!3833 rewrite op register func

Merge pull request !3833 from fuzhiye/mindspore

mindspore/lite/src/kernel_factory.cc

@@ -43,19 +43,11 @@ LiteKernel *KernelFactory::GetKernel(const std::vector<tensor::Tensor *> &inputs
     MS_LOG(ERROR) << "PopulateParameter return nullptr, type: " << schema::EnumNamePrimitiveType(primitive->Type());
     return nullptr;
   }
-  auto creator = KernelRegistry::GetInstance()->GetKernelCreator(key);
+  auto creator = KernelRegistry::GetInstance()->GetCreator(key);
   if (creator != nullptr) {
-    auto *kernel = creator(inputs, outputs, parameter, ctx, key);
-    if (kernel != nullptr) {
-      return kernel;
-    } else {
-      MS_LOG(ERROR) << "Creator kernel failed for " << schema::EnumNamePrimitiveType(key.type);
-      return nullptr;
-    }
-  } else {
-    MS_LOG(ERROR) << "Can not find OpCreator for " << schema::EnumNamePrimitiveType(key.type);
-    return nullptr;
+    auto kernel = creator(inputs, outputs, parameter, ctx, key);
+    return kernel;
   }
+  return nullptr;
 }
 }  // namespace mindspore::lite
-

mindspore/lite/src/kernel_factory.h

@@ -38,4 +38,3 @@ class KernelFactory {
 }  // namespace mindspore::lite
 
 #endif  // MINDSPORE_LITE_SRC_KERNEL_FACTORY_H_
-

mindspore/lite/src/kernel_registry.cc

@@ -13,47 +13,105 @@
  * See the License for the specific language governing permissions and
  * limitations under the License.
  */
-
 #include "src/kernel_registry.h"
+#include "include/errorcode.h"
+#include "ir/dtype/type_id.h"
+#ifdef ENABLE_ARM64
+#include <asm/hwcap.h>
+#include "common/utils.h"
+#include "utils/log_adapter.h"
+#include "src/runtime/kernel/arm/opclib/optimized_kernel.h"
+#endif
 
+using mindspore::kernel::kCPU;
+using mindspore::kernel::KERNEL_ARCH;
 using mindspore::kernel::KernelCreator;
 using mindspore::kernel::KernelKey;
-using mindspore::kernel::KERNEL_ARCH;
+using mindspore::kernel::kKernelArch_MAX;
+using mindspore::kernel::kKernelArch_MIN;
+using mindspore::schema::PrimitiveType_MAX;
+using mindspore::schema::PrimitiveType_MIN;
 
 namespace mindspore::lite {
 KernelRegistry::KernelRegistry() {}
 
-KernelRegistry::~KernelRegistry() {}
+KernelRegistry::~KernelRegistry() { FreeCreatorArray(); }
 
 KernelRegistry *KernelRegistry::GetInstance() {
   static KernelRegistry instance;
   return &instance;
 }
 
-KernelCreator KernelRegistry::GetKernelCreator(const KernelKey &desc) {
-  auto it = creators.find(desc);
-  if (it != creators.end()) {
-    return it->second;
+int KernelRegistry::Init() {
+  lock_.lock();
+  if (creator_arrays_ != nullptr) {
+      lock_.unlock();
+      return RET_OK;
+  }
+  device_type_length_ = kKernelArch_MAX - kKernelArch_MIN;
+  data_type_length_ = kNumberTypeEnd - kNumberTypeBegin;
+  op_type_length_ = PrimitiveType_MAX - PrimitiveType_MIN;
+  // malloc an array contain creator functions of kernel
+  auto total_len = device_type_length_ * data_type_length_ * op_type_length_;
+  creator_arrays_ = (kernel::KernelCreator *)malloc(total_len * sizeof(kernel::KernelCreator));
+  if (creator_arrays_ == nullptr) {
+    MS_LOG(ERROR) << "malloc creator_arrays_ failed.";
+    lock_.unlock();
+    return RET_ERROR;
+  }
+  for (int i = 0; i < total_len; ++i) {
+    creator_arrays_[i] = nullptr;
+  }
+#ifdef ENABLE_ARM64
+  void *optimized_lib_handler = OptimizeModule::GetInstance()->optimized_op_handler_;
+  if (optimized_lib_handler != nullptr) {
+    MS_LOG(INFO) << "load optimize lib success.";
+  } else {
+    MS_LOG(INFO) << "load optimize lib failed.";
+  }
+#endif
+  lock_.unlock();
+  return RET_OK;
+}
+
+void KernelRegistry::FreeCreatorArray() {
+  if (creator_arrays_ != nullptr) {
+    free(creator_arrays_);
+    creator_arrays_ = nullptr;
   }
+}
 
-  // if not find, use cpu kernel
-  KernelKey cpuDesc {kernel::KERNEL_ARCH::kCPU, desc.type};
-  it = creators.find(cpuDesc);
-  if (it != creators.end()) {
-    return it->second;
+kernel::KernelCreator KernelRegistry::GetCreator(const KernelKey &desc) {
+  int index = GetCreatorFuncIndex(desc);
+  auto it = creator_arrays_[index];
+  if (it != nullptr) {
+    return it;
   }
   return nullptr;
 }
 
-void KernelRegistry::RegKernel(const KernelKey desc, KernelCreator creator) { creators[desc] = creator; }
+int KernelRegistry::GetCreatorFuncIndex(const kernel::KernelKey desc) {
+  int index;
+  int device_index = static_cast<int>(desc.arch);
+  int dType_index = static_cast<int>(desc.data_type);
+  int op_index = static_cast<int>(desc.type);
+  index = device_index * data_type_length_ * op_type_length_ + dType_index * op_type_length_ + op_index;
+  return index;
+}
 
-void KernelRegistry::RegKernel(const KERNEL_ARCH arch, const schema::PrimitiveType type, KernelCreator creator) {
-  KernelKey desc = {arch, type};
-  creators[desc] = creator;
+void KernelRegistry::RegKernel(const KernelKey desc, kernel::KernelCreator creator) {
+  int index = GetCreatorFuncIndex(desc);
+  creator_arrays_[index] = creator;
+}
+
+void KernelRegistry::RegKernel(const KERNEL_ARCH arch, const TypeId data_type, const schema::PrimitiveType op_type,
+                               kernel::KernelCreator creator) {
+  KernelKey desc = {arch, data_type, op_type};
+  int index = GetCreatorFuncIndex(desc);
+  creator_arrays_[index] = creator;
 }
 
 bool KernelRegistry::Merge(const std::unordered_map<KernelKey, KernelCreator> &newCreators) { return false; }
 
-const std::map<KernelKey, KernelCreator> &KernelRegistry::GetKernelCreators() { return creators; }
+const kernel::KernelCreator *KernelRegistry::GetCreatorArrays() { return creator_arrays_; }
 }  // namespace mindspore::lite
-

mindspore/lite/src/kernel_registry.h

@@ -30,16 +30,22 @@ class KernelRegistry {
   virtual ~KernelRegistry();
 
   static KernelRegistry *GetInstance();
-  virtual kernel::KernelCreator GetKernelCreator(const kernel::KernelKey &desc);
-
-  const std::map<kernel::KernelKey, kernel::KernelCreator> &GetKernelCreators();
-
+  int Init();
+  void FreeCreatorArray();
+  virtual kernel::KernelCreator GetCreator(const kernel::KernelKey &desc);
+  const kernel::KernelCreator *GetCreatorArrays();
+  int GetCreatorFuncIndex(const kernel::KernelKey desc);
   void RegKernel(const kernel::KernelKey desc, kernel::KernelCreator creator);
-  void RegKernel(const kernel::KERNEL_ARCH arch, const schema::PrimitiveType type, kernel::KernelCreator creator);
+  void RegKernel(const kernel::KERNEL_ARCH arch, const TypeId data_type, const schema::PrimitiveType type,
+                 kernel::KernelCreator creator);
   bool Merge(const std::unordered_map<kernel::KernelKey, kernel::KernelCreator> &newCreators);
 
  protected:
-  std::map<kernel::KernelKey, kernel::KernelCreator> creators;
+  kernel::KernelCreator *creator_arrays_ = nullptr;
+  int device_type_length_;
+  int data_type_length_;
+  int op_type_length_;
+  std::mutex lock_;
 };
 
 class KernelRegistrar {
@@ -48,14 +54,14 @@ class KernelRegistrar {
     KernelRegistry::GetInstance()->RegKernel(desc, creator);
   }
 
-  KernelRegistrar(const kernel::KERNEL_ARCH arch, const schema::PrimitiveType type, kernel::KernelCreator creator) {
-    KernelRegistry::GetInstance()->RegKernel(arch, type, creator);
+  KernelRegistrar(const kernel::KERNEL_ARCH arch, const TypeId data_type, const schema::PrimitiveType op_type,
+                  kernel::KernelCreator creator) {
+    KernelRegistry::GetInstance()->RegKernel(arch, data_type, op_type, creator);
   }
 };
 
-#define REG_KERNEL(arch, type, kernelCreater) \
-  static KernelRegistrar g_##arch##type##kernelReg(arch, type, kernelCreater);
+#define REG_KERNEL(arch, data_type, op_type, kernelCreater) \
+  static KernelRegistrar g_##arch##data_type##op_type##kernelReg(arch, data_type, op_type, kernelCreater);
 }  // namespace mindspore::lite
 
 #endif  // MINDSPORE_LITE_SRC_KERNEL_REGISTRY_H_
-

mindspore/lite/src/lite_kernel.h

@@ -18,7 +18,7 @@
 #define MINDSPORE_LITE_SRC_LITE_KERNEL_H_
 #include <vector>
 #include <string>
-#ifdef ENABLE_FP16
+#ifdef ENABLE_ARM
 #include <arm_neon.h>
 #endif
 #include "src/runtime/kernel/arm/opclib/op_base.h"
@@ -35,14 +35,17 @@ using FLOAT_t = float;
 
 // using mindspore::kernel::AddressPtr;
 namespace mindspore::kernel {
-enum KERNEL_ARCH { kCPU, kGPU, kNPU, kInferShape };
+enum KERNEL_ARCH { kCPU, kGPU, kNPU, kKernelArch_MIN = kCPU, kKernelArch_MAX = kNPU };
 struct KernelKey {
   KERNEL_ARCH arch;
+  TypeId data_type;
   schema::PrimitiveType type;
 
   bool operator<(const KernelKey &dst) const {
     if (arch != dst.arch) {
       return arch < dst.arch;
+    } else if (data_type != dst.data_type) {
+      return data_type < dst.data_type;
     } else {
       return type < dst.type;
     }
@@ -179,4 +182,3 @@ class LiteKernelUtil {
 }  // namespace mindspore::kernel
 
 #endif  // MINDSPORE_LITE_SRC_LITE_KERNEL_H_
-

mindspore/lite/src/lite_session.cc

@@ -24,6 +24,7 @@
 #include "src/executor.h"
 #include "src/common/utils.h"
 #include "src/common/graph_util.h"
+#include "src/kernel_registry.h"
 #if SUPPORT_GPU
 #include "src/runtime/opencl/opencl_runtime.h"
 #endif
@@ -197,7 +198,11 @@ void LiteSession::Init(Context *context) {
   this->context->deviceCtx.type = context->deviceCtx.type;
   this->context->allocator = std::make_shared<DefaultAllocator>();
   ConfigThreadPool(context->cpuBindMode, context->threadNum);
-
+  auto ret = KernelRegistry::GetInstance()->Init();
+  if (ret != RET_OK) {
+    MS_LOG(ERROR) << "KernelRegistry Init Failed.";
+    return;
+  }
 #if SUPPORT_GPU
   if (context->deviceCtx.type == DT_GPU) {
     auto opencl_runtime = lite::opencl::OpenCLRuntime::GetInstance();
@@ -228,6 +233,7 @@ LiteSession::~LiteSession() {
     delete kernel;
   }
 }
+
 std::vector<mindspore::tensor::MSTensor *> LiteSession::GetInputsByName(std::string name) {
   return input_map[name];
 }

mindspore/lite/src/runtime/kernel/arm/CMakeLists.txt

@@ -25,13 +25,5 @@ if (PLATFORM_ARM32)
     set(KERNEL_SRC ${KERNEL_SRC} ${ASSEMBLY_SRC})
 endif()
 
-if (ENABLE_FP16)
-    file(GLOB FP6_SRC
-            ${CMAKE_CURRENT_SOURCE_DIR}/fp16/*.cc
-            ${CMAKE_CURRENT_SOURCE_DIR}/opclib/fp16/*.cc
-            )
-    set(KERNEL_SRC ${KERNEL_SRC} ${FP6_SRC})
-endif ()
-
 add_library(cpu_kernel_mid_ OBJECT ${KERNEL_SRC})
 add_subdirectory(opclib)

mindspore/lite/src/runtime/kernel/arm/base/concat_base.cc

@@ -36,7 +36,8 @@ int ConcatBaseCPUKernel::Init() {
 
 kernel::LiteKernel *CpuConcatInt8KernelCreator(const std::vector<lite::tensor::Tensor *> &inputs,
                                                const std::vector<lite::tensor::Tensor *> &outputs,
-                                               OpParameter *opParameter, const Context *ctx) {
+                                               OpParameter *opParameter, const Context *ctx,
+                                               const kernel::KernelKey &desc) {
   if (opParameter == nullptr) {
     MS_LOG(ERROR) << "Input opParameter is nullptr!";
     return nullptr;
@@ -47,12 +48,20 @@ kernel::LiteKernel *CpuConcatInt8KernelCreator(const std::vector<lite::tensor::T
     MS_LOG(ERROR) << "new ConcatCPUKernel fail!";
     return nullptr;
   }
+  auto ret = kernel->Init();
+  if (ret != RET_OK) {
+    delete kernel;
+    MS_LOG(ERROR) << "Init kernel failed, name: " << opParameter->name_ << ", type: "
+                  << schema::EnumNamePrimitiveType(static_cast<schema::PrimitiveType>(opParameter->type_));
+    return nullptr;
+  }
   return kernel;
 }
 
-kernel::LiteKernel *CpuConcatFp32OrInt32KernelCreator(const std::vector<lite::tensor::Tensor *> &inputs,
+kernel::LiteKernel *CpuConcatInt32KernelCreator(const std::vector<lite::tensor::Tensor *> &inputs,
                                                       const std::vector<lite::tensor::Tensor *> &outputs,
-                                                      OpParameter *opParameter, const Context *ctx) {
+                                                      OpParameter *opParameter, const Context *ctx,
+                                                const kernel::KernelKey &desc) {
   if (opParameter == nullptr) {
     MS_LOG(ERROR) << "Input opParameter is nullptr!";
     return nullptr;
@@ -63,33 +72,28 @@ kernel::LiteKernel *CpuConcatFp32OrInt32KernelCreator(const std::vector<lite::te
     MS_LOG(ERROR) << "new ConcatCPUKernel fail!";
     return nullptr;
   }
+  auto ret = kernel->Init();
+  if (ret != RET_OK) {
+    delete kernel;
+    MS_LOG(ERROR) << "Init kernel failed, name: " << opParameter->name_ << ", type: "
+                  << schema::EnumNamePrimitiveType(static_cast<schema::PrimitiveType>(opParameter->type_));
+    return nullptr;
+  }
   return kernel;
 }
 
-kernel::LiteKernel *CpuConcatKernelCreator(const std::vector<lite::tensor::Tensor *> &inputs,
-                                           const std::vector<lite::tensor::Tensor *> &outputs,
-                                           OpParameter *opParameter,
-                                           const lite::Context *ctx, const kernel::KernelKey &desc) {
-  MS_ASSERT(opParameter != nullptr);
-  MS_ASSERT(desc.type == schema::PrimitiveType_Concat);
-  auto input_tensor = inputs.at(kInputIndex);
-  auto data_type = input_tensor->data_type();
-  kernel::LiteKernel *kernel = nullptr;
-  switch (data_type) {
-    case kNumberTypeInt8:
-    case kNumberTypeUInt8:
-      kernel = CpuConcatInt8KernelCreator(inputs, outputs, opParameter, ctx);
-      break;
-    case kNumberTypeInt32:
-    case kNumberTypeFloat32:
-      kernel = CpuConcatFp32OrInt32KernelCreator(inputs, outputs, opParameter, ctx);
-      break;
-    default:
-      break;
+kernel::LiteKernel *CpuConcatFp32KernelCreator(const std::vector<lite::tensor::Tensor *> &inputs,
+                                                      const std::vector<lite::tensor::Tensor *> &outputs,
+                                                      OpParameter *opParameter, const Context *ctx,
+                                               const kernel::KernelKey &desc) {;
+  if (opParameter == nullptr) {
+    MS_LOG(ERROR) << "Input opParameter is nullptr!";
+    return nullptr;
   }
-
+  MS_ASSERT(desc.type == schema::PrimitiveType_Concat);
+  auto *kernel = new(std::nothrow) ConcatCPUKernel(opParameter, inputs, outputs, ctx);
   if (kernel == nullptr) {
-    MS_LOG(ERROR) << "kernel is nullptr.";
+    MS_LOG(ERROR) << "new ConcatCPUKernel fail!";
     return nullptr;
   }
   auto ret = kernel->Init();
@@ -102,6 +106,8 @@ kernel::LiteKernel *CpuConcatKernelCreator(const std::vector<lite::tensor::Tenso
   return kernel;
 }
 
-REG_KERNEL(kCPU, PrimitiveType_Concat, CpuConcatKernelCreator)
+REG_KERNEL(kCPU, kNumberTypeInt8, PrimitiveType_Concat, CpuConcatInt8KernelCreator)
+REG_KERNEL(kCPU, kNumberTypeInt32, PrimitiveType_Concat, CpuConcatInt32KernelCreator)
+REG_KERNEL(kCPU, kNumberTypeFloat32, PrimitiveType_Concat, CpuConcatFp32KernelCreator)
 }  // namespace mindspore::kernel
 

mindspore/lite/src/runtime/kernel/arm/base/convolution_base.cc

@@ -15,24 +15,6 @@
  */
 
 #include "src/runtime/kernel/arm/base/convolution_base.h"
-#include "src/runtime/kernel/arm/fp32/convolution.h"
-#include "src/runtime/kernel/arm/fp32/convolution_winograd.h"
-#include "src/runtime/kernel/arm/fp32/deconvolution.h"
-#include "src/runtime/kernel/arm/fp32/convolution_1x1.h"
-#include "src/runtime/kernel/arm/fp32/convolution_3x3.h"
-#include "src/runtime/kernel/arm/fp32/convolution_depthwise.h"
-#include "src/runtime/kernel/arm/fp32/deconvolution_depthwise.h"
-#ifdef ENABLE_FP16
-#include "src/runtime/kernel/arm/fp16/convolution_fp16.h"
-#include "src/runtime/kernel/arm/fp16/convolution_3x3_fp16.h"
-#include "src/runtime/kernel/arm/fp16/convolution_depthwise_fp16.h"
-#include "src/runtime/kernel/arm/fp16/deconvolution_depthwise_fp16.h"
-#endif
-#include "src/runtime/kernel/arm/int8/deconvolution_int8.h"
-#include "src/runtime/kernel/arm/int8/convolution_int8.h"
-#include "src/runtime/kernel/arm/int8/convolution_3x3_int8.h"
-#include "src/runtime/kernel/arm/int8/convolution_depthwise_int8.h"
-#include "src/runtime/kernel/arm/int8/deconvolution_depthwise_int8.h"
 #include "schema/model_generated.h"
 #include "src/kernel_factory.h"
 #include "include/errorcode.h"
@@ -42,10 +24,6 @@ using mindspore::lite::RET_ERROR;
 using mindspore::lite::RET_OK;
 using mindspore::schema::ActivationType;
 using mindspore::schema::PadMode;
-using mindspore::schema::PrimitiveType_Conv2D;
-using mindspore::schema::PrimitiveType_DeConv2D;
-using mindspore::schema::PrimitiveType_DeDepthwiseConv2D;
-using mindspore::schema::PrimitiveType_DepthwiseConv2D;
 
 namespace mindspore::kernel {
 ConvolutionBaseCPUKernel::~ConvolutionBaseCPUKernel() {
@@ -192,352 +170,4 @@ void ComputeQuantOutRange(ConvParameter *conv_param) {
   conv_param->conv_quant_arg_.out_act_min_[0] = min;
   conv_param->conv_quant_arg_.out_act_max_[0] = max;
 }
-
-void CheckIfUseWinograd(bool *use_winograd, int *output_unit, ConvParameter *conv_param,
-                        InputTransformUnitFunc input_trans_func, OutputTransformUnitFunc output_trans_func) {
-  if (conv_param->kernel_w_ == conv_param->kernel_h_ && conv_param->dilation_h_ == 1 && conv_param->dilation_w_ == 1 &&
-      conv_param->stride_h_ == 1 && conv_param->stride_w_ == 1) {
-    *output_unit = SelectOutputUnit(conv_param);
-    if (*output_unit > 1) {
-      *use_winograd = true;
-      int input_unit = conv_param->kernel_h_ + *output_unit - 1;
-      input_trans_func = GetInputTransFunc(input_unit);
-      if (input_trans_func == nullptr) {
-        MS_LOG(INFO) << "No matching input trans func. Turn back to common conv.";
-        *use_winograd = false;
-      }
-      output_trans_func = GetOutputTransFunc(input_unit, *output_unit);
-      if (output_trans_func == nullptr) {
-        MS_LOG(INFO) << "No matching output trans func. Turn back to common conv.";
-        *use_winograd = false;
-      }
-    } else {
-      *use_winograd = false;
-    }
-  } else {
-    *use_winograd = false;
-  }
-}
-
-bool CheckSupportFP16() {
-  bool support_fp16 = false;
-#ifdef ENABLE_ARM64
-  void *optimize_op_handler = OptimizeModule::GetInstance()->optimized_op_handler_;
-  if (optimize_op_handler != nullptr) {
-    support_fp16 = true;
-    MS_LOG(INFO) << "Support FP16.";
-  } else {
-    support_fp16 = false;
-    MS_LOG(INFO) << "Your machine doesn't support fp16, return back to float32 kernel.";
-  }
-#endif
-  return support_fp16;
-}
-
-kernel::LiteKernel *CpuConvFloatKernelCreator(const std::vector<lite::tensor::Tensor *> &inputs,
-                                              const std::vector<lite::tensor::Tensor *> &outputs,
-                                              OpParameter *opParameter, const Context *ctx) {
-  auto conv_param = reinterpret_cast<ConvParameter *>(opParameter);
-  int kernel_h = conv_param->kernel_h_;
-  int kernel_w = conv_param->kernel_w_;
-  int stride_h = conv_param->stride_h_;
-  int stride_w = conv_param->stride_w_;
-  int dilation_h = conv_param->dilation_h_;
-  int dilation_w = conv_param->dilation_w_;
-  conv_param->input_h_ = inputs.front()->Height();
-  conv_param->input_w_ = inputs.front()->Width();
-  conv_param->output_h_ = outputs.front()->Height();
-  conv_param->output_w_ = outputs.front()->Width();
-  bool use_winograd;
-  int out_unit;
-  InputTransformUnitFunc input_trans_func = nullptr;
-  OutputTransformUnitFunc output_trans_func = nullptr;
-  CheckIfUseWinograd(&use_winograd, &out_unit, conv_param, input_trans_func, output_trans_func);
-  bool support_fp16 = CheckSupportFP16();
-
-  if (kernel_h == 1 && kernel_w == 1) {
-    auto kernel = new (std::nothrow) Convolution1x1CPUKernel(opParameter, inputs, outputs, ctx);
-    return kernel;
-  } else if (kernel_h == 3 && kernel_w == 3 && stride_h == 1 && stride_w == 1 && dilation_h == 1 && dilation_w == 1) {
-    if (support_fp16) {
-#ifdef ENABLE_FP16
-      auto kernel = new (std::nothrow) Convolution3x3FP16CPUKernel(opParameter, inputs, outputs, ctx);
-      return kernel;
-#endif
-    }
-    auto kernel = new (std::nothrow) Convolution3x3CPUKernel(opParameter, inputs, outputs, ctx);
-    return kernel;
-  } else if (use_winograd) {
-    auto kernel = new (std::nothrow) ConvolutionWinogradCPUKernel(opParameter, inputs, outputs, ctx, out_unit);
-    return kernel;
-  } else {
-    if (support_fp16) {
-#ifdef ENABLE_FP16
-      auto kernel = new (std::nothrow) ConvolutionFP16CPUKernel(opParameter, inputs, outputs, ctx);
-      return kernel;
-#endif
-    }
-    auto kernel = new (std::nothrow) ConvolutionCPUKernel(opParameter, inputs, outputs, ctx);
-    return kernel;
-  }
-}
-
-kernel::LiteKernel *CpuConvInt8KernelCreator(const std::vector<lite::tensor::Tensor *> &inputs,
-                                             const std::vector<lite::tensor::Tensor *> &outputs,
-                                             OpParameter *opParameter, const Context *ctx) {
-  auto conv_param = reinterpret_cast<ConvParameter *>(opParameter);
-  int kernel_h = conv_param->kernel_h_;
-  int kernel_w = conv_param->kernel_w_;
-  int stride_h = conv_param->stride_h_;
-  int stride_w = conv_param->stride_w_;
-  int dilation_h = conv_param->dilation_h_;
-  int dilation_w = conv_param->dilation_w_;
-
-  if (kernel_h == 3 && kernel_w == 3 && stride_h == 1 && stride_w == 1 && dilation_h == 1 && dilation_w == 1) {
-    auto kernel = new (std::nothrow) Convolution3x3Int8CPUKernel(opParameter, inputs, outputs, ctx);
-    return kernel;
-  } else {
-    auto kernel = new (std::nothrow) ConvolutionInt8CPUKernel(opParameter, inputs, outputs, ctx);
-    return kernel;
-  }
-}
-
-kernel::LiteKernel *CpuConvKernelCreator(const std::vector<lite::tensor::Tensor *> &inputs,
-                                         const std::vector<lite::tensor::Tensor *> &outputs, OpParameter *opParameter,
-                                         const lite::Context *ctx, const kernel::KernelKey &desc) {
-  MS_ASSERT(opParameter != nullptr);
-  MS_ASSERT(desc.type == schema::PrimitiveType_Conv2D);
-  auto input_tensor = inputs.at(kInputIndex);
-  auto data_type = input_tensor->data_type();
-  kernel::LiteKernel *kernel = nullptr;
-  switch (data_type) {
-    case kNumberTypeInt8:
-      kernel = CpuConvInt8KernelCreator(inputs, outputs, opParameter, ctx);
-      break;
-    case kNumberTypeFloat32:
-      kernel = CpuConvFloatKernelCreator(inputs, outputs, opParameter, ctx);
-      break;
-    default:
-      break;
-  }
-
-  if (kernel == nullptr) {
-    MS_LOG(ERROR) << "kernel is nullptr.";
-    return nullptr;
-  }
-  auto ret = kernel->Init();
-  if (ret != RET_OK) {
-    delete kernel;
-    MS_LOG(ERROR) << "Init kernel failed, name: " << opParameter->name_ << ", type: "
-                  << schema::EnumNamePrimitiveType(static_cast<schema::PrimitiveType>(opParameter->type_));
-    return nullptr;
-  }
-  return kernel;
-}
-
-kernel::LiteKernel *CpuConvDwFp32KernelCreator(const std::vector<lite::tensor::Tensor *> &inputs,
-                                               const std::vector<lite::tensor::Tensor *> &outputs,
-                                               OpParameter *opParameter, const Context *ctx) {
-  auto kernel = new (std::nothrow) ConvolutionDepthwiseCPUKernel(opParameter, inputs, outputs, ctx);
-  if (kernel == nullptr) {
-    MS_LOG(ERROR) << "kernel is nullptr.";
-    return nullptr;
-  }
-  return kernel;
-}
-
-#ifdef ENABLE_FP16
-kernel::LiteKernel *CpuConvDwFp16KernelCreator(const std::vector<lite::tensor::Tensor *> &inputs,
-                                               const std::vector<lite::tensor::Tensor *> &outputs,
-                                               OpParameter *opParameter, const Context *ctx) {
-  auto kernel = new (std::nothrow) ConvolutionDepthwiseFp16CPUKernel(opParameter, inputs, outputs, ctx);
-  if (kernel == nullptr) {
-    MS_LOG(ERROR) << "kernel is nullptr.";
-    return nullptr;
-  }
-  return kernel;
-}
-#endif
-
-kernel::LiteKernel *CpuConvDwInt8KernelCreator(const std::vector<lite::tensor::Tensor *> &inputs,
-                                               const std::vector<lite::tensor::Tensor *> &outputs,
-                                               OpParameter *opParameter, const Context *ctx) {
-  auto kernel = new (std::nothrow) ConvolutionDepthwiseInt8CPUKernel(opParameter, inputs, outputs, ctx);
-  if (kernel == nullptr) {
-    MS_LOG(ERROR) << "kernel is nullptr.";
-    return nullptr;
-  }
-  return kernel;
-}
-
-kernel::LiteKernel *CpuConvDwKernelCreator(const std::vector<lite::tensor::Tensor *> &inputs,
-                                           const std::vector<lite::tensor::Tensor *> &outputs, OpParameter *opParameter,
-                                           const lite::Context *ctx, const kernel::KernelKey &desc) {
-  MS_ASSERT(opParameter != nullptr);
-  MS_ASSERT(desc.type == schema::PrimitiveType_DepthwiseConv2D);
-  auto input_tensor = inputs.at(kInputIndex);
-  auto data_type = input_tensor->data_type();
-  kernel::LiteKernel *kernel = nullptr;
-  switch (data_type) {
-    case kNumberTypeInt8:
-      kernel = CpuConvDwInt8KernelCreator(inputs, outputs, opParameter, ctx);
-      break;
-    case kNumberTypeFloat32:
-#ifdef ENABLE_FP16
-      kernel = CpuConvDwFp16KernelCreator(inputs, outputs, opParameter, ctx);
-#else
-      kernel = CpuConvDwFp32KernelCreator(inputs, outputs, opParameter, ctx);
-#endif
-      break;
-    default:
-      break;
-  }
-  if (kernel == nullptr) {
-    MS_LOG(ERROR) << "kernel is nullptr.";
-    return nullptr;
-  }
-  auto ret = kernel->Init();
-  if (ret != RET_OK) {
-    delete kernel;
-    MS_LOG(ERROR) << "Init kernel failed, name: " << opParameter->name_ << ", type: "
-                  << schema::EnumNamePrimitiveType(static_cast<schema::PrimitiveType>(opParameter->type_));
-    return nullptr;
-  }
-  return kernel;
-}
-
-kernel::LiteKernel *CpuDeconvDwFp32KernelCreator(const std::vector<lite::tensor::Tensor *> &inputs,
-                                                 const std::vector<lite::tensor::Tensor *> &outputs,
-                                                 OpParameter *opParameter, const lite::Context *ctx) {
-  auto kernel = new (std::nothrow) DeconvolutionDepthwiseCPUKernel(opParameter, inputs, outputs, ctx);
-  if (kernel == nullptr) {
-    MS_LOG(ERROR) << "kernel is nullptr.";
-    return nullptr;
-  }
-  return kernel;
-}
-
-#ifdef ENABLE_FP16
-kernel::LiteKernel *CpuDeconvDwFp16KernelCreator(const std::vector<lite::tensor::Tensor *> &inputs,
-                                                 const std::vector<lite::tensor::Tensor *> &outputs,
-                                                 OpParameter *opParameter, const lite::Context *ctx) {
-  auto kernel = new (std::nothrow) DeconvolutionDepthwiseFp16CPUKernel(opParameter, inputs, outputs, ctx);
-  if (kernel == nullptr) {
-    MS_LOG(ERROR) << "kernel is nullptr.";
-    return nullptr;
-  }
-  return kernel;
-}
-#endif
-
-kernel::LiteKernel *CpuDeconvDwInt8KernelCreator(const std::vector<lite::tensor::Tensor *> &inputs,
-                                                 const std::vector<lite::tensor::Tensor *> &outputs,
-                                                 OpParameter *opParameter, const lite::Context *ctx) {
-  auto kernel = new (std::nothrow) DeconvolutionDepthwiseInt8CPUKernel(opParameter, inputs, outputs, ctx);
-  if (kernel == nullptr) {
-    MS_LOG(ERROR) << "kernel is nullptr.";
-    return nullptr;
-  }
-  return kernel;
-}
-
-kernel::LiteKernel *CpuDeconvDwKernelCreator(const std::vector<lite::tensor::Tensor *> &inputs,
-                                             const std::vector<lite::tensor::Tensor *> &outputs,
-                                             OpParameter *opParameter, const lite::Context *ctx,
-                                             const kernel::KernelKey &desc) {
-  MS_ASSERT(opParameter != nullptr);
-  MS_ASSERT(desc.type == schema::PrimitiveType_DepthwiseConv2D);
-  auto input_tensor = inputs.at(kInputIndex);
-  auto data_type = input_tensor->data_type();
-  kernel::LiteKernel *kernel = nullptr;
-  switch (data_type) {
-    case kNumberTypeInt8:
-      kernel = CpuDeconvDwInt8KernelCreator(inputs, outputs, opParameter, ctx);
-      break;
-    case kNumberTypeFloat32:
-#ifdef ENABLE_FP16
-      kernel = CpuDeconvDwFp16KernelCreator(inputs, outputs, opParameter, ctx);
-#else
-      kernel = CpuDeconvDwFp32KernelCreator(inputs, outputs, opParameter, ctx);
-#endif
-      break;
-    default:
-      break;
-  }
-  if (kernel == nullptr) {
-    MS_LOG(ERROR) << "kernel is nullptr.";
-    return nullptr;
-  }
-  auto ret = kernel->Init();
-  if (ret != RET_OK) {
-    delete kernel;
-    MS_LOG(ERROR) << "Init kernel failed, name: " << opParameter->name_ << ", type: "
-                  << schema::EnumNamePrimitiveType(static_cast<schema::PrimitiveType>(opParameter->type_));
-    return nullptr;
-  }
-  return kernel;
-}
-
-kernel::LiteKernel *CpuDeConvFp32KernelCreator(const std::vector<lite::tensor::Tensor *> &inputs,
-                                               const std::vector<lite::tensor::Tensor *> &outputs,
-                                               OpParameter *opParameter, const lite::Context *ctx) {
-  auto kernel = new (std::nothrow) DeConvolutionCPUKernel(opParameter, inputs, outputs, ctx);
-  if (kernel == nullptr) {
-    MS_LOG(ERROR) << "kernel is nullptr.";
-    return nullptr;
-  }
-  return kernel;
-}
-
-kernel::LiteKernel *CpuDeConvInt8KernelCreator(const std::vector<lite::tensor::Tensor *> &inputs,
-                                               const std::vector<lite::tensor::Tensor *> &outputs,
-                                               OpParameter *opParameter, const lite::Context *ctx) {
-  auto kernel = new (std::nothrow) DeConvInt8CPUKernel(opParameter, inputs, outputs, ctx);
-  if (kernel == nullptr) {
-    MS_LOG(ERROR) << "kernel is nullptr.";
-    return nullptr;
-  }
-  return kernel;
-}
-
-kernel::LiteKernel *CpuDeConvKernelCreator(const std::vector<lite::tensor::Tensor *> &inputs,
-                                           const std::vector<lite::tensor::Tensor *> &outputs, OpParameter *opParameter,
-                                           const lite::Context *ctx, const kernel::KernelKey &desc) {
-  MS_ASSERT(opParameter != nullptr);
-  MS_ASSERT(desc.type == schema::PrimitiveType_DeConv2D);
-  auto input_tensor = inputs.at(kInputIndex);
-  auto data_type = input_tensor->data_type();
-  kernel::LiteKernel *kernel = nullptr;
-  switch (data_type) {
-    case kNumberTypeInt8:
-      kernel = CpuDeConvInt8KernelCreator(inputs, outputs, opParameter, ctx);
-      break;
-#ifdef ENABLE_FP16
-    case kNumberTypeFloat16:
-      break;
-#endif
-    case kNumberTypeFloat32:
-      kernel = CpuDeConvFp32KernelCreator(inputs, outputs, opParameter, ctx);
-      break;
-    default:
-      break;
-  }
-  if (kernel == nullptr) {
-    MS_LOG(ERROR) << "kernel is nullptr.";
-    return nullptr;
-  }
-  auto ret = kernel->Init();
-  if (ret != RET_OK) {
-    delete kernel;
-    MS_LOG(ERROR) << "Init kernel failed, name: " << opParameter->name_ << ", type: "
-                  << schema::EnumNamePrimitiveType(static_cast<schema::PrimitiveType>(opParameter->type_));
-    return nullptr;
-  }
-  return kernel;
-}
-
-REG_KERNEL(kCPU, PrimitiveType_Conv2D, CpuConvKernelCreator)
-REG_KERNEL(kCPU, PrimitiveType_DeConv2D, CpuDeConvKernelCreator)
-REG_KERNEL(kCPU, PrimitiveType_DepthwiseConv2D, CpuConvDwKernelCreator)
-REG_KERNEL(kCPU, PrimitiveType_DeDepthwiseConv2D, CpuDeconvDwKernelCreator)
 }  // namespace mindspore::kernel

mindspore/lite/src/runtime/kernel/arm/base/convolution_base.h

@@ -28,7 +28,6 @@
 #include "src/lite_kernel.h"
 #include "include/context.h"
 #include "src/runtime/kernel/arm/base/layout_transform.h"
-#include "src/runtime/kernel/arm/opclib/optimized_kernel.h"
 
 using mindspore::lite::Context;
 using mindspore::schema::PadMode;

mindspore/lite/src/runtime/kernel/arm/base/fullconnection_base.cc

@@ -32,39 +32,38 @@ int FullconnectionBaseCPUKernel::Init() {
   return RET_OK;
 }
 
-kernel::LiteKernel *CpuFullConnectionKernelCreator(const std::vector<lite::tensor::Tensor *> &inputs,
-                                                   const std::vector<lite::tensor::Tensor *> &outputs,
-                                                   OpParameter *opParameter, const lite::Context *ctx,
-                                                   const kernel::KernelKey &desc) {
+kernel::LiteKernel *CpuFullConnectionInt8KernelCreator(const std::vector<lite::tensor::Tensor *> &inputs,
+                                                       const std::vector<lite::tensor::Tensor *> &outputs,
+                                                       OpParameter *opParameter, const lite::Context *ctx,
+                                                       const kernel::KernelKey &desc) {
   MS_ASSERT(opParameter != nullptr);
   MS_ASSERT(desc.type == schema::PrimitiveType_Concat);
-  auto input_tensor = inputs.at(kInputIndex);
-  auto data_type = input_tensor->data_type();
-  kernel::LiteKernel *kernel = nullptr;
-  switch (data_type) {
-    case kNumberTypeInt8:
-    case kNumberTypeUInt8: {
-      kernel = new (std::nothrow) FullconnectionInt8CPUKernel(opParameter, inputs, outputs, ctx);
-      if (!kernel) {
-        MS_LOG(ERROR) << "kernel is nullptr.";
-        return nullptr;
-      }
-      break;
-    }
-
-    case kNumberTypeFloat32: {
-      kernel = new (std::nothrow) FullconnectionCPUKernel(opParameter, inputs, outputs, ctx);
-      if (!kernel) {
-        MS_LOG(ERROR) << "kernel is nullptr.";
-        return nullptr;
-      }
-      break;
-    }
-
-    default:
-      break;
+  auto kernel = new (std::nothrow) FullconnectionInt8CPUKernel(opParameter, inputs, outputs, ctx);
+  if (!kernel) {
+    MS_LOG(ERROR) << "kernel is nullptr.";
+    return nullptr;
+  }
+  auto ret = kernel->Init();
+  if (ret != RET_OK) {
+    delete kernel;
+    MS_LOG(ERROR) << "Init kernel failed, name: " << opParameter->name_ << ", type: "
+                  << schema::EnumNamePrimitiveType(static_cast<schema::PrimitiveType>(opParameter->type_));
+    return nullptr;
   }
+  return kernel;
+}
 
+kernel::LiteKernel *CpuFullConnectionFp32KernelCreator(const std::vector<lite::tensor::Tensor *> &inputs,
+                                                       const std::vector<lite::tensor::Tensor *> &outputs,
+                                                       OpParameter *opParameter, const lite::Context *ctx,
+                                                       const kernel::KernelKey &desc) {
+  MS_ASSERT(opParameter != nullptr);
+  MS_ASSERT(desc.type == schema::PrimitiveType_Concat);
+  auto kernel = new (std::nothrow) FullconnectionCPUKernel(opParameter, inputs, outputs, ctx);
+  if (!kernel) {
+    MS_LOG(ERROR) << "kernel is nullptr.";
+    return nullptr;
+  }
   auto ret = kernel->Init();
   if (ret != RET_OK) {
     delete kernel;
@@ -75,5 +74,6 @@ kernel::LiteKernel *CpuFullConnectionKernelCreator(const std::vector<lite::tenso
   return kernel;
 }
 
-REG_KERNEL(kCPU, PrimitiveType_FullConnection, CpuFullConnectionKernelCreator)
+REG_KERNEL(kCPU, kNumberTypeInt8, PrimitiveType_FullConnection, CpuFullConnectionInt8KernelCreator)
+REG_KERNEL(kCPU, kNumberTypeFloat32, PrimitiveType_FullConnection, CpuFullConnectionFp32KernelCreator)
 }  // namespace mindspore::kernel

mindspore/lite/src/runtime/kernel/arm/base/layout_transform.cc

@@ -15,6 +15,7 @@
  */
 
 #include "src/runtime/kernel/arm/base/layout_transform.h"
+#include "mindspore/core/utils/log_adapter.h"
 
 using mindspore::schema::Format;
 namespace mindspore::kernel {

mindspore/lite/src/runtime/kernel/arm/base/layout_transform.h

@@ -21,7 +21,8 @@
 #include <arm_neon.h>
 #endif
 #include "src/runtime/kernel/arm/opclib/pack.h"
-#include "src/ir/tensor.h"
+#include "ir/dtype/type_id.h"
+#include "schema/ops_generated.h"
 
 namespace mindspore::kernel {
 typedef void (*LayoutConvertor)(const void *src, void *dst, int batch, int plane, int channel);

mindspore/lite/src/runtime/kernel/arm/base/pad.cc

@@ -32,11 +32,20 @@ kernel::LiteKernel *CpuPadInt8KernelCreator(const std::vector<lite::tensor::Tens
                                             const std::vector<lite::tensor::Tensor *> &outputs,
                                             OpParameter *opParameter, const lite::Context *ctx,
                                             const kernel::KernelKey &desc) {
+  MS_ASSERT(opParameter != nullptr);
+  MS_ASSERT(desc.type == schema::PrimitiveType_Pad);
   auto *kernel = new (std::nothrow) PadInt8CPUKernel(opParameter, inputs, outputs, ctx);
   if (kernel == nullptr) {
     MS_LOG(ERROR) << "new PadCPUKernel failed.";
     return nullptr;
   }
+  auto ret = kernel->Init();
+  if (ret != RET_OK) {
+    delete kernel;
+    MS_LOG(ERROR) << "Init kernel failed, name: " << opParameter->name_ << ", type: "
+                  << schema::EnumNamePrimitiveType(static_cast<schema::PrimitiveType>(opParameter->type_));
+    return nullptr;
+  }
   return kernel;
 }
 
@@ -44,38 +53,13 @@ kernel::LiteKernel *CpuPadFp32KernelCreator(const std::vector<lite::tensor::Tens
                                             const std::vector<lite::tensor::Tensor *> &outputs,
                                             OpParameter *opParameter, const lite::Context *ctx,
                                             const kernel::KernelKey &desc) {
+  MS_ASSERT(opParameter != nullptr);
+  MS_ASSERT(desc.type == schema::PrimitiveType_Pad);
   auto *kernel = new (std::nothrow) PadCPUKernel(opParameter, inputs, outputs, ctx);
   if (kernel == nullptr) {
     MS_LOG(ERROR) << "new PadCPUKernel failed.";
     return nullptr;
   }
-  return kernel;
-}
-
-kernel::LiteKernel *CpuPadKernelCreator(const std::vector<lite::tensor::Tensor *> &inputs,
-                                        const std::vector<lite::tensor::Tensor *> &outputs, OpParameter *opParameter,
-                                        const lite::Context *ctx, const kernel::KernelKey &desc) {
-  MS_ASSERT(opParameter != nullptr);
-  MS_ASSERT(desc.type == schema::PrimitiveType_Concat);
-  auto input_tensor = inputs.at(kInputIndex);
-  auto data_type = input_tensor->data_type();
-  kernel::LiteKernel *kernel = nullptr;
-  switch (data_type) {
-    case kNumberTypeInt8:
-      kernel = CpuPadInt8KernelCreator(inputs, outputs, opParameter, ctx, desc);
-      break;
-    case kNumberTypeFloat32:
-      kernel = CpuPadFp32KernelCreator(inputs, outputs, opParameter, ctx, desc);
-      break;
-    default:
-      break;
-  }
-
-  if (kernel == nullptr) {
-    MS_LOG(ERROR) << "kernel is nullptr.";
-    return nullptr;
-  }
-
   auto ret = kernel->Init();
   if (ret != RET_OK) {
     delete kernel;
@@ -86,5 +70,6 @@ kernel::LiteKernel *CpuPadKernelCreator(const std::vector<lite::tensor::Tensor *
   return kernel;
 }
 
-REG_KERNEL(kCPU, PrimitiveType_Pad, CpuPadKernelCreator)
+REG_KERNEL(kCPU, kNumberTypeInt8, PrimitiveType_Pad, CpuPadInt8KernelCreator)
+REG_KERNEL(kCPU, kNumberTypeFloat32, PrimitiveType_Pad, CpuPadFp32KernelCreator)
 }  // namespace mindspore::kernel

mindspore/lite/src/runtime/kernel/arm/base/pooling_base.cc

@@ -81,7 +81,8 @@ int PoolingBaseCPUKernel::Init() {
 
 kernel::LiteKernel *CpuPoolingInt8KernelCreator(const std::vector<lite::tensor::Tensor *> &inputs,
                                                 const std::vector<lite::tensor::Tensor *> &outputs,
-                                                OpParameter *opParameter, const Context *ctx) {
+                                                OpParameter *opParameter, const Context *ctx,
+                                                const kernel::KernelKey &desc) {
   if (opParameter == nullptr) {
     MS_LOG(ERROR) << "Input opParameter is nullptr!";
     return nullptr;
@@ -92,12 +93,20 @@ kernel::LiteKernel *CpuPoolingInt8KernelCreator(const std::vector<lite::tensor::
     MS_LOG(ERROR) << "new PoolingInt8CPUKernel fail!";
     return nullptr;
   }
+  auto ret = kernel->Init();
+  if (ret != RET_OK) {
+    delete kernel;
+    MS_LOG(ERROR) << "Init kernel failed, name: " << opParameter->name_ << ", type: "
+                  << schema::EnumNamePrimitiveType(static_cast<schema::PrimitiveType>(opParameter->type_));
+    return nullptr;
+  }
   return kernel;
 }
 
 kernel::LiteKernel *CpuPoolingFp32KernelCreator(const std::vector<lite::tensor::Tensor *> &inputs,
                                                 const std::vector<lite::tensor::Tensor *> &outputs,
-                                                OpParameter *opParameter, const Context *ctx) {
+                                                OpParameter *opParameter, const Context *ctx,
+                                                const kernel::KernelKey &desc) {
   if (opParameter == nullptr) {
     MS_LOG(ERROR) << "Input opParameter is nullptr!";
     return nullptr;
@@ -108,34 +117,6 @@ kernel::LiteKernel *CpuPoolingFp32KernelCreator(const std::vector<lite::tensor::
     MS_LOG(ERROR) << "new PoolingCPUKernel fail!";
     return nullptr;
   }
-  return kernel;
-}
-
-kernel::LiteKernel *CpuPoolingKernelCreator(const std::vector<lite::tensor::Tensor *> &inputs,
-                                            const std::vector<lite::tensor::Tensor *> &outputs,
-                                            OpParameter *opParameter, const lite::Context *ctx,
-                                            const kernel::KernelKey &desc) {
-  MS_ASSERT(opParameter != nullptr);
-  MS_ASSERT(desc.type == schema::PrimitiveType_Pooing);
-  auto input_tensor = inputs.at(kInputIndex);
-  auto data_type = input_tensor->data_type();
-  kernel::LiteKernel *kernel = nullptr;
-  switch (data_type) {
-    case kNumberTypeInt8:
-    case kNumberTypeUInt8:
-      kernel = CpuPoolingInt8KernelCreator(inputs, outputs, opParameter, ctx);
-      break;
-    case kNumberTypeFloat32:
-      kernel = CpuPoolingFp32KernelCreator(inputs, outputs, opParameter, ctx);
-      break;
-    default:
-      break;
-  }
-
-  if (kernel == nullptr) {
-    MS_LOG(ERROR) << "kernel is nullptr.";
-    return nullptr;
-  }
   auto ret = kernel->Init();
   if (ret != RET_OK) {
     delete kernel;
@@ -146,5 +127,6 @@ kernel::LiteKernel *CpuPoolingKernelCreator(const std::vector<lite::tensor::Tens
   return kernel;
 }
 
-REG_KERNEL(kCPU, PrimitiveType_Pooling, CpuPoolingKernelCreator)
+REG_KERNEL(kCPU, kNumberTypeInt8, PrimitiveType_Pooling, CpuPoolingInt8KernelCreator)
+REG_KERNEL(kCPU, kNumberTypeFloat32, PrimitiveType_Pooling, CpuPoolingFp32KernelCreator)
 }  // namespace mindspore::kernel

mindspore/lite/src/runtime/kernel/arm/base/reshape_base.cc

@@ -35,60 +35,64 @@ int ReshapeBaseCPUKernel::Init() {
 
 kernel::LiteKernel *CpuReshapeInt8KernelCreator(const std::vector<lite::tensor::Tensor *> &inputs,
                                                 const std::vector<lite::tensor::Tensor *> &outputs,
-                                                OpParameter *opParameter, const Context *ctx) {
+                                                OpParameter *opParameter, const Context *ctx,
+                                                const kernel::KernelKey &desc) {
   if (opParameter == nullptr) {
     MS_LOG(ERROR) << "Input opParameter is nullptr!";
     return nullptr;
   }
-  MS_ASSERT(desc.type == schema::PrimitiveType_Concat);
-  auto *kernel = new(std::nothrow) ReshapeInt8CPUKernel(opParameter, inputs, outputs, ctx);
+  MS_ASSERT(desc.type == schema::PrimitiveType_Reshape);
+  auto *kernel = new (std::nothrow) ReshapeInt8CPUKernel(opParameter, inputs, outputs, ctx);
   if (kernel == nullptr) {
     MS_LOG(ERROR) << "new ConcatCPUKernel fail!";
     return nullptr;
   }
+  auto ret = kernel->Init();
+  if (ret != RET_OK) {
+    delete kernel;
+    MS_LOG(ERROR) << "Init kernel failed, name: " << opParameter->name_ << ", type: "
+                  << schema::EnumNamePrimitiveType(static_cast<schema::PrimitiveType>(opParameter->type_));
+    return nullptr;
+  }
   return kernel;
 }
 
-kernel::LiteKernel *CpuReshapeFp32OrInt32KernelCreator(const std::vector<lite::tensor::Tensor *> &inputs,
-                                                       const std::vector<lite::tensor::Tensor *> &outputs,
-                                                       OpParameter *opParameter, const Context *ctx) {
+kernel::LiteKernel *CpuReshapeInt32KernelCreator(const std::vector<lite::tensor::Tensor *> &inputs,
+                                                 const std::vector<lite::tensor::Tensor *> &outputs,
+                                                 OpParameter *opParameter, const Context *ctx,
+                                                 const kernel::KernelKey &desc) {
   if (opParameter == nullptr) {
     MS_LOG(ERROR) << "Input opParameter is nullptr!";
     return nullptr;
   }
-  MS_ASSERT(desc.type == schema::PrimitiveType_Concat);
-  auto *kernel = new(std::nothrow) ReshapeCPUKernel(opParameter, inputs, outputs, ctx);
+  MS_ASSERT(desc.type == schema::PrimitiveType_Reshape);
+  auto *kernel = new (std::nothrow) ReshapeCPUKernel(opParameter, inputs, outputs, ctx);
   if (kernel == nullptr) {
     MS_LOG(ERROR) << "new ConcatCPUKernel fail!";
     return nullptr;
   }
+  auto ret = kernel->Init();
+  if (ret != RET_OK) {
+    delete kernel;
+    MS_LOG(ERROR) << "Init kernel failed, name: " << opParameter->name_ << ", type: "
+                  << schema::EnumNamePrimitiveType(static_cast<schema::PrimitiveType>(opParameter->type_));
+    return nullptr;
+  }
   return kernel;
 }
 
-kernel::LiteKernel *CpuReshapeKernelCreator(const std::vector<lite::tensor::Tensor *> &inputs,
-                                            const std::vector<lite::tensor::Tensor *> &outputs,
-                                            OpParameter *opParameter, const lite::Context *ctx,
-                                            const kernel::KernelKey &desc) {
-  MS_ASSERT(opParameter != nullptr);
-  MS_ASSERT(desc.type == schema::PrimitiveType_Concat);
-  auto input_tensor = inputs.at(kInputIndex);
-  auto data_type = input_tensor->data_type();
-  kernel::LiteKernel *kernel = nullptr;
-  switch (data_type) {
-    case kNumberTypeInt8:
-    case kNumberTypeUInt8:
-      kernel = CpuReshapeInt8KernelCreator(inputs, outputs, opParameter, ctx);
-      break;
-    case kNumberTypeInt32:
-    case kNumberTypeFloat32:
-      kernel = CpuReshapeFp32OrInt32KernelCreator(inputs, outputs, opParameter, ctx);
-      break;
-    default:
-      break;
+kernel::LiteKernel *CpuReshapeFp32KernelCreator(const std::vector<lite::tensor::Tensor *> &inputs,
+                                                const std::vector<lite::tensor::Tensor *> &outputs,
+                                                OpParameter *opParameter, const Context *ctx,
+                                                const kernel::KernelKey &desc) {
+  if (opParameter == nullptr) {
+    MS_LOG(ERROR) << "Input opParameter is nullptr!";
+    return nullptr;
   }
-
+  MS_ASSERT(desc.type == schema::PrimitiveType_Reshape);
+  auto *kernel = new (std::nothrow) ReshapeCPUKernel(opParameter, inputs, outputs, ctx);
   if (kernel == nullptr) {
-    MS_LOG(ERROR) << "kernel is nullptr.";
+    MS_LOG(ERROR) << "new ConcatCPUKernel fail!";
     return nullptr;
   }
   auto ret = kernel->Init();
@@ -101,6 +105,7 @@ kernel::LiteKernel *CpuReshapeKernelCreator(const std::vector<lite::tensor::Tens
   return kernel;
 }
 
-REG_KERNEL(kCPU, PrimitiveType_Reshape, CpuReshapeKernelCreator)
+REG_KERNEL(kCPU, kNumberTypeInt8, PrimitiveType_Reshape, CpuReshapeInt8KernelCreator)
+REG_KERNEL(kCPU, kNumberTypeInt32, PrimitiveType_Reshape, CpuReshapeInt32KernelCreator)
+REG_KERNEL(kCPU, kNumberTypeFloat32, PrimitiveType_Reshape, CpuReshapeFp32KernelCreator)
 }  // namespace mindspore::kernel
-

mindspore/lite/src/runtime/kernel/arm/fp16/convolution_3x3_fp16.cc

@@ -16,6 +16,8 @@
 
 #include "src/runtime/kernel/arm/fp16/convolution_3x3_fp16.h"
 #include "src/runtime/kernel/arm/opclib/fp16/conv_fp16.h"
+#include "src/runtime/kernel/arm/opclib/fp16/winograd_transform_fp16.h"
+#include "src/runtime/kernel/arm/opclib/fp16/pack_fp16.h"
 #include "src/runtime/kernel/arm/base/layout_transform.h"
 #include "schema/model_generated.h"
 #include "src/kernel_registry.h"
@@ -165,7 +167,6 @@ void Convolution3x3FP16CPUKernel::ConfigInputOutput() {
 }
 
 int Convolution3x3FP16CPUKernel::Init() {
-  ConvolutionBaseCPUKernel::Init();
   auto ret = ConvolutionBaseCPUKernel::Init();
   if (ret != RET_OK) {
     MS_LOG(ERROR) << "ConvolutionBase init failed.";

mindspore/lite/src/runtime/kernel/arm/fp16/convolution_3x3_fp16.h

@@ -20,8 +20,6 @@
 #include <arm_neon.h>
 #include <vector>
 #include "src/lite_kernel.h"
-
-#include "src/runtime/kernel/arm/opclib/winograd_transform.h"
 #include "src/runtime/kernel/arm/base/convolution_base.h"
 #include "src/runtime/kernel/arm/opclib/optimized_kernel.h"
 

mindspore/lite/src/runtime/kernel/arm/fp16/convolution_depthwise_fp16.cc

@@ -15,6 +15,7 @@
  */
 
 #include "src/runtime/kernel/arm/fp16/convolution_depthwise_fp16.h"
+#include "src/runtime/kernel/arm/opclib/fp16/pack_fp16.h"
 #include "schema/model_generated.h"
 #include "src/kernel_registry.h"
 #include "include/errorcode.h"
@@ -161,4 +162,27 @@ int ConvolutionDepthwiseFp16CPUKernel::Run() {
 
   return RET_OK;
 }
+
+kernel::LiteKernel *CpuConvDwFp16KernelCreator(const std::vector<lite::tensor::Tensor *> &inputs,
+                                               const std::vector<lite::tensor::Tensor *> &outputs,
+                                               OpParameter *opParameter, const Context *ctx,
+                                               const kernel::KernelKey &desc) {
+  MS_ASSERT(opParameter != nullptr);
+  MS_ASSERT(desc.type == schema::PrimitiveType_DepthwiseConv2D);
+  auto kernel = new (std::nothrow) ConvolutionDepthwiseFp16CPUKernel(opParameter, inputs, outputs, ctx);
+  if (kernel == nullptr) {
+    MS_LOG(ERROR) << "kernel is nullptr.";
+    return nullptr;
+  }
+  auto ret = kernel->Init();
+  if (ret != RET_OK) {
+    delete kernel;
+    MS_LOG(ERROR) << "Init kernel failed, name: " << opParameter->name_ << ", type: "
+                  << schema::EnumNamePrimitiveType(static_cast<schema::PrimitiveType>(opParameter->type_));
+    return nullptr;
+  }
+  return kernel;
+}
+
+REG_KERNEL(kCPU, kNumberTypeFloat16, PrimitiveType_DepthwiseConv2D, CpuConvDwFp16KernelCreator)
 }  // namespace mindspore::kernel

mindspore/lite/src/runtime/kernel/arm/fp16/convolution_fp16.cc

@@ -15,7 +15,9 @@
  */
 
 #include "src/runtime/kernel/arm/fp16/convolution_fp16.h"
+#include "src/runtime/kernel/arm/fp16/convolution_3x3_fp16.h"
 #include "src/runtime/kernel/arm/opclib/fp16/conv_fp16.h"
+#include "src/runtime/kernel/arm/opclib/fp16/pack_fp16.h"
 #include "src/runtime/kernel/arm/base/layout_transform.h"
 #include "schema/model_generated.h"
 #include "src/kernel_registry.h"
@@ -218,5 +220,42 @@ int ConvolutionFP16CPUKernel::Run() {
   }
   return RET_OK;
 }
-}  // namespace mindspore::kernel
 
+kernel::LiteKernel *CpuConvFp16KernelCreator(const std::vector<lite::tensor::Tensor *> &inputs,
+                                             const std::vector<lite::tensor::Tensor *> &outputs,
+                                             OpParameter *opParameter, const Context *ctx,
+                                             const kernel::KernelKey &desc) {
+  MS_ASSERT(opParameter != nullptr);
+  MS_ASSERT(desc.type == schema::PrimitiveType_Conv2D);
+  auto conv_param = reinterpret_cast<ConvParameter *>(opParameter);
+  int kernel_h = conv_param->kernel_h_;
+  int kernel_w = conv_param->kernel_w_;
+  int stride_h = conv_param->stride_h_;
+  int stride_w = conv_param->stride_w_;
+  int dilation_h = conv_param->dilation_h_;
+  int dilation_w = conv_param->dilation_w_;
+  conv_param->input_h_ = inputs.front()->Height();
+  conv_param->input_w_ = inputs.front()->Width();
+  conv_param->output_h_ = outputs.front()->Height();
+  conv_param->output_w_ = outputs.front()->Width();
+  kernel::LiteKernel *kernel;
+  if (kernel_h == 3 && kernel_w == 3 && stride_h == 1 && stride_w == 1 && dilation_h == 1 && dilation_w == 1) {
+    kernel = new (std::nothrow) kernel::Convolution3x3FP16CPUKernel(opParameter, inputs, outputs, ctx);
+  } else {
+    kernel = new (std::nothrow) kernel::ConvolutionFP16CPUKernel(opParameter, inputs, outputs, ctx);
+  }
+  if (kernel == nullptr) {
+    MS_LOG(ERROR) << "Create conv fp16 kernel failed.";
+    return nullptr;
+  }
+  auto ret = kernel->Init();
+  if (ret != RET_OK) {
+    delete kernel;
+    MS_LOG(ERROR) << "Init kernel failed, name: " << opParameter->name_ << ", type: "
+                  << schema::EnumNamePrimitiveType(static_cast<schema::PrimitiveType>(opParameter->type_));
+    return nullptr;
+  }
+  return kernel;
+}
+REG_KERNEL(kCPU, kNumberTypeFloat16, PrimitiveType_Conv2D, CpuConvFp16KernelCreator)
+}  // namespace mindspore::kernel

mindspore/lite/src/runtime/kernel/arm/fp16/deconvolution_depthwise_fp16.cc

@@ -15,6 +15,7 @@
  */
 
 #include "src/runtime/kernel/arm/fp16/deconvolution_depthwise_fp16.h"
+#include "src/runtime/kernel/arm/opclib/fp16/pack_fp16.h"
 #include "schema/model_generated.h"
 #include "src/kernel_registry.h"
 #include "include/errorcode.h"
@@ -24,7 +25,7 @@ using mindspore::kernel::KERNEL_ARCH::kCPU;
 using mindspore::lite::KernelRegistrar;
 using mindspore::lite::RET_ERROR;
 using mindspore::lite::RET_OK;
-using mindspore::schema::PrimitiveType_DepthwiseConv2D;
+using mindspore::schema::PrimitiveType_DeDepthwiseConv2D;
 
 namespace mindspore::kernel {
 int DeconvolutionDepthwiseFp16CPUKernel::InitSlideParam() {
@@ -171,4 +172,27 @@ int DeconvolutionDepthwiseFp16CPUKernel::Run() {
                           conv_param_->output_h_ * conv_param_->output_w_, conv_param_->output_channel_);
   return RET_OK;
 }
+
+kernel::LiteKernel *CpuDeconvDwFp16KernelCreator(const std::vector<lite::tensor::Tensor *> &inputs,
+                                                 const std::vector<lite::tensor::Tensor *> &outputs,
+                                                 OpParameter *opParameter, const lite::Context *ctx,
+                                                 const kernel::KernelKey &desc) {
+  MS_ASSERT(opParameter != nullptr);
+  MS_ASSERT(desc.type == schema::PrimitiveType_DeDepthwiseConv2D);
+  auto kernel = new (std::nothrow) DeconvolutionDepthwiseFp16CPUKernel(opParameter, inputs, outputs, ctx);
+  if (kernel == nullptr) {
+    MS_LOG(ERROR) << "kernel is nullptr.";
+    return nullptr;
+  }
+  auto ret = kernel->Init();
+  if (ret != RET_OK) {
+    delete kernel;
+    MS_LOG(ERROR) << "Init kernel failed, name: " << opParameter->name_ << ", type: "
+                  << schema::EnumNamePrimitiveType(static_cast<schema::PrimitiveType>(opParameter->type_));
+    return nullptr;
+  }
+  return kernel;
+}
+
+REG_KERNEL(kCPU, kNumberTypeFloat16, PrimitiveType_DeDepthwiseConv2D, CpuDeconvDwFp16KernelCreator)
 }  // namespace mindspore::kernel

mindspore/lite/src/runtime/kernel/arm/fp32/activation.cc

@@ -93,6 +93,10 @@ kernel::LiteKernel *CpuActivationFp32KernelCreator(const std::vector<lite::tenso
   MS_ASSERT(opParameter != nullptr);
   MS_ASSERT(desc.type == schema::PrimitiveType_Activation);
   auto *kernel = new (std::nothrow) ActivationCPUKernel(opParameter, inputs, outputs, ctx);
+  if (kernel == nullptr) {
+    MS_LOG(ERROR) << "kernel is nullptr.";
+    return nullptr;
+  }
   auto ret = kernel->Init();
   if (ret != RET_OK) {
     MS_LOG(ERROR) << "Init kernel failed, name: " << opParameter->name_ << ", type: "
@@ -101,6 +105,5 @@ kernel::LiteKernel *CpuActivationFp32KernelCreator(const std::vector<lite::tenso
   return kernel;
 }
 
-REG_KERNEL(kCPU, PrimitiveType_Activation, CpuActivationFp32KernelCreator)
+REG_KERNEL(kCPU, kNumberTypeFloat32, PrimitiveType_Activation, CpuActivationFp32KernelCreator)
 }  // namespace mindspore::kernel
-

mindspore/lite/src/runtime/kernel/arm/fp32/addn.cc

@@ -100,7 +100,7 @@ kernel::LiteKernel *CpuAddNFp32KernelCreator(const std::vector<lite::tensor::Ten
     MS_LOG(ERROR) << "Input context is nullptr!";
     return nullptr;
   }
-
+  MS_ASSERT(desc.type == schema::PrimitiveType_AddN);
   op_parameter->thread_num_ = ctx->threadNum;
   auto *kernel = new (std::nothrow) AddNCPUKernel(op_parameter, inputs, outputs);
   if (kernel == nullptr) {
@@ -117,5 +117,5 @@ kernel::LiteKernel *CpuAddNFp32KernelCreator(const std::vector<lite::tensor::Ten
   return kernel;
 }
 
-REG_KERNEL(kCPU, PrimitiveType_AddN, CpuAddNFp32KernelCreator)
+REG_KERNEL(kCPU, kNumberTypeFloat32, PrimitiveType_AddN, CpuAddNFp32KernelCreator)
 }  // namespace mindspore::kernel

mindspore/lite/src/runtime/kernel/arm/fp32/argminmax.cc

@@ -92,9 +92,6 @@ kernel::LiteKernel *CpuArgMinMaxFp32KernelCreator(const std::vector<lite::tensor
   return kernel;
 }
 
-REG_KERNEL(kCPU, PrimitiveType_ArgMax, CpuArgMinMaxFp32KernelCreator)
-REG_KERNEL(kCPU, PrimitiveType_ArgMin, CpuArgMinMaxFp32KernelCreator)
+REG_KERNEL(kCPU, kNumberTypeFloat32, PrimitiveType_ArgMax, CpuArgMinMaxFp32KernelCreator)
+REG_KERNEL(kCPU, kNumberTypeFloat32, PrimitiveType_ArgMin, CpuArgMinMaxFp32KernelCreator)
 }  // namespace mindspore::kernel
-
-
-

mindspore/lite/src/runtime/kernel/arm/fp32/arithmetic.cc

@@ -112,25 +112,8 @@ kernel::LiteKernel *CpuArithmeticFp32KernelCreator(const std::vector<lite::tenso
                                                    const std::vector<lite::tensor::Tensor *> &outputs,
                                                    OpParameter *parameter, const lite::Context *ctx,
                                                    const kernel::KernelKey &desc) {
-  MS_ASSERT(parameter);
-  MS_ASSERT(inputs.at(0));
-  auto data_type = inputs.at(0)->data_type();
-  kernel::LiteKernel *kernel = nullptr;
-  switch (data_type) {
-    case kNumberTypeFloat32:
-      kernel = new (std::nothrow) ArithmeticCPUKernel(parameter, inputs, outputs, ctx);
-      break;
-    case kNumberTypeInt8:
-      if (desc.type == schema::PrimitiveType_Add) {
-        kernel = new (std::nothrow) QuantizedAddCPUKernel(parameter, inputs, outputs, ctx);
-      } else if (desc.type == schema::PrimitiveType_Mul) {
-        kernel = new (std::nothrow) MulInt8CPUKernel(parameter, inputs, outputs, ctx);
-      } else {
-      }
-      break;
-    default:
-      break;
-  }
+  MS_ASSERT(parameter != nullptr);
+  auto kernel = new (std::nothrow) ArithmeticCPUKernel(parameter, inputs, outputs, ctx);
   if (kernel == nullptr) {
     MS_LOG(ERROR) << "Create kernel failed, name: " << parameter->name_;
     return nullptr;
@@ -145,24 +128,23 @@ kernel::LiteKernel *CpuArithmeticFp32KernelCreator(const std::vector<lite::tenso
   return kernel;
 }
 
-REG_KERNEL(kCPU, PrimitiveType_Mul, CpuArithmeticFp32KernelCreator)
-REG_KERNEL(kCPU, PrimitiveType_Add, CpuArithmeticFp32KernelCreator)
-REG_KERNEL(kCPU, PrimitiveType_Sub, CpuArithmeticFp32KernelCreator)
-REG_KERNEL(kCPU, PrimitiveType_Div, CpuArithmeticFp32KernelCreator)
-REG_KERNEL(kCPU, PrimitiveType_LogicalAnd, CpuArithmeticFp32KernelCreator)
-REG_KERNEL(kCPU, PrimitiveType_LogicalOr, CpuArithmeticFp32KernelCreator)
-REG_KERNEL(kCPU, PrimitiveType_Maximum, CpuArithmeticFp32KernelCreator)
-REG_KERNEL(kCPU, PrimitiveType_Minimum, CpuArithmeticFp32KernelCreator)
-REG_KERNEL(kCPU, PrimitiveType_FloorDiv, CpuArithmeticFp32KernelCreator)
-REG_KERNEL(kCPU, PrimitiveType_FloorMod, CpuArithmeticFp32KernelCreator)
-REG_KERNEL(kCPU, PrimitiveType_SquaredDifference, CpuArithmeticFp32KernelCreator)
-REG_KERNEL(kCPU, PrimitiveType_Equal, CpuArithmeticFp32KernelCreator)
-REG_KERNEL(kCPU, PrimitiveType_NotEqual, CpuArithmeticFp32KernelCreator)
-REG_KERNEL(kCPU, PrimitiveType_Less, CpuArithmeticFp32KernelCreator)
-REG_KERNEL(kCPU, PrimitiveType_LessEqual, CpuArithmeticFp32KernelCreator)
-REG_KERNEL(kCPU, PrimitiveType_Greater, CpuArithmeticFp32KernelCreator)
-REG_KERNEL(kCPU, PrimitiveType_GreaterEqual, CpuArithmeticFp32KernelCreator)
-REG_KERNEL(kCPU, PrimitiveType_Eltwise, CpuArithmeticFp32KernelCreator)
+REG_KERNEL(kCPU, kNumberTypeFloat32, PrimitiveType_Mul, CpuArithmeticFp32KernelCreator)
+REG_KERNEL(kCPU, kNumberTypeFloat32, PrimitiveType_Add, CpuArithmeticFp32KernelCreator)
+REG_KERNEL(kCPU, kNumberTypeFloat32, PrimitiveType_Sub, CpuArithmeticFp32KernelCreator)
+REG_KERNEL(kCPU, kNumberTypeFloat32, PrimitiveType_Div, CpuArithmeticFp32KernelCreator)
+REG_KERNEL(kCPU, kNumberTypeFloat32, PrimitiveType_LogicalAnd, CpuArithmeticFp32KernelCreator)
+REG_KERNEL(kCPU, kNumberTypeFloat32, PrimitiveType_LogicalOr, CpuArithmeticFp32KernelCreator)
+REG_KERNEL(kCPU, kNumberTypeFloat32, PrimitiveType_Maximum, CpuArithmeticFp32KernelCreator)
+REG_KERNEL(kCPU, kNumberTypeFloat32, PrimitiveType_Minimum, CpuArithmeticFp32KernelCreator)
+REG_KERNEL(kCPU, kNumberTypeFloat32, PrimitiveType_FloorDiv, CpuArithmeticFp32KernelCreator)
+REG_KERNEL(kCPU, kNumberTypeFloat32, PrimitiveType_FloorMod, CpuArithmeticFp32KernelCreator)
+REG_KERNEL(kCPU, kNumberTypeFloat32, PrimitiveType_SquaredDifference, CpuArithmeticFp32KernelCreator)
+REG_KERNEL(kCPU, kNumberTypeFloat32, PrimitiveType_Equal, CpuArithmeticFp32KernelCreator)
+REG_KERNEL(kCPU, kNumberTypeFloat32, PrimitiveType_NotEqual, CpuArithmeticFp32KernelCreator)
+REG_KERNEL(kCPU, kNumberTypeFloat32, PrimitiveType_Less, CpuArithmeticFp32KernelCreator)
+REG_KERNEL(kCPU, kNumberTypeFloat32, PrimitiveType_LessEqual, CpuArithmeticFp32KernelCreator)
+REG_KERNEL(kCPU, kNumberTypeFloat32, PrimitiveType_Greater, CpuArithmeticFp32KernelCreator)
+REG_KERNEL(kCPU, kNumberTypeFloat32, PrimitiveType_GreaterEqual, CpuArithmeticFp32KernelCreator)
+REG_KERNEL(kCPU, kNumberTypeFloat32, PrimitiveType_Eltwise, CpuArithmeticFp32KernelCreator)
 
 }  // namespace mindspore::kernel
-

mindspore/lite/src/runtime/kernel/arm/fp32/arithmetic_self.cc

@@ -101,16 +101,15 @@ kernel::LiteKernel *CpuArithmeticSelfFp32KernelCreator(const std::vector<lite::t
   return kernel;
 }
 
-REG_KERNEL(kCPU, PrimitiveType_Abs, CpuArithmeticSelfFp32KernelCreator)
-REG_KERNEL(kCPU, PrimitiveType_Cos, CpuArithmeticSelfFp32KernelCreator)
-REG_KERNEL(kCPU, PrimitiveType_Exp, CpuArithmeticSelfFp32KernelCreator)
-REG_KERNEL(kCPU, PrimitiveType_Log, CpuArithmeticSelfFp32KernelCreator)
-REG_KERNEL(kCPU, PrimitiveType_Square, CpuArithmeticSelfFp32KernelCreator)
-REG_KERNEL(kCPU, PrimitiveType_Sqrt, CpuArithmeticSelfFp32KernelCreator)
-REG_KERNEL(kCPU, PrimitiveType_Rsqrt, CpuArithmeticSelfFp32KernelCreator)
-REG_KERNEL(kCPU, PrimitiveType_Sin, CpuArithmeticSelfFp32KernelCreator)
-REG_KERNEL(kCPU, PrimitiveType_LogicalNot, CpuArithmeticSelfFp32KernelCreator)
-REG_KERNEL(kCPU, PrimitiveType_Floor, CpuArithmeticSelfFp32KernelCreator)
-REG_KERNEL(kCPU, PrimitiveType_Ceil, CpuArithmeticSelfFp32KernelCreator)
+REG_KERNEL(kCPU, kNumberTypeFloat32, PrimitiveType_Abs, CpuArithmeticSelfFp32KernelCreator)
+REG_KERNEL(kCPU, kNumberTypeFloat32, PrimitiveType_Cos, CpuArithmeticSelfFp32KernelCreator)
+REG_KERNEL(kCPU, kNumberTypeFloat32, PrimitiveType_Exp, CpuArithmeticSelfFp32KernelCreator)
+REG_KERNEL(kCPU, kNumberTypeFloat32, PrimitiveType_Log, CpuArithmeticSelfFp32KernelCreator)
+REG_KERNEL(kCPU, kNumberTypeFloat32, PrimitiveType_Square, CpuArithmeticSelfFp32KernelCreator)
+REG_KERNEL(kCPU, kNumberTypeFloat32, PrimitiveType_Sqrt, CpuArithmeticSelfFp32KernelCreator)
+REG_KERNEL(kCPU, kNumberTypeFloat32, PrimitiveType_Rsqrt, CpuArithmeticSelfFp32KernelCreator)
+REG_KERNEL(kCPU, kNumberTypeFloat32, PrimitiveType_Sin, CpuArithmeticSelfFp32KernelCreator)
+REG_KERNEL(kCPU, kNumberTypeFloat32, PrimitiveType_LogicalNot, CpuArithmeticSelfFp32KernelCreator)
+REG_KERNEL(kCPU, kNumberTypeFloat32, PrimitiveType_Floor, CpuArithmeticSelfFp32KernelCreator)
+REG_KERNEL(kCPU, kNumberTypeFloat32, PrimitiveType_Ceil, CpuArithmeticSelfFp32KernelCreator)
 }  // namespace mindspore::kernel
-

mindspore/lite/src/runtime/kernel/arm/fp32/batch_to_space.cc

@@ -69,6 +69,7 @@ kernel::LiteKernel *CpuBatchToSpaceFp32KernelCreator(const std::vector<lite::ten
     MS_LOG(ERROR) << "Input op_parameter is nullptr!";
     return nullptr;
   }
+  MS_ASSERT(desc.type == schema::PrimitiveType_BatchToSpace);
   auto *kernel = new (std::nothrow) BatchToSpaceCPUKernel(op_parameter, inputs, outputs);
   if (kernel == nullptr) {
     MS_LOG(ERROR) << "new BatchToSpaceCPUKernel fail!";
@@ -85,7 +86,5 @@ kernel::LiteKernel *CpuBatchToSpaceFp32KernelCreator(const std::vector<lite::ten
   return kernel;
 }
 
-REG_KERNEL(kCPU, PrimitiveType_BatchToSpace, CpuBatchToSpaceFp32KernelCreator)
+REG_KERNEL(kCPU, kNumberTypeFloat32, PrimitiveType_BatchToSpace, CpuBatchToSpaceFp32KernelCreator)
 }  // namespace mindspore::kernel
-
-

mindspore/lite/src/runtime/kernel/arm/fp32/bias.cc

@@ -52,7 +52,7 @@ int BiasCPUKernel::Init() {
     bias_param_->in_shape1_[i] = 1;
     bias_param_->out_shape_[i] = dims[i];
   }
-  bias_param_->in_shape1_[bias_param_->ndim_ - 1] = dims[bias_param_->ndim_  - 1];
+  bias_param_->in_shape1_[bias_param_->ndim_ - 1] = dims[bias_param_->ndim_ - 1];
   return RET_OK;
 }
 
@@ -61,19 +61,7 @@ kernel::LiteKernel *CpuBiasFp32KernelCreator(const std::vector<lite::tensor::Ten
                                              const lite::Context *ctx, const kernel::KernelKey &desc) {
   MS_ASSERT(parameter != nullptr);
   MS_ASSERT(desc.type == schema::PrimitiveType_BiasAdd);
-  MS_ASSERT(inputs.at(0));
-  auto data_type = inputs.at(0)->data_type();
-  kernel::LiteKernel *kernel = nullptr;
-  switch (data_type) {
-    case kNumberTypeFloat32:
-      kernel = new (std::nothrow) BiasCPUKernel(parameter, inputs, outputs);
-      break;
-    case kNumberTypeInt8:
-      kernel = new (std::nothrow) BiasAddInt8CPUKernel(parameter, inputs, outputs, ctx);
-      break;
-    default:
-      break;
-  }
+  auto kernel = new (std::nothrow) BiasCPUKernel(parameter, inputs, outputs);
   if (kernel == nullptr) {
     MS_LOG(ERROR) << "Create kernel failed, name: " << parameter->name_;
     return nullptr;
@@ -89,6 +77,5 @@ kernel::LiteKernel *CpuBiasFp32KernelCreator(const std::vector<lite::tensor::Ten
   return kernel;
 }
 
-REG_KERNEL(kCPU, PrimitiveType_BiasAdd, CpuBiasFp32KernelCreator)
+REG_KERNEL(kCPU, kNumberTypeFloat32, PrimitiveType_BiasAdd, CpuBiasFp32KernelCreator)
 }  // namespace mindspore::kernel
-

mindspore/lite/src/runtime/kernel/arm/fp32/broadcast_to.cc

@@ -56,6 +56,7 @@ kernel::LiteKernel *CpuBroadcastToFp32KernelCreator(const std::vector<lite::tens
     MS_LOG(ERROR) << "Input op_parameter is nullptr!";
     return nullptr;
   }
+  MS_ASSERT(desc.type == schema::PrimitiveType_BroadcastTo);
   auto *kernel = new (std::nothrow) BroadcastToCPUKernel(op_parameter, inputs, outputs);
   if (kernel == nullptr) {
     MS_LOG(ERROR) << "new BroadcastToCPUKernel fail!";
@@ -72,7 +73,5 @@ kernel::LiteKernel *CpuBroadcastToFp32KernelCreator(const std::vector<lite::tens
   return kernel;
 }
 
-REG_KERNEL(kCPU, PrimitiveType_BroadcastTo, CpuBroadcastToFp32KernelCreator)
+REG_KERNEL(kCPU, kNumberTypeFloat32, PrimitiveType_BroadcastTo, CpuBroadcastToFp32KernelCreator)
 }  // namespace mindspore::kernel
-
-

mindspore/lite/src/runtime/kernel/arm/fp32/cast.cc

@@ -114,5 +114,5 @@ kernel::LiteKernel *CpuCastFp32KernelCreator(const std::vector<lite::tensor::Ten
   return kernel;
 }
 
-REG_KERNEL(kCPU, PrimitiveType_Cast, CpuCastFp32KernelCreator)
+REG_KERNEL(kCPU, kNumberTypeFloat32, PrimitiveType_Cast, CpuCastFp32KernelCreator)
 }  // namespace mindspore::kernel

mindspore/lite/src/runtime/kernel/arm/fp32/convolution.cc

@@ -15,6 +15,9 @@
  */
 
 #include "src/runtime/kernel/arm/fp32/convolution.h"
+#include "src/runtime/kernel/arm/fp32/convolution_1x1.h"
+#include "src/runtime/kernel/arm/fp32/convolution_3x3.h"
+#include "src/runtime/kernel/arm/fp32/convolution_winograd.h"
 #include "src/runtime/kernel/arm/opclib/fp32/conv.h"
 #include "schema/model_generated.h"
 #include "src/kernel_factory.h"
@@ -204,4 +207,78 @@ int ConvolutionCPUKernel::Run() {
   }
   return RET_OK;
 }
+
+void CheckIfUseWinograd(bool *use_winograd, int *output_unit, ConvParameter *conv_param,
+                        InputTransformUnitFunc input_trans_func, OutputTransformUnitFunc output_trans_func) {
+  if (conv_param->kernel_w_ == conv_param->kernel_h_ && conv_param->dilation_h_ == 1 && conv_param->dilation_w_ == 1 &&
+      conv_param->stride_h_ == 1 && conv_param->stride_w_ == 1) {
+    *output_unit = SelectOutputUnit(conv_param);
+    if (*output_unit > 1) {
+      *use_winograd = true;
+      int input_unit = conv_param->kernel_h_ + *output_unit - 1;
+      input_trans_func = GetInputTransFunc(input_unit);
+      if (input_trans_func == nullptr) {
+        MS_LOG(INFO) << "No matching input trans func. Turn back to common conv.";
+        *use_winograd = false;
+      }
+      output_trans_func = GetOutputTransFunc(input_unit, *output_unit);
+      if (output_trans_func == nullptr) {
+        MS_LOG(INFO) << "No matching output trans func. Turn back to common conv.";
+        *use_winograd = false;
+      }
+    } else {
+      *use_winograd = false;
+    }
+  } else {
+    *use_winograd = false;
+  }
+}
+
+kernel::LiteKernel *CpuConvFp32KernelCreator(const std::vector<lite::tensor::Tensor *> &inputs,
+                                             const std::vector<lite::tensor::Tensor *> &outputs,
+                                             OpParameter *opParameter, const Context *ctx,
+                                             const kernel::KernelKey &desc) {
+  MS_ASSERT(opParameter != nullptr);
+  MS_ASSERT(desc.type == schema::PrimitiveType_Conv2D);
+  auto conv_param = reinterpret_cast<ConvParameter *>(opParameter);
+  int kernel_h = conv_param->kernel_h_;
+  int kernel_w = conv_param->kernel_w_;
+  int stride_h = conv_param->stride_h_;
+  int stride_w = conv_param->stride_w_;
+  int dilation_h = conv_param->dilation_h_;
+  int dilation_w = conv_param->dilation_w_;
+  conv_param->input_h_ = inputs.front()->Height();
+  conv_param->input_w_ = inputs.front()->Width();
+  conv_param->output_h_ = outputs.front()->Height();
+  conv_param->output_w_ = outputs.front()->Width();
+  bool use_winograd;
+  int out_unit;
+  InputTransformUnitFunc input_trans_func = nullptr;
+  OutputTransformUnitFunc output_trans_func = nullptr;
+  CheckIfUseWinograd(&use_winograd, &out_unit, conv_param, input_trans_func, output_trans_func);
+  kernel::LiteKernel *kernel;
+  if (kernel_h == 1 && kernel_w == 1) {
+    kernel = new (std::nothrow) kernel::Convolution1x1CPUKernel(opParameter, inputs, outputs, ctx);
+  } else if (kernel_h == 3 && kernel_w == 3 && stride_h == 1 && stride_w == 1 && dilation_h == 1 && dilation_w == 1) {
+    kernel = new (std::nothrow) kernel::Convolution3x3CPUKernel(opParameter, inputs, outputs, ctx);
+  } else if (use_winograd) {
+    kernel = new (std::nothrow) kernel::ConvolutionWinogradCPUKernel(opParameter, inputs, outputs, ctx, out_unit);
+  } else {
+    kernel = new (std::nothrow) kernel::ConvolutionCPUKernel(opParameter, inputs, outputs, ctx);
+  }
+  if (kernel == nullptr) {
+    MS_LOG(ERROR) << "kernel is nullptr.";
+    return nullptr;
+  }
+  auto ret = kernel->Init();
+  if (ret != RET_OK) {
+    delete kernel;
+    MS_LOG(ERROR) << "Init kernel failed, name: " << opParameter->name_ << ", type: "
+                  << schema::EnumNamePrimitiveType(static_cast<schema::PrimitiveType>(opParameter->type_));
+    return nullptr;
+  }
+  return kernel;
+}
+
+REG_KERNEL(kCPU, kNumberTypeFloat32, PrimitiveType_Conv2D, CpuConvFp32KernelCreator)
 }  // namespace mindspore::kernel

mindspore/lite/src/runtime/kernel/arm/fp32/convolution_depthwise.cc

@@ -145,5 +145,29 @@ int ConvolutionDepthwiseCPUKernel::Run() {
   }
   return RET_OK;
 }
+
+
+kernel::LiteKernel *CpuConvDwFp32KernelCreator(const std::vector<lite::tensor::Tensor *> &inputs,
+                                               const std::vector<lite::tensor::Tensor *> &outputs,
+                                               OpParameter *opParameter, const Context *ctx,
+                                               const kernel::KernelKey &desc) {
+  MS_ASSERT(opParameter != nullptr);
+  MS_ASSERT(desc.type == schema::PrimitiveType_DepthwiseConv2D);
+  auto kernel = new (std::nothrow) kernel::ConvolutionDepthwiseCPUKernel(opParameter, inputs, outputs, ctx);
+  if (kernel == nullptr) {
+    MS_LOG(ERROR) << "kernel is nullptr.";
+    return nullptr;
+  }
+  auto ret = kernel->Init();
+  if (ret != RET_OK) {
+    delete kernel;
+    MS_LOG(ERROR) << "Init kernel failed, name: " << opParameter->name_ << ", type: "
+                  << schema::EnumNamePrimitiveType(static_cast<schema::PrimitiveType>(opParameter->type_));
+    return nullptr;
+  }
+  return kernel;
+}
+
+REG_KERNEL(kCPU, kNumberTypeFloat32, PrimitiveType_DepthwiseConv2D, CpuConvDwFp32KernelCreator)
 }  // namespace mindspore::kernel
 

mindspore/lite/src/runtime/kernel/arm/fp32/crop.cc

@@ -37,12 +37,12 @@ int CropLaunch(int thread_id, LiteParallelGroupEnv *penv, void *cdata) {
   auto kernel = reinterpret_cast<CropCPUKernel *>(cdata);
   return kernel->CropParallelRun(thread_id);
 }
-}
+}  // namespace
 
 int CropCPUKernel::Init() {
   schema::Format input0_format = inputs_[0]->GetFormat();
   if (input0_format != schema::Format_NCHW && input0_format != schema::Format_NHWC) {
-     MS_LOG(ERROR) << "Unsupport format " << input0_format;
+    MS_LOG(ERROR) << "Unsupport format " << input0_format;
     return RET_FORMAT_ERR;
   }
   outputs_[0]->SetFormat(input0_format);
@@ -90,7 +90,7 @@ kernel::LiteKernel *CpuCropFp32KernelCreator(const std::vector<lite::tensor::Ten
     MS_LOG(ERROR) << "Input context is nullptr!";
     return nullptr;
   }
-
+  MS_ASSERT(desc.type == schema::PrimitiveType_Crop);
   op_parameter->thread_num_ = ctx->threadNum;
   auto *kernel = new (std::nothrow) CropCPUKernel(op_parameter, inputs, outputs);
   if (kernel == nullptr) {
@@ -108,5 +108,5 @@ kernel::LiteKernel *CpuCropFp32KernelCreator(const std::vector<lite::tensor::Ten
   return kernel;
 }
 
-REG_KERNEL(kCPU, PrimitiveType_Crop, CpuCropFp32KernelCreator)
+REG_KERNEL(kCPU, kNumberTypeFloat32, PrimitiveType_Crop, CpuCropFp32KernelCreator)
 }  // namespace mindspore::kernel

mindspore/lite/src/runtime/kernel/arm/fp32/deconvolution.cc

@@ -225,4 +225,28 @@ int DeConvolutionCPUKernel::Run() {
   }
   return RET_OK;
 }
+
+
+kernel::LiteKernel *CpuDeConvFp32KernelCreator(const std::vector<lite::tensor::Tensor *> &inputs,
+                                               const std::vector<lite::tensor::Tensor *> &outputs,
+                                               OpParameter *opParameter, const lite::Context *ctx,
+                                               const kernel::KernelKey &desc) {
+  MS_ASSERT(opParameter != nullptr);
+  MS_ASSERT(desc.type == schema::PrimitiveType_DeConv2D);
+  auto kernel = new (std::nothrow) kernel::DeConvolutionCPUKernel(opParameter, inputs, outputs, ctx);
+  if (kernel == nullptr) {
+    MS_LOG(ERROR) << "kernel is nullptr.";
+    return nullptr;
+  }
+  auto ret = kernel->Init();
+  if (ret != RET_OK) {
+    delete kernel;
+    MS_LOG(ERROR) << "Init kernel failed, name: " << opParameter->name_ << ", type: "
+                  << schema::EnumNamePrimitiveType(static_cast<schema::PrimitiveType>(opParameter->type_));
+    return nullptr;
+  }
+  return kernel;
+}
+
+REG_KERNEL(kCPU, kNumberTypeFloat32, PrimitiveType_DeConv2D, CpuDeConvFp32KernelCreator)
 }  // namespace mindspore::kernel

mindspore/lite/src/runtime/kernel/arm/fp32/deconvolution_depthwise.cc

@@ -24,7 +24,7 @@ using mindspore::kernel::KERNEL_ARCH::kCPU;
 using mindspore::lite::KernelRegistrar;
 using mindspore::lite::RET_ERROR;
 using mindspore::lite::RET_OK;
-using mindspore::schema::PrimitiveType_DepthwiseConv2D;
+using mindspore::schema::PrimitiveType_DeDepthwiseConv2D;
 
 namespace mindspore::kernel {
 int DeconvolutionDepthwiseCPUKernel::InitSlideParam() {
@@ -158,5 +158,28 @@ int DeconvolutionDepthwiseCPUKernel::Run() {
   }
   return RET_OK;
 }
+
+kernel::LiteKernel *CpuDeconvDwFp32KernelCreator(const std::vector<lite::tensor::Tensor *> &inputs,
+                                                 const std::vector<lite::tensor::Tensor *> &outputs,
+                                                 OpParameter *opParameter, const lite::Context *ctx,
+                                                 const kernel::KernelKey &desc) {
+  MS_ASSERT(opParameter != nullptr);
+  MS_ASSERT(desc.type == schema::PrimitiveType_DeDepthwiseConv2D);
+  auto kernel = new (std::nothrow) kernel::DeconvolutionDepthwiseCPUKernel(opParameter, inputs, outputs, ctx);
+  if (kernel == nullptr) {
+    MS_LOG(ERROR) << "kernel is nullptr.";
+    return nullptr;
+  }
+  auto ret = kernel->Init();
+  if (ret != RET_OK) {
+    delete kernel;
+    MS_LOG(ERROR) << "Init kernel failed, name: " << opParameter->name_ << ", type: "
+                  << schema::EnumNamePrimitiveType(static_cast<schema::PrimitiveType>(opParameter->type_));
+    return nullptr;
+  }
+  return kernel;
+}
+
+REG_KERNEL(kCPU, kNumberTypeFloat32, PrimitiveType_DeDepthwiseConv2D, CpuDeconvDwFp32KernelCreator)
 }  // namespace mindspore::kernel
 

mindspore/lite/src/runtime/kernel/arm/fp32/depth_to_space.cc

@@ -67,6 +67,7 @@ kernel::LiteKernel *CpuDepthToSpaceFp32KernelCreator(const std::vector<lite::ten
     MS_LOG(ERROR) << "Input opParameter is nullptr!";
     return nullptr;
   }
+  MS_ASSERT(desc.type == schema::PrimitiveType_DepthToSpace);
   auto *kernel = new (std::nothrow) DepthToSpaceCPUKernel(opParameter, inputs, outputs);
   if (kernel == nullptr) {
     MS_LOG(ERROR) << "new DepthToSpaceCPUKernel fail!";
@@ -83,6 +84,6 @@ kernel::LiteKernel *CpuDepthToSpaceFp32KernelCreator(const std::vector<lite::ten
   return kernel;
 }
 
-REG_KERNEL(kCPU, PrimitiveType_DepthToSpace, CpuDepthToSpaceFp32KernelCreator)
+REG_KERNEL(kCPU, kNumberTypeFloat32, PrimitiveType_DepthToSpace, CpuDepthToSpaceFp32KernelCreator)
 }  // namespace mindspore::kernel
 

mindspore/lite/src/runtime/kernel/arm/fp32/expandDims.cc

@@ -82,6 +82,10 @@ kernel::LiteKernel *CpuExpandsDimsFp32KernelCreator(const std::vector<lite::tens
   MS_ASSERT(opParameter != nullptr);
   MS_ASSERT(desc.type == schema::PrimitiveType_ExpandDims);
   auto *kernel = new (std::nothrow) ExpandDimsCPUKernel(opParameter, inputs, outputs, ctx);
+  if (kernel == nullptr) {
+    MS_LOG(ERROR) << "new ExpandDimsCPUKernel fail!";
+    return nullptr;
+  }
   auto ret = kernel->Init();
   if (ret != RET_OK) {
     MS_LOG(ERROR) << "Init kernel failed, name: " << opParameter->name_ << ", type: "
@@ -92,6 +96,6 @@ kernel::LiteKernel *CpuExpandsDimsFp32KernelCreator(const std::vector<lite::tens
   return kernel;
 }
 
-REG_KERNEL(kCPU, PrimitiveType_ExpandDims, CpuExpandsDimsFp32KernelCreator)
+REG_KERNEL(kCPU, kNumberTypeFloat32, PrimitiveType_ExpandDims, CpuExpandsDimsFp32KernelCreator)
 }  // namespace mindspore::kernel
 

mindspore/lite/src/runtime/kernel/arm/fp32/fill.cc

@@ -92,6 +92,10 @@ kernel::LiteKernel *CpuFillFp32KernelCreator(const std::vector<lite::tensor::Ten
   }
   MS_ASSERT(desc.type == schema::PrimitiveType_Fill);
   auto *kernel = new (std::nothrow) FillCPUKernel(opParameter, inputs, outputs, ctx);
+  if (kernel == nullptr) {
+    MS_LOG(ERROR) << "new FillCPUKernel fail!";
+    return nullptr;
+  }
   auto ret = kernel->Init();
   if (ret != RET_OK) {
     MS_LOG(ERROR) << "Init kernel failed, name: " << opParameter->name_ << ", type: "
@@ -102,6 +106,6 @@ kernel::LiteKernel *CpuFillFp32KernelCreator(const std::vector<lite::tensor::Ten
   return kernel;
 }
 
-REG_KERNEL(kCPU, PrimitiveType_Fill, CpuFillFp32KernelCreator)
+REG_KERNEL(kCPU, kNumberTypeFloat32, PrimitiveType_Fill, CpuFillFp32KernelCreator)
 }  // namespace mindspore::kernel
 

mindspore/lite/src/runtime/kernel/arm/fp32/flatten.cc

@@ -56,6 +56,10 @@ kernel::LiteKernel *CpuFlattenFp32KernelCreator(const std::vector<lite::tensor::
   }
   MS_ASSERT(desc.type == schema::PrimitiveType_Flatten);
   auto *kernel = new (std::nothrow) FlattenCPUKernel(opParameter, inputs, outputs, ctx);
+  if (kernel == nullptr) {
+    MS_LOG(ERROR) << "new FlattenCPUKernel fail!";
+    return nullptr;
+  }
   auto ret = kernel->Init();
   if (ret != RET_OK) {
     MS_LOG(ERROR) << "Init kernel failed, name: " << opParameter->name_ << ", type: "
@@ -66,6 +70,6 @@ kernel::LiteKernel *CpuFlattenFp32KernelCreator(const std::vector<lite::tensor::
   return kernel;
 }
 
-REG_KERNEL(kCPU, PrimitiveType_Flatten, CpuFlattenFp32KernelCreator)
+REG_KERNEL(kCPU, kNumberTypeFloat32, PrimitiveType_Flatten, CpuFlattenFp32KernelCreator)
 }  // namespace mindspore::kernel
 

mindspore/lite/src/runtime/kernel/arm/fp32/fused_batchnorm.cc

@@ -55,14 +55,19 @@ kernel::LiteKernel *CpuFusedBatchnormKernelCreator(const std::vector<lite::tenso
   MS_ASSERT(opParameter != nullptr);
   MS_ASSERT(desc.type == schema::PrimitiveType_FusedBatchNorm);
   auto *kernel = new (std::nothrow) FusedBatchnormCPUKernel(opParameter, inputs, outputs);
+  if (kernel == nullptr) {
+    MS_LOG(ERROR) << "new FusedBatchnormCPUKernel fail!";
+    return nullptr;
+  }
   auto ret = kernel->Init();
   if (ret != RET_OK) {
     MS_LOG(ERROR) << "Init kernel failed, name: " << opParameter->name_ << ", type: "
                   << schema::EnumNamePrimitiveType(static_cast<schema::PrimitiveType>(opParameter->type_));
+    return nullptr;
   }
   return kernel;
 }
 
-REG_KERNEL(kCPU, PrimitiveType_FusedBatchNorm, CpuFusedBatchnormKernelCreator)
+REG_KERNEL(kCPU, kNumberTypeFloat32, PrimitiveType_FusedBatchNorm, CpuFusedBatchnormKernelCreator)
 }  // namespace mindspore::kernel
 

mindspore/lite/src/runtime/kernel/arm/fp32/gather.cc

@@ -121,6 +121,6 @@ kernel::LiteKernel *CpuGatherFp32KernelCreator(const std::vector<lite::tensor::T
   return kernel;
 }
 
-REG_KERNEL(kCPU, PrimitiveType_Gather, CpuGatherFp32KernelCreator)
+REG_KERNEL(kCPU, kNumberTypeFloat32, PrimitiveType_Gather, CpuGatherFp32KernelCreator)
 }  // namespace mindspore::kernel
 

mindspore/lite/src/runtime/kernel/arm/fp32/gatherNd.cc

@@ -143,6 +143,6 @@ kernel::LiteKernel *CpuGatherNdFp32KernelCreator(const std::vector<lite::tensor:
   return kernel;
 }
 
-REG_KERNEL(kCPU, PrimitiveType_GatherNd, CpuGatherNdFp32KernelCreator)
+REG_KERNEL(kCPU, kNumberTypeFloat32, PrimitiveType_GatherNd, CpuGatherNdFp32KernelCreator)
 }  // namespace mindspore::kernel
 

mindspore/lite/src/runtime/kernel/arm/fp32/local_response_norm.cc

@@ -95,6 +95,10 @@ kernel::LiteKernel *CpuLocalResponseNormFp32KernelCreator(const std::vector<lite
   MS_ASSERT(desc.type == schema::PrimitiveType_LocalResponseNormalization);
 
   auto *kernel = new (std::nothrow) LocalResponseNormCPUKernel(opParameter, inputs, outputs, ctx);
+  if (kernel == nullptr) {
+    MS_LOG(ERROR) << "new LocalResponseNormCPUKernel fail!";
+    return nullptr;
+  }
   auto ret = kernel->Init();
   if (ret != RET_OK) {
     MS_LOG(ERROR) << "Init kernel failed, name: " << opParameter->name_ << ", type: "
@@ -105,6 +109,5 @@ kernel::LiteKernel *CpuLocalResponseNormFp32KernelCreator(const std::vector<lite
   return kernel;
 }
 
-REG_KERNEL(kCPU, PrimitiveType_LocalResponseNormalization, CpuLocalResponseNormFp32KernelCreator)
+REG_KERNEL(kCPU, kNumberTypeFloat32, PrimitiveType_LocalResponseNormalization, CpuLocalResponseNormFp32KernelCreator)
 }  // namespace mindspore::kernel
-

mindspore/lite/src/runtime/kernel/arm/fp32/matmul.cc

@@ -40,6 +40,10 @@ kernel::LiteKernel *CpuMatmulFp32KernelCreator(const std::vector<lite::tensor::T
                                                const kernel::KernelKey &desc) {
   MS_ASSERT(desc.type == schema::PrimitiveType_MatMul);
   auto *kernel = new (std::nothrow) MatmulCPUKernel(opParameter, inputs, outputs);
+  if (kernel == nullptr) {
+    MS_LOG(ERROR) << "new MatmulCPUKernel fail!";
+    return nullptr;
+  }
   auto ret = kernel->Init();
   if (ret != RET_OK) {
     MS_LOG(ERROR) << "Init kernel failed, name: " << opParameter->name_ << ", type: "
@@ -48,6 +52,6 @@ kernel::LiteKernel *CpuMatmulFp32KernelCreator(const std::vector<lite::tensor::T
   return kernel;
 }
 
-REG_KERNEL(kCPU, PrimitiveType_MatMul, CpuMatmulFp32KernelCreator)
+REG_KERNEL(kCPU, kNumberTypeFloat32, PrimitiveType_MatMul, CpuMatmulFp32KernelCreator)
 }  // namespace mindspore::kernel
 

mindspore/lite/src/runtime/kernel/arm/fp32/nchw2nhwc.cc

@@ -43,6 +43,10 @@ kernel::LiteKernel *CpuNchw2NhwcFp32KernelCreator(const std::vector<lite::tensor
   MS_ASSERT(opParameter != nullptr);
   MS_ASSERT(desc.type == schema::PrimitiveType_Nchw2Nhwc);
   auto *kernel = new (std::nothrow) Nchw2NhwcCPUKernel(opParameter, inputs, outputs);
+  if (kernel == nullptr) {
+    MS_LOG(ERROR) << "new Nchw2NhwcCPUKernel fail!";
+    return nullptr;
+  }
   auto ret = kernel->Init();
   if (ret != RET_OK) {
     MS_LOG(ERROR) << "Init kernel failed, name: " << opParameter->name_ << ", type: "
@@ -53,6 +57,6 @@ kernel::LiteKernel *CpuNchw2NhwcFp32KernelCreator(const std::vector<lite::tensor
   return kernel;
 }
 
-REG_KERNEL(kCPU, PrimitiveType_Nchw2Nhwc, CpuNchw2NhwcFp32KernelCreator)
+REG_KERNEL(kCPU, kNumberTypeFloat32, PrimitiveType_Nchw2Nhwc, CpuNchw2NhwcFp32KernelCreator)
 }  // namespace mindspore::kernel
 

mindspore/lite/src/runtime/kernel/arm/fp32/nhwc2nchw.cc

@@ -43,6 +43,10 @@ kernel::LiteKernel *CpuNhwc2NchwFp32KernelCreator(const std::vector<lite::tensor
   MS_ASSERT(opParameter != nullptr);
   MS_ASSERT(desc.type == schema::PrimitiveType_Nhwc2Nchw);
   auto *kernel = new (std::nothrow) Nhwc2NchwCPUKernel(opParameter, inputs, outputs);
+  if (kernel == nullptr) {
+    MS_LOG(ERROR) << "new Nhwc2NchwCPUKernel fail!";
+    return nullptr;
+  }
   auto ret = kernel->Init();
   if (ret != RET_OK) {
     MS_LOG(ERROR) << "Init kernel failed, name: " << opParameter->name_ << ", type: "
@@ -53,6 +57,6 @@ kernel::LiteKernel *CpuNhwc2NchwFp32KernelCreator(const std::vector<lite::tensor
   return kernel;
 }
 
-REG_KERNEL(kCPU, PrimitiveType_Nhwc2Nchw, CpuNhwc2NchwFp32KernelCreator)
+REG_KERNEL(kCPU, kNumberTypeFloat32, PrimitiveType_Nhwc2Nchw, CpuNhwc2NchwFp32KernelCreator)
 }  // namespace mindspore::kernel
 

mindspore/lite/src/runtime/kernel/arm/fp32/one_hot.cc

@@ -183,5 +183,5 @@ kernel::LiteKernel *CpuOneHotFp32KernelCreator(const std::vector<lite::tensor::T
   return kernel;
 }
 
-REG_KERNEL(kCPU, PrimitiveType_OneHot, CpuOneHotFp32KernelCreator)
+REG_KERNEL(kCPU, kNumberTypeFloat32, PrimitiveType_OneHot, CpuOneHotFp32KernelCreator)
 }  // namespace mindspore::kernel

mindspore/lite/src/runtime/kernel/arm/fp32/power.cc

@@ -68,6 +68,10 @@ kernel::LiteKernel *CpuPowerFp32KernelCreator(const std::vector<lite::tensor::Te
   MS_ASSERT(desc.type == schema::PrimitiveType_Power);
   auto *kernel =
     new (std::nothrow) PowerCPUKernel(reinterpret_cast<PowerParameter *>(opParameter), inputs, outputs, ctx);
+  if (kernel == nullptr) {
+    MS_LOG(ERROR) << "new PowerCPUKernel fail!";
+    return nullptr;
+  }
   auto ret = kernel->Init();
   if (ret != RET_OK) {
     MS_LOG(ERROR) << "Init kernel failed, name: " << opParameter->name_ << ", type: "
@@ -76,5 +80,5 @@ kernel::LiteKernel *CpuPowerFp32KernelCreator(const std::vector<lite::tensor::Te
   return kernel;
 }
 
-REG_KERNEL(kCPU, PrimitiveType_Power, CpuPowerFp32KernelCreator)
+REG_KERNEL(kCPU, kNumberTypeFloat32, PrimitiveType_Power, CpuPowerFp32KernelCreator)
 }  // namespace mindspore::kernel

mindspore/lite/src/runtime/kernel/arm/fp32/prelu.cc

@@ -70,7 +70,7 @@ kernel::LiteKernel *CpuPReluFp32KernelCreator(const std::vector<lite::tensor::Te
     MS_LOG(ERROR) << "input opParameter is nullptr!";
     return nullptr;
   }
-
+  MS_ASSERT(desc.type == schema::PrimitiveType_Prelu);
   auto *kernel = new (std::nothrow) PReluCPUKernel(opParameter, inputs, outputs, ctx);
   if (kernel == nullptr) {
     MS_LOG(ERROR) << "new PReluCPUKernel fail!";
@@ -86,6 +86,6 @@ kernel::LiteKernel *CpuPReluFp32KernelCreator(const std::vector<lite::tensor::Te
   return kernel;
 }
 
-REG_KERNEL(kCPU, PrimitiveType_Prelu, CpuPReluFp32KernelCreator)
+REG_KERNEL(kCPU, kNumberTypeFloat32, PrimitiveType_Prelu, CpuPReluFp32KernelCreator)
 }  // namespace mindspore::kernel
 

mindspore/lite/src/runtime/kernel/arm/fp32/range.cc

@@ -54,6 +54,10 @@ kernel::LiteKernel *CpuRangeFp32KernelCreator(const std::vector<lite::tensor::Te
   MS_ASSERT(desc.type == schema::PrimitiveType_Range);
 
   auto *kernel = new (std::nothrow) RangeCPUKernel(opParameter, inputs, outputs);
+  if (kernel == nullptr) {
+    MS_LOG(ERROR) << "new RangeCPUKernel fail!";
+    return nullptr;
+  }
   auto ret = kernel->Init();
   if (ret != RET_OK) {
     MS_LOG(ERROR) << "Init kernel failed, name: " << opParameter->name_ << ", type: "
@@ -64,7 +68,7 @@ kernel::LiteKernel *CpuRangeFp32KernelCreator(const std::vector<lite::tensor::Te
   return kernel;
 }
 
-REG_KERNEL(kCPU, PrimitiveType_Range, CpuRangeFp32KernelCreator)
+REG_KERNEL(kCPU, kNumberTypeFloat32, PrimitiveType_Range, CpuRangeFp32KernelCreator)
 
 }  // namespace mindspore::kernel
 

mindspore/lite/src/runtime/kernel/arm/fp32/rank.cc

@@ -53,6 +53,10 @@ kernel::LiteKernel *CpuRankFp32KernelCreator(const std::vector<lite::tensor::Ten
   MS_ASSERT(desc.type == schema::PrimitiveType_Rank);
 
   auto *kernel = new (std::nothrow) RankCPUKernel(opParameter, inputs, outputs);
+  if (kernel == nullptr) {
+    MS_LOG(ERROR) << "new RankCPUKernel fail!";
+    return nullptr;
+  }
   auto ret = kernel->Init();
   if (ret != RET_OK) {
     MS_LOG(ERROR) << "Init kernel failed, name: " << opParameter->name_ << ", type: "
@@ -63,6 +67,6 @@ kernel::LiteKernel *CpuRankFp32KernelCreator(const std::vector<lite::tensor::Ten
   return kernel;
 }
 
-REG_KERNEL(kCPU, PrimitiveType_Rank, CpuRankFp32KernelCreator)
+REG_KERNEL(kCPU, kNumberTypeFloat32, PrimitiveType_Rank, CpuRankFp32KernelCreator)
 }  // namespace mindspore::kernel
 

mindspore/lite/src/runtime/kernel/arm/fp32/reduce.cc

@@ -241,7 +241,8 @@ int ReduceCPUKernel::MallocTmpBuffer() {
     data_buffers_.emplace_back(buffer);
     input_shape[axis] = 1;
   }
+  return RET_OK;
 }
 
-REG_KERNEL(kCPU, PrimitiveType_Reduce, CpuReduceFp32KernelCreator)
+REG_KERNEL(kCPU, kNumberTypeFloat32, PrimitiveType_Reduce, CpuReduceFp32KernelCreator)
 }  // namespace mindspore::kernel

mindspore/lite/src/runtime/kernel/arm/fp32/resize.cc

@@ -221,6 +221,7 @@ kernel::LiteKernel *CpuResizeFp32KernelCreator(const std::vector<lite::tensor::T
     MS_LOG(ERROR) << "Input opParameter is nullptr!";
     return nullptr;
   }
+  MS_ASSERT(desc.type == schema::PrimitiveType_Resize);
   auto *kernel = new (std::nothrow) ResizeCPUKernel(opParameter, inputs, outputs, ctx);
   if (kernel == nullptr) {
     MS_LOG(ERROR) << "new ResizeCPUKernel fail!";
@@ -237,6 +238,6 @@ kernel::LiteKernel *CpuResizeFp32KernelCreator(const std::vector<lite::tensor::T
   return kernel;
 }
 
-REG_KERNEL(kCPU, PrimitiveType_Resize, CpuResizeFp32KernelCreator)
+REG_KERNEL(kCPU, kNumberTypeFloat32, PrimitiveType_Resize, CpuResizeFp32KernelCreator)
 }  // namespace mindspore::kernel
 

mindspore/lite/src/runtime/kernel/arm/fp32/reverse.cc

@@ -139,6 +139,7 @@ kernel::LiteKernel *CpuReverseFp32KernelCreator(const std::vector<lite::tensor::
     MS_LOG(ERROR) << "opParameter is NULL! ";
     return nullptr;
   }
+  MS_ASSERT(desc.type == schema::PrimitiveType_Reverse);
   auto *kernel = new (std::nothrow) ReverseCPUKernel(opParameter, inputs, outputs, ctx);
   if (kernel == nullptr) {
     MS_LOG(ERROR) << "Kernel is NULL! name: " << opParameter->name_ << ", type: "
@@ -156,6 +157,6 @@ kernel::LiteKernel *CpuReverseFp32KernelCreator(const std::vector<lite::tensor::
   return kernel;
 }
 
-REG_KERNEL(kCPU, PrimitiveType_Reverse, CpuReverseFp32KernelCreator)
+REG_KERNEL(kCPU, kNumberTypeFloat32, PrimitiveType_Reverse, CpuReverseFp32KernelCreator)
 }  // namespace mindspore::kernel
 

mindspore/lite/src/runtime/kernel/arm/fp32/reverse_sequence.cc

@@ -96,6 +96,7 @@ kernel::LiteKernel *CpuReverseSequenceFp32KernelCreator(const std::vector<lite::
                                                         OpParameter *parameter, const lite::Context *ctx,
                                                         const KernelKey &desc) {
   MS_ASSERT(parameter != nullptr);
+  MS_ASSERT(desc.type == schema::PrimitiveType_ReverseSequence);
   auto *kernel = new (std::nothrow) ReverseSequenceCPUKernel(parameter, inputs, outputs);
   if (kernel == nullptr) {
     MS_LOG(ERROR) << "Create kernel failed, name: " << parameter->name_;
@@ -111,6 +112,6 @@ kernel::LiteKernel *CpuReverseSequenceFp32KernelCreator(const std::vector<lite::
   return kernel;
 }
 
-REG_KERNEL(kCPU, PrimitiveType_ReverseSequence, CpuReverseSequenceFp32KernelCreator)
+REG_KERNEL(kCPU, kNumberTypeFloat32, PrimitiveType_ReverseSequence, CpuReverseSequenceFp32KernelCreator)
 }  // namespace mindspore::kernel
 

mindspore/lite/src/runtime/kernel/arm/fp32/scale.cc

@@ -164,5 +164,5 @@ kernel::LiteKernel *CpuScaleFp32KernelCreator(const std::vector<lite::tensor::Te
   return kernel;
 }
 
-REG_KERNEL(kCPU, PrimitiveType_Scale, CpuScaleFp32KernelCreator)
+REG_KERNEL(kCPU, kNumberTypeFloat32, PrimitiveType_Scale, CpuScaleFp32KernelCreator)
 }  // namespace mindspore::kernel

mindspore/lite/src/runtime/kernel/arm/fp32/scatter_nd.cc

@@ -181,6 +181,6 @@ kernel::LiteKernel *CpuScatterNDFp32KernelCreator(const std::vector<lite::tensor
   return kernel;
 }
 
-REG_KERNEL(kCPU, PrimitiveType_ScatterND, CpuScatterNDFp32KernelCreator)
+REG_KERNEL(kCPU, kNumberTypeFloat32, PrimitiveType_ScatterND, CpuScatterNDFp32KernelCreator)
 }  // namespace mindspore::kernel
 

mindspore/lite/src/runtime/kernel/arm/fp32/shape.cc

@@ -79,6 +79,6 @@ kernel::LiteKernel *CpuShapeFp32KernelCreator(const std::vector<lite::tensor::Te
   return kernel;
 }
 
-REG_KERNEL(kCPU, PrimitiveType_Shape, CpuShapeFp32KernelCreator)
+REG_KERNEL(kCPU, kNumberTypeFloat32, PrimitiveType_Shape, CpuShapeFp32KernelCreator)
 }  // namespace mindspore::kernel
 

mindspore/lite/src/runtime/kernel/arm/fp32/slice.cc

@@ -23,8 +23,8 @@
 
 using mindspore::lite::KernelRegistrar;
 using mindspore::lite::RET_ERROR;
-using mindspore::lite::RET_OK;
 using mindspore::lite::RET_NULL_PTR;
+using mindspore::lite::RET_OK;
 using mindspore::schema::PrimitiveType_Slice;
 
 namespace mindspore::kernel {
@@ -37,7 +37,7 @@ int SliceLaunch(int thread_id, LiteParallelGroupEnv *penv, void *cdata) {
   auto kernel = reinterpret_cast<SliceCPUKernel *>(cdata);
   return kernel->SliceParallelRun(thread_id);
 }
-}
+}  // namespace
 
 int SliceCPUKernel::Init() {
   auto *param = reinterpret_cast<SliceParameter *>(opParameter);
@@ -106,7 +106,7 @@ kernel::LiteKernel *CpuSliceFp32KernelCreator(const std::vector<lite::tensor::Te
     MS_LOG(ERROR) << "Input context is nullptr!";
     return nullptr;
   }
-
+  MS_ASSERT(desc.type == schema::PrimitiveType_Slice);
   op_parameter->thread_num_ = ctx->threadNum;
   auto *kernel = new (std::nothrow) SliceCPUKernel(op_parameter, inputs, outputs);
   if (kernel == nullptr) {
@@ -124,5 +124,5 @@ kernel::LiteKernel *CpuSliceFp32KernelCreator(const std::vector<lite::tensor::Te
   return kernel;
 }
 
-REG_KERNEL(kCPU, PrimitiveType_Slice, CpuSliceFp32KernelCreator)
+REG_KERNEL(kCPU, kNumberTypeFloat32, PrimitiveType_Slice, CpuSliceFp32KernelCreator)
 }  // namespace mindspore::kernel

mindspore/lite/src/runtime/kernel/arm/fp32/softmax.cc

@@ -64,6 +64,10 @@ kernel::LiteKernel *CpuSoftmaxFp32KernelCreator(const std::vector<lite::tensor::
   MS_ASSERT(opParameter != nullptr);
   MS_ASSERT(desc.type == schema::PrimitiveType_SoftMax);
   auto *kernel = new (std::nothrow) SoftmaxCPUKernel(opParameter, inputs, outputs);
+  if (kernel == nullptr) {
+    MS_LOG(ERROR) << "new SoftmaxCPUKernel fail!";
+    return nullptr;
+  }
   auto ret = kernel->Init();
   if (ret != RET_OK) {
     MS_LOG(ERROR) << "Init kernel failed, name: " << opParameter->name_ << ", type: "
@@ -74,6 +78,6 @@ kernel::LiteKernel *CpuSoftmaxFp32KernelCreator(const std::vector<lite::tensor::
   return kernel;
 }
 
-REG_KERNEL(kCPU, PrimitiveType_SoftMax, CpuSoftmaxFp32KernelCreator)
+REG_KERNEL(kCPU, kNumberTypeFloat32, PrimitiveType_SoftMax, CpuSoftmaxFp32KernelCreator)
 }  // namespace mindspore::kernel
 

mindspore/lite/src/runtime/kernel/arm/fp32/sparse_to_dense.cc

@@ -81,7 +81,7 @@ kernel::LiteKernel *CpuSparseToDenseFp32KernelCreator(const std::vector<lite::te
     MS_LOG(ERROR) << "input opParameter is nullptr!";
     return nullptr;
   }
-
+  MS_ASSERT(desc.type == schema::PrimitiveType_SparseToDense);
   auto *kernel = new (std::nothrow) SparseToDenseCPUKernel(opParameter, inputs, outputs, ctx);
   if (kernel == nullptr) {
     MS_LOG(ERROR) << "new SparseToDenseCPUKernel fail!";
@@ -97,6 +97,6 @@ kernel::LiteKernel *CpuSparseToDenseFp32KernelCreator(const std::vector<lite::te
   return kernel;
 }
 
-REG_KERNEL(kCPU, PrimitiveType_SparseToDense, CpuSparseToDenseFp32KernelCreator)
+REG_KERNEL(kCPU, kNumberTypeFloat32, PrimitiveType_SparseToDense, CpuSparseToDenseFp32KernelCreator)
 }  // namespace mindspore::kernel
 

mindspore/lite/src/runtime/kernel/arm/fp32/split.cc

@@ -126,5 +126,5 @@ kernel::LiteKernel *CpuSplitFp32KernelCreator(const std::vector<lite::tensor::Te
   return kernel;
 }
 
-REG_KERNEL(kCPU, PrimitiveType_Split, CpuSplitFp32KernelCreator)
+REG_KERNEL(kCPU, kNumberTypeFloat32, PrimitiveType_Split, CpuSplitFp32KernelCreator)
 }  // namespace mindspore::kernel

mindspore/lite/src/runtime/kernel/arm/fp32/squeeze.cc

@@ -74,6 +74,6 @@ kernel::LiteKernel *CpuSqueezeFp32KernelCreator(const std::vector<lite::tensor::
   return kernel;
 }
 
-REG_KERNEL(kCPU, PrimitiveType_Squeeze, CpuSqueezeFp32KernelCreator)
+REG_KERNEL(kCPU, kNumberTypeFloat32, PrimitiveType_Squeeze, CpuSqueezeFp32KernelCreator)
 }  // namespace mindspore::kernel
 

mindspore/lite/src/runtime/kernel/arm/fp32/stack.cc

@@ -92,6 +92,7 @@ kernel::LiteKernel *CpuStackFp32KernelCreator(const std::vector<lite::tensor::Te
     MS_LOG(ERROR) << "Input op_parameter is nullptr!";
     return nullptr;
   }
+  MS_ASSERT(desc.type == schema::PrimitiveType_Stack);
   auto *kernel = new (std::nothrow) StackCPUKernel(op_parameter, inputs, outputs);
   if (kernel == nullptr) {
     MS_LOG(ERROR) << "new StackCPUKernel fail!";
@@ -108,6 +109,5 @@ kernel::LiteKernel *CpuStackFp32KernelCreator(const std::vector<lite::tensor::Te
   return kernel;
 }
 
-REG_KERNEL(kCPU, PrimitiveType_Stack, CpuStackFp32KernelCreator)
+REG_KERNEL(kCPU, kNumberTypeFloat32, PrimitiveType_Stack, CpuStackFp32KernelCreator)
 }  // namespace mindspore::kernel
-

mindspore/lite/src/runtime/kernel/arm/fp32/strided_slice.cc

@@ -82,5 +82,5 @@ kernel::LiteKernel *CpuStridedSliceFp32KernelCreator(const std::vector<lite::ten
   return kernel;
 }
 
-REG_KERNEL(kCPU, PrimitiveType_StridedSlice, CpuStridedSliceFp32KernelCreator)
+REG_KERNEL(kCPU, kNumberTypeFloat32, PrimitiveType_StridedSlice, CpuStridedSliceFp32KernelCreator)
 }  // namespace mindspore::kernel

mindspore/lite/src/runtime/kernel/arm/fp32/tile.cc

@@ -77,6 +77,6 @@ kernel::LiteKernel *CpuTileFp32KernelCreator(const std::vector<lite::tensor::Ten
   return kernel;
 }
 
-REG_KERNEL(kCPU, PrimitiveType_Tile, CpuTileFp32KernelCreator)
+REG_KERNEL(kCPU, kNumberTypeFloat32, PrimitiveType_Tile, CpuTileFp32KernelCreator)
 }  // namespace mindspore::kernel
 

mindspore/lite/src/runtime/kernel/arm/fp32/topk.cc

@@ -53,10 +53,13 @@ int TopKCPUKernel::Run() {
 kernel::LiteKernel *CpuTopKFp32KernelCreator(const std::vector<lite::tensor::Tensor *> &inputs,
                                              const std::vector<lite::tensor::Tensor *> &outputs, OpParameter *parameter,
                                              const lite::Context *ctx, const KernelKey &desc) {
-  MS_EXCEPTION_IF_NULL(parameter);
+  MS_ASSERT(parameter != nullptr);
   MS_ASSERT(desc.type == PrimitiveType_Tile);
   auto *kernel = new (std::nothrow) TopKCPUKernel(parameter, inputs, outputs);
-  MS_EXCEPTION_IF_NULL(kernel);
+  if (kernel == nullptr) {
+    MS_LOG(ERROR) << "new TopKCPUKernel fail!";
+    return nullptr;
+  }
 
   auto ret = kernel->Init();
   if (ret != RET_OK) {
@@ -68,6 +71,6 @@ kernel::LiteKernel *CpuTopKFp32KernelCreator(const std::vector<lite::tensor::Ten
   return kernel;
 }
 
-REG_KERNEL(kCPU, PrimitiveType_TopK, CpuTopKFp32KernelCreator)
+REG_KERNEL(kCPU, kNumberTypeFloat32, PrimitiveType_TopK, CpuTopKFp32KernelCreator)
 }  // namespace mindspore::kernel
 

mindspore/lite/src/runtime/kernel/arm/fp32/transpose.cc

@@ -95,6 +95,6 @@ kernel::LiteKernel *CpuTransposeFp32KernelCreator(const std::vector<lite::tensor
   return kernel;
 }
 
-REG_KERNEL(kCPU, PrimitiveType_Transpose, CpuTransposeFp32KernelCreator)
+REG_KERNEL(kCPU, kNumberTypeFloat32, PrimitiveType_Transpose, CpuTransposeFp32KernelCreator)
 }  // namespace mindspore::kernel
 

mindspore/lite/src/runtime/kernel/arm/fp32/unique.cc

@@ -62,6 +62,6 @@ kernel::LiteKernel *CpuUniqueFp32KernelCreator(const std::vector<lite::tensor::T
   return kernel;
 }
 
-REG_KERNEL(kCPU, PrimitiveType_Unique, CpuUniqueFp32KernelCreator)
+REG_KERNEL(kCPU, kNumberTypeFloat32, PrimitiveType_Unique, CpuUniqueFp32KernelCreator)
 }  // namespace mindspore::kernel
 

mindspore/lite/src/runtime/kernel/arm/fp32/unsqueeze.cc

@@ -81,6 +81,10 @@ kernel::LiteKernel *CpuUnsqueezeFp32KernelCreator(const std::vector<lite::tensor
   MS_ASSERT(opParameter != nullptr);
   MS_ASSERT(desc.type == schema::PrimitiveType_Unsqueeze);
   auto *kernel = new (std::nothrow) UnsqueezeCPUKernel(opParameter, inputs, outputs, ctx);
+  if (kernel == nullptr) {
+    MS_LOG(ERROR) << "new AddNCPUKernel fail!";
+    return nullptr;
+  }
   auto ret = kernel->Init();
   if (ret != RET_OK) {
     MS_LOG(ERROR) << "Init kernel failed, name: " << opParameter->name_ << ", type: "
@@ -91,6 +95,6 @@ kernel::LiteKernel *CpuUnsqueezeFp32KernelCreator(const std::vector<lite::tensor
   return kernel;
 }
 
-REG_KERNEL(kCPU, PrimitiveType_Unsqueeze, CpuUnsqueezeFp32KernelCreator)
+REG_KERNEL(kCPU, kNumberTypeFloat32, PrimitiveType_Unsqueeze, CpuUnsqueezeFp32KernelCreator)
 }  // namespace mindspore::kernel
 

mindspore/lite/src/runtime/kernel/arm/fp32/unstack.cc

@@ -86,6 +86,5 @@ kernel::LiteKernel *CpuUnstackFp32KernelCreator(const std::vector<lite::tensor::
   return kernel;
 }
 
-REG_KERNEL(kCPU, PrimitiveType_Unstack, CpuUnstackFp32KernelCreator)
+REG_KERNEL(kCPU, kNumberTypeFloat32, PrimitiveType_Unstack, CpuUnstackFp32KernelCreator)
 }  // namespace mindspore::kernel
-

mindspore/lite/src/runtime/kernel/arm/fp32/where.cc

@@ -90,7 +90,7 @@ kernel::LiteKernel *CpuWhereFp32KernelCreator(const std::vector<lite::tensor::Te
     MS_LOG(ERROR) << "input opParameter is nullptr!";
     return nullptr;
   }
-
+  MS_ASSERT(desc.type == schema::PrimitiveType_Where);
   auto *kernel = new (std::nothrow) WhereCPUKernel(opParameter, inputs, outputs, ctx);
   if (kernel == nullptr) {
     MS_LOG(ERROR) << "new WhereCPUKernel fail!";
@@ -106,5 +106,5 @@ kernel::LiteKernel *CpuWhereFp32KernelCreator(const std::vector<lite::tensor::Te
   return kernel;
 }
 
-REG_KERNEL(kCPU, PrimitiveType_Where, CpuWhereFp32KernelCreator)
+REG_KERNEL(kCPU, kNumberTypeFloat32, PrimitiveType_Where, CpuWhereFp32KernelCreator)
 }  // namespace mindspore::kernel

mindspore/lite/src/runtime/kernel/arm/fp32/zeroslike.cc

@@ -48,6 +48,7 @@ kernel::LiteKernel *CpuZerosLikeFp32KernelCreator(const std::vector<lite::tensor
     MS_LOG(ERROR) << "input opParameter is nullptr!";
     return nullptr;
   }
+  MS_ASSERT(desc.type == schema::PrimitiveType_ZerosLike);
   auto *kernel = new (std::nothrow) ZerosLikeCPUKernel(opParameter, inputs, outputs);
   if (kernel == nullptr) {
     MS_LOG(ERROR) << "new ZerosLikeCPUKernel fail!";
@@ -63,6 +64,6 @@ kernel::LiteKernel *CpuZerosLikeFp32KernelCreator(const std::vector<lite::tensor
   return kernel;
 }
 
-REG_KERNEL(kCPU, PrimitiveType_ZerosLike, CpuZerosLikeFp32KernelCreator)
+REG_KERNEL(kCPU, kNumberTypeFloat32, PrimitiveType_ZerosLike, CpuZerosLikeFp32KernelCreator)
 }  // namespace mindspore::kernel
 

mindspore/lite/src/runtime/kernel/arm/int8/add_int8.cc

@@ -130,8 +130,10 @@ kernel::LiteKernel *CpuAddInt8KernelCreator(const std::vector<lite::tensor::Tens
   }
   MS_ASSERT(desc.type == PrimitiveType_Add);
   auto *kernel = new (std::nothrow) QuantizedAddCPUKernel(parameter, inputs, outputs, ctx);
-  MS_EXCEPTION_IF_NULL(kernel);
-
+  if (kernel == nullptr) {
+    MS_LOG(ERROR) << "kernel is nullptr.";
+    return nullptr;
+  }
   auto ret = kernel->Init();
   if (0 != ret) {
     MS_LOG(ERROR) << "Init kernel failed, name: " << parameter->name_
@@ -142,5 +144,6 @@ kernel::LiteKernel *CpuAddInt8KernelCreator(const std::vector<lite::tensor::Tens
   return kernel;
 }
 
+REG_KERNEL(kCPU, kNumberTypeInt8, PrimitiveType_Add, CpuAddInt8KernelCreator)
 }  // namespace mindspore::kernel
 

mindspore/lite/src/runtime/kernel/arm/int8/bias_add_int8.cc

@@ -80,5 +80,6 @@ kernel::LiteKernel *CpuBiasAddInt8KernelCreator(const std::vector<lite::tensor::
   return kernel;
 }
 
+REG_KERNEL(kCPU, kNumberTypeInt8, PrimitiveType_BiasAdd, CpuBiasAddInt8KernelCreator)
 }  // namespace mindspore::kernel
 

mindspore/lite/src/runtime/kernel/arm/int8/convolution_depthwise_int8.cc

@@ -25,6 +25,7 @@ using mindspore::kernel::KERNEL_ARCH::kCPU;
 using mindspore::lite::KernelRegistrar;
 using mindspore::lite::RET_ERROR;
 using mindspore::lite::RET_OK;
+using mindspore::schema::PrimitiveType_DepthwiseConv2D;
 
 namespace mindspore::kernel {
 int ConvolutionDepthwiseInt8CPUKernel::InitWeightBias() {
@@ -143,4 +144,27 @@ int ConvolutionDepthwiseInt8CPUKernel::Run() {
   }
   return RET_OK;
 }
+
+kernel::LiteKernel *CpuConvDwInt8KernelCreator(const std::vector<lite::tensor::Tensor *> &inputs,
+                                               const std::vector<lite::tensor::Tensor *> &outputs,
+                                               OpParameter *opParameter, const Context *ctx,
+                                               const kernel::KernelKey &desc) {
+  MS_ASSERT(opParameter != nullptr);
+  MS_ASSERT(desc.type == schema::PrimitiveType_DepthwiseConv2D);
+  auto kernel = new (std::nothrow) kernel::ConvolutionDepthwiseInt8CPUKernel(opParameter, inputs, outputs, ctx);
+  if (kernel == nullptr) {
+    MS_LOG(ERROR) << "kernel is nullptr.";
+    return nullptr;
+  }
+  auto ret = kernel->Init();
+  if (ret != RET_OK) {
+    delete kernel;
+    MS_LOG(ERROR) << "Init kernel failed, name: " << opParameter->name_ << ", type: "
+                  << schema::EnumNamePrimitiveType(static_cast<schema::PrimitiveType>(opParameter->type_));
+    return nullptr;
+  }
+  return kernel;
+}
+
+REG_KERNEL(kCPU, kNumberTypeInt8, PrimitiveType_DepthwiseConv2D, CpuConvDwInt8KernelCreator)
 }  // namespace mindspore::kernel

mindspore/lite/src/runtime/kernel/arm/int8/convolution_int8.cc

@@ -15,6 +15,7 @@
  */
 
 #include "src/runtime/kernel/arm/int8/convolution_int8.h"
+#include "src/runtime/kernel/arm/int8/convolution_3x3_int8.h"
 #include "src/runtime/kernel/arm/opclib/int8/conv_int8.h"
 #include "src/runtime/kernel/arm/base/layout_transform.h"
 #include "schema/model_generated.h"
@@ -36,7 +37,7 @@ void ConvolutionInt8CPUKernel::CheckSupportOptimize() {
   support_optimize_ = false;
 #endif
 
-#ifdef __aarch64__
+#ifdef ENABLE_ARM64
   void *optimize_op_handler = OptimizeModule::GetInstance()->optimized_op_handler_;
   if (optimize_op_handler != nullptr) {
     dlerror();
@@ -383,4 +384,39 @@ int ConvolutionInt8CPUKernel::Run() {
   }
   return RET_OK;
 }
+
+kernel::LiteKernel *CpuConvInt8KernelCreator(const std::vector<lite::tensor::Tensor *> &inputs,
+                                             const std::vector<lite::tensor::Tensor *> &outputs,
+                                             OpParameter *opParameter, const Context *ctx,
+                                             const kernel::KernelKey &desc) {
+  MS_ASSERT(opParameter != nullptr);
+  MS_ASSERT(desc.type == schema::PrimitiveType_Conv2D);
+  auto conv_param = reinterpret_cast<ConvParameter *>(opParameter);
+  int kernel_h = conv_param->kernel_h_;
+  int kernel_w = conv_param->kernel_w_;
+  int stride_h = conv_param->stride_h_;
+  int stride_w = conv_param->stride_w_;
+  int dilation_h = conv_param->dilation_h_;
+  int dilation_w = conv_param->dilation_w_;
+  kernel::LiteKernel *kernel;
+  if (kernel_h == 3 && kernel_w == 3 && stride_h == 1 && stride_w == 1 && dilation_h == 1 && dilation_w == 1) {
+    kernel = new (std::nothrow) kernel::Convolution3x3Int8CPUKernel(opParameter, inputs, outputs, ctx);
+  } else {
+    kernel = new (std::nothrow) kernel::ConvolutionInt8CPUKernel(opParameter, inputs, outputs, ctx);
+  }
+  if (kernel == nullptr) {
+    MS_LOG(ERROR) << "kernel is nullptr.";
+    return nullptr;
+  }
+  auto ret = kernel->Init();
+  if (ret != RET_OK) {
+    delete kernel;
+    MS_LOG(ERROR) << "Init kernel failed, name: " << opParameter->name_ << ", type: "
+                  << schema::EnumNamePrimitiveType(static_cast<schema::PrimitiveType>(opParameter->type_));
+    return nullptr;
+  }
+  return kernel;
+}
+
+REG_KERNEL(kCPU, kNumberTypeInt8, PrimitiveType_Conv2D, CpuConvInt8KernelCreator)
 }  // namespace mindspore::kernel

mindspore/lite/src/runtime/kernel/arm/int8/convolution_int8.h

@@ -72,4 +72,3 @@ class ConvolutionInt8CPUKernel : public ConvolutionBaseCPUKernel {
 }  // namespace mindspore::kernel
 
 #endif  // MINDSPORE_LITE_SRC_RUNTIME_KERNEL_ARM_INT8_CONVOLUTION_INT8_H_
-

mindspore/lite/src/runtime/kernel/arm/int8/deconvolution_depthwise_int8.cc

@@ -25,6 +25,7 @@ using mindspore::kernel::KERNEL_ARCH::kCPU;
 using mindspore::lite::KernelRegistrar;
 using mindspore::lite::RET_ERROR;
 using mindspore::lite::RET_OK;
+using mindspore::schema::PrimitiveType_DeDepthwiseConv2D;
 
 namespace mindspore::kernel {
 int DeconvolutionDepthwiseInt8CPUKernel::InitWeightBias() {
@@ -63,9 +64,9 @@ int DeconvolutionDepthwiseInt8CPUKernel::InitSlideParam() {
 
   sliding->in_h_step_ = conv_param_->input_w_ * C4NUM;
   sliding->in_sh_step_ = conv_param_->input_w_ * C4NUM * conv_param_->stride_h_;    // stride H
-  sliding->in_sw_step_ = C4NUM * conv_param_->stride_h_;                           // stride W
+  sliding->in_sw_step_ = C4NUM * conv_param_->stride_h_;                            // stride W
   sliding->in_kh_step_ = conv_param_->input_w_ * C4NUM * conv_param_->dilation_h_;  // kernel H
-  sliding->in_kw_step_ = C4NUM * conv_param_->dilation_w_;                         // kernel W
+  sliding->in_kw_step_ = C4NUM * conv_param_->dilation_w_;                          // kernel W
   return RET_OK;
 }
 
@@ -171,4 +172,27 @@ int DeconvolutionDepthwiseInt8CPUKernel::Run() {
   }
   return RET_OK;
 }
+
+kernel::LiteKernel *CpuDeconvDwInt8KernelCreator(const std::vector<lite::tensor::Tensor *> &inputs,
+                                                 const std::vector<lite::tensor::Tensor *> &outputs,
+                                                 OpParameter *opParameter, const lite::Context *ctx,
+                                                 const kernel::KernelKey &desc) {
+  MS_ASSERT(opParameter != nullptr);
+  MS_ASSERT(desc.type == schema::PrimitiveType_DeDepthwiseConv2D);
+  auto kernel = new (std::nothrow) kernel::DeconvolutionDepthwiseInt8CPUKernel(opParameter, inputs, outputs, ctx);
+  if (kernel == nullptr) {
+    MS_LOG(ERROR) << "kernel is nullptr.";
+    return nullptr;
+  }
+  auto ret = kernel->Init();
+  if (ret != RET_OK) {
+    delete kernel;
+    MS_LOG(ERROR) << "Init kernel failed, name: " << opParameter->name_ << ", type: "
+                  << schema::EnumNamePrimitiveType(static_cast<schema::PrimitiveType>(opParameter->type_));
+    return nullptr;
+  }
+  return kernel;
+}
+
+REG_KERNEL(kCPU, kNumberTypeInt8, PrimitiveType_DeDepthwiseConv2D, CpuDeconvDwInt8KernelCreator)
 }  // namespace mindspore::kernel

mindspore/lite/src/runtime/kernel/arm/int8/deconvolution_int8.cc

@@ -17,8 +17,10 @@
 #include "src/runtime/kernel/arm/int8/deconvolution_int8.h"
 #include "src/runtime/kernel/arm/opclib/quantization/fixed_point.h"
 #include "src/runtime/runtime_api.h"
+#include "src/kernel_registry.h"
 
 using mindspore::kernel::KERNEL_ARCH::kCPU;
+using mindspore::lite::KernelRegistrar;
 using mindspore::lite::RET_ERROR;
 using mindspore::lite::RET_MEMORY_FAILED;
 using mindspore::lite::RET_OK;
@@ -216,5 +218,27 @@ int DeConvInt8CPUKernel::Run() {
 
   return RET_OK;
 }
-}  // namespace mindspore::kernel
 
+kernel::LiteKernel *CpuDeConvInt8KernelCreator(const std::vector<lite::tensor::Tensor *> &inputs,
+                                               const std::vector<lite::tensor::Tensor *> &outputs,
+                                               OpParameter *opParameter, const lite::Context *ctx,
+                                               const kernel::KernelKey &desc) {
+  MS_ASSERT(opParameter != nullptr);
+  MS_ASSERT(desc.type == schema::PrimitiveType_DeConv2D);
+  auto kernel = new (std::nothrow) kernel::DeConvInt8CPUKernel(opParameter, inputs, outputs, ctx);
+  if (kernel == nullptr) {
+    MS_LOG(ERROR) << "kernel is nullptr.";
+    return nullptr;
+  }
+  auto ret = kernel->Init();
+  if (ret != RET_OK) {
+    delete kernel;
+    MS_LOG(ERROR) << "Init kernel failed, name: " << opParameter->name_ << ", type: "
+                  << schema::EnumNamePrimitiveType(static_cast<schema::PrimitiveType>(opParameter->type_));
+    return nullptr;
+  }
+  return kernel;
+}
+
+REG_KERNEL(kCPU, kNumberTypeInt8, PrimitiveType_DeConv2D, CpuDeConvInt8KernelCreator)
+}  // namespace mindspore::kernel

mindspore/lite/src/runtime/kernel/arm/int8/mul_int8.cc

@@ -129,4 +129,5 @@ kernel::LiteKernel *CpuMulInt8KernelCreator(const std::vector<lite::tensor::Tens
   return kernel;
 }
 
+REG_KERNEL(kCPU, kNumberTypeInt8, PrimitiveType_Mul, CpuMulInt8KernelCreator)
 }  // namespace mindspore::kernel

mindspore/lite/src/runtime/kernel/arm/opclib/CMakeLists.txt

@@ -10,14 +10,13 @@ file(GLOB OPTIMIZED_ASSEMBLY
         ${OPTIMIZED_OP_DIR}/assembly/opt/*.S
         )
 
-
 file(GLOB FP16_SRC
-        #        ${OPTIMIZED_OP_DIR}/fp16/*.cc
-        #        ${OPTIMIZED_OP_DIR}/../fp16/*.cc
+        ${OPTIMIZED_OP_DIR}/fp16/*.cc
+        ${OPTIMIZED_OP_DIR}/../fp16/*.cc
         )
 
 ########################### share library build ########################
-set(OPTIMIZED_OPS "opt_op_handler.c")
+set(OPTIMIZED_OPS ${OPTIMIZED_OP_DIR}/opt_op_handler.c)
 
 set_property(SOURCE ${OPTIMIZED_ASSEMBLY} PROPERTY LANGUAGE C)
 list(APPEND OPTIMIZED_OPS ${OPTIMIZED_ASSEMBLY} ${FP16_SRC})
@@ -27,6 +26,10 @@ if (PLATFORM_ARM64)
     set(CMAKE_C_FLAGS "${CMAKE_C_FLAGS} -march=armv8.2-a+dotprod+fp16")
     set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -march=armv8.2-a+dotprod+fp16")
     add_library(optimize SHARED ${OPTIMIZED_OPS})
+    target_link_libraries(
+            optimize
+            mindspore-lite
+    )
     set_target_properties(optimize PROPERTIES CLEAN_DIRECT_OUTPUT 1)
 
     add_custom_command(TARGET optimize POST_BUILD

mindspore/lite/src/runtime/kernel/arm/opclib/assembly/arm32/IndirectGemmInt16to32_8x4.S

@@ -1,5 +1,4 @@
-#ifdef __arm__
-#ifndef __aarch64__
+#ifdef ENABLE_ARM32
 
 .text
 .align 5
@@ -236,5 +235,4 @@ IndirectGemmInt16to32_8x4:
     pop {r4-r8, r10, pc}
 
 #endif
-#endif
 

mindspore/lite/src/runtime/kernel/arm/opclib/fp16/conv_depthwise_fp16.cc

@@ -15,7 +15,6 @@
  */
 
 #include "src/runtime/kernel/arm/opclib/fp16/conv_depthwise_fp16.h"
-#ifdef ENABLE_FP16
 #include <arm_neon.h>
 
 /*conv depthwise fp16 begin*/
@@ -299,4 +298,3 @@ void DeconvDwC8Fp16(float16_t *output_data, const float16_t *input_data, const f
 }
 /*deconv depthwise fp16 end*/
 
-#endif

mindspore/lite/src/runtime/kernel/arm/opclib/fp16/conv_depthwise_fp16.h

@@ -20,7 +20,6 @@
 #include "src/runtime/kernel/arm/opclib/conv_parameter.h"
 #include "src/runtime/kernel/arm/opclib/fp32/conv_depthwise.h"
 
-#ifdef ENABLE_FP16
 void ConvDwC8Fp16(float16_t *output_data, const float16_t *input_data, const float16_t *weight_data,
                   const float16_t *bias_data, const ConvParameter *conv_param, const SlidingWindowParam *sliding,
                   int task_id);
@@ -28,6 +27,5 @@ void ConvDwC8Fp16(float16_t *output_data, const float16_t *input_data, const flo
 void DeconvDwC8Fp16(float16_t *output_data, const float16_t *input_data, const float16_t *weight_data,
                     const float16_t *bias_data, const ConvParameter *conv_param, const SlidingWindowParam *sliding,
                     int task_id);
-#endif
 
 #endif  // MINDSPORE_LITE_SRC_RUNTIME_KERNEL_ARM_OPCLIB_FP16_CONV_DEPTHWISE_FP16_H_

mindspore/lite/src/runtime/kernel/arm/opclib/fp16/conv_fp16.cc

@@ -15,20 +15,17 @@
  */
 #include "src/runtime/kernel/arm/opclib/fp16/conv_fp16.h"
 #include <string.h>
-#include "src/runtime/kernel/arm/opclib/pack.h"
-#include "src/runtime/kernel/arm/opclib/winograd_transform.h"
+#include "src/runtime/kernel/arm/opclib/fp16/pack_fp16.h"
+#include "src/runtime/kernel/arm/opclib/fp16/winograd_transform_fp16.h"
 
 extern "C" {
 #ifdef ENABLE_ARM64
-#ifdef ENABLE_FP16
 void IndirectGemmFp16_16x8(float16_t *output, float16_t *input, float16_t *weight, float16_t *bias, size_t step,
                            size_t ic4, size_t oc8, size_t offset, size_t mode, size_t writeC4, size_t relu,
                            size_t relu6);
 #endif
-#endif
 }
 
-#ifdef ENABLE_FP16
 #ifndef ENABLE_NEON
 void IndirectGemmFp16_16x8(float16_t *output, float16_t *input, float16_t *weight, float16_t *bias, size_t step,
                            size_t ic4, size_t out_channel, size_t offset, size_t mode, size_t writeC4, size_t relu,
@@ -215,5 +212,5 @@ void Conv3x3Fp16(float16_t *input_data, float16_t *transed_weight, const float16
     }
   }
 }
-#endif
+
 

mindspore/lite/src/runtime/kernel/arm/opclib/fp16/conv_fp16.h

@@ -16,12 +16,9 @@
 #ifndef MINDSPORE_LITE_SRC_RUNTIME_KERNEL_ARM_OPCLIB_FP16_CONV_FP16_H_
 #define MINDSPORE_LITE_SRC_RUNTIME_KERNEL_ARM_OPCLIB_FP16_CONV_FP16_H_
 
-#ifdef ENABLE_FP16
 #include <arm_neon.h>
-#endif
 #include "src/runtime/kernel/arm/opclib/conv_parameter.h"
 
-#ifdef ENABLE_FP16
 #ifndef ENABLE_NEON
 void IndirectGemmFp16_16x8(float16_t *output, float16_t *input, float16_t *weight, float16_t *bias, size_t step,
                            size_t ic4, size_t oc8, size_t offset, size_t mode, size_t writeC4, size_t relu,
@@ -36,7 +33,6 @@ void ConvFp16(float16_t *input_data, float16_t *packed_input, float16_t *packed_
 void Conv3x3Fp16(float16_t *input_data, float16_t *transed_weight, const float16_t *bias_data, float16_t *output_data,
                  float16_t *tile_buffer, float16_t *block_unit_buffer, float16_t *tmp_dst_buffer, float16_t *tmp_out,
                  int task_id, ConvParameter *conv_param);
-#endif
 
 #endif  // MINDSPORE_LITE_SRC_RUNTIME_KERNEL_ARM_OPCLIB_FP16_CONV_FP16_H_
 

mindspore/lite/src/runtime/kernel/arm/opclib/fp16/pack_fp16.cc

@@ -0,0 +1,342 @@
+/**
+ * Copyright 2020 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/kernel/arm/opclib/fp16/pack_fp16.h"
+#include <cstring>
+#include <cstdlib>
+
+void Im2ColPackUnitFp16(float16_t *input_data, ConvParameter *conv_param, float16_t *packed_input, int real_cal_num,
+                        int block_index) {
+  // input format : nhwc
+  int kernel_h = conv_param->kernel_h_;
+  int kernel_w = conv_param->kernel_w_;
+  int stride_h = conv_param->stride_h_;
+  int stride_w = conv_param->stride_w_;
+  int pad_h = conv_param->pad_h_;
+  int pad_w = conv_param->pad_w_;
+  int dilation_h = conv_param->dilation_h_;
+  int dilation_w = conv_param->dilation_w_;
+  int in_channel = conv_param->input_channel_;
+  int in_h = conv_param->input_h_;
+  int in_w = conv_param->input_w_;
+  int out_w = conv_param->output_w_;
+  int channel_block = UP_DIV(in_channel, 4);
+  int kernel_plane = kernel_h * kernel_w;
+
+  for (int i = 0; i < real_cal_num; i++) {
+    int block_start = block_index + i;
+    int input_h = block_start / out_w * stride_h - pad_h;
+    int input_w = block_start % out_w * stride_w - pad_w;
+    for (int j = 0; j < kernel_h; j++) {
+      int input_y = input_h + j * dilation_h;
+      if (input_y < 0 || input_y >= in_h) {
+        continue;
+      }
+      int input_y_stride = input_y * in_w * channel_block * C4NUM;
+      for (int n = 0; n < kernel_w; n++) {
+        int input_x = input_w + n * dilation_w;
+        if (input_x < 0 || input_x >= in_w) {
+          continue;
+        }
+        int input_x_stride = input_y_stride + input_x * channel_block * C4NUM;
+        int input_plane_offset = (j * kernel_w + n) * 16 * C4NUM * channel_block + i * C4NUM;
+        for (int m = 0; m < channel_block; m++) {
+          int channel_block_stride = input_x_stride + m * C4NUM;
+          int channel_block_offset = input_plane_offset + m * 16 * C4NUM;
+          (packed_input + channel_block_offset)[0] = (input_data + channel_block_stride)[0];
+          (packed_input + channel_block_offset)[1] = (input_data + channel_block_stride)[1];
+          (packed_input + channel_block_offset)[2] = (input_data + channel_block_stride)[2];
+          (packed_input + channel_block_offset)[3] = (input_data + channel_block_stride)[3];
+        }  // channel_block loop
+      }    // kernel_w loop
+    }      // kernel_h loop
+  }        // tile num loop
+}
+
+void PackWeightFp16(float16_t *weight_data, ConvParameter *conv_param, float16_t *packed_weight) {
+  // original weight format : ohwi
+  int tile_num = 8;
+  int inchannel_block = 4;
+  int kernel_h = conv_param->kernel_h_;
+  int kernel_w = conv_param->kernel_w_;
+  int in_channel = conv_param->input_channel_;
+  int out_channel = conv_param->output_channel_;
+  int kernel_block = UP_DIV(out_channel, tile_num);
+  int channel_block = UP_DIV(in_channel, inchannel_block);
+  int kernel_plane = kernel_h * kernel_w;
+  int pack_weight_size = kernel_block * channel_block * tile_num * inchannel_block * kernel_plane;
+
+  int unit_size = tile_num * inchannel_block;
+  int block_size = pack_weight_size / kernel_block;
+
+  for (int m = 0; m < kernel_plane; m++) {
+    int kernel_plane_stride = m * in_channel;
+    int packed_kernel_plane_stride = m * unit_size * channel_block;
+    for (int i = 0; i < channel_block; i++) {
+      int channel_block_stride = kernel_plane_stride + i * inchannel_block;
+      int packed_channel_block_size = packed_kernel_plane_stride + i * unit_size;
+      int ic_remainder = in_channel - i * inchannel_block;
+      int real_ic_num = ic_remainder < inchannel_block ? ic_remainder : inchannel_block;
+      for (int h = 0; h < real_ic_num; h++) {
+        int block_stride = channel_block_stride + h;
+        int packed_block_stride = packed_channel_block_size + h * tile_num;
+        for (int j = 0; j < kernel_block; j++) {
+          int kernel_block_stride = block_stride + j * tile_num * kernel_plane * in_channel;
+          int packed_kernel_block_size = packed_block_stride + j * block_size;
+          int oc_remainder = out_channel - j * tile_num;
+          int real_oc_num = oc_remainder < tile_num ? oc_remainder : tile_num;
+          for (int k = 0; k < real_oc_num; k++) {
+            float16_t *origin_data_ptr = weight_data + kernel_block_stride + k * kernel_plane * in_channel;
+            float16_t *packed_data_ptr = packed_weight + packed_kernel_block_size + k;
+            *packed_data_ptr = *origin_data_ptr;
+          }
+        }  // kernel block loop
+      }    // inchannel block loop
+    }      // channel block loop
+  }        // kernel plane loop
+}
+
+void PackWeightToC8Fp16(const float16_t *origin_weight_data, float16_t *packed_weight_data, ConvParameter *conv_param) {
+  // origin weight format : ohwi
+  int input_channel = conv_param->input_channel_;
+  int ic8 = UP_DIV(input_channel, C8NUM);
+  int output_channel = conv_param->output_channel_;
+  int kernel_plane = conv_param->kernel_h_ * conv_param->kernel_w_;
+
+  for (int k = 0; k < kernel_plane; k++) {
+    int src_kernel_offset = k * input_channel;
+    int dst_kernel_offset = k * C8NUM;
+    for (int o = 0; o < output_channel; o++) {
+      int src_oc_offset = src_kernel_offset + o * kernel_plane * input_channel;
+      int dst_oc_offset = dst_kernel_offset + o * ic8 * kernel_plane * C8NUM;
+      for (int i = 0; i < input_channel; i++) {
+        int c8_block_num = i / C8NUM;
+        int c8_block_rem = i % C8NUM;
+        int src_ic_offset = src_oc_offset + i;
+        int dst_ic_offset = dst_oc_offset + c8_block_num * kernel_plane * C8NUM + c8_block_rem;
+        (packed_weight_data + dst_ic_offset)[0] = (origin_weight_data + src_ic_offset)[0];
+      }
+    }
+  }
+}
+
+void PackWeightToC4Fp16(const float16_t *origin_weight_data, float16_t *packed_weight_data, ConvParameter *conv_param) {
+  // origin weight format : ohwi
+  int input_channel = conv_param->input_channel_;
+  int ic4 = UP_DIV(input_channel, C4NUM);
+  int output_channel = conv_param->output_channel_;
+  int kernel_plane = conv_param->kernel_h_ * conv_param->kernel_w_;
+
+  for (int k = 0; k < kernel_plane; k++) {
+    int src_kernel_offset = k * input_channel;
+    int dst_kernel_offset = k * C4NUM;
+    for (int o = 0; o < output_channel; o++) {
+      int src_oc_offset = src_kernel_offset + o * kernel_plane * input_channel;
+      int dst_oc_offset = dst_kernel_offset + o * ic4 * kernel_plane * C4NUM;
+      for (int i = 0; i < input_channel; i++) {
+        int c4_block_num = i / C4NUM;
+        int c4_block_rem = i % C4NUM;
+        int src_ic_offset = src_oc_offset + i;
+        int dst_ic_offset = dst_oc_offset + c4_block_num * kernel_plane * C4NUM + c4_block_rem;
+        (packed_weight_data + dst_ic_offset)[0] = (origin_weight_data + src_ic_offset)[0];
+      }
+    }
+  }
+}
+
+void PackNHWCToNC4HW4Fp16(const void *src, void *dst, int batch, int plane, int channel) {
+  int c4 = UP_DIV(channel, C4NUM);
+  for (int b = 0; b < batch; b++) {
+    int src_oc_offset = b * plane * channel;
+    int dst_oc_offset = b * plane * c4 * C4NUM;
+    for (int k = 0; k < plane; k++) {
+      int src_kernel_offset = src_oc_offset + k * channel;
+      int dst_kernel_offset = dst_oc_offset + k * C4NUM;
+      for (int i = 0; i < channel; i++) {
+        int c4_block_num = i / C4NUM;
+        int c4_block_rem = i % C4NUM;
+        int src_ic_offset = src_kernel_offset + i;
+        int dst_ic_offset = dst_kernel_offset + c4_block_num * plane * C4NUM + c4_block_rem;
+        ((float16_t *)dst + dst_ic_offset)[0] = ((float16_t *)src + src_ic_offset)[0];
+      }
+    }
+  }
+}
+
+void PackNCHWToNC4HW4Fp16(const void *src, void *dst, int batch, int plane, int channel) {
+  int c4 = UP_DIV(channel, C4NUM);
+  for (int b = 0; b < batch; b++) {
+    int src_offset = b * plane * channel;
+    int dst_offset = b * plane * c4 * C4NUM;
+    for (int c = 0; c < channel; c++) {
+      int c4_block_num = c / C4NUM;
+      int c4_block_rem = c % C4NUM;
+      int src_c_offset = src_offset + c * plane;
+      int dst_c_offset = dst_offset + c4_block_num * plane * C4NUM;
+      for (int k = 0; k < plane; k++) {
+        int src_kernel_offset = src_c_offset + k;
+        int dst_kernel_offset = dst_c_offset + C4NUM * k + c4_block_rem;
+        ((float16_t *)dst + dst_kernel_offset)[0] = ((float16_t *)src + src_kernel_offset)[0];
+      }
+    }
+  }
+}
+
+void PackNHWCToNHWC4Fp16(const void *src, void *dst, int batch, int plane, int channel) {
+  int ic4 = UP_DIV(channel, C4NUM);
+  int nhwc4_batch_unit_offset = ic4 * C4NUM * plane;
+  int ic_remainder_ = channel % C4NUM;
+  if (ic_remainder_ != 0) {
+    int nhwc4_batch_offset = 0;
+    for (int b = 0; b < batch; b++) {
+      int batch_offset = b * channel * plane;
+      for (int i = 0; i < plane; i++) {
+        memcpy((float16_t *)dst + nhwc4_batch_offset + i * ic4 * C4NUM, (float16_t *)src + batch_offset + i * channel,
+               channel * sizeof(float16_t));
+      }
+      nhwc4_batch_offset += nhwc4_batch_unit_offset;
+    }
+  } else {
+    size_t ori_input_size = batch * plane * channel * sizeof(float16_t);
+    memcpy(dst, src, ori_input_size);
+  }
+}
+
+void PackNCHWToNHWC4Fp16(const void *src, void *dst, int batch, int plane, int channel) {
+  int nhwc4_batch_offset = 0;
+  int ic4 = UP_DIV(channel, C4NUM);
+  int nhwc4_batch_unit_offset = ic4 * C4NUM * plane;
+
+  for (int b = 0; b < batch; b++) {
+    int batch_offset = b * channel * plane;
+    for (int c = 0; c < channel; c++) {
+      int src_c_offset = batch_offset + c * plane;
+      int dst_c_offset = nhwc4_batch_offset + c;
+      for (int i = 0; i < plane; i++) {
+        int src_plane_offset = src_c_offset + i;
+        int dst_plane_offset = dst_c_offset + i * ic4 * C4NUM;
+        ((float16_t *)dst)[dst_plane_offset] = ((float16_t *)src)[src_plane_offset];
+      }
+    }
+    nhwc4_batch_offset += nhwc4_batch_unit_offset;
+  }
+}
+
+void PackNC4HW4ToNHWC4Fp16(const void *src, void *dst, int batch, int plane, int channel) {
+  int c4 = UP_DIV(channel, C4NUM);
+  for (int b = 0; b < batch; b++) {
+    int src_offset = b * plane * c4 * C4NUM;
+    int dst_offset = b * plane * channel;
+    for (int c = 0; c < channel; c++) {
+      int c4_block_num = c / C4NUM;
+      int c4_block_res = c % C4NUM;
+      int src_c_offset = src_offset + c4_block_num * plane * C4NUM + c4_block_res;
+      int dst_c_offset = dst_offset + c4_block_num * C4NUM + c4_block_res;
+      for (int k = 0; k < plane; k++) {
+        int src_kernel_offset = src_c_offset + k * C4NUM;
+        int dst_kernel_offset = dst_c_offset + k * c4 * C4NUM;
+        ((float16_t *)dst + dst_kernel_offset)[0] = ((float16_t *)src + src_kernel_offset)[0];
+      }
+    }
+  }
+}
+
+void PackNC4HW4ToNHWCFp16(const void *src, void *dst, int batch, int plane, int channel) {
+  int c4 = UP_DIV(channel, C4NUM);
+  for (int b = 0; b < batch; b++) {
+    int src_offset = b * plane * c4 * C4NUM;
+    int dst_offset = b * plane * channel;
+    for (int c = 0; c < channel; c++) {
+      int c4_block_num = c / C4NUM;
+      int c4_block_res = c % C4NUM;
+      int src_c_offset = src_offset + c4_block_num * plane * C4NUM + c4_block_res;
+      int dst_c_offset = dst_offset + c;
+      for (int k = 0; k < plane; k++) {
+        int src_kernel_offset = src_c_offset + k * C4NUM;
+        int dst_kernel_offset = dst_c_offset + k * channel;
+        ((float16_t *)dst + dst_kernel_offset)[0] = ((float16_t *)src + src_kernel_offset)[0];
+      }
+    }
+  }
+}
+
+void PackNC4HW4ToNCHWFp16(const void *src, void *dst, int batch, int plane, int channel) {
+  int c4 = UP_DIV(channel, C4NUM);
+  for (int b = 0; b < batch; b++) {
+    int src_offset = b * plane * c4 * C4NUM;
+    int dst_offset = b * plane * channel;
+    for (int c = 0; c < channel; c++) {
+      int c4_block_num = c / C4NUM;
+      int c4_block_res = c % C4NUM;
+      int src_c_offset = src_offset + c4_block_num * plane * C4NUM + c4_block_res;
+      int dst_c_offset = dst_offset + c * plane;
+      for (int k = 0; k < plane; k++) {
+        int src_kernel_offset = src_c_offset + k * C4NUM;
+        int dst_kernel_offset = dst_c_offset + k;
+        ((float16_t *)dst + dst_kernel_offset)[0] = ((float16_t *)src + src_kernel_offset)[0];
+      }
+    }
+  }
+}
+
+void PackNCHWFp32ToNC8HW8Fp16(float *src, float16_t *dst, int batch, int plane, int channel) {
+  int c8 = UP_DIV(channel, C8NUM);
+  for (int b = 0; b < batch; b++) {
+    int src_offset = b * plane * channel;
+    int dst_offset = b * plane * c8 * C8NUM;
+    for (int c = 0; c < channel; c++) {
+      int c8_block_num = c / C8NUM;
+      int c8_block_rem = c % C8NUM;
+      int src_c_offset = src_offset + c * plane;
+      int dst_c_offset = dst_offset + c8_block_num * plane * C8NUM;
+      for (int k = 0; k < plane; k++) {
+        int src_kernel_offset = src_c_offset + k;
+        int dst_kernel_offset = dst_c_offset + C8NUM * k + c8_block_rem;
+        (dst + dst_kernel_offset)[0] = (float16_t)(src + src_kernel_offset)[0];
+      }
+    }
+  }
+}
+
+void PackNHWCFp32ToNHWC8Fp16(float *src, float16_t *dst, int batch, int plane, int channel) {
+  int c8 = UP_DIV(channel, C8NUM);
+  int nhwc8_batch_unit_offset = c8 * C8NUM * plane;
+  int nhwc8_batch_offset = 0;
+  for (int b = 0; b < batch; b++) {
+    int batch_offset = b * channel * plane;
+    for (int i = 0; i < plane; i++) {
+      for (int c = 0; c < channel; c++) {
+        (dst + nhwc8_batch_offset + i * c8 * C8NUM)[c] = (float16_t)(src + batch_offset + i * channel)[c];
+      }
+    }
+    nhwc8_batch_offset += nhwc8_batch_unit_offset;
+  }
+}
+
+void PackNHWC8Fp16ToNHWCFp32(float16_t *src, float *dst, int batch, int plane, int channel) {
+  int c8 = UP_DIV(channel, C8NUM);
+  int nhwc_batch_unit_offset = channel * plane;
+  int nhwc_batch_offset = 0;
+  for (int b = 0; b < batch; b++) {
+    int batch_offset = b * c8 * C8NUM * plane;
+    for (int i = 0; i < plane; i++) {
+      for (int c = 0; c < channel; c++) {
+        (dst + nhwc_batch_offset + i * channel)[c] = (float)(src + batch_offset + i * c8 * C8NUM)[c];
+      }
+    }
+    nhwc_batch_offset += nhwc_batch_unit_offset;
+  }
+}

mindspore/lite/src/runtime/kernel/arm/opclib/fp16/pack_fp16.h

@@ -0,0 +1,57 @@
+/**
+ * Copyright 2020 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_KERNEL_ARM_OPCLIB_FP16_PACK_FP16_H_
+#define MINDSPORE_LITE_SRC_RUNTIME_KERNEL_ARM_OPCLIB_FP16_PACK_FP16_H_
+
+#ifdef ENABLE_NEON
+#include <arm_neon.h>
+#endif
+#include "src/runtime/kernel/arm/opclib/conv_parameter.h"
+#include "src/runtime/kernel/arm/opclib/op_base.h"
+
+void Im2ColPackUnitFp16(float16_t *input_data, ConvParameter *conv_param, float16_t *packed_input, int real_cal_num,
+                        int block_index);
+
+void PackWeightFp16(float16_t *weight_data, ConvParameter *conv_param, float16_t *packed_weight);
+
+void PackWeightToC8Fp16(const float16_t *origin_weight_data, float16_t *packed_weight_data, ConvParameter *conv_param);
+
+void PackWeightToC4Fp16(const float16_t *origin_weight_data, float16_t *packed_weight_data, ConvParameter *conv_param);
+
+void PackNHWCToNC4HW4Fp16(const void *src, void *dst, int batch, int plane, int channel);
+
+void PackNCHWToNC4HW4Fp16(const void *src, void *dst, int batch, int plane, int channel);
+
+void PackNHWCToNHWC4Fp16(const void *src, void *dst, int batch, int plane, int channel);
+
+void PackNCHWToNHWC4Fp16(const void *src, void *dst, int batch, int plane, int channel);
+
+void PackNC4HW4ToNHWC4Fp16(const void *src, void *dst, int batch, int plane, int channel);
+
+void PackNC4HW4ToNHWCFp16(const void *src, void *dst, int batch, int plane, int channel);
+
+void PackNC4HW4ToNCHWFp16(const void *src, void *dst, int batch, int plane, int channel);
+
+void PackNC8HW8ToNHWCFp16(const void *src, void *dst, int batch, int plane, int channel);
+
+void PackNCHWFp32ToNC8HW8Fp16(float *src, float16_t *dst, int batch, int plane, int channel);
+
+void PackNHWCFp32ToNHWC8Fp16(float *src, float16_t *dst, int batch, int plane, int channel);
+
+void PackNHWC8Fp16ToNHWCFp32(float16_t *src, float *dst, int batch, int plane, int channel);
+
+#endif  // MINDSPORE_LITE_SRC_RUNTIME_KERNEL_ARM_OPCLIB_FP16_PACK_FP16_H_

mindspore/lite/src/runtime/kernel/arm/opclib/fp16/winograd_transform_fp16.cc

@@ -0,0 +1,527 @@
+/**
+ * Copyright 2020 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/kernel/arm/opclib/fp16/winograd_transform_fp16.h"
+
+// for fp16 convolution 3x3 filter/input/output transform F(4,3)
+void Conv3x3Fp16InputUnit(float16_t *tmp_data, float16_t *trans_input_data, size_t step) {
+  float16x4_t d00 = vld1_f16(tmp_data);
+  float16x4_t d01 = vld1_f16(tmp_data + 4);
+  float16x4_t d02 = vld1_f16(tmp_data + 2 * 4);
+  float16x4_t d03 = vld1_f16(tmp_data + 3 * 4);
+  float16x4_t d04 = vld1_f16(tmp_data + 4 * 4);
+  float16x4_t d05 = vld1_f16(tmp_data + 5 * 4);
+
+  float16x4_t d10 = vld1_f16(tmp_data + 6 * 4);
+  float16x4_t d11 = vld1_f16(tmp_data + 7 * 4);
+  float16x4_t d12 = vld1_f16(tmp_data + 8 * 4);
+  float16x4_t d13 = vld1_f16(tmp_data + 9 * 4);
+  float16x4_t d14 = vld1_f16(tmp_data + 10 * 4);
+  float16x4_t d15 = vld1_f16(tmp_data + 11 * 4);
+
+  float16x4_t d20 = vld1_f16(tmp_data + 12 * 4);
+  float16x4_t d21 = vld1_f16(tmp_data + 13 * 4);
+  float16x4_t d22 = vld1_f16(tmp_data + 14 * 4);
+  float16x4_t d23 = vld1_f16(tmp_data + 15 * 4);
+  float16x4_t d24 = vld1_f16(tmp_data + 16 * 4);
+  float16x4_t d25 = vld1_f16(tmp_data + 17 * 4);
+
+  float16x4_t d30 = vld1_f16(tmp_data + 18 * 4);
+  float16x4_t d31 = vld1_f16(tmp_data + 19 * 4);
+  float16x4_t d32 = vld1_f16(tmp_data + 20 * 4);
+  float16x4_t d33 = vld1_f16(tmp_data + 21 * 4);
+  float16x4_t d34 = vld1_f16(tmp_data + 22 * 4);
+  float16x4_t d35 = vld1_f16(tmp_data + 23 * 4);
+
+  float16x4_t d40 = vld1_f16(tmp_data + 24 * 4);
+  float16x4_t d41 = vld1_f16(tmp_data + 25 * 4);
+  float16x4_t d42 = vld1_f16(tmp_data + 26 * 4);
+  float16x4_t d43 = vld1_f16(tmp_data + 27 * 4);
+  float16x4_t d44 = vld1_f16(tmp_data + 28 * 4);
+  float16x4_t d45 = vld1_f16(tmp_data + 29 * 4);
+
+  float16x4_t d50 = vld1_f16(tmp_data + 30 * 4);
+  float16x4_t d51 = vld1_f16(tmp_data + 31 * 4);
+  float16x4_t d52 = vld1_f16(tmp_data + 32 * 4);
+  float16x4_t d53 = vld1_f16(tmp_data + 33 * 4);
+  float16x4_t d54 = vld1_f16(tmp_data + 34 * 4);
+  float16x4_t d55 = vld1_f16(tmp_data + 35 * 4);
+
+  float16x4_t t00 = vadd_f16(vsub_f16(vmul_n_f16(d00, 4), vmul_n_f16(d20, 5)), d40);
+  float16x4_t t01 = vadd_f16(vsub_f16(vmul_n_f16(d01, 4), vmul_n_f16(d21, 5)), d41);
+  float16x4_t t02 = vadd_f16(vsub_f16(vmul_n_f16(d02, 4), vmul_n_f16(d22, 5)), d42);
+  float16x4_t t03 = vadd_f16(vsub_f16(vmul_n_f16(d03, 4), vmul_n_f16(d23, 5)), d43);
+  float16x4_t t04 = vadd_f16(vsub_f16(vmul_n_f16(d04, 4), vmul_n_f16(d24, 5)), d44);
+  float16x4_t t05 = vadd_f16(vsub_f16(vmul_n_f16(d05, 4), vmul_n_f16(d25, 5)), d45);
+
+  float16x4_t t10 = vadd_f16(vadd_f16(d30, d40), vmul_n_f16(vadd_f16(d10, d20), -4));
+  float16x4_t t11 = vadd_f16(vadd_f16(d31, d41), vmul_n_f16(vadd_f16(d11, d21), -4));
+  float16x4_t t12 = vadd_f16(vadd_f16(d32, d42), vmul_n_f16(vadd_f16(d12, d22), -4));
+  float16x4_t t13 = vadd_f16(vadd_f16(d33, d43), vmul_n_f16(vadd_f16(d13, d23), -4));
+  float16x4_t t14 = vadd_f16(vadd_f16(d34, d44), vmul_n_f16(vadd_f16(d14, d24), -4));
+  float16x4_t t15 = vadd_f16(vadd_f16(d35, d45), vmul_n_f16(vadd_f16(d15, d25), -4));
+
+  float16x4_t t20 = vadd_f16(vsub_f16(d40, d30), vmul_n_f16(vsub_f16(d10, d20), 4));
+  float16x4_t t21 = vadd_f16(vsub_f16(d41, d31), vmul_n_f16(vsub_f16(d11, d21), 4));
+  float16x4_t t22 = vadd_f16(vsub_f16(d42, d32), vmul_n_f16(vsub_f16(d12, d22), 4));
+  float16x4_t t23 = vadd_f16(vsub_f16(d43, d33), vmul_n_f16(vsub_f16(d13, d23), 4));
+  float16x4_t t24 = vadd_f16(vsub_f16(d44, d34), vmul_n_f16(vsub_f16(d14, d24), 4));
+  float16x4_t t25 = vadd_f16(vsub_f16(d45, d35), vmul_n_f16(vsub_f16(d15, d25), 4));
+
+  float16x4_t t30 = vadd_f16(vsub_f16(d40, d20), vmul_n_f16(vsub_f16(d30, d10), 2));
+  float16x4_t t31 = vadd_f16(vsub_f16(d41, d21), vmul_n_f16(vsub_f16(d31, d11), 2));
+  float16x4_t t32 = vadd_f16(vsub_f16(d42, d22), vmul_n_f16(vsub_f16(d32, d12), 2));
+  float16x4_t t33 = vadd_f16(vsub_f16(d43, d23), vmul_n_f16(vsub_f16(d33, d13), 2));
+  float16x4_t t34 = vadd_f16(vsub_f16(d44, d24), vmul_n_f16(vsub_f16(d34, d14), 2));
+  float16x4_t t35 = vadd_f16(vsub_f16(d45, d25), vmul_n_f16(vsub_f16(d35, d15), 2));
+
+  float16x4_t t40 = vadd_f16(vsub_f16(d40, d20), vmul_n_f16(vsub_f16(d10, d30), 2));
+  float16x4_t t41 = vadd_f16(vsub_f16(d41, d21), vmul_n_f16(vsub_f16(d11, d31), 2));
+  float16x4_t t42 = vadd_f16(vsub_f16(d42, d22), vmul_n_f16(vsub_f16(d12, d32), 2));
+  float16x4_t t43 = vadd_f16(vsub_f16(d43, d23), vmul_n_f16(vsub_f16(d13, d33), 2));
+  float16x4_t t44 = vadd_f16(vsub_f16(d44, d24), vmul_n_f16(vsub_f16(d14, d34), 2));
+  float16x4_t t45 = vadd_f16(vsub_f16(d45, d25), vmul_n_f16(vsub_f16(d15, d35), 2));
+
+  float16x4_t t50 = vadd_f16(vsub_f16(vmul_n_f16(d10, 4), vmul_n_f16(d30, 5)), d50);
+  float16x4_t t51 = vadd_f16(vsub_f16(vmul_n_f16(d11, 4), vmul_n_f16(d31, 5)), d51);
+  float16x4_t t52 = vadd_f16(vsub_f16(vmul_n_f16(d12, 4), vmul_n_f16(d32, 5)), d52);
+  float16x4_t t53 = vadd_f16(vsub_f16(vmul_n_f16(d13, 4), vmul_n_f16(d33, 5)), d53);
+  float16x4_t t54 = vadd_f16(vsub_f16(vmul_n_f16(d14, 4), vmul_n_f16(d34, 5)), d54);
+  float16x4_t t55 = vadd_f16(vsub_f16(vmul_n_f16(d15, 4), vmul_n_f16(d35, 5)), d55);
+
+  float16x4_t m00 = vadd_f16(vsub_f16(vmul_n_f16(t00, 4), vmul_n_f16(t02, 5)), t04);
+  float16x4_t m01 = vadd_f16(vadd_f16(t03, t04), vmul_n_f16(vadd_f16(t01, t02), -4));
+  float16x4_t m02 = vadd_f16(vsub_f16(t04, t03), vmul_n_f16(vsub_f16(t01, t02), 4));
+  float16x4_t m03 = vadd_f16(vsub_f16(t04, t02), vmul_n_f16(vsub_f16(t03, t01), 2));
+  float16x4_t m04 = vadd_f16(vsub_f16(t04, t02), vmul_n_f16(vsub_f16(t01, t03), 2));
+  float16x4_t m05 = vadd_f16(vsub_f16(vmul_n_f16(t01, 4), vmul_n_f16(t03, 5)), t05);
+
+  float16x4_t m10 = vadd_f16(vsub_f16(vmul_n_f16(t10, 4), vmul_n_f16(t12, 5)), t14);
+  float16x4_t m11 = vadd_f16(vadd_f16(t13, t14), vmul_n_f16(vadd_f16(t11, t12), -4));
+  float16x4_t m12 = vadd_f16(vsub_f16(t14, t13), vmul_n_f16(vsub_f16(t11, t12), 4));
+  float16x4_t m13 = vadd_f16(vsub_f16(t14, t12), vmul_n_f16(vsub_f16(t13, t11), 2));
+  float16x4_t m14 = vadd_f16(vsub_f16(t14, t12), vmul_n_f16(vsub_f16(t11, t13), 2));
+  float16x4_t m15 = vadd_f16(vsub_f16(vmul_n_f16(t11, 4), vmul_n_f16(t13, 5)), t15);
+
+  float16x4_t m20 = vadd_f16(vsub_f16(vmul_n_f16(t20, 4), vmul_n_f16(t22, 5)), t24);
+  float16x4_t m21 = vadd_f16(vadd_f16(t23, t24), vmul_n_f16(vadd_f16(t21, t22), -4));
+  float16x4_t m22 = vadd_f16(vsub_f16(t24, t23), vmul_n_f16(vsub_f16(t21, t22), 4));
+  float16x4_t m23 = vadd_f16(vsub_f16(t24, t22), vmul_n_f16(vsub_f16(t23, t21), 2));
+  float16x4_t m24 = vadd_f16(vsub_f16(t24, t22), vmul_n_f16(vsub_f16(t21, t23), 2));
+  float16x4_t m25 = vadd_f16(vsub_f16(vmul_n_f16(t21, 4), vmul_n_f16(t23, 5)), t25);
+
+  float16x4_t m30 = vadd_f16(vsub_f16(vmul_n_f16(t30, 4), vmul_n_f16(t32, 5)), t34);
+  float16x4_t m31 = vadd_f16(vadd_f16(t33, t34), vmul_n_f16(vadd_f16(t31, t32), -4));
+  float16x4_t m32 = vadd_f16(vsub_f16(t34, t33), vmul_n_f16(vsub_f16(t31, t32), 4));
+  float16x4_t m33 = vadd_f16(vsub_f16(t34, t32), vmul_n_f16(vsub_f16(t33, t31), 2));
+  float16x4_t m34 = vadd_f16(vsub_f16(t34, t32), vmul_n_f16(vsub_f16(t31, t33), 2));
+  float16x4_t m35 = vadd_f16(vsub_f16(vmul_n_f16(t31, 4), vmul_n_f16(t33, 5)), t35);
+
+  float16x4_t m40 = vadd_f16(vsub_f16(vmul_n_f16(t40, 4), vmul_n_f16(t42, 5)), t44);
+  float16x4_t m41 = vadd_f16(vadd_f16(t43, t44), vmul_n_f16(vadd_f16(t41, t42), -4));
+  float16x4_t m42 = vadd_f16(vsub_f16(t44, t43), vmul_n_f16(vsub_f16(t41, t42), 4));
+  float16x4_t m43 = vadd_f16(vsub_f16(t44, t42), vmul_n_f16(vsub_f16(t43, t41), 2));
+  float16x4_t m44 = vadd_f16(vsub_f16(t44, t42), vmul_n_f16(vsub_f16(t41, t43), 2));
+  float16x4_t m45 = vadd_f16(vsub_f16(vmul_n_f16(t41, 4), vmul_n_f16(t43, 5)), t45);
+
+  float16x4_t m50 = vadd_f16(vsub_f16(vmul_n_f16(t50, 4), vmul_n_f16(t52, 5)), t54);
+  float16x4_t m51 = vadd_f16(vadd_f16(t53, t54), vmul_n_f16(vadd_f16(t51, t52), -4));
+  float16x4_t m52 = vadd_f16(vsub_f16(t54, t53), vmul_n_f16(vsub_f16(t51, t52), 4));
+  float16x4_t m53 = vadd_f16(vsub_f16(t54, t52), vmul_n_f16(vsub_f16(t53, t51), 2));
+  float16x4_t m54 = vadd_f16(vsub_f16(t54, t52), vmul_n_f16(vsub_f16(t51, t53), 2));
+  float16x4_t m55 = vadd_f16(vsub_f16(vmul_n_f16(t51, 4), vmul_n_f16(t53, 5)), t55);
+
+  vst1_f16(trans_input_data, m00);
+  vst1_f16(trans_input_data + step, m01);
+  vst1_f16(trans_input_data + 2 * step, m02);
+  vst1_f16(trans_input_data + 3 * step, m03);
+  vst1_f16(trans_input_data + 4 * step, m04);
+  vst1_f16(trans_input_data + 5 * step, m05);
+
+  vst1_f16(trans_input_data + 6 * step, m10);
+  vst1_f16(trans_input_data + 7 * step, m11);
+  vst1_f16(trans_input_data + 8 * step, m12);
+  vst1_f16(trans_input_data + 9 * step, m13);
+  vst1_f16(trans_input_data + 10 * step, m14);
+  vst1_f16(trans_input_data + 11 * step, m15);
+
+  vst1_f16(trans_input_data + 12 * step, m20);
+  vst1_f16(trans_input_data + 13 * step, m21);
+  vst1_f16(trans_input_data + 14 * step, m22);
+  vst1_f16(trans_input_data + 15 * step, m23);
+  vst1_f16(trans_input_data + 16 * step, m24);
+  vst1_f16(trans_input_data + 17 * step, m25);
+
+  vst1_f16(trans_input_data + 18 * step, m30);
+  vst1_f16(trans_input_data + 19 * step, m31);
+  vst1_f16(trans_input_data + 20 * step, m32);
+  vst1_f16(trans_input_data + 21 * step, m33);
+  vst1_f16(trans_input_data + 22 * step, m34);
+  vst1_f16(trans_input_data + 23 * step, m35);
+
+  vst1_f16(trans_input_data + 24 * step, m40);
+  vst1_f16(trans_input_data + 25 * step, m41);
+  vst1_f16(trans_input_data + 26 * step, m42);
+  vst1_f16(trans_input_data + 27 * step, m43);
+  vst1_f16(trans_input_data + 28 * step, m44);
+  vst1_f16(trans_input_data + 29 * step, m45);
+
+  vst1_f16(trans_input_data + 30 * step, m50);
+  vst1_f16(trans_input_data + 31 * step, m51);
+  vst1_f16(trans_input_data + 32 * step, m52);
+  vst1_f16(trans_input_data + 33 * step, m53);
+  vst1_f16(trans_input_data + 34 * step, m54);
+  vst1_f16(trans_input_data + 35 * step, m55);
+}
+
+void Conv3x3Fp16InputTransform(const float16_t *input_data, float16_t *trans_input, float16_t *tmp_data,
+                               int start_index, int real_cal_num, int out_w_block, ConvParameter *conv_param) {
+  // input data format : nhwc
+  int output_unit = 4;
+  int input_channel = conv_param->input_channel_;
+  int input_width = conv_param->input_w_;
+  int input_height = conv_param->input_h_;
+  int pad_w = conv_param->pad_w_;
+  int pad_h = conv_param->pad_h_;
+  int ic4 = UP_DIV(input_channel, C4NUM);
+
+  for (int cal_id = 0; cal_id < real_cal_num; cal_id++) {
+    int x_id = start_index + cal_id;
+    int origin_x = (x_id % out_w_block) * output_unit - pad_w;
+    int origin_y = (x_id / out_w_block) * output_unit - pad_h;
+    int real_x_start = origin_x > 0 ? 0 : -origin_x;
+    int real_x_end = (origin_x + 6) < input_width ? 6 : (input_width - origin_x);
+    int real_y_start = origin_y > 0 ? 0 : -origin_y;
+    int real_y_end = (origin_y + 6) < input_height ? 6 : (input_height - origin_y);
+
+    int src_plane_offset = input_channel * (origin_y * input_width + origin_x);
+    int dst_plane_offset = cal_id * C4NUM;
+    for (int ic = 0; ic < ic4; ic++) {
+      // clear tmp buffer
+      memset(tmp_data, 0, 6 * 6 * C4NUM * sizeof(float16_t));
+
+      // get real input block with padding
+      int src_ic4_offset = src_plane_offset + ic * C4NUM;
+      for (int interval = real_y_start; interval < real_y_end; interval++) {
+        int src_y_offset = src_ic4_offset + interval * input_width * input_channel + real_x_start * input_channel;
+        int dst_y_offset = interval * 6 * C4NUM + real_x_start * C4NUM;
+        for (int j = 0; j < (real_x_end - real_x_start); j++) {
+          int src_x_offset = src_y_offset + j * input_channel;
+          int dst_x_offset = dst_y_offset + j * C4NUM;
+          float16_t *src_addr = (float16_t *)(input_data) + src_x_offset;
+          float16_t *dst_addr = tmp_data + dst_x_offset;
+          dst_addr[0] = src_addr[0];
+          dst_addr[1] = src_addr[1];
+          dst_addr[2] = src_addr[2];
+          dst_addr[3] = src_addr[3];
+        }
+      }
+
+      // todo
+      // input transform
+      int dst_ic4_offset = dst_plane_offset + ic * 16 * C4NUM;
+      size_t dst_step = ic4 * C4NUM * 16;
+      float16_t *trans_input_ptr = trans_input + dst_ic4_offset;
+      Conv3x3Fp16InputUnit(tmp_data, trans_input_ptr, dst_step);
+    }
+  }
+}
+
+void Conv3x3Fp16FilterTransform(const float16_t *weight_data, float16_t *trans_weight, int iC4, int output_channel,
+                                int kernel_plane) {
+  int dst_step = iC4 * C4NUM * 8;
+  for (int o = 0; o < output_channel; o++) {
+    int oc8_block_num = o / C8NUM;
+    int oc8_block_rem = o % C8NUM;
+    int src_oc_offset = o * iC4 * C4NUM * kernel_plane;
+    int dst_oc_offset = oc8_block_num * C8NUM * iC4 * C4NUM * 36 + oc8_block_rem;
+    for (int i = 0; i < iC4; i++) {
+      const float16_t *src_ic4_ptr = weight_data + src_oc_offset + i * kernel_plane * C4NUM;
+      float16_t *dst_ic4_ptr = trans_weight + dst_oc_offset + i * 8 * C4NUM;
+      float16x4_t g00 = vld1_f16(src_ic4_ptr);
+      float16x4_t g01 = vld1_f16(src_ic4_ptr + 4);
+      float16x4_t g02 = vld1_f16(src_ic4_ptr + 2 * 4);
+      float16x4_t g10 = vld1_f16(src_ic4_ptr + 3 * 4);
+      float16x4_t g11 = vld1_f16(src_ic4_ptr + 4 * 4);
+      float16x4_t g12 = vld1_f16(src_ic4_ptr + 5 * 4);
+      float16x4_t g20 = vld1_f16(src_ic4_ptr + 6 * 4);
+      float16x4_t g21 = vld1_f16(src_ic4_ptr + 7 * 4);
+      float16x4_t g22 = vld1_f16(src_ic4_ptr + 8 * 4);
+
+      float16x4_t dst00 = vmul_n_f16(g00, 0.25);
+      float16x4_t dst01 = vmul_n_f16(g01, 0.25);
+      float16x4_t dst02 = vmul_n_f16(g02, 0.25);
+
+      float16x4_t dst10 = vmul_n_f16(vadd_f16(g00, vadd_f16(g10, g20)), -0.1666666666667);
+      float16x4_t dst11 = vmul_n_f16(vadd_f16(g01, vadd_f16(g11, g21)), -0.1666666666667);
+      float16x4_t dst12 = vmul_n_f16(vadd_f16(g02, vadd_f16(g12, g22)), -0.1666666666667);
+
+      float16x4_t dst20 = vmul_n_f16(vsub_f16(vadd_f16(g00, g20), g10), -0.1666666666667);
+      float16x4_t dst21 = vmul_n_f16(vsub_f16(vadd_f16(g01, g21), g11), -0.1666666666667);
+      float16x4_t dst22 = vmul_n_f16(vsub_f16(vadd_f16(g02, g22), g12), -0.1666666666667);
+
+      float16x4_t dst30 = vadd_f16(vmul_n_f16(g10, 0.08333333333333),
+                                   vadd_f16(vmul_n_f16(g00, 0.04166666666667), vmul_n_f16(g20, 0.1666666666667)));
+      float16x4_t dst31 = vadd_f16(vmul_n_f16(g11, 0.08333333333333),
+                                   vadd_f16(vmul_n_f16(g01, 0.04166666666667), vmul_n_f16(g21, 0.1666666666667)));
+      float16x4_t dst32 = vadd_f16(vmul_n_f16(g12, 0.08333333333333),
+                                   vadd_f16(vmul_n_f16(g02, 0.04166666666667), vmul_n_f16(g22, 0.1666666666667)));
+
+      float16x4_t dst40 = vsub_f16(vadd_f16(vmul_n_f16(g00, 0.04166666666667), vmul_n_f16(g20, 0.1666666666667)),
+                                   vmul_n_f16(g10, 0.08333333333333));
+      float16x4_t dst41 = vsub_f16(vadd_f16(vmul_n_f16(g01, 0.04166666666667), vmul_n_f16(g21, 0.1666666666667)),
+                                   vmul_n_f16(g11, 0.08333333333333));
+      float16x4_t dst42 = vsub_f16(vadd_f16(vmul_n_f16(g02, 0.04166666666667), vmul_n_f16(g22, 0.1666666666667)),
+                                   vmul_n_f16(g12, 0.08333333333333));
+
+      float16x4_t dst50 = g20;
+      float16x4_t dst51 = g21;
+      float16x4_t dst52 = g22;
+
+      float16x4_t m00 = vmul_n_f16(dst00, 0.25);
+      float16x4_t m01 = vmul_n_f16(vadd_f16(dst00, vadd_f16(dst01, dst02)), -0.1666666666667);
+      float16x4_t m02 = vmul_n_f16(vsub_f16(vadd_f16(dst00, dst02), dst01), -0.1666666666667);
+      float16x4_t m03 = vadd_f16(vmul_n_f16(dst01, 0.08333333333333),
+                                 vadd_f16(vmul_n_f16(dst00, 0.04166666666667), vmul_n_f16(dst02, 0.1666666666667)));
+      float16x4_t m04 = vsub_f16(vadd_f16(vmul_n_f16(dst00, 0.04166666666667), vmul_n_f16(dst02, 0.1666666666667)),
+                                 vmul_n_f16(dst01, 0.08333333333333));
+      float16x4_t m05 = dst02;
+
+      float16x4_t m10 = vmul_n_f16(dst10, 0.25);
+      float16x4_t m11 = vmul_n_f16(vadd_f16(dst10, vadd_f16(dst11, dst12)), -0.1666666666667);
+      float16x4_t m12 = vmul_n_f16(vsub_f16(vadd_f16(dst10, dst12), dst11), -0.1666666666667);
+      float16x4_t m13 = vadd_f16(vmul_n_f16(dst11, 0.08333333333333),
+                                 vadd_f16(vmul_n_f16(dst10, 0.04166666666667), vmul_n_f16(dst12, 0.1666666666667)));
+      float16x4_t m14 = vsub_f16(vadd_f16(vmul_n_f16(dst10, 0.04166666666667), vmul_n_f16(dst12, 0.1666666666667)),
+                                 vmul_n_f16(dst11, 0.08333333333333));
+      float16x4_t m15 = dst12;
+
+      float16x4_t m20 = vmul_n_f16(dst20, 0.25);
+      float16x4_t m21 = vmul_n_f16(vadd_f16(dst20, vadd_f16(dst21, dst22)), -0.1666666666667);
+      float16x4_t m22 = vmul_n_f16(vsub_f16(vadd_f16(dst20, dst22), dst21), -0.1666666666667);
+      float16x4_t m23 = vadd_f16(vmul_n_f16(dst21, 0.08333333333333),
+                                 vadd_f16(vmul_n_f16(dst20, 0.04166666666667), vmul_n_f16(dst22, 0.1666666666667)));
+      float16x4_t m24 = vsub_f16(vadd_f16(vmul_n_f16(dst20, 0.04166666666667), vmul_n_f16(dst22, 0.1666666666667)),
+                                 vmul_n_f16(dst21, 0.08333333333333));
+      float16x4_t m25 = dst22;
+
+      float16x4_t m30 = vmul_n_f16(dst30, 0.25);
+      float16x4_t m31 = vmul_n_f16(vadd_f16(dst30, vadd_f16(dst31, dst32)), -0.1666666666667);
+      float16x4_t m32 = vmul_n_f16(vsub_f16(vadd_f16(dst30, dst32), dst31), -0.1666666666667);
+      float16x4_t m33 = vadd_f16(vmul_n_f16(dst31, 0.08333333333333),
+                                 vadd_f16(vmul_n_f16(dst30, 0.04166666666667), vmul_n_f16(dst32, 0.1666666666667)));
+      float16x4_t m34 = vsub_f16(vadd_f16(vmul_n_f16(dst30, 0.04166666666667), vmul_n_f16(dst32, 0.1666666666667)),
+                                 vmul_n_f16(dst31, 0.08333333333333));
+      float16x4_t m35 = dst32;
+
+      float16x4_t m40 = vmul_n_f16(dst40, 0.25);
+      float16x4_t m41 = vmul_n_f16(vadd_f16(dst40, vadd_f16(dst41, dst42)), -0.1666666666667);
+      float16x4_t m42 = vmul_n_f16(vsub_f16(vadd_f16(dst40, dst42), dst41), -0.1666666666667);
+      float16x4_t m43 = vadd_f16(vmul_n_f16(dst41, 0.08333333333333),
+                                 vadd_f16(vmul_n_f16(dst40, 0.04166666666667), vmul_n_f16(dst42, 0.1666666666667)));
+      float16x4_t m44 = vsub_f16(vadd_f16(vmul_n_f16(dst40, 0.04166666666667), vmul_n_f16(dst42, 0.1666666666667)),
+                                 vmul_n_f16(dst41, 0.08333333333333));
+      float16x4_t m45 = dst42;
+
+      float16x4_t m50 = vmul_n_f16(dst50, 0.25);
+      float16x4_t m51 = vmul_n_f16(vadd_f16(dst50, vadd_f16(dst51, dst52)), -0.1666666666667);
+      float16x4_t m52 = vmul_n_f16(vsub_f16(vadd_f16(dst50, dst52), dst51), -0.1666666666667);
+      float16x4_t m53 = vadd_f16(vmul_n_f16(dst51, 0.08333333333333),
+                                 vadd_f16(vmul_n_f16(dst50, 0.04166666666667), vmul_n_f16(dst52, 0.1666666666667)));
+      float16x4_t m54 = vsub_f16(vadd_f16(vmul_n_f16(dst50, 0.04166666666667), vmul_n_f16(dst52, 0.1666666666667)),
+                                 vmul_n_f16(dst51, 0.08333333333333));
+      float16x4_t m55 = dst52;
+
+      for (int j = 0; j < 4; j++) {
+        dst_ic4_ptr[j * 8] = m00[j];
+        dst_ic4_ptr[j * 8 + dst_step] = m01[j];
+        dst_ic4_ptr[j * 8 + 2 * dst_step] = m02[j];
+        dst_ic4_ptr[j * 8 + 3 * dst_step] = m03[j];
+        dst_ic4_ptr[j * 8 + 4 * dst_step] = m04[j];
+        dst_ic4_ptr[j * 8 + 5 * dst_step] = m05[j];
+        dst_ic4_ptr[j * 8 + 6 * dst_step] = m10[j];
+        dst_ic4_ptr[j * 8 + 7 * dst_step] = m11[j];
+        dst_ic4_ptr[j * 8 + 8 * dst_step] = m12[j];
+        dst_ic4_ptr[j * 8 + 9 * dst_step] = m13[j];
+        dst_ic4_ptr[j * 8 + 10 * dst_step] = m14[j];
+        dst_ic4_ptr[j * 8 + 11 * dst_step] = m15[j];
+        dst_ic4_ptr[j * 8 + 12 * dst_step] = m20[j];
+        dst_ic4_ptr[j * 8 + 13 * dst_step] = m21[j];
+        dst_ic4_ptr[j * 8 + 14 * dst_step] = m22[j];
+        dst_ic4_ptr[j * 8 + 15 * dst_step] = m23[j];
+        dst_ic4_ptr[j * 8 + 16 * dst_step] = m24[j];
+        dst_ic4_ptr[j * 8 + 17 * dst_step] = m25[j];
+        dst_ic4_ptr[j * 8 + 18 * dst_step] = m30[j];
+        dst_ic4_ptr[j * 8 + 19 * dst_step] = m31[j];
+        dst_ic4_ptr[j * 8 + 20 * dst_step] = m32[j];
+        dst_ic4_ptr[j * 8 + 21 * dst_step] = m33[j];
+        dst_ic4_ptr[j * 8 + 22 * dst_step] = m34[j];
+        dst_ic4_ptr[j * 8 + 23 * dst_step] = m35[j];
+        dst_ic4_ptr[j * 8 + 24 * dst_step] = m40[j];
+        dst_ic4_ptr[j * 8 + 25 * dst_step] = m41[j];
+        dst_ic4_ptr[j * 8 + 26 * dst_step] = m42[j];
+        dst_ic4_ptr[j * 8 + 27 * dst_step] = m43[j];
+        dst_ic4_ptr[j * 8 + 28 * dst_step] = m44[j];
+        dst_ic4_ptr[j * 8 + 29 * dst_step] = m45[j];
+        dst_ic4_ptr[j * 8 + 30 * dst_step] = m50[j];
+        dst_ic4_ptr[j * 8 + 31 * dst_step] = m51[j];
+        dst_ic4_ptr[j * 8 + 32 * dst_step] = m52[j];
+        dst_ic4_ptr[j * 8 + 33 * dst_step] = m53[j];
+        dst_ic4_ptr[j * 8 + 34 * dst_step] = m54[j];
+        dst_ic4_ptr[j * 8 + 35 * dst_step] = m55[j];
+      }
+    }
+  }
+}
+
+void Conv3x3Fp16OutputUnit(const float16_t *gemm_out, const float16_t *bias_data, float16_t *output_data,
+                           int output_w) {
+  float16x8_t s00 = vld1q_f16(gemm_out);
+  float16x8_t s01 = vld1q_f16(gemm_out + 8);
+  float16x8_t s02 = vld1q_f16(gemm_out + 16);
+  float16x8_t s03 = vld1q_f16(gemm_out + 24);
+  float16x8_t s04 = vld1q_f16(gemm_out + 32);
+  float16x8_t s05 = vld1q_f16(gemm_out + 40);
+
+  float16x8_t s10 = vld1q_f16(gemm_out + 48);
+  float16x8_t s11 = vld1q_f16(gemm_out + 56);
+  float16x8_t s12 = vld1q_f16(gemm_out + 64);
+  float16x8_t s13 = vld1q_f16(gemm_out + 72);
+  float16x8_t s14 = vld1q_f16(gemm_out + 80);
+  float16x8_t s15 = vld1q_f16(gemm_out + 88);
+
+  float16x8_t s20 = vld1q_f16(gemm_out + 96);
+  float16x8_t s21 = vld1q_f16(gemm_out + 104);
+  float16x8_t s22 = vld1q_f16(gemm_out + 112);
+  float16x8_t s23 = vld1q_f16(gemm_out + 120);
+  float16x8_t s24 = vld1q_f16(gemm_out + 128);
+  float16x8_t s25 = vld1q_f16(gemm_out + 136);
+
+  float16x8_t s30 = vld1q_f16(gemm_out + 144);
+  float16x8_t s31 = vld1q_f16(gemm_out + 152);
+  float16x8_t s32 = vld1q_f16(gemm_out + 160);
+  float16x8_t s33 = vld1q_f16(gemm_out + 168);
+  float16x8_t s34 = vld1q_f16(gemm_out + 176);
+  float16x8_t s35 = vld1q_f16(gemm_out + 184);
+
+  float16x8_t s40 = vld1q_f16(gemm_out + 192);
+  float16x8_t s41 = vld1q_f16(gemm_out + 200);
+  float16x8_t s42 = vld1q_f16(gemm_out + 208);
+  float16x8_t s43 = vld1q_f16(gemm_out + 216);
+  float16x8_t s44 = vld1q_f16(gemm_out + 224);
+  float16x8_t s45 = vld1q_f16(gemm_out + 232);
+
+  float16x8_t s50 = vld1q_f16(gemm_out + 240);
+  float16x8_t s51 = vld1q_f16(gemm_out + 248);
+  float16x8_t s52 = vld1q_f16(gemm_out + 256);
+  float16x8_t s53 = vld1q_f16(gemm_out + 264);
+  float16x8_t s54 = vld1q_f16(gemm_out + 272);
+  float16x8_t s55 = vld1q_f16(gemm_out + 280);
+
+  float16x8_t t00 = vaddq_f16(vaddq_f16(vaddq_f16(s00, s10), vaddq_f16(s20, s30)), s40);
+  float16x8_t t01 = vaddq_f16(vaddq_f16(vaddq_f16(s01, s11), vaddq_f16(s21, s31)), s41);
+  float16x8_t t02 = vaddq_f16(vaddq_f16(vaddq_f16(s02, s12), vaddq_f16(s22, s32)), s42);
+  float16x8_t t03 = vaddq_f16(vaddq_f16(vaddq_f16(s03, s13), vaddq_f16(s23, s33)), s43);
+  float16x8_t t04 = vaddq_f16(vaddq_f16(vaddq_f16(s04, s14), vaddq_f16(s24, s34)), s44);
+  float16x8_t t05 = vaddq_f16(vaddq_f16(vaddq_f16(s05, s15), vaddq_f16(s25, s35)), s45);
+
+  float16x8_t t10 = vaddq_f16(vsubq_f16(s10, s20), vmulq_n_f16(vsubq_f16(s30, s40), 2));
+  float16x8_t t11 = vaddq_f16(vsubq_f16(s11, s21), vmulq_n_f16(vsubq_f16(s31, s41), 2));
+  float16x8_t t12 = vaddq_f16(vsubq_f16(s12, s22), vmulq_n_f16(vsubq_f16(s32, s42), 2));
+  float16x8_t t13 = vaddq_f16(vsubq_f16(s13, s23), vmulq_n_f16(vsubq_f16(s33, s43), 2));
+  float16x8_t t14 = vaddq_f16(vsubq_f16(s14, s24), vmulq_n_f16(vsubq_f16(s34, s44), 2));
+  float16x8_t t15 = vaddq_f16(vsubq_f16(s15, s25), vmulq_n_f16(vsubq_f16(s35, s45), 2));
+
+  float16x8_t t20 = vaddq_f16(vaddq_f16(s10, s20), vmulq_n_f16(vaddq_f16(s30, s40), 4));
+  float16x8_t t21 = vaddq_f16(vaddq_f16(s11, s21), vmulq_n_f16(vaddq_f16(s31, s41), 4));
+  float16x8_t t22 = vaddq_f16(vaddq_f16(s12, s22), vmulq_n_f16(vaddq_f16(s32, s42), 4));
+  float16x8_t t23 = vaddq_f16(vaddq_f16(s13, s23), vmulq_n_f16(vaddq_f16(s33, s43), 4));
+  float16x8_t t24 = vaddq_f16(vaddq_f16(s14, s24), vmulq_n_f16(vaddq_f16(s34, s44), 4));
+  float16x8_t t25 = vaddq_f16(vaddq_f16(s15, s25), vmulq_n_f16(vaddq_f16(s35, s45), 4));
+
+  float16x8_t t30 = vaddq_f16(vaddq_f16(vsubq_f16(s10, s20), vmulq_n_f16(vsubq_f16(s30, s40), 8)), s50);
+  float16x8_t t31 = vaddq_f16(vaddq_f16(vsubq_f16(s11, s21), vmulq_n_f16(vsubq_f16(s31, s41), 8)), s51);
+  float16x8_t t32 = vaddq_f16(vaddq_f16(vsubq_f16(s12, s22), vmulq_n_f16(vsubq_f16(s32, s42), 8)), s52);
+  float16x8_t t33 = vaddq_f16(vaddq_f16(vsubq_f16(s13, s23), vmulq_n_f16(vsubq_f16(s33, s43), 8)), s53);
+  float16x8_t t34 = vaddq_f16(vaddq_f16(vsubq_f16(s14, s24), vmulq_n_f16(vsubq_f16(s34, s44), 8)), s54);
+  float16x8_t t35 = vaddq_f16(vaddq_f16(vsubq_f16(s15, s25), vmulq_n_f16(vsubq_f16(s35, s45), 8)), s55);
+
+  float16x8_t d00 = vaddq_f16(vaddq_f16(vaddq_f16(t00, t01), vaddq_f16(t02, t03)), t04);
+  float16x8_t d01 = vaddq_f16(vsubq_f16(t01, t02), vmulq_n_f16(vsubq_f16(t03, t04), 2));
+  float16x8_t d02 = vaddq_f16(vaddq_f16(t01, t02), vmulq_n_f16(vaddq_f16(t03, t04), 4));
+  float16x8_t d03 = vaddq_f16(vaddq_f16(vsubq_f16(t01, t02), vmulq_n_f16(vsubq_f16(t03, t04), 8)), t05);
+
+  float16x8_t d10 = vaddq_f16(vaddq_f16(vaddq_f16(t10, t11), vaddq_f16(t12, t13)), t14);
+  float16x8_t d11 = vaddq_f16(vsubq_f16(t11, t12), vmulq_n_f16(vsubq_f16(t13, t14), 2));
+  float16x8_t d12 = vaddq_f16(vaddq_f16(t11, t12), vmulq_n_f16(vaddq_f16(t13, t14), 4));
+  float16x8_t d13 = vaddq_f16(vaddq_f16(vsubq_f16(t11, t12), vmulq_n_f16(vsubq_f16(t13, t14), 8)), t15);
+
+  float16x8_t d20 = vaddq_f16(vaddq_f16(vaddq_f16(t20, t21), vaddq_f16(t22, t23)), t24);
+  float16x8_t d21 = vaddq_f16(vsubq_f16(t21, t22), vmulq_n_f16(vsubq_f16(t23, t24), 2));
+  float16x8_t d22 = vaddq_f16(vaddq_f16(t21, t22), vmulq_n_f16(vaddq_f16(t23, t24), 4));
+  float16x8_t d23 = vaddq_f16(vaddq_f16(vsubq_f16(t21, t22), vmulq_n_f16(vsubq_f16(t23, t24), 8)), t25);
+
+  float16x8_t d30 = vaddq_f16(vaddq_f16(vaddq_f16(t30, t31), vaddq_f16(t32, t33)), t34);
+  float16x8_t d31 = vaddq_f16(vsubq_f16(t31, t32), vmulq_n_f16(vsubq_f16(t33, t34), 2));
+  float16x8_t d32 = vaddq_f16(vaddq_f16(t31, t32), vmulq_n_f16(vaddq_f16(t33, t34), 4));
+  float16x8_t d33 = vaddq_f16(vaddq_f16(vsubq_f16(t31, t32), vmulq_n_f16(vsubq_f16(t33, t34), 8)), t35);
+
+  vst1q_f16(output_data, d00);
+  vst1q_f16(output_data + 8, d01);
+  vst1q_f16(output_data + 16, d02);
+  vst1q_f16(output_data + 24, d03);
+
+  vst1q_f16(output_data + output_w * 8, d10);
+  vst1q_f16(output_data + output_w * 8 + 8, d11);
+  vst1q_f16(output_data + output_w * 8 + 16, d12);
+  vst1q_f16(output_data + output_w * 8 + 24, d13);
+
+  vst1q_f16(output_data + 2 * output_w * 8, d20);
+  vst1q_f16(output_data + 2 * output_w * 8 + 8, d21);
+  vst1q_f16(output_data + 2 * output_w * 8 + 16, d22);
+  vst1q_f16(output_data + 2 * output_w * 8 + 24, d23);
+
+  vst1q_f16(output_data + 3 * output_w * 8, d30);
+  vst1q_f16(output_data + 3 * output_w * 8 + 8, d31);
+  vst1q_f16(output_data + 3 * output_w * 8 + 16, d32);
+  vst1q_f16(output_data + 3 * output_w * 8 + 24, d33);
+}
+
+void Conv3x3Fp16OutputTransform(const float16_t *gemm_out, float16_t *out_data, const float16_t *bias_data,
+                                int start_index, int real_cal_num, int out_w_block, ConvParameter *conv_param) {
+  int output_channel = conv_param->output_channel_;
+  int output_w = conv_param->output_w_;
+  int output_h = conv_param->output_h_;
+  int oc8 = UP_DIV(output_channel, C8NUM);
+
+  for (int i = 0; i < real_cal_num; i++) {
+    int out_w_index = (start_index + i) % out_w_block;
+    int out_h_index = (start_index + i) / out_w_block;
+    int src_tile_offset = i * oc8 * C8NUM * 36;
+    int dst_tile_offset = 8 * (out_w_index * 4 + out_h_index * 4 * output_w);
+
+    for (int j = 0; j < oc8; j++) {
+      int src_oc8_offset = src_tile_offset + j * 36 * C8NUM;
+      int dst_oc8_offset = dst_tile_offset + j * C8NUM * output_h * output_w;
+      const float16_t *src_ptr = gemm_out + src_oc8_offset;
+      const float16_t *bias_ptr = bias_data + j * C8NUM;
+      float16_t *dst_ptr = out_data + dst_oc8_offset;
+
+      // output transform
+      Conv3x3Fp16OutputUnit(src_ptr, bias_ptr, dst_ptr, output_w);
+    }
+  }
+}

mindspore/lite/src/runtime/kernel/arm/opclib/fp16/winograd_transform_fp16.h

@@ -0,0 +1,39 @@
+/**
+ * Copyright 2020 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_KERNEL_ARM_OPCLIB_FP16_WINOGRAD_TRANSFORM_FP16_H_
+#define MINDSPORE_LITE_SRC_RUNTIME_KERNEL_ARM_OPCLIB_FP16_WINOGRAD_TRANSFORM_FP16_H_
+
+#include <arm_neon.h>
+#include <string.h>
+#include "src/runtime/kernel/arm/opclib/fp16/pack_fp16.h"
+#include "src/runtime/kernel/arm/opclib/fp16/conv_fp16.h"
+
+// for fp16 convolution 3x3 filter/input/output transform
+void Conv3x3Fp16InputUnit(float16_t *tmp_data, float16_t *trans_input_data, size_t step);
+
+void Conv3x3Fp16InputTransform(const float16_t *input_data, float16_t *trans_input, float16_t *tmp_data,
+                               int start_index, int real_cal_num, int out_w_block, ConvParameter *conv_param);
+
+void Conv3x3Fp16FilterTransform(const float16_t *weight_data, float16_t *trans_weight, int iC8, int output_channel,
+                                int kernel_plane);
+
+void Conv3x3Fp16OutputUnit(const float16_t *gemm_out, const float16_t *bias_data, float16_t *output_data, int output_w);
+
+void Conv3x3Fp16OutputTransform(const float16_t *gemm_out, float16_t *out_data, const float16_t *bias_data,
+                                int start_index, int real_cal_num, int out_w_block, ConvParameter *conv_param);
+
+#endif  // MINDSPORE_LITE_SRC_RUNTIME_KERNEL_ARM_OPCLIB_FP16_WINOGRAD_TRANSFORM_FP16_H_

mindspore/lite/src/runtime/kernel/arm/opclib/optimized_kernel.h

@@ -29,11 +29,24 @@ class OptimizeModule {
  public:
   OptimizeModule() {
     bool support_optimize_ops = false;
-
+    bool support_fp16 = false;
 #ifdef __ANDROID__
     int hwcap_type = 16;
     uint32_t hwcap = getHwCap(hwcap_type);
-#if defined(__aarch64__)
+#ifdef ENABLE_ARM64
+    if (hwcap & HWCAP_FPHP) {
+#elif defined(__arm__)
+    if (hwcap & HWCAP_HALF) {
+#endif
+      MS_LOG(INFO) << "Hw cap support FP16, hwcap: 0x" << hwcap;
+      support_fp16 = true;
+#ifdef ENABLE_ARM64
+    }
+#elif defined(__arm__)
+    }
+#endif
+
+#ifdef ENABLE_ARM64
     if (hwcap & HWCAP_ASIMDDP) {
       printf("Hw cap support SMID Dot Product, hwcap: 0x%x \n", hwcap);
       support_optimize_ops = true;
@@ -42,7 +55,7 @@ class OptimizeModule {
     }
 #endif
 #endif
-    if (!support_optimize_ops) {
+    if ((!support_optimize_ops) && (!support_fp16)) {
       return;
     }
     optimized_op_handler_ = dlopen(OPTIMIZE_SHARED_LIBRARY_PATH, RTLD_LAZY);
@@ -61,4 +74,3 @@ class OptimizeModule {
 };
 
 #endif  // MINDSPORE_LITE_SRC_RUNTIME_KERNEL_ARM_OPCLIB_OPTIMIZED_KERNEL_H_
-

mindspore/lite/src/runtime/kernel/arm/opclib/pack.cc

@@ -18,331 +18,6 @@
 #include <cstring>
 #include <cstdlib>
 
-#ifdef ENABLE_FP16
-void Im2ColPackUnitFp16(float16_t *input_data, ConvParameter *conv_param, float16_t *packed_input, int real_cal_num,
-                        int block_index) {
-  // input format : nhwc
-  int kernel_h = conv_param->kernel_h_;
-  int kernel_w = conv_param->kernel_w_;
-  int stride_h = conv_param->stride_h_;
-  int stride_w = conv_param->stride_w_;
-  int pad_h = conv_param->pad_h_;
-  int pad_w = conv_param->pad_w_;
-  int dilation_h = conv_param->dilation_h_;
-  int dilation_w = conv_param->dilation_w_;
-  int in_channel = conv_param->input_channel_;
-  int in_h = conv_param->input_h_;
-  int in_w = conv_param->input_w_;
-  int out_w = conv_param->output_w_;
-  int channel_block = UP_DIV(in_channel, 4);
-  int kernel_plane = kernel_h * kernel_w;
-
-  for (int i = 0; i < real_cal_num; i++) {
-    int block_start = block_index + i;
-    int input_h = block_start / out_w * stride_h - pad_h;
-    int input_w = block_start % out_w * stride_w - pad_w;
-    for (int j = 0; j < kernel_h; j++) {
-      int input_y = input_h + j * dilation_h;
-      if (input_y < 0 || input_y >= in_h) {
-        continue;
-      }
-      int input_y_stride = input_y * in_w * channel_block * C4NUM;
-      for (int n = 0; n < kernel_w; n++) {
-        int input_x = input_w + n * dilation_w;
-        if (input_x < 0 || input_x >= in_w) {
-          continue;
-        }
-        int input_x_stride = input_y_stride + input_x * channel_block * C4NUM;
-        int input_plane_offset = (j * kernel_w + n) * 16 * C4NUM * channel_block + i * C4NUM;
-        for (int m = 0; m < channel_block; m++) {
-          int channel_block_stride = input_x_stride + m * C4NUM;
-          int channel_block_offset = input_plane_offset + m * 16 * C4NUM;
-          (packed_input + channel_block_offset)[0] = (input_data + channel_block_stride)[0];
-          (packed_input + channel_block_offset)[1] = (input_data + channel_block_stride)[1];
-          (packed_input + channel_block_offset)[2] = (input_data + channel_block_stride)[2];
-          (packed_input + channel_block_offset)[3] = (input_data + channel_block_stride)[3];
-        }  // channel_block loop
-      }    // kernel_w loop
-    }      // kernel_h loop
-  }        // tile num loop
-}
-
-void PackWeightFp16(float16_t *weight_data, ConvParameter *conv_param, float16_t *packed_weight) {
-  // original weight format : ohwi
-  int tile_num = 8;
-  int inchannel_block = 4;
-  int kernel_h = conv_param->kernel_h_;
-  int kernel_w = conv_param->kernel_w_;
-  int in_channel = conv_param->input_channel_;
-  int out_channel = conv_param->output_channel_;
-  int kernel_block = UP_DIV(out_channel, tile_num);
-  int channel_block = UP_DIV(in_channel, inchannel_block);
-  int kernel_plane = kernel_h * kernel_w;
-  int pack_weight_size = kernel_block * channel_block * tile_num * inchannel_block * kernel_plane;
-
-  int unit_size = tile_num * inchannel_block;
-  int block_size = pack_weight_size / kernel_block;
-
-  for (int m = 0; m < kernel_plane; m++) {
-    int kernel_plane_stride = m * in_channel;
-    int packed_kernel_plane_stride = m * unit_size * channel_block;
-    for (int i = 0; i < channel_block; i++) {
-      int channel_block_stride = kernel_plane_stride + i * inchannel_block;
-      int packed_channel_block_size = packed_kernel_plane_stride + i * unit_size;
-      int ic_remainder = in_channel - i * inchannel_block;
-      int real_ic_num = ic_remainder < inchannel_block ? ic_remainder : inchannel_block;
-      for (int h = 0; h < real_ic_num; h++) {
-        int block_stride = channel_block_stride + h;
-        int packed_block_stride = packed_channel_block_size + h * tile_num;
-        for (int j = 0; j < kernel_block; j++) {
-          int kernel_block_stride = block_stride + j * tile_num * kernel_plane * in_channel;
-          int packed_kernel_block_size = packed_block_stride + j * block_size;
-          int oc_remainder = out_channel - j * tile_num;
-          int real_oc_num = oc_remainder < tile_num ? oc_remainder : tile_num;
-          for (int k = 0; k < real_oc_num; k++) {
-            float16_t *origin_data_ptr = weight_data + kernel_block_stride + k * kernel_plane * in_channel;
-            float16_t *packed_data_ptr = packed_weight + packed_kernel_block_size + k;
-            *packed_data_ptr = *origin_data_ptr;
-          }
-        }  // kernel block loop
-      }    // inchannel block loop
-    }      // channel block loop
-  }        // kernel plane loop
-}
-
-void PackWeightToC8Fp16(const float16_t *origin_weight_data, float16_t *packed_weight_data, ConvParameter *conv_param) {
-  // origin weight format : ohwi
-  int input_channel = conv_param->input_channel_;
-  int ic8 = UP_DIV(input_channel, C8NUM);
-  int output_channel = conv_param->output_channel_;
-  int kernel_plane = conv_param->kernel_h_ * conv_param->kernel_w_;
-
-  for (int k = 0; k < kernel_plane; k++) {
-    int src_kernel_offset = k * input_channel;
-    int dst_kernel_offset = k * C8NUM;
-    for (int o = 0; o < output_channel; o++) {
-      int src_oc_offset = src_kernel_offset + o * kernel_plane * input_channel;
-      int dst_oc_offset = dst_kernel_offset + o * ic8 * kernel_plane * C8NUM;
-      for (int i = 0; i < input_channel; i++) {
-        int c8_block_num = i / C8NUM;
-        int c8_block_rem = i % C8NUM;
-        int src_ic_offset = src_oc_offset + i;
-        int dst_ic_offset = dst_oc_offset + c8_block_num * kernel_plane * C8NUM + c8_block_rem;
-        (packed_weight_data + dst_ic_offset)[0] = (origin_weight_data + src_ic_offset)[0];
-      }
-    }
-  }
-}
-
-void PackWeightToC4Fp16(const float16_t *origin_weight_data, float16_t *packed_weight_data, ConvParameter *conv_param) {
-  // origin weight format : ohwi
-  int input_channel = conv_param->input_channel_;
-  int ic4 = UP_DIV(input_channel, C4NUM);
-  int output_channel = conv_param->output_channel_;
-  int kernel_plane = conv_param->kernel_h_ * conv_param->kernel_w_;
-
-  for (int k = 0; k < kernel_plane; k++) {
-    int src_kernel_offset = k * input_channel;
-    int dst_kernel_offset = k * C4NUM;
-    for (int o = 0; o < output_channel; o++) {
-      int src_oc_offset = src_kernel_offset + o * kernel_plane * input_channel;
-      int dst_oc_offset = dst_kernel_offset + o * ic4 * kernel_plane * C4NUM;
-      for (int i = 0; i < input_channel; i++) {
-        int c4_block_num = i / C4NUM;
-        int c4_block_rem = i % C4NUM;
-        int src_ic_offset = src_oc_offset + i;
-        int dst_ic_offset = dst_oc_offset + c4_block_num * kernel_plane * C4NUM + c4_block_rem;
-        (packed_weight_data + dst_ic_offset)[0] = (origin_weight_data + src_ic_offset)[0];
-      }
-    }
-  }
-}
-
-void PackNHWCToNC4HW4Fp16(const void *src, void *dst, int batch, int plane, int channel) {
-  int c4 = UP_DIV(channel, C4NUM);
-  for (int b = 0; b < batch; b++) {
-    int src_oc_offset = b * plane * channel;
-    int dst_oc_offset = b * plane * c4 * C4NUM;
-    for (int k = 0; k < plane; k++) {
-      int src_kernel_offset = src_oc_offset + k * channel;
-      int dst_kernel_offset = dst_oc_offset + k * C4NUM;
-      for (int i = 0; i < channel; i++) {
-        int c4_block_num = i / C4NUM;
-        int c4_block_rem = i % C4NUM;
-        int src_ic_offset = src_kernel_offset + i;
-        int dst_ic_offset = dst_kernel_offset + c4_block_num * plane * C4NUM + c4_block_rem;
-        ((float16_t *)dst + dst_ic_offset)[0] = ((float16_t *)src + src_ic_offset)[0];
-      }
-    }
-  }
-}
-
-void PackNCHWToNC4HW4Fp16(const void *src, void *dst, int batch, int plane, int channel) {
-  int c4 = UP_DIV(channel, C4NUM);
-  for (int b = 0; b < batch; b++) {
-    int src_offset = b * plane * channel;
-    int dst_offset = b * plane * c4 * C4NUM;
-    for (int c = 0; c < channel; c++) {
-      int c4_block_num = c / C4NUM;
-      int c4_block_rem = c % C4NUM;
-      int src_c_offset = src_offset + c * plane;
-      int dst_c_offset = dst_offset + c4_block_num * plane * C4NUM;
-      for (int k = 0; k < plane; k++) {
-        int src_kernel_offset = src_c_offset + k;
-        int dst_kernel_offset = dst_c_offset + C4NUM * k + c4_block_rem;
-        ((float16_t *)dst + dst_kernel_offset)[0] = ((float16_t *)src + src_kernel_offset)[0];
-      }
-    }
-  }
-}
-
-void PackNHWCToNHWC4Fp16(const void *src, void *dst, int batch, int plane, int channel) {
-  int ic4 = UP_DIV(channel, C4NUM);
-  int nhwc4_batch_unit_offset = ic4 * C4NUM * plane;
-  int ic_remainder_ = channel % C4NUM;
-  if (ic_remainder_ != 0) {
-    int nhwc4_batch_offset = 0;
-    for (int b = 0; b < batch; b++) {
-      int batch_offset = b * channel * plane;
-      for (int i = 0; i < plane; i++) {
-        memcpy((float16_t *)dst + nhwc4_batch_offset + i * ic4 * C4NUM, (float16_t *)src + batch_offset + i * channel,
-               channel * sizeof(float16_t));
-      }
-      nhwc4_batch_offset += nhwc4_batch_unit_offset;
-    }
-  } else {
-    size_t ori_input_size = batch * plane * channel * sizeof(float16_t);
-    memcpy(dst, src, ori_input_size);
-  }
-}
-
-void PackNCHWToNHWC4Fp16(const void *src, void *dst, int batch, int plane, int channel) {
-  int nhwc4_batch_offset = 0;
-  int ic4 = UP_DIV(channel, C4NUM);
-  int nhwc4_batch_unit_offset = ic4 * C4NUM * plane;
-
-  for (int b = 0; b < batch; b++) {
-    int batch_offset = b * channel * plane;
-    for (int c = 0; c < channel; c++) {
-      int src_c_offset = batch_offset + c * plane;
-      int dst_c_offset = nhwc4_batch_offset + c;
-      for (int i = 0; i < plane; i++) {
-        int src_plane_offset = src_c_offset + i;
-        int dst_plane_offset = dst_c_offset + i * ic4 * C4NUM;
-        ((float16_t *)dst)[dst_plane_offset] = ((float16_t *)src)[src_plane_offset];
-      }
-    }
-    nhwc4_batch_offset += nhwc4_batch_unit_offset;
-  }
-}
-
-void PackNC4HW4ToNHWC4Fp16(const void *src, void *dst, int batch, int plane, int channel) {
-  int c4 = UP_DIV(channel, C4NUM);
-  for (int b = 0; b < batch; b++) {
-    int src_offset = b * plane * c4 * C4NUM;
-    int dst_offset = b * plane * channel;
-    for (int c = 0; c < channel; c++) {
-      int c4_block_num = c / C4NUM;
-      int c4_block_res = c % C4NUM;
-      int src_c_offset = src_offset + c4_block_num * plane * C4NUM + c4_block_res;
-      int dst_c_offset = dst_offset + c4_block_num * C4NUM + c4_block_res;
-      for (int k = 0; k < plane; k++) {
-        int src_kernel_offset = src_c_offset + k * C4NUM;
-        int dst_kernel_offset = dst_c_offset + k * c4 * C4NUM;
-        ((float16_t *)dst + dst_kernel_offset)[0] = ((float16_t *)src + src_kernel_offset)[0];
-      }
-    }
-  }
-}
-
-void PackNC4HW4ToNHWCFp16(const void *src, void *dst, int batch, int plane, int channel) {
-  int c4 = UP_DIV(channel, C4NUM);
-  for (int b = 0; b < batch; b++) {
-    int src_offset = b * plane * c4 * C4NUM;
-    int dst_offset = b * plane * channel;
-    for (int c = 0; c < channel; c++) {
-      int c4_block_num = c / C4NUM;
-      int c4_block_res = c % C4NUM;
-      int src_c_offset = src_offset + c4_block_num * plane * C4NUM + c4_block_res;
-      int dst_c_offset = dst_offset + c;
-      for (int k = 0; k < plane; k++) {
-        int src_kernel_offset = src_c_offset + k * C4NUM;
-        int dst_kernel_offset = dst_c_offset + k * channel;
-        ((float16_t *)dst + dst_kernel_offset)[0] = ((float16_t *)src + src_kernel_offset)[0];
-      }
-    }
-  }
-}
-
-void PackNC4HW4ToNCHWFp16(const void *src, void *dst, int batch, int plane, int channel) {
-  int c4 = UP_DIV(channel, C4NUM);
-  for (int b = 0; b < batch; b++) {
-    int src_offset = b * plane * c4 * C4NUM;
-    int dst_offset = b * plane * channel;
-    for (int c = 0; c < channel; c++) {
-      int c4_block_num = c / C4NUM;
-      int c4_block_res = c % C4NUM;
-      int src_c_offset = src_offset + c4_block_num * plane * C4NUM + c4_block_res;
-      int dst_c_offset = dst_offset + c * plane;
-      for (int k = 0; k < plane; k++) {
-        int src_kernel_offset = src_c_offset + k * C4NUM;
-        int dst_kernel_offset = dst_c_offset + k;
-        ((float16_t *)dst + dst_kernel_offset)[0] = ((float16_t *)src + src_kernel_offset)[0];
-      }
-    }
-  }
-}
-
-void PackNCHWFp32ToNC8HW8Fp16(float *src, float16_t *dst, int batch, int plane, int channel) {
-  int c8 = UP_DIV(channel, C8NUM);
-  for (int b = 0; b < batch; b++) {
-    int src_offset = b * plane * channel;
-    int dst_offset = b * plane * c8 * C8NUM;
-    for (int c = 0; c < channel; c++) {
-      int c8_block_num = c / C8NUM;
-      int c8_block_rem = c % C8NUM;
-      int src_c_offset = src_offset + c * plane;
-      int dst_c_offset = dst_offset + c8_block_num * plane * C8NUM;
-      for (int k = 0; k < plane; k++) {
-        int src_kernel_offset = src_c_offset + k;
-        int dst_kernel_offset = dst_c_offset + C8NUM * k + c8_block_rem;
-        (dst + dst_kernel_offset)[0] = (float16_t)(src + src_kernel_offset)[0];
-      }
-    }
-  }
-}
-
-void PackNHWCFp32ToNHWC8Fp16(float *src, float16_t *dst, int batch, int plane, int channel) {
-  int c8 = UP_DIV(channel, C8NUM);
-  int nhwc8_batch_unit_offset = c8 * C8NUM * plane;
-  int nhwc8_batch_offset = 0;
-  for (int b = 0; b < batch; b++) {
-    int batch_offset = b * channel * plane;
-    for (int i = 0; i < plane; i++) {
-      for (int c = 0; c < channel; c++) {
-        (dst + nhwc8_batch_offset + i * c8 * C8NUM)[c] = (float16_t)(src + batch_offset + i * channel)[c];
-      }
-    }
-    nhwc8_batch_offset += nhwc8_batch_unit_offset;
-  }
-}
-
-void PackNHWC8Fp16ToNHWCFp32(float16_t *src, float *dst, int batch, int plane, int channel) {
-  int c8 = UP_DIV(channel, C8NUM);
-  int nhwc_batch_unit_offset = channel * plane;
-  int nhwc_batch_offset = 0;
-  for (int b = 0; b < batch; b++) {
-    int batch_offset = b * c8 * C8NUM * plane;
-    for (int i = 0; i < plane; i++) {
-      for (int c = 0; c < channel; c++) {
-        (dst + nhwc_batch_offset + i * channel)[c] = (float)(src + batch_offset + i * c8 * C8NUM)[c];
-      }
-    }
-    nhwc_batch_offset += nhwc_batch_unit_offset;
-  }
-}
-#endif
-
 void PackWeightFp32(float *weight_data, ConvParameter *conv_param, float *packed_weight) {
   // original weight format : ohwi
   // todo pack weight for arm32 platform

mindspore/lite/src/runtime/kernel/arm/opclib/pack.h

@@ -23,38 +23,6 @@
 #include "src/runtime/kernel/arm/opclib/conv_parameter.h"
 #include "src/runtime/kernel/arm/opclib/op_base.h"
 
-#ifdef ENABLE_FP16
-void Im2ColPackUnitFp16(float16_t *input_data, ConvParameter *conv_param, float16_t *packed_input, int real_cal_num,
-                        int block_index);
-
-void PackWeightFp16(float16_t *weight_data, ConvParameter *conv_param, float16_t *packed_weight);
-
-void PackWeightToC8Fp16(const float16_t *origin_weight_data, float16_t *packed_weight_data, ConvParameter *conv_param);
-
-void PackWeightToC4Fp16(const float16_t *origin_weight_data, float16_t *packed_weight_data, ConvParameter *conv_param);
-
-void PackNHWCToNC4HW4Fp16(const void *src, void *dst, int batch, int plane, int channel);
-
-void PackNCHWToNC4HW4Fp16(const void *src, void *dst, int batch, int plane, int channel);
-
-void PackNHWCToNHWC4Fp16(const void *src, void *dst, int batch, int plane, int channel);
-
-void PackNCHWToNHWC4Fp16(const void *src, void *dst, int batch, int plane, int channel);
-
-void PackNC4HW4ToNHWC4Fp16(const void *src, void *dst, int batch, int plane, int channel);
-
-void PackNC4HW4ToNHWCFp16(const void *src, void *dst, int batch, int plane, int channel);
-
-void PackNC4HW4ToNCHWFp16(const void *src, void *dst, int batch, int plane, int channel);
-
-void PackNC8HW8ToNHWCFp16(const void *src, void *dst, int batch, int plane, int channel);
-
-void PackNCHWFp32ToNC8HW8Fp16(float *src, float16_t *dst, int batch, int plane, int channel);
-
-void PackNHWCFp32ToNHWC8Fp16(float *src, float16_t *dst, int batch, int plane, int channel);
-
-void PackNHWC8Fp16ToNHWCFp32(float16_t *src, float *dst, int batch, int plane, int channel);
-#endif
 void Im2ColPackUnitFp32(const float *input_data, ConvParameter *conv_param, float *packed_input, int real_cal_num,
                         int block_index);
 

mindspore/lite/src/runtime/kernel/arm/opclib/winograd_transform.cc

@@ -675,518 +675,6 @@ void Conv3x3Fp32OutputTransform(const float *gemm_out, float *out_data, const fl
   }
 }
 
-#ifdef ENABLE_FP16
-// for fp16 convolution 3x3 filter/input/output transform F(4,3)
-void Conv3x3Fp16InputUnit(float16_t *tmp_data, float16_t *trans_input_data, size_t step) {
-  float16x4_t d00 = vld1_f16(tmp_data);
-  float16x4_t d01 = vld1_f16(tmp_data + 4);
-  float16x4_t d02 = vld1_f16(tmp_data + 2 * 4);
-  float16x4_t d03 = vld1_f16(tmp_data + 3 * 4);
-  float16x4_t d04 = vld1_f16(tmp_data + 4 * 4);
-  float16x4_t d05 = vld1_f16(tmp_data + 5 * 4);
-
-  float16x4_t d10 = vld1_f16(tmp_data + 6 * 4);
-  float16x4_t d11 = vld1_f16(tmp_data + 7 * 4);
-  float16x4_t d12 = vld1_f16(tmp_data + 8 * 4);
-  float16x4_t d13 = vld1_f16(tmp_data + 9 * 4);
-  float16x4_t d14 = vld1_f16(tmp_data + 10 * 4);
-  float16x4_t d15 = vld1_f16(tmp_data + 11 * 4);
-
-  float16x4_t d20 = vld1_f16(tmp_data + 12 * 4);
-  float16x4_t d21 = vld1_f16(tmp_data + 13 * 4);
-  float16x4_t d22 = vld1_f16(tmp_data + 14 * 4);
-  float16x4_t d23 = vld1_f16(tmp_data + 15 * 4);
-  float16x4_t d24 = vld1_f16(tmp_data + 16 * 4);
-  float16x4_t d25 = vld1_f16(tmp_data + 17 * 4);
-
-  float16x4_t d30 = vld1_f16(tmp_data + 18 * 4);
-  float16x4_t d31 = vld1_f16(tmp_data + 19 * 4);
-  float16x4_t d32 = vld1_f16(tmp_data + 20 * 4);
-  float16x4_t d33 = vld1_f16(tmp_data + 21 * 4);
-  float16x4_t d34 = vld1_f16(tmp_data + 22 * 4);
-  float16x4_t d35 = vld1_f16(tmp_data + 23 * 4);
-
-  float16x4_t d40 = vld1_f16(tmp_data + 24 * 4);
-  float16x4_t d41 = vld1_f16(tmp_data + 25 * 4);
-  float16x4_t d42 = vld1_f16(tmp_data + 26 * 4);
-  float16x4_t d43 = vld1_f16(tmp_data + 27 * 4);
-  float16x4_t d44 = vld1_f16(tmp_data + 28 * 4);
-  float16x4_t d45 = vld1_f16(tmp_data + 29 * 4);
-
-  float16x4_t d50 = vld1_f16(tmp_data + 30 * 4);
-  float16x4_t d51 = vld1_f16(tmp_data + 31 * 4);
-  float16x4_t d52 = vld1_f16(tmp_data + 32 * 4);
-  float16x4_t d53 = vld1_f16(tmp_data + 33 * 4);
-  float16x4_t d54 = vld1_f16(tmp_data + 34 * 4);
-  float16x4_t d55 = vld1_f16(tmp_data + 35 * 4);
-
-  float16x4_t t00 = vadd_f16(vsub_f16(vmul_n_f16(d00, 4), vmul_n_f16(d20, 5)), d40);
-  float16x4_t t01 = vadd_f16(vsub_f16(vmul_n_f16(d01, 4), vmul_n_f16(d21, 5)), d41);
-  float16x4_t t02 = vadd_f16(vsub_f16(vmul_n_f16(d02, 4), vmul_n_f16(d22, 5)), d42);
-  float16x4_t t03 = vadd_f16(vsub_f16(vmul_n_f16(d03, 4), vmul_n_f16(d23, 5)), d43);
-  float16x4_t t04 = vadd_f16(vsub_f16(vmul_n_f16(d04, 4), vmul_n_f16(d24, 5)), d44);
-  float16x4_t t05 = vadd_f16(vsub_f16(vmul_n_f16(d05, 4), vmul_n_f16(d25, 5)), d45);
-
-  float16x4_t t10 = vadd_f16(vadd_f16(d30, d40), vmul_n_f16(vadd_f16(d10, d20), -4));
-  float16x4_t t11 = vadd_f16(vadd_f16(d31, d41), vmul_n_f16(vadd_f16(d11, d21), -4));
-  float16x4_t t12 = vadd_f16(vadd_f16(d32, d42), vmul_n_f16(vadd_f16(d12, d22), -4));
-  float16x4_t t13 = vadd_f16(vadd_f16(d33, d43), vmul_n_f16(vadd_f16(d13, d23), -4));
-  float16x4_t t14 = vadd_f16(vadd_f16(d34, d44), vmul_n_f16(vadd_f16(d14, d24), -4));
-  float16x4_t t15 = vadd_f16(vadd_f16(d35, d45), vmul_n_f16(vadd_f16(d15, d25), -4));
-
-  float16x4_t t20 = vadd_f16(vsub_f16(d40, d30), vmul_n_f16(vsub_f16(d10, d20), 4));
-  float16x4_t t21 = vadd_f16(vsub_f16(d41, d31), vmul_n_f16(vsub_f16(d11, d21), 4));
-  float16x4_t t22 = vadd_f16(vsub_f16(d42, d32), vmul_n_f16(vsub_f16(d12, d22), 4));
-  float16x4_t t23 = vadd_f16(vsub_f16(d43, d33), vmul_n_f16(vsub_f16(d13, d23), 4));
-  float16x4_t t24 = vadd_f16(vsub_f16(d44, d34), vmul_n_f16(vsub_f16(d14, d24), 4));
-  float16x4_t t25 = vadd_f16(vsub_f16(d45, d35), vmul_n_f16(vsub_f16(d15, d25), 4));
-
-  float16x4_t t30 = vadd_f16(vsub_f16(d40, d20), vmul_n_f16(vsub_f16(d30, d10), 2));
-  float16x4_t t31 = vadd_f16(vsub_f16(d41, d21), vmul_n_f16(vsub_f16(d31, d11), 2));
-  float16x4_t t32 = vadd_f16(vsub_f16(d42, d22), vmul_n_f16(vsub_f16(d32, d12), 2));
-  float16x4_t t33 = vadd_f16(vsub_f16(d43, d23), vmul_n_f16(vsub_f16(d33, d13), 2));
-  float16x4_t t34 = vadd_f16(vsub_f16(d44, d24), vmul_n_f16(vsub_f16(d34, d14), 2));
-  float16x4_t t35 = vadd_f16(vsub_f16(d45, d25), vmul_n_f16(vsub_f16(d35, d15), 2));
-
-  float16x4_t t40 = vadd_f16(vsub_f16(d40, d20), vmul_n_f16(vsub_f16(d10, d30), 2));
-  float16x4_t t41 = vadd_f16(vsub_f16(d41, d21), vmul_n_f16(vsub_f16(d11, d31), 2));
-  float16x4_t t42 = vadd_f16(vsub_f16(d42, d22), vmul_n_f16(vsub_f16(d12, d32), 2));
-  float16x4_t t43 = vadd_f16(vsub_f16(d43, d23), vmul_n_f16(vsub_f16(d13, d33), 2));
-  float16x4_t t44 = vadd_f16(vsub_f16(d44, d24), vmul_n_f16(vsub_f16(d14, d34), 2));
-  float16x4_t t45 = vadd_f16(vsub_f16(d45, d25), vmul_n_f16(vsub_f16(d15, d35), 2));
-
-  float16x4_t t50 = vadd_f16(vsub_f16(vmul_n_f16(d10, 4), vmul_n_f16(d30, 5)), d50);
-  float16x4_t t51 = vadd_f16(vsub_f16(vmul_n_f16(d11, 4), vmul_n_f16(d31, 5)), d51);
-  float16x4_t t52 = vadd_f16(vsub_f16(vmul_n_f16(d12, 4), vmul_n_f16(d32, 5)), d52);
-  float16x4_t t53 = vadd_f16(vsub_f16(vmul_n_f16(d13, 4), vmul_n_f16(d33, 5)), d53);
-  float16x4_t t54 = vadd_f16(vsub_f16(vmul_n_f16(d14, 4), vmul_n_f16(d34, 5)), d54);
-  float16x4_t t55 = vadd_f16(vsub_f16(vmul_n_f16(d15, 4), vmul_n_f16(d35, 5)), d55);
-
-  float16x4_t m00 = vadd_f16(vsub_f16(vmul_n_f16(t00, 4), vmul_n_f16(t02, 5)), t04);
-  float16x4_t m01 = vadd_f16(vadd_f16(t03, t04), vmul_n_f16(vadd_f16(t01, t02), -4));
-  float16x4_t m02 = vadd_f16(vsub_f16(t04, t03), vmul_n_f16(vsub_f16(t01, t02), 4));
-  float16x4_t m03 = vadd_f16(vsub_f16(t04, t02), vmul_n_f16(vsub_f16(t03, t01), 2));
-  float16x4_t m04 = vadd_f16(vsub_f16(t04, t02), vmul_n_f16(vsub_f16(t01, t03), 2));
-  float16x4_t m05 = vadd_f16(vsub_f16(vmul_n_f16(t01, 4), vmul_n_f16(t03, 5)), t05);
-
-  float16x4_t m10 = vadd_f16(vsub_f16(vmul_n_f16(t10, 4), vmul_n_f16(t12, 5)), t14);
-  float16x4_t m11 = vadd_f16(vadd_f16(t13, t14), vmul_n_f16(vadd_f16(t11, t12), -4));
-  float16x4_t m12 = vadd_f16(vsub_f16(t14, t13), vmul_n_f16(vsub_f16(t11, t12), 4));
-  float16x4_t m13 = vadd_f16(vsub_f16(t14, t12), vmul_n_f16(vsub_f16(t13, t11), 2));
-  float16x4_t m14 = vadd_f16(vsub_f16(t14, t12), vmul_n_f16(vsub_f16(t11, t13), 2));
-  float16x4_t m15 = vadd_f16(vsub_f16(vmul_n_f16(t11, 4), vmul_n_f16(t13, 5)), t15);
-
-  float16x4_t m20 = vadd_f16(vsub_f16(vmul_n_f16(t20, 4), vmul_n_f16(t22, 5)), t24);
-  float16x4_t m21 = vadd_f16(vadd_f16(t23, t24), vmul_n_f16(vadd_f16(t21, t22), -4));
-  float16x4_t m22 = vadd_f16(vsub_f16(t24, t23), vmul_n_f16(vsub_f16(t21, t22), 4));
-  float16x4_t m23 = vadd_f16(vsub_f16(t24, t22), vmul_n_f16(vsub_f16(t23, t21), 2));
-  float16x4_t m24 = vadd_f16(vsub_f16(t24, t22), vmul_n_f16(vsub_f16(t21, t23), 2));
-  float16x4_t m25 = vadd_f16(vsub_f16(vmul_n_f16(t21, 4), vmul_n_f16(t23, 5)), t25);
-
-  float16x4_t m30 = vadd_f16(vsub_f16(vmul_n_f16(t30, 4), vmul_n_f16(t32, 5)), t34);
-  float16x4_t m31 = vadd_f16(vadd_f16(t33, t34), vmul_n_f16(vadd_f16(t31, t32), -4));
-  float16x4_t m32 = vadd_f16(vsub_f16(t34, t33), vmul_n_f16(vsub_f16(t31, t32), 4));
-  float16x4_t m33 = vadd_f16(vsub_f16(t34, t32), vmul_n_f16(vsub_f16(t33, t31), 2));
-  float16x4_t m34 = vadd_f16(vsub_f16(t34, t32), vmul_n_f16(vsub_f16(t31, t33), 2));
-  float16x4_t m35 = vadd_f16(vsub_f16(vmul_n_f16(t31, 4), vmul_n_f16(t33, 5)), t35);
-
-  float16x4_t m40 = vadd_f16(vsub_f16(vmul_n_f16(t40, 4), vmul_n_f16(t42, 5)), t44);
-  float16x4_t m41 = vadd_f16(vadd_f16(t43, t44), vmul_n_f16(vadd_f16(t41, t42), -4));
-  float16x4_t m42 = vadd_f16(vsub_f16(t44, t43), vmul_n_f16(vsub_f16(t41, t42), 4));
-  float16x4_t m43 = vadd_f16(vsub_f16(t44, t42), vmul_n_f16(vsub_f16(t43, t41), 2));
-  float16x4_t m44 = vadd_f16(vsub_f16(t44, t42), vmul_n_f16(vsub_f16(t41, t43), 2));
-  float16x4_t m45 = vadd_f16(vsub_f16(vmul_n_f16(t41, 4), vmul_n_f16(t43, 5)), t45);
-
-  float16x4_t m50 = vadd_f16(vsub_f16(vmul_n_f16(t50, 4), vmul_n_f16(t52, 5)), t54);
-  float16x4_t m51 = vadd_f16(vadd_f16(t53, t54), vmul_n_f16(vadd_f16(t51, t52), -4));
-  float16x4_t m52 = vadd_f16(vsub_f16(t54, t53), vmul_n_f16(vsub_f16(t51, t52), 4));
-  float16x4_t m53 = vadd_f16(vsub_f16(t54, t52), vmul_n_f16(vsub_f16(t53, t51), 2));
-  float16x4_t m54 = vadd_f16(vsub_f16(t54, t52), vmul_n_f16(vsub_f16(t51, t53), 2));
-  float16x4_t m55 = vadd_f16(vsub_f16(vmul_n_f16(t51, 4), vmul_n_f16(t53, 5)), t55);
-
-  vst1_f16(trans_input_data, m00);
-  vst1_f16(trans_input_data + step, m01);
-  vst1_f16(trans_input_data + 2 * step, m02);
-  vst1_f16(trans_input_data + 3 * step, m03);
-  vst1_f16(trans_input_data + 4 * step, m04);
-  vst1_f16(trans_input_data + 5 * step, m05);
-
-  vst1_f16(trans_input_data + 6 * step, m10);
-  vst1_f16(trans_input_data + 7 * step, m11);
-  vst1_f16(trans_input_data + 8 * step, m12);
-  vst1_f16(trans_input_data + 9 * step, m13);
-  vst1_f16(trans_input_data + 10 * step, m14);
-  vst1_f16(trans_input_data + 11 * step, m15);
-
-  vst1_f16(trans_input_data + 12 * step, m20);
-  vst1_f16(trans_input_data + 13 * step, m21);
-  vst1_f16(trans_input_data + 14 * step, m22);
-  vst1_f16(trans_input_data + 15 * step, m23);
-  vst1_f16(trans_input_data + 16 * step, m24);
-  vst1_f16(trans_input_data + 17 * step, m25);
-
-  vst1_f16(trans_input_data + 18 * step, m30);
-  vst1_f16(trans_input_data + 19 * step, m31);
-  vst1_f16(trans_input_data + 20 * step, m32);
-  vst1_f16(trans_input_data + 21 * step, m33);
-  vst1_f16(trans_input_data + 22 * step, m34);
-  vst1_f16(trans_input_data + 23 * step, m35);
-
-  vst1_f16(trans_input_data + 24 * step, m40);
-  vst1_f16(trans_input_data + 25 * step, m41);
-  vst1_f16(trans_input_data + 26 * step, m42);
-  vst1_f16(trans_input_data + 27 * step, m43);
-  vst1_f16(trans_input_data + 28 * step, m44);
-  vst1_f16(trans_input_data + 29 * step, m45);
-
-  vst1_f16(trans_input_data + 30 * step, m50);
-  vst1_f16(trans_input_data + 31 * step, m51);
-  vst1_f16(trans_input_data + 32 * step, m52);
-  vst1_f16(trans_input_data + 33 * step, m53);
-  vst1_f16(trans_input_data + 34 * step, m54);
-  vst1_f16(trans_input_data + 35 * step, m55);
-}
-
-void Conv3x3Fp16InputTransform(const float16_t *input_data, float16_t *trans_input, float16_t *tmp_data,
-                               int start_index, int real_cal_num, int out_w_block, ConvParameter *conv_param) {
-  // input data format : nhwc
-  int output_unit = 4;
-  int input_channel = conv_param->input_channel_;
-  int input_width = conv_param->input_w_;
-  int input_height = conv_param->input_h_;
-  int pad_w = conv_param->pad_w_;
-  int pad_h = conv_param->pad_h_;
-  int ic4 = UP_DIV(input_channel, C4NUM);
-
-  for (int cal_id = 0; cal_id < real_cal_num; cal_id++) {
-    int x_id = start_index + cal_id;
-    int origin_x = (x_id % out_w_block) * output_unit - pad_w;
-    int origin_y = (x_id / out_w_block) * output_unit - pad_h;
-    int real_x_start = origin_x > 0 ? 0 : -origin_x;
-    int real_x_end = (origin_x + 6) < input_width ? 6 : (input_width - origin_x);
-    int real_y_start = origin_y > 0 ? 0 : -origin_y;
-    int real_y_end = (origin_y + 6) < input_height ? 6 : (input_height - origin_y);
-
-    int src_plane_offset = input_channel * (origin_y * input_width + origin_x);
-    int dst_plane_offset = cal_id * C4NUM;
-    for (int ic = 0; ic < ic4; ic++) {
-      // clear tmp buffer
-      memset(tmp_data, 0, 6 * 6 * C4NUM * sizeof(float16_t));
-
-      // get real input block with padding
-      int src_ic4_offset = src_plane_offset + ic * C4NUM;
-      for (int interval = real_y_start; interval < real_y_end; interval++) {
-        int src_y_offset = src_ic4_offset + interval * input_width * input_channel + real_x_start * input_channel;
-        int dst_y_offset = interval * 6 * C4NUM + real_x_start * C4NUM;
-        for (int j = 0; j < (real_x_end - real_x_start); j++) {
-          int src_x_offset = src_y_offset + j * input_channel;
-          int dst_x_offset = dst_y_offset + j * C4NUM;
-          float16_t *src_addr = (float16_t *)(input_data) + src_x_offset;
-          float16_t *dst_addr = tmp_data + dst_x_offset;
-          dst_addr[0] = src_addr[0];
-          dst_addr[1] = src_addr[1];
-          dst_addr[2] = src_addr[2];
-          dst_addr[3] = src_addr[3];
-        }
-      }
-
-      // todo
-      // input transform
-      int dst_ic4_offset = dst_plane_offset + ic * 16 * C4NUM;
-      size_t dst_step = ic4 * C4NUM * 16;
-      float16_t *trans_input_ptr = trans_input + dst_ic4_offset;
-      Conv3x3Fp16InputUnit(tmp_data, trans_input_ptr, dst_step);
-    }
-  }
-}
-
-void Conv3x3Fp16FilterTransform(const float16_t *weight_data, float16_t *trans_weight, int iC4, int output_channel,
-                                int kernel_plane) {
-  int dst_step = iC4 * C4NUM * 8;
-  for (int o = 0; o < output_channel; o++) {
-    int oc8_block_num = o / C8NUM;
-    int oc8_block_rem = o % C8NUM;
-    int src_oc_offset = o * iC4 * C4NUM * kernel_plane;
-    int dst_oc_offset = oc8_block_num * C8NUM * iC4 * C4NUM * 36 + oc8_block_rem;
-    for (int i = 0; i < iC4; i++) {
-      const float16_t *src_ic4_ptr = weight_data + src_oc_offset + i * kernel_plane * C4NUM;
-      float16_t *dst_ic4_ptr = trans_weight + dst_oc_offset + i * 8 * C4NUM;
-      float16x4_t g00 = vld1_f16(src_ic4_ptr);
-      float16x4_t g01 = vld1_f16(src_ic4_ptr + 4);
-      float16x4_t g02 = vld1_f16(src_ic4_ptr + 2 * 4);
-      float16x4_t g10 = vld1_f16(src_ic4_ptr + 3 * 4);
-      float16x4_t g11 = vld1_f16(src_ic4_ptr + 4 * 4);
-      float16x4_t g12 = vld1_f16(src_ic4_ptr + 5 * 4);
-      float16x4_t g20 = vld1_f16(src_ic4_ptr + 6 * 4);
-      float16x4_t g21 = vld1_f16(src_ic4_ptr + 7 * 4);
-      float16x4_t g22 = vld1_f16(src_ic4_ptr + 8 * 4);
-
-      float16x4_t dst00 = vmul_n_f16(g00, 0.25);
-      float16x4_t dst01 = vmul_n_f16(g01, 0.25);
-      float16x4_t dst02 = vmul_n_f16(g02, 0.25);
-
-      float16x4_t dst10 = vmul_n_f16(vadd_f16(g00, vadd_f16(g10, g20)), -0.1666666666667);
-      float16x4_t dst11 = vmul_n_f16(vadd_f16(g01, vadd_f16(g11, g21)), -0.1666666666667);
-      float16x4_t dst12 = vmul_n_f16(vadd_f16(g02, vadd_f16(g12, g22)), -0.1666666666667);
-
-      float16x4_t dst20 = vmul_n_f16(vsub_f16(vadd_f16(g00, g20), g10), -0.1666666666667);
-      float16x4_t dst21 = vmul_n_f16(vsub_f16(vadd_f16(g01, g21), g11), -0.1666666666667);
-      float16x4_t dst22 = vmul_n_f16(vsub_f16(vadd_f16(g02, g22), g12), -0.1666666666667);
-
-      float16x4_t dst30 = vadd_f16(vmul_n_f16(g10, 0.08333333333333),
-                                   vadd_f16(vmul_n_f16(g00, 0.04166666666667), vmul_n_f16(g20, 0.1666666666667)));
-      float16x4_t dst31 = vadd_f16(vmul_n_f16(g11, 0.08333333333333),
-                                   vadd_f16(vmul_n_f16(g01, 0.04166666666667), vmul_n_f16(g21, 0.1666666666667)));
-      float16x4_t dst32 = vadd_f16(vmul_n_f16(g12, 0.08333333333333),
-                                   vadd_f16(vmul_n_f16(g02, 0.04166666666667), vmul_n_f16(g22, 0.1666666666667)));
-
-      float16x4_t dst40 = vsub_f16(vadd_f16(vmul_n_f16(g00, 0.04166666666667), vmul_n_f16(g20, 0.1666666666667)),
-                                   vmul_n_f16(g10, 0.08333333333333));
-      float16x4_t dst41 = vsub_f16(vadd_f16(vmul_n_f16(g01, 0.04166666666667), vmul_n_f16(g21, 0.1666666666667)),
-                                   vmul_n_f16(g11, 0.08333333333333));
-      float16x4_t dst42 = vsub_f16(vadd_f16(vmul_n_f16(g02, 0.04166666666667), vmul_n_f16(g22, 0.1666666666667)),
-                                   vmul_n_f16(g12, 0.08333333333333));
-
-      float16x4_t dst50 = g20;
-      float16x4_t dst51 = g21;
-      float16x4_t dst52 = g22;
-
-      float16x4_t m00 = vmul_n_f16(dst00, 0.25);
-      float16x4_t m01 = vmul_n_f16(vadd_f16(dst00, vadd_f16(dst01, dst02)), -0.1666666666667);
-      float16x4_t m02 = vmul_n_f16(vsub_f16(vadd_f16(dst00, dst02), dst01), -0.1666666666667);
-      float16x4_t m03 = vadd_f16(vmul_n_f16(dst01, 0.08333333333333),
-                                 vadd_f16(vmul_n_f16(dst00, 0.04166666666667), vmul_n_f16(dst02, 0.1666666666667)));
-      float16x4_t m04 = vsub_f16(vadd_f16(vmul_n_f16(dst00, 0.04166666666667), vmul_n_f16(dst02, 0.1666666666667)),
-                                 vmul_n_f16(dst01, 0.08333333333333));
-      float16x4_t m05 = dst02;
-
-      float16x4_t m10 = vmul_n_f16(dst10, 0.25);
-      float16x4_t m11 = vmul_n_f16(vadd_f16(dst10, vadd_f16(dst11, dst12)), -0.1666666666667);
-      float16x4_t m12 = vmul_n_f16(vsub_f16(vadd_f16(dst10, dst12), dst11), -0.1666666666667);
-      float16x4_t m13 = vadd_f16(vmul_n_f16(dst11, 0.08333333333333),
-                                 vadd_f16(vmul_n_f16(dst10, 0.04166666666667), vmul_n_f16(dst12, 0.1666666666667)));
-      float16x4_t m14 = vsub_f16(vadd_f16(vmul_n_f16(dst10, 0.04166666666667), vmul_n_f16(dst12, 0.1666666666667)),
-                                 vmul_n_f16(dst11, 0.08333333333333));
-      float16x4_t m15 = dst12;
-
-      float16x4_t m20 = vmul_n_f16(dst20, 0.25);
-      float16x4_t m21 = vmul_n_f16(vadd_f16(dst20, vadd_f16(dst21, dst22)), -0.1666666666667);
-      float16x4_t m22 = vmul_n_f16(vsub_f16(vadd_f16(dst20, dst22), dst21), -0.1666666666667);
-      float16x4_t m23 = vadd_f16(vmul_n_f16(dst21, 0.08333333333333),
-                                 vadd_f16(vmul_n_f16(dst20, 0.04166666666667), vmul_n_f16(dst22, 0.1666666666667)));
-      float16x4_t m24 = vsub_f16(vadd_f16(vmul_n_f16(dst20, 0.04166666666667), vmul_n_f16(dst22, 0.1666666666667)),
-                                 vmul_n_f16(dst21, 0.08333333333333));
-      float16x4_t m25 = dst22;
-
-      float16x4_t m30 = vmul_n_f16(dst30, 0.25);
-      float16x4_t m31 = vmul_n_f16(vadd_f16(dst30, vadd_f16(dst31, dst32)), -0.1666666666667);
-      float16x4_t m32 = vmul_n_f16(vsub_f16(vadd_f16(dst30, dst32), dst31), -0.1666666666667);
-      float16x4_t m33 = vadd_f16(vmul_n_f16(dst31, 0.08333333333333),
-                                 vadd_f16(vmul_n_f16(dst30, 0.04166666666667), vmul_n_f16(dst32, 0.1666666666667)));
-      float16x4_t m34 = vsub_f16(vadd_f16(vmul_n_f16(dst30, 0.04166666666667), vmul_n_f16(dst32, 0.1666666666667)),
-                                 vmul_n_f16(dst31, 0.08333333333333));
-      float16x4_t m35 = dst32;
-
-      float16x4_t m40 = vmul_n_f16(dst40, 0.25);
-      float16x4_t m41 = vmul_n_f16(vadd_f16(dst40, vadd_f16(dst41, dst42)), -0.1666666666667);
-      float16x4_t m42 = vmul_n_f16(vsub_f16(vadd_f16(dst40, dst42), dst41), -0.1666666666667);
-      float16x4_t m43 = vadd_f16(vmul_n_f16(dst41, 0.08333333333333),
-                                 vadd_f16(vmul_n_f16(dst40, 0.04166666666667), vmul_n_f16(dst42, 0.1666666666667)));
-      float16x4_t m44 = vsub_f16(vadd_f16(vmul_n_f16(dst40, 0.04166666666667), vmul_n_f16(dst42, 0.1666666666667)),
-                                 vmul_n_f16(dst41, 0.08333333333333));
-      float16x4_t m45 = dst42;
-
-      float16x4_t m50 = vmul_n_f16(dst50, 0.25);
-      float16x4_t m51 = vmul_n_f16(vadd_f16(dst50, vadd_f16(dst51, dst52)), -0.1666666666667);
-      float16x4_t m52 = vmul_n_f16(vsub_f16(vadd_f16(dst50, dst52), dst51), -0.1666666666667);
-      float16x4_t m53 = vadd_f16(vmul_n_f16(dst51, 0.08333333333333),
-                                 vadd_f16(vmul_n_f16(dst50, 0.04166666666667), vmul_n_f16(dst52, 0.1666666666667)));
-      float16x4_t m54 = vsub_f16(vadd_f16(vmul_n_f16(dst50, 0.04166666666667), vmul_n_f16(dst52, 0.1666666666667)),
-                                 vmul_n_f16(dst51, 0.08333333333333));
-      float16x4_t m55 = dst52;
-
-      for (int j = 0; j < 4; j++) {
-        dst_ic4_ptr[j * 8] = m00[j];
-        dst_ic4_ptr[j * 8 + dst_step] = m01[j];
-        dst_ic4_ptr[j * 8 + 2 * dst_step] = m02[j];
-        dst_ic4_ptr[j * 8 + 3 * dst_step] = m03[j];
-        dst_ic4_ptr[j * 8 + 4 * dst_step] = m04[j];
-        dst_ic4_ptr[j * 8 + 5 * dst_step] = m05[j];
-        dst_ic4_ptr[j * 8 + 6 * dst_step] = m10[j];
-        dst_ic4_ptr[j * 8 + 7 * dst_step] = m11[j];
-        dst_ic4_ptr[j * 8 + 8 * dst_step] = m12[j];
-        dst_ic4_ptr[j * 8 + 9 * dst_step] = m13[j];
-        dst_ic4_ptr[j * 8 + 10 * dst_step] = m14[j];
-        dst_ic4_ptr[j * 8 + 11 * dst_step] = m15[j];
-        dst_ic4_ptr[j * 8 + 12 * dst_step] = m20[j];
-        dst_ic4_ptr[j * 8 + 13 * dst_step] = m21[j];
-        dst_ic4_ptr[j * 8 + 14 * dst_step] = m22[j];
-        dst_ic4_ptr[j * 8 + 15 * dst_step] = m23[j];
-        dst_ic4_ptr[j * 8 + 16 * dst_step] = m24[j];
-        dst_ic4_ptr[j * 8 + 17 * dst_step] = m25[j];
-        dst_ic4_ptr[j * 8 + 18 * dst_step] = m30[j];
-        dst_ic4_ptr[j * 8 + 19 * dst_step] = m31[j];
-        dst_ic4_ptr[j * 8 + 20 * dst_step] = m32[j];
-        dst_ic4_ptr[j * 8 + 21 * dst_step] = m33[j];
-        dst_ic4_ptr[j * 8 + 22 * dst_step] = m34[j];
-        dst_ic4_ptr[j * 8 + 23 * dst_step] = m35[j];
-        dst_ic4_ptr[j * 8 + 24 * dst_step] = m40[j];
-        dst_ic4_ptr[j * 8 + 25 * dst_step] = m41[j];
-        dst_ic4_ptr[j * 8 + 26 * dst_step] = m42[j];
-        dst_ic4_ptr[j * 8 + 27 * dst_step] = m43[j];
-        dst_ic4_ptr[j * 8 + 28 * dst_step] = m44[j];
-        dst_ic4_ptr[j * 8 + 29 * dst_step] = m45[j];
-        dst_ic4_ptr[j * 8 + 30 * dst_step] = m50[j];
-        dst_ic4_ptr[j * 8 + 31 * dst_step] = m51[j];
-        dst_ic4_ptr[j * 8 + 32 * dst_step] = m52[j];
-        dst_ic4_ptr[j * 8 + 33 * dst_step] = m53[j];
-        dst_ic4_ptr[j * 8 + 34 * dst_step] = m54[j];
-        dst_ic4_ptr[j * 8 + 35 * dst_step] = m55[j];
-      }
-    }
-  }
-}
-
-void Conv3x3Fp16OutputUnit(const float16_t *gemm_out, const float16_t *bias_data, float16_t *output_data,
-                           int output_w) {
-  float16x8_t s00 = vld1q_f16(gemm_out);
-  float16x8_t s01 = vld1q_f16(gemm_out + 8);
-  float16x8_t s02 = vld1q_f16(gemm_out + 16);
-  float16x8_t s03 = vld1q_f16(gemm_out + 24);
-  float16x8_t s04 = vld1q_f16(gemm_out + 32);
-  float16x8_t s05 = vld1q_f16(gemm_out + 40);
-
-  float16x8_t s10 = vld1q_f16(gemm_out + 48);
-  float16x8_t s11 = vld1q_f16(gemm_out + 56);
-  float16x8_t s12 = vld1q_f16(gemm_out + 64);
-  float16x8_t s13 = vld1q_f16(gemm_out + 72);
-  float16x8_t s14 = vld1q_f16(gemm_out + 80);
-  float16x8_t s15 = vld1q_f16(gemm_out + 88);
-
-  float16x8_t s20 = vld1q_f16(gemm_out + 96);
-  float16x8_t s21 = vld1q_f16(gemm_out + 104);
-  float16x8_t s22 = vld1q_f16(gemm_out + 112);
-  float16x8_t s23 = vld1q_f16(gemm_out + 120);
-  float16x8_t s24 = vld1q_f16(gemm_out + 128);
-  float16x8_t s25 = vld1q_f16(gemm_out + 136);
-
-  float16x8_t s30 = vld1q_f16(gemm_out + 144);
-  float16x8_t s31 = vld1q_f16(gemm_out + 152);
-  float16x8_t s32 = vld1q_f16(gemm_out + 160);
-  float16x8_t s33 = vld1q_f16(gemm_out + 168);
-  float16x8_t s34 = vld1q_f16(gemm_out + 176);
-  float16x8_t s35 = vld1q_f16(gemm_out + 184);
-
-  float16x8_t s40 = vld1q_f16(gemm_out + 192);
-  float16x8_t s41 = vld1q_f16(gemm_out + 200);
-  float16x8_t s42 = vld1q_f16(gemm_out + 208);
-  float16x8_t s43 = vld1q_f16(gemm_out + 216);
-  float16x8_t s44 = vld1q_f16(gemm_out + 224);
-  float16x8_t s45 = vld1q_f16(gemm_out + 232);
-
-  float16x8_t s50 = vld1q_f16(gemm_out + 240);
-  float16x8_t s51 = vld1q_f16(gemm_out + 248);
-  float16x8_t s52 = vld1q_f16(gemm_out + 256);
-  float16x8_t s53 = vld1q_f16(gemm_out + 264);
-  float16x8_t s54 = vld1q_f16(gemm_out + 272);
-  float16x8_t s55 = vld1q_f16(gemm_out + 280);
-
-  float16x8_t t00 = vaddq_f16(vaddq_f16(vaddq_f16(s00, s10), vaddq_f16(s20, s30)), s40);
-  float16x8_t t01 = vaddq_f16(vaddq_f16(vaddq_f16(s01, s11), vaddq_f16(s21, s31)), s41);
-  float16x8_t t02 = vaddq_f16(vaddq_f16(vaddq_f16(s02, s12), vaddq_f16(s22, s32)), s42);
-  float16x8_t t03 = vaddq_f16(vaddq_f16(vaddq_f16(s03, s13), vaddq_f16(s23, s33)), s43);
-  float16x8_t t04 = vaddq_f16(vaddq_f16(vaddq_f16(s04, s14), vaddq_f16(s24, s34)), s44);
-  float16x8_t t05 = vaddq_f16(vaddq_f16(vaddq_f16(s05, s15), vaddq_f16(s25, s35)), s45);
-
-  float16x8_t t10 = vaddq_f16(vsubq_f16(s10, s20), vmulq_n_f16(vsubq_f16(s30, s40), 2));
-  float16x8_t t11 = vaddq_f16(vsubq_f16(s11, s21), vmulq_n_f16(vsubq_f16(s31, s41), 2));
-  float16x8_t t12 = vaddq_f16(vsubq_f16(s12, s22), vmulq_n_f16(vsubq_f16(s32, s42), 2));
-  float16x8_t t13 = vaddq_f16(vsubq_f16(s13, s23), vmulq_n_f16(vsubq_f16(s33, s43), 2));
-  float16x8_t t14 = vaddq_f16(vsubq_f16(s14, s24), vmulq_n_f16(vsubq_f16(s34, s44), 2));
-  float16x8_t t15 = vaddq_f16(vsubq_f16(s15, s25), vmulq_n_f16(vsubq_f16(s35, s45), 2));
-
-  float16x8_t t20 = vaddq_f16(vaddq_f16(s10, s20), vmulq_n_f16(vaddq_f16(s30, s40), 4));
-  float16x8_t t21 = vaddq_f16(vaddq_f16(s11, s21), vmulq_n_f16(vaddq_f16(s31, s41), 4));
-  float16x8_t t22 = vaddq_f16(vaddq_f16(s12, s22), vmulq_n_f16(vaddq_f16(s32, s42), 4));
-  float16x8_t t23 = vaddq_f16(vaddq_f16(s13, s23), vmulq_n_f16(vaddq_f16(s33, s43), 4));
-  float16x8_t t24 = vaddq_f16(vaddq_f16(s14, s24), vmulq_n_f16(vaddq_f16(s34, s44), 4));
-  float16x8_t t25 = vaddq_f16(vaddq_f16(s15, s25), vmulq_n_f16(vaddq_f16(s35, s45), 4));
-
-  float16x8_t t30 = vaddq_f16(vaddq_f16(vsubq_f16(s10, s20), vmulq_n_f16(vsubq_f16(s30, s40), 8)), s50);
-  float16x8_t t31 = vaddq_f16(vaddq_f16(vsubq_f16(s11, s21), vmulq_n_f16(vsubq_f16(s31, s41), 8)), s51);
-  float16x8_t t32 = vaddq_f16(vaddq_f16(vsubq_f16(s12, s22), vmulq_n_f16(vsubq_f16(s32, s42), 8)), s52);
-  float16x8_t t33 = vaddq_f16(vaddq_f16(vsubq_f16(s13, s23), vmulq_n_f16(vsubq_f16(s33, s43), 8)), s53);
-  float16x8_t t34 = vaddq_f16(vaddq_f16(vsubq_f16(s14, s24), vmulq_n_f16(vsubq_f16(s34, s44), 8)), s54);
-  float16x8_t t35 = vaddq_f16(vaddq_f16(vsubq_f16(s15, s25), vmulq_n_f16(vsubq_f16(s35, s45), 8)), s55);
-
-  float16x8_t d00 = vaddq_f16(vaddq_f16(vaddq_f16(t00, t01), vaddq_f16(t02, t03)), t04);
-  float16x8_t d01 = vaddq_f16(vsubq_f16(t01, t02), vmulq_n_f16(vsubq_f16(t03, t04), 2));
-  float16x8_t d02 = vaddq_f16(vaddq_f16(t01, t02), vmulq_n_f16(vaddq_f16(t03, t04), 4));
-  float16x8_t d03 = vaddq_f16(vaddq_f16(vsubq_f16(t01, t02), vmulq_n_f16(vsubq_f16(t03, t04), 8)), t05);
-
-  float16x8_t d10 = vaddq_f16(vaddq_f16(vaddq_f16(t10, t11), vaddq_f16(t12, t13)), t14);
-  float16x8_t d11 = vaddq_f16(vsubq_f16(t11, t12), vmulq_n_f16(vsubq_f16(t13, t14), 2));
-  float16x8_t d12 = vaddq_f16(vaddq_f16(t11, t12), vmulq_n_f16(vaddq_f16(t13, t14), 4));
-  float16x8_t d13 = vaddq_f16(vaddq_f16(vsubq_f16(t11, t12), vmulq_n_f16(vsubq_f16(t13, t14), 8)), t15);
-
-  float16x8_t d20 = vaddq_f16(vaddq_f16(vaddq_f16(t20, t21), vaddq_f16(t22, t23)), t24);
-  float16x8_t d21 = vaddq_f16(vsubq_f16(t21, t22), vmulq_n_f16(vsubq_f16(t23, t24), 2));
-  float16x8_t d22 = vaddq_f16(vaddq_f16(t21, t22), vmulq_n_f16(vaddq_f16(t23, t24), 4));
-  float16x8_t d23 = vaddq_f16(vaddq_f16(vsubq_f16(t21, t22), vmulq_n_f16(vsubq_f16(t23, t24), 8)), t25);
-
-  float16x8_t d30 = vaddq_f16(vaddq_f16(vaddq_f16(t30, t31), vaddq_f16(t32, t33)), t34);
-  float16x8_t d31 = vaddq_f16(vsubq_f16(t31, t32), vmulq_n_f16(vsubq_f16(t33, t34), 2));
-  float16x8_t d32 = vaddq_f16(vaddq_f16(t31, t32), vmulq_n_f16(vaddq_f16(t33, t34), 4));
-  float16x8_t d33 = vaddq_f16(vaddq_f16(vsubq_f16(t31, t32), vmulq_n_f16(vsubq_f16(t33, t34), 8)), t35);
-
-  vst1q_f16(output_data, d00);
-  vst1q_f16(output_data + 8, d01);
-  vst1q_f16(output_data + 16, d02);
-  vst1q_f16(output_data + 24, d03);
-
-  vst1q_f16(output_data + output_w * 8, d10);
-  vst1q_f16(output_data + output_w * 8 + 8, d11);
-  vst1q_f16(output_data + output_w * 8 + 16, d12);
-  vst1q_f16(output_data + output_w * 8 + 24, d13);
-
-  vst1q_f16(output_data + 2 * output_w * 8, d20);
-  vst1q_f16(output_data + 2 * output_w * 8 + 8, d21);
-  vst1q_f16(output_data + 2 * output_w * 8 + 16, d22);
-  vst1q_f16(output_data + 2 * output_w * 8 + 24, d23);
-
-  vst1q_f16(output_data + 3 * output_w * 8, d30);
-  vst1q_f16(output_data + 3 * output_w * 8 + 8, d31);
-  vst1q_f16(output_data + 3 * output_w * 8 + 16, d32);
-  vst1q_f16(output_data + 3 * output_w * 8 + 24, d33);
-}
-
-void Conv3x3Fp16OutputTransform(const float16_t *gemm_out, float16_t *out_data, const float16_t *bias_data,
-                                int start_index, int real_cal_num, int out_w_block, ConvParameter *conv_param) {
-  int output_channel = conv_param->output_channel_;
-  int output_w = conv_param->output_w_;
-  int output_h = conv_param->output_h_;
-  int oc8 = UP_DIV(output_channel, C8NUM);
-
-  for (int i = 0; i < real_cal_num; i++) {
-    int out_w_index = (start_index + i) % out_w_block;
-    int out_h_index = (start_index + i) / out_w_block;
-    int src_tile_offset = i * oc8 * C8NUM * 36;
-    int dst_tile_offset = 8 * (out_w_index * 4 + out_h_index * 4 * output_w);
-
-    for (int j = 0; j < oc8; j++) {
-      int src_oc8_offset = src_tile_offset + j * 36 * C8NUM;
-      int dst_oc8_offset = dst_tile_offset + j * C8NUM * output_h * output_w;
-      const float16_t *src_ptr = gemm_out + src_oc8_offset;
-      const float16_t *bias_ptr = bias_data + j * C8NUM;
-      float16_t *dst_ptr = out_data + dst_oc8_offset;
-
-      // output transform
-      Conv3x3Fp16OutputUnit(src_ptr, bias_ptr, dst_ptr, output_w);
-    }
-  }
-}
-#endif
-
 // int8 conv3x3
 void Conv3x3Uint8InputUnit(int16_t *tmp_data, int16_t *trans_input_data, size_t step, int input_zp) {
 #ifdef ENABLE_ARM

mindspore/lite/src/runtime/kernel/arm/opclib/winograd_transform.h

@@ -51,22 +51,6 @@ void Conv3x3Fp32OutputUnit(const float *gemm_out, const float *bias_data, float
 void Conv3x3Fp32OutputTransform(const float *gemm_out, float *out_data, const float *bias_data, int start_index,
                                 int real_cal_num, int out_w_block, ConvParameter *conv_param);
 
-#ifdef ENABLE_FP16
-// for fp16 convolution 3x3 filter/input/output transform
-void Conv3x3Fp16InputUnit(float16_t *tmp_data, float16_t *trans_input_data, size_t step);
-
-void Conv3x3Fp16InputTransform(const float16_t *input_data, float16_t *trans_input, float16_t *tmp_data,
-                               int start_index, int real_cal_num, int out_w_block, ConvParameter *conv_param);
-
-void Conv3x3Fp16FilterTransform(const float16_t *weight_data, float16_t *trans_weight, int iC8, int output_channel,
-                                int kernel_plane);
-
-void Conv3x3Fp16OutputUnit(const float16_t *gemm_out, const float16_t *bias_data, float16_t *output_data, int output_w);
-
-void Conv3x3Fp16OutputTransform(const float16_t *gemm_out, float16_t *out_data, const float16_t *bias_data,
-                                int start_index, int real_cal_num, int out_w_block, ConvParameter *conv_param);
-#endif
-
 // for int8 convolution 3x3 filter/input/output transform
 void Conv3x3Uint8InputUnit(int16_t *tmp_data, int16_t *trans_input_data, size_t step, int input_zp);
 

mindspore/lite/src/runtime/kernel/opencl/kernel/arithmetic.cc

@@ -127,9 +127,9 @@ kernel::LiteKernel *OpenCLArithmeticKernelCreator(const std::vector<lite::tensor
   return kernel;
 }
 
-REG_KERNEL(kGPU, PrimitiveType_Mul, OpenCLArithmeticKernelCreator)
-REG_KERNEL(kGPU, PrimitiveType_Add, OpenCLArithmeticKernelCreator)
-REG_KERNEL(kGPU, PrimitiveType_Sub, OpenCLArithmeticKernelCreator)
-REG_KERNEL(kGPU, PrimitiveType_Div, OpenCLArithmeticKernelCreator)
+REG_KERNEL(kGPU, kNumberTypeFloat32, PrimitiveType_Mul, OpenCLArithmeticKernelCreator)
+REG_KERNEL(kGPU, kNumberTypeFloat32, PrimitiveType_Add, OpenCLArithmeticKernelCreator)
+REG_KERNEL(kGPU, kNumberTypeFloat32, PrimitiveType_Sub, OpenCLArithmeticKernelCreator)
+REG_KERNEL(kGPU, kNumberTypeFloat32, PrimitiveType_Div, OpenCLArithmeticKernelCreator)
 }  // namespace mindspore::kernel
 

mindspore/lite/src/runtime/kernel/opencl/kernel/concat.cc

@@ -134,6 +134,6 @@ kernel::LiteKernel *OpenCLConcatKernelCreator(const std::vector<lite::tensor::Te
   return kernel;
 }
 
-REG_KERNEL(kGPU, PrimitiveType_Concat, OpenCLConcatKernelCreator);
+REG_KERNEL(kGPU, kNumberTypeFloat32, PrimitiveType_Concat, OpenCLConcatKernelCreator);
 }  // namespace mindspore::kernel
 

mindspore/lite/src/runtime/kernel/opencl/kernel/conv2d_transpose.cc

@@ -178,6 +178,6 @@ kernel::LiteKernel *OpenCLConv2dTransposeKernelCreator(const std::vector<lite::t
   return kernel;
 }
 
-REG_KERNEL(kGPU, PrimitiveType_DeConv2D, OpenCLConv2dTransposeKernelCreator)
+REG_KERNEL(kGPU, kNumberTypeFloat32, PrimitiveType_DeConv2D, OpenCLConv2dTransposeKernelCreator)
 }  // namespace mindspore::kernel