin() + 1));
-  auto in_shape = in_tensors[0]->shape();  // default format: nhwc, RunPass not called
+  auto in_shape = in_tensors[0].Shape();  // default format: nhwc, RunPass not called
   if (stride_h > in_shape[1] || stride_w > in_shape[2]) {
     MS_LOG(WARNING) << "Npu convolution does not support stride greater than input size.";
     return RET_NOT_SUPPORT;
@@ -61,8 +61,8 @@ int ConvolutionNPUOp::SetConvParam(const schema::Conv2DFusion *conv_prim) {
   return RET_OK;
 }
 
-int ConvolutionNPUOp::Init(const schema::Primitive *primitive, const std::vector<tensor::MSTensor *> &in_tensors,
-                           const std::vector<tensor::MSTensor *> &out_tensors) {
+int ConvolutionNPUOp::Init(const schema::Primitive *primitive, const std::vector<mindspore::MSTensor> &in_tensors,
+                           const std::vector<mindspore::MSTensor> &out_tensors) {
   // set conv attr param
   conv_ = new (std::nothrow) hiai::op::Convolution(name_ + "_conv");
   if (conv_ == nullptr) {
@@ -90,8 +90,8 @@ int ConvolutionNPUOp::Init(const schema::Primitive *primitive, const std::vector
   return RET_OK;
 }
 
-int ConvolutionNPUOp::SetNPUInputs(const std::vector<tensor::MSTensor *> &in_tensors,
-                                   const std::vector<tensor::MSTensor *> &out_tensors,
+int ConvolutionNPUOp::SetNPUInputs(const std::vector<mindspore::MSTensor> &in_tensors,
+                                   const std::vector<mindspore::MSTensor> &out_tensors,
                                    const std::vector<ge::Operator *> &npu_inputs) {
   auto ret = InitWeightConst(in_tensors);
   if (ret != RET_OK) {
@@ -125,30 +125,30 @@ ConvolutionNPUOp::~ConvolutionNPUOp() {
     conv_ = nullptr;
   }
 }
-
-NPUOp *GetNPUConvOp(const schema::Primitive *primitive, const std::vector<tensor::MSTensor *> &in_tensors,
-                    const std::vector<tensor::MSTensor *> &out_tensors, std::string name) {
-  auto shape = out_tensors.front()->shape();
+NPUOp *GetNPUConvOp(const schema::Primitive *primitive, const std::vector<mindspore::MSTensor> &in_tensors,
+                    const std::vector<mindspore::MSTensor> &out_tensors, std::string name) {
+  auto shape = out_tensors.front().Shape();
   if (std::find(shape.begin(), shape.end(), -1) != shape.end()) {
     MS_LOG(ERROR) << "NPU does not support runtime inference shape.";
     return nullptr;
   }
 
-  if (in_tensors[0]->shape().size() > 4) {
+  if (in_tensors[0].Shape().size() > 4) {
     MS_LOG(ERROR) << "Npu does not support input tensor dims greater than 4";
     return nullptr;
   }
 
-  if (in_tensors[0]->data_type() != kNumberTypeFloat32 && in_tensors[0]->data_type() != kNumberTypeFloat16) {
-    MS_LOG(ERROR) << "Npu does not support datatype " << in_tensors[0]->data_type();
+  if (in_tensors[0].DataType() != DataType::kNumberTypeFloat32 &&
+      in_tensors[0].DataType() != DataType::kNumberTypeFloat16) {
+    MS_LOG(ERROR) << "Npu does not support datatype " << static_cast<int>(in_tensors[0].DataType());
     return nullptr;
   }
 
   NPUOp *op = nullptr;
   auto conv_prim = primitive->value_as_Conv2DFusion();
   auto group = static_cast<int>(conv_prim->group());
-  auto input_channel = in_tensors.front()->shape()[3];
-  auto output_channel = out_tensors.front()->shape()[3];
+  auto input_channel = in_tensors.front().Shape()[3];
+  auto output_channel = out_tensors.front().Shape()[3];
   if (group == input_channel && group == output_channel) {
     op = new (std::nothrow) ConvolutionDepthwiseNPUOp(primitive, in_tensors, out_tensors, name);
   } else {

mindspore/lite/src/delegate/npu/op/convolution_npu.h

@@ -23,20 +23,20 @@
 namespace mindspore {
 class ConvolutionNPUOp : public ConvolutionBaseNPUOp {
  public:
-  ConvolutionNPUOp(const schema::Primitive *primitive, const std::vector<tensor::MSTensor *> &in_tensors,
-                   const std::vector<tensor::MSTensor *> &out_tensors, std::string name)
+  ConvolutionNPUOp(const schema::Primitive *primitive, const std::vector<mindspore::MSTensor> &in_tensors,
+                   const std::vector<mindspore::MSTensor> &out_tensors, std::string name)
       : ConvolutionBaseNPUOp(primitive, in_tensors, out_tensors, name) {}
 
   ~ConvolutionNPUOp() override;
 
-  int IsSupport(const schema::Primitive *primitive, const std::vector<tensor::MSTensor *> &in_tensors,
-                const std::vector<tensor::MSTensor *> &out_tensors) override;
+  int IsSupport(const schema::Primitive *primitive, const std::vector<mindspore::MSTensor> &in_tensors,
+                const std::vector<mindspore::MSTensor> &out_tensors) override;
 
-  int Init(const schema::Primitive *primitive, const std::vector<tensor::MSTensor *> &in_tensors,
-           const std::vector<tensor::MSTensor *> &out_tensors) override;
+  int Init(const schema::Primitive *primitive, const std::vector<mindspore::MSTensor> &in_tensors,
+           const std::vector<mindspore::MSTensor> &out_tensors) override;
 
-  int SetNPUInputs(const std::vector<tensor::MSTensor *> &in_tensors,
-                   const std::vector<tensor::MSTensor *> &out_tensors,
+  int SetNPUInputs(const std::vector<mindspore::MSTensor> &in_tensors,
+                   const std::vector<mindspore::MSTensor> &out_tensors,
                    const std::vector<ge::Operator *> &npu_inputs) override;
 
   ge::Operator *GetNPUOp() override;
@@ -47,7 +47,7 @@ class ConvolutionNPUOp : public ConvolutionBaseNPUOp {
   hiai::op::Convolution *conv_ = nullptr;
 };
 
-NPUOp *GetNPUConvOp(const schema::Primitive *primitive, const std::vector<tensor::MSTensor *> &in_tensors,
-                    const std::vector<tensor::MSTensor *> &out_tensors, std::string name);
+NPUOp *GetNPUConvOp(const schema::Primitive *primitive, const std::vector<mindspore::MSTensor> &in_tensors,
+                    const std::vector<mindspore::MSTensor> &out_tensors, std::string name);
 }  // namespace mindspore
 #endif  // MINDSPORE_LITE_SRC_RUNTIME_DELEGATE_NPU_OP_CONVOLUTION_NPU_H_

mindspore/lite/src/delegate/npu/op/crop_and_resize_npu.cc

@@ -16,8 +16,9 @@
 
 #include "src/delegate/npu/op/crop_and_resize_npu.h"
 namespace mindspore {
-int CropAndResizeNPUOp::IsSupport(const schema::Primitive *primitive, const std::vector<tensor::MSTensor *> &in_tensors,
-                                  const std::vector<tensor::MSTensor *> &out_tensors) {
+int CropAndResizeNPUOp::IsSupport(const schema::Primitive *primitive,
+                                  const std::vector<mindspore::MSTensor> &in_tensors,
+                                  const std::vector<mindspore::MSTensor> &out_tensors) {
   if (in_tensors.size() < 4) {
     MS_LOG(WARNING) << "NPU CropAndResize got nput inputs size < 4";
     return RET_NOT_SUPPORT;
@@ -37,8 +38,8 @@ int CropAndResizeNPUOp::IsSupport(const schema::Primitive *primitive, const std:
   return RET_OK;
 }
 
-int CropAndResizeNPUOp::Init(const schema::Primitive *primitive, const std::vector<tensor::MSTensor *> &in_tensors,
-                             const std::vector<tensor::MSTensor *> &out_tensors) {
+int CropAndResizeNPUOp::Init(const schema::Primitive *primitive, const std::vector<mindspore::MSTensor> &in_tensors,
+                             const std::vector<mindspore::MSTensor> &out_tensors) {
   crop_and_resize_ = new (std::nothrow) hiai::op::CropAndResize(name_);
   if (crop_and_resize_ == nullptr) {
     MS_LOG(ERROR) << name_ << " op is nullptr";
@@ -62,8 +63,8 @@ int CropAndResizeNPUOp::Init(const schema::Primitive *primitive, const std::vect
   return RET_OK;
 }
 
-int CropAndResizeNPUOp::SetNPUInputs(const std::vector<tensor::MSTensor *> &in_tensors,
-                                     const std::vector<tensor::MSTensor *> &out_tensors,
+int CropAndResizeNPUOp::SetNPUInputs(const std::vector<mindspore::MSTensor> &in_tensors,
+                                     const std::vector<mindspore::MSTensor> &out_tensors,
                                      const std::vector<ge::Operator *> &npu_inputs) {
   crop_and_resize_->set_input_x(*npu_inputs[0]);
   crop_and_resize_->set_input_boxes(*npu_inputs[1]);

mindspore/lite/src/delegate/npu/op/crop_and_resize_npu.h

@@ -24,20 +24,20 @@
 namespace mindspore {
 class CropAndResizeNPUOp : public NPUOp {
  public:
-  CropAndResizeNPUOp(const schema::Primitive *primitive, const std::vector<tensor::MSTensor *> &in_tensors,
-                     const std::vector<tensor::MSTensor *> &out_tensors, std::string name)
+  CropAndResizeNPUOp(const schema::Primitive *primitive, const std::vector<mindspore::MSTensor> &in_tensors,
+                     const std::vector<mindspore::MSTensor> &out_tensors, std::string name)
       : NPUOp(primitive, in_tensors, out_tensors, name) {}
 
   ~CropAndResizeNPUOp() override;
 
-  int IsSupport(const schema::Primitive *primitive, const std::vector<tensor::MSTensor *> &in_tensors,
-                const std::vector<tensor::MSTensor *> &out_tensors) override;
+  int IsSupport(const schema::Primitive *primitive, const std::vector<mindspore::MSTensor> &in_tensors,
+                const std::vector<mindspore::MSTensor> &out_tensors) override;
 
-  int Init(const schema::Primitive *primitive, const std::vector<tensor::MSTensor *> &in_tensors,
-           const std::vector<tensor::MSTensor *> &out_tensors) override;
+  int Init(const schema::Primitive *primitive, const std::vector<mindspore::MSTensor> &in_tensors,
+           const std::vector<mindspore::MSTensor> &out_tensors) override;
 
-  int SetNPUInputs(const std::vector<tensor::MSTensor *> &in_tensors,
-                   const std::vector<tensor::MSTensor *> &out_tensors,
+  int SetNPUInputs(const std::vector<mindspore::MSTensor> &in_tensors,
+                   const std::vector<mindspore::MSTensor> &out_tensors,
                    const std::vector<ge::Operator *> &npu_inputs) override;
 
   ge::Operator *GetNPUOp() override;

mindspore/lite/src/delegate/npu/op/deconvolution_npu.cc

@@ -18,8 +18,9 @@
 #include "src/delegate/npu/npu_converter_utils.h"
 
 namespace mindspore {
-int DeconvolutionNPUOp::IsSupport(const schema::Primitive *primitive, const std::vector<tensor::MSTensor *> &in_tensors,
-                                  const std::vector<tensor::MSTensor *> &out_tensors) {
+int DeconvolutionNPUOp::IsSupport(const schema::Primitive *primitive,
+                                  const std::vector<mindspore::MSTensor> &in_tensors,
+                                  const std::vector<mindspore::MSTensor> &out_tensors) {
   auto deconv_prim = primitive->value_as_Conv2dTransposeFusion();
   if (deconv_prim == nullptr) {
     MS_LOG(ERROR) << "Get null primitive value for op ." << name_;
@@ -59,8 +60,8 @@ int DeconvolutionNPUOp::SetDeconvParam(const schema::Conv2dTransposeFusion *conv
   return RET_OK;
 }
 
-int DeconvolutionNPUOp::Init(const schema::Primitive *primitive, const std::vector<tensor::MSTensor *> &in_tensors,
-                             const std::vector<tensor::MSTensor *> &out_tensors) {
+int DeconvolutionNPUOp::Init(const schema::Primitive *primitive, const std::vector<mindspore::MSTensor> &in_tensors,
+                             const std::vector<mindspore::MSTensor> &out_tensors) {
   // set deconv attr param
   deconv_ = new (std::nothrow) hiai::op::ConvTranspose(name_ + "_deconv");
   if (deconv_ == nullptr) {
@@ -89,8 +90,8 @@ int DeconvolutionNPUOp::Init(const schema::Primitive *primitive, const std::vect
   return RET_OK;
 }
 
-int DeconvolutionNPUOp::SetNPUInputs(const std::vector<tensor::MSTensor *> &in_tensors,
-                                     const std::vector<tensor::MSTensor *> &out_tensors,
+int DeconvolutionNPUOp::SetNPUInputs(const std::vector<mindspore::MSTensor> &in_tensors,
+                                     const std::vector<mindspore::MSTensor> &out_tensors,
                                      const std::vector<ge::Operator *> &npu_inputs) {
   auto ret = InitWeightConst(in_tensors);
   if (ret != RET_OK) {

mindspore/lite/src/delegate/npu/op/deconvolution_npu.h

@@ -24,19 +24,19 @@
 namespace mindspore {
 class DeconvolutionNPUOp : public ConvolutionBaseNPUOp {
  public:
-  DeconvolutionNPUOp(const schema::Primitive *primitive, const std::vector<tensor::MSTensor *> &in_tensors,
-                     const std::vector<tensor::MSTensor *> &out_tensors, std::string name)
+  DeconvolutionNPUOp(const schema::Primitive *primitive, const std::vector<mindspore::MSTensor> &in_tensors,
+                     const std::vector<mindspore::MSTensor> &out_tensors, std::string name)
       : ConvolutionBaseNPUOp(primitive, in_tensors, out_tensors, name) {}
   ~DeconvolutionNPUOp() override;
 
-  int IsSupport(const schema::Primitive *primitive, const std::vector<tensor::MSTensor *> &in_tensors,
-                const std::vector<tensor::MSTensor *> &out_tensors) override;
+  int IsSupport(const schema::Primitive *primitive, const std::vector<mindspore::MSTensor> &in_tensors,
+                const std::vector<mindspore::MSTensor> &out_tensors) override;
 
-  int Init(const schema::Primitive *primitive, const std::vector<tensor::MSTensor *> &in_tensors,
-           const std::vector<tensor::MSTensor *> &out_tensors) override;
+  int Init(const schema::Primitive *primitive, const std::vector<mindspore::MSTensor> &in_tensors,
+           const std::vector<mindspore::MSTensor> &out_tensors) override;
 
-  int SetNPUInputs(const std::vector<tensor::MSTensor *> &in_tensors,
-                   const std::vector<tensor::MSTensor *> &out_tensors,
+  int SetNPUInputs(const std::vector<mindspore::MSTensor> &in_tensors,
+                   const std::vector<mindspore::MSTensor> &out_tensors,
                    const std::vector<ge::Operator *> &npu_inputs) override;
 
   ge::Operator *GetNPUOp() override;

mindspore/lite/src/delegate/npu/op/eltwise_npu.cc

@@ -20,8 +20,8 @@
 #include "src/delegate/npu/npu_converter_utils.h"
 
 namespace mindspore {
-int EltwiseNPUOp::Init(const schema::Primitive *primitive, const std::vector<tensor::MSTensor *> &in_tensors,
-                       const std::vector<tensor::MSTensor *> &out_tensors) {
+int EltwiseNPUOp::Init(const schema::Primitive *primitive, const std::vector<mindspore::MSTensor> &in_tensors,
+                       const std::vector<mindspore::MSTensor> &out_tensors) {
   eltwise_ = new (std::nothrow) hiai::op::Eltwise(name_);
   if (eltwise_ == nullptr) {
     MS_LOG(ERROR) << name_ << " op is nullptr";
@@ -39,8 +39,8 @@ int EltwiseNPUOp::Init(const schema::Primitive *primitive, const std::vector<ten
   return RET_OK;
 }
 
-int EltwiseNPUOp::SetNPUInputs(const std::vector<tensor::MSTensor *> &in_tensors,
-                               const std::vector<tensor::MSTensor *> &out_tensors,
+int EltwiseNPUOp::SetNPUInputs(const std::vector<mindspore::MSTensor> &in_tensors,
+                               const std::vector<mindspore::MSTensor> &out_tensors,
                                const std::vector<ge::Operator *> &npu_inputs) {
   for (int i = 0; i < npu_inputs.size(); ++i) {
     eltwise_->set_dynamic_input_x(i + 1, *npu_inputs[i]);

mindspore/lite/src/delegate/npu/op/eltwise_npu.h

@@ -24,22 +24,22 @@
 namespace mindspore {
 class EltwiseNPUOp : public NPUOp {
  public:
-  EltwiseNPUOp(const schema::Primitive *primitive, const std::vector<tensor::MSTensor *> &in_tensors,
-               const std::vector<tensor::MSTensor *> &out_tensors, std::string name)
+  EltwiseNPUOp(const schema::Primitive *primitive, const std::vector<mindspore::MSTensor> &in_tensors,
+               const std::vector<mindspore::MSTensor> &out_tensors, std::string name)
       : NPUOp(primitive, in_tensors, out_tensors, name) {}
 
   ~EltwiseNPUOp() override;
 
-  int IsSupport(const schema::Primitive *primitive, const std::vector<tensor::MSTensor *> &in_tensors,
-                const std::vector<tensor::MSTensor *> &out_tensors) override {
+  int IsSupport(const schema::Primitive *primitive, const std::vector<mindspore::MSTensor> &in_tensors,
+                const std::vector<mindspore::MSTensor> &out_tensors) override {
     return RET_OK;
   }
 
-  int Init(const schema::Primitive *primitive, const std::vector<tensor::MSTensor *> &in_tensors,
-           const std::vector<tensor::MSTensor *> &out_tensors) override;
+  int Init(const schema::Primitive *primitive, const std::vector<mindspore::MSTensor> &in_tensors,
+           const std::vector<mindspore::MSTensor> &out_tensors) override;
 
-  int SetNPUInputs(const std::vector<tensor::MSTensor *> &in_tensors,
-                   const std::vector<tensor::MSTensor *> &out_tensors,
+  int SetNPUInputs(const std::vector<mindspore::MSTensor> &in_tensors,
+                   const std::vector<mindspore::MSTensor> &out_tensors,
                    const std::vector<ge::Operator *> &npu_inputs) override;
 
   ge::Operator *GetNPUOp() override;

mindspore/lite/src/delegate/npu/op/expand_dims_npu.cc

@@ -19,8 +19,8 @@
 #include "src/delegate/npu/npu_converter_utils.h"
 
 namespace mindspore {
-int ExpandDimsNPUOp::Init(const schema::Primitive *primitive, const std::vector<tensor::MSTensor *> &in_tensors,
-                          const std::vector<tensor::MSTensor *> &out_tensors) {
+int ExpandDimsNPUOp::Init(const schema::Primitive *primitive, const std::vector<mindspore::MSTensor> &in_tensors,
+                          const std::vector<mindspore::MSTensor> &out_tensors) {
   expand_dims_ = new (std::nothrow) hiai::op::ExpandDims(name_);
   if (expand_dims_ == nullptr) {
     MS_LOG(ERROR) << name_ << " op is nullptr";
@@ -29,8 +29,8 @@ int ExpandDimsNPUOp::Init(const schema::Primitive *primitive, const std::vector<
   return RET_OK;
 }
 
-int ExpandDimsNPUOp::SetNPUInputs(const std::vector<tensor::MSTensor *> &in_tensors,
-                                  const std::vector<tensor::MSTensor *> &out_tensors,
+int ExpandDimsNPUOp::SetNPUInputs(const std::vector<mindspore::MSTensor> &in_tensors,
+                                  const std::vector<mindspore::MSTensor> &out_tensors,
                                   const std::vector<ge::Operator *> &npu_inputs) {
   expand_dims_->set_input_x(*npu_inputs[0]);
   expand_dims_->set_input_axis(*npu_inputs[1]);

mindspore/lite/src/delegate/npu/op/expand_dims_npu.h

@@ -24,21 +24,21 @@
 namespace mindspore {
 class ExpandDimsNPUOp : public NPUOp {
  public:
-  ExpandDimsNPUOp(const schema::Primitive *primitive, const std::vector<tensor::MSTensor *> &in_tensors,
-                  const std::vector<tensor::MSTensor *> &out_tensors, std::string name)
+  ExpandDimsNPUOp(const schema::Primitive *primitive, const std::vector<mindspore::MSTensor> &in_tensors,
+                  const std::vector<mindspore::MSTensor> &out_tensors, std::string name)
       : NPUOp(primitive, in_tensors, out_tensors, name) {}
   ~ExpandDimsNPUOp() override;
 
-  int IsSupport(const schema::Primitive *primitive, const std::vector<tensor::MSTensor *> &in_tensors,
-                const std::vector<tensor::MSTensor *> &out_tensors) override {
+  int IsSupport(const schema::Primitive *primitive, const std::vector<mindspore::MSTensor> &in_tensors,
+                const std::vector<mindspore::MSTensor> &out_tensors) override {
     return RET_OK;
   }
 
-  int Init(const schema::Primitive *primitive, const std::vector<tensor::MSTensor *> &in_tensors,
-           const std::vector<tensor::MSTensor *> &out_tensors) override;
+  int Init(const schema::Primitive *primitive, const std::vector<mindspore::MSTensor> &in_tensors,
+           const std::vector<mindspore::MSTensor> &out_tensors) override;
 
-  int SetNPUInputs(const std::vector<tensor::MSTensor *> &in_tensors,
-                   const std::vector<tensor::MSTensor *> &out_tensors,
+  int SetNPUInputs(const std::vector<mindspore::MSTensor> &in_tensors,
+                   const std::vector<mindspore::MSTensor> &out_tensors,
                    const std::vector<ge::Operator *> &npu_inputs) override;
 
   ge::Operator *GetNPUOp() override;

mindspore/lite/src/delegate/npu/op/fullconnection_npu.cc

@@ -19,15 +19,15 @@
 #include "src/delegate/npu/npu_converter_utils.h"
 
 namespace mindspore {
-int FullconnectionNPUOp::Init(const schema::Primitive *primitive, const std::vector<tensor::MSTensor *> &in_tensors,
-                              const std::vector<tensor::MSTensor *> &out_tensors) {
+int FullconnectionNPUOp::Init(const schema::Primitive *primitive, const std::vector<mindspore::MSTensor> &in_tensors,
+                              const std::vector<mindspore::MSTensor> &out_tensors) {
   auto fc_prim = primitive->value_as_FullConnection();
   if (fc_prim == nullptr) {
     MS_LOG(ERROR) << "Get null primitive value for op ." << name_;
     return RET_ERROR;
   }
   act_type_ = fc_prim->activation_type();
-  auto input_shape = in_tensors[0]->shape();
+  auto input_shape = in_tensors[0].Shape();
   reshape_ = new (std::nothrow) hiai::op::Reshape(name_ + "_reshape");
   if (reshape_ == nullptr) {
     MS_LOG(ERROR) << "New reshape operator for fullconnection op " << name_ << " failed.";
@@ -39,7 +39,7 @@ int FullconnectionNPUOp::Init(const schema::Primitive *primitive, const std::vec
     col *= input_shape[i];
   }
   reshape_op_ = new (std::nothrow) hiai::op::Const(name_ + "_reshape_data");
-  vector<int> reshape_data = {input_shape[0], col};
+  vector<int> reshape_data = {static_cast<int>(input_shape[0]), col};
   ge::TensorDesc reshape_tensor_desc(ge::Shape({2}), ge::FORMAT_NCHW, ge::DT_FLOAT);
   ge::TensorPtr reshape_tensor = std::make_shared<hiai::Tensor>(reshape_tensor_desc);
   reshape_tensor->SetData(reinterpret_cast<uint8_t *>(reshape_data.data()), 2 * sizeof(float));
@@ -54,8 +54,8 @@ int FullconnectionNPUOp::Init(const schema::Primitive *primitive, const std::vec
   return RET_OK;
 }
 
-int FullconnectionNPUOp::SetNPUInputs(const std::vector<tensor::MSTensor *> &in_tensors,
-                                      const std::vector<tensor::MSTensor *> &out_tensors,
+int FullconnectionNPUOp::SetNPUInputs(const std::vector<mindspore::MSTensor> &in_tensors,
+                                      const std::vector<mindspore::MSTensor> &out_tensors,
                                       const std::vector<ge::Operator *> &npu_inputs) {
   reshape_->set_input_x(*npu_inputs[0]);
   fc_->set_input_x1(*reshape_);

mindspore/lite/src/delegate/npu/op/fullconnection_npu.h

@@ -24,22 +24,22 @@
 namespace mindspore {
 class FullconnectionNPUOp : public ConvolutionBaseNPUOp {
  public:
-  FullconnectionNPUOp(const schema::Primitive *primitive, const std::vector<tensor::MSTensor *> &in_tensors,
-                      const std::vector<tensor::MSTensor *> &out_tensors, std::string name)
+  FullconnectionNPUOp(const schema::Primitive *primitive, const std::vector<mindspore::MSTensor> &in_tensors,
+                      const std::vector<mindspore::MSTensor> &out_tensors, std::string name)
       : ConvolutionBaseNPUOp(primitive, in_tensors, out_tensors, name) {}
 
   ~FullconnectionNPUOp() override;
 
-  int IsSupport(const schema::Primitive *primitive, const std::vector<tensor::MSTensor *> &in_tensors,
-                const std::vector<tensor::MSTensor *> &out_tensors) override {
+  int IsSupport(const schema::Primitive *primitive, const std::vector<mindspore::MSTensor> &in_tensors,
+                const std::vector<mindspore::MSTensor> &out_tensors) override {
     return RET_OK;
   }
 
-  int Init(const schema::Primitive *primitive, const std::vector<tensor::MSTensor *> &in_tensors,
-           const std::vector<tensor::MSTensor *> &out_tensors) override;
+  int Init(const schema::Primitive *primitive, const std::vector<mindspore::MSTensor> &in_tensors,
+           const std::vector<mindspore::MSTensor> &out_tensors) override;
 
-  int SetNPUInputs(const std::vector<tensor::MSTensor *> &in_tensors,
-                   const std::vector<tensor::MSTensor *> &out_tensors,
+  int SetNPUInputs(const std::vector<mindspore::MSTensor> &in_tensors,
+                   const std::vector<mindspore::MSTensor> &out_tensors,
                    const std::vector<ge::Operator *> &npu_inputs) override;
 
   ge::Operator *GetNPUOp() override;

mindspore/lite/src/delegate/npu/op/gather_npu.cc

@@ -17,14 +17,14 @@
 #include "src/delegate/npu/op/gather_npu.h"
 
 namespace mindspore {
-int GatherNPUOp::IsSupport(const schema::Primitive *primitive, const std::vector<tensor::MSTensor *> &in_tensors,
-                           const std::vector<tensor::MSTensor *> &out_tensors) {
-  if (in_tensors[1]->data_type() != kNumberTypeInt32) {
+int GatherNPUOp::IsSupport(const schema::Primitive *primitive, const std::vector<mindspore::MSTensor> &in_tensors,
+                           const std::vector<mindspore::MSTensor> &out_tensors) {
+  if (in_tensors[1].DataType() != DataType::kNumberTypeInt32) {
     MS_LOG(WARNING) << "Gather indices only support Int32";
     return RET_NOT_SUPPORT;
   }
-  if (in_tensors.size() >= 3 && in_tensors[2]->ElementsNum() == 1) {
-    axis_ = static_cast<int *>(in_tensors[2]->data())[0];
+  if (in_tensors.size() >= 3 && in_tensors[2].ElementNum() == 1) {
+    axis_ = static_cast<const int *>(in_tensors[2].Data().get())[0];
   } else {
     MS_LOG(WARNING) << "NPU axis is attribute.";
     return RET_NOT_SUPPORT;
@@ -32,8 +32,8 @@ int GatherNPUOp::IsSupport(const schema::Primitive *primitive, const std::vector
   return RET_OK;
 }
 
-int GatherNPUOp::Init(const schema::Primitive *primitive, const std::vector<tensor::MSTensor *> &in_tensors,
-                      const std::vector<tensor::MSTensor *> &out_tensors) {
+int GatherNPUOp::Init(const schema::Primitive *primitive, const std::vector<mindspore::MSTensor> &in_tensors,
+                      const std::vector<mindspore::MSTensor> &out_tensors) {
   gather_ = new (std::nothrow) hiai::op::GatherV2D(name_);
   if (gather_ == nullptr) {
     MS_LOG(ERROR) << name_ << " op is nullptr";
@@ -43,8 +43,8 @@ int GatherNPUOp::Init(const schema::Primitive *primitive, const std::vector<tens
   return RET_OK;
 }
 
-int GatherNPUOp::SetNPUInputs(const std::vector<tensor::MSTensor *> &in_tensors,
-                              const std::vector<tensor::MSTensor *> &out_tensors,
+int GatherNPUOp::SetNPUInputs(const std::vector<mindspore::MSTensor> &in_tensors,
+                              const std::vector<mindspore::MSTensor> &out_tensors,
                               const std::vector<ge::Operator *> &npu_inputs) {
   gather_->set_input_x(*npu_inputs[0]);
   gather_->set_input_indices(*npu_inputs[1]);

mindspore/lite/src/delegate/npu/op/gather_npu.h

@@ -24,20 +24,20 @@
 namespace mindspore {
 class GatherNPUOp : public NPUOp {
  public:
-  GatherNPUOp(const schema::Primitive *primitive, const std::vector<tensor::MSTensor *> &in_tensors,
-              const std::vector<tensor::MSTensor *> &out_tensors, std::string name)
+  GatherNPUOp(const schema::Primitive *primitive, const std::vector<mindspore::MSTensor> &in_tensors,
+              const std::vector<mindspore::MSTensor> &out_tensors, std::string name)
       : NPUOp(primitive, in_tensors, out_tensors, name) {}
 
   ~GatherNPUOp() override;
 
-  int IsSupport(const schema::Primitive *primitive, const std::vector<tensor::MSTensor *> &in_tensors,
-                const std::vector<tensor::MSTensor *> &out_tensors) override;
+  int IsSupport(const schema::Primitive *primitive, const std::vector<mindspore::MSTensor> &in_tensors,
+                const std::vector<mindspore::MSTensor> &out_tensors) override;
 
-  int Init(const schema::Primitive *primitive, const std::vector<tensor::MSTensor *> &in_tensors,
-           const std::vector<tensor::MSTensor *> &out_tensors) override;
+  int Init(const schema::Primitive *primitive, const std::vector<mindspore::MSTensor> &in_tensors,
+           const std::vector<mindspore::MSTensor> &out_tensors) override;
 
-  int SetNPUInputs(const std::vector<tensor::MSTensor *> &in_tensors,
-                   const std::vector<tensor::MSTensor *> &out_tensors,
+  int SetNPUInputs(const std::vector<mindspore::MSTensor> &in_tensors,
+                   const std::vector<mindspore::MSTensor> &out_tensors,
                    const std::vector<ge::Operator *> &npu_inputs) override;
 
   ge::Operator *GetNPUOp() override;

mindspore/lite/src/delegate/npu/op/instance_norm_npu.cc

@@ -15,12 +15,11 @@
  */
 
 #include "src/delegate/npu/op/instance_norm_npu.h"
-#include <memory>
 #include "src/delegate/npu/npu_converter_utils.h"
 
 namespace mindspore {
-int InstanceNormNPUOp::Init(const schema::Primitive *primitive, const std::vector<tensor::MSTensor *> &in_tensors,
-                            const std::vector<tensor::MSTensor *> &out_tensors) {
+int InstanceNormNPUOp::Init(const schema::Primitive *primitive, const std::vector<mindspore::MSTensor> &in_tensors,
+                            const std::vector<mindspore::MSTensor> &out_tensors) {
   instance_norm_ = new (std::nothrow) hiai::op::InstanceNorm(name_);
   if (instance_norm_ == nullptr) {
     MS_LOG(ERROR) << "New instance norm npu operator for op " << name_ << " failed.";
@@ -35,12 +34,12 @@ int InstanceNormNPUOp::Init(const schema::Primitive *primitive, const std::vecto
   return RET_OK;
 }
 
-int InstanceNormNPUOp::SetNPUInputs(const std::vector<tensor::MSTensor *> &in_tensors,
-                                    const std::vector<tensor::MSTensor *> &out_tensors,
+int InstanceNormNPUOp::SetNPUInputs(const std::vector<mindspore::MSTensor> &in_tensors,
+                                    const std::vector<mindspore::MSTensor> &out_tensors,
                                     const std::vector<ge::Operator *> &npu_inputs) {
   instance_norm_->set_input_x(*npu_inputs[0]);
 
-  auto gamma_shape = in_tensors[1]->shape();
+  auto gamma_shape = in_tensors[1].Shape();
   auto gamma_tensor = ConverterToNPUTensor(in_tensors[1]);
   if (gamma_tensor == nullptr) {
     MS_LOG(ERROR) << "Get gamma_tensor failed.";
@@ -56,7 +55,7 @@ int InstanceNormNPUOp::SetNPUInputs(const std::vector<tensor::MSTensor *> &in_te
   gamma_->set_attr_value(gamma_tensor);
   instance_norm_->set_input_gamma(*gamma_);
 
-  auto beta_shape = in_tensors[2]->shape();
+  auto beta_shape = in_tensors[2].Shape();
   auto beta_tensor = ConverterToNPUTensor(in_tensors[2]);
   if (beta_tensor == nullptr) {
     MS_LOG(ERROR) << "Get beta_tensor failed.";

mindspore/lite/src/delegate/npu/op/instance_norm_npu.h

@@ -24,22 +24,22 @@
 namespace mindspore {
 class InstanceNormNPUOp : public NPUOp {
  public:
-  InstanceNormNPUOp(const schema::Primitive *primitive, const std::vector<tensor::MSTensor *> &in_tensors,
-                    const std::vector<tensor::MSTensor *> &out_tensors, std::string name)
+  InstanceNormNPUOp(const schema::Primitive *primitive, const std::vector<mindspore::MSTensor> &in_tensors,
+                    const std::vector<mindspore::MSTensor> &out_tensors, std::string name)
       : NPUOp(primitive, in_tensors, out_tensors, name) {}
 
   ~InstanceNormNPUOp() override;
 
-  int IsSupport(const schema::Primitive *primitive, const std::vector<tensor::MSTensor *> &in_tensors,
-                const std::vector<tensor::MSTensor *> &out_tensors) override {
+  int IsSupport(const schema::Primitive *primitive, const std::vector<mindspore::MSTensor> &in_tensors,
+                const std::vector<mindspore::MSTensor> &out_tensors) override {
     return RET_OK;
   }
 
-  int Init(const schema::Primitive *primitive, const std::vector<tensor::MSTensor *> &in_tensors,
-           const std::vector<tensor::MSTensor *> &out_tensors) override;
+  int Init(const schema::Primitive *primitive, const std::vector<mindspore::MSTensor> &in_tensors,
+           const std::vector<mindspore::MSTensor> &out_tensors) override;
 
-  int SetNPUInputs(const std::vector<tensor::MSTensor *> &in_tensors,
-                   const std::vector<tensor::MSTensor *> &out_tensors,
+  int SetNPUInputs(const std::vector<mindspore::MSTensor> &in_tensors,
+                   const std::vector<mindspore::MSTensor> &out_tensors,
                    const std::vector<ge::Operator *> &npu_inputs) override;
 
   ge::Operator *GetNPUOp() override;

mindspore/lite/src/delegate/npu/op/matmul_npu.cc

@@ -15,21 +15,20 @@
  */
 
 #include "src/delegate/npu/op/matmul_npu.h"
-#include <memory>
 #include "src/delegate/npu/npu_converter_utils.h"
 namespace mindspore {
-int MatMulNPUOp::IsSupport(const schema::Primitive *primitive, const std::vector<tensor::MSTensor *> &in_tensors,
-                           const std::vector<tensor::MSTensor *> &out_tensors) {
+int MatMulNPUOp::IsSupport(const schema::Primitive *primitive, const std::vector<mindspore::MSTensor> &in_tensors,
+                           const std::vector<mindspore::MSTensor> &out_tensors) {
   if (in_tensors.size() == 3) {
-    if (in_tensors[2]->shape().size() != 1) {
+    if (in_tensors[2].Shape().size() != 1) {
       return RET_NOT_SUPPORT;
     }
   }
   return RET_OK;
 }
 
-int MatMulNPUOp::Init(const schema::Primitive *primitive, const std::vector<tensor::MSTensor *> &in_tensors,
-                      const std::vector<tensor::MSTensor *> &out_tensors) {
+int MatMulNPUOp::Init(const schema::Primitive *primitive, const std::vector<mindspore::MSTensor> &in_tensors,
+                      const std::vector<mindspore::MSTensor> &out_tensors) {
   matmul_ = new (std::nothrow) hiai::op::MatMul(name_);
   if (matmul_ == nullptr) {
     MS_LOG(ERROR) << "New matmul npu operator for op " << name_ << " failed.";
@@ -48,8 +47,8 @@ int MatMulNPUOp::Init(const schema::Primitive *primitive, const std::vector<tens
   return RET_OK;
 }
 
-int MatMulNPUOp::SetNPUInputs(const std::vector<tensor::MSTensor *> &in_tensors,
-                              const std::vector<tensor::MSTensor *> &out_tensors,
+int MatMulNPUOp::SetNPUInputs(const std::vector<mindspore::MSTensor> &in_tensors,
+                              const std::vector<mindspore::MSTensor> &out_tensors,
                               const std::vector<ge::Operator *> &npu_inputs) {
   matmul_->set_input_x1(*npu_inputs[0]);
   matmul_->set_input_x2(*npu_inputs[1]);
@@ -60,7 +59,7 @@ int MatMulNPUOp::SetNPUInputs(const std::vector<tensor::MSTensor *> &in_tensors,
       return RET_ERROR;
     }
     add_op_->set_input_x1(*matmul_);
-    auto bias_shape = in_tensors[2]->shape();
+    auto bias_shape = in_tensors[2].Shape();
     auto bias_tensor = ConverterToNPUTensor(in_tensors[2]);
     if (bias_tensor == nullptr) {
       MS_LOG(ERROR) << "Get bias_tensor failed.";
@@ -68,7 +67,7 @@ int MatMulNPUOp::SetNPUInputs(const std::vector<tensor::MSTensor *> &in_tensors,
     }
 
     ge::TensorDesc bias_tensor_desc(ConverterToNPUShape({1, bias_shape[0], 1, 1}));
-    if (out_tensors[0]->shape().size() == 2) {
+    if (out_tensors[0].Shape().size() == 2) {
       bias_tensor_desc.SetShape(ConverterToNPUShape({1, bias_shape[0]}));
     }
     bias_tensor->SetTensorDesc(bias_tensor_desc);

mindspore/lite/src/delegate/npu/op/matmul_npu.h

@@ -24,20 +24,20 @@
 namespace mindspore {
 class MatMulNPUOp : public NPUOp {
  public:
-  MatMulNPUOp(const schema::Primitive *primitive, const std::vector<tensor::MSTensor *> &in_tensors,
-              const std::vector<tensor::MSTensor *> &out_tensors, std::string name)
+  MatMulNPUOp(const schema::Primitive *primitive, const std::vector<mindspore::MSTensor> &in_tensors,
+              const std::vector<mindspore::MSTensor> &out_tensors, std::string name)
       : NPUOp(primitive, in_tensors, out_tensors, name) {}
 
   ~MatMulNPUOp() override;
 
-  int IsSupport(const schema::Primitive *primitive, const std::vector<tensor::MSTensor *> &in_tensors,
-                const std::vector<tensor::MSTensor *> &out_tensors) override;
+  int IsSupport(const schema::Primitive *primitive, const std::vector<mindspore::MSTensor> &in_tensors,
+                const std::vector<mindspore::MSTensor> &out_tensors) override;
 
-  int Init(const schema::Primitive *primitive, const std::vector<tensor::MSTensor *> &in_tensors,
-           const std::vector<tensor::MSTensor *> &out_tensors) override;
+  int Init(const schema::Primitive *primitive, const std::vector<mindspore::MSTensor> &in_tensors,
+           const std::vector<mindspore::MSTensor> &out_tensors) override;
 
-  int SetNPUInputs(const std::vector<tensor::MSTensor *> &in_tensors,
-                   const std::vector<tensor::MSTensor *> &out_tensors,
+  int SetNPUInputs(const std::vector<mindspore::MSTensor> &in_tensors,
+                   const std::vector<mindspore::MSTensor> &out_tensors,
                    const std::vector<ge::Operator *> &npu_inputs) override;
 
   ge::Operator *GetNPUOp() override;

mindspore/lite/src/delegate/npu/op/max_pooling_npu.cc

@@ -16,8 +16,8 @@
 
 #include "src/delegate/npu/op/max_pooling_npu.h"
 namespace mindspore {
-int MaxPoolingNPUOp::IsSupport(const schema::Primitive *primitive, const std::vector<tensor::MSTensor *> &in_tensors,
-                               const std::vector<tensor::MSTensor *> &out_tensors) {
+int MaxPoolingNPUOp::IsSupport(const schema::Primitive *primitive, const std::vector<mindspore::MSTensor> &in_tensors,
+                               const std::vector<mindspore::MSTensor> &out_tensors) {
   auto pooling_prim = primitive->value_as_MaxPoolFusion();
   if (pooling_prim == nullptr) {
     MS_LOG(ERROR) << "Get null primitive value for op ." << name_;
@@ -71,8 +71,8 @@ int MaxPoolingNPUOp::SetPoolingParam(const schema::MaxPoolFusion *pooling_prim)
   return RET_OK;
 }
 
-int MaxPoolingNPUOp::Init(const schema::Primitive *primitive, const std::vector<tensor::MSTensor *> &in_tensors,
-                          const std::vector<tensor::MSTensor *> &out_tensors) {
+int MaxPoolingNPUOp::Init(const schema::Primitive *primitive, const std::vector<mindspore::MSTensor> &in_tensors,
+                          const std::vector<mindspore::MSTensor> &out_tensors) {
   pooling_ = new (std::nothrow) hiai::op::PoolingD(name_ + "_pooling");
   if (pooling_ == nullptr) {
     MS_LOG(ERROR) << "New pooling npu operator for op " << name_ << " failed.";
@@ -99,8 +99,8 @@ int MaxPoolingNPUOp::Init(const schema::Primitive *primitive, const std::vector<
   return RET_OK;
 }
 
-int MaxPoolingNPUOp::SetNPUInputs(const std::vector<tensor::MSTensor *> &in_tensors,
-                                  const std::vector<tensor::MSTensor *> &out_tensors,
+int MaxPoolingNPUOp::SetNPUInputs(const std::vector<mindspore::MSTensor> &in_tensors,
+                                  const std::vector<mindspore::MSTensor> &out_tensors,
                                   const std::vector<ge::Operator *> &npu_inputs) {
   pooling_->set_input_x(*npu_inputs[0]);
   return RET_OK;

mindspore/lite/src/delegate/npu/op/max_pooling_npu.h

@@ -18,25 +18,24 @@
 #define MINDSPORE_LITE_SRC_RUNTIME_DELEGATE_NPU_OP_MAX_POOLING_NPU_H_
 #include <vector>
 #include <string>
-#include "include/graph/op/all_ops.h"
 #include "src/delegate/npu/op/convolution_base_npu.h"
 namespace mindspore {
 class MaxPoolingNPUOp : public ConvolutionBaseNPUOp {
  public:
-  MaxPoolingNPUOp(const schema::Primitive *primitive, const std::vector<tensor::MSTensor *> &in_tensors,
-                  const std::vector<tensor::MSTensor *> &out_tensors, std::string name)
+  MaxPoolingNPUOp(const schema::Primitive *primitive, const std::vector<mindspore::MSTensor> &in_tensors,
+                  const std::vector<mindspore::MSTensor> &out_tensors, std::string name)
       : ConvolutionBaseNPUOp(primitive, in_tensors, out_tensors, name) {}
 
   ~MaxPoolingNPUOp() override;
 
-  int IsSupport(const schema::Primitive *primitive, const std::vector<tensor::MSTensor *> &in_tensors,
-                const std::vector<tensor::MSTensor *> &out_tensors) override;
+  int IsSupport(const schema::Primitive *primitive, const std::vector<mindspore::MSTensor> &in_tensors,
+                const std::vector<mindspore::MSTensor> &out_tensors) override;
 
-  int Init(const schema::Primitive *primitive, const std::vector<tensor::MSTensor *> &in_tensors,
-           const std::vector<tensor::MSTensor *> &out_tensors) override;
+  int Init(const schema::Primitive *primitive, const std::vector<mindspore::MSTensor> &in_tensors,
+           const std::vector<mindspore::MSTensor> &out_tensors) override;
 
-  int SetNPUInputs(const std::vector<tensor::MSTensor *> &in_tensors,
-                   const std::vector<tensor::MSTensor *> &out_tensors,
+  int SetNPUInputs(const std::vector<mindspore::MSTensor> &in_tensors,
+                   const std::vector<mindspore::MSTensor> &out_tensors,
                    const std::vector<ge::Operator *> &npu_inputs) override;
 
   ge::Operator *GetNPUOp() override;

mindspore/lite/src/delegate/npu/op/npu_op.h

@@ -21,19 +21,20 @@
 #include <string>
 #include <set>
 #include <unordered_map>
-#include "include/errorcode.h"
-#include "include/ms_tensor.h"
+#include "include/graph/graph.h"
 #include "schema/model_generated.h"
+#include "include/errorcode.h"
+#include "include/api/types.h"
+#include "include/api/data_type.h"
 #include "src/common/log_adapter.h"
-#include "include/graph/graph.h"
 using mindspore::lite::RET_ERROR;
 using mindspore::lite::RET_NOT_SUPPORT;
 using mindspore::lite::RET_OK;
 namespace mindspore {
 class NPUOp {
  public:
-  NPUOp(const schema::Primitive *primitive, const std::vector<tensor::MSTensor *> &in_tensors,
-        const std::vector<tensor::MSTensor *> &out_tensors, std::string name)
+  NPUOp(const schema::Primitive *primitive, const std::vector<mindspore::MSTensor> &in_tensors,
+        const std::vector<mindspore::MSTensor> &out_tensors, std::string name)
       : inputs_(std::move(in_tensors)), outputs_(std::move(out_tensors)), name_(name) {
     if (primitive != nullptr) {
       type_ = primitive->value_type();
@@ -42,24 +43,24 @@ class NPUOp {
 
   virtual ~NPUOp() = default;
 
-  virtual int IsSupport(const schema::Primitive *primitive, const std::vector<tensor::MSTensor *> &in_tensors,
-                        const std::vector<tensor::MSTensor *> &out_tensors) {
+  virtual int IsSupport(const schema::Primitive *primitive, const std::vector<mindspore::MSTensor> &in_tensors,
+                        const std::vector<mindspore::MSTensor> &out_tensors) {
     return RET_ERROR;
   }
 
-  virtual int Init(const schema::Primitive *primitive, const std::vector<tensor::MSTensor *> &in_tensors,
-                   const std::vector<tensor::MSTensor *> &out_tensors) {
+  virtual int Init(const schema::Primitive *primitive, const std::vector<mindspore::MSTensor> &in_tensors,
+                   const std::vector<mindspore::MSTensor> &out_tensors) {
     return RET_ERROR;
   }
 
-  virtual int SetNPUInputs(const std::vector<tensor::MSTensor *> &in_tensors,
-                           const std::vector<tensor::MSTensor *> &out_tensors,
+  virtual int SetNPUInputs(const std::vector<mindspore::MSTensor> &in_tensors,
+                           const std::vector<mindspore::MSTensor> &out_tensors,
                            const std::vector<ge::Operator *> &npu_inputs) {
     return RET_ERROR;
   }
 
-  virtual int SetNPUInputs(const std::vector<tensor::MSTensor *> &in_tensors,
-                           const std::vector<tensor::MSTensor *> &out_tensors,
+  virtual int SetNPUInputs(const std::vector<mindspore::MSTensor> &in_tensors,
+                           const std::vector<mindspore::MSTensor> &out_tensors,
                            const std::vector<ge::Operator *> &npu_inputs,
                            const std::unordered_map<int, std::pair<ge::Operator *, int>> &index2_multi_out_index) {
     if (index2_multi_out_index.empty()) {
@@ -70,18 +71,18 @@ class NPUOp {
 
   virtual ge::Operator *GetNPUOp() { return nullptr; }
 
-  void set_inputs(const std::vector<mindspore::tensor::MSTensor *> &in_tensors) { this->inputs_ = in_tensors; }
+  void set_inputs(const std::vector<mindspore::MSTensor> &in_tensors) { this->inputs_ = in_tensors; }
 
-  void set_input(mindspore::tensor::MSTensor *in_tensor, int index) {
+  void set_input(mindspore::MSTensor in_tensor, int index) {
     MS_ASSERT(index < inputs_.size());
     this->inputs_[index] = in_tensor;
   }
 
-  void set_outputs(const std::vector<mindspore::tensor::MSTensor *> &out_tensors) { this->outputs_ = out_tensors; }
+  void set_outputs(const std::vector<mindspore::MSTensor> &out_tensors) { this->outputs_ = out_tensors; }
 
-  const std::vector<mindspore::tensor::MSTensor *> &inputs() { return this->inputs_; }
+  const std::vector<mindspore::MSTensor> &inputs() { return this->inputs_; }
 
-  const std::vector<mindspore::tensor::MSTensor *> &outputs() { return this->outputs_; }
+  const std::vector<mindspore::MSTensor> &outputs() { return this->outputs_; }
 
   void set_in_ops(const std::vector<NPUOp *> &in_ops) { this->in_ops_ = in_ops; }
 
@@ -98,37 +99,37 @@ class NPUOp {
   void set_name(const std::string &name) { this->name_ = name; }
 
  protected:
-  std::vector<mindspore::tensor::MSTensor *> inputs_;
-  std::vector<mindspore::tensor::MSTensor *> outputs_;
+  std::vector<mindspore::MSTensor> inputs_;
+  std::vector<mindspore::MSTensor> outputs_;
   std::vector<NPUOp *> in_ops_;
   std::vector<NPUOp *> out_ops_;
   schema::PrimitiveType type_ = schema::PrimitiveType_NONE;
   std::string name_;
 };
 
-typedef NPUOp *(*NPUGetOp)(const schema::Primitive *primitive, const std::vector<tensor::MSTensor *> &in_tensors,
-                           const std::vector<tensor::MSTensor *> &out_tensors, std::string name);
+typedef NPUOp *(*NPUGetOp)(const schema::Primitive *primitive, const std::vector<mindspore::MSTensor> &in_tensors,
+                           const std::vector<mindspore::MSTensor> &out_tensors, std::string name);
 
 template <class T>
-NPUOp *GetNPUOp(const schema::Primitive *primitive, const std::vector<tensor::MSTensor *> &in_tensors,
-                const std::vector<tensor::MSTensor *> &out_tensors, std::string name) {
-  auto shape = out_tensors.front()->shape();
+NPUOp *GetNPUOp(const schema::Primitive *primitive, const std::vector<mindspore::MSTensor> &in_tensors,
+                const std::vector<mindspore::MSTensor> &out_tensors, std::string name) {
+  auto shape = out_tensors.front().Shape();
   if (std::find(shape.begin(), shape.end(), -1) != shape.end()) {
     MS_LOG(ERROR) << "NPU does not support runtime inference shape.";
     return nullptr;
   }
 
-  if (in_tensors[0]->shape().size() > 4) {
+  if (in_tensors[0].Shape().size() > 4) {
     MS_LOG(ERROR) << "Npu does not support input tensor dims greater than 4";
     return nullptr;
   }
 
   std::set<schema::PrimitiveType> int32_lists = {schema::PrimitiveType_Cast, schema::PrimitiveType_StridedSlice};
-  auto support_int32 = in_tensors[0]->data_type() == kNumberTypeInt32 &&
+  auto support_int32 = in_tensors[0].DataType() == DataType::kNumberTypeInt32 &&
                        find(int32_lists.begin(), int32_lists.end(), primitive->value_type()) != int32_lists.end();
-  if (in_tensors[0]->data_type() != kNumberTypeFloat32 && in_tensors[0]->data_type() != kNumberTypeFloat16 &&
-      !support_int32) {
-    MS_LOG(ERROR) << "Npu does not support datatype " << in_tensors[0]->data_type() << " for op type "
+  if (in_tensors[0].DataType() != DataType::kNumberTypeFloat32 &&
+      in_tensors[0].DataType() != DataType::kNumberTypeFloat16 && !support_int32) {
+    MS_LOG(ERROR) << "Npu does not support datatype " << static_cast<int>(in_tensors[0].DataType()) << " for op type "
                   << primitive->value_type();
     return nullptr;
   }

mindspore/lite/src/delegate/npu/op/pad_npu.cc

@@ -19,8 +19,8 @@
 #include "src/delegate/npu/npu_converter_utils.h"
 
 namespace mindspore {
-int PadNPUOp::IsSupport(const schema::Primitive *primitive, const std::vector<tensor::MSTensor *> &in_tensors,
-                        const std::vector<tensor::MSTensor *> &out_tensors) {
+int PadNPUOp::IsSupport(const schema::Primitive *primitive, const std::vector<mindspore::MSTensor> &in_tensors,
+                        const std::vector<mindspore::MSTensor> &out_tensors) {
   auto pad_prim = primitive->value_as_PadFusion();
   if (pad_prim == nullptr) {
     MS_LOG(ERROR) << "Get null primitive value for op ." << name_;
@@ -33,15 +33,15 @@ int PadNPUOp::IsSupport(const schema::Primitive *primitive, const std::vector<te
   if (pad_prim->paddings() != nullptr) {
     return RET_OK;
   }
-  if (in_tensors.size() >= 2 && in_tensors[1]->data() != nullptr) {
+  if (in_tensors.size() >= 2 && in_tensors[1].Data() != nullptr) {
     return RET_OK;
   }
   MS_LOG(WARNING) << "NPU pad only support constant pad size.";
   return RET_ERROR;
 }
 
-int PadNPUOp::Init(const schema::Primitive *primitive, const std::vector<tensor::MSTensor *> &in_tensors,
-                   const std::vector<tensor::MSTensor *> &out_tensors) {
+int PadNPUOp::Init(const schema::Primitive *primitive, const std::vector<mindspore::MSTensor> &in_tensors,
+                   const std::vector<mindspore::MSTensor> &out_tensors) {
   pad_ = new (std::nothrow) hiai::op::PadV2(name_);
   if (pad_ == nullptr) {
     MS_LOG(ERROR) << name_ << " op is nullptr";
@@ -67,9 +67,9 @@ int PadNPUOp::Init(const schema::Primitive *primitive, const std::vector<tensor:
       auto paddings = std::vector<int64_t>(paddings_data->begin(), paddings_data->end());
       paddings_vec_.insert(paddings_vec_.end(), paddings.begin(), paddings.end());
     }
-  } else if (in_tensors.size() >= 2 && in_tensors[1]->data() != nullptr) {
-    for (int i = 0; i < in_tensors[1]->ElementsNum(); i++) {
-      paddings_vec_.push_back(static_cast<int *>(in_tensors[1]->data())[i]);
+  } else if (in_tensors.size() >= 2 && in_tensors[1].Data() != nullptr) {
+    for (int i = 0; i < in_tensors[1].ElementNum(); i++) {
+      paddings_vec_.push_back(static_cast<const int *>(in_tensors[1].Data().get())[i]);
     }
   } else {
     MS_LOG(ERROR) << "NPU pad only support constant pad size.";
@@ -86,8 +86,8 @@ int PadNPUOp::Init(const schema::Primitive *primitive, const std::vector<tensor:
   return RET_OK;
 }
 
-int PadNPUOp::SetNPUInputs(const std::vector<tensor::MSTensor *> &in_tensors,
-                           const std::vector<tensor::MSTensor *> &out_tensors,
+int PadNPUOp::SetNPUInputs(const std::vector<mindspore::MSTensor> &in_tensors,
+                           const std::vector<mindspore::MSTensor> &out_tensors,
                            const std::vector<ge::Operator *> &npu_inputs) {
   int size = static_cast<int>(paddings_vec_.size() / 2);
   ge::TensorDesc padding_tensor_desc(ge::Shape({size, 2}), ge::FORMAT_NCHW, ge::DT_INT32);

mindspore/lite/src/delegate/npu/op/pad_npu.h

@@ -24,20 +24,20 @@
 namespace mindspore {
 class PadNPUOp : public NPUOp {
  public:
-  PadNPUOp(const schema::Primitive *primitive, const std::vector<tensor::MSTensor *> &in_tensors,
-           const std::vector<tensor::MSTensor *> &out_tensors, std::string name)
+  PadNPUOp(const schema::Primitive *primitive, const std::vector<mindspore::MSTensor> &in_tensors,
+           const std::vector<mindspore::MSTensor> &out_tensors, std::string name)
       : NPUOp(primitive, in_tensors, out_tensors, name) {}
 
   ~PadNPUOp() override;
 
-  int IsSupport(const schema::Primitive *primitive, const std::vector<tensor::MSTensor *> &in_tensors,
-                const std::vector<tensor::MSTensor *> &out_tensors) override;
+  int IsSupport(const schema::Primitive *primitive, const std::vector<mindspore::MSTensor> &in_tensors,
+                const std::vector<mindspore::MSTensor> &out_tensors) override;
 
-  int Init(const schema::Primitive *primitive, const std::vector<tensor::MSTensor *> &in_tensors,
-           const std::vector<tensor::MSTensor *> &out_tensors) override;
+  int Init(const schema::Primitive *primitive, const std::vector<mindspore::MSTensor> &in_tensors,
+           const std::vector<mindspore::MSTensor> &out_tensors) override;
 
-  int SetNPUInputs(const std::vector<tensor::MSTensor *> &in_tensors,
-                   const std::vector<tensor::MSTensor *> &out_tensors,
+  int SetNPUInputs(const std::vector<mindspore::MSTensor> &in_tensors,
+                   const std::vector<mindspore::MSTensor> &out_tensors,
                    const std::vector<ge::Operator *> &npu_inputs) override;
 
   ge::Operator *GetNPUOp() override;

mindspore/lite/src/delegate/npu/op/reduce_npu.cc

@@ -15,11 +15,10 @@
  */
 
 #include "src/delegate/npu/op/reduce_npu.h"
-#include <memory>
 
 namespace mindspore {
-int ReduceNPUOp::IsSupport(const schema::Primitive *primitive, const std::vector<tensor::MSTensor *> &in_tensors,
-                           const std::vector<tensor::MSTensor *> &out_tensors) {
+int ReduceNPUOp::IsSupport(const schema::Primitive *primitive, const std::vector<mindspore::MSTensor> &in_tensors,
+                           const std::vector<mindspore::MSTensor> &out_tensors) {
   auto reduce_prim = primitive->value_as_ReduceFusion();
   if (reduce_prim == nullptr) {
     MS_LOG(ERROR) << "Get null primitive value for op ." << name_;
@@ -37,8 +36,8 @@ int ReduceNPUOp::IsSupport(const schema::Primitive *primitive, const std::vector
   return RET_OK;
 }
 
-int ReduceNPUOp::Init(const schema::Primitive *primitive, const std::vector<tensor::MSTensor *> &in_tensors,
-                      const std::vector<tensor::MSTensor *> &out_tensors) {
+int ReduceNPUOp::Init(const schema::Primitive *primitive, const std::vector<mindspore::MSTensor> &in_tensors,
+                      const std::vector<mindspore::MSTensor> &out_tensors) {
   auto reduce_prim = primitive->value_as_ReduceFusion();
   if (reduce_prim == nullptr) {
     MS_LOG(ERROR) << "Get null primitive value for op ." << name_;
@@ -59,8 +58,8 @@ int ReduceNPUOp::Init(const schema::Primitive *primitive, const std::vector<tens
   return RET_OK;
 }
 
-int ReduceNPUOp::SetNPUInputs(const std::vector<tensor::MSTensor *> &in_tensors,
-                              const std::vector<tensor::MSTensor *> &out_tensors,
+int ReduceNPUOp::SetNPUInputs(const std::vector<mindspore::MSTensor> &in_tensors,
+                              const std::vector<mindspore::MSTensor> &out_tensors,
                               const std::vector<ge::Operator *> &npu_inputs) {
   if (reduce_mode_ == schema::ReduceMode_ReduceMean) {
     auto reduce_mean = reinterpret_cast<hiai::op::ReduceMean *>(reduce_);

mindspore/lite/src/delegate/npu/op/reduce_npu.h

@@ -24,20 +24,20 @@
 namespace mindspore {
 class ReduceNPUOp : public NPUOp {
  public:
-  ReduceNPUOp(const schema::Primitive *primitive, const std::vector<tensor::MSTensor *> &in_tensors,
-              const std::vector<tensor::MSTensor *> &out_tensors, std::string name)
+  ReduceNPUOp(const schema::Primitive *primitive, const std::vector<mindspore::MSTensor> &in_tensors,
+              const std::vector<mindspore::MSTensor> &out_tensors, std::string name)
       : NPUOp(primitive, in_tensors, out_tensors, name) {}
 
   ~ReduceNPUOp() override;
 
-  int IsSupport(const schema::Primitive *primitive, const std::vector<tensor::MSTensor *> &in_tensors,
-                const std::vector<tensor::MSTensor *> &out_tensors) override;
+  int IsSupport(const schema::Primitive *primitive, const std::vector<mindspore::MSTensor> &in_tensors,
+                const std::vector<mindspore::MSTensor> &out_tensors) override;
 
-  int Init(const schema::Primitive *primitive, const std::vector<tensor::MSTensor *> &in_tensors,
-           const std::vector<tensor::MSTensor *> &out_tensors) override;
+  int Init(const schema::Primitive *primitive, const std::vector<mindspore::MSTensor> &in_tensors,
+           const std::vector<mindspore::MSTensor> &out_tensors) override;
 
-  int SetNPUInputs(const std::vector<tensor::MSTensor *> &in_tensors,
-                   const std::vector<tensor::MSTensor *> &out_tensors,
+  int SetNPUInputs(const std::vector<mindspore::MSTensor> &in_tensors,
+                   const std::vector<mindspore::MSTensor> &out_tensors,
                    const std::vector<ge::Operator *> &npu_inputs) override;
 
   ge::Operator *GetNPUOp() override;

mindspore/lite/src/delegate/npu/op/reshape_npu.cc

@@ -15,26 +15,25 @@
  */
 
 #include "src/delegate/npu/op/reshape_npu.h"
-#include <memory>
 #include "include/graph/op/all_ops.h"
 #include "src/delegate/npu/npu_converter_utils.h"
 namespace mindspore {
-int ReshapeNPUOp::IsSupport(const schema::Primitive *primitive, const std::vector<tensor::MSTensor *> &in_tensors,
-                            const std::vector<tensor::MSTensor *> &out_tensors) {
+int ReshapeNPUOp::IsSupport(const schema::Primitive *primitive, const std::vector<mindspore::MSTensor> &in_tensors,
+                            const std::vector<mindspore::MSTensor> &out_tensors) {
   if (in_tensors.size() != 2) {
     MS_LOG(WARNING) << "Npu op should have w2 input tensors.";
     return RET_NOT_SUPPORT;
   }
   auto shape_tensor = in_tensors.at(1);
-  if (shape_tensor->data() == nullptr) {
+  if (shape_tensor.Data() == nullptr) {
     MS_LOG(WARNING) << "Npu reshape op only supports const shape.";
     return RET_NOT_SUPPORT;
   }
   return RET_OK;
 }
 
-int ReshapeNPUOp::Init(const schema::Primitive *primitive, const std::vector<tensor::MSTensor *> &in_tensors,
-                       const std::vector<tensor::MSTensor *> &out_tensors) {
+int ReshapeNPUOp::Init(const schema::Primitive *primitive, const std::vector<mindspore::MSTensor> &in_tensors,
+                       const std::vector<mindspore::MSTensor> &out_tensors) {
   reshape_ = new (std::nothrow) hiai::op::Reshape(name_);
   if (reshape_ == nullptr) {
     MS_LOG(ERROR) << name_ << " op is nullptr";
@@ -43,8 +42,8 @@ int ReshapeNPUOp::Init(const schema::Primitive *primitive, const std::vector<ten
   return RET_OK;
 }
 
-int ReshapeNPUOp::SetNPUInputs(const std::vector<tensor::MSTensor *> &in_tensors,
-                               const std::vector<tensor::MSTensor *> &out_tensors,
+int ReshapeNPUOp::SetNPUInputs(const std::vector<mindspore::MSTensor> &in_tensors,
+                               const std::vector<mindspore::MSTensor> &out_tensors,
                                const std::vector<ge::Operator *> &npu_inputs) {
   reshape_->set_input_x(*npu_inputs[0]);
   reshape_->set_input_shape(*npu_inputs[1]);

mindspore/lite/src/delegate/npu/op/reshape_npu.h

@@ -23,20 +23,20 @@
 namespace mindspore {
 class ReshapeNPUOp : public NPUOp {
  public:
-  ReshapeNPUOp(const schema::Primitive *primitive, const std::vector<tensor::MSTensor *> &in_tensors,
-               const std::vector<tensor::MSTensor *> &out_tensors, std::string name)
+  ReshapeNPUOp(const schema::Primitive *primitive, const std::vector<mindspore::MSTensor> &in_tensors,
+               const std::vector<mindspore::MSTensor> &out_tensors, std::string name)
       : NPUOp(primitive, in_tensors, out_tensors, name) {}
 
   ~ReshapeNPUOp() override;
 
-  int Init(const schema::Primitive *primitive, const std::vector<tensor::MSTensor *> &in_tensors,
-           const std::vector<tensor::MSTensor *> &out_tensors) override;
+  int Init(const schema::Primitive *primitive, const std::vector<mindspore::MSTensor> &in_tensors,
+           const std::vector<mindspore::MSTensor> &out_tensors) override;
 
-  int IsSupport(const schema::Primitive *primitive, const std::vector<tensor::MSTensor *> &in_tensors,
-                const std::vector<tensor::MSTensor *> &out_tensors) override;
+  int IsSupport(const schema::Primitive *primitive, const std::vector<mindspore::MSTensor> &in_tensors,
+                const std::vector<mindspore::MSTensor> &out_tensors) override;
 
-  int SetNPUInputs(const std::vector<tensor::MSTensor *> &in_tensors,
-                   const std::vector<tensor::MSTensor *> &out_tensors,
+  int SetNPUInputs(const std::vector<mindspore::MSTensor> &in_tensors,
+                   const std::vector<mindspore::MSTensor> &out_tensors,
                    const std::vector<ge::Operator *> &npu_inputs) override;
 
   ge::Operator *GetNPUOp() override;

mindspore/lite/src/delegate/npu/op/resize_npu.cc

@@ -19,8 +19,8 @@
 #include "src/delegate/npu/npu_converter_utils.h"
 
 namespace mindspore {
-int ResizeNPUOp::IsSupport(const schema::Primitive *primitive, const std::vector<tensor::MSTensor *> &in_tensors,
-                           const std::vector<tensor::MSTensor *> &out_tensors) {
+int ResizeNPUOp::IsSupport(const schema::Primitive *primitive, const std::vector<mindspore::MSTensor> &in_tensors,
+                           const std::vector<mindspore::MSTensor> &out_tensors) {
   auto resize_prim = primitive->value_as_Resize();
   if (resize_prim == nullptr) {
     MS_LOG(ERROR) << "Get null primitive value for op ." << name_;
@@ -32,16 +32,15 @@ int ResizeNPUOp::IsSupport(const schema::Primitive *primitive, const std::vector
     return RET_NOT_SUPPORT;
   }
 
-  if (in_tensors[0]->shape()[1] > out_tensors[0]->shape()[1] ||
-      in_tensors[0]->shape()[2] > out_tensors[0]->shape()[2]) {
+  if (in_tensors[0].Shape()[1] > out_tensors[0].Shape()[1] || in_tensors[0].Shape()[2] > out_tensors[0].Shape()[2]) {
     MS_LOG(WARNING) << "Npu resize does not support reduction.";
     return RET_NOT_SUPPORT;
   }
   return RET_OK;
 }
 
-int ResizeNPUOp::Init(const schema::Primitive *primitive, const std::vector<tensor::MSTensor *> &in_tensors,
-                      const std::vector<tensor::MSTensor *> &out_tensors) {
+int ResizeNPUOp::Init(const schema::Primitive *primitive, const std::vector<mindspore::MSTensor> &in_tensors,
+                      const std::vector<mindspore::MSTensor> &out_tensors) {
   auto resize_prim = primitive->value_as_Resize();
   if (resize_prim == nullptr) {
     MS_LOG(ERROR) << "Get null primitive value for op ." << name_;
@@ -51,13 +50,13 @@ int ResizeNPUOp::Init(const schema::Primitive *primitive, const std::vector<tens
     new_height_ = resize_prim->new_height();
     new_width_ = resize_prim->new_width();
   } else if (in_tensors.size() == 2) {
-    auto out_size = in_tensors.at(1)->data();
+    auto out_size = in_tensors.at(1).Data();
     if (out_size == nullptr) {
       MS_LOG(ERROR) << "Out size is not assigned";
       return RET_ERROR;
     }
-    new_height_ = out_tensors.at(0)->shape().at(1);
-    new_width_ = out_tensors.at(0)->shape().at(2);
+    new_height_ = out_tensors.at(0).Shape().at(1);
+    new_width_ = out_tensors.at(0).Shape().at(2);
   } else {
     MS_LOG(ERROR) << "Get resize op new_height and new_width error.";
     return RET_ERROR;
@@ -97,8 +96,8 @@ int ResizeNPUOp::Init(const schema::Primitive *primitive, const std::vector<tens
   return RET_OK;
 }
 
-int ResizeNPUOp::SetNPUInputs(const std::vector<tensor::MSTensor *> &in_tensors,
-                              const std::vector<tensor::MSTensor *> &out_tensors,
+int ResizeNPUOp::SetNPUInputs(const std::vector<mindspore::MSTensor> &in_tensors,
+                              const std::vector<mindspore::MSTensor> &out_tensors,
                               const std::vector<ge::Operator *> &npu_inputs) {
   if (resize_method_ == schema::ResizeMethod_LINEAR) {
     auto resize_bilinear = reinterpret_cast<hiai::op::ResizeBilinearV2 *>(resize_);

mindspore/lite/src/delegate/npu/op/resize_npu.h

@@ -24,20 +24,20 @@
 namespace mindspore {
 class ResizeNPUOp : public NPUOp {
  public:
-  ResizeNPUOp(const schema::Primitive *primitive, const std::vector<tensor::MSTensor *> &in_tensors,
-              const std::vector<tensor::MSTensor *> &out_tensors, std::string name)
+  ResizeNPUOp(const schema::Primitive *primitive, const std::vector<mindspore::MSTensor> &in_tensors,
+              const std::vector<mindspore::MSTensor> &out_tensors, std::string name)
       : NPUOp(primitive, in_tensors, out_tensors, name) {}
 
   ~ResizeNPUOp() override;
 
-  int IsSupport(const schema::Primitive *primitive, const std::vector<tensor::MSTensor *> &in_tensors,
-                const std::vector<tensor::MSTensor *> &out_tensors) override;
+  int IsSupport(const schema::Primitive *primitive, const std::vector<mindspore::MSTensor> &in_tensors,
+                const std::vector<mindspore::MSTensor> &out_tensors) override;
 
-  int Init(const schema::Primitive *primitive, const std::vector<tensor::MSTensor *> &in_tensors,
-           const std::vector<tensor::MSTensor *> &out_tensors) override;
+  int Init(const schema::Primitive *primitive, const std::vector<mindspore::MSTensor> &in_tensors,
+           const std::vector<mindspore::MSTensor> &out_tensors) override;
 
-  int SetNPUInputs(const std::vector<tensor::MSTensor *> &in_tensors,
-                   const std::vector<tensor::MSTensor *> &out_tensors,
+  int SetNPUInputs(const std::vector<mindspore::MSTensor> &in_tensors,
+                   const std::vector<mindspore::MSTensor> &out_tensors,
                    const std::vector<ge::Operator *> &npu_inputs) override;
 
   ge::Operator *GetNPUOp() override;

mindspore/lite/src/delegate/npu/op/scale_npu.cc

@@ -15,12 +15,11 @@
  */
 
 #include "src/delegate/npu/op/scale_npu.h"
-#include <memory>
 #include "src/delegate/npu/npu_converter_utils.h"
 
 namespace mindspore {
-int ScaleNPUOp::IsSupport(const schema::Primitive *primitive, const std::vector<tensor::MSTensor *> &in_tensors,
-                          const std::vector<tensor::MSTensor *> &out_tensors) {
+int ScaleNPUOp::IsSupport(const schema::Primitive *primitive, const std::vector<mindspore::MSTensor> &in_tensors,
+                          const std::vector<mindspore::MSTensor> &out_tensors) {
   auto scale_prim = primitive->value_as_ScaleFusion();
   if (scale_prim == nullptr) {
     MS_LOG(ERROR) << "Get null primitive value for op ." << name_;
@@ -28,7 +27,7 @@ int ScaleNPUOp::IsSupport(const schema::Primitive *primitive, const std::vector<
   }
   axis_ = scale_prim->axis();
   if (axis_ < 0) {
-    axis_ = axis_ + in_tensors[0]->shape().size();
+    axis_ = axis_ + in_tensors[0].Shape().size();
   }
   if (axis_ != 1 && axis_ != 3) {
     MS_LOG(WARNING) << "Npu scale axis attr only support 1 or channel, now is " << axis_;
@@ -37,8 +36,8 @@ int ScaleNPUOp::IsSupport(const schema::Primitive *primitive, const std::vector<
   return RET_OK;
 }
 
-int ScaleNPUOp::Init(const schema::Primitive *primitive, const std::vector<tensor::MSTensor *> &in_tensors,
-                     const std::vector<tensor::MSTensor *> &out_tensors) {
+int ScaleNPUOp::Init(const schema::Primitive *primitive, const std::vector<mindspore::MSTensor> &in_tensors,
+                     const std::vector<mindspore::MSTensor> &out_tensors) {
   op_ = new (std::nothrow) hiai::op::Scale(name_);
   if (op_ == nullptr) {
     MS_LOG(ERROR) << name_ << " op is nullptr";
@@ -62,12 +61,12 @@ int ScaleNPUOp::Init(const schema::Primitive *primitive, const std::vector<tenso
   return RET_OK;
 }
 
-int ScaleNPUOp::SetNPUInputs(const std::vector<tensor::MSTensor *> &in_tensors,
-                             const std::vector<tensor::MSTensor *> &out_tensors,
+int ScaleNPUOp::SetNPUInputs(const std::vector<mindspore::MSTensor> &in_tensors,
+                             const std::vector<mindspore::MSTensor> &out_tensors,
                              const std::vector<ge::Operator *> &npu_inputs) {
   op_->set_input_x(*npu_inputs.at(0));
   MS_ASSERT(in_tensors.size() > 1);
-  auto scale_shape = in_tensors[1]->shape();
+  auto scale_shape = in_tensors[1].Shape();
   auto scale_tensor = ConverterToNPUTensor(in_tensors[1]);
   if (scale_tensor == nullptr) {
     MS_LOG(ERROR) << "Get scale_tensor failed.";
@@ -84,7 +83,7 @@ int ScaleNPUOp::SetNPUInputs(const std::vector<tensor::MSTensor *> &in_tensors,
   op_->set_input_scale(*scale_);
 
   if (in_tensors.size() > 2 && in_tensors[2] != nullptr) {
-    auto bias_shape = in_tensors[2]->shape();
+    auto bias_shape = in_tensors[2].Shape();
     auto bias_tensor = ConverterToNPUTensor(in_tensors[2]);
     if (bias_tensor == nullptr) {
       MS_LOG(ERROR) << "Get bias_tensor failed.";

mindspore/lite/src/delegate/npu/op/scale_npu.h

@@ -25,20 +25,20 @@
 namespace mindspore {
 class ScaleNPUOp : public NPUOp {
  public:
-  ScaleNPUOp(const schema::Primitive *primitive, const std::vector<tensor::MSTensor *> &in_tensors,
-             const std::vector<tensor::MSTensor *> &out_tensors, std::string name)
+  ScaleNPUOp(const schema::Primitive *primitive, const std::vector<mindspore::MSTensor> &in_tensors,
+             const std::vector<mindspore::MSTensor> &out_tensors, std::string name)
       : NPUOp(primitive, in_tensors, out_tensors, name) {}
 
   ~ScaleNPUOp() override;
 
-  int IsSupport(const schema::Primitive *primitive, const std::vector<tensor::MSTensor *> &in_tensors,
-                const std::vector<tensor::MSTensor *> &out_tensors) override;
+  int IsSupport(const schema::Primitive *primitive, const std::vector<mindspore::MSTensor> &in_tensors,
+                const std::vector<mindspore::MSTensor> &out_tensors) override;
 
-  int Init(const schema::Primitive *primitive, const std::vector<tensor::MSTensor *> &in_tensors,
-           const std::vector<tensor::MSTensor *> &out_tensors) override;
+  int Init(const schema::Primitive *primitive, const std::vector<mindspore::MSTensor> &in_tensors,
+           const std::vector<mindspore::MSTensor> &out_tensors) override;
 
-  int SetNPUInputs(const std::vector<tensor::MSTensor *> &in_tensors,
-                   const std::vector<tensor::MSTensor *> &out_tensors,
+  int SetNPUInputs(const std::vector<mindspore::MSTensor> &in_tensors,
+                   const std::vector<mindspore::MSTensor> &out_tensors,
                    const std::vector<ge::Operator *> &npu_inputs) override;
 
   ge::Operator *GetNPUOp() override;

mindspore/lite/src/delegate/npu/op/slice_npu.cc

@@ -18,8 +18,8 @@
 #include "src/delegate/npu/npu_converter_utils.h"
 
 namespace mindspore {
-int SliceNPUOp::Init(const schema::Primitive *primitive, const std::vector<tensor::MSTensor *> &in_tensors,
-                     const std::vector<tensor::MSTensor *> &out_tensors) {
+int SliceNPUOp::Init(const schema::Primitive *primitive, const std::vector<mindspore::MSTensor> &in_tensors,
+                     const std::vector<mindspore::MSTensor> &out_tensors) {
   slice_ = new (std::nothrow) hiai::op::Slice(name_);
   if (slice_ == nullptr) {
     MS_LOG(ERROR) << name_ << " op is nullptr";
@@ -28,8 +28,8 @@ int SliceNPUOp::Init(const schema::Primitive *primitive, const std::vector<tenso
   return RET_OK;
 }
 
-int SliceNPUOp::SetNPUInputs(const std::vector<tensor::MSTensor *> &in_tensors,
-                             const std::vector<tensor::MSTensor *> &out_tensors,
+int SliceNPUOp::SetNPUInputs(const std::vector<mindspore::MSTensor> &in_tensors,
+                             const std::vector<mindspore::MSTensor> &out_tensors,
                              const std::vector<ge::Operator *> &npu_inputs) {
   slice_->set_input_x(*npu_inputs[0]);
   slice_->set_input_offsets(*npu_inputs[1]);

mindspore/lite/src/delegate/npu/op/slice_npu.h

@@ -24,22 +24,22 @@
 namespace mindspore {
 class SliceNPUOp : public NPUOp {
  public:
-  SliceNPUOp(const schema::Primitive *primitive, const std::vector<tensor::MSTensor *> &in_tensors,
-             const std::vector<tensor::MSTensor *> &out_tensors, std::string name)
+  SliceNPUOp(const schema::Primitive *primitive, const std::vector<mindspore::MSTensor> &in_tensors,
+             const std::vector<mindspore::MSTensor> &out_tensors, std::string name)
       : NPUOp(primitive, in_tensors, out_tensors, name) {}
 
   ~SliceNPUOp() override;
 
-  int IsSupport(const schema::Primitive *primitive, const std::vector<tensor::MSTensor *> &in_tensors,
-                const std::vector<tensor::MSTensor *> &out_tensors) override {
+  int IsSupport(const schema::Primitive *primitive, const std::vector<mindspore::MSTensor> &in_tensors,
+                const std::vector<mindspore::MSTensor> &out_tensors) override {
     return RET_OK;
   }
 
-  int Init(const schema::Primitive *primitive, const std::vector<tensor::MSTensor *> &in_tensors,
-           const std::vector<tensor::MSTensor *> &out_tensors) override;
+  int Init(const schema::Primitive *primitive, const std::vector<mindspore::MSTensor> &in_tensors,
+           const std::vector<mindspore::MSTensor> &out_tensors) override;
 
-  int SetNPUInputs(const std::vector<tensor::MSTensor *> &in_tensors,
-                   const std::vector<tensor::MSTensor *> &out_tensors,
+  int SetNPUInputs(const std::vector<mindspore::MSTensor> &in_tensors,
+                   const std::vector<mindspore::MSTensor> &out_tensors,
                    const std::vector<ge::Operator *> &npu_inputs) override;
 
   ge::Operator *GetNPUOp() override;

mindspore/lite/src/delegate/npu/op/softmax_npu.cc

@@ -16,8 +16,8 @@
 
 #include "src/delegate/npu/op/softmax_npu.h"
 namespace mindspore {
-int SoftmaxNPUOp::Init(const schema::Primitive *primitive, const std::vector<tensor::MSTensor *> &in_tensors,
-                       const std::vector<tensor::MSTensor *> &out_tensors) {
+int SoftmaxNPUOp::Init(const schema::Primitive *primitive, const std::vector<mindspore::MSTensor> &in_tensors,
+                       const std::vector<mindspore::MSTensor> &out_tensors) {
   softmax_ = new (std::nothrow) hiai::op::Softmax(name_);
   if (softmax_ == nullptr) {
     MS_LOG(ERROR) << name_ << " op is nullptr";
@@ -30,15 +30,15 @@ int SoftmaxNPUOp::Init(const schema::Primitive *primitive, const std::vector<ten
   }
   auto axis = static_cast<int>(*(softmax_prim->axis()->begin()));
   if (axis == -1) {
-    softmax_->set_attr_axis(in_tensors[0]->shape().size() + axis);
+    softmax_->set_attr_axis(in_tensors[0].Shape().size() + axis);
   } else {
     softmax_->set_attr_axis(axis);
   }
   return RET_OK;
 }
 
-int SoftmaxNPUOp::SetNPUInputs(const std::vector<tensor::MSTensor *> &in_tensors,
-                               const std::vector<tensor::MSTensor *> &out_tensors,
+int SoftmaxNPUOp::SetNPUInputs(const std::vector<mindspore::MSTensor> &in_tensors,
+                               const std::vector<mindspore::MSTensor> &out_tensors,
                                const std::vector<ge::Operator *> &npu_inputs) {
   softmax_->set_input_x(*npu_inputs[0]);
   return RET_OK;

mindspore/lite/src/delegate/npu/op/softmax_npu.h

@@ -24,22 +24,22 @@
 namespace mindspore {
 class SoftmaxNPUOp : public NPUOp {
  public:
-  SoftmaxNPUOp(const schema::Primitive *primitive, const std::vector<tensor::MSTensor *> &in_tensors,
-               const std::vector<tensor::MSTensor *> &out_tensors, std::string name)
+  SoftmaxNPUOp(const schema::Primitive *primitive, const std::vector<mindspore::MSTensor> &in_tensors,
+               const std::vector<mindspore::MSTensor> &out_tensors, std::string name)
       : NPUOp(primitive, in_tensors, out_tensors, name) {}
 
   ~SoftmaxNPUOp() override;
 
-  int IsSupport(const schema::Primitive *primitive, const std::vector<tensor::MSTensor *> &in_tensors,
-                const std::vector<tensor::MSTensor *> &out_tensors) override {
+  int IsSupport(const schema::Primitive *primitive, const std::vector<mindspore::MSTensor> &in_tensors,
+                const std::vector<mindspore::MSTensor> &out_tensors) override {
     return RET_OK;
   }
 
-  int Init(const schema::Primitive *primitive, const std::vector<tensor::MSTensor *> &in_tensors,
-           const std::vector<tensor::MSTensor *> &out_tensors) override;
+  int Init(const schema::Primitive *primitive, const std::vector<mindspore::MSTensor> &in_tensors,
+           const std::vector<mindspore::MSTensor> &out_tensors) override;
 
-  int SetNPUInputs(const std::vector<tensor::MSTensor *> &in_tensors,
-                   const std::vector<tensor::MSTensor *> &out_tensors,
+  int SetNPUInputs(const std::vector<mindspore::MSTensor> &in_tensors,
+                   const std::vector<mindspore::MSTensor> &out_tensors,
                    const std::vector<ge::Operator *> &npu_inputs) override;
 
   ge::Operator *GetNPUOp() override;

mindspore/lite/src/delegate/npu/op/split_npu.cc

@@ -19,8 +19,8 @@
 #include "src/delegate/npu/npu_converter_utils.h"
 
 namespace mindspore {
-int SplitNPUOp::Init(const schema::Primitive *primitive, const std::vector<tensor::MSTensor *> &in_tensors,
-                     const std::vector<tensor::MSTensor *> &out_tensors) {
+int SplitNPUOp::Init(const schema::Primitive *primitive, const std::vector<mindspore::MSTensor> &in_tensors,
+                     const std::vector<mindspore::MSTensor> &out_tensors) {
   split_ = new (std::nothrow) hiai::op::SplitV(name_);
   if (split_ == nullptr) {
     MS_LOG(ERROR) << "New split npu operator for op " << name_ << " failed.";
@@ -53,8 +53,8 @@ int SplitNPUOp::Init(const schema::Primitive *primitive, const std::vector<tenso
   return RET_OK;
 }
 
-int SplitNPUOp::SetNPUInputs(const std::vector<tensor::MSTensor *> &in_tensors,
-                             const std::vector<tensor::MSTensor *> &out_tensors,
+int SplitNPUOp::SetNPUInputs(const std::vector<mindspore::MSTensor> &in_tensors,
+                             const std::vector<mindspore::MSTensor> &out_tensors,
                              const std::vector<ge::Operator *> &npu_inputs) {
   ge::TensorDesc split_dim_tensor_desc(ge::Shape({1}), ge::FORMAT_NCHW, ge::DT_INT32);
   ge::TensorPtr split_dim_tensor = std::make_shared<hiai::Tensor>(split_dim_tensor_desc);

mindspore/lite/src/delegate/npu/op/split_npu.h

@@ -24,22 +24,22 @@
 namespace mindspore {
 class SplitNPUOp : public NPUOp {
  public:
-  SplitNPUOp(const schema::Primitive *primitive, const std::vector<tensor::MSTensor *> &in_tensors,
-             const std::vector<tensor::MSTensor *> &out_tensors, std::string name)
+  SplitNPUOp(const schema::Primitive *primitive, const std::vector<mindspore::MSTensor> &in_tensors,
+             const std::vector<mindspore::MSTensor> &out_tensors, std::string name)
       : NPUOp(primitive, in_tensors, out_tensors, name) {}
 
   ~SplitNPUOp();
 
-  int IsSupport(const schema::Primitive *primitive, const std::vector<tensor::MSTensor *> &in_tensors,
-                const std::vector<tensor::MSTensor *> &out_tensors) override {
+  int IsSupport(const schema::Primitive *primitive, const std::vector<mindspore::MSTensor> &in_tensors,
+                const std::vector<mindspore::MSTensor> &out_tensors) override {
     return RET_OK;
   }
 
-  int Init(const schema::Primitive *primitive, const std::vector<tensor::MSTensor *> &in_tensors,
-           const std::vector<tensor::MSTensor *> &out_tensors) override;
+  int Init(const schema::Primitive *primitive, const std::vector<mindspore::MSTensor> &in_tensors,
+           const std::vector<mindspore::MSTensor> &out_tensors) override;
 
-  int SetNPUInputs(const std::vector<tensor::MSTensor *> &in_tensors,
-                   const std::vector<tensor::MSTensor *> &out_tensors,
+  int SetNPUInputs(const std::vector<mindspore::MSTensor> &in_tensors,
+                   const std::vector<mindspore::MSTensor> &out_tensors,
                    const std::vector<ge::Operator *> &npu_inputs) override;
 
   int HandleAxis();