mindspore support coreml

.jenkins/check/config/filter_cpplint.txt

@@ -63,6 +63,7 @@
 "mindspore/mindspore/lite/src/runtime/thread_pool.c"                                               "runtime/arrays"
 "mindspore/mindspore/lite/src/runtime/thread_pool.c"                                               "runtime/int"
 "mindspore/mindspore/lite/src/common/ops/ops_def.cc"                                               "runtime/int"
+"mindspore/mindspore/lite/src/runtime/delegate/coreml/coreml_executor.h"                           "readability/casting"
 "mindspore/mindspore/lite/examples/runtime_gpu_extend/src/cl"                                      "legal/copyright"
 "mindspore/mindspore/lite/examples/runtime_gpu_extend/src/cl"                                      "readability/casting"
 "mindspore/mindspore/lite/examples/runtime_gpu_extend/src/cl"                                      "readability/fn_size"

cmake/external_libs/protobuf_arm.cmake

@@ -50,6 +50,22 @@ else()
     set(PROTOBUF_PATCH_ROOT ${CMAKE_SOURCE_DIR}/third_party/patch/protobuf)
 endif()
 
+if(APPLE)
+    mindspore_add_pkg(protobuf_arm
+            VER 3.13.0
+            LIBS protobuf
+            URL ${REQ_URL}
+            MD5 ${MD5}
+            CMAKE_PATH cmake/
+            CMAKE_OPTION
+            -Dprotobuf_BUILD_TESTS=OFF
+            -Dprotobuf_BUILD_SHARED_LIBS=OFF
+            -DCMAKE_BUILD_TYPE=Release
+            -Dprotobuf_WITH_ZLIB=OFF
+            -DCMAKE_OSX_SYSROOT=${CMAKE_OSX_SYSROOT}
+            -DCMAKE_OSX_DEPLOYMENT_TARGET=${CMAKE_OSX_DEPLOYMENT_TARGET}
+            PATCHES ${PROTOBUF_PATCH_ROOT}/CVE-2021-22570.patch)
+else()
     mindspore_add_pkg(protobuf_arm
             VER 3.13.0
             LIBS protobuf
@@ -64,6 +80,7 @@ mindspore_add_pkg(protobuf_arm
             -DCMAKE_CXX_COMPILER=${CMAKE_CXX_COMPILER}
             -Dprotobuf_WITH_ZLIB=OFF
             PATCHES ${PROTOBUF_PATCH_ROOT}/CVE-2021-22570.patch)
+endif()
 
 include_directories(${protobuf_arm_INC})
 add_library(mindspore::protobuf_arm ALIAS protobuf_arm::protobuf)

mindspore/lite/CMakeLists.txt

@@ -79,6 +79,9 @@ endif()
 if(DEFINED ENV{MSLITE_ENABLE_NPU})
     set(MSLITE_ENABLE_NPU $ENV{MSLITE_ENABLE_NPU})
 endif()
+if(DEFINED ENV{MSLITE_ENABLE_COREML})
+    set(MSLITE_ENABLE_COREML $ENV{MSLITE_ENABLE_COREML})
+endif()
 if(DEFINED ENV{MSLITE_ENABLE_TRAIN})
     set(MSLITE_ENABLE_TRAIN $ENV{MSLITE_ENABLE_TRAIN})
 endif()
@@ -280,6 +283,10 @@ else()
     set(MSLITE_ENABLE_NPU off)
 endif()
 
+if(NOT APPLE)
+    set(MSLITE_ENABLE_COREML off)
+endif()
+
 if(DEFINED ENV{MSLITE_ENABLE_RUNTIME_GLOG})
     set(MSLITE_ENABLE_RUNTIME_GLOG $ENV{MSLITE_ENABLE_RUNTIME_GLOG})
 endif()
@@ -379,6 +386,7 @@ message(STATUS "************MindSpore Lite Build Option:************")
 message(STATUS "\tMSLITE_GPU_BACKEND                         = \t${MSLITE_GPU_BACKEND}")
 message(STATUS "\tMSLITE_REGISTRY_DEVICE                     = \t${MSLITE_REGISTRY_DEVICE}")
 message(STATUS "\tMSLITE_ENABLE_NPU                          = \t${MSLITE_ENABLE_NPU}")
+message(STATUS "\tMSLITE_ENABLE_COREML                       = \t${MSLITE_ENABLE_COREML}")
 message(STATUS "\tMSLITE_ENABLE_TRAIN                        = \t${MSLITE_ENABLE_TRAIN}")
 message(STATUS "\tMSLITE_MICRO_PLATFORM                      = \t${MSLITE_MICRO_PLATFORM}")
 message(STATUS "\tMSLITE_ENABLE_SSE                          = \t${MSLITE_ENABLE_SSE}")
@@ -430,10 +438,10 @@ if(MSLITE_ENABLE_EXPERIMENTAL_KERNEL)
     add_compile_definitions(MSLITE_ENABLE_EXPERIMENTAL_KERNEL)
 endif()
 
-if(((MSLITE_GPU_BACKEND STREQUAL tensorrt) OR MSLITE_ENABLE_NPU) AND (
+if(((MSLITE_GPU_BACKEND STREQUAL tensorrt) OR MSLITE_ENABLE_NPU OR MSLITE_ENABLE_COREML) AND (
         NOT MSLITE_ENABLE_DELEGATE))
-    message(FATAL_ERROR "If MSLITE_ENABLE_DELEGATE use is configured as off, MSLITE_ENABLE_NPU must also be configured
-    as off and MSLITE_GPU_BACKEND nor can it be configured as tensorrt.")
+    message(FATAL_ERROR "If MSLITE_ENABLE_DELEGATE use is configured as off, MSLITE_ENABLE_NPU and MSLITE_ENABLE_COREML
+    must also be configured as off and MSLITE_GPU_BACKEND nor can it be configured as tensorrt.")
 endif()
 
 if(MSLITE_ENABLE_HIGH_PERFORMANCE)
@@ -554,6 +562,27 @@ if(MSLITE_GPU_BACKEND STREQUAL opencl)
     set(MSLITE_DEPS_OPENCL on)
 endif()
 
+function(find_required_package pkg_name)
+    find_package(${pkg_name})
+    if(NOT ${pkg_name}_FOUND)
+        message(FATAL_ERROR "Required package ${pkg_name} not found, "
+                "please install the package and try building MindSpore again.")
+    endif()
+endfunction()
+
+if(MSLITE_ENABLE_COREML)
+    if(PLATFORM_ARM32)
+        message(FATAL_ERROR "CoreML not support arm32 platform!")
+    endif()
+    add_compile_definitions(ENABLE_COREML)
+    find_required_package(Patch)
+    include(${TOP_DIR}/cmake/external_libs/protobuf.cmake)
+    set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -arch arm64")
+    set(CMAKE_C_FLAGS "${CMAKE_C_FLAGS} -arch arm64")
+    set(CMAKE_OSX_DEPLOYMENT_TARGET "11.0")
+    include(${TOP_DIR}/cmake/external_libs/protobuf_arm.cmake)
+endif()
+
 if(MSLITE_ENABLE_CONVERTER OR MSLITE_MINDDATA_IMPLEMENT STREQUAL "full" OR MSLITE_MINDDATA_IMPLEMENT STREQUAL "wrapper"
     OR MSLITE_ENABLE_TOOLS OR MSLITE_ENABLE_KERNEL_EXECUTOR)
     # include(${TOP_DIR}/cmake/external_libs/json.cmake)
@@ -631,14 +660,6 @@ if((MSLITE_ENABLE_CONVERTER OR MSLITE_ENABLE_RUNTIME_GLOG))
     set(MSLITE_DEPS_GLOG on)
 endif()
 
-function(find_required_package pkg_name)
-    find_package(${pkg_name})
-    if(NOT ${pkg_name}_FOUND)
-        message(FATAL_ERROR "Required package ${pkg_name} not found, "
-                "please install the package and try building MindSpore again.")
-    endif()
-endfunction()
-
 if(MSLITE_ENABLE_CONVERTER OR MSLITE_ENABLE_KERNEL_EXECUTOR)
     find_required_package(Patch)
     # include(${TOP_DIR}/cmake/external_libs/protobuf.cmake)

mindspore/lite/build_lite.sh

@@ -429,6 +429,10 @@ build_lite() {
           mkdir -p ${BASEPATH}/output
           cp -r ${BASEPATH}/mindspore/lite/build/src/Release-*/mindspore-lite.framework ${BASEPATH}/output/mindspore-lite.framework
           cd ${BASEPATH}/output
+          local protobuf_arm_lib=${BASEPATH}/mindspore/lite/build/_deps/protobuf_arm-src/_build/libprotobuf-lite.a
+          if [ -e "$protobuf_arm_lib" ]; then
+            cp $protobuf_arm_lib ${BASEPATH}/output/mindspore-lite.framework/
+          fi
           tar -zcvf ${pkg_name}.tar.gz mindspore-lite.framework/
           sha256sum ${pkg_name}.tar.gz > ${pkg_name}.tar.gz.sha256
           rm -r mindspore-lite.framework

mindspore/lite/src/CMakeLists.txt

@@ -446,6 +446,10 @@ if(APPLE)
             ${MINDSPORE_LITE_PUB_HDRS_MINDAPI_HDRS}
             ${MINDSPORE_LITE_PUB_HDRS_IR_HDRS}
             )
+    if(MSLITE_ENABLE_COREML)
+        add_subdirectory(runtime/delegate/coreml)
+        target_link_libraries(mindspore-lite_static coreml_proto_mid coreml_kernel_mid)
+    endif()
     add_dependencies(mindspore-lite_static fbs_inner_src)
 else()
     add_library(mindspore-lite_static STATIC $<TARGET_OBJECTS:lite_src_mid>)

mindspore/lite/src/runtime/delegate/coreml/CMakeLists.txt

@@ -0,0 +1,15 @@
+file(GLOB PROTO_FILE "" ${TOP_DIR}/third_party/proto/coreml/*.proto)
+ms_protobuf_generate(PROTO_SRCS PROTO_HDRS ${PROTO_FILE})
+add_library(coreml_proto_mid OBJECT ${PROTO_SRCS})
+include_directories(${CMAKE_BINARY_DIR}/proto)
+
+file(GLOB_RECURSE COREML_RUNTIME_SRC
+        ${CMAKE_CURRENT_SOURCE_DIR}/*.mm
+        ${CMAKE_CURRENT_SOURCE_DIR}/*.cc
+        ${CMAKE_CURRENT_SOURCE_DIR}/op/*.cc
+        ${CMAKE_CURRENT_SOURCE_DIR}/pass/*.cc
+        ${CMAKE_CURRENT_SOURCE_DIR}/../delegate_utils.cc
+        )
+add_library(coreml_kernel_mid OBJECT ${COREML_RUNTIME_SRC})
+add_dependencies(coreml_kernel_mid fbs_src)
+target_link_libraries(coreml_kernel_mid coreml_proto_mid)

mindspore/lite/src/runtime/delegate/coreml/coreml_delegate.h

@@ -0,0 +1,55 @@
+/**
+ * Copyright 2022 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_RUNTIME_DELEGATE_COREML_COREML_DELEGATE_H_
+#define MINDSPORE_LITE_SRC_RUNTIME_DELEGATE_COREML_COREML_DELEGATE_H_
+
+#include <vector>
+#include <map>
+#include "include/api/delegate.h"
+#include "include/context.h"
+#include "src/runtime/delegate/coreml/op/coreml_op.h"
+#include "src/runtime/delegate/coreml/pass/coreml_pass_manager.h"
+
+namespace mindspore {
+class CoreMLDelegate : public Delegate {
+ public:
+  CoreMLDelegate() = default;
+
+  ~CoreMLDelegate() override;
+
+  bool IsSupportCoreML() const;
+
+  Status Init() override;
+
+  Status Build(DelegateModel<schema::Primitive> *model) override;
+
+ protected:
+  CoreMLOp *GetOP(kernel::Kernel *kernel, const schema::Primitive *primitive);
+
+  kernel::Kernel *CreateCoreMLGraph(const std::vector<CoreMLOp *> &ops, DelegateModel<schema::Primitive> *model,
+                                    KernelIter from, KernelIter end);
+
+  Status AddPasses();
+
+ protected:
+  int graph_index_ = 0;
+  CoreMLPassManager *pass_manager_ = nullptr;
+  std::map<schema::PrimitiveType, CoreMLGetOp> op_func_lists_;
+};
+}  // namespace mindspore
+
+#endif  // MINDSPORE_LITE_SRC_RUNTIME_DELEGATE_COREML_COREML_DELEGATE_H_

mindspore/lite/src/runtime/delegate/coreml/coreml_delegate.mm

@@ -0,0 +1,233 @@
+/**
+ * Copyright 2022 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/runtime/delegate/coreml/coreml_delegate.h"
+#include "include/errorcode.h"
+#include "src/common/prim_util.h"
+#include "src/runtime/delegate/coreml/op/coreml_op.h"
+#include "src/runtime/delegate/coreml/op/activation_coreml.h"
+#include "src/runtime/delegate/coreml/op/transpose_coreml.h"
+#include "src/runtime/delegate/coreml/op/convolution_coreml.h"
+#include "src/runtime/delegate/coreml/op/deconvolution_coreml.h"
+#include "src/runtime/delegate/coreml/op/avg_pooling_coreml.h"
+#include "src/runtime/delegate/coreml/op/max_pooling_coreml.h"
+#include "src/runtime/delegate/coreml/op/arithmetic_coreml.h"
+#include "src/runtime/delegate/coreml/op/resize_coreml.h"
+#include "src/runtime/delegate/coreml/op/reshape_coreml.h"
+#include "src/runtime/delegate/coreml/op/matmul_coreml.h"
+#include "src/runtime/delegate/coreml/op/concat_coreml.h"
+#include "src/runtime/delegate/coreml/op/unsqueeze_coreml.h"
+#include "src/runtime/delegate/coreml/op/gather_coreml.h"
+#include "src/runtime/delegate/coreml/op/shape_coreml.h"
+#include "src/runtime/delegate/coreml/op/softmax_coreml.h"
+#include "src/runtime/delegate/coreml/op/flatten_coreml.h"
+#include "src/runtime/delegate/coreml/coreml_graph.h"
+#include "src/runtime/delegate/delegate_utils.h"
+#include "src/runtime/delegate/coreml/pass/coreml_format_trans_pass.h"
+#include "src/runtime/delegate/coreml/pass/coreml_trans_extend_pass.h"
+#include "src/runtime/delegate/coreml/pass/coreml_fusion_pass.h"
+
+using mindspore::lite::RET_ERROR;
+using mindspore::lite::RET_OK;
+
+namespace mindspore {
+CoreMLDelegate::~CoreMLDelegate() {
+  if (pass_manager_ != nullptr) {
+    pass_manager_->Clear();
+    delete pass_manager_;
+    pass_manager_ = nullptr;
+  }
+}
+
+bool CoreMLDelegate::IsSupportCoreML() {
+  if (@available(iOS 11, *)) {
+    return true;
+  }
+  return false;
+}
+
+Status CoreMLDelegate::AddPasses() {
+  auto format_trans_pass = new (std::nothrow) CoreMLFormatTransPass();
+  if (format_trans_pass == nullptr) {
+    MS_LOG(ERROR) << "New CoreMLFormatTransPass failed.";
+    return mindspore::kLiteNullptr;
+  }
+  pass_manager_->AddPass(format_trans_pass);
+
+  auto trans_extend_pass = new (std::nothrow) CoreMLTransExtendPass();
+  if (trans_extend_pass == nullptr) {
+    MS_LOG(ERROR) << "New CoreMLTransExtendPass failed.";
+    return mindspore::kLiteNullptr;
+  }
+  pass_manager_->AddPass(trans_extend_pass);
+
+  auto fusion_pass = new (std::nothrow) CoreMLFusionPass();
+  if (fusion_pass == nullptr) {
+    MS_LOG(ERROR) << "New CoreMLFusionPass failed.";
+    return mindspore::kLiteNullptr;
+  }
+  pass_manager_->AddPass(fusion_pass);
+  return mindspore::kSuccess;
+}
+
+Status CoreMLDelegate::Init() {
+  if (!IsSupportCoreML()) {
+    return mindspore::kLiteNotSupport;
+  }
+  pass_manager_ = new (std::nothrow) CoreMLPassManager();
+  if (pass_manager_ == nullptr) {
+    MS_LOG(ERROR) << "New coreml pass manager failed.";
+    return mindspore::kLiteNullptr;
+  }
+  auto ret = AddPasses();
+  if (ret != mindspore::kSuccess) {
+    MS_LOG(ERROR) << "add passes for coreml pass manager failed.";
+    return ret;
+  }
+  op_func_lists_.clear();
+  op_func_lists_ = {
+    {schema::PrimitiveType_Activation, GetCoreMLOp<ActivationCoreMLOp>},
+    {schema::PrimitiveType_Transpose, GetCoreMLOp<TransposeCoreMLOp>},
+    {schema::PrimitiveType_Conv2DFusion, GetCoreMLOp<ConvolutionCoreMLOp>},
+    {schema::PrimitiveType_Conv2dTransposeFusion, GetCoreMLOp<DeconvolutionCoreMLOp>},
+    {schema::PrimitiveType_AvgPoolFusion, GetCoreMLOp<AvgPoolingCoreMLOp>},
+    {schema::PrimitiveType_MaxPoolFusion, GetCoreMLOp<MaxPoolingCoreMLOp>},
+    {schema::PrimitiveType_AddFusion, GetCoreMLOp<ArithmeticCoreMLOp>},
+    {schema::PrimitiveType_MulFusion, GetCoreMLOp<ArithmeticCoreMLOp>},
+    {schema::PrimitiveType_Reshape, GetCoreMLOp<ReshapeCoreMLOp>},
+    {schema::PrimitiveType_Resize, GetCoreMLOp<ResizeCoreMLOp>},
+    {schema::PrimitiveType_Concat, GetCoreMLOp<ConcatCoreMLOp>},
+    {schema::PrimitiveType_Shape, GetCoreMLOp<ShapeCoreMLOp>},
+    {schema::PrimitiveType_Gather, GetCoreMLOp<GatherCoreMLOp>},
+    {schema::PrimitiveType_Unsqueeze, GetCoreMLOp<UnsqueezeCoreMLOp>},
+    {schema::PrimitiveType_MatMulFusion, GetCoreMLOp<MatMulCoreMLOp>},
+    {schema::PrimitiveType_Softmax, GetCoreMLOp<SoftmaxCoreMLOp>},
+    {schema::PrimitiveType_Flatten, GetCoreMLOp<FlattenCoreMLOp>},
+  };
+  return mindspore::kSuccess;
+}
+
+Status CoreMLDelegate::Build(DelegateModel<schema::Primitive> *model) {
+  KernelIter from, end;
+  std::vector<CoreMLOp *> coreml_ops;
+  for (KernelIter iter = model->BeginKernelIterator(); iter != model->EndKernelIterator(); iter++) {
+    kernel::Kernel *kernel = *iter;
+    auto coreml_op = GetOP(kernel, model->GetPrimitive(kernel));
+    if (coreml_op != nullptr) {
+      // If coreml_op does not equal nullptr, this kernel can be supported by delegate
+      if (coreml_ops.size() == 0) {
+        from = iter;
+      }
+      coreml_ops.push_back(coreml_op);
+      end = iter;
+    } else {
+      if (!coreml_ops.empty()) {
+        auto coreml_graph_kernel = CreateCoreMLGraph(coreml_ops, model, from, end);
+        if (coreml_graph_kernel == nullptr) {
+          MS_LOG(ERROR) << "Create CoreML Graph failed.";
+          return mindspore::kLiteNullptr;
+        }
+        iter = model->Replace(from, end + 1, coreml_graph_kernel);
+        coreml_ops.clear();
+      }
+    }
+  }
+  if (!coreml_ops.empty()) {
+    auto coreml_graph_kernel = CreateCoreMLGraph(coreml_ops, model, from, end);
+    if (coreml_graph_kernel == nullptr) {
+      MS_LOG(ERROR) << "Create CoreML Graph failed.";
+      return mindspore::kLiteNullptr;
+    }
+    model->Replace(from, end + 1, coreml_graph_kernel);
+    coreml_ops.clear();
+  }
+  MS_LOG(ERROR) << "CoreML graph build success!";
+  return mindspore::kSuccess;
+}
+
+CoreMLOp *CoreMLDelegate::GetOP(kernel::Kernel *kernel, const schema::Primitive *primitive) {
+  if (primitive == nullptr) {
+    MS_LOG(ERROR) << "primitive is NULL!";
+    return nullptr;
+  }
+  if (kernel == nullptr) {
+    MS_LOG(ERROR) << "kernel is NULL!";
+    return nullptr;
+  }
+  auto name = kernel->name();
+  CoreMLOp *coreml_op = nullptr;
+  auto node_type = primitive->value_type();
+  if (op_func_lists_.find(node_type) != op_func_lists_.end()) {
+    coreml_op = op_func_lists_[node_type](primitive, kernel->inputs(), kernel->outputs(), name);
+  } else {
+    MS_LOG(DEBUG) << "Unsupported op type for CoreML.";
+    return nullptr;
+  }
+
+  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 (int i = 0; i < kernel->outputs().size(); i++) {
+    mindspore::MSTensor tensor = kernel->outputs()[i];
+    if (tensor.DataType() == DataType::kNumberTypeFloat16) {
+      tensor.SetDataType(DataType::kNumberTypeFloat32);
+    }
+  }
+
+  if (coreml_op != nullptr) {
+    MS_LOG(DEBUG) << "kernel: [" << kernel->name().c_str() << "] op success. "
+                  << "op_type: " << lite::PrimitiveCurVersionTypeName(kernel->type());
+  }
+  return coreml_op;
+}
+
+kernel::Kernel *CoreMLDelegate::CreateCoreMLGraph(const std::vector<CoreMLOp *> &ops,
+                                                  DelegateModel<schema::Primitive> *model, KernelIter from,
+                                                  KernelIter end) {
+  auto in_tensors = lite::GetGraphInTensors(ops, nullptr);
+  auto out_tensors = lite::GraphOutTensors<CoreMLOp>(ops, model, from, end);
+  auto graph_kernel = new (std::nothrow) CoreMLGraph(ops, in_tensors, out_tensors);
+  if (graph_kernel == nullptr) {
+    MS_LOG(ERROR) << "New CoreML Graph failed.";
+    return nullptr;
+  }
+  graph_kernel->set_name("CoreMLGraph" + std::to_string(graph_index_++));
+
+  // 1. For every op, find pre and next ops
+  lite::FindPreNextOps<CoreMLOp>(ops);
+
+  // 2. Run pass
+  auto ret = pass_manager_->RunPass(graph_kernel);
+  if (ret != RET_OK) {
+    delete graph_kernel;
+    MS_LOG(ERROR) << "CoreML Graph run pass failed. This function mainly solves the problem that the format is "
+                     "inconsistent and requires interpolation transpose operators.";
+    return nullptr;
+  }
+
+  // 3. CoreMLGraph init, build and compile the MLModel
+  ret = graph_kernel->Init();
+  if (ret != RET_OK) {
+    delete graph_kernel;
+    MS_LOG(ERROR) << "CoreML subgraph Init failed.";
+    return nullptr;
+  }
+  return graph_kernel;
+}
+}  // namespace mindspore

mindspore/lite/src/runtime/delegate/coreml/coreml_executor.h

@@ -0,0 +1,50 @@
+/**
+* Copyright 2022 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_RUNTIME_DELEGATE_COREML_COREML_EXECUTOR_H_
+#define MINDSPORE_LITE_SRC_RUNTIME_DELEGATE_COREML_COREML_EXECUTOR_H_
+
+#import <CoreML/CoreML.h>
+#import <Foundation/Foundation.h>
+#include <string>
+#include <vector>
+#include "include/api/types.h"
+
+API_AVAILABLE(ios(11))
+@interface InputFeatureProvider : NSObject <MLFeatureProvider> {
+  const std::vector<mindspore::MSTensor>* _inputs;
+  NSSet* _featureNames;
+}
+
+- (instancetype)initWithInputs:(const std::vector<mindspore::MSTensor>*)inputs
+                 coreMLVersion:(int)coreMLVersion;
+- (NSSet<NSString*>*)featureNames;
+- (MLFeatureValue *)featureValueForName:(NSString *)featureName;
+
+@property(nonatomic, readonly) int coreMLVersion;
+@end
+
+API_AVAILABLE(ios(11))
+@interface CoreMLExecutor : NSObject
+
+- (bool)ExecuteWithInputs:(const std::vector<mindspore::MSTensor>&)inputs
+                  outputs:(const std::vector<mindspore::MSTensor>&)outputs;
+
+- (bool)loadModelC:(NSURL*)compileUrl;
+
+@property MLModel* model;
+@property(nonatomic, readonly) int coreMLVersion;
+@end
+#endif  // MINDSPORE_LITE_SRC_RUNTIME_DELEGATE_COREML_COREML_EXECUTOR_H_

mindspore/lite/src/runtime/delegate/coreml/coreml_executor.mm

@@ -0,0 +1,198 @@
+/**
+ * Copyright 2022 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.
+ */
+
+#import "src/runtime/delegate/coreml/coreml_executor.h"
+#include <fstream>
+#include <iostream>
+
+namespace {
+// The subgraph split can cause the change of tensor name. This function is used to get the original name.
+std::string GetOrgFeatureName(const std::string &input_name) {
+  auto org_name = input_name;
+  std::string pattern_1 = "_duplicate_";
+  auto pos_1 = input_name.find(pattern_1);
+  if (pos_1 != std::string::npos) {
+    org_name = input_name.substr(pos_1 + pattern_1.length());
+    return org_name;
+  }
+  std::string pattern_2 = "_duplicate";
+  auto pos_2 = input_name.find(pattern_2);
+  if (pos_2 != std::string::npos) {
+    org_name = input_name.substr(0, pos_2);
+    return org_name;
+  }
+  return org_name;
+}
+}  // namespace
+
+@implementation InputFeatureProvider
+
+- (instancetype)initWithInputs:(const std::vector<mindspore::MSTensor>*)inputs
+                 coreMLVersion:(int)coreMLVersion {
+  self = [super init];
+  _inputs = inputs;
+  _coreMLVersion = coreMLVersion;
+  NSMutableArray* names = [[NSMutableArray alloc] init];
+  for (auto& input : *_inputs) {
+    auto input_name = GetOrgFeatureName(input.Name());
+    [names addObject:[NSString stringWithCString:input_name.c_str()
+                                        encoding:[NSString defaultCStringEncoding]]];
+  }
+  _featureNames = [NSSet setWithArray:names];
+  return self;
+}
+
+- (NSSet<NSString*>*)featureNames{ return _featureNames; }
+
+- (MLFeatureValue*)featureValueForName:(NSString*)featureName {
+  for (auto input : *_inputs) {
+    auto input_name = GetOrgFeatureName(input.Name());
+    if ([featureName cStringUsingEncoding:NSUTF8StringEncoding] == input_name) {
+      NSArray* shape;
+      NSArray* strides;
+      int tensorRank = input.Shape().size();
+      switch(tensorRank) {
+        case 1:
+          shape = @[
+            @(input.Shape()[0])
+          ];
+          strides = @[
+            @1
+          ];
+          break;
+        case 2:
+          shape = @[
+            @(input.Shape()[0]),
+            @(input.Shape()[1])
+          ];
+          strides = @[
+            @(input.Shape()[1]),
+            @1
+          ];
+          break;
+        case 3:
+          shape = @[
+            @(input.Shape()[0]),
+            @(input.Shape()[1]),
+            @(input.Shape()[2])
+          ];
+          strides = @[
+            @(input.Shape()[2] * input.Shape()[1]),
+            @(input.Shape()[2]),
+            @1
+          ];
+          break;
+        case 4:
+          shape = @[
+            @(input.Shape()[0]),
+            @(input.Shape()[1]),
+            @(input.Shape()[2]),
+            @(input.Shape()[3])
+          ];
+          strides = @[
+            @(input.Shape()[3] * input.Shape()[2] * input.Shape()[1]),
+            @(input.Shape()[3] * input.Shape()[2]),
+            @(input.Shape()[3]),
+            @1
+          ];
+          break;
+        default:
+          NSLog(@"The rank of input tensor:%@ is unsupported!", featureName);
+      }
+
+      NSError* error = nil;
+      MLMultiArray* mlArray = [[MLMultiArray alloc] initWithDataPointer:(float*)input.MutableData()
+                                                                  shape:shape
+                                                               dataType:MLMultiArrayDataTypeFloat32
+                                                                strides:strides
+                                                            deallocator:(^(void* bytes){
+                                                                        })error:&error];
+      if (error != nil) {
+        NSLog(@"Failed to create MLMultiArray for input tensor %@ error: %@!", featureName,
+              [error localizedDescription]);
+        return nil;
+      }
+      auto* mlFeatureValue = [MLFeatureValue featureValueWithMultiArray:mlArray];
+      return mlFeatureValue;
+    }
+  }
+
+  NSLog(@"Input tensor %@ not found!", featureName);
+  return nil;
+}
+@end
+
+@implementation CoreMLExecutor
+
+- (bool)ExecuteWithInputs:(const std::vector<mindspore::MSTensor>&)inputs
+                 outputs:(const std::vector<mindspore::MSTensor>&)outputs {
+  if (_model == nil) {
+    return NO;
+  }
+  _coreMLVersion = 3;
+  NSError* error = nil;
+  //Initialize the CoreML feature provider with input MSTensor
+  InputFeatureProvider* inputFeature =
+      [[InputFeatureProvider alloc] initWithInputs:&inputs coreMLVersion:[self coreMLVersion]];
+  if (inputFeature == nil) {
+    NSLog(@"inputFeature initialization failed.");
+    return NO;
+  }
+  //Inference configuration, auto use GPU by default
+  MLPredictionOptions* options = [[MLPredictionOptions alloc] init];
+
+  //inference with specific input
+  id<MLFeatureProvider> outputFeature = [_model predictionFromFeatures:inputFeature
+                                                               options:options
+                                                                 error:&error];
+  if (error != nil) {
+    NSLog(@"Execute model failed, error code: %@", [error localizedDescription]);
+    return NO;
+  }
+  NSSet<NSString*>* outputFeatureNames = [outputFeature featureNames];
+  for (auto output : outputs) {
+    auto orgOutputName = GetOrgFeatureName(output.Name());
+    NSString* outputName = [NSString stringWithCString:orgOutputName.c_str()
+                                              encoding:[NSString defaultCStringEncoding]];
+    MLFeatureValue* outputValue =
+        [outputFeature featureValueForName:[outputFeatureNames member:outputName]];
+    auto* data = [outputValue multiArrayValue];
+    float* outputData = (float*)data.dataPointer;
+    if (outputData == nullptr) {
+      NSLog(@"Output data is null!");
+      return NO;
+    }
+    memcpy(output.MutableData(), outputData, output.DataSize());
+  }
+  return YES;
+}
+
+- (bool)loadModelC:(NSURL*)compileUrl {
+  NSError* error = nil;
+  if (@available(iOS 12.0, *)) {
+    MLModelConfiguration* config = [MLModelConfiguration alloc];
+    config.computeUnits = MLComputeUnitsAll;
+    _model = [MLModel modelWithContentsOfURL:compileUrl configuration:config error:&error];
+  } else {
+    _model = [MLModel modelWithContentsOfURL:compileUrl error:&error];
+  }
+  if (error != NULL) {
+    NSLog(@"Create MLModel failed, error code: %@", [error localizedDescription]);
+    return NO;
+  }
+  return YES;
+}
+@end

mindspore/lite/src/runtime/delegate/coreml/coreml_executor_wrapper.h

@@ -0,0 +1,43 @@
+/**
+ * Copyright 2022 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_RUNTIME_DELEGATE_COREML_COREML_EXECUTOR_WRAPPER_H_
+#define MINDSPORE_LITE_SRC_RUNTIME_DELEGATE_COREML_COREML_EXECUTOR_WRAPPER_H_
+#include <vector>
+#include <string>
+#include "include/errorcode.h"
+#include "include/api/types.h"
+
+namespace mindspore {
+class CoreMLExecutorWrapper {
+ public:
+  CoreMLExecutorWrapper();
+
+  ~CoreMLExecutorWrapper();
+
+  int Run(const std::vector<mindspore::MSTensor> &in_tensors, const std::vector<mindspore::MSTensor> &out_tensors);
+
+  int CompileMLModel(const std::string &modelPath);
+
+  int CleanTmpFile();
+
+ private:
+  void *coreml_executor_ = nullptr;
+  std::string mlmodel_path_;
+  std::string mlmodelc_path_;
+};
+}  // namespace mindspore
+#endif  // MINDSPORE_LITE_SRC_RUNTIME_DELEGATE_COREML_COREML_EXECUTOR_WRAPPER_H_

mindspore/lite/src/runtime/delegate/coreml/coreml_executor_wrapper.mm

@@ -0,0 +1,93 @@
+/**
+ * Copyright 2022 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/runtime/delegate/coreml/coreml_executor_wrapper.h"
+#import "src/runtime/delegate/coreml/coreml_executor.h"
+
+namespace mindspore {
+CoreMLExecutorWrapper::CoreMLExecutorWrapper() {
+  if (coreml_executor_ == nullptr) {
+    //cast object-c ptr to c ptr, and transfer its ownership to a c object to avoid auto release
+    coreml_executor_ = (__bridge_retained void*)[CoreMLExecutor new];
+  }
+}
+
+CoreMLExecutorWrapper::~CoreMLExecutorWrapper() {
+  //cast c ptr to object-c ptr, and transfer its ownership to an ARC object which is able to auto release
+  auto arc_executor = (__bridge_transfer CoreMLExecutor*)coreml_executor_;
+  (void)arc_executor;
+  coreml_executor_ = nullptr;
+}
+
+int CoreMLExecutorWrapper::CompileMLModel(const std::string &modelPath) {
+  mlmodel_path_ = modelPath;
+  NSString *MLModelSrcPath = [NSString stringWithCString:modelPath.c_str() encoding:[NSString defaultCStringEncoding]];
+  NSError *error = nil;
+  NSURL *MLModelCURL = [MLModel compileModelAtURL:[NSURL fileURLWithPath:MLModelSrcPath] error:nil];
+  if (error) {
+    NSLog(@"Compile MLModel to MLModelC Error: %@", error);
+    (void)CleanTmpFile();
+    return lite::RET_ERROR;
+  }
+  mlmodelc_path_ = [[MLModelCURL path] UTF8String];
+  bool success = [(__bridge id)coreml_executor_ loadModelC:MLModelCURL];
+  if (!success) {
+    NSLog(@"Load MLModelC failed!");
+    (void)CleanTmpFile();
+    return lite::RET_ERROR;
+  }
+  auto ret = CleanTmpFile();
+  if (ret != lite::RET_OK) {
+    NSLog(@"Clean temp model file failed!");
+  }
+  return lite::RET_OK;
+}
+
+int CoreMLExecutorWrapper::Run(const std::vector<mindspore::MSTensor> &in_tensors,
+                               const std::vector<mindspore::MSTensor> &out_tensors){
+  auto success = [(__bridge id)coreml_executor_ ExecuteWithInputs:in_tensors outputs:out_tensors];
+  if (!success) {
+    NSLog(@"coreML model execute failed!");
+    return lite::RET_ERROR;
+  }
+  NSLog(@"coreML model execute success!");
+  return lite::RET_OK;
+}
+
+int CoreMLExecutorWrapper::CleanTmpFile() {
+  NSError* error = nil;
+  NSString *mlModelPath = [NSString stringWithCString:mlmodel_path_.c_str() encoding:[NSString defaultCStringEncoding]];
+  NSString *mlModelCPath = [NSString stringWithCString:mlmodelc_path_.c_str() encoding:[NSString defaultCStringEncoding]];
+  NSFileManager *fileManager = [NSFileManager defaultManager];
+  bool isDir = NO;
+  if ([fileManager fileExistsAtPath:mlModelPath isDirectory:&isDir] && isDir) {
+    [fileManager removeItemAtPath:mlModelPath error:&error];
+    if (error != nil) {
+      NSLog(@"Failed cleaning up model: %@", [error localizedDescription]);
+      return lite::RET_ERROR;
+    }
+  }
+  isDir = NO;
+  if ([fileManager fileExistsAtPath:mlModelCPath isDirectory:&isDir] && isDir) {
+    [fileManager removeItemAtPath:mlModelCPath error:&error];
+    if (error != nil) {
+      NSLog(@"Failed cleaning up compiled model: %@", [error localizedDescription]);
+      return lite::RET_ERROR;
+    }
+  }
+  return lite::RET_OK;
+}
+}  // namespace mindspore

mindspore/lite/src/runtime/delegate/coreml/coreml_graph.cc

@@ -0,0 +1,171 @@
+/**
+ * Copyright 2022 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/runtime/delegate/coreml/coreml_graph.h"
+#include <fstream>
+namespace mindspore {
+CoreMLGraph::~CoreMLGraph() {
+  for (auto *kernel : all_kernels_) {
+    delete kernel;
+  }
+  for (auto *op : coreml_ops_) {
+    delete op;
+  }
+  for (auto tensor : insert_tensors_) {
+    MSTensor::DestroyTensorPtr(tensor);
+  }
+  delete ml_model_;
+  delete executor_wrapper_;
+}
+
+void CoreMLGraph::set_input(mindspore::MSTensor in_tensor, int index) {
+  MS_ASSERT(static_cast<size_t>(index) < inputs_.size());
+  auto origin_tensor = this->inputs_[index];
+  for (auto kernel : all_kernels_) {
+    for (size_t i = 0; i < kernel->inputs().size(); i++) {
+      if (kernel->inputs()[i] == origin_tensor) {
+        kernel->set_input(in_tensor, i);
+      }
+    }
+  }
+  this->inputs_[index] = in_tensor;
+}
+
+void CoreMLGraph::set_output(mindspore::MSTensor out_tensor, int index) {
+  MS_ASSERT(static_cast<size_t>(index) < outputs_.size());
+  auto origin_tensor = this->outputs_[index];
+  for (auto kernel : all_kernels_) {
+    for (size_t i = 0; i < kernel->outputs().size(); i++) {
+      if (kernel->outputs()[i] == origin_tensor) {
+        kernel->set_output(out_tensor, i);
+      }
+    }
+  }
+  this->outputs_[index] = out_tensor;
+}
+
+int CoreMLGraph::Init() {
+  ml_model_ = BuildMLModel();
+  if (ml_model_ == nullptr) {
+    MS_LOG(ERROR) << "Build CoreML model failed.";
+    return RET_ERROR;
+  }
+  auto model_path = SaveMLModel();
+  executor_wrapper_ = new (std::nothrow) CoreMLExecutorWrapper();
+  if (executor_wrapper_ == nullptr) {
+    MS_LOG(ERROR) << "Create CoreML executor wrapper failed.";
+    return RET_ERROR;
+  }
+  auto ret = executor_wrapper_->CompileMLModel(model_path);
+  if (ret != RET_OK) {
+    MS_LOG(ERROR) << "Compile coreML model failed!";
+    return RET_ERROR;
+  }
+  return RET_OK;
+}
+
+CoreML::Specification::Model *CoreMLGraph::BuildMLModel() {
+  auto *model = new (std::nothrow) CoreML::Specification::Model();
+  model->set_specificationversion(kCoreMLVersion4);
+  model->mutable_neuralnetwork()->set_arrayinputshapemapping(CoreML::Specification::EXACT_ARRAY_MAPPING);
+  auto *network = model->mutable_neuralnetwork();
+  for (auto &op : coreml_ops_) {
+    auto ret = op->BuildLayer();
+    if (ret != RET_OK) {
+      MS_LOG(ERROR) << "Failed to build layer for op: " << op->name();
+      delete model;
+      model = nullptr;
+      return nullptr;
+    }
+    op->SetMLOpInOut();
+    auto layers = op->GetLayers();
+    if (layers.empty()) {
+      MS_LOG(ERROR) << "No layer found for op: " << op->name();
+      delete model;
+      model = nullptr;
+      return nullptr;
+    }
+    for (auto layer : layers) {
+      MS_ASSERT(layer != nullptr);
+      network->mutable_layers()->AddAllocated(layer);
+    }
+  }
+  auto ret = SetMLModelInOut(model);
+  if (ret != RET_OK) {
+    MS_LOG(ERROR) << "Set model input output failed.";
+    delete model;
+    model = nullptr;
+    return nullptr;
+  }
+  return model;
+}
+
+int CoreMLGraph::SetMLModelInOut(CoreML::Specification::Model *model) {
+  MS_ASSERT(model != nullptr);
+  auto model_desc = model->mutable_description();
+  for (const auto &in_tensor : this->inputs_) {
+    // add input
+    auto input = model_desc->add_input();
+    input->set_name(in_tensor.Name());
+    auto in_multi_array = input->mutable_type()->mutable_multiarraytype();
+    if (in_tensor.DataType() == DataType::kNumberTypeFloat32) {
+      in_multi_array->set_datatype(CoreML::Specification::ArrayFeatureType::FLOAT32);
+    } else if (in_tensor.DataType() == DataType::kNumberTypeInt32) {
+      in_multi_array->set_datatype(CoreML::Specification::ArrayFeatureType::INT32);
+    } else {
+      MS_LOG(ERROR) << "Unsupported model input data type: " << static_cast<int>(in_tensor.DataType());
+      return RET_ERROR;
+    }
+    for (int64_t i : in_tensor.Shape()) {
+      in_multi_array->add_shape(static_cast<uint64_t>(i));
+    }
+  }
+  for (const auto &out_tensor : this->outputs_) {
+    // add output
+    auto output = model_desc->add_output();
+    output->set_name(out_tensor.Name());
+    auto out_multi_array = output->mutable_type()->mutable_multiarraytype();
+    if (out_tensor.DataType() == DataType::kNumberTypeFloat32) {
+      out_multi_array->set_datatype(CoreML::Specification::ArrayFeatureType::FLOAT32);
+    } else if (out_tensor.DataType() == DataType::kNumberTypeInt32) {
+      out_multi_array->set_datatype(CoreML::Specification::ArrayFeatureType::INT32);
+    } else {
+      MS_LOG(ERROR) << "Unsupported model output data type: " << static_cast<int>(out_tensor.DataType());
+      return RET_ERROR;
+    }
+    for (int64_t i : out_tensor.Shape()) {
+      out_multi_array->add_shape(static_cast<uint64_t>(i));
+    }
+  }
+  return RET_OK;
+}
+
+std::string CoreMLGraph::SaveMLModel() {
+  MS_ASSERT(ml_model_ != nullptr);
+  std::string model_name = this->name() + ".mlmodel";
+  auto model_path = std::string(getenv("HOME")) + "/tmp/" + model_name;
+  std::ofstream file_stream(model_path, std::ios::out | std::ios::binary);
+  ml_model_->SerializeToOstream(&file_stream);
+  MS_LOG(ERROR) << "Build CoreML model success!";
+  return model_path;
+}
+
+int CoreMLGraph::Execute() {
+  auto ret = executor_wrapper_->Run(inputs(), outputs());
+  MS_LOG(INFO) << "run model success!";
+  return ret;
+}
+}  // namespace mindspore

mindspore/lite/src/runtime/delegate/coreml/coreml_graph.h

@@ -0,0 +1,75 @@
+/**
+ * Copyright 2022 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_RUNTIME_DELEGATE_COREML_COREML_GRAPH_H_
+#define MINDSPORE_LITE_SRC_RUNTIME_DELEGATE_COREML_COREML_GRAPH_H_
+
+#include <vector>
+#include <queue>
+#include <map>
+#include <string>
+#include <utility>
+#include "proto/Model.pb.h"
+#include "proto/NeuralNetwork.pb.h"
+#include "include/api/kernel.h"
+#include "src/runtime/delegate/coreml/op/coreml_op.h"
+#include "src/runtime/delegate/coreml/coreml_executor_wrapper.h"
+
+namespace mindspore {
+constexpr int kCoreMLVersion4 = 4;
+class CoreMLGraph : public kernel::Kernel {
+ public:
+  CoreMLGraph(std::vector<CoreMLOp *> coreml_ops, const std::vector<mindspore::MSTensor> &inputs,
+              const std::vector<mindspore::MSTensor> &outputs)
+      : kernel::Kernel(inputs, outputs, nullptr, nullptr), coreml_ops_(std::move(coreml_ops)) {}
+
+  ~CoreMLGraph() override;
+
+  int Init();
+
+  int Prepare() override { return lite::RET_OK; }
+
+  int Execute() override;
+
+  int ReSize() override {
+    MS_LOG(ERROR) << "CoreML does not support the resize function temporarily.";
+    return lite::RET_ERROR;
+  }
+
+  void set_input(mindspore::MSTensor in_tensor, int index) override;
+
+  void set_output(mindspore::MSTensor out_tensor, int index) override;
+
+  std::vector<CoreMLOp *> *GetOps() { return &coreml_ops_; }
+
+  std::vector<mindspore::MSTensor *> *GetInsertTensors() { return &insert_tensors_; }
+
+ protected:
+  CoreML::Specification::Model *BuildMLModel();
+
+  int SetMLModelInOut(CoreML::Specification::Model *model);
+
+  std::string SaveMLModel();
+
+  std::vector<CoreMLOp *> coreml_ops_{};
+  std::vector<kernel::Kernel *> all_kernels_{};
+  CoreML::Specification::Model *ml_model_ = nullptr;
+  CoreMLExecutorWrapper *executor_wrapper_ = nullptr;
+  std::vector<mindspore::MSTensor *> insert_tensors_;
+};
+}  // namespace mindspore
+
+#endif  // MINDSPORE_LITE_SRC_RUNTIME_DELEGATE_COREML_COREML_GRAPH_H_

mindspore/lite/src/runtime/delegate/coreml/op/activation_coreml.cc

@@ -0,0 +1,60 @@
+/**
+ * Copyright 2022 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/runtime/delegate/coreml/op/activation_coreml.h"
+namespace mindspore {
+int ActivationCoreMLOp::IsSupport() {
+  auto act_prim = op_primitive_->value_as_Activation();
+  if (act_prim == nullptr) {
+    MS_LOG(ERROR) << "Get null primitive value for op ." << name_;
+    return RET_ERROR;
+  }
+  act_type_ = act_prim->activation_type();
+  if (act_type_ != schema::ActivationType_RELU && act_type_ != schema::ActivationType_RELU6 &&
+      act_type_ != schema::ActivationType_SIGMOID && act_type_ != schema::ActivationType_TANH &&
+      act_type_ != schema::ActivationType_HSIGMOID && act_type_ != schema::ActivationType_LEAKY_RELU &&
+      act_type_ != schema::ActivationType_SWISH && act_type_ != schema::ActivationType_ELU) {
+    MS_LOG(WARNING) << "Unsupported activation type for activation op " << name_ << "when running coreML.";
+    return RET_NOT_SUPPORT;
+  }
+  return RET_OK;
+}
+
+int ActivationCoreMLOp::BuildLayer() {
+  MS_ASSERT(op_ != nullptr);
+  switch (act_type_) {
+    case schema::ActivationType_RELU:
+      op_->mutable_activation()->mutable_relu();
+      break;
+    case schema::ActivationType_RELU6: {
+      auto clip_param = act_op_->mutable_clip();
+      clip_param->set_minval(0);
+      clip_param->set_maxval(kValueThreshold6);
+      break;
+    }
+    case schema::ActivationType_TANH:
+      op_->mutable_activation()->mutable_tanh();
+      break;
+    case schema::ActivationType_SIGMOID:
+      op_->mutable_activation()->mutable_sigmoid();
+      break;
+    default:
+      MS_LOG(ERROR) << "Unsupported activation type.";
+      return RET_ERROR;
+  }
+  return RET_OK;
+}
+}  // namespace mindspore

mindspore/lite/src/runtime/delegate/coreml/op/activation_coreml.h

@@ -0,0 +1,37 @@
+/**
+ * Copyright 2022 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_RUNTIME_DELEGATE_COREML_OP_ACTIVATION_COREML_H_
+#define MINDSPORE_LITE_SRC_RUNTIME_DELEGATE_COREML_OP_ACTIVATION_COREML_H_
+
+#include <vector>
+#include <string>
+#include "src/runtime/delegate/coreml/op/coreml_op.h"
+namespace mindspore {
+class ActivationCoreMLOp : public CoreMLOp {
+ public:
+  ActivationCoreMLOp(const schema::Primitive *primitive, const std::vector<mindspore::MSTensor> &in_tensors,
+                     const std::vector<mindspore::MSTensor> &out_tensors, std::string name)
+      : CoreMLOp(primitive, in_tensors, out_tensors, name) {}
+
+  int IsSupport() override;
+
+  int BuildLayer() override;
+
+ private:
+  schema::ActivationType act_type_ = schema::ActivationType_NO_ACTIVATION;
+};
+}  // namespace mindspore
+#endif  // MINDSPORE_LITE_SRC_RUNTIME_DELEGATE_COREML_OP_ACTIVATION_COREML_H_

mindspore/lite/src/runtime/delegate/coreml/op/arithmetic_coreml.cc

@@ -0,0 +1,104 @@
+/**
+ * Copyright 2022 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/runtime/delegate/coreml/op/arithmetic_coreml.h"
+namespace mindspore {
+int ArithmeticCoreMLOp::IsSupport() {
+  MS_CHECK_TRUE_MSG(in_tensors_.size() == kInputSize1, RET_NOT_SUPPORT, "Arithmetic op only support two inputs.");
+  auto input_a = in_tensors_.at(0);
+  auto input_b = in_tensors_.at(1);
+  if ((input_a.IsConst() && input_a.ElementNum() == 1) || (input_b.IsConst() && input_b.ElementNum() == 1)) {
+    use_normal_ = true;
+  }
+  return RET_OK;
+}
+
+int ArithmeticCoreMLOp::BuildLayer() {
+  if (use_normal_) {
+    auto ret = BuildNormalArithmetic();
+    if (ret != RET_OK) {
+      MS_LOG(ERROR) << "Build normal arithmetic layer failed for op: " << name_;
+      return RET_ERROR;
+    }
+    return RET_OK;
+  }
+  auto ret = BuildBroadcastableArithmetic();
+  if (ret != RET_OK) {
+    MS_LOG(ERROR) << "Build broadcastable arithmetic layer failed for op: " << name_;
+    return RET_ERROR;
+  }
+  return RET_OK;
+}
+
+int ArithmeticCoreMLOp::BuildNormalArithmetic() {
+  MS_ASSERT(op_ != nullptr);
+  switch (type_) {
+    case schema::PrimitiveType_AddFusion: {
+      auto add_param = op_->mutable_add();
+      SetNormalConst<CoreML::Specification::AddLayerParams>(add_param);
+      break;
+    }
+    case schema::PrimitiveType_MulFusion: {
+      auto mul_param = op_->mutable_multiply();
+      SetNormalConst<CoreML::Specification::MultiplyLayerParams>(mul_param);
+      break;
+    }
+    default:
+      MS_LOG(ERROR) << "Unsupported arithmetic type.";
+      return RET_ERROR;
+  }
+  return RET_OK;
+}
+
+int ArithmeticCoreMLOp::BuildBroadcastableArithmetic() {
+  MS_ASSERT(op_ != nullptr);
+  switch (type_) {
+    case schema::PrimitiveType_AddFusion:
+      (void)op_->mutable_addbroadcastable();
+      break;
+    case schema::PrimitiveType_MulFusion:
+      (void)op_->mutable_multiplybroadcastable();
+      break;
+    default:
+      MS_LOG(ERROR) << "Unsupported arithmetic type.";
+      return RET_ERROR;
+  }
+  for (const auto &in_tensor : in_tensors_) {
+    if (in_tensor.IsConst()) {
+      auto ret = SetConstInput(in_tensor);
+      if (ret != RET_OK) {
+        MS_LOG(ERROR) << "Set const input failed for op: " << name_;
+        return RET_ERROR;
+      }
+    }
+  }
+  return RET_OK;
+}
+
+void ArithmeticCoreMLOp::SetMLOpInOut() {
+  MS_ASSERT(op_ != nullptr);
+  for (const auto &in_tensor : in_tensors_) {
+    if (in_tensor.IsConst() && !use_normal_) {
+      // const op has not input
+      const_ops_[in_tensor.Name()]->add_output(in_tensor.Name());
+    }
+    if (!(in_tensor.IsConst() && use_normal_)) {
+      op_->add_input(in_tensor.Name());
+    }
+  }
+  op_->add_output(out_tensors_[0].Name());
+}
+}  // namespace mindspore

mindspore/lite/src/runtime/delegate/coreml/op/arithmetic_coreml.h

@@ -0,0 +1,58 @@
+/**
+ * Copyright 2022 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_RUNTIME_DELEGATE_COREML_OP_ARITHMETIC_COREML_H_
+#define MINDSPORE_LITE_SRC_RUNTIME_DELEGATE_COREML_OP_ARITHMETIC_COREML_H_
+
+#include <vector>
+#include <string>
+#include "src/runtime/delegate/coreml/op/coreml_op.h"
+namespace mindspore {
+class ArithmeticCoreMLOp : public CoreMLOp {
+ public:
+  ArithmeticCoreMLOp(const schema::Primitive *primitive, const std::vector<mindspore::MSTensor> &in_tensors,
+                     const std::vector<mindspore::MSTensor> &out_tensors, std::string name)
+      : CoreMLOp(primitive, in_tensors, out_tensors, name) {}
+
+  int IsSupport() override;
+
+  int BuildLayer() override;
+
+  int BuildNormalArithmetic();
+
+  int BuildBroadcastableArithmetic();
+
+ protected:
+  void SetMLOpInOut() override;
+
+  template <typename T>
+  void SetNormalConst(T *arithmetic_param) {
+    const void *org_data = nullptr;
+    if (in_tensors_[0].IsConst()) {
+      org_data = in_tensors_[0].Data().get();
+    } else if (in_tensors_[1].IsConst()) {
+      org_data = in_tensors_[1].Data().get();
+    }
+    if (org_data != nullptr) {
+      auto const_data = reinterpret_cast<const float *>(org_data);
+      arithmetic_param->set_alpha(const_data[0]);
+    }
+  }
+
+ protected:
+  bool use_normal_ = false;
+};
+}  // namespace mindspore
+#endif  // MINDSPORE_LITE_SRC_RUNTIME_DELEGATE_COREML_OP_ARITHMETIC_COREML_H_

mindspore/lite/src/runtime/delegate/coreml/op/avg_pooling_coreml.cc

@@ -0,0 +1,72 @@
+/**
+ * Copyright 2022 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/runtime/delegate/coreml/op/avg_pooling_coreml.h"
+namespace mindspore {
+int AvgPoolingCoreMLOp::InitParams() {
+  pooling_prim_ = op_primitive_->value_as_AvgPoolFusion();
+  if (pooling_prim_ == nullptr) {
+    MS_LOG(ERROR) << "Get null primitive value for op ." << name_;
+    return RET_ERROR;
+  }
+  return RET_OK;
+}
+
+int AvgPoolingCoreMLOp::BuildLayer() {
+  MS_ASSERT(op_ != nullptr);
+  auto pooling_param = op_->mutable_pooling();
+  pooling_param->set_type(CoreML::Specification::PoolingLayerParams::AVERAGE);
+  if (pooling_prim_->global()) {
+    pooling_param->set_globalpooling(true);
+    pooling_param->mutable_valid();
+    return RET_OK;
+  }
+  pooling_param->set_avgpoolexcludepadding(true);
+  auto kernel_h = static_cast<int>(*(pooling_prim_->kernel_size()->begin()));
+  auto kernel_w = static_cast<int>(*(pooling_prim_->kernel_size()->begin() + 1));
+  auto stride_h = static_cast<int>(*(pooling_prim_->strides()->begin()));
+  auto stride_w = static_cast<int>(*(pooling_prim_->strides()->begin() + 1));
+  pooling_param->add_stride(stride_h);
+  pooling_param->add_stride(stride_w);
+  pooling_param->add_kernelsize(kernel_h);
+  pooling_param->add_kernelsize(kernel_w);
+  if (pooling_prim_->pad_mode() == schema::PadMode_SAME) {
+    pooling_param->mutable_same();
+  } else {
+    pooling_param->mutable_valid();
+    if (pooling_prim_->pad() != nullptr) {
+      auto pad_u = static_cast<int>(*(pooling_prim_->pad()->begin() + PAD_UP));
+      auto pad_d = static_cast<int>(*(pooling_prim_->pad()->begin() + PAD_DOWN));
+      auto pad_l = static_cast<int>(*(pooling_prim_->pad()->begin() + PAD_LEFT));
+      auto pad_r = static_cast<int>(*(pooling_prim_->pad()->begin() + PAD_RIGHT));
+      auto ret = SetPadding({pad_u, pad_d, pad_l, pad_r});
+      if (ret != RET_OK) {
+        MS_LOG(ERROR) << "Fail to set padding for op: " << name_;
+        return RET_ERROR;
+      }
+    }
+  }
+  auto act_type = pooling_prim_->activation_type();
+  if (act_type != schema::ActivationType_NO_ACTIVATION) {
+    auto ret = SetActivation(act_type);
+    if (ret != RET_OK) {
+      MS_LOG(ERROR) << "Set pooling activation failed for op: " << name_;
+      return RET_ERROR;
+    }
+  }
+  return RET_OK;
+}
+}  // namespace mindspore

mindspore/lite/src/runtime/delegate/coreml/op/avg_pooling_coreml.h

@@ -0,0 +1,39 @@
+/**
+ * Copyright 2022 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_RUNTIME_DELEGATE_COREML_OP_AVG_POOLING_COREML_H_
+#define MINDSPORE_LITE_SRC_RUNTIME_DELEGATE_COREML_OP_AVG_POOLING_COREML_H_
+
+#include <vector>
+#include <string>
+#include <utility>
+#include <unordered_map>
+#include "src/runtime/delegate/coreml/op/coreml_op.h"
+namespace mindspore {
+class AvgPoolingCoreMLOp : public CoreMLOp {
+ public:
+  AvgPoolingCoreMLOp(const schema::Primitive *primitive, const std::vector<mindspore::MSTensor> &in_tensors,
+                     const std::vector<mindspore::MSTensor> &out_tensors, std::string name)
+      : CoreMLOp(primitive, in_tensors, out_tensors, name) {}
+
+  int InitParams() override;
+
+  int BuildLayer() override;
+
+ private:
+  const schema::AvgPoolFusion *pooling_prim_ = nullptr;
+};
+}  // namespace mindspore
+#endif  // MINDSPORE_LITE_SRC_RUNTIME_DELEGATE_COREML_OP_AVG_POOLING_COREML_H_

mindspore/lite/src/runtime/delegate/coreml/op/concat_coreml.cc

@@ -0,0 +1,72 @@
+/**
+ * Copyright 2022 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/runtime/delegate/coreml/op/concat_coreml.h"
+namespace mindspore {
+int ConcatCoreMLOp::IsSupport() {
+  MS_CHECK_GE(in_tensors_.size(), kInputSize1, RET_NOT_SUPPORT);
+  if (std::any_of(in_tensors_.begin(), in_tensors_.end(), [](mindspore::MSTensor &tensor) {
+        return tensor.IsConst() && tensor.DataType() != DataType::kNumberTypeInt32 &&
+               tensor.DataType() != DataType::kNumberTypeFloat32;
+      })) {
+    MS_LOG(ERROR) << "The datatype of CoreML Concat op's constant inputs must be int or float, op name: " << name_;
+    return RET_NOT_SUPPORT;
+  }
+  return RET_OK;
+}
+
+int ConcatCoreMLOp::InitParams() {
+  concat_prim_ = op_primitive_->value_as_Concat();
+  if (concat_prim_ == nullptr) {
+    MS_LOG(ERROR) << "Get null primitive value for op ." << name_;
+    return RET_ERROR;
+  }
+  axis_ = concat_prim_->axis();
+  return RET_OK;
+}
+
+int ConcatCoreMLOp::BuildLayer() {
+  MS_ASSERT(op_ != nullptr);
+  op_->mutable_concatnd()->set_axis(axis_);
+  for (const auto &in_tensor : in_tensors_) {
+    if (in_tensor.IsConst()) {
+      auto ret = SetConstInput(in_tensor);
+      if (ret != RET_OK) {
+        MS_LOG(ERROR) << "Set const input failed for op: " << name_;
+        return RET_ERROR;
+      }
+    }
+  }
+  return RET_OK;
+}
+
+int ConcatCoreMLOp::HandleAxis() {
+  axis_ = NCHW2NHWC_PERM[axis_];
+  return RET_OK;
+}
+
+void ConcatCoreMLOp::SetMLOpInOut() {
+  MS_ASSERT(op_ != nullptr);
+  for (const auto &in_tensor : in_tensors_) {
+    if (in_tensor.IsConst()) {
+      // const op has not input
+      const_ops_[in_tensor.Name()]->add_output(in_tensor.Name());
+    }
+    op_->add_input(in_tensor.Name());
+  }
+  op_->add_output(out_tensors_[0].Name());
+}
+}  // namespace mindspore

mindspore/lite/src/runtime/delegate/coreml/op/concat_coreml.h

@@ -0,0 +1,44 @@
+/**
+ * Copyright 2022 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_RUNTIME_DELEGATE_COREML_OP_CONCAT_COREML_H_
+#define MINDSPORE_LITE_SRC_RUNTIME_DELEGATE_COREML_OP_CONCAT_COREML_H_
+
+#include <vector>
+#include <string>
+#include "src/runtime/delegate/coreml/op/coreml_op.h"
+namespace mindspore {
+class ConcatCoreMLOp : public CoreMLOp {
+ public:
+  ConcatCoreMLOp(const schema::Primitive *primitive, const std::vector<mindspore::MSTensor> &in_tensors,
+                 const std::vector<mindspore::MSTensor> &out_tensors, std::string name)
+      : CoreMLOp(primitive, in_tensors, out_tensors, name) {}
+
+  int IsSupport() override;
+
+  int InitParams() override;
+
+  int HandleAxis() override;
+
+  int BuildLayer() override;
+
+  void SetMLOpInOut() override;
+
+ private:
+  int axis_;
+  const schema::Concat *concat_prim_ = nullptr;
+};
+}  // namespace mindspore
+#endif  // MINDSPORE_LITE_SRC_RUNTIME_DELEGATE_COREML_OP_CONCAT_COREML_H_

mindspore/lite/src/runtime/delegate/coreml/op/convolution_base_coreml.cc

@@ -0,0 +1,89 @@
+/**
+ * Copyright 2022 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/runtime/delegate/coreml/op/convolution_base_coreml.h"
+#include "src/runtime/delegate/delegate_utils.h"
+namespace mindspore {
+int ConvolutionBaseCoreMLOp::SetConvWeight() {
+  auto weight_tensor = in_tensors_.at(kWeightIndex);
+  auto weight_shape = weight_tensor.Shape();
+  conv_param_->set_kernelchannels(weight_shape.at(MS_WT_CIN));
+  conv_param_->set_outputchannels(weight_shape.at(MS_WT_COUT));
+  conv_param_->add_kernelsize(weight_shape.at(MS_WT_H));
+  conv_param_->add_kernelsize(weight_shape.at(MS_WT_W));
+
+  // transpose the weight, (c_out, h, w, c_in) -> (c_out, c_in, h, w)
+  auto org_weight = weight_tensor.Data().get();
+  MS_ASSERT(org_weight != nullptr);
+  if (weight_tensor.DataType() == DataType::kNumberTypeFloat32) {
+    auto *ml_weight_container = conv_param_->mutable_weights()->mutable_floatvalue();
+    ml_weight_container->Resize(weight_tensor.ElementNum(), 0);
+    auto *ml_weight = reinterpret_cast<void *>(ml_weight_container->mutable_data());
+    lite::PackNHWCToNCHWFp32(org_weight, ml_weight, weight_shape[MS_WT_COUT],
+                             weight_shape[MS_WT_H] * weight_shape[MS_WT_W], weight_shape[MS_WT_CIN]);
+  } else {
+    MS_LOG(ERROR) << "Unsupported data type of weight tensor for CoreML convolution.";
+    return RET_ERROR;
+  }
+  return RET_OK;
+}
+
+int ConvolutionBaseCoreMLOp::SetConvBias() {
+  if (in_tensors_.size() >= kInputSize2) {
+    auto bias_tensor = in_tensors_.at(kBiasIndex);
+    auto org_bias = bias_tensor.Data().get();
+    conv_param_->set_hasbias(true);
+    if (bias_tensor.DataType() == DataType::kNumberTypeFloat32) {
+      auto *ml_bias_container = conv_param_->mutable_bias()->mutable_floatvalue();
+      ml_bias_container->Resize(bias_tensor.ElementNum(), 0);
+      auto *ml_bias = reinterpret_cast<void *>(ml_bias_container->mutable_data());
+      memcpy(ml_bias, org_bias, bias_tensor.DataSize());
+    } else {
+      MS_LOG(ERROR) << "Unsupported data type of bias tensor for CoreML convolution.";
+      return RET_ERROR;
+    }
+  }
+  return RET_OK;
+}
+
+int ConvolutionBaseCoreMLOp::BuildLayer() {
+  MS_ASSERT(op_ != nullptr);
+  conv_param_ = op_->mutable_convolution();
+  auto ret = SetConvParam();
+  if (ret != RET_OK) {
+    MS_LOG(ERROR) << "Set conv param failed for op: " << name_;
+    return RET_ERROR;
+  }
+  ret = SetConvWeight();
+  if (ret != RET_OK) {
+    MS_LOG(ERROR) << "Set conv weight failed for op: " << name_;
+    return RET_ERROR;
+  }
+  ret = SetConvBias();
+  if (ret != RET_OK) {
+    MS_LOG(ERROR) << "Set conv bias failed for op: " << name_;
+    return RET_ERROR;
+  }
+  if (act_type_ != schema::ActivationType_NO_ACTIVATION) {
+    ret = SetActivation(act_type_);
+    if (ret != RET_OK) {
+      MS_LOG(ERROR) << "Set conv activation failed for op: " << name_;
+      return RET_ERROR;
+    }
+  }
+  return RET_OK;
+}
+}  // namespace mindspore

mindspore/lite/src/runtime/delegate/coreml/op/convolution_base_coreml.h

@@ -0,0 +1,61 @@
+/**
+ * Copyright 2022 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_RUNTIME_DELEGATE_COREML_OP_CONVOLUTION_BASE_COREML_H_
+#define MINDSPORE_LITE_SRC_RUNTIME_DELEGATE_COREML_OP_CONVOLUTION_BASE_COREML_H_
+
+#include <vector>
+#include <string>
+#include <utility>
+#include <memory>
+#include <unordered_map>
+#include "src/runtime/delegate/coreml/op/coreml_op.h"
+namespace mindspore {
+class ConvolutionBaseCoreMLOp : public CoreMLOp {
+ public:
+  ConvolutionBaseCoreMLOp(const schema::Primitive *primitive, const std::vector<mindspore::MSTensor> &in_tensors,
+                          const std::vector<mindspore::MSTensor> &out_tensors, std::string name)
+      : CoreMLOp(primitive, in_tensors, out_tensors, name) {
+    input_h_ = static_cast<int>(in_tensors.at(0).Shape().at(kNHWC_H));
+    input_w_ = static_cast<int>(in_tensors.at(0).Shape().at(kNHWC_W));
+    kernel_h_ = static_cast<int>(in_tensors.at(1).Shape().at(MS_WT_H));
+    kernel_w_ = static_cast<int>(in_tensors.at(1).Shape().at(MS_WT_W));
+    output_h_ = static_cast<int>(out_tensors.at(0).Shape().at(kNHWC_H));
+    output_w_ = static_cast<int>(out_tensors.at(0).Shape().at(kNHWC_W));
+  }
+
+  int BuildLayer() override;
+
+ protected:
+  virtual int SetConvParam() { return RET_OK; }
+
+  virtual int SetConvWeight();
+
+  virtual int SetConvBias();
+
+ protected:
+  int input_h_;
+  int input_w_;
+  int kernel_h_;
+  int kernel_w_;
+  int output_h_;
+  int output_w_;
+  CoreML::Specification::ConvolutionLayerParams *conv_param_ = nullptr;
+  schema::ActivationType act_type_ = schema::ActivationType_NO_ACTIVATION;
+  std::unique_ptr<CoreML::Specification::NeuralNetworkLayer> trans_in_op_ = nullptr;
+  std::unique_ptr<CoreML::Specification::NeuralNetworkLayer> trans_out_op_ = nullptr;
+};
+}  // namespace mindspore
+#endif  // MINDSPORE_LITE_SRC_RUNTIME_DELEGATE_COREML_OP_CONVOLUTION_BASE_COREML_H_

mindspore/lite/src/runtime/delegate/coreml/op/convolution_coreml.cc

@@ -0,0 +1,71 @@
+/**
+ * Copyright 2022 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/runtime/delegate/coreml/op/convolution_coreml.h"
+#include <cmath>
+#include "src/runtime/delegate/delegate_utils.h"
+namespace mindspore {
+int ConvolutionCoreMLOp::IsSupport() {
+  if (!in_tensors_[kWeightIndex].IsConst()) {
+    MS_LOG(WARNING) << "CoreML convolution does not support dynamic weight.";
+    return RET_NOT_SUPPORT;
+  }
+  conv_prim_ = op_primitive_->value_as_Conv2DFusion();
+  if (conv_prim_ == nullptr) {
+    MS_LOG(ERROR) << "Get null primitive value for op ." << name_;
+    return RET_ERROR;
+  }
+  CHECK_NULL_RETURN(conv_prim_->stride());
+  stride_h_ = static_cast<int>(*(conv_prim_->stride()->begin()));
+  stride_w_ = static_cast<int>(*(conv_prim_->stride()->begin() + 1));
+  CHECK_NULL_RETURN(conv_prim_->dilation());
+  dilation_h_ = static_cast<int>(*(conv_prim_->dilation()->begin()));
+  dilation_w_ = static_cast<int>(*(conv_prim_->dilation()->begin() + 1));
+  // org conv format: NHWC
+  if (stride_h_ > in_tensors_[0].Shape()[kNHWC_H] || stride_w_ > in_tensors_[0].Shape()[kNHWC_W]) {
+    MS_LOG(WARNING) << "CoreML convolution does not support stride greater than input size.";
+    return RET_NOT_SUPPORT;
+  }
+  return RET_OK;
+}
+
+int ConvolutionCoreMLOp::SetConvParam() {
+  auto group = static_cast<int>(conv_prim_->group());
+  conv_param_->set_ngroups(group);
+  conv_param_->add_stride(stride_h_);
+  conv_param_->add_stride(stride_w_);
+  conv_param_->add_dilationfactor(dilation_h_);
+  conv_param_->add_dilationfactor(dilation_w_);
+  if (conv_prim_->pad_mode() == schema::PadMode_SAME) {
+    conv_param_->mutable_same();
+  } else {
+    conv_param_->mutable_valid();
+    if (conv_prim_->pad_list() != nullptr) {
+      auto pad_u = static_cast<int>(*(conv_prim_->pad_list()->begin() + PAD_UP));
+      auto pad_d = static_cast<int>(*(conv_prim_->pad_list()->begin() + PAD_DOWN));
+      auto pad_l = static_cast<int>(*(conv_prim_->pad_list()->begin() + PAD_LEFT));
+      auto pad_r = static_cast<int>(*(conv_prim_->pad_list()->begin() + PAD_RIGHT));
+      auto ret = SetPadding({pad_u, pad_d, pad_l, pad_r});
+      if (ret != RET_OK) {
+        MS_LOG(ERROR) << "Fail to set padding for op: " << name_;
+        return RET_ERROR;
+      }
+    }
+  }
+  act_type_ = conv_prim_->activation_type();
+  return RET_OK;
+}
+}  // namespace mindspore

mindspore/lite/src/runtime/delegate/coreml/op/convolution_coreml.h

@@ -0,0 +1,46 @@
+/**
+ * Copyright 2022 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_RUNTIME_DELEGATE_COREML_OP_CONVOLUTION_COREML_H_
+#define MINDSPORE_LITE_SRC_RUNTIME_DELEGATE_COREML_OP_CONVOLUTION_COREML_H_
+
+#include <vector>
+#include <string>
+#include <utility>
+#include <unordered_map>
+#include "src/runtime/delegate/coreml/op/convolution_base_coreml.h"
+namespace mindspore {
+class ConvolutionCoreMLOp : public ConvolutionBaseCoreMLOp {
+ public:
+  ConvolutionCoreMLOp(const schema::Primitive *primitive, const std::vector<mindspore::MSTensor> &in_tensors,
+                      const std::vector<mindspore::MSTensor> &out_tensors, std::string name)
+      : ConvolutionBaseCoreMLOp(primitive, in_tensors, out_tensors, name) {}
+
+  int IsSupport() override;
+
+ private:
+  schema::PadMode GetPadMode();
+
+  int SetConvParam() override;
+
+ private:
+  int stride_h_;
+  int stride_w_;
+  int dilation_h_;
+  int dilation_w_;
+  const schema::Conv2DFusion *conv_prim_ = nullptr;
+};
+}  // namespace mindspore
+#endif  // MINDSPORE_LITE_SRC_RUNTIME_DELEGATE_COREML_OP_CONVOLUTION_COREML_H_

mindspore/lite/src/runtime/delegate/coreml/op/coreml_op.cc

@@ -0,0 +1,155 @@
+/**
+ * Copyright 2022 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/runtime/delegate/coreml/op/coreml_op.h"
+#include "nnacl/base/cast_base.h"
+namespace mindspore {
+int CoreMLOp::Init() {
+  auto ret = InitParams();
+  if (ret != RET_OK) {
+    MS_LOG(ERROR) << "CoreML op " << name_ << "'s parameter initialization failed.";
+    return RET_ERROR;
+  }
+  op_ = std::make_unique<CoreML::Specification::NeuralNetworkLayer>();
+  if (op_ == nullptr) {
+    MS_LOG(ERROR) << "New CoreML op " << name_ << " failed.";
+    return RET_ERROR;
+  }
+  op_->set_name("CoreML_" + name_);
+  return RET_OK;
+}
+
+int CoreMLOp::SetActivation(schema::ActivationType act_type) {
+  act_op_ = std::make_unique<CoreML::Specification::NeuralNetworkLayer>();
+  if (act_op_ == nullptr) {
+    MS_LOG(ERROR) << "New CoreML op " << name_ << "_activation failed.";
+    return RET_ERROR;
+  }
+  act_op_->set_name("CoreML_" + name_ + "_activation");
+  switch (act_type) {
+    case schema::ActivationType_RELU:
+      act_op_->mutable_activation()->mutable_relu();
+      break;
+    case schema::ActivationType_RELU6: {
+      auto clip_param = act_op_->mutable_clip();
+      clip_param->set_minval(0);
+      clip_param->set_maxval(kValueThreshold6);
+      break;
+    }
+    case schema::ActivationType_TANH:
+      act_op_->mutable_activation()->mutable_tanh();
+      break;
+    case schema::ActivationType_SIGMOID:
+      act_op_->mutable_activation()->mutable_sigmoid();
+      break;
+    default:
+      MS_LOG(ERROR) << "Unsupported activation type.";
+      return RET_ERROR;
+  }
+  return RET_OK;
+}
+
+int CoreMLOp::SetPadding(std::vector<int> pad_list) {
+  pad_op_ = std::make_unique<CoreML::Specification::NeuralNetworkLayer>();
+  if (pad_op_ == nullptr) {
+    MS_LOG(ERROR) << "New CoreML op " << name_ << "_pad failed.";
+    return RET_ERROR;
+  }
+  pad_op_->set_name("CoreML_" + name_ + "_pad");
+  auto pad_param = pad_op_->mutable_padding();
+  pad_param->mutable_constant();
+  auto height_border = pad_param->mutable_paddingamounts()->add_borderamounts();
+  auto width_border = pad_param->mutable_paddingamounts()->add_borderamounts();
+  height_border->set_startedgesize(pad_list[PAD_UP]);
+  height_border->set_endedgesize(pad_list[PAD_DOWN]);
+  width_border->set_startedgesize(pad_list[PAD_LEFT]);
+  width_border->set_endedgesize(pad_list[PAD_RIGHT]);
+  return RET_OK;
+}
+
+int CoreMLOp::SetConstInput(const mindspore::MSTensor &in_tensor) {
+  MS_CHECK_TRUE_MSG(in_tensor.IsConst(), RET_ERROR, "Only constant tensor can be set as CoreML Const op.");
+  std::string const_op_name = "CoreML_" + in_tensor.Name() + "_const";
+  auto const_op = std::make_unique<CoreML::Specification::NeuralNetworkLayer>();
+  if (const_op == nullptr) {
+    MS_LOG(ERROR) << "New CoreML const op " << const_op_name << " for op " << name_ << " failed.";
+    return RET_ERROR;
+  }
+  const_op->set_name(const_op_name);
+  auto const_param = const_op->mutable_loadconstantnd();
+  for (auto i : in_tensor.Shape()) {
+    const_param->add_shape(static_cast<uint64_t>(i));
+  }
+  if (in_tensor.Shape().empty()) {
+    const_param->add_shape(1);
+  }
+  // set const data
+  auto org_data = in_tensor.Data().get();
+  auto *ml_data_container = const_param->mutable_data()->mutable_floatvalue();
+  ml_data_container->Resize(in_tensor.ElementNum(), 0);
+  auto *ml_data = reinterpret_cast<float *>(ml_data_container->mutable_data());
+  if (in_tensor.DataType() == DataType::kNumberTypeInt32) {
+    Int32ToFloat32(reinterpret_cast<const int *>(org_data), ml_data, in_tensor.ElementNum());
+  } else if (in_tensor.DataType() == DataType::kNumberTypeFloat32) {
+    memcpy(ml_data, org_data, in_tensor.DataSize());
+  } else {
+    MS_LOG(ERROR) << "Unsupported const input data type: " << static_cast<int>(in_tensor.DataType());
+    return RET_ERROR;
+  }
+  const_ops_[in_tensor.Name()] = std::move(const_op);
+  return RET_OK;
+}
+
+void CoreMLOp::SetMLOpInOut() {
+  MS_ASSERT(op_ != nullptr);
+  auto input_name = in_tensors_.at(0).Name();
+  if (pad_op_ != nullptr) {
+    std::string pad_name = op_->name() + "_pad_0";
+    pad_op_->add_input(input_name);
+    pad_op_->add_output(pad_name);
+    op_->add_input(pad_name);
+  } else {
+    op_->add_input(input_name);
+  }
+  auto output_name = out_tensors_.at(0).Name();
+  if (act_op_ != nullptr) {
+    std::string act_name = op_->name() + "_act_0";
+    op_->add_output(act_name);
+    act_op_->add_input(act_name);
+    act_op_->add_output(output_name);
+  } else {
+    op_->add_output(output_name);
+  }
+}
+
+std::vector<CoreML::Specification::NeuralNetworkLayer *> CoreMLOp::GetLayers() {
+  MS_ASSERT(op_ != nullptr);
+  std::vector<CoreML::Specification::NeuralNetworkLayer *> ret_ops;
+  if (pad_op_ != nullptr) {
+    ret_ops.push_back(pad_op_.release());
+  }
+  if (!const_ops_.empty()) {
+    for (auto it = const_ops_.begin(); it != const_ops_.end(); it++) {
+      ret_ops.push_back(it->second.release());
+    }
+  }
+  ret_ops.push_back(op_.release());
+  if (act_op_ != nullptr) {
+    ret_ops.push_back(act_op_.release());
+  }
+  return ret_ops;
+}
+}  // namespace mindspore

mindspore/lite/src/runtime/delegate/coreml/op/coreml_op.h

@@ -0,0 +1,155 @@
+/**
+ * Copyright 2022 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_RUNTIME_DELEGATE_COREML_OP_COREML_OP_
+#define MINDSPORE_LITE_SRC_RUNTIME_DELEGATE_COREML_OP_COREML_OP_
+#include <utility>
+#include <vector>
+#include <string>
+#include <set>
+#include <memory>
+#include <unordered_map>
+#include "proto/Model.pb.h"
+#include "proto/NeuralNetwork.pb.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 "src/common/log_util.h"
+#include "nnacl/op_base.h"
+using mindspore::lite::RET_ERROR;
+using mindspore::lite::RET_NOT_SUPPORT;
+using mindspore::lite::RET_OK;
+namespace mindspore {
+inline const std::vector<int> NHWC2NCHW_PERM = {0, 3, 1, 2};
+inline const std::vector<int> NCHW2NHWC_PERM = {0, 2, 3, 1};
+enum COREML_WEIGHT_SHAPE { ML_WT_COUT = 0, ML_WT_CIN = 1, ML_WT_H = 2, ML_WT_W = 3 };
+enum MSLITE_WEIGHT_SHAPE { MS_WT_COUT = 0, MS_WT_H = 1, MS_WT_W = 2, MS_WT_CIN = 3 };
+enum PAD { PAD_UP = 0, PAD_DOWN = 1, PAD_LEFT = 2, PAD_RIGHT = 3 };
+constexpr int REPEAT_TIMES2 = 2;
+class CoreMLOp {
+ public:
+  CoreMLOp(const schema::Primitive *primitive, std::vector<mindspore::MSTensor> in_tensors,
+           std::vector<mindspore::MSTensor> out_tensors, std::string name)
+      : op_primitive_(primitive),
+        in_tensors_(std::move(in_tensors)),
+        out_tensors_(std::move(out_tensors)),
+        name_(std::move(name)) {
+    if (primitive != nullptr) {
+      type_ = primitive->value_type();
+    }
+  }
+
+  // the op will be managed by coreml model, no need to manually deconstruct
+  virtual ~CoreMLOp() = default;
+
+  virtual int IsSupport() { return RET_OK; }
+
+  virtual int Init();
+
+  virtual int InitParams() { return RET_OK; }
+
+  virtual int HandleAxis() { return RET_OK; }
+
+  virtual int BuildLayer() { return RET_OK; }
+
+  // override this method if the op has tensor which does not need to add to graph，e.g.，const tensor.
+  virtual void SetMLOpInOut();
+
+  // Transfer the ownership of op to coreml model; Multiple layers are possible to be build for one op, thus using
+  // vector as return.
+  virtual std::vector<CoreML::Specification::NeuralNetworkLayer *> GetLayers();
+
+  virtual int SetActivation(schema::ActivationType act_type);
+
+  virtual int SetPadding(std::vector<int> pad_list);
+
+  virtual int SetConstInput(const mindspore::MSTensor &in_tensor);
+
+  void set_inputs(const std::vector<mindspore::MSTensor> &in_tensors) { this->in_tensors_ = in_tensors; }
+
+  void set_input(const mindspore::MSTensor &in_tensor, int index) {
+    MS_ASSERT(static_cast<size_t>(index) < in_tensors_.size());
+    this->in_tensors_[index] = in_tensor;
+  }
+
+  void set_outputs(const std::vector<mindspore::MSTensor> &out_tensors) { this->out_tensors_ = out_tensors; }
+
+  const std::vector<mindspore::MSTensor> &inputs() { return this->in_tensors_; }
+
+  const std::vector<mindspore::MSTensor> &outputs() { return this->out_tensors_; }
+
+  void set_in_ops(const std::vector<CoreMLOp *> &in_ops) { this->in_ops_ = in_ops; }
+
+  void set_out_ops(const std::vector<CoreMLOp *> &out_ops) { this->out_ops_ = out_ops; }
+
+  const std::vector<CoreMLOp *> &in_ops() const { return this->in_ops_; }
+
+  const std::vector<CoreMLOp *> &out_ops() const { return this->out_ops_; }
+
+  schema::PrimitiveType type() const { return type_; }
+
+  std::string name() const { return this->name_; }
+
+  void set_name(const std::string &name) { this->name_ = name; }
+
+ protected:
+  const schema::Primitive *op_primitive_ = nullptr;
+  std::vector<mindspore::MSTensor> in_tensors_;
+  std::vector<mindspore::MSTensor> out_tensors_;
+  std::vector<CoreMLOp *> in_ops_;
+  std::vector<CoreMLOp *> out_ops_;
+  schema::PrimitiveType type_ = schema::PrimitiveType_NONE;
+  std::string name_;
+  std::unique_ptr<CoreML::Specification::NeuralNetworkLayer> op_ = nullptr;
+  std::unique_ptr<CoreML::Specification::NeuralNetworkLayer> pad_op_ = nullptr;
+  std::unique_ptr<CoreML::Specification::NeuralNetworkLayer> act_op_ = nullptr;
+  std::unordered_map<std::string, std::unique_ptr<CoreML::Specification::NeuralNetworkLayer>> const_ops_ = {};
+};
+
+typedef CoreMLOp *(*CoreMLGetOp)(const schema::Primitive *primitive, const std::vector<mindspore::MSTensor> &in_tensors,
+                                 const std::vector<mindspore::MSTensor> &out_tensors, const std::string &name);
+
+template <class T>
+CoreMLOp *GetCoreMLOp(const schema::Primitive *primitive, const std::vector<mindspore::MSTensor> &in_tensors,
+                      const std::vector<mindspore::MSTensor> &out_tensors, const std::string &name) {
+  auto shape = out_tensors.front().Shape();
+  if (std::find(shape.begin(), shape.end(), -1) != shape.end()) {
+    MS_LOG(ERROR) << "CoreML does not support runtime inference shape.";
+    return nullptr;
+  }
+  auto *op = new (std::nothrow) T(primitive, in_tensors, out_tensors, name);
+  if (op == nullptr) {
+    MS_LOG(ERROR) << "op is nullptr.";
+    return nullptr;
+  }
+  auto ret = op->IsSupport();
+  if (ret != RET_OK) {
+    MS_LOG(WARNING) << "CoreML op is not supported.";
+    delete op;
+    return nullptr;
+  }
+  ret = op->Init();
+  if (ret != RET_OK) {
+    MS_LOG(WARNING) << "CoreML op init failed.";
+    delete op;
+    return nullptr;
+  }
+  return op;
+}
+}  // namespace mindspore
+#endif  // MINDSPORE_LITE_SRC_RUNTIME_DELEGATE_COREML_OP_COREML_OP_

mindspore/lite/src/runtime/delegate/coreml/op/deconvolution_coreml.cc

@@ -0,0 +1,70 @@
+/**
+ * Copyright 2022 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/runtime/delegate/coreml/op/deconvolution_coreml.h"
+#include "src/runtime/delegate/delegate_utils.h"
+namespace mindspore {
+int DeconvolutionCoreMLOp::IsSupport() {
+  if (!in_tensors_[kWeightIndex].IsConst()) {
+    MS_LOG(WARNING) << "CoreML deconvolution does not support dynamic weight.";
+    return RET_NOT_SUPPORT;
+  }
+  deconv_prim_ = op_primitive_->value_as_Conv2dTransposeFusion();
+  if (deconv_prim_ == nullptr) {
+    MS_LOG(ERROR) << "Get null primitive value for op ." << name_;
+    return RET_ERROR;
+  }
+  if (static_cast<int>(deconv_prim_->group()) != 1) {
+    MS_LOG(WARNING) << "Only support group equals 1 for npu deconvolution op";
+    return RET_NOT_SUPPORT;
+  }
+  return RET_OK;
+}
+
+int DeconvolutionCoreMLOp::SetConvParam() {
+  conv_param_->set_isdeconvolution(true);
+  CHECK_NULL_RETURN(deconv_prim_->stride());
+  auto stride_h = static_cast<int>(*(deconv_prim_->stride()->begin()));
+  auto stride_w = static_cast<int>(*(deconv_prim_->stride()->begin() + 1));
+  conv_param_->add_stride(stride_h);
+  conv_param_->add_stride(stride_w);
+  CHECK_NULL_RETURN(deconv_prim_->dilation());
+  auto dilation_h = static_cast<int>(*(deconv_prim_->dilation()->begin()));
+  auto dilation_w = static_cast<int>(*(deconv_prim_->dilation()->begin() + 1));
+  conv_param_->add_dilationfactor(dilation_h);
+  conv_param_->add_dilationfactor(dilation_w);
+  conv_param_->add_outputshape(output_h_);
+  conv_param_->add_outputshape(output_w_);
+  if (deconv_prim_->pad_mode() == schema::PadMode_SAME) {
+    conv_param_->mutable_same();
+  } else {
+    conv_param_->mutable_valid();
+    if (deconv_prim_->pad_list() != nullptr) {
+      auto pad_u = static_cast<int>(*(deconv_prim_->pad_list()->begin() + PAD_UP));
+      auto pad_d = static_cast<int>(*(deconv_prim_->pad_list()->begin() + PAD_DOWN));
+      auto pad_l = static_cast<int>(*(deconv_prim_->pad_list()->begin() + PAD_LEFT));
+      auto pad_r = static_cast<int>(*(deconv_prim_->pad_list()->begin() + PAD_RIGHT));
+      auto ret = SetPadding({pad_u, pad_d, pad_l, pad_r});
+      if (ret != RET_OK) {
+        MS_LOG(ERROR) << "Fail to set padding for op: " << name_;
+        return RET_ERROR;
+      }
+    }
+  }
+  act_type_ = deconv_prim_->activation_type();
+  return RET_OK;
+}
+}  // namespace mindspore

mindspore/lite/src/runtime/delegate/coreml/op/deconvolution_coreml.h

@@ -0,0 +1,42 @@
+/**
+ * Copyright 2022 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_RUNTIME_DELEGATE_COREML_OP_DECONVOLUTION_COREML_H_
+#define MINDSPORE_LITE_SRC_RUNTIME_DELEGATE_COREML_OP_DECONVOLUTION_COREML_H_
+
+#include <vector>
+#include <string>
+#include <utility>
+#include <unordered_map>
+#include "src/runtime/delegate/coreml/op/convolution_base_coreml.h"
+namespace mindspore {
+class DeconvolutionCoreMLOp : public ConvolutionBaseCoreMLOp {
+ public:
+  DeconvolutionCoreMLOp(const schema::Primitive *primitive, const std::vector<mindspore::MSTensor> &in_tensors,
+                        const std::vector<mindspore::MSTensor> &out_tensors, std::string name)
+      : ConvolutionBaseCoreMLOp(primitive, in_tensors, out_tensors, name) {}
+
+  int IsSupport() override;
+
+ private:
+  schema::PadMode GetPadMode();
+
+  int SetConvParam() override;
+
+ private:
+  const schema::Conv2dTransposeFusion *deconv_prim_ = nullptr;
+};
+}  // namespace mindspore
+#endif  // MINDSPORE_LITE_SRC_RUNTIME_DELEGATE_COREML_OP_DECONVOLUTION_COREML_H_

mindspore/lite/src/runtime/delegate/coreml/op/flatten_coreml.cc

@@ -0,0 +1,24 @@
+/**
+ * Copyright 2022 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/runtime/delegate/coreml/op/flatten_coreml.h"
+namespace mindspore {
+int FlattenCoreMLOp::BuildLayer() {
+  MS_ASSERT(op_ != nullptr);
+  (void)op_->mutable_flattento2d();
+  return RET_OK;
+}
+}  // namespace mindspore

mindspore/lite/src/runtime/delegate/coreml/op/flatten_coreml.h

@@ -0,0 +1,32 @@
+/**
+ * Copyright 2022 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_RUNTIME_DELEGATE_COREML_OP_FLATTEN_COREML_H_
+#define MINDSPORE_LITE_SRC_RUNTIME_DELEGATE_COREML_OP_FLATTEN_COREML_H_
+
+#include <vector>
+#include <string>
+#include "src/runtime/delegate/coreml/op/coreml_op.h"
+namespace mindspore {
+class FlattenCoreMLOp : public CoreMLOp {
+ public:
+  FlattenCoreMLOp(const schema::Primitive *primitive, const std::vector<mindspore::MSTensor> &in_tensors,
+                  const std::vector<mindspore::MSTensor> &out_tensors, std::string name)
+      : CoreMLOp(primitive, in_tensors, out_tensors, name) {}
+
+  int BuildLayer() override;
+};
+}  // namespace mindspore
+#endif  // MINDSPORE_LITE_SRC_RUNTIME_DELEGATE_COREML_OP_FLATTEN_COREML_H_

mindspore/lite/src/runtime/delegate/coreml/op/gather_coreml.cc

@@ -0,0 +1,51 @@
+/**
+ * Copyright 2022 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/runtime/delegate/coreml/op/gather_coreml.h"
+namespace mindspore {
+int GatherCoreMLOp::IsSupport() {
+  MS_CHECK_GE(in_tensors_.size(), kInputSize2, RET_NOT_SUPPORT);
+  return RET_OK;
+}
+
+int GatherCoreMLOp::BuildLayer() {
+  MS_ASSERT(op_ != nullptr);
+  auto gather_params = op_->mutable_gather();
+  CHECK_NULL_RETURN(in_tensors_[THIRD_INPUT].Data());
+  auto axis_data = reinterpret_cast<const int *>(in_tensors_[THIRD_INPUT].Data().get());
+  gather_params->set_axis(axis_data[0]);
+  auto indices_tensor = in_tensors_[SECOND_INPUT];
+  if (indices_tensor.IsConst()) {
+    auto ret = SetConstInput(indices_tensor);
+    if (ret != RET_OK) {
+      MS_LOG(ERROR) << "Set const input failed for op: " << name_;
+      return RET_ERROR;
+    }
+  }
+  return RET_OK;
+}
+
+void GatherCoreMLOp::SetMLOpInOut() {
+  MS_ASSERT(op_ != nullptr);
+  op_->add_input(in_tensors_[FIRST_INPUT].Name());
+  auto indices_tensor = in_tensors_[SECOND_INPUT];
+  if (indices_tensor.IsConst()) {
+    const_ops_[indices_tensor.Name()]->add_output(indices_tensor.Name());
+  }
+  op_->add_input(indices_tensor.Name());
+  op_->add_output(out_tensors_[0].Name());
+}
+}  // namespace mindspore

mindspore/lite/src/runtime/delegate/coreml/op/gather_coreml.h

@@ -0,0 +1,36 @@
+/**
+ * Copyright 2022 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_RUNTIME_DELEGATE_COREML_OP_GATHER_COREML_H_
+#define MINDSPORE_LITE_SRC_RUNTIME_DELEGATE_COREML_OP_GATHER_COREML_H_
+
+#include <vector>
+#include <string>
+#include "src/runtime/delegate/coreml/op/coreml_op.h"
+namespace mindspore {
+class GatherCoreMLOp : public CoreMLOp {
+ public:
+  GatherCoreMLOp(const schema::Primitive *primitive, const std::vector<mindspore::MSTensor> &in_tensors,
+                 const std::vector<mindspore::MSTensor> &out_tensors, std::string name)
+      : CoreMLOp(primitive, in_tensors, out_tensors, name) {}
+
+  int IsSupport() override;
+
+  int BuildLayer() override;
+
+  void SetMLOpInOut() override;
+};
+}  // namespace mindspore
+#endif  // MINDSPORE_LITE_SRC_RUNTIME_DELEGATE_COREML_OP_GATHER_COREML_H_

mindspore/lite/src/runtime/delegate/coreml/op/matmul_coreml.cc

@@ -0,0 +1,184 @@
+/**
+ * Copyright 2022 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/runtime/delegate/coreml/op/matmul_coreml.h"
+namespace mindspore {
+int MatMulCoreMLOp::IsSupport() {
+  MS_CHECK_GE(in_tensors_.size(), kInputSize1, RET_NOT_SUPPORT);
+  if (in_tensors_.size() > kInputSize1 && !in_tensors_.at(SECOND_INPUT).IsConst()) {
+    MS_LOG(WARNING) << "Bias for CoreML matmul is supported only when the second input is a constant.";
+    return RET_NOT_SUPPORT;
+  }
+  return RET_OK;
+}
+
+int MatMulCoreMLOp::InitParams() {
+  matmul_prim_ = op_primitive_->value_as_MatMulFusion();
+  if (matmul_prim_ == nullptr) {
+    MS_LOG(ERROR) << "Get null primitive value for op ." << name_;
+    return RET_ERROR;
+  }
+  return RET_OK;
+}
+
+int MatMulCoreMLOp::BuildLayer() {
+  MS_ASSERT(op_ != nullptr);
+  matmul_param_ = op_->mutable_batchedmatmul();
+  matmul_param_->set_transposea(matmul_prim_->transpose_a());
+  matmul_param_->set_transposeb(matmul_prim_->transpose_b());
+  if (in_tensors_.at(SECOND_INPUT).IsConst()) {
+    if (matmul_prim_->transpose_b()) {
+      auto ret = ConstMatMulWithTransB();
+      if (ret != RET_OK) {
+        MS_LOG(ERROR) << "Build MatMul layer with const input and true TransposeB failed for op: " << name_;
+        return RET_ERROR;
+      }
+    } else {
+      // CoreML will automatically transpose the const input even though transposeB is false.
+      auto ret = ConstMatMulWithoutTransB();
+      if (ret != RET_OK) {
+        MS_LOG(ERROR) << "Build MatMul layer with const input and false TransposeB failed for op: " << name_;
+        return RET_ERROR;
+      }
+    }
+  }
+  auto act_type = matmul_prim_->activation_type();
+  if (act_type != schema::ActivationType_NO_ACTIVATION) {
+    auto ret = SetActivation(act_type);
+    if (ret != RET_OK) {
+      MS_LOG(ERROR) << "Set matmul activation failed for op: " << name_;
+      return RET_ERROR;
+    }
+  }
+  return RET_OK;
+}
+
+int MatMulCoreMLOp::ConstMatMulWithTransB() {
+  MS_ASSERT(matmul_param_ != nullptr);
+  auto input_b = in_tensors_.at(SECOND_INPUT);
+  auto dim_b = input_b.Shape().size();
+  int64_t in_channel =
+    matmul_prim_->transpose_b() ? input_b.Shape()[dim_b - DIMENSION_1D] : input_b.Shape()[dim_b - DIMENSION_2D];
+  int64_t out_channel =
+    matmul_prim_->transpose_b() ? input_b.Shape()[dim_b - DIMENSION_2D] : input_b.Shape()[dim_b - DIMENSION_1D];
+  matmul_param_->set_weightmatrixfirstdimension(in_channel);
+  matmul_param_->set_weightmatrixseconddimension(out_channel);
+  auto org_weight = input_b.Data().get();
+  if (input_b.DataType() == DataType::kNumberTypeFloat32) {
+    auto *ml_weight_container = matmul_param_->mutable_weights()->mutable_floatvalue();
+    ml_weight_container->Resize(input_b.ElementNum(), 0);
+    auto *ml_weight = reinterpret_cast<void *>(ml_weight_container->mutable_data());
+    memcpy(ml_weight, org_weight, input_b.DataSize());
+  } else {
+    MS_LOG(ERROR) << "Unsupported data type of weight tensor for CoreML convolution.";
+    return RET_ERROR;
+  }
+  if (in_tensors_.size() > kInputSize1) {
+    auto bias_tensor = in_tensors_.at(THIRD_INPUT);
+    auto org_bias = bias_tensor.Data().get();
+    matmul_param_->set_hasbias(true);
+    if (bias_tensor.DataType() == DataType::kNumberTypeFloat32) {
+      auto *ml_bias_container = matmul_param_->mutable_bias()->mutable_floatvalue();
+      ml_bias_container->Resize(bias_tensor.ElementNum(), 0);
+      auto *ml_bias = reinterpret_cast<void *>(ml_bias_container->mutable_data());
+      memcpy(ml_bias, org_bias, bias_tensor.DataSize());
+    } else {
+      MS_LOG(ERROR) << "Unsupported data type of bias tensor for CoreML convolution.";
+      return RET_ERROR;
+    }
+  }
+  return RET_OK;
+}
+
+int MatMulCoreMLOp::ConstMatMulWithoutTransB() {
+  MS_ASSERT(matmul_param_ != nullptr);
+  auto ret = SetConstInput(in_tensors_[SECOND_INPUT]);
+  if (ret != RET_OK) {
+    MS_LOG(ERROR) << "Set const input failed for op: " << name_;
+    return RET_ERROR;
+  }
+  if (in_tensors_.size() > kInputSize1) {
+    // when the second input is not const anymore, the bias param will be invalid.
+    bias_op_ = std::make_unique<CoreML::Specification::NeuralNetworkLayer>();
+    if (bias_op_ == nullptr) {
+      MS_LOG(ERROR) << "New CoreML op " << name_ << "_bias failed.";
+      return RET_ERROR;
+    }
+    bias_op_->set_name("CoreML_" + name_ + "_bias");
+    (void)bias_op_->mutable_addbroadcastable();
+    ret = SetConstInput(in_tensors_[THIRD_INPUT]);
+    if (ret != RET_OK) {
+      MS_LOG(ERROR) << "Set const input failed for op: " << name_;
+      return RET_ERROR;
+    }
+  }
+  return RET_OK;
+}
+
+void MatMulCoreMLOp::SetMLOpInOut() {
+  MS_ASSERT(op_ != nullptr);
+  op_->add_input(in_tensors_.at(FIRST_INPUT).Name());
+  auto input_b_name = in_tensors_.at(SECOND_INPUT).Name();
+  auto output_name = out_tensors_.at(0).Name();
+  if (!in_tensors_.at(SECOND_INPUT).IsConst()) {
+    op_->add_input(input_b_name);
+  } else if (!const_ops_.empty()) {
+    const_ops_[input_b_name]->add_output(input_b_name);
+    op_->add_input(input_b_name);
+    if (bias_op_ != nullptr) {
+      std::string bias_name = op_->name() + "_bias_0";
+      op_->add_output(bias_name);
+      bias_op_->add_input(bias_name);
+      auto input_c_name = in_tensors_.at(THIRD_INPUT).Name();
+      const_ops_[input_c_name]->add_output(input_c_name);
+      bias_op_->add_input(input_c_name);
+    }
+  }
+  if (act_op_ != nullptr) {
+    std::string act_name = op_->name() + "_act_0";
+    if (bias_op_ != nullptr) {
+      bias_op_->add_output(act_name);
+    } else {
+      op_->add_output(act_name);
+    }
+    act_op_->add_input(act_name);
+    act_op_->add_output(output_name);
+    return;
+  }
+  if (bias_op_ != nullptr) {
+    bias_op_->add_output(output_name);
+  } else {
+    op_->add_output(output_name);
+  }
+  return;
+}
+
+std::vector<CoreML::Specification::NeuralNetworkLayer *> MatMulCoreMLOp::GetLayers() {
+  MS_ASSERT(op_ != nullptr);
+  std::vector<CoreML::Specification::NeuralNetworkLayer *> ret_ops;
+  for (auto it = const_ops_.begin(); it != const_ops_.end(); it++) {
+    ret_ops.push_back(it->second.release());
+  }
+  ret_ops.push_back(op_.release());
+  if (bias_op_ != nullptr) {
+    ret_ops.push_back(bias_op_.release());
+  }
+  if (act_op_ != nullptr) {
+    ret_ops.push_back(act_op_.release());
+  }
+  return ret_ops;
+}
+}  // namespace mindspore

mindspore/lite/src/runtime/delegate/coreml/op/matmul_coreml.h

@@ -0,0 +1,50 @@
+/**
+ * Copyright 2022 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_RUNTIME_DELEGATE_COREML_OP_MATMUL_COREML_H_
+#define MINDSPORE_LITE_SRC_RUNTIME_DELEGATE_COREML_OP_MATMUL_COREML_H_
+
+#include <vector>
+#include <string>
+#include <memory>
+#include "src/runtime/delegate/coreml/op/coreml_op.h"
+namespace mindspore {
+class MatMulCoreMLOp : public CoreMLOp {
+ public:
+  MatMulCoreMLOp(const schema::Primitive *primitive, const std::vector<mindspore::MSTensor> &in_tensors,
+                 const std::vector<mindspore::MSTensor> &out_tensors, std::string name)
+      : CoreMLOp(primitive, in_tensors, out_tensors, name) {}
+
+  int IsSupport() override;
+
+  int InitParams() override;
+
+  int BuildLayer() override;
+
+  std::vector<CoreML::Specification::NeuralNetworkLayer *> GetLayers() override;
+
+  void SetMLOpInOut() override;
+
+  int ConstMatMulWithTransB();
+
+  int ConstMatMulWithoutTransB();
+
+ protected:
+  const schema::MatMulFusion *matmul_prim_ = nullptr;
+  CoreML::Specification::BatchedMatMulLayerParams *matmul_param_ = nullptr;
+  std::unique_ptr<CoreML::Specification::NeuralNetworkLayer> bias_op_ = nullptr;
+};
+}  // namespace mindspore
+#endif  // MINDSPORE_LITE_SRC_RUNTIME_DELEGATE_COREML_OP_MATMUL_COREML_H_

mindspore/lite/src/runtime/delegate/coreml/op/max_pooling_coreml.cc

@@ -0,0 +1,71 @@
+/**
+ * Copyright 2022 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/runtime/delegate/coreml/op/max_pooling_coreml.h"
+namespace mindspore {
+int MaxPoolingCoreMLOp::InitParams() {
+  pooling_prim_ = op_primitive_->value_as_MaxPoolFusion();
+  if (pooling_prim_ == nullptr) {
+    MS_LOG(ERROR) << "Get null primitive value for op ." << name_;
+    return RET_ERROR;
+  }
+  return RET_OK;
+}
+
+int MaxPoolingCoreMLOp::BuildLayer() {
+  MS_ASSERT(op_ != nullptr);
+  auto pooling_param = op_->mutable_pooling();
+  pooling_param->set_type(CoreML::Specification::PoolingLayerParams::MAX);
+  if (pooling_prim_->global()) {
+    pooling_param->set_globalpooling(true);
+    pooling_param->mutable_valid();
+    return RET_OK;
+  }
+  auto kernel_h = static_cast<int>(*(pooling_prim_->kernel_size()->begin()));
+  auto kernel_w = static_cast<int>(*(pooling_prim_->kernel_size()->begin() + 1));
+  auto stride_h = static_cast<int>(*(pooling_prim_->strides()->begin()));
+  auto stride_w = static_cast<int>(*(pooling_prim_->strides()->begin() + 1));
+  pooling_param->add_stride(stride_h);
+  pooling_param->add_stride(stride_w);
+  pooling_param->add_kernelsize(kernel_h);
+  pooling_param->add_kernelsize(kernel_w);
+  if (pooling_prim_->pad_mode() == schema::PadMode_SAME) {
+    pooling_param->mutable_same();
+  } else {
+    pooling_param->mutable_valid();
+    if (pooling_prim_->pad() != nullptr) {
+      auto pad_u = static_cast<int>(*(pooling_prim_->pad()->begin() + PAD_UP));
+      auto pad_d = static_cast<int>(*(pooling_prim_->pad()->begin() + PAD_DOWN));
+      auto pad_l = static_cast<int>(*(pooling_prim_->pad()->begin() + PAD_LEFT));
+      auto pad_r = static_cast<int>(*(pooling_prim_->pad()->begin() + PAD_RIGHT));
+      auto ret = SetPadding({pad_u, pad_d, pad_l, pad_r});
+      if (ret != RET_OK) {
+        MS_LOG(ERROR) << "Fail to set padding for op: " << name_;
+        return RET_ERROR;
+      }
+    }
+  }
+  auto act_type = pooling_prim_->activation_type();
+  if (act_type != schema::ActivationType_NO_ACTIVATION) {
+    auto ret = SetActivation(act_type);
+    if (ret != RET_OK) {
+      MS_LOG(ERROR) << "Set pooling activation failed for op: " << name_;
+      return RET_ERROR;
+    }
+  }
+  return RET_OK;
+}
+}  // namespace mindspore

mindspore/lite/src/runtime/delegate/coreml/op/max_pooling_coreml.h

@@ -0,0 +1,39 @@
+/**
+ * Copyright 2022 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_RUNTIME_DELEGATE_COREML_OP_MAX_POOLING_COREML_H_
+#define MINDSPORE_LITE_SRC_RUNTIME_DELEGATE_COREML_OP_MAX_POOLING_COREML_H_
+
+#include <vector>
+#include <string>
+#include <utility>
+#include <unordered_map>
+#include "src/runtime/delegate/coreml/op/coreml_op.h"
+namespace mindspore {
+class MaxPoolingCoreMLOp : public CoreMLOp {
+ public:
+  MaxPoolingCoreMLOp(const schema::Primitive *primitive, const std::vector<mindspore::MSTensor> &in_tensors,
+                     const std::vector<mindspore::MSTensor> &out_tensors, std::string name)
+      : CoreMLOp(primitive, in_tensors, out_tensors, name) {}
+
+  int InitParams() override;
+
+  int BuildLayer() override;
+
+ private:
+  const schema::MaxPoolFusion *pooling_prim_ = nullptr;
+};
+}  // namespace mindspore
+#endif  // MINDSPORE_LITE_SRC_RUNTIME_DELEGATE_COREML_OP_MAX_POOLING_COREML_H_

mindspore/lite/src/runtime/delegate/coreml/op/reshape_coreml.cc

@@ -0,0 +1,48 @@
+/**
+ * Copyright 2022 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 shapeissions and
+ * limitations under the License.
+ */
+
+#include "src/runtime/delegate/coreml/op/reshape_coreml.h"
+namespace mindspore {
+int ReshapeCoreMLOp::IsSupport() {
+  MS_CHECK_GE(in_tensors_.size(), kInputSize1, RET_NOT_SUPPORT);
+  return RET_OK;
+}
+
+int ReshapeCoreMLOp::BuildLayer() {
+  MS_ASSERT(op_ != nullptr);
+  auto shape_tensor = in_tensors_.at(1);
+  if (shape_tensor.IsConst()) {
+    auto shape_dim = shape_tensor.ElementNum();
+    auto shape_data = reinterpret_cast<const int *>(shape_tensor.Data().get());
+    auto shape_param = op_->mutable_reshapestatic();
+    for (int i = 0; i < shape_dim; i++) {
+      shape_param->add_targetshape(shape_data[i]);
+    }
+  } else {
+    op_->mutable_reshapedynamic();
+  }
+  return RET_OK;
+}
+
+void ReshapeCoreMLOp::SetMLOpInOut() {
+  MS_ASSERT(op_ != nullptr);
+  op_->add_input(in_tensors_[0].Name());
+  if (!in_tensors_[1].IsConst()) {
+    op_->add_input(in_tensors_[1].Name());
+  }
+  op_->add_output(out_tensors_[0].Name());
+}
+}  // namespace mindspore

mindspore/lite/src/runtime/delegate/coreml/op/reshape_coreml.h

@@ -0,0 +1,36 @@
+/**
+ * Copyright 2022 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_RUNTIME_DELEGATE_COREML_OP_RESHAPE_COREML_H_
+#define MINDSPORE_LITE_SRC_RUNTIME_DELEGATE_COREML_OP_RESHAPE_COREML_H_
+
+#include <vector>
+#include <string>
+#include "src/runtime/delegate/coreml/op/coreml_op.h"
+namespace mindspore {
+class ReshapeCoreMLOp : public CoreMLOp {
+ public:
+  ReshapeCoreMLOp(const schema::Primitive *primitive, const std::vector<mindspore::MSTensor> &in_tensors,
+                  const std::vector<mindspore::MSTensor> &out_tensors, std::string name)
+      : CoreMLOp(primitive, in_tensors, out_tensors, name) {}
+
+  int IsSupport() override;
+
+  int BuildLayer() override;
+
+  void SetMLOpInOut() override;
+};
+}  // namespace mindspore
+#endif  // MINDSPORE_LITE_SRC_RUNTIME_DELEGATE_COREML_OP_RESHAPE_COREML_H_

mindspore/lite/src/runtime/delegate/coreml/op/resize_coreml.cc

@@ -0,0 +1,83 @@
+/**
+ * Copyright 2022 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/runtime/delegate/coreml/op/resize_coreml.h"
+namespace mindspore {
+int ResizeCoreMLOp::IsSupport() {
+  resize_prim_ = op_primitive_->value_as_Resize();
+  if (resize_prim_ == nullptr) {
+    MS_LOG(ERROR) << "Get null primitive value for op ." << name_;
+    return RET_ERROR;
+  }
+  auto resize_method = resize_prim_->method();
+  if (resize_method != schema::ResizeMethod_LINEAR && resize_method != schema::ResizeMethod_NEAREST) {
+    MS_LOG(WARNING) << "Unsupported resize method type: " << resize_method;
+    return RET_NOT_SUPPORT;
+  }
+  if (resize_method != schema::ResizeMethod_LINEAR ||
+      resize_prim_->coordinate_transform_mode() != schema::CoordinateTransformMode_ALIGN_CORNERS) {
+    use_upsample_ = true;
+    if (in_tensors_.size() != kInputSize1 || !in_tensors_[1].IsConst() || in_tensors_[1].ElementNum() != C2NUM) {
+      MS_LOG(WARNING) << "The second input must be a constant with two scale values of height and width when using "
+                         "CoreML upsample layer for op: "
+                      << name_;
+      return RET_NOT_SUPPORT;
+    }
+  }
+  return RET_OK;
+}
+
+int ResizeCoreMLOp::BuildLayer() {
+  MS_ASSERT(op_ != nullptr);
+  if (use_upsample_) {
+    auto resize_param = op_->mutable_upsample();
+    MS_CHECK_GE(in_tensors_.size(), kInputSize1, RET_NOT_SUPPORT);
+    auto scale_tensor = in_tensors_.at(1);
+    auto scale_data = scale_tensor.Data().get();
+    if (scale_tensor.DataType() == DataType::kNumberTypeInt32) {
+      resize_param->add_scalingfactor(static_cast<const int *>(scale_data)[0]);
+      resize_param->add_scalingfactor(static_cast<const int *>(scale_data)[1]);
+    } else if (scale_tensor.DataType() == DataType::kNumberTypeFloat32) {
+      resize_param->add_fractionalscalingfactor(static_cast<const float *>(scale_data)[0]);
+      resize_param->add_fractionalscalingfactor(static_cast<const float *>(scale_data)[1]);
+    } else {
+      MS_LOG(ERROR) << "Unsupported Resize scale data type: " << static_cast<int>(scale_tensor.DataType());
+      return RET_ERROR;
+    }
+    if (resize_prim_->method() == schema::ResizeMethod_LINEAR) {
+      resize_param->set_mode(CoreML::Specification::UpsampleLayerParams_InterpolationMode_BILINEAR);
+      if (resize_prim_->coordinate_transform_mode() == schema::CoordinateTransformMode_ALIGN_CORNERS) {
+        resize_param->set_linearupsamplemode(
+          CoreML::Specification::UpsampleLayerParams_LinearUpsampleMode_ALIGN_CORNERS_TRUE);
+      } else {
+        resize_param->set_linearupsamplemode(
+          CoreML::Specification::UpsampleLayerParams_LinearUpsampleMode_ALIGN_CORNERS_FALSE);
+      }
+    } else if (resize_prim_->method() == schema::ResizeMethod_NEAREST) {
+      resize_param->set_mode(CoreML::Specification::UpsampleLayerParams_InterpolationMode_NN);
+    }
+    return RET_OK;
+  }
+  // Using resize_bilinear op. The op executed with NCHW format.
+  auto out_height = static_cast<int>(out_tensors_.at(0).Shape().at(kNCHW_H));
+  auto out_width = static_cast<int>(out_tensors_.at(0).Shape().at(kNCHW_W));
+  auto resize_param = op_->mutable_resizebilinear();
+  resize_param->add_targetsize(out_height);
+  resize_param->add_targetsize(out_width);
+  resize_param->mutable_mode()->set_samplingmethod(CoreML::Specification::SamplingMode::STRICT_ALIGN_ENDPOINTS_MODE);
+  return RET_OK;
+}
+}  // namespace mindspore

mindspore/lite/src/runtime/delegate/coreml/op/resize_coreml.h

@@ -0,0 +1,38 @@
+/**
+ * Copyright 2022 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_RUNTIME_DELEGATE_COREML_OP_RESIZE_COREML_H_
+#define MINDSPORE_LITE_SRC_RUNTIME_DELEGATE_COREML_OP_RESIZE_COREML_H_
+
+#include <vector>
+#include <string>
+#include "src/runtime/delegate/coreml/op/coreml_op.h"
+namespace mindspore {
+class ResizeCoreMLOp : public CoreMLOp {
+ public:
+  ResizeCoreMLOp(const schema::Primitive *primitive, const std::vector<mindspore::MSTensor> &in_tensors,
+                 const std::vector<mindspore::MSTensor> &out_tensors, std::string name)
+      : CoreMLOp(primitive, in_tensors, out_tensors, name) {}
+
+  int IsSupport() override;
+
+  int BuildLayer() override;
+
+ protected:
+  const schema::Resize *resize_prim_ = nullptr;
+  bool use_upsample_ = false;
+};
+}  // namespace mindspore
+#endif  // MINDSPORE_LITE_SRC_RUNTIME_DELEGATE_COREML_OP_RESIZE_COREML_H_

mindspore/lite/src/runtime/delegate/coreml/op/shape_coreml.cc

@@ -0,0 +1,24 @@
+/**
+ * Copyright 2022 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/runtime/delegate/coreml/op/shape_coreml.h"
+namespace mindspore {
+int ShapeCoreMLOp::BuildLayer() {
+  MS_ASSERT(op_ != nullptr);
+  (void)op_->mutable_getshape();
+  return RET_OK;
+}
+}  // namespace mindspore

mindspore/lite/src/runtime/delegate/coreml/op/shape_coreml.h

@@ -0,0 +1,32 @@
+/**
+ * Copyright 2022 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_RUNTIME_DELEGATE_COREML_OP_SHAPE_COREML_H_
+#define MINDSPORE_LITE_SRC_RUNTIME_DELEGATE_COREML_OP_SHAPE_COREML_H_
+
+#include <vector>
+#include <string>
+#include "src/runtime/delegate/coreml/op/coreml_op.h"
+namespace mindspore {
+class ShapeCoreMLOp : public CoreMLOp {
+ public:
+  ShapeCoreMLOp(const schema::Primitive *primitive, const std::vector<mindspore::MSTensor> &in_tensors,
+                const std::vector<mindspore::MSTensor> &out_tensors, std::string name)
+      : CoreMLOp(primitive, in_tensors, out_tensors, name) {}
+
+  int BuildLayer() override;
+};
+}  // namespace mindspore
+#endif  // MINDSPORE_LITE_SRC_RUNTIME_DELEGATE_COREML_OP_SHAPE_COREML_H_

mindspore/lite/src/runtime/delegate/coreml/op/softmax_coreml.cc

@@ -0,0 +1,41 @@
+/**
+ * Copyright 2022 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/runtime/delegate/coreml/op/softmax_coreml.h"
+namespace mindspore {
+int SoftmaxCoreMLOp::InitParams() {
+  softmax_prim_ = op_primitive_->value_as_Softmax();
+  if (softmax_prim_ == nullptr) {
+    MS_LOG(ERROR) << "Get null primitive value for op ." << name_;
+    return RET_ERROR;
+  }
+  MS_CHECK_TRUE_MSG(softmax_prim_->axis() != nullptr, RET_ERROR, "Softmax axis is null!");
+  axis_ = static_cast<int>(*(softmax_prim_->axis()->begin()));
+  return RET_OK;
+}
+
+int SoftmaxCoreMLOp::BuildLayer() {
+  MS_ASSERT(op_ != nullptr && softmax_prim_ != nullptr);
+  auto softmax_param = op_->mutable_softmaxnd();
+  softmax_param->set_axis(axis_);
+  return RET_OK;
+}
+
+int SoftmaxCoreMLOp::HandleAxis() {
+  axis_ = NCHW2NHWC_PERM[axis_];
+  return RET_OK;
+}
+}  // namespace mindspore

mindspore/lite/src/runtime/delegate/coreml/op/softmax_coreml.h

@@ -0,0 +1,40 @@
+/**
+ * Copyright 2022 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_RUNTIME_DELEGATE_COREML_OP_SOFTMAX_COREML_H_
+#define MINDSPORE_LITE_SRC_RUNTIME_DELEGATE_COREML_OP_SOFTMAX_COREML_H_
+
+#include <vector>
+#include <string>
+#include "src/runtime/delegate/coreml/op/coreml_op.h"
+namespace mindspore {
+class SoftmaxCoreMLOp : public CoreMLOp {
+ public:
+  SoftmaxCoreMLOp(const schema::Primitive *primitive, const std::vector<mindspore::MSTensor> &in_tensors,
+                  const std::vector<mindspore::MSTensor> &out_tensors, std::string name)
+      : CoreMLOp(primitive, in_tensors, out_tensors, name) {}
+
+  int InitParams() override;
+
+  int BuildLayer() override;
+
+  int HandleAxis() override;
+
+ private:
+  const schema::Softmax *softmax_prim_;
+  int axis_;
+};
+}  // namespace mindspore
+#endif  // MINDSPORE_LITE_SRC_RUNTIME_DELEGATE_COREML_OP_SOFTMAX_COREML_H_

mindspore/lite/src/runtime/delegate/coreml/op/transpose_coreml.cc

@@ -0,0 +1,37 @@
+/**
+ * Copyright 2022 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/runtime/delegate/coreml/op/transpose_coreml.h"
+namespace mindspore {
+int TransposeCoreMLOp::IsSupport() {
+  MS_CHECK_GE(in_tensors_.size(), kInputSize1, RET_NOT_SUPPORT);
+  auto perm_tensor = in_tensors_.at(1);
+  if (!perm_tensor.IsConst()) {
+    MS_LOG(WARNING) << "CoreML transpose must get fixed axis values.";
+    return RET_NOT_SUPPORT;
+  }
+  return RET_OK;
+}
+
+int TransposeCoreMLOp::BuildLayer() {
+  MS_ASSERT(op_ != nullptr);
+  auto transpose_param = op_->mutable_transpose();
+  for (auto perm : perm_) {
+    transpose_param->add_axes(perm);
+  }
+  return RET_OK;
+}
+}  // namespace mindspore

mindspore/lite/src/runtime/delegate/coreml/op/transpose_coreml.h

@@ -0,0 +1,54 @@
+/**
+ * Copyright 2022 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_RUNTIME_DELEGATE_COREML_OP_TRANSPOSE_COREML_H_
+#define MINDSPORE_LITE_SRC_RUNTIME_DELEGATE_COREML_OP_TRANSPOSE_COREML_H_
+
+#include <vector>
+#include <string>
+#include "src/runtime/delegate/coreml/op/coreml_op.h"
+namespace mindspore {
+class TransposeCoreMLOp : public CoreMLOp {
+ public:
+  TransposeCoreMLOp(const schema::Primitive *primitive, const std::vector<mindspore::MSTensor> &in_tensors,
+                    const std::vector<mindspore::MSTensor> &out_tensors, std::string name)
+      : CoreMLOp(primitive, in_tensors, out_tensors, name) {
+    MS_ASSERT(in_tensors.size() == kInputSize1);
+    auto perm_tensor = in_tensors.at(1);
+    auto perm_num = perm_tensor.ElementNum();
+    auto perm_data = reinterpret_cast<const int *>(perm_tensor.Data().get());
+    for (size_t i = 0; i < perm_num; i++) {
+      perm_.push_back(perm_data[i]);
+    }
+  }
+
+  TransposeCoreMLOp(const std::vector<mindspore::MSTensor> &in_tensors,
+                    const std::vector<mindspore::MSTensor> &out_tensors, std::vector<int> perm, std::string name)
+      : CoreMLOp(nullptr, in_tensors, out_tensors, name) {
+    perm_ = perm;
+    type_ = schema::PrimitiveType_Transpose;
+  }
+
+  int IsSupport() override;
+
+  int BuildLayer() override;
+
+  std::vector<int> GetPerm() { return perm_; }
+
+ protected:
+  std::vector<int> perm_;
+};
+}  // namespace mindspore
+#endif  // MINDSPORE_LITE_SRC_RUNTIME_DELEGATE_COREML_OP_TRANSPOSE_COREML_H_

mindspore/lite/src/runtime/delegate/coreml/op/unsqueeze_coreml.cc

@@ -0,0 +1,38 @@
+/**
+ * Copyright 2022 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/runtime/delegate/coreml/op/unsqueeze_coreml.h"
+namespace mindspore {
+int UnsqueezeCoreMLOp::InitParams() {
+  unsqueeze_prim_ = op_primitive_->value_as_Unsqueeze();
+  if (unsqueeze_prim_ == nullptr) {
+    MS_LOG(ERROR) << "Get null primitive value for op ." << name_;
+    return RET_ERROR;
+  }
+  return RET_OK;
+}
+
+int UnsqueezeCoreMLOp::BuildLayer() {
+  MS_ASSERT(op_ != nullptr && unsqueeze_prim_ != nullptr);
+  auto expanddims_param = op_->mutable_expanddims();
+  MS_CHECK_TRUE_MSG(unsqueeze_prim_->axis() != nullptr, RET_ERROR, "Unsqueeze axis is null!");
+  auto axes = std::vector<int>(unsqueeze_prim_->axis()->begin(), unsqueeze_prim_->axis()->end());
+  for (auto axis : axes) {
+    expanddims_param->add_axes(axis);
+  }
+  return RET_OK;
+}
+}  // namespace mindspore

mindspore/lite/src/runtime/delegate/coreml/op/unsqueeze_coreml.h

@@ -0,0 +1,37 @@
+/**
+ * Copyright 2022 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_RUNTIME_DELEGATE_COREML_OP_UNSQUEEZE_COREML_H_
+#define MINDSPORE_LITE_SRC_RUNTIME_DELEGATE_COREML_OP_UNSQUEEZE_COREML_H_
+
+#include <vector>
+#include <string>
+#include "src/runtime/delegate/coreml/op/coreml_op.h"
+namespace mindspore {
+class UnsqueezeCoreMLOp : public CoreMLOp {
+ public:
+  UnsqueezeCoreMLOp(const schema::Primitive *primitive, const std::vector<mindspore::MSTensor> &in_tensors,
+                    const std::vector<mindspore::MSTensor> &out_tensors, std::string name)
+      : CoreMLOp(primitive, in_tensors, out_tensors, name) {}
+
+  int InitParams() override;
+
+  int BuildLayer() override;
+
+ private:
+  const schema::Unsqueeze *unsqueeze_prim_;
+};
+}  // namespace mindspore
+#endif  // MINDSPORE_LITE_SRC_RUNTIME_DELEGATE_COREML_OP_UNSQUEEZE_COREML_H_

mindspore/lite/src/runtime/delegate/coreml/pass/coreml_base_pass.h

@@ -0,0 +1,37 @@
+/**
+ * Copyright 2022 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_RUNTIME_DELEGATE_COREML_PASS_COREML_BASE_PASS_H_
+#define MINDSPORE_LITE_SRC_RUNTIME_DELEGATE_COREML_PASS_COREML_BASE_PASS_H_
+#include <string>
+#include "src/runtime/delegate/coreml/coreml_graph.h"
+
+namespace mindspore {
+class CoreMLBasePass {
+ public:
+  virtual int Run(CoreMLGraph *subgraph) = 0;
+
+  virtual ~CoreMLBasePass() = default;
+
+  std::string name() { return name_; }
+
+ protected:
+  std::string name_;
+  CoreMLGraph *subgraph_ = nullptr;
+};
+}  // namespace mindspore
+
+#endif  // MINDSPORE_LITE_SRC_RUNTIME_DELEGATE_COREML_PASS_COREML_BASE_PASS_H_

mindspore/lite/src/runtime/delegate/coreml/pass/coreml_format_trans_pass.cc

@@ -0,0 +1,211 @@
+/**
+ * Copyright 2022 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/runtime/delegate/coreml/pass/coreml_format_trans_pass.h"
+#include <vector>
+#include "src/runtime/delegate/coreml/pass/coreml_pass_utils.h"
+
+using mindspore::lite::RET_ERROR;
+using mindspore::lite::RET_OK;
+
+namespace mindspore {
+std::set<mindspore::schema::PrimitiveType> nchw_nodes = {
+  schema::PrimitiveType_Conv2DFusion,  schema::PrimitiveType_Conv2dTransposeFusion, schema::PrimitiveType_Resize,
+  schema::PrimitiveType_MaxPoolFusion, schema::PrimitiveType_AvgPoolFusion,         schema::PrimitiveType_ScaleFusion,
+  schema::PrimitiveType_CropAndResize, schema::PrimitiveType_InstanceNorm};
+
+int CoreMLFormatTransPass::InsertPreNodes(CoreMLOp *op, std::vector<CoreMLOp *> *trans_ops) {
+  bool is_input_op = op->in_ops().empty();
+  // not always single input (like CropAndResize), but we care about the input with 4d.
+  auto it = std::find_if(op->in_ops().begin(), op->in_ops().end(), [](CoreMLOp *k) {
+    return k->outputs().size() > 0 && k->outputs()[0].Shape().size() == COMM_SHAPE_SIZE;
+  });
+  if (!is_input_op && it == op->in_ops().end()) {
+    MS_LOG(ERROR) << "CoreML Transform pass does not find in op with 4d output";
+    return RET_ERROR;
+  }
+  if (is_input_op || nchw_nodes.find((*it)->type()) == nchw_nodes.end()) {
+    CoreMLOp *pre_op = nullptr;
+    if (!is_input_op) {
+      pre_op = *it;
+    }
+
+    // Create pre transform op's out tensor.
+    auto name = op->name() + "_pre_trans" + "_Nhwc2Nchw_" + std::to_string(total++);
+    auto nhwc_shape = op->inputs()[0].Shape();
+    std::vector<int64_t> nchw_shape = {nhwc_shape[kNHWC_N], nhwc_shape[kNHWC_C], nhwc_shape[kNHWC_H],
+                                       nhwc_shape[kNHWC_W]};
+    auto tensor =
+      mindspore::MSTensor::CreateTensor(name + "/output0", op->inputs()[0].DataType(), nchw_shape, nullptr, 0);
+    if (tensor == nullptr) {
+      MS_LOG(ERROR) << "New nchw tensor failed when inserting pre nhwc2nchw op.";
+      return RET_ERROR;
+    }
+    tensor->SetFormat(Format::NCHW);
+    std::vector<mindspore::MSTensor> pre_trans_outputs = {*tensor};
+    all_tensors_->push_back(tensor);
+
+    // Create pre transform op: Nhwc2Nchw
+    auto *trans_op = CoreMLPassUtils::CreateNhwc2NchwOp({op->inputs()[0]}, pre_trans_outputs, name);
+    if (trans_op == nullptr) {
+      MS_LOG(ERROR) << "Create Nhwc2Nchw transpose op failed.";
+      return RET_ERROR;
+    }
+    trans_ops->push_back(trans_op);
+
+    // Set in_ops, out_ops, inputs, outputs for transform op
+    std::vector<CoreMLOp *> pre_trans_in_ops;
+    if (!is_input_op) {
+      pre_trans_in_ops = {pre_op};
+    }
+    CoreMLPassUtils::UpdateOp(trans_op, pre_trans_in_ops, {op}, trans_op->inputs(), pre_trans_outputs);
+
+    if (pre_op != nullptr) {
+      CoreMLPassUtils::UpdateNH2NCTransNodePreOp(pre_op, trans_op, op);
+    }
+    CoreMLPassUtils::UpdateNH2NCTransNodePostOp(trans_op, op);
+  }
+  return RET_OK;
+}
+
+int CoreMLFormatTransPass::InsertPostNodes(CoreMLOp *op, std::vector<CoreMLOp *> *trans_ops) {
+  bool is_output_op = false;
+  if (op->out_ops().empty() ||
+      find(subgraph_->outputs().begin(), subgraph_->outputs().end(), op->outputs()[0]) != subgraph_->outputs().end()) {
+    is_output_op = true;
+  }
+  // Get the post op that need insert trans op.
+  // If no need for inserting trans op, the post op must be coreml and in trans_nodes.
+  std::vector<CoreMLOp *> post_insert_ops;
+  std::vector<CoreMLOp *> post_non_insert_ops;
+  for (int i = 0; i < op->out_ops().size(); i++) {
+    auto post_op = op->out_ops()[i];
+    if (nchw_nodes.find(post_op->type()) == nchw_nodes.end()) {
+      post_insert_ops.push_back(post_op);
+    } else {
+      post_non_insert_ops.push_back(post_op);
+    }
+  }
+  if (!is_output_op && post_insert_ops.empty()) {
+    return RET_OK;
+  }
+  // Create post transform op's in tensor.
+  auto name = op->name() + "_post_trans" + "_Nchw2Nhwc" + std::to_string(total++);
+
+  auto nhwc_shape = op->outputs()[0].Shape();
+  std::vector<int64_t> nchw_shape = {nhwc_shape[kNHWC_N], nhwc_shape[kNHWC_C], nhwc_shape[kNHWC_H],
+                                     nhwc_shape[kNHWC_W]};
+  auto nc2nh_tensor =
+    mindspore::MSTensor::CreateTensor(name + "/input0", op->outputs()[0].DataType(), nchw_shape, nullptr, 0);
+  if (nc2nh_tensor == nullptr) {
+    MS_LOG(ERROR) << "New nchw tensor failed when inserting post nchw2nhwc op.";
+    return RET_ERROR;
+  }
+  nc2nh_tensor->SetFormat(Format::NCHW);
+  all_tensors_->push_back(nc2nh_tensor);
+
+  if (is_output_op) {
+    std::vector<mindspore::MSTensor> nc2nh_outputs{op->outputs().at(0)};
+    // Create post transform op: Nchw2Nhwc
+    auto *post_trans_op = CoreMLPassUtils::CreateNchw2NhwcOp({*nc2nh_tensor}, nc2nh_outputs, name);
+    if (post_trans_op == nullptr) {
+      MS_LOG(ERROR) << "Create Nchw2Nhwc transpose op failed.";
+      return RET_ERROR;
+    }
+    // Set in_ops, out_ops, inputs, outputs for transform op
+    CoreMLPassUtils::UpdateOp(post_trans_op, {op}, {}, post_trans_op->inputs(), post_trans_op->outputs());
+    trans_ops->push_back(post_trans_op);
+  }
+  // for each to-be-insert out op, create one transpose op, one perm tensor, one out tensor
+  // but using same one in_tensor.
+  for (auto i = 0; i < post_insert_ops.size(); ++i) {
+    auto post_insert_op = post_insert_ops.at(i);
+    // nc2nh op out tensor: abandon original out_tensor, all ops use newly created out tensor.
+    std::vector<mindspore::MSTensor> nc2nh_outputs{};
+    auto origin_out_tensor = op->outputs().at(0);
+    auto out_tensor_name = op->name() + "_post_trans" + "_Nchw2Nhwc_" + std::to_string(i) + "_out_tensor";
+    auto out_tensor = mindspore::MSTensor::CreateTensor(out_tensor_name, origin_out_tensor.DataType(),
+                                                        origin_out_tensor.Shape(), nullptr, 0);
+    if (out_tensor == nullptr) {
+      MS_LOG(ERROR) << "New nhwc tensor failed when inserting post nchw2nhwc op.";
+      return RET_ERROR;
+    }
+    out_tensor->SetFormat(Format::NHWC);
+    all_tensors_->push_back(out_tensor);
+    nc2nh_outputs.push_back(*out_tensor);
+
+    // Create post transform op: Nchw2Nhwc
+    auto *post_trans_op =
+      CoreMLPassUtils::CreateNchw2NhwcOp({*nc2nh_tensor}, nc2nh_outputs, name + "_" + std::to_string(i));
+    if (post_trans_op == nullptr) {
+      MS_LOG(ERROR) << "Create Nchw2Nhwc transpose op failed.";
+      return RET_ERROR;
+    }
+    // Set in_ops, out_ops, inputs, outputs for transform op
+    CoreMLPassUtils::UpdateOp(post_trans_op, {op}, {post_insert_op}, post_trans_op->inputs(), post_trans_op->outputs());
+    trans_ops->push_back(post_trans_op);
+    // update post op inputs in_ops
+    CoreMLPassUtils::UpdateNC2NHTransNodePostOp(op, post_trans_op, post_insert_op, origin_out_tensor);
+  }
+  // for those non-insert post ops, update their in_tensor
+  for (auto non_insert_op : post_non_insert_ops) {
+    auto inputs = non_insert_op->inputs();
+    std::replace(inputs.begin(), inputs.end(), op->outputs().at(0), *nc2nh_tensor);
+    non_insert_op->set_inputs(inputs);
+  }
+  // update origin op's out tensor and out op
+  CoreMLPassUtils::UpdateNC2NHTransNodePreOp(op, *trans_ops, post_insert_ops);
+  return RET_OK;
+}
+
+int CoreMLFormatTransPass::Run(CoreMLGraph *subgraph) {
+  subgraph_ = subgraph;
+  all_ops_ = subgraph_->GetOps();
+  all_tensors_ = subgraph_->GetInsertTensors();
+  for (size_t i = 0; i < all_ops_->size();) {
+    auto op = (*all_ops_)[i];
+    if (nchw_nodes.find(op->type()) == nchw_nodes.end()) {
+      i++;
+      continue;
+    }
+    if (op->type() == schema::PrimitiveType_InstanceNorm && op->inputs().front().format() == mindspore::Format::NCHW) {
+      i++;
+      continue;
+    }
+    // insert pre_ops before op in vector
+    // modify loop index add (pre_ops.size() + 1) to the post_ops insert location
+    std::vector<CoreMLOp *> pre_ops;
+    auto ret = InsertPreNodes(op, &pre_ops);
+    if (ret != RET_OK) {
+      MS_LOG(ERROR) << "Insert nhwc2nchw op before op " << op->name() << " failed.";
+      return RET_ERROR;
+    }
+    all_ops_->insert(all_ops_->begin() + i, pre_ops.begin(), pre_ops.end());
+    i += (pre_ops.size() + 1);
+
+    // insert post_ops after op in vector
+    // modify loop index add post_ops.size() to the next op in the origin vector
+    std::vector<CoreMLOp *> post_ops;
+    ret = InsertPostNodes(op, &post_ops);
+    if (ret != RET_OK) {
+      MS_LOG(ERROR) << "Insert nchw2nhwc op after op " << op->name() << " failed.";
+      return RET_ERROR;
+    }
+    all_ops_->insert(all_ops_->begin() + i, post_ops.begin(), post_ops.end());
+    i += post_ops.size();
+  }
+  return RET_OK;
+}
+}  // namespace mindspore

mindspore/lite/src/runtime/delegate/coreml/pass/coreml_format_trans_pass.h

@@ -0,0 +1,43 @@
+/**
+ * Copyright 2022 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_RUNTIME_DELEGATE_COREML_PASS_COREML_FORMAT_TRANS_PASS_H_
+#define MINDSPORE_LITE_SRC_RUNTIME_DELEGATE_COREML_PASS_COREML_FORMAT_TRANS_PASS_H_
+
+#include <set>
+#include <vector>
+#include "src/runtime/delegate/coreml/op/coreml_op.h"
+#include "src/runtime/delegate/coreml/pass/coreml_base_pass.h"
+
+namespace mindspore {
+class CoreMLFormatTransPass : public CoreMLBasePass {
+ public:
+  CoreMLFormatTransPass() { name_ = "CoreMLFormatTransPass"; }
+
+  int Run(CoreMLGraph *subgraph) override;
+
+ private:
+  int InsertPreNodes(CoreMLOp *op, std::vector<CoreMLOp *> *trans_ops);
+
+  int InsertPostNodes(CoreMLOp *op, std::vector<CoreMLOp *> *trans_ops);
+
+ private:
+  int total = 0;
+  std::vector<CoreMLOp *> *all_ops_ = nullptr;
+  std::vector<mindspore::MSTensor *> *all_tensors_ = nullptr;
+};
+}  // namespace mindspore
+#endif  // MINDSPORE_LITE_SRC_RUNTIME_DELEGATE_COREML_PASS_COREML_FORMAT_TRANS_PASS_H_

mindspore/lite/src/runtime/delegate/coreml/pass/coreml_fusion_pass.cc

@@ -0,0 +1,384 @@
+/**
+ * Copyright 2022 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/runtime/delegate/coreml/pass/coreml_fusion_pass.h"
+#include <vector>
+#include "src/runtime/delegate/coreml/pass/coreml_pass_utils.h"
+
+using mindspore::lite::RET_ERROR;
+using mindspore::lite::RET_OK;
+
+namespace mindspore {
+bool CheckFusion(CoreMLOp *cur_op, const std::vector<mindspore::MSTensor> &graph_outputs) {
+  if (cur_op->in_ops().empty() || cur_op->out_ops().empty()) {
+    return false;
+  }
+  auto pre_flag = std::all_of(cur_op->in_ops().begin(), cur_op->in_ops().end(), [](CoreMLOp *in_op) {
+    return CoreMLPassUtils::IsNchw2Nhwc(in_op) && in_op->out_ops().size() == 1;
+  });
+  if (!pre_flag) {
+    return false;
+  }
+  auto post_flag = std::all_of(cur_op->out_ops().begin(), cur_op->out_ops().end(),
+                               [](CoreMLOp *out_op) { return CoreMLPassUtils::IsNhwc2Nchw(out_op); });
+  if (!post_flag) {
+    return false;
+  }
+  for (auto out_op : cur_op->out_ops()) {
+    // If the pattern is "nc2nh->cur_op->nh2nc" while the output tensors of "cur_op" and "nh2nc" are both graph output,
+    // the trans ops can not be fused since it will cause the missing of graph output.
+    if (out_op->out_ops().empty() &&
+        std::find(graph_outputs.begin(), graph_outputs.end(), out_op->inputs().at(0)) != graph_outputs.end()) {
+      return false;
+    }
+  }
+  return true;
+}
+
+bool CheckFormatFusion(CoreMLOp *cur_op) {
+  if (cur_op->out_ops().empty()) {
+    return false;
+  }
+  if (CoreMLPassUtils::IsNhwc2Nchw(cur_op)) {
+    return std::all_of(cur_op->out_ops().begin(), cur_op->out_ops().end(),
+                       [](CoreMLOp *op) { return CoreMLPassUtils::IsNchw2Nhwc(op); });
+  }
+  if (CoreMLPassUtils::IsNchw2Nhwc(cur_op)) {
+    return std::all_of(cur_op->out_ops().begin(), cur_op->out_ops().end(),
+                       [](CoreMLOp *op) { return CoreMLPassUtils::IsNhwc2Nchw(op); });
+  }
+  return false;
+}
+
+void CoreMLFusionPass::RemoveAndFreeOp(CoreMLOp *cur_op) {
+  auto itr = find(all_ops_->begin(), all_ops_->end(), cur_op);
+  if (itr != all_ops_->end()) {
+    all_ops_->erase(itr);
+  }
+  delete cur_op;
+}
+
+int CoreMLFusionPass::UpdatePreOps(CoreMLOp *cur_op) {
+  auto cur_in_ops = cur_op->in_ops();
+  for (auto in_op : cur_op->in_ops()) {
+    // graph in op
+    if (in_op->in_ops().empty()) {
+      cur_in_ops.erase(find(cur_in_ops.begin(), cur_in_ops.end(), in_op));
+    } else {
+      auto pre_op = in_op->in_ops()[0];
+      auto pre_out_ops = pre_op->out_ops();
+      for (size_t i = 0; i < pre_out_ops.size(); i++) {
+        if (pre_out_ops[i] == in_op) {
+          pre_out_ops[i] = cur_op;
+          break;
+        }
+      }
+      pre_op->set_out_ops(pre_out_ops);
+
+      for (size_t i = 0; i < cur_in_ops.size(); i++) {
+        if (cur_in_ops[i] == in_op) {
+          cur_in_ops[i] = pre_op;
+          break;
+        }
+      }
+    }
+    RemoveAndFreeOp(in_op);
+  }
+  cur_op->set_in_ops(cur_in_ops);
+  return RET_OK;
+}
+
+int CoreMLFusionPass::UpdatePostOps(CoreMLOp *cur_op) {
+  auto cur_out_ops = cur_op->out_ops();
+  for (auto out_op : cur_op->out_ops()) {
+    // graph out op
+    if (out_op->out_ops().empty()) {
+      cur_out_ops.erase(find(cur_out_ops.begin(), cur_out_ops.end(), out_op));
+    } else {
+      auto post_op = out_op->out_ops()[0];
+      auto post_in_ops = post_op->in_ops();
+      for (size_t i = 0; i < post_in_ops.size(); i++) {
+        if (post_in_ops[i] == out_op) {
+          post_in_ops[i] = cur_op;
+          break;
+        }
+      }
+      post_op->set_in_ops(post_in_ops);
+
+      for (size_t i = 0; i < cur_out_ops.size(); i++) {
+        if (cur_out_ops[i] == out_op) {
+          cur_out_ops[i] = post_op;
+          break;
+        }
+      }
+    }
+    RemoveAndFreeOp(out_op);
+  }
+  cur_op->set_out_ops(cur_out_ops);
+  return RET_OK;
+}
+
+int UpdatePreTensors(CoreMLOp *cur_op) {
+  auto in_tensors_vec = cur_op->inputs();
+  for (auto in_op : cur_op->in_ops()) {
+    if (in_op->inputs().empty() || in_op->outputs().empty()) {
+      MS_LOG(ERROR) << "in_tensors or out_tensors of input op is empty.";
+      return RET_ERROR;
+    }
+    mindspore::MSTensor cur_tensor;
+    auto in_tensor = in_op->inputs()[0];
+    auto out_tensor = in_op->outputs()[0];
+    if (!in_op->in_ops().empty()) {
+      auto pre_op = in_op->in_ops()[0];
+      for (size_t i = 0; i < pre_op->outputs().size(); i++) {
+        if (pre_op->outputs()[i] == in_tensor) {
+          cur_tensor = pre_op->outputs()[i];
+          break;
+        }
+      }
+    } else {
+      // graph input
+      cur_tensor = in_tensor;
+    }
+
+    for (size_t i = 0; i < in_tensors_vec.size(); i++) {
+      if (in_tensors_vec[i] == out_tensor) {
+        in_tensors_vec[i] = cur_tensor;
+      }
+    }
+  }
+  cur_op->set_inputs(in_tensors_vec);
+  return RET_OK;
+}
+
+int UpdatePostTensors(CoreMLOp *cur_op) {
+  mindspore::MSTensor new_post_input;
+  for (auto out_op : cur_op->out_ops()) {
+    auto in_tensor = out_op->inputs()[0];
+    auto out_tensor = out_op->outputs()[0];
+    auto nhwc_shape = in_tensor.Shape();
+    if (in_tensor.format() == Format::NHWC) {
+      MS_CHECK_TRUE_MSG(nhwc_shape.size() == COMM_SHAPE_SIZE, RET_ERROR, "Invalid transpose dim size!");
+      in_tensor.SetShape({nhwc_shape[kNHWC_N], nhwc_shape[kNHWC_C], nhwc_shape[kNHWC_H], nhwc_shape[kNHWC_W]});
+      in_tensor.SetFormat(Format::NCHW);
+    }
+    // out_op is a graph output op
+    if (out_op->out_ops().empty()) {
+      auto out_tensors_vec = cur_op->outputs();
+      for (size_t i = 0; i < out_tensors_vec.size(); i++) {
+        if (out_tensors_vec[i] == in_tensor) {
+          out_tensors_vec[i] = out_op->outputs()[0];
+        }
+      }
+      cur_op->set_outputs(out_tensors_vec);
+      // exist other out_ops using the same tensor as the current out_op, note that the other out_op has likely been
+      // updated, which mean it may be not a Transpose op anymore.
+      for (auto other_out_op : cur_op->out_ops()) {
+        auto other_in_tensors_vec = other_out_op->inputs();
+        for (size_t i = 0; i < other_in_tensors_vec.size(); i++) {
+          if (other_in_tensors_vec[i] == in_tensor) {
+            other_in_tensors_vec[i] = out_op->outputs()[0];
+          }
+        }
+        other_out_op->set_inputs(other_in_tensors_vec);
+      }
+    }
+    // out_op is not a graph out op
+    for (auto post_op : out_op->out_ops()) {
+      auto in_tensors_vec = post_op->inputs();
+      for (size_t i = 0; i < in_tensors_vec.size(); i++) {
+        if (in_tensors_vec[i] == out_tensor) {
+          in_tensors_vec[i] = in_tensor;
+        }
+      }
+      post_op->set_inputs(in_tensors_vec);
+    }
+  }
+  return RET_OK;
+}
+
+int CoreMLFusionPass::UpdateOp(CoreMLOp *cur_op) {
+  if (cur_op == nullptr) {
+    MS_LOG(ERROR) << "kernel is nullptr.";
+    return RET_ERROR;
+  }
+  auto ret = UpdatePreTensors(cur_op);
+  if (ret != RET_OK) {
+    MS_LOG(ERROR) << "UpdatePreTensors failed.";
+    return RET_ERROR;
+  }
+  ret = UpdatePostTensors(cur_op);
+  if (ret != RET_OK) {
+    MS_LOG(ERROR) << "UpdatePostTensors failed.";
+    return RET_ERROR;
+  }
+  ret = UpdatePreOps(cur_op);
+  if (ret != RET_OK) {
+    MS_LOG(ERROR) << "UpdatePreOps failed.";
+    return RET_ERROR;
+  }
+  ret = UpdatePostOps(cur_op);
+  if (ret != RET_OK) {
+    MS_LOG(ERROR) << "UpdatePostOps failed.";
+    return RET_ERROR;
+  }
+  return RET_OK;
+}
+
+int CoreMLFusionPass::CommonFusion(CoreMLOp *cur_op) {
+  if (cur_op == nullptr) {
+    return RET_ERROR;
+  }
+  auto ret = UpdateOp(cur_op);
+  if (ret != RET_OK) {
+    MS_LOG(ERROR) << "UpdateOp failed.";
+    return RET_ERROR;
+  }
+  ret = cur_op->HandleAxis();
+  if (ret != RET_OK) {
+    MS_LOG(ERROR) << "HandleAxis failed.";
+    return ret;
+  }
+  return RET_OK;
+}
+
+void UpdateOutOpsOfPreOp(CoreMLOp *cur_op, bool found_graph_out_tensor, const mindspore::MSTensor &graph_out_tensor,
+                         const std::vector<CoreMLOp *> &pre_insert_ops) {
+  MS_ASSERT(cur_op != nullptr);
+  auto is_graph_input = cur_op->in_ops().empty();
+  auto cur_op_in_tensor = cur_op->inputs()[0];
+  if (!is_graph_input) {
+    auto pre_op = cur_op->in_ops()[0];
+    auto pre_out_ops = pre_op->out_ops();
+    size_t cur_op_index = 0;
+    for (size_t index = 0; index < pre_out_ops.size(); index++) {
+      if (pre_out_ops[index] == cur_op) {
+        pre_out_ops.erase(pre_out_ops.begin() + index);
+        cur_op_index = index;
+        index--;
+      } else if (found_graph_out_tensor) {
+        // only in this case, the output of pre_op is specified to 2nd trans op's output and pre_out_ops need update.
+        auto tensors_vec = pre_out_ops[index]->inputs();
+        for (size_t i = 0; i < tensors_vec.size(); i++) {
+          if (tensors_vec[i] == cur_op_in_tensor) {
+            tensors_vec[i] = graph_out_tensor;
+            break;
+          }
+        }
+        pre_out_ops[index]->set_inputs(tensors_vec);
+      }
+    }
+    pre_out_ops.insert(pre_out_ops.begin() + cur_op_index, pre_insert_ops.begin(), pre_insert_ops.end());
+    pre_op->set_out_ops(pre_out_ops);
+  }
+  return;
+}
+
+int CoreMLFusionPass::FormatFusion(CoreMLOp *cur_op) {
+  CHECK_NULL_RETURN(cur_op);
+  auto is_graph_input = cur_op->in_ops().empty();
+  auto cur_op_in_tensor = cur_op->inputs()[0];
+  std::vector<CoreMLOp *> pre_insert_ops;
+  CoreMLOp *pre_op = nullptr;
+  if (!is_graph_input) {
+    pre_op = cur_op->in_ops()[0];
+  }
+  mindspore::MSTensor graph_out_tensor;
+  bool found_graph_out_tensor = false;
+  auto graph_outputs = subgraph_->outputs();
+  // if the output of second trans op(s) is graph output, find it out and use it as the pre-op's output.
+  for (const auto &sec_op : cur_op->out_ops()) {
+    if (std::find(graph_outputs.begin(), graph_outputs.end(), sec_op->outputs()[0]) != graph_outputs.end()) {
+      graph_out_tensor = sec_op->outputs()[0];
+      if (!is_graph_input) {
+        found_graph_out_tensor = true;
+        // cur_op is the first trans op, it's input op num and input tensor num must be 1
+        pre_op->set_outputs({graph_out_tensor});
+        // in fp16 mode, tensor data type fp16 need to be changed back.
+        auto tensor = pre_op->outputs()[0];
+        if (tensor.DataType() == DataType::kNumberTypeFloat16) {
+          tensor.SetDataType(DataType::kNumberTypeFloat32);
+        }
+        break;
+      } else {
+        MS_LOG(WARNING) << "Existing graph output equivalent to graph input, which is unsupported now.";
+        return RET_OK;
+      }
+    }
+  }
+  for (const auto &trans_op : cur_op->out_ops()) {
+    for (const auto &post_op : trans_op->out_ops()) {
+      // update tensor
+      auto tensors_vec = post_op->inputs();
+      for (size_t i = 0; i < tensors_vec.size(); i++) {
+        if (tensors_vec[i] == trans_op->outputs()[0]) {
+          tensors_vec[i] = found_graph_out_tensor ? graph_out_tensor : cur_op_in_tensor;
+          break;
+        }
+      }
+      post_op->set_inputs(tensors_vec);
+
+      // update op
+      auto post_in_ops = post_op->in_ops();
+      for (size_t i = 0; i < post_in_ops.size(); i++) {
+        if (post_in_ops[i] == trans_op) {
+          if (is_graph_input) {
+            post_in_ops.erase(post_in_ops.begin() + i);
+          } else {
+            post_in_ops[i] = pre_op;
+          }
+          break;
+        }
+      }
+      post_op->set_in_ops(post_in_ops);
+      pre_insert_ops.push_back(post_op);
+    }
+    RemoveAndFreeOp(trans_op);
+  }
+  UpdateOutOpsOfPreOp(cur_op, found_graph_out_tensor, graph_out_tensor, pre_insert_ops);
+  RemoveAndFreeOp(cur_op);
+  return RET_OK;
+}
+
+int CoreMLFusionPass::Run(CoreMLGraph *subgraph) {
+  subgraph_ = subgraph;
+  all_ops_ = subgraph->GetOps();
+  for (size_t i = 0; i < all_ops_->size(); i++) {
+    auto cur_op = (*all_ops_)[i];
+    auto ret = RET_OK;
+    if (CheckFusion(cur_op, subgraph->outputs())) {
+      i -= cur_op->in_ops().size();
+      ret = CommonFusion(cur_op);
+    }
+    if (ret != RET_OK) {
+      MS_LOG(ERROR) << "Fusion failed.";
+      return RET_ERROR;
+    }
+  }
+  for (size_t i = 0; i < all_ops_->size(); ++i) {
+    auto cur_op = (*all_ops_)[i];
+    if (CheckFormatFusion(cur_op)) {
+      i--;
+      auto ret = FormatFusion(cur_op);
+      if (ret != RET_OK) {
+        MS_LOG(ERROR) << "FormatFusion failed.";
+        return RET_ERROR;
+      }
+    }
+  }
+  return RET_OK;
+}
+}  // namespace mindspore

mindspore/lite/src/runtime/delegate/coreml/pass/coreml_fusion_pass.h

@@ -0,0 +1,42 @@
+/**
+ * Copyright 2022 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_RUNTIME_DELEGATE_COREML_PASS_COREML_FUSION_PASS_H_
+#define MINDSPORE_LITE_SRC_RUNTIME_DELEGATE_COREML_PASS_COREML_FUSION_PASS_H_
+#include <vector>
+#include "src/runtime/delegate/coreml/op/coreml_op.h"
+#include "src/runtime/delegate/coreml/pass/coreml_base_pass.h"
+
+namespace mindspore {
+class CoreMLFusionPass : public CoreMLBasePass {
+ public:
+  CoreMLFusionPass() { name_ = "CoreMLFusionPass"; }
+
+  int Run(CoreMLGraph *subgraph) override;
+
+ protected:
+  int UpdatePreOps(CoreMLOp *cur_op);
+  int UpdatePostOps(CoreMLOp *cur_op);
+  void RemoveAndFreeOp(CoreMLOp *cur_op);
+  int UpdateOp(CoreMLOp *cur_op);
+  int CommonFusion(CoreMLOp *cur_op);
+  int FormatFusion(CoreMLOp *cur_op);
+
+ private:
+  std::vector<CoreMLOp *> *all_ops_;
+};
+}  // namespace mindspore
+#endif  // MINDSPORE_LITE_SRC_RUNTIME_DELEGATE_COREML_PASS_COREML_FUSION_PASS_H_

mindspore/lite/src/runtime/delegate/coreml/pass/coreml_pass_manager.cc

@@ -0,0 +1,44 @@
+/**
+ * Copyright 2022 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/runtime/delegate/coreml/pass/coreml_pass_manager.h"
+#include "include/errorcode.h"
+#include "src/common/log_adapter.h"
+
+using mindspore::lite::RET_ERROR;
+using mindspore::lite::RET_OK;
+
+namespace mindspore {
+void CoreMLPassManager::AddPass(CoreMLBasePass *pass) { all_pass_.push_back(pass); }
+
+int CoreMLPassManager::RunPass(CoreMLGraph *subgraph) {
+  for (auto pass : all_pass_) {
+    auto ret = pass->Run(subgraph);
+    if (ret != RET_OK) {
+      MS_LOG(ERROR) << "CoreML Pass Run failed. Pass name is:" << pass->name() << " for subgraph " << subgraph->name();
+      return ret;
+    }
+  }
+  return RET_OK;
+}
+
+void CoreMLPassManager::Clear() {
+  for (auto pass : all_pass_) {
+    delete pass;
+  }
+  all_pass_.clear();
+}
+}  // namespace mindspore

mindspore/lite/src/runtime/delegate/coreml/pass/coreml_pass_manager.h

@@ -0,0 +1,41 @@
+/**
+ * Copyright 2022 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_RUNTIME_DELEGATE_COREML_PASS_COREML_PASS_MANAGER_H_
+#define MINDSPORE_LITE_SRC_RUNTIME_DELEGATE_COREML_PASS_COREML_PASS_MANAGER_H_
+#include <vector>
+#include "src/runtime/delegate/coreml/pass/coreml_base_pass.h"
+namespace mindspore {
+class CoreMLPassManager {
+ public:
+  static CoreMLPassManager *GetInstance() {
+    static CoreMLPassManager pass_manager;
+    return &pass_manager;
+  }
+
+  ~CoreMLPassManager() { Clear(); }
+
+  void AddPass(CoreMLBasePass *pass);
+
+  int RunPass(CoreMLGraph *subgraph);
+
+  void Clear();
+
+ private:
+  std::vector<CoreMLBasePass *> all_pass_{};
+};
+}  // namespace mindspore
+#endif  // MINDSPORE_LITE_SRC_RUNTIME_DELEGATE_COREML_PASS_COREML_PASS_MANAGER_H_

mindspore/lite/src/runtime/delegate/coreml/pass/coreml_pass_utils.cc

@@ -0,0 +1,178 @@
+/**
+ * Copyright 2022 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/runtime/delegate/coreml/pass/coreml_pass_utils.h"
+#include <algorithm>
+#include "src/runtime/delegate/coreml/op/transpose_coreml.h"
+
+namespace mindspore {
+CoreMLOp *CoreMLPassUtils::CreateNchw2NhwcOp(const std::vector<mindspore::MSTensor> &in_tensors,
+                                             const std::vector<mindspore::MSTensor> &out_tensors,
+                                             const std::string &name) {
+  auto trans_op = new (std::nothrow) TransposeCoreMLOp(in_tensors, out_tensors, NCHW2NHWC_PERM, name);
+  if (trans_op == nullptr) {
+    MS_LOG(ERROR) << "New Nchw2Nhwc CoreMLOp failed.";
+    return nullptr;
+  }
+  auto ret = trans_op->Init();
+  if (ret != RET_OK) {
+    MS_LOG(ERROR) << "Nchw2Nhwc transpose op init failed.";
+    return nullptr;
+  }
+  return trans_op;
+}
+
+CoreMLOp *CoreMLPassUtils::CreateNhwc2NchwOp(const std::vector<mindspore::MSTensor> &in_tensors,
+                                             const std::vector<mindspore::MSTensor> &out_tensors,
+                                             const std::string &name) {
+  auto trans_op = new (std::nothrow) TransposeCoreMLOp(in_tensors, out_tensors, NHWC2NCHW_PERM, name);
+  if (trans_op == nullptr) {
+    MS_LOG(ERROR) << "New Nhwc2Nchw CoreMLOp failed.";
+    return nullptr;
+  }
+  auto ret = trans_op->Init();
+  if (ret != RET_OK) {
+    MS_LOG(ERROR) << "Nhwc2Nchw transpose op init failed.";
+    return nullptr;
+  }
+  return trans_op;
+}
+
+void CoreMLPassUtils::UpdateOp(CoreMLOp *op, const std::vector<CoreMLOp *> &in_ops,
+                               const std::vector<CoreMLOp *> &out_ops,
+                               const std::vector<mindspore::MSTensor> &in_tensors,
+                               const std::vector<mindspore::MSTensor> &outputs) {
+  op->set_inputs(in_tensors);
+  op->set_outputs(outputs);
+  op->set_in_ops(in_ops);
+  op->set_out_ops(out_ops);
+}
+
+void CoreMLPassUtils::UpdateNH2NCTransNodePreOp(CoreMLOp *pre_op, CoreMLOp *trans_op, CoreMLOp *op) {
+  // For op before trans, update the out_ops; the output tensor of op is the input tensor of trans, no need to update.
+  std::vector<CoreMLOp *> out_ops = pre_op->out_ops();
+  if (op == nullptr) {
+    out_ops.emplace_back(trans_op);
+  } else {
+    for (size_t i = 0; i < out_ops.size(); i++) {
+      if (out_ops[i] == op) {
+        out_ops[i] = trans_op;
+        break;
+      }
+    }
+  }
+  pre_op->set_out_ops(out_ops);
+}
+
+void CoreMLPassUtils::UpdateNC2NHTransNodePreOp(CoreMLOp *pre_op, const std::vector<CoreMLOp *> &trans_ops,
+                                                const std::vector<CoreMLOp *> &ops) {
+  // For op before trans, there may be multiple outputs.
+  auto cur_out_ops = pre_op->out_ops();
+  for (size_t i = 0; i < ops.size(); i++) {
+    auto itr = find(cur_out_ops.begin(), cur_out_ops.end(), ops[i]);
+    if (itr != cur_out_ops.end()) {
+      cur_out_ops.erase(itr);
+    }
+  }
+  std::copy(trans_ops.begin(), trans_ops.end(), std::back_inserter(cur_out_ops));
+  pre_op->set_out_ops(cur_out_ops);
+  // For op before trans, the output tensor is used for output tensor of trans, so replace the output tensor
+  // with the input tensor of trans.
+  pre_op->set_outputs({trans_ops.at(0)->inputs().at(0)});
+}
+
+void CoreMLPassUtils::UpdateNH2NCTransNodePostOp(CoreMLOp *trans_op, CoreMLOp *post_op) {
+  auto cur_in_tensors = post_op->inputs();
+  cur_in_tensors[0] = trans_op->outputs()[0];
+  post_op->set_inputs(cur_in_tensors);
+  post_op->set_in_ops({trans_op});
+}
+
+void CoreMLPassUtils::UpdateNC2NHTransNodePostOp(CoreMLOp *op, CoreMLOp *trans_op, CoreMLOp *post_op,
+                                                 const mindspore::MSTensor &org_in_tensor) {
+  // The input tensor should be replaced with the output tensor of trans_op.
+  auto post_in_tensors = post_op->inputs();
+  std::replace(post_in_tensors.begin(), post_in_tensors.end(), org_in_tensor, trans_op->outputs().at(0));
+  post_op->set_inputs(post_in_tensors);
+
+  // For post_op after trans, op in in_ops should be replaced with trans_op.
+  auto post_in_ops = post_op->in_ops();
+  if (op == nullptr) {
+    post_in_ops.push_back(trans_op);
+  } else {
+    std::replace(post_in_ops.begin(), post_in_ops.end(), op, trans_op);
+  }
+  post_op->set_in_ops(post_in_ops);
+}
+
+bool CoreMLPassUtils::IsNhwc2Nchw(CoreMLOp *op) {
+  if (op == nullptr) {
+    return false;
+  }
+  if (op->type() != schema::PrimitiveType_Transpose) {
+    return false;
+  }
+  auto transpose_op = static_cast<TransposeCoreMLOp *>(op);
+  std::vector<int> perm = transpose_op->GetPerm();
+  std::vector<int> nh2nc_perm = {0, 3, 1, 2};
+  if (perm != nh2nc_perm) {
+    return false;
+  }
+  return true;
+}
+
+bool CoreMLPassUtils::IsNchw2Nhwc(CoreMLOp *op) {
+  if (op == nullptr) {
+    return false;
+  }
+  if (op->type() != schema::PrimitiveType_Transpose) {
+    return false;
+  }
+  auto transpose_op = static_cast<TransposeCoreMLOp *>(op);
+  std::vector<int> perm = transpose_op->GetPerm();
+  std::vector<int> nc2nh_perm = {0, 2, 3, 1};
+  if (perm != nc2nh_perm) {
+    return false;
+  }
+  return true;
+}
+
+CoreMLOp *CoreMLPassUtils::OpInputFromOp(CoreMLOp *op, mindspore::MSTensor in_tensor) {
+  // given op and input tensor index, get which op output this tensor.
+  // If input tensor is graph input, return nullptr.
+  if (op == nullptr) {
+    return nullptr;
+  }
+  auto in_ops = op->in_ops();
+  auto output_contain = [in_tensor](CoreMLOp *in_op) {
+    auto outputs = in_op->outputs();
+    return std::find(outputs.begin(), outputs.end(), in_tensor) != outputs.end();
+  };
+  auto it = std::find_if(in_ops.begin(), in_ops.end(), output_contain);
+  if (it == in_ops.end()) {
+    return nullptr;
+  }
+  return *it;
+}
+
+std::vector<mindspore::MSTensor> CoreMLPassUtils::GetNonConstInputs(CoreMLOp *op) {
+  MS_CHECK_TRUE_MSG(op != nullptr, {}, "Input op is null!");
+  std::vector<mindspore::MSTensor> non_const_in_tensors;
+  std::copy_if(op->inputs().begin(), op->inputs().end(), std::back_inserter(non_const_in_tensors),
+               [](const auto &tensor) { return !tensor.IsConst(); });
+  return non_const_in_tensors;
+}
+}  // namespace mindspore

mindspore/lite/src/runtime/delegate/coreml/pass/coreml_pass_utils.h

@@ -0,0 +1,56 @@
+/**
+ * Copyright 2022 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_RUNTIME_DELEGATE_COREML_PASS_COREML_PASS_UTILS_H_
+#define MINDSPORE_LITE_SRC_RUNTIME_DELEGATE_COREML_PASS_COREML_PASS_UTILS_H_
+#include <vector>
+#include <set>
+#include <string>
+#include <unordered_map>
+#include "src/runtime/delegate/coreml/op/coreml_op.h"
+#include "src/runtime/delegate/coreml/op/transpose_coreml.h"
+
+namespace mindspore {
+class CoreMLPassUtils {
+ public:
+  static CoreMLOp *CreateNchw2NhwcOp(const std::vector<mindspore::MSTensor> &in_tensors,
+                                     const std::vector<mindspore::MSTensor> &out_tensors, const std::string &name);
+
+  static CoreMLOp *CreateNhwc2NchwOp(const std::vector<mindspore::MSTensor> &in_tensors,
+                                     const std::vector<mindspore::MSTensor> &out_tensors, const std::string &name);
+
+  static void UpdateOp(CoreMLOp *op, const std::vector<CoreMLOp *> &in_ops, const std::vector<CoreMLOp *> &out_ops,
+                       const std::vector<mindspore::MSTensor> &in_tensors,
+                       const std::vector<mindspore::MSTensor> &out_tensors);
+
+  static void UpdateNH2NCTransNodePreOp(CoreMLOp *pre_op, CoreMLOp *trans_op, CoreMLOp *op);
+
+  static void UpdateNC2NHTransNodePreOp(CoreMLOp *pre_op, const std::vector<CoreMLOp *> &trans_ops,
+                                        const std::vector<CoreMLOp *> &ops);
+
+  static void UpdateNH2NCTransNodePostOp(CoreMLOp *trans_op, CoreMLOp *post_op);
+
+  static void UpdateNC2NHTransNodePostOp(CoreMLOp *op, CoreMLOp *trans_op, CoreMLOp *post_op,
+                                         const mindspore::MSTensor &org_in_tensor);
+
+  static bool IsNhwc2Nchw(CoreMLOp *op);
+
+  static bool IsNchw2Nhwc(CoreMLOp *op);
+  static CoreMLOp *OpInputFromOp(CoreMLOp *op, mindspore::MSTensor in_tensor);
+  static std::vector<mindspore::MSTensor> GetNonConstInputs(CoreMLOp *op);
+};
+}  // namespace mindspore
+#endif  // MINDSPORE_LITE_SRC_RUNTIME_DELEGATE_COREML_PASS_COREML_PASS_UTILS_H_

mindspore/lite/src/runtime/delegate/coreml/pass/coreml_trans_extend_pass.cc

@@ -0,0 +1,316 @@
+/**
+ * Copyright 2020-2022 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/runtime/delegate/coreml/pass/coreml_trans_extend_pass.h"
+#include <algorithm>
+#include <set>
+#include <string>
+#include "src/runtime/delegate/coreml/pass/coreml_pass_utils.h"
+
+using mindspore::lite::RET_ERROR;
+using mindspore::lite::RET_OK;
+
+namespace mindspore {
+std::set<mindspore::schema::PrimitiveType> format_depend_nodes = {
+  schema::PrimitiveType_Conv2DFusion,  schema::PrimitiveType_Conv2dTransposeFusion,
+  schema::PrimitiveType_MaxPoolFusion, schema::PrimitiveType_AvgPoolFusion,
+  schema::PrimitiveType_CropAndResize, schema::PrimitiveType_InstanceNorm,
+  schema::PrimitiveType_ArgMaxFusion,  schema::PrimitiveType_FullConnection,
+  schema::PrimitiveType_ScaleFusion,   schema::PrimitiveType_ExpandDims,
+  schema::PrimitiveType_Unsqueeze,     schema::PrimitiveType_SliceFusion,
+  schema::PrimitiveType_BroadcastTo,   schema::PrimitiveType_TileFusion,
+  schema::PrimitiveType_Resize,        schema::PrimitiveType_MatMulFusion,
+  schema::PrimitiveType_Gather,        schema::PrimitiveType_Gather,
+  schema::PrimitiveType_Squeeze,       schema::PrimitiveType_Reshape,
+  schema::PrimitiveType_Unsqueeze,     schema::PrimitiveType_Transpose,
+};
+
+// this pass goal is to minimize subgraphs generated
+// by inserting nchw2nhwc or nhwc2nchw before or after the operator (e.g. concat, add, etc..) together with
+// fusion pass. If transpose inserted are more than half of input output, we will insert remaining input
+// output with transpose and hopefully do a fusion pass. Otherwise, we don't insert anything.
+
+// Typically concat accept output from nchw2nhwc, we fill other input with nh2nc and nc2nh so that inputs to concat are
+// format same and then fusion all nchw2nhwc op.
+// e.g.
+// original     (conv->nchw2nhwc, add(format nhwc)) -> concat-> (nhwc2nchw->conv)
+// current pass (conv->nchw2nhwc, add->nhwc2nchw->nchw2nhwc) -> concat -> (nhwc2nchw->conv)
+// fusion pass  (conv, add->nhwc2nchw) -> concat -> conv
+// original 2 cpusubgraph, after 2 pass, only 1 cpu subgraph
+
+// Such ops require inputs all have same format, could be nchw or nhwc or other format.
+// Their inputs outputs may not be 4d, or are already format ok,
+// so we won't insert nc2nh or nh2nc when op's in ops and out ops contains no nc2nh or nh2nc.
+// This pass should be run after npu_transform_pass, which insert transpose for nchw-input-limited op like conv2d.
+
+InsertState CoreMLTransExtendPass::GetInsertState(CoreMLOp *op) {
+  // filter out irrelevant op
+  if (format_depend_nodes.find(op->type()) != format_depend_nodes.end()) {
+    return InsertState::InsertNone;
+  }
+  // current op is target op
+  // Use out ops to count the out lines from current op since a single tensor can be used by multiple out ops. Besides,
+  // a tensor can be used by out ops and graph output at the same time, there will be one more line in this case.
+  std::vector<mindspore::MSTensor> inputs = CoreMLPassUtils::GetNonConstInputs(op);
+  size_t in_out_tensor_num =
+    inputs.size() + std::max(std::max(op->out_ops().size(), static_cast<size_t>(1)), op->outputs().size());
+  size_t transpose_input_num = 0;
+  size_t transpose_output_num = 0;
+  size_t graph_input_num = 0;
+  size_t graph_output_num = 0;
+  bool need_pre_insert = false;
+  bool need_post_insert = false;
+  // count number of input tensor from nc2nh and output tensor to nh2nc
+  for (size_t i = 0; i < inputs.size(); ++i) {
+    auto in_op = CoreMLPassUtils::OpInputFromOp(op, inputs.at(i));
+    if (CoreMLPassUtils::IsNchw2Nhwc(in_op)) {
+      transpose_input_num++;
+    } else {
+      need_pre_insert = true;
+    }
+    if (in_op == nullptr) {
+      graph_input_num++;
+    }
+  }
+  auto graph_output = subgraph_->outputs();
+  for (auto output : op->outputs()) {
+    if (std::find(graph_output.begin(), graph_output.end(), output) != graph_output.end()) {
+      graph_output_num++;
+      need_post_insert = true;
+    }
+  }
+  for (const auto out_op : op->out_ops()) {
+    for (auto out_op_input : out_op->inputs()) {
+      if (std::find(graph_output.begin(), graph_output.end(), out_op_input) != graph_output.end()) {
+        in_out_tensor_num++;
+      }
+    }
+    if (CoreMLPassUtils::IsNhwc2Nchw(out_op)) {
+      transpose_output_num++;
+    } else {
+      need_post_insert = true;
+    }
+  }
+
+  // won't insert any thing if num of transpose tensor is smaller than half of total op inputs and op outputs, unless
+  // current op is the graph input or output op, since we should avoid to build a single op subgraph in this case.
+  // won't insert if total input output are all transpose tensor, the fusion pass will handle this.
+  size_t transpose_tensor_num = transpose_input_num + transpose_output_num;
+  size_t connected_in_out_tensor_num = in_out_tensor_num - graph_output_num - graph_input_num;
+  if (transpose_tensor_num == 0 || transpose_tensor_num * REPEAT_TIMES2 < connected_in_out_tensor_num ||
+      transpose_tensor_num == in_out_tensor_num) {
+    return InsertState::InsertNone;
+  }
+  InsertState ret = (need_pre_insert && need_post_insert)
+                      ? InsertState::BothInsert
+                      : (need_pre_insert ? InsertState::PreInsert
+                                         : (need_post_insert ? InsertState::PostInsert : InsertState::InsertNone));
+
+  return ret;
+}
+
+int CoreMLTransExtendPass::InsertTransNode(CoreMLOp *op, CoreMLOp *post_op, const mindspore::MSTensor &trans_in_tensor,
+                                           std::vector<CoreMLOp *> *trans_ops) {
+  MS_ASSERT(op != nullptr || post_op != nullptr);
+  std::string op_name;
+  std::vector<CoreMLOp *> in_ops;
+  std::vector<CoreMLOp *> out_ops;
+  if (op != nullptr) {
+    op_name = op->name() + "_post";
+    in_ops.emplace_back(op);
+  }
+  if (post_op != nullptr) {
+    op_name = post_op->name() + "_pre";
+    out_ops.emplace_back(post_op);
+  }
+  auto nhwc_shape = trans_in_tensor.Shape();
+  std::vector<int64_t> nchw_shape = {nhwc_shape[kNHWC_N], nhwc_shape[kNHWC_C], nhwc_shape[kNHWC_H],
+                                     nhwc_shape[kNHWC_W]};
+
+  auto nh2nc_name = op_name + "_nh2nc_" + std::to_string(total++);
+  auto nh2nc_tensor =
+    mindspore::MSTensor::CreateTensor(nh2nc_name + "/output0", trans_in_tensor.DataType(), nchw_shape, nullptr, 0);
+  if (nh2nc_tensor == nullptr) {
+    MS_LOG(ERROR) << "New nchw tensor failed when inserting nchw2nhwc op.";
+    return RET_ERROR;
+  }
+  nh2nc_tensor->SetFormat(Format::NCHW);
+  std::vector<mindspore::MSTensor> nh2nc_tensors = {*nh2nc_tensor};
+  all_tensors_->push_back(nh2nc_tensor);
+
+  auto nc2nh_name = op_name + "_nc2nh_" + std::to_string(total++);
+  auto nc2nh_tensor =
+    mindspore::MSTensor::CreateTensor(nc2nh_name + "/output0", trans_in_tensor.DataType(), nhwc_shape, nullptr, 0);
+  if (nc2nh_tensor == nullptr) {
+    MS_LOG(ERROR) << "New nhwc tensor failed when inserting nhwc2nchw op.";
+    return RET_ERROR;
+  }
+  nc2nh_tensor->SetFormat(Format::NHWC);
+  std::vector<mindspore::MSTensor> nc2nh_tensors = {*nc2nh_tensor};
+  all_tensors_->push_back(nc2nh_tensor);
+
+  auto *nh2nc_op = CoreMLPassUtils::CreateNhwc2NchwOp({trans_in_tensor}, nh2nc_tensors, nh2nc_name);
+  trans_ops->push_back(nh2nc_op);
+
+  auto *nc2nh_op = CoreMLPassUtils::CreateNchw2NhwcOp(nh2nc_tensors, nc2nh_tensors, nc2nh_name);
+  trans_ops->push_back(nc2nh_op);
+
+  CoreMLPassUtils::UpdateOp(nh2nc_op, in_ops, {nc2nh_op}, {trans_in_tensor}, nh2nc_tensors);
+  CoreMLPassUtils::UpdateOp(nc2nh_op, {nh2nc_op}, out_ops, {nh2nc_tensors[0]}, nc2nh_tensors);
+  if (op != nullptr) {
+    CoreMLPassUtils::UpdateNH2NCTransNodePreOp(op, nh2nc_op, post_op);
+  }
+  if (post_op != nullptr) {
+    CoreMLPassUtils::UpdateNC2NHTransNodePostOp(op, nc2nh_op, post_op, trans_in_tensor);
+  } else {
+    // post_op nullptr mean output, we remain graph output tensor name unchanged
+    auto graph_output_name = trans_in_tensor.Name();
+    nc2nh_tensor->SetTensorName(graph_output_name + "_after_" + name_);
+  }
+  return RET_OK;
+}
+
+int CoreMLTransExtendPass::InsertPreNodes(CoreMLOp *op, std::vector<CoreMLOp *> *trans_ops) {
+  int ret = RET_OK;
+  auto inputs = CoreMLPassUtils::GetNonConstInputs(op);
+  for (auto tensor : inputs) {
+    if (tensor.Shape().size() < COMM_SHAPE_SIZE) {
+      continue;
+    }
+    // the input tensor can only come from a single op
+    auto pre_op = CoreMLPassUtils::OpInputFromOp(op, tensor);
+    if (CoreMLPassUtils::IsNchw2Nhwc(pre_op)) {
+      continue;
+    }
+    // if this tensor is input of graph, pre_op is nullptr.;
+    ret = InsertTransNode(pre_op, op, tensor, trans_ops);
+    if (ret != RET_OK) {
+      MS_LOG(ERROR) << "Insert nhwc2nchw op and nchw2nhwc op before op " << op->name() << " failed.";
+      return ret;
+    }
+  }
+  return ret;
+}
+
+int CoreMLTransExtendPass::InsertPostNodes(CoreMLOp *op, std::vector<CoreMLOp *> *trans_ops) {
+  int ret = RET_OK;
+  for (size_t idx = 0; idx < op->outputs().size(); idx++) {
+    auto out_tensor = op->outputs().at(idx);
+    if (out_tensor.Shape().size() < COMM_SHAPE_SIZE) {
+      continue;
+    }
+    if (std::find(subgraph_->outputs().begin(), subgraph_->outputs().end(), out_tensor) != subgraph_->outputs().end()) {
+      // the case that op's out tensor is graph output
+      ret = InsertTransNode(op, nullptr, op->outputs().at(idx), trans_ops);
+      if (ret != RET_OK) {
+        MS_LOG(ERROR) << "Insert nhwc2nchw op and nchw2nhwc op after op " << op->name() << " failed.";
+        return RET_ERROR;
+      }
+      // use origin output as the last trans op's output in order to avoid the lost of the output tensor after transpose
+      // fusion. The input of the cur_op's out_op will be updated in the loop below.
+      auto last_trans = trans_ops->back();
+      auto trans_output = last_trans->outputs();
+      auto cur_outputs = op->outputs();
+      cur_outputs[idx] = last_trans->outputs()[0];
+      trans_output[0] = op->outputs()[idx];
+      last_trans->set_outputs(trans_output);
+      op->set_outputs(cur_outputs);
+    }
+
+    // besides of being as graph outputs, the output tensors also can connected with multiple ops.
+    for (auto post_op : op->out_ops()) {
+      auto post_op_input = post_op->inputs();
+      auto it = std::find(post_op_input.begin(), post_op_input.end(), out_tensor);
+      if (it == post_op_input.end()) {
+        continue;
+      }
+      auto related_idx = it - post_op_input.begin();
+      post_op_input[related_idx] = op->outputs().at(idx);
+      post_op->set_inputs(post_op_input);
+
+      if (CoreMLPassUtils::IsNhwc2Nchw(post_op)) {
+        continue;
+      }
+      // the case that op's out tensor is one of post_op's input tensor
+      ret = InsertTransNode(op, post_op, op->outputs().at(idx), trans_ops);
+      if (ret != RET_OK) {
+        MS_LOG(ERROR) << "Insert nhwc2nchw op and nchw2nhwc op after op " << op->name() << " failed.";
+        return ret;
+      }
+    }
+  }
+  return ret;
+}
+
+int CoreMLTransExtendPass::Run(CoreMLGraph *subgraph) {
+  subgraph_ = subgraph;
+  all_ops_ = subgraph_->GetOps();
+  all_tensors_ = subgraph_->GetInsertTensors();
+  std::vector<CoreMLOp *> insert_ops;
+  for (int j = 0; j < REPEAT_TIMES2; ++j) {
+    for (size_t i = 0; i < all_ops_->size(); i++) {
+      auto op = (*all_ops_)[i];
+      auto insert_state = GetInsertState(op);
+      insert_ops.clear();
+      // If the every output op is nhwc2nchw, insert
+      // modify loop index add post_ops.size() to the next op in the origin vector
+      switch (insert_state) {
+        case InsertState::PreInsert: {
+          auto ret = InsertPreNodes(op, &insert_ops);
+          if (ret != RET_OK) {
+            MS_LOG(ERROR) << "Insert nhwc2nchw op and nchw2nhwc op before op " << op->name() << " failed.";
+            return RET_ERROR;
+          }
+          all_ops_->insert(all_ops_->begin() + i, insert_ops.begin(), insert_ops.end());
+          i += insert_ops.size();
+          break;
+        }
+        case InsertState::PostInsert: {
+          auto ret = InsertPostNodes(op, &insert_ops);
+          if (ret != RET_OK) {
+            MS_LOG(ERROR) << "Insert nhwc2nchw op and nchw2nhwc op after op " << op->name() << " failed.";
+            return RET_ERROR;
+          }
+          all_ops_->insert(all_ops_->begin() + i + 1, insert_ops.begin(), insert_ops.end());
+          i += insert_ops.size();
+          break;
+        }
+        case InsertState::BothInsert: {
+          auto ret = InsertPreNodes(op, &insert_ops);
+          if (ret != RET_OK) {
+            MS_LOG(ERROR) << "Insert nhwc2nchw op and nchw2nhwc op before op " << op->name() << " failed.";
+            return RET_ERROR;
+          }
+          all_ops_->insert(all_ops_->begin() + i, insert_ops.begin(), insert_ops.end());
+          i += insert_ops.size();
+
+          insert_ops.clear();
+          ret = InsertPostNodes(op, &insert_ops);
+          if (ret != RET_OK) {
+            MS_LOG(ERROR) << "Insert nhwc2nchw op and nchw2nhwc op after op " << op->name() << " failed.";
+            return RET_ERROR;
+          }
+          all_ops_->insert(all_ops_->begin() + i + 1, insert_ops.begin(), insert_ops.end());
+          i += insert_ops.size();
+          break;
+        }
+        default:
+          MS_LOG(DEBUG) << "Insert Nothing on op " << op->name();
+      }
+    }
+  }
+  return RET_OK;
+}
+}  // namespace mindspore

mindspore/lite/src/runtime/delegate/coreml/pass/coreml_trans_extend_pass.h

@@ -0,0 +1,45 @@
+/**
+ * Copyright 2020-2022 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_RUNTIME_DELEGATE_COREML_PASS_COREML_TRANS_EXTEND_PASS_H_
+#define MINDSPORE_LITE_SRC_RUNTIME_DELEGATE_COREML_PASS_COREML_TRANS_EXTEND_PASS_H_
+#include <vector>
+#include "src/runtime/delegate/coreml/op/coreml_op.h"
+#include "src/runtime/delegate/coreml/pass/coreml_base_pass.h"
+namespace mindspore {
+enum class InsertState { InsertNone, PreInsert, PostInsert, BothInsert };
+class CoreMLTransExtendPass : public CoreMLBasePass {
+ public:
+  CoreMLTransExtendPass() { name_ = "CoreMLTransExtendPass"; }
+
+  int Run(CoreMLGraph *subgraph) override;
+
+ private:
+  InsertState GetInsertState(CoreMLOp *op);
+  bool IsNeedInsert(size_t transpose_tensor_num, size_t graph_input_num, size_t graph_output_num,
+                    size_t in_out_tensor_num, bool need_pre_insert, bool need_post_insert);
+  int InsertPreNodes(CoreMLOp *op, std::vector<CoreMLOp *> *trans_ops);
+  int InsertPostNodes(CoreMLOp *op, std::vector<CoreMLOp *> *trans_ops);
+  int InsertTransNode(CoreMLOp *op, CoreMLOp *post_op, const mindspore::MSTensor &trans_in_tensor,
+                      std::vector<CoreMLOp *> *trans_ops);
+
+ private:
+  int total = 0;
+  std::vector<CoreMLOp *> *all_ops_ = nullptr;
+  std::vector<mindspore::MSTensor *> *all_tensors_ = nullptr;
+};
+}  // namespace mindspore
+#endif  // MINDSPORE_LITE_SRC_RUNTIME_DELEGATE_COREML_PASS_COREML_TRANS_EXTEND_PASS_H_

mindspore/lite/src/runtime/delegate/delegate_utils.cc

@@ -15,8 +15,49 @@
  */
 
 #include "src/runtime/delegate/delegate_utils.h"
-namespace mindspore::lite {
+#include "nnacl/fp32/pack_fp32.h"
+namespace mindspore {
+namespace lite {
+void PackNHWCToNCHWFp32(const void *src, void *dst, int batches, int plane, int channel) {
+  int hw8 = plane / C8NUM * C8NUM;
+  int batch = plane * channel;
+  for (int n = 0; n < batches; n++) {
+    const float *src_batch = (const float *)src + n * batch;
+    float *dst_batch = reinterpret_cast<float *>(dst) + n * batch;
+    int hw = 0;
+    for (; hw < hw8; hw += C8NUM) {
+      int c = 0;
+#ifdef ENABLE_ARM64
+      for (; c <= channel - C8NUM; c += C8NUM) {
+        const float *src_ptr = src_batch + hw * channel + c;
+        float *dst_ptr = dst_batch + c * plane + hw;
+        Transpose8X8Fp32Arm64(src_ptr, dst_ptr, channel, plane);
+      }
+#endif
+      for (; c < channel; c++) {
+        const float *src_ptr = src_batch + hw * channel + c;
+        float *dst_ptr = dst_batch + c * plane + hw;
+        for (size_t i = 0; i < C8NUM; i++) {
+          dst_ptr[i] = src_ptr[i * channel];
+        }
+      }
+    }
+    for (; hw < plane; hw++) {
+      const float *src_ptr = src_batch + hw * channel;
+      float *dst_ptr = dst_batch + hw;
+      for (size_t i = 0; i < channel; i++) {
+        dst_ptr[i * plane] = src_ptr[i];
+      }
+    }
+  }
+}
+
+void PackNCHWToNHWCFp32(const void *src, void *dst, int batch, int plane, int channel) {
+  return PackNHWCToNCHWFp32(src, dst, batch, channel, plane);
+}
+
 bool IsSubGraphInputTensor(const std::vector<mindspore::MSTensor> &inputs, mindspore::MSTensor input) {
   return std::find(inputs.begin(), inputs.end(), input) != inputs.end();
 }
-}  // namespace mindspore::lite
+}  // namespace lite
+}  // namespace mindspore

mindspore/lite/src/runtime/delegate/delegate_utils.h

@@ -23,6 +23,10 @@
 namespace mindspore::lite {
 bool IsSubGraphInputTensor(const std::vector<mindspore::MSTensor> &inputs, mindspore::MSTensor input);
 
+void PackNHWCToNCHWFp32(const void *src, void *dst, int batches, int plane, int channel);
+
+void PackNCHWToNHWCFp32(const void *src, void *dst, int batch, int plane, int channel);
+
 template <typename T>
 std::vector<mindspore::MSTensor> GetGraphInTensors(std::vector<T *> ops, std::vector<size_t> *input_index) {
   std::vector<mindspore::MSTensor> inputs;

mindspore/lite/src/runtime/delegate/npu/CMakeLists.txt

@@ -4,6 +4,7 @@ file(GLOB_RECURSE NPU_RUNTIME_SRC
         ${CMAKE_CURRENT_SOURCE_DIR}/*.cc
         ${CMAKE_CURRENT_SOURCE_DIR}/op/*.cc
         ${CMAKE_CURRENT_SOURCE_DIR}/pass/*.cc
+        ${CMAKE_CURRENT_SOURCE_DIR}/../delegate_utils.cc
         )
 add_library(hiai SHARED IMPORTED)
 set_target_properties(hiai PROPERTIES IMPORTED_LOCATION

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

@@ -286,38 +286,10 @@ NPUOp *NPUDelegate::GetOP(kernel::Kernel *kernel, const schema::Primitive *primi
   return npu_op;
 }
 
-std::vector<mindspore::MSTensor> GraphOutTensors(const std::vector<NPUOp *> &ops,
-                                                 DelegateModel<schema::Primitive> *model, KernelIter from,
-                                                 KernelIter end) {
-  auto out_tensors = lite::GetGraphOutTensors(ops);
-  std::vector<mindspore::MSTensor> all_out_tensors;
-  for (auto op : ops) {
-    for (const auto &out_tensor : op->outputs()) {
-      if (find(out_tensors.begin(), out_tensors.end(), out_tensor) == out_tensors.end()) {
-        all_out_tensors.push_back(out_tensor);
-      }
-    }
-  }
-
-  for (auto iter = model->BeginKernelIterator(); iter != model->EndKernelIterator(); iter++) {
-    if (iter >= from && iter <= end) {
-      continue;
-    }
-    // The input of other kernels is the output of the current subgraph kernel.
-    for (const auto &in_tensor : (*iter)->inputs()) {
-      if (find(all_out_tensors.begin(), all_out_tensors.end(), in_tensor) != all_out_tensors.end() &&
-          find(out_tensors.begin(), out_tensors.end(), in_tensor) == out_tensors.end()) {
-        out_tensors.push_back(in_tensor);
-      }
-    }
-  }
-  return out_tensors;
-}
-
 kernel::Kernel *NPUDelegate::CreateNPUGraph(const std::vector<NPUOp *> &ops, DelegateModel<schema::Primitive> *model,
                                             KernelIter from, KernelIter end) {
   auto in_tensors = lite::GetGraphInTensors(ops, nullptr);
-  auto out_tensors = GraphOutTensors(ops, model, from, end);
+  auto out_tensors = lite::GraphOutTensors<NPUOp>(ops, model, from, end);
   auto graph_kernel = new (std::nothrow) NPUGraph(ops, npu_manager_, in_tensors, out_tensors);
   if (graph_kernel == nullptr) {
     MS_LOG(DEBUG) << "New NPU Graph failed.";
@@ -326,12 +298,14 @@ kernel::Kernel *NPUDelegate::CreateNPUGraph(const std::vector<NPUOp *> &ops, Del
   // 1. For every op, find pre and next ops
   auto ret = graph_kernel->FindPreNextOps();
   if (ret != RET_OK) {
+    delete graph_kernel;
     MS_LOG(DEBUG) << "NPU Graph find input and output ops for every op failed.";
     return nullptr;
   }
   // 2. Pass
   ret = pass_manager_->RunPass(graph_kernel);
   if (ret != RET_OK) {
+    delete graph_kernel;
     MS_LOG(DEBUG) << "NPU Graph run pass failed. This function mainly solves the problem that the format is "
                      "inconsistent and requires interpolation transpose operators.";
     return nullptr;
@@ -339,6 +313,7 @@ kernel::Kernel *NPUDelegate::CreateNPUGraph(const std::vector<NPUOp *> &ops, Del
   // 3. NPUGraph init, create subgraph_kernel and transpose_kernel
   ret = graph_kernel->Init();
   if (ret != RET_OK) {
+    delete graph_kernel;
     MS_LOG(DEBUG) << "NPU subgraph Init failed.";
     return nullptr;
   }

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

@@ -17,6 +17,7 @@
 #include "src/runtime/delegate/npu/op/convolution_base_npu.h"
 #include "src/runtime/delegate/npu/npu_converter_utils.h"
 #include "src/runtime/delegate/npu/transpose_kernel.h"
+#include "src/runtime/delegate/delegate_utils.h"
 #include "nnacl/int8/pack_int8.h"
 
 namespace mindspore {
@@ -72,7 +73,8 @@ int ConvolutionBaseNPUOp::InitWeightConst(const std::vector<mindspore::MSTensor>
     // weight fp16->fp32
     Float16ToFloat32(reinterpret_cast<const float16_t *>(origin_weight), reinterpret_cast<float *>(fp32_weight_),
                      inputs[1].ElementNum());
-    PackNHWCToNCHWFp32(fp32_weight_, nchw_weight_, w_shape[NHWC_N], w_shape[NHWC_H] * w_shape[NHWC_W], w_shape[NHWC_C]);
+    lite::PackNHWCToNCHWFp32(fp32_weight_, nchw_weight_, w_shape[NHWC_N], w_shape[NHWC_H] * w_shape[NHWC_W],
+                             w_shape[NHWC_C]);
 #else
     MS_LOG(ERROR) << "This platform does not support fp16.";
     FreeTmpWeight();
@@ -84,7 +86,7 @@ int ConvolutionBaseNPUOp::InitWeightConst(const std::vector<mindspore::MSTensor>
       MS_LOG(ERROR) << "Malloc buffer failed.";
       return RET_ERROR;
     }
-    PackNHWCToNCHWFp32(origin_weight, nchw_weight_, w_shape[NHWC_N], w_shape[NHWC_H] * w_shape[NHWC_W],
+    lite::PackNHWCToNCHWFp32(origin_weight, nchw_weight_, w_shape[NHWC_N], w_shape[NHWC_H] * w_shape[NHWC_W],
                              w_shape[NHWC_C]);
   } else if (inputs[1].DataType() == DataType::kNumberTypeInt8) {
     nchw_weight_ = malloc(inputs[1].ElementNum() * sizeof(int8_t));

mindspore/lite/src/runtime/delegate/npu/transpose_kernel.cc

@@ -17,46 +17,9 @@
 #include "src/runtime/delegate/npu/transpose_kernel.h"
 #include "src/runtime/delegate/npu/npu_converter_utils.h"
 #include "src/runtime/delegate/npu/op/npu_op.h"
+#include "src/runtime/delegate/delegate_utils.h"
 #include "nnacl/fp32/pack_fp32.h"
 namespace mindspore {
-void PackNHWCToNCHWFp32(const void *src, void *dst, int batches, int plane, int channel) {
-  int hw8 = plane / C8NUM * C8NUM;
-  int batch = plane * channel;
-  for (int n = 0; n < batches; n++) {
-    const float *src_batch = (const float *)src + n * batch;
-    float *dst_batch = reinterpret_cast<float *>(dst) + n * batch;
-    int hw = 0;
-    for (; hw < hw8; hw += C8NUM) {
-      int c = 0;
-#ifdef ENABLE_ARM64
-      for (; c <= channel - C8NUM; c += C8NUM) {
-        const float *src_ptr = src_batch + hw * channel + c;
-        float *dst_ptr = dst_batch + c * plane + hw;
-        Transpose8X8Fp32Arm64(src_ptr, dst_ptr, channel, plane);
-      }
-#endif
-      for (; c < channel; c++) {
-        const float *src_ptr = src_batch + hw * channel + c;
-        float *dst_ptr = dst_batch + c * plane + hw;
-        for (size_t i = 0; i < C8NUM; i++) {
-          dst_ptr[i] = src_ptr[i * channel];
-        }
-      }
-    }
-    for (; hw < plane; hw++) {
-      const float *src_ptr = src_batch + hw * channel;
-      float *dst_ptr = dst_batch + hw;
-      for (size_t i = 0; i < channel; i++) {
-        dst_ptr[i * plane] = src_ptr[i];
-      }
-    }
-  }
-}
-
-void PackNCHWToNHWCFp32(const void *src, void *dst, int batch, int plane, int channel) {
-  return PackNHWCToNCHWFp32(src, dst, batch, channel, plane);
-}
-
 int TransposeNPUKernel::Execute() {
   if (perm_ != NHWC2NCHW_PERM && perm_ != NCHW2NHWC_PERM) {
     MS_LOG(ERROR) << "NPU transpose op only supports nhwc->nchw or nchw->nhwc.";
@@ -74,9 +37,9 @@ int TransposeNPUKernel::Execute() {
   auto output = out_tensor.MutableData();
   MS_ASSERT(output);
   if (perm_ == NHWC2NCHW_PERM) {
-    PackNHWCToNCHWFp32(input, output, shape[NHWC_N], shape[NHWC_H] * shape[NHWC_W], shape[NHWC_C]);
+    lite::PackNHWCToNCHWFp32(input, output, shape[NHWC_N], shape[NHWC_H] * shape[NHWC_W], shape[NHWC_C]);
   } else if (perm_ == NCHW2NHWC_PERM) {
-    PackNCHWToNHWCFp32(input, output, shape[NCHW_N], shape[NCHW_H] * shape[NCHW_W], shape[NCHW_C]);
+    lite::PackNCHWToNHWCFp32(input, output, shape[NCHW_N], shape[NCHW_H] * shape[NCHW_W], shape[NCHW_C]);
   } else {
     MS_LOG(ERROR) << "NPU transpose op only supports nhwc->nchw or nchw->nhwc.";
     return RET_ERROR;

mindspore/lite/src/runtime/lite_session.cc

@@ -52,6 +52,9 @@
 #if GPU_TENSORRT
 #include "src/extendrt/delegate/tensorrt/tensorrt_delegate.h"
 #endif
+#ifdef ENABLE_COREML
+#include "src/runtime/delegate/coreml/coreml_delegate.h"
+#endif
 #include "src/runtime/runtime_convert.h"
 #include "extendrt/mindir_loader/model_loader.h"
 
@@ -820,21 +823,38 @@ int LiteSession::CreateNPUDelegate() {
   return RET_OK;
 }
 
+int LiteSession::CreateCoreMLDelegate() {
+#ifdef ENABLE_COREML
+  delegate_ = std::make_shared<CoreMLDelegate>();
+  if (delegate_ == nullptr) {
+    MS_LOG(ERROR) << "New delegate_ failed";
+    return RET_ERROR;
+  }
+  delegate_device_type_ = DT_CPU;
+  this->context_->delegate = delegate_;
+#endif
+  return RET_OK;
+}
+
 int LiteSession::DelegateInit() {
 #ifndef DELEGATE_CLIP
   if (context_->delegate != nullptr) {
     delegate_ = context_->delegate;
     delegate_device_type_ = -1;
   } else {
+    auto ret = CreateCoreMLDelegate();
+    if (ret != RET_OK) {
+      return ret;
+    }
     if (context_->IsDeviceTypeEnabled(DT_NPU)) {
-      auto ret = CreateNPUDelegate();
+      ret = CreateNPUDelegate();
       if (ret != RET_OK) {
         return ret;
       }
     }
 
     if (context_->IsDeviceTypeEnabled(DT_GPU)) {
-      auto ret = CreateTensorRTDelegate();
+      ret = CreateTensorRTDelegate();
       if (ret != RET_OK) {
         return ret;
       }

mindspore/lite/src/runtime/lite_session.h

@@ -150,6 +150,7 @@ class LiteSession {
   int ContextInit(InnerContext *context);
   int CreateTensorRTDelegate();
   int CreateNPUDelegate();
+  int CreateCoreMLDelegate();
   int DelegateInit();
   int InitGPURuntime();
 

mindspore/lite/tools/cropper/build_cropper_config.sh

@@ -367,7 +367,8 @@ npu_files=()
 while IFS='' read -r line; do npu_files+=("$line"); done < <(ls mindspore/lite/src/runtime/delegate/npu/*.cc)
 while IFS='' read -r line; do npu_files+=("$line"); done < <(ls mindspore/lite/src/runtime/delegate/npu/op/*.cc)
 while IFS='' read -r line; do npu_files+=("$line"); done < <(ls mindspore/lite/src/runtime/delegate/npu/pass/*.cc)
-
+npu_others_files=("mindspore/lite/src/runtime/delegate/delegate_utils.cc")
+npu_files=("${npu_files[@]}" "${npu_others_files[@]}")
 # shellcheck disable=SC2068
 for file in ${npu_files[@]}; do
   file=$(echo ${file} | awk -F '/' '{print $NF}')

third_party/proto/coreml/DataStructures.proto

@@ -0,0 +1,95 @@
+// Copyright (c) 2017, Apple Inc. All rights reserved.
+//
+// Use of this source code is governed by a BSD-3-clause license that can be
+// found in LICENSE.txt or at https://opensource.org/licenses/BSD-3-Clause
+
+syntax = "proto3";
+option optimize_for = LITE_RUNTIME;
+
+import public "FeatureTypes.proto";
+
+package CoreML.Specification;
+
+/**
+ * A mapping from a string
+ * to a 64-bit integer.
+ */
+message StringToInt64Map {
+    map<string, int64> map = 1;
+}
+
+/**
+ * A mapping from a 64-bit integer
+ * to a string.
+ */
+message Int64ToStringMap {
+    map<int64, string> map = 1;
+}
+
+/**
+ * A mapping from a string
+ * to a double-precision floating point number.
+ */
+message StringToDoubleMap {
+    map<string, double> map = 1;
+}
+
+/**
+ * A mapping from a 64-bit integer
+ * to a double-precision floating point number.
+ */
+message Int64ToDoubleMap {
+    map<int64, double> map = 1;
+}
+
+/**
+ * A vector of strings.
+ */
+message StringVector {
+    repeated string vector = 1;
+}
+
+/**
+ * A vector of 64-bit integers.
+ */
+message Int64Vector {
+    repeated int64 vector = 1;
+}
+
+/**
+ * A vector of floating point numbers.
+ */
+message FloatVector {
+    repeated float vector = 1;
+}
+
+/**
+ * A vector of double-precision floating point numbers.
+ */
+message DoubleVector {
+    repeated double vector = 1;
+}
+
+/**
+ * A range of int64 values
+ */
+message Int64Range {
+    int64 minValue = 1;
+    int64 maxValue = 2;
+}
+
+/**
+ * A set of int64 values
+ */
+message Int64Set {
+    repeated int64 values = 1;
+}
+
+/**
+ * A range of double values
+ */
+message DoubleRange {
+    double minValue = 1;
+    double maxValue = 2;
+}
+

third_party/proto/coreml/FeatureTypes.proto

@@ -0,0 +1,224 @@
+// Copyright (c) 2017, Apple Inc. All rights reserved.
+//
+// Use of this source code is governed by a BSD-3-clause license that can be
+// found in LICENSE.txt or at https://opensource.org/licenses/BSD-3-Clause
+
+syntax = "proto3";
+option optimize_for = LITE_RUNTIME;
+
+package CoreML.Specification;
+
+/**
+ * The 64-bit integer feature type.
+ */
+message Int64FeatureType {}
+
+/**
+ * The double-precision floating point number feature type.
+ */
+message DoubleFeatureType {}
+
+/**
+ * The string feature type.
+ */
+message StringFeatureType {}
+
+
+message SizeRange {
+    uint64 lowerBound = 1;
+    int64 upperBound = 2; // negative value means unbound otherwise upperbound is included in range
+}
+
+/**
+ * The image feature type.
+ */
+message ImageFeatureType {
+    // Assumes raw (decompressed) format
+    enum ColorSpace {
+        INVALID_COLOR_SPACE = 0;
+        GRAYSCALE = 10; //  8 bits per pixel
+        RGB = 20;       // 32 bits per pixel: RGBA with A channel ignored
+        BGR = 30;       // 32 bits per pixel: BGRA with A channel ignored
+    }
+
+    message ImageSize {
+        uint64 width = 1;
+        uint64 height = 2;
+    }
+
+    message EnumeratedImageSizes {
+        repeated ImageSize sizes = 1;
+    }
+
+    message ImageSizeRange {
+        SizeRange widthRange = 1;
+        SizeRange heightRange = 2;
+    }
+
+    // The required or default image size is width x height
+    //
+    // If specificationVersion <= 2 or SizeFlexibility is empty,
+    // width x height is the required fixed image size
+    //
+    // If SizeFlexibility is present, width x height indicate a "default"
+    // image size which must be consistent with the flexibilty specified
+
+    int64 width = 1;
+    int64 height = 2;
+
+    // For specification version >= 3 you can specify image size flexibility.
+
+    oneof SizeFlexibility {
+
+        // Use enumeratedSizes for a set of distinct fixed sizes
+        // e.g. portrait or landscape: [80 x 100, 100 x 8]
+        //
+        // If the width x height fields above are specified then they must be
+        // one of the sizes listed.
+        //
+        // If width and height are not specified above then the default width
+        // and height will be enumeratedSizes[0]
+        //
+        // Must be non-empty
+
+        EnumeratedImageSizes enumeratedSizes = 21;
+
+        // Use imageSizeRange to allow for ranges of values
+        // e.g. any image greater than 10 x 20: [10..<max] x [20..<max]
+        //
+        // If width and height are specified above they must fall in the range
+        // specified in imageSizeRange. They will be treated as the default size.
+        //
+        // If width and height are not specified above then the default width
+        // and height will be imageSizeRange.widthRange.lowerBound x imageSizeRange.heightRange.lowerBound
+
+        ImageSizeRange imageSizeRange = 31;
+    }
+
+    ColorSpace colorSpace = 3;
+}
+
+/**
+ * The array feature type.
+ */
+message ArrayFeatureType {
+
+    enum ArrayDataType {
+        INVALID_ARRAY_DATA_TYPE = 0;
+        FLOAT32 = 65568; // 0x10000 | 32
+        DOUBLE = 65600;  // 0x10000 | 64
+        INT32 = 131104;  // 0x20000 | 32
+    }
+
+    // The required or default shape
+    //
+    // If specificationVersion <= 2 or ShapeFlexibility is empty,
+    // shape is the required fixed shape
+    //
+    // If ShapeFlexibility is present, shape indicate a "default"
+    // shape which must be consistent with the flexibilty specified
+
+    repeated int64 shape = 1;
+
+    ArrayDataType dataType = 2;
+
+    message Shape {
+        repeated int64 shape = 1;
+    }
+
+    message EnumeratedShapes {
+        repeated Shape shapes = 1;
+    }
+
+    message ShapeRange {
+        // sizeRanges.size() must be length 1 or 3
+        // sizeRanges[d] specifies the allowed range for dimension d
+        repeated SizeRange sizeRanges = 1;
+    }
+
+    // For specification version >= 3 you can specify image size flexibility.
+
+    oneof ShapeFlexibility {
+
+        // Use enumeratedShapes for a set of distinct fixed shapes
+        //
+        // If the shape field is specified then it must be
+        // one of the enumerated shapes.
+        ///
+        // If shape is not specifed, the "default" shape will be considered
+        // enumeratedShapes[0]
+        //
+        // Must be non-empty
+
+        EnumeratedShapes enumeratedShapes = 21;
+
+        // Use shapeRange to allow the size of each dimension vary within
+        // indpendently specified ranges
+        //
+        // If you specify shape above it must fall in the range
+        // specified in shapeRanges. It will be treated as the default shape.
+        //
+        // If you don't specify shape above then the default shape will
+        // have shape[d] = shapeRange.sizeRanges[d].lowerBound
+
+        ShapeRange shapeRange = 31;
+
+    }
+
+    oneof defaultOptionalValue {
+        int32 intDefaultValue = 41;
+        float floatDefaultValue = 51;
+        double doubleDefaultValue = 61;
+    }
+
+}
+
+/**
+ * The dictionary feature type.
+ */
+message DictionaryFeatureType {
+    /**
+     *  Key/value type tags, with the following restrictions:
+     *  - ``keyType`` must be a hashable type
+     *  - ``valueType`` is assumed to be a ``double``
+     */
+    oneof KeyType {
+        Int64FeatureType int64KeyType = 1;
+        StringFeatureType stringKeyType = 2;
+    }
+}
+
+/**
+ * The Sequence feature type.
+ */
+message SequenceFeatureType {
+
+    /**
+     * Currently only categorical int64 and String sequences are supported
+     */
+    oneof Type {
+        Int64FeatureType int64Type = 1;
+        StringFeatureType stringType = 3;
+    }
+
+    // Range of allowed size/length/count of sequence
+    SizeRange sizeRange = 101;
+}
+
+/**
+ * A feature, which may be optional.
+ */
+message FeatureType {
+    oneof Type {
+        Int64FeatureType int64Type = 1;
+        DoubleFeatureType doubleType = 2;
+        StringFeatureType stringType = 3;
+        ImageFeatureType imageType = 4;
+        ArrayFeatureType multiArrayType = 5;
+        DictionaryFeatureType dictionaryType = 6;
+        SequenceFeatureType sequenceType = 7;
+    }
+
+    bool isOptional = 1000;
+}
+

third_party/proto/coreml/Model.proto

@@ -0,0 +1,164 @@
+// Copyright (c) 2017, Apple Inc. All rights reserved.
+//
+// Use of this source code is governed by a BSD-3-clause license that can be
+// found in LICENSE.txt or at https://opensource.org/licenses/BSD-3-Clause
+
+
+syntax = "proto3";
+option optimize_for = LITE_RUNTIME;
+
+import public "NeuralNetwork.proto";
+
+package CoreML.Specification;
+
+/**
+ * A feature description,
+ * consisting of a name, short description, and type.
+ */
+message FeatureDescription {
+    string name = 1;
+    string shortDescription = 2;
+    FeatureType type = 3;
+}
+
+/**
+ * Model metadata,
+ * consisting of a short description, a version string,
+ * an author, a license, and any other user defined
+ * key/value meta data.
+ */
+message Metadata {
+    string shortDescription = 1;
+    string versionString = 2;
+    string author = 3;
+    string license = 4;
+    map<string, string> userDefined = 100;
+}
+
+/**
+ * A description of a model,
+ * consisting of descriptions of its input and output features.
+ * Both regressor and classifier models require the name of the
+ * primary predicted output feature (``predictedFeatureName``).
+ * Classifier models can specify the output feature containing
+ * probabilities for the predicted classes
+ * (``predictedProbabilitiesName``).
+ */
+message ModelDescription {
+    repeated FeatureDescription input = 1;
+    repeated FeatureDescription output = 10;
+
+    // [Required for regressor and classifier models]: the name
+    // to give to an output feature containing the prediction.
+    string predictedFeatureName = 11;
+
+    // [Optional for classifier models]: the name to give to an
+    // output feature containing a dictionary mapping class
+    // labels to their predicted probabilities. If not specified,
+    // the dictionary will not be returned by the model.
+    string predictedProbabilitiesName = 12;
+
+    repeated FeatureDescription trainingInput = 50;
+
+    Metadata metadata = 100;
+}
+
+message SerializedModel {
+    // Identifier whose content describes the model type of the serialized protocol buffer message.
+    string identifier = 1;
+
+    // Must be a valid serialized protocol buffer of the above specified type.
+    bytes model = 2;
+}
+
+/**
+ * A Core ML model,
+ * consisting of a specification version,
+ * a model description, and a model type.
+ *
+ * Core ML model compatibility is indicated by
+ * a monotonically increasing specification version number,
+ * which is incremented anytime a backward-incompatible change is made
+ * (this is functionally equivalent to the MAJOR version number
+ * described by `Semantic Versioning 2.0.0 <http://semver.org/>`_).
+ *
+ * Specification Versions : OS Availability (Core ML Version)
+ *
+ * 1 : iOS 11, macOS 10.13, tvOS 11, watchOS 4 (Core ML 1)
+ * - Feedforward & Recurrent Neural Networks
+ * - General Linear Models
+ * - Tree Ensembles
+ * - Support Vector Machines
+ * - Pipelines
+ * - Feature Engineering
+ *
+ * 2 : iOS 11.2, macOS 10.13.2, tvOS 11.2, watchOS 4.2 (Core ML 1.2)
+ * - Custom Layers for Neural Networks
+ * - Float 16 support for Neural Network layers
+ *
+ * 3 : iOS 12, macOS 10.14, tvOS 12, watchOS 5 (Core ML 2)
+ * - Flexible shapes and image sizes
+ * - Categorical sequences
+ * - Core ML Vision Feature Print, Text Classifier, Word Tagger
+ * - Non Max Suppression
+ * - Crop and Resize Bilinear NN layers
+ * - Custom Models
+ *
+ * 4 : iOS 13, macOS 10.15, tvOS 13, watchOS 6 (Core ML 3)
+ * - Updatable models
+ * - Exact shape / general rank mapping for neural networks
+ * - Large expansion of supported neural network layers
+ *   - Generalized operations
+ *   - Control flow
+ *   - Dynamic layers
+ *   - See NeuralNetwork.proto
+ * - Nearest Neighbor Classifier
+ * - Sound Analysis Prepreocessing
+ * - Recommender
+ * - Linked Model
+ * - NLP Gazeteer
+ * - NLP WordEmbedding
+ *
+ * 5 : iOS 14, macOS 11, tvOS 14, watchOS 7 (Core ML 4)
+ * - Model Deployment
+ * - Model Encryption
+ * - Unified converter API with PyTorch and Tensorflow 2 Support in coremltools 4
+ * - MIL builder for neural networks and composite ops in coremltools 4
+ * - New layers in neural network:
+ *      - CumSum
+ *      - OneHot
+ *      - ClampedReLu
+ *      - ArgSort
+ *      - SliceBySize
+ *      - Convolution3D
+ *      - Pool3D
+ *      - Bilinear Upsample with align corners and fractional factors
+ *      - PixelShuffle
+ *      - MatMul with int8 weights and int8 activations
+ *      - Concat interleave
+ *      - See NeuralNetwork.proto
+ * - Enhanced Xcode model view with interactive previews
+ * - Enhanced Xcode Playground support for Core ML models
+ *
+ */
+message Model {
+    int32 specificationVersion = 1;
+    ModelDescription description = 2;
+    
+    /*
+     * Following model types support on-device update:
+     *
+     * - NeuralNetworkClassifier
+     * - NeuralNetworkRegressor
+     * - NeuralNetwork
+     * - KNearestNeighborsClassifier
+     */
+    bool isUpdatable = 10;
+    
+    // start at 200 here
+    // model specific parameters:
+    oneof Type {
+        // generic models start at 500
+        NeuralNetwork neuralNetwork = 500;
+    }
+}

third_party/proto/coreml/Parameters.proto

@@ -0,0 +1,52 @@
+// Copyright (c) 2017, Apple Inc. All rights reserved.
+//
+// Use of this source code is governed by a BSD-3-clause license that can be
+// found in LICENSE.txt or at https://opensource.org/licenses/BSD-3-Clause
+
+syntax = "proto3";
+option optimize_for = LITE_RUNTIME;
+
+import public "DataStructures.proto";
+
+package CoreML.Specification;
+
+/**
+ * Int64 parameter,
+ * consisting of a default int64 value, and allowed range or set of values
+ * value is unbounded if AllowedValues is not set.
+ */
+message Int64Parameter {
+    int64 defaultValue = 1;
+    oneof AllowedValues {
+        Int64Range range = 10;
+        Int64Set set = 11;
+    }
+}
+
+/**
+ * Double parameter,
+ * consisting of a default double value, and allowed range of values
+ * value is unbounded if AllowedValues is not set.
+ */
+message DoubleParameter {
+    double defaultValue = 1;
+    oneof AllowedValues {
+        DoubleRange range = 10;
+    }
+}
+
+/**
+ * String parameter,
+ * A default string value must be provided
+ */
+message StringParameter {
+    string defaultValue = 1;
+}
+
+/**
+ * String parameter,
+ * A default bool value must be provided
+ */
+message BoolParameter {
+    bool defaultValue = 1;
+}