diff --git a/mindspore/lite/src/runtime/kernel/arm/fp32/group_convolution_fp32.cc b/mindspore/lite/src/runtime/kernel/arm/base/group_convolution.cc
similarity index 94%
rename from mindspore/lite/src/runtime/kernel/arm/fp32/group_convolution_fp32.cc
rename to mindspore/lite/src/runtime/kernel/arm/base/group_convolution.cc
index 087d57c282..02f3342d95 100644
--- a/mindspore/lite/src/runtime/kernel/arm/fp32/group_convolution_fp32.cc
+++ b/mindspore/lite/src/runtime/kernel/arm/base/group_convolution.cc
@@ -14,7 +14,7 @@
  * limitations under the License.
  */
 
-#include "src/runtime/kernel/arm/fp32/group_convolution_fp32.h"
+#include "src/runtime/kernel/arm/base/group_convolution.h"
 #include "src/runtime/infer_manager.h"
 #include "include/errorcode.h"
 
@@ -139,7 +139,7 @@ void GroupConvolutionCPUKernel::SeparateInput(int group_id) {
   int sub_in_channel = conv_param_->input_channel_;
   int ori_in_channel = sub_in_channel * group_num_;
   auto sub_in_data = reinterpret_cast<float *>(group_convs_.at(group_id)->in_tensors().front()->data_c());
-  float *src_ptr = ori_in_data_ + group_id * sub_in_channel;
+  float *src_ptr = reinterpret_cast<float *>(ori_in_data_) + group_id * sub_in_channel;
   float *dst_ptr = sub_in_data;
   for (int i = 0; i < in_plane; ++i) {
     memcpy(dst_ptr, src_ptr, sub_in_channel * sizeof(float));
@@ -155,7 +155,7 @@ void GroupConvolutionCPUKernel::PostConcat(int group_id) {
   int ori_out_channel = sub_out_channel * group_num_;
   auto sub_out_data = reinterpret_cast<float *>(group_convs_.at(group_id)->out_tensors().front()->data_c());
   float *src_ptr = sub_out_data;
-  float *dst_ptr = ori_out_data_ + group_id * sub_out_channel;
+  float *dst_ptr = reinterpret_cast<float *>(ori_out_data_) + group_id * sub_out_channel;
   for (int i = 0; i < out_plane; ++i) {
     memcpy(dst_ptr, src_ptr, sub_out_channel * sizeof(float));
     src_ptr += sub_out_channel;
@@ -164,8 +164,8 @@ void GroupConvolutionCPUKernel::PostConcat(int group_id) {
 }
 
 int GroupConvolutionCPUKernel::Run() {
-  ori_in_data_ = reinterpret_cast<float *>(in_tensors().front()->data_c());
-  ori_out_data_ = reinterpret_cast<float *>(out_tensors().front()->data_c());
+  ori_in_data_ = in_tensors().front()->data_c();
+  ori_out_data_ = out_tensors().front()->data_c();
   for (int i = 0; i < group_num_; ++i) {
     // first, separate group conv input into several parts. This step must be in runtime stage.
     SeparateInput(i);
diff --git a/mindspore/lite/src/runtime/kernel/arm/fp32/group_convolution_fp32.h b/mindspore/lite/src/runtime/kernel/arm/base/group_convolution.h
similarity index 85%
rename from mindspore/lite/src/runtime/kernel/arm/fp32/group_convolution_fp32.h
rename to mindspore/lite/src/runtime/kernel/arm/base/group_convolution.h
index 543646d742..165ca94a5d 100644
--- a/mindspore/lite/src/runtime/kernel/arm/fp32/group_convolution_fp32.h
+++ b/mindspore/lite/src/runtime/kernel/arm/base/group_convolution.h
@@ -14,8 +14,8 @@
  * limitations under the License.
  */
 
-#ifndef MINDSPORE_LITE_SRC_RUNTIME_KERNEL_ARM_FP32_GROUP_CONVOLUTION_H_
-#define MINDSPORE_LITE_SRC_RUNTIME_KERNEL_ARM_FP32_GROUP_CONVOLUTION_H_
+#ifndef MINDSPORE_LITE_SRC_RUNTIME_KERNEL_ARM_GROUP_CONVOLUTION_H_
+#define MINDSPORE_LITE_SRC_RUNTIME_KERNEL_ARM_GROUP_CONVOLUTION_H_
 
 #include <utility>
 #include <vector>
@@ -47,11 +47,9 @@ class GroupConvolutionCPUKernel : public ConvolutionBaseCPUKernel {
  protected:
   std::vector<kernel::LiteKernel *> group_convs_;
   const int group_num_;
-
- private:
-  float *ori_in_data_ = nullptr;   // do not free
-  float *ori_out_data_ = nullptr;  // do not free
+  void *ori_in_data_ = nullptr;   // do not free
+  void *ori_out_data_ = nullptr;  // do not free
 };
 }  // namespace mindspore::kernel
 
-#endif  // MINDSPORE_LITE_SRC_RUNTIME_KERNEL_ARM_FP32_GROUP_CONVOLUTION_H_
+#endif  // MINDSPORE_LITE_SRC_RUNTIME_KERNEL_ARM_GROUP_CONVOLUTION_H_
diff --git a/mindspore/lite/src/runtime/kernel/arm/base/group_convolution_creator.cc b/mindspore/lite/src/runtime/kernel/arm/base/group_convolution_creator.cc
new file mode 100644
index 0000000000..5e447781a5
--- /dev/null
+++ b/mindspore/lite/src/runtime/kernel/arm/base/group_convolution_creator.cc
@@ -0,0 +1,246 @@
+/**
+ * Copyright 2020-2021 Huawei Technologies Co., Ltd
+ *
+ * Licensed under the Apache License, Version 2.0 (the "License");
+ * you may not use this file except in compliance with the License.
+ * You may obtain a copy of the License at
+ *
+ * http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS,
+ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License.
+ */
+
+#include "src/runtime/kernel/arm/base/group_convolution_creator.h"
+#include "src/runtime/kernel/arm/base/group_convolution.h"
+#include "src/runtime/kernel/arm/int8/convolution_int8_creator.h"
+#include "src/runtime/kernel/arm/fp32/convolution_delegate_fp32.h"
+#include "src/runtime/kernel/arm/int8/group_convolution_int8.h"
+
+namespace mindspore::kernel {
+void CopyTensorQuantParam(lite::Tensor *dst, lite::Tensor *src) {
+  for (size_t i = 0; i < src->quant_params().size(); i++) {
+    dst->AddQuantParam(src->quant_params().at(i));
+  }
+}
+
+ConvParameter *CreateNewConvParameter(ConvParameter *parameter) {
+  auto conv_parameter = reinterpret_cast<ConvParameter *>(malloc(sizeof(ConvParameter)));
+  if (conv_parameter == nullptr) {
+    MS_LOG(ERROR) << "Malloc new conv parameter failed.";
+    return nullptr;
+  }
+  memcpy(conv_parameter, parameter, sizeof(ConvParameter));
+  return conv_parameter;
+}
+
+void FreeMemory(ConvParameter *conv_param, const std::vector<lite::Tensor *> &new_inputs,
+                const std::vector<lite::Tensor *> &new_outputs) {
+  if (conv_param != nullptr) {
+    free(conv_param);
+  }
+  for (auto &in_tensor : new_inputs) {
+    delete in_tensor;
+  }
+  for (auto &out_tensor : new_outputs) {
+    delete out_tensor;
+  }
+}
+
+static inline lite::Tensor *TensorMalloc(lite::Tensor *tensor) {
+  if (tensor->MallocData() != lite::RET_OK) {
+    delete tensor;
+    MS_LOG(ERROR) << "malloc tensor data failed.";
+    return nullptr;
+  }
+  return tensor;
+}
+
+lite::Tensor *CreateConstTensor(lite::Tensor *tensor, const std::vector<int> &shape, const int index) {
+  auto new_tensor = new (std::nothrow)
+    lite::Tensor(tensor->data_type(), shape, schema::Format_NHWC, lite::Tensor::Category::CONST_TENSOR);
+  if (new_tensor == nullptr) {
+    MS_LOG(ERROR) << "Create new_tensor failed.";
+    return nullptr;
+  }
+  auto ret = new_tensor->MallocData();
+  if (ret != lite::RET_OK) {
+    delete new_tensor;
+    MS_LOG(ERROR) << "Malloc new_tensor failed.";
+    return nullptr;
+  }
+  memcpy(new_tensor->data_c(), reinterpret_cast<char *>(tensor->data_c()) + index * new_tensor->Size(),
+         new_tensor->Size());
+  return new_tensor;
+}
+
+lite::Tensor *CreateVarTensor(const TensorInfo &tensor_info, bool inferred) {
+  auto tensor = new (std::nothrow) lite::Tensor();
+  if (!tensor) {
+    MS_LOG(ERROR) << "new tensor failed.";
+    return nullptr;
+  }
+  tensor->set_data_type(tensor_info.data_type_);
+  tensor->set_format(tensor_info.format_);
+  tensor->set_category(tensor_info.tensor_type_);
+  if (tensor_info.is_in_) {
+    tensor->set_shape(tensor_info.shape_);
+  }
+
+  if (inferred) {
+    // set shape of out tensor
+    if (!tensor_info.is_in_) {
+      tensor->set_shape(tensor_info.shape_);
+    }
+    return TensorMalloc(tensor);
+  }
+  return tensor;
+}
+
+/* Class GroupConv Creator Implement Part*/
+void GroupConvCreator::CopyQuantParam(std::vector<lite::Tensor *> *tensors) {
+  for (size_t j = 0; j < origin_inputs_.size(); ++j) {
+    CopyTensorQuantParam(tensors->at(j), origin_inputs_.at(j));
+  }
+}
+bool GroupConvCreator::CheckIfValidPoint(void *ptr) {
+  if (ptr == nullptr) {
+    for (auto &sub_conv : group_convs_) {
+      delete sub_conv;
+    }
+    return false;
+  }
+  return true;
+}
+
+int GroupConvCreator::NewInputTensor(std::vector<lite::Tensor *> *tensors) {
+  auto in_tensor = CreateVarTensor(
+    {input_shape_, schema::Format_NHWC, origin_inputs_.at(0)->data_type(), lite::Tensor::Category::VAR, true},
+    infered_);
+  if (!CheckIfValidPoint(in_tensor)) {
+    return lite::RET_ERROR;
+  }
+  tensors->emplace_back(in_tensor);
+  return lite::RET_OK;
+}
+int GroupConvCreator::NewOutputTensor(std::vector<lite::Tensor *> *tensors, lite::Tensor *output) {
+  auto out_tensor =
+    CreateVarTensor({output_shape_, output->format(), output->data_type(), output->category(), false}, infered_);
+  if (!CheckIfValidPoint(out_tensor)) {
+    return lite::RET_ERROR;
+  }
+  if (is_quant_) {
+    CopyTensorQuantParam(out_tensor, output);
+  }
+  tensors->emplace_back(out_tensor);
+  return lite::RET_OK;
+}
+
+int GroupConvCreator::NewConstTensor(std::vector<lite::Tensor *> *tensors, int group_id) {
+  std::vector<std::pair<int, std::vector<int>>> const_tensor_list{std::make_pair(kWeightIndex, filter_shape_)};
+  if (origin_inputs_.size() == 3) {
+    const_tensor_list.emplace_back(std::make_pair(kBiasIndex, bias_shape_));
+  }
+  for (auto &info : const_tensor_list) {
+    auto const_tensor = CreateConstTensor(origin_inputs_.at(info.first), info.second, group_id);
+    if (!CheckIfValidPoint(const_tensor)) {
+      return lite::RET_ERROR;
+    }
+    tensors->emplace_back(const_tensor);
+  }
+  return lite::RET_OK;
+}
+void GroupConvCreator::SetShapeOfTensors() {
+  int new_in_channel = origin_inputs_.at(kWeightIndex)->Channel();
+  int new_out_channel;
+  if (conv_param_->group_ == 0) {
+    MS_LOG(ERROR) << "Divisor 'group' cannot be 0.";
+    return;
+  } else {
+    new_out_channel = origin_inputs_.at(kWeightIndex)->Batch() / conv_param_->group_;
+  }
+
+  /* set shape */
+  set_filter_shape({new_out_channel, conv_param_->kernel_h_, conv_param_->kernel_w_, new_in_channel});
+  set_bias_shape({new_out_channel});
+  if (infered_) {
+    conv_param_->input_channel_ = new_in_channel;
+    conv_param_->output_channel_ = new_out_channel;
+    set_input_shape({origin_inputs_.front()->Batch(), origin_inputs_.front()->Height(), origin_inputs_.front()->Width(),
+                     new_in_channel});
+    set_output_shape({origin_inputs_.front()->Batch(), origin_outputs_.front()->Height(),
+                      origin_outputs_.front()->Width(), new_out_channel});
+  }
+}
+
+int GroupConvCreator::CreatGroupConv() {
+  for (int i = 0; i < conv_param_->group_; ++i) {
+    auto new_conv_parameter = CreateNewConvParameter(conv_param_);
+    if (!CheckIfValidPoint(new_conv_parameter)) {
+      return lite::RET_ERROR;
+    }
+    // create new input for each group
+    std::vector<lite::Tensor *> new_inputs;
+    if (NewInputTensor(&new_inputs) != lite::RET_OK) {
+      MS_LOG(ERROR) << "new input tensor failed.";
+      FreeMemory(new_conv_parameter, new_inputs, {});
+      return lite::RET_ERROR;
+    }
+    // const tensor
+    if (NewConstTensor(&new_inputs, i) != lite::RET_OK) {
+      MS_LOG(ERROR) << "new const tensor failed.";
+      FreeMemory(new_conv_parameter, new_inputs, {});
+      return lite::RET_ERROR;
+    }
+    // create new output tensor
+    std::vector<lite::Tensor *> new_outputs;
+    for (auto &output : origin_outputs_) {
+      if (NewOutputTensor(&new_outputs, output) != lite::RET_OK) {
+        MS_LOG(ERROR) << "new output tensor failed.";
+        FreeMemory(new_conv_parameter, new_inputs, new_outputs);
+        return lite::RET_ERROR;
+      }
+    }
+
+    if (is_quant_) {
+      CopyQuantParam(&new_inputs);
+      group_convs_.emplace_back(CpuConvInt8KernelSelect(new_inputs, new_outputs,
+                                                        reinterpret_cast<OpParameter *>(new_conv_parameter), context_));
+    } else {
+      group_convs_.emplace_back(new (std::nothrow) kernel::ConvolutionDelegateCPUKernel(
+        reinterpret_cast<OpParameter *>(new_conv_parameter), new_inputs, new_outputs, context_));
+    }
+  }
+  return lite::RET_OK;
+}
+
+kernel::LiteKernel *CpuGroupConvFp32KernelCreator(const std::vector<lite::Tensor *> &inputs,
+                                                  const std::vector<lite::Tensor *> &outputs, OpParameter *op_parameter,
+                                                  const lite::InnerContext *ctx) {
+  GroupConvCreator group_conv_creator(inputs, outputs, op_parameter, ctx, false);
+  group_conv_creator.SetShapeOfTensors();
+  if (group_conv_creator.CreatGroupConv() != lite::RET_OK) {
+    MS_LOG(ERROR) << "Create fp32 group conv failed.";
+    return nullptr;
+  }
+  return new (std::nothrow)
+    GroupConvolutionCPUKernel(op_parameter, inputs, outputs, ctx, group_conv_creator.get_group_conv(),
+                              reinterpret_cast<ConvParameter *>(op_parameter)->group_);
+}
+
+kernel::LiteKernel *CpuGroupConvInt8KernelCreator(const std::vector<lite::Tensor *> &inputs,
+                                                  const std::vector<lite::Tensor *> &outputs, OpParameter *op_parameter,
+                                                  const lite::InnerContext *ctx, int group) {
+  GroupConvCreator group_conv_creator(inputs, outputs, op_parameter, ctx, true);
+  group_conv_creator.SetShapeOfTensors();
+  if (group_conv_creator.CreatGroupConv() != lite::RET_OK) {
+    MS_LOG(ERROR) << "Create int8 group conv failed.";
+    return nullptr;
+  }
+  return new (std::nothrow)
+    GroupConvolutionInt8CPUKernel(op_parameter, inputs, outputs, ctx, group_conv_creator.get_group_conv(), group);
+}
+}  // namespace mindspore::kernel
diff --git a/mindspore/lite/src/runtime/kernel/arm/base/group_convolution_creator.h b/mindspore/lite/src/runtime/kernel/arm/base/group_convolution_creator.h
new file mode 100644
index 0000000000..5c0e616ede
--- /dev/null
+++ b/mindspore/lite/src/runtime/kernel/arm/base/group_convolution_creator.h
@@ -0,0 +1,88 @@
+/**
+ * Copyright 2020-2021 Huawei Technologies Co., Ltd
+ *
+ * Licensed under the Apache License, Version 2.0 (the "License");
+ * you may not use this file except in compliance with the License.
+ * You may obtain a copy of the License at
+ *
+ * http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS,
+ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License.
+ */
+
+#ifndef MINDSPORE_LITE_SRC_RUNTIME_KERNEL_ARM_GROUP_CONVOLUTION_CREATOR_H_
+#define MINDSPORE_LITE_SRC_RUNTIME_KERNEL_ARM_GROUP_CONVOLUTION_CREATOR_H_
+
+#include <utility>
+#include <vector>
+#include "src/lite_kernel.h"
+#include "nnacl/conv_parameter.h"
+
+namespace mindspore::kernel {
+struct TensorInfo {
+  std::vector<int> shape_;
+  schema::Format format_;
+  TypeId data_type_;
+  lite::Tensor::Category tensor_type_;
+  bool is_in_;
+};
+
+class GroupConvCreator {
+ public:
+  GroupConvCreator(std::vector<lite::Tensor *> inputs, std::vector<lite::Tensor *> outputs, OpParameter *op_parameter,
+                   const lite::InnerContext *ctx, bool is_quant)
+      : origin_inputs_(std::move(inputs)),
+        origin_outputs_(std::move(outputs)),
+        context_(ctx),
+        infered_(op_parameter->infer_flag_),
+        is_quant_(is_quant) {
+    conv_param_ = reinterpret_cast<ConvParameter *>(op_parameter);
+  }
+
+  ~GroupConvCreator() = default;
+
+ public:
+  void SetShapeOfTensors();
+  int CreatGroupConv();
+  std::vector<kernel::LiteKernel *> get_group_conv() { return group_convs_; }
+
+ protected:
+  void set_input_shape(const std::vector<int> &shape) { input_shape_ = shape; }
+  void set_output_shape(const std::vector<int> &shape) { output_shape_ = shape; }
+  void set_filter_shape(const std::vector<int> &shape) { filter_shape_ = shape; }
+  void set_bias_shape(const std::vector<int> &shape) { bias_shape_ = shape; }
+  void CopyQuantParam(std::vector<lite::Tensor *> *tensors);
+  bool CheckIfValidPoint(void *ptr);
+  int NewInputTensor(std::vector<lite::Tensor *> *tensors);
+  int NewConstTensor(std::vector<lite::Tensor *> *tensors, int group_id);
+  int NewOutputTensor(std::vector<lite::Tensor *> *tensors, lite::Tensor *output);
+
+ private:
+  std::vector<lite::Tensor *> origin_inputs_;
+  std::vector<lite::Tensor *> origin_outputs_;
+  std::vector<kernel::LiteKernel *> group_convs_;
+  std::vector<int> input_shape_;
+  std::vector<int> output_shape_;
+  std::vector<int> filter_shape_;
+  std::vector<int> bias_shape_;
+  const lite::InnerContext *context_;
+  ConvParameter *conv_param_;
+  bool infered_;
+  bool is_quant_;
+};
+
+LiteKernel *CpuGroupConvFp32KernelCreator(const std::vector<lite::Tensor *> &inputs,
+                                          const std::vector<lite::Tensor *> &outputs, OpParameter *op_parameter,
+                                          const lite::InnerContext *ctx);
+
+LiteKernel *CpuGroupConvInt8KernelCreator(const std::vector<lite::Tensor *> &inputs,
+                                          const std::vector<lite::Tensor *> &outputs, OpParameter *op_parameter,
+                                          const lite::InnerContext *ctx, int group);
+
+}  // namespace mindspore::kernel
+
+#endif  // MINDSPORE_LITE_SRC_RUNTIME_KERNEL_ARM_GROUP_CONVOLUTION_CREATOR_H_
diff --git a/mindspore/lite/src/runtime/kernel/arm/fp16/convolution_delegate_fp16.cc b/mindspore/lite/src/runtime/kernel/arm/fp16/convolution_delegate_fp16.cc
index b9c3e5c16f..b1b880e9fd 100644
--- a/mindspore/lite/src/runtime/kernel/arm/fp16/convolution_delegate_fp16.cc
+++ b/mindspore/lite/src/runtime/kernel/arm/fp16/convolution_delegate_fp16.cc
@@ -16,7 +16,6 @@
 
 #include "src/runtime/kernel/arm/fp16/convolution_delegate_fp16.h"
 #include <vector>
-#include "src/runtime/kernel/arm/fp32/convolution_delegate_fp32.h"
 #include "src/runtime/kernel/arm/fp16/convolution_fp16.h"
 #include "src/runtime/kernel/arm/fp16/convolution_winograd_fp16.h"
 #include "src/runtime/kernel/arm/fp16/convolution_1x1_fp16.h"
diff --git a/mindspore/lite/src/runtime/kernel/arm/fp32/convolution_creator_manager.cc b/mindspore/lite/src/runtime/kernel/arm/fp32/convolution_creator_manager.cc
deleted file mode 100644
index 90aa2ee686..0000000000
--- a/mindspore/lite/src/runtime/kernel/arm/fp32/convolution_creator_manager.cc
+++ /dev/null
@@ -1,212 +0,0 @@
-/**
- * Copyright 2020-2021 Huawei Technologies Co., Ltd
- *
- * Licensed under the Apache License, Version 2.0 (the "License");
- * you may not use this file except in compliance with the License.
- * You may obtain a copy of the License at
- *
- * http://www.apache.org/licenses/LICENSE-2.0
- *
- * Unless required by applicable law or agreed to in writing, software
- * distributed under the License is distributed on an "AS IS" BASIS,
- * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
- * See the License for the specific language governing permissions and
- * limitations under the License.
- */
-
-#include <vector>
-#include "src/runtime/kernel/arm/fp32/convolution_creator_manager.h"
-#include "src/runtime/kernel/arm/fp32/convolution_delegate_fp32.h"
-#include "src/runtime/kernel/arm/fp32/group_convolution_fp32.h"
-#include "src/runtime/kernel/arm/fp32/convolution_depthwise_fp32.h"
-#include "src/runtime/kernel/arm/fp32/convolution_depthwise_3x3_fp32.h"
-#include "src/runtime/kernel/arm/fp32/convolution_depthwise_slidewindow_fp32.h"
-#include "src/runtime/kernel/arm/fp32/convolution_depthwise_indirect_fp32.h"
-#include "src/runtime/kernel/arm/int8/convolution_int8.h"
-#include "src/runtime/kernel/arm/int8/convolution_1x1_int8.h"
-#include "src/runtime/kernel/arm/int8/convolution_3x3_int8.h"
-#include "nnacl/conv_parameter.h"
-
-namespace mindspore::lite {
-using mindspore::lite::Format::Format_NHWC;
-
-static inline lite::Tensor *TensorMalloc(lite::Tensor *tensor) {
-  if (tensor->MallocData() != RET_OK) {
-    delete tensor;
-    MS_LOG(ERROR) << "malloc tensor data failed.";
-    return nullptr;
-  }
-  return tensor;
-}
-
-lite::Tensor *CreateConstTensor(lite::Tensor *tensor, const std::vector<int> &shape, const int index) {
-  auto new_tensor =
-    new (std::nothrow) lite::Tensor(tensor->data_type(), shape, Format_NHWC, lite::Tensor::Category::CONST_TENSOR);
-  if (new_tensor == nullptr) {
-    MS_LOG(ERROR) << "Create new_tensor failed.";
-    return nullptr;
-  }
-  auto ret = new_tensor->MallocData();
-  if (ret != RET_OK) {
-    delete new_tensor;
-    MS_LOG(ERROR) << "Malloc new_tensor failed.";
-    return nullptr;
-  }
-  memcpy(new_tensor->data_c(), reinterpret_cast<char *>(tensor->data_c()) + index * new_tensor->Size(),
-         new_tensor->Size());
-  return new_tensor;
-}
-
-lite::Tensor *CreateVarTensor(const TensorInfo &tensor_info, bool inferred) {
-  auto tensor = new (std::nothrow) lite::Tensor();
-  if (!tensor) {
-    MS_LOG(ERROR) << "new tensor failed.";
-    return nullptr;
-  }
-  tensor->set_data_type(tensor_info.data_type_);
-  tensor->set_format(tensor_info.format_);
-  tensor->set_category(tensor_info.tensor_type_);
-  if (tensor_info.is_in_) {
-    tensor->set_shape(tensor_info.shape_);
-  }
-
-  if (inferred) {
-    // set shape of out tensor
-    if (!tensor_info.is_in_) {
-      tensor->set_shape(tensor_info.shape_);
-    }
-    return TensorMalloc(tensor);
-  }
-  return tensor;
-}
-
-/* Kernel creator func part */
-kernel::LiteKernel *CpuConvInt8KernelSelect(const std::vector<lite::Tensor *> &inputs,
-                                            const std::vector<lite::Tensor *> &outputs, OpParameter *op_parameter,
-                                            const InnerContext *ctx) {
-  auto conv_param = reinterpret_cast<ConvParameter *>(op_parameter);
-  kernel::LiteKernel *kernel = nullptr;
-  if (conv_param->kernel_h_ == 3 && conv_param->kernel_w_ == 3 && conv_param->stride_h_ == 1 &&
-      conv_param->stride_w_ == 1 && conv_param->dilation_h_ == 1 && conv_param->dilation_w_ == 1) {
-#ifdef ENABLE_ARM64
-    if (mindspore::lite::IsSupportSDot()) {
-      kernel = new (std::nothrow) kernel::ConvolutionInt8CPUKernel(op_parameter, inputs, outputs, ctx);
-    } else {
-      kernel = new (std::nothrow) kernel::Convolution3x3Int8CPUKernel(op_parameter, inputs, outputs, ctx);
-    }
-#else
-    kernel = new (std::nothrow) kernel::Convolution3x3Int8CPUKernel(op_parameter, inputs, outputs, ctx);
-#endif
-  } else if (conv_param->kernel_h_ == 1 && conv_param->kernel_w_ == 1) {
-    kernel = new (std::nothrow) kernel::Convolution1x1Int8CPUKernel(op_parameter, inputs, outputs, ctx);
-  } else {
-    kernel = new (std::nothrow) kernel::ConvolutionInt8CPUKernel(op_parameter, inputs, outputs, ctx);
-  }
-  return kernel;
-}
-
-kernel::LiteKernel *DispatchConvDw(const std::vector<lite::Tensor *> &inputs,
-                                   const std::vector<lite::Tensor *> &outputs, OpParameter *opParameter,
-                                   const InnerContext *ctx) {
-  auto conv_param = reinterpret_cast<ConvParameter *>(opParameter);
-  kernel::LiteKernel *kernel = nullptr;
-  if (opParameter != nullptr && opParameter->infer_flag_) {
-#if defined(ENABLE_ARM) || (defined(ENABLE_SSE) && !defined(ENABLE_AVX))
-    if (CheckConvDw1DWinograd(conv_param, ctx->thread_num_)) {
-      kernel = new (std::nothrow) kernel::ConvolutionDepthwise3x3CPUKernel(opParameter, inputs, outputs, ctx);
-    }
-#endif
-#if defined(ENABLE_ARM64) || defined(ENABLE_AVX)
-    if (kernel == nullptr && CheckConvDwUseIndirectBuffer(conv_param)) {
-      kernel = new (std::nothrow) kernel::ConvolutionDepthwiseIndirectCPUKernel(opParameter, inputs, outputs, ctx);
-    }
-#endif
-    if (kernel == nullptr && conv_param->input_channel_ < 32) {
-      kernel = new (std::nothrow) kernel::ConvolutionDepthwiseSWCPUKernel(opParameter, inputs, outputs, ctx);
-    }
-  }
-  if (kernel == nullptr) {
-    kernel = new (std::nothrow) kernel::ConvolutionDepthwiseCPUKernel(opParameter, inputs, outputs, ctx);
-  }
-  return kernel;
-}
-
-kernel::LiteKernel *DispatchGroupConv(const std::vector<lite::Tensor *> &inputs,
-                                      const std::vector<lite::Tensor *> &outputs, OpParameter *op_parameter,
-                                      const InnerContext *ctx) {
-  GroupConvCreator group_conv_creator(inputs, outputs, op_parameter, ctx, false);
-  group_conv_creator.SetShapeOfTensors();
-  if (group_conv_creator.CreatGroupConv() != RET_OK) {
-    MS_LOG(ERROR) << "Create group conv failed.";
-    return nullptr;
-  }
-  return new (std::nothrow)
-    kernel::GroupConvolutionCPUKernel(op_parameter, inputs, outputs, ctx, group_conv_creator.get_group_conv(),
-                                      reinterpret_cast<ConvParameter *>(op_parameter)->group_);
-}
-
-/* Class GroupConv Creator Implement Part*/
-void GroupConvCreator::SetShapeOfTensors() {
-  int new_in_channel = origin_inputs_.at(kWeightIndex)->Channel();
-  int new_out_channel;
-  if (conv_param_->group_ == 0) {
-    MS_LOG(ERROR) << "Divisor 'group' cannot be 0.";
-    return;
-  } else {
-    new_out_channel = origin_inputs_.at(kWeightIndex)->Batch() / conv_param_->group_;
-  }
-
-  /* set shape */
-  set_filter_shape({new_out_channel, conv_param_->kernel_h_, conv_param_->kernel_w_, new_in_channel});
-  set_bias_shape({new_out_channel});
-  if (infered_) {
-    conv_param_->input_channel_ = new_in_channel;
-    conv_param_->output_channel_ = new_out_channel;
-    set_input_shape({origin_inputs_.front()->Batch(), origin_inputs_.front()->Height(), origin_inputs_.front()->Width(),
-                     new_in_channel});
-    set_output_shape({origin_inputs_.front()->Batch(), origin_outputs_.front()->Height(),
-                      origin_outputs_.front()->Width(), new_out_channel});
-  }
-}
-
-int GroupConvCreator::CreatGroupConv() {
-  for (int i = 0; i < conv_param_->group_; ++i) {
-    auto new_conv_parameter = CreateNewConvParameter(conv_param_);
-    if (!CheckIfValidPoint(new_conv_parameter)) {
-      return RET_ERROR;
-    }
-    // create new input for each group
-    std::vector<lite::Tensor *> new_inputs;
-    if (NewInputTensor(&new_inputs) != RET_OK) {
-      MS_LOG(ERROR) << "new input tensor failed.";
-      FreeMemory(new_conv_parameter, new_inputs, {});
-      return RET_ERROR;
-    }
-    // const tensor
-    if (NewConstTensor(&new_inputs, i) != RET_OK) {
-      MS_LOG(ERROR) << "new const tensor failed.";
-      FreeMemory(new_conv_parameter, new_inputs, {});
-      return RET_ERROR;
-    }
-    // create new output tensor
-    std::vector<lite::Tensor *> new_outputs;
-    for (auto &output : origin_outputs_) {
-      if (NewOutputTensor(&new_outputs, output) != RET_OK) {
-        MS_LOG(ERROR) << "new output tensor failed.";
-        FreeMemory(new_conv_parameter, new_inputs, new_outputs);
-        return RET_ERROR;
-      }
-    }
-
-    if (is_quant_) {
-      CopyQuantParam(&new_inputs);
-      group_convs_.emplace_back(CpuConvInt8KernelSelect(new_inputs, new_outputs,
-                                                        reinterpret_cast<OpParameter *>(new_conv_parameter), context_));
-    } else {
-      group_convs_.emplace_back(new (std::nothrow) kernel::ConvolutionDelegateCPUKernel(
-        reinterpret_cast<OpParameter *>(new_conv_parameter), new_inputs, new_outputs, context_));
-    }
-  }
-  return RET_OK;
-}
-}  // namespace mindspore::lite
diff --git a/mindspore/lite/src/runtime/kernel/arm/fp32/convolution_creator_manager.h b/mindspore/lite/src/runtime/kernel/arm/fp32/convolution_creator_manager.h
deleted file mode 100644
index e6b194d8a4..0000000000
--- a/mindspore/lite/src/runtime/kernel/arm/fp32/convolution_creator_manager.h
+++ /dev/null
@@ -1,180 +0,0 @@
-/**
- * Copyright 2020-2021 Huawei Technologies Co., Ltd
- *
- * Licensed under the Apache License, Version 2.0 (the "License");
- * you may not use this file except in compliance with the License.
- * You may obtain a copy of the License at
- *
- * http://www.apache.org/licenses/LICENSE-2.0
- *
- * Unless required by applicable law or agreed to in writing, software
- * distributed under the License is distributed on an "AS IS" BASIS,
- * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
- * See the License for the specific language governing permissions and
- * limitations under the License.
- */
-
-#ifndef MINDSPORE_LITE_SRC_RUNTIME_KERNEL_ARM_FP32_CONVOLUTION_CREATOR_H_
-#define MINDSPORE_LITE_SRC_RUNTIME_KERNEL_ARM_FP32_CONVOLUTION_CREATOR_H_
-
-#include <utility>
-#include <vector>
-#include "src/lite_kernel.h"
-#include "nnacl/conv_parameter.h"
-
-namespace mindspore::lite {
-using Category = lite::Tensor::Category;
-using Format = mindspore::schema::Format;
-struct TensorInfo {
-  std::vector<int> shape_;
-  Format format_;
-  TypeId data_type_;
-  Category tensor_type_;
-  bool is_in_;
-};
-
-inline void CopyTensorQuantParam(lite::Tensor *dst, lite::Tensor *src) {
-  for (size_t i = 0; i < src->quant_params().size(); i++) {
-    dst->AddQuantParam(src->quant_params().at(i));
-  }
-}
-
-inline ConvParameter *CreateNewConvParameter(ConvParameter *parameter) {
-  auto conv_parameter = reinterpret_cast<ConvParameter *>(malloc(sizeof(ConvParameter)));
-  if (conv_parameter == nullptr) {
-    MS_LOG(ERROR) << "Malloc new conv parameter failed.";
-    return nullptr;
-  }
-  memcpy(conv_parameter, parameter, sizeof(ConvParameter));
-  return conv_parameter;
-}
-
-inline void FreeMemory(ConvParameter *conv_param, const std::vector<lite::Tensor *> &new_inputs,
-                       const std::vector<lite::Tensor *> &new_outputs) {
-  if (conv_param != nullptr) {
-    free(conv_param);
-  }
-  for (auto &in_tensor : new_inputs) {
-    delete in_tensor;
-  }
-  for (auto &out_tensor : new_outputs) {
-    delete out_tensor;
-  }
-}
-
-lite::Tensor *CreateVarTensor(const TensorInfo &tensor_info, bool inferred);
-
-lite::Tensor *CreateConstTensor(lite::Tensor *tensor, const std::vector<int> &shape, int index);
-
-kernel::LiteKernel *CpuConvInt8KernelSelect(const std::vector<lite::Tensor *> &inputs,
-                                            const std::vector<lite::Tensor *> &outputs, OpParameter *op_parameter,
-                                            const InnerContext *ctx);
-
-kernel::LiteKernel *DispatchConvDw(const std::vector<lite::Tensor *> &inputs,
-                                   const std::vector<lite::Tensor *> &outputs, OpParameter *opParameter,
-                                   const InnerContext *ctx);
-
-kernel::LiteKernel *DispatchGroupConv(const std::vector<lite::Tensor *> &inputs,
-                                      const std::vector<lite::Tensor *> &outputs, OpParameter *op_parameter,
-                                      const InnerContext *ctx);
-
-class GroupConvCreator {
- public:
-  GroupConvCreator(std::vector<lite::Tensor *> inputs, std::vector<lite::Tensor *> outputs, OpParameter *op_parameter,
-                   const InnerContext *ctx, bool is_quant)
-      : origin_inputs_(std::move(inputs)),
-        origin_outputs_(std::move(outputs)),
-        context_(ctx),
-        infered_(op_parameter->infer_flag_),
-        is_quant_(is_quant) {
-    conv_param_ = reinterpret_cast<ConvParameter *>(op_parameter);
-  }
-
-  ~GroupConvCreator() = default;
-
- public:
-  void SetShapeOfTensors();
-  void set_input_shape(const std::vector<int> &shape) { input_shape_ = shape; }
-  void set_output_shape(const std::vector<int> &shape) { output_shape_ = shape; }
-  void set_filter_shape(const std::vector<int> &shape) { filter_shape_ = shape; }
-  void set_bias_shape(const std::vector<int> &shape) { bias_shape_ = shape; }
-
-  std::vector<kernel::LiteKernel *> get_group_conv() { return group_convs_; }
-  int CreatGroupConv();
-
- protected:
-  void FreeSubConv() {
-    for (auto &sub_conv : group_convs_) {
-      delete sub_conv;
-    }
-  }
-
-  bool CheckIfValidPoint(void *ptr) {
-    if (ptr == nullptr) {
-      MS_LOG(ERROR) << "pointer is nullptr.";
-      FreeSubConv();
-      return false;
-    }
-    return true;
-  }
-  int NewInputTensor(std::vector<lite::Tensor *> *tensors) {
-    auto in_tensor = CreateVarTensor(
-      {input_shape_, Format::Format_NHWC, origin_inputs_.at(0)->data_type(), Category::VAR, true}, infered_);
-    if (!CheckIfValidPoint(in_tensor)) {
-      return RET_ERROR;
-    }
-    tensors->emplace_back(in_tensor);
-    return RET_OK;
-  }
-
-  int NewConstTensor(std::vector<lite::Tensor *> *tensors, int group_id) {
-    std::vector<std::pair<int, std::vector<int>>> const_tensor_list{std::make_pair(kWeightIndex, filter_shape_)};
-    if (origin_inputs_.size() == 3) {
-      const_tensor_list.emplace_back(std::make_pair(kBiasIndex, bias_shape_));
-    }
-    for (auto &info : const_tensor_list) {
-      auto const_tensor = CreateConstTensor(origin_inputs_.at(info.first), info.second, group_id);
-      if (!CheckIfValidPoint(const_tensor)) {
-        return RET_ERROR;
-      }
-      tensors->emplace_back(const_tensor);
-    }
-    return RET_OK;
-  }
-
-  int NewOutputTensor(std::vector<lite::Tensor *> *tensors, lite::Tensor *output) {
-    auto out_tensor =
-      CreateVarTensor({output_shape_, output->format(), output->data_type(), output->category(), false}, infered_);
-    if (!CheckIfValidPoint(out_tensor)) {
-      return RET_ERROR;
-    }
-    if (is_quant_) {
-      CopyTensorQuantParam(out_tensor, output);
-    }
-    tensors->emplace_back(out_tensor);
-    return RET_OK;
-  }
-
-  void CopyQuantParam(std::vector<lite::Tensor *> *tensors) {
-    for (size_t j = 0; j < origin_inputs_.size(); ++j) {
-      CopyTensorQuantParam(tensors->at(j), origin_inputs_.at(j));
-    }
-  }
-
- private:
-  std::vector<lite::Tensor *> origin_inputs_;
-  std::vector<lite::Tensor *> origin_outputs_;
-  std::vector<kernel::LiteKernel *> group_convs_;
-  std::vector<int> input_shape_;
-  std::vector<int> output_shape_;
-  std::vector<int> filter_shape_;
-  std::vector<int> bias_shape_;
-  const InnerContext *context_;
-  ConvParameter *conv_param_;
-  bool infered_;
-  bool is_quant_;
-};
-
-}  // namespace mindspore::lite
-
-#endif  // MINDSPORE_LITE_SRC_RUNTIME_KERNEL_ARM_FP32_CONVOLUTION_CREATOR_H_
diff --git a/mindspore/lite/src/runtime/kernel/arm/fp32/convolution_delegate_fp32.cc b/mindspore/lite/src/runtime/kernel/arm/fp32/convolution_delegate_fp32.cc
index 122308cf47..4b39ff7c47 100644
--- a/mindspore/lite/src/runtime/kernel/arm/fp32/convolution_delegate_fp32.cc
+++ b/mindspore/lite/src/runtime/kernel/arm/fp32/convolution_delegate_fp32.cc
@@ -14,12 +14,16 @@
  * limitations under the License.
  */
 
-#include "src/kernel_registry.h"
 #include "src/runtime/kernel/arm/fp32/convolution_delegate_fp32.h"
-#include "src/runtime/kernel/arm/fp32/convolution_creator_manager.h"
+#include "src/kernel_registry.h"
 #include "src/runtime/kernel/arm/fp32/convolution_fp32.h"
 #include "src/runtime/kernel/arm/fp32/convolution_1x1_fp32.h"
 #include "src/runtime/kernel/arm/fp32/convolution_winograd_fp32.h"
+#include "src/runtime/kernel/arm/fp32/convolution_depthwise_fp32.h"
+#include "src/runtime/kernel/arm/fp32/convolution_depthwise_3x3_fp32.h"
+#include "src/runtime/kernel/arm/fp32/convolution_depthwise_slidewindow_fp32.h"
+#include "src/runtime/kernel/arm/fp32/convolution_depthwise_indirect_fp32.h"
+#include "src/runtime/kernel/arm/base/group_convolution_creator.h"
 #include "schema/model_generated.h"
 #include "include/errorcode.h"
 
@@ -157,6 +161,32 @@ kernel::LiteKernel *ConvolutionDelegateCPUKernel::CpuConvFp32KernelSelect() {
   return kernel;
 }
 
+kernel::LiteKernel *DispatchConvDw(const std::vector<lite::Tensor *> &inputs,
+                                   const std::vector<lite::Tensor *> &outputs, OpParameter *opParameter,
+                                   const InnerContext *ctx) {
+  auto conv_param = reinterpret_cast<ConvParameter *>(opParameter);
+  kernel::LiteKernel *kernel = nullptr;
+  if (opParameter != nullptr && opParameter->infer_flag_) {
+#if defined(ENABLE_ARM) || (defined(ENABLE_SSE) && !defined(ENABLE_AVX))
+    if (CheckConvDw1DWinograd(conv_param, ctx->thread_num_)) {
+      kernel = new (std::nothrow) kernel::ConvolutionDepthwise3x3CPUKernel(opParameter, inputs, outputs, ctx);
+    }
+#endif
+#if defined(ENABLE_ARM64) || defined(ENABLE_AVX)
+    if (kernel == nullptr && CheckConvDwUseIndirectBuffer(conv_param)) {
+      kernel = new (std::nothrow) kernel::ConvolutionDepthwiseIndirectCPUKernel(opParameter, inputs, outputs, ctx);
+    }
+#endif
+    if (kernel == nullptr && conv_param->input_channel_ < 32) {
+      kernel = new (std::nothrow) kernel::ConvolutionDepthwiseSWCPUKernel(opParameter, inputs, outputs, ctx);
+    }
+  }
+  if (kernel == nullptr) {
+    kernel = new (std::nothrow) kernel::ConvolutionDepthwiseCPUKernel(opParameter, inputs, outputs, ctx);
+  }
+  return kernel;
+}
+
 /* creator func */
 kernel::LiteKernel *CpuConvFp32KernelCreator(const std::vector<lite::Tensor *> &inputs,
                                              const std::vector<lite::Tensor *> &outputs, OpParameter *op_parameter,
@@ -172,7 +202,7 @@ kernel::LiteKernel *CpuConvFp32KernelCreator(const std::vector<lite::Tensor *> &
   } else if (conv_param->group_ == conv_param->input_channel_ && conv_param->group_ == conv_param->output_channel_) {
     kernel = DispatchConvDw(inputs, outputs, op_parameter, ctx);
   } else {
-    kernel = DispatchGroupConv(inputs, outputs, op_parameter, ctx);
+    kernel = CpuGroupConvFp32KernelCreator(inputs, outputs, op_parameter, ctx);
   }
 
   if (kernel == nullptr) {
diff --git a/mindspore/lite/src/runtime/kernel/arm/fp32/convolution_delegate_fp32.h b/mindspore/lite/src/runtime/kernel/arm/fp32/convolution_delegate_fp32.h
index 1836772ed6..9cfaba5c7b 100644
--- a/mindspore/lite/src/runtime/kernel/arm/fp32/convolution_delegate_fp32.h
+++ b/mindspore/lite/src/runtime/kernel/arm/fp32/convolution_delegate_fp32.h
@@ -18,7 +18,6 @@
 
 #include <vector>
 #include "src/lite_kernel.h"
-#include "src/runtime/kernel/arm/fp32/convolution_creator_manager.h"
 #include "nnacl/conv_parameter.h"
 #include "nnacl/op_base.h"
 
diff --git a/mindspore/lite/src/runtime/kernel/arm/int8/convolution_int8.cc b/mindspore/lite/src/runtime/kernel/arm/int8/convolution_int8.cc
index ed7805ff69..e647b28ce3 100644
--- a/mindspore/lite/src/runtime/kernel/arm/int8/convolution_int8.cc
+++ b/mindspore/lite/src/runtime/kernel/arm/int8/convolution_int8.cc
@@ -19,24 +19,13 @@
 #include "nnacl/int8/conv_int8.h"
 #include "schema/model_generated.h"
 #include "src/kernel_registry.h"
-#include "src/runtime/kernel/arm/fp32/convolution_creator_manager.h"
-#include "src/runtime/kernel/arm/int8/convolution_1x1_int8.h"
-#include "src/runtime/kernel/arm/int8/convolution_3x3_int8.h"
-#include "src/runtime/kernel/arm/int8/group_convolution_int8.h"
-#include "src/runtime/kernel/arm/int8/convolution_depthwise_int8.h"
-#include "src/runtime/kernel/arm/int8/convolution_depthwise_3x3_int8.h"
-#include "src/runtime/kernel/arm/int8/convolution_depthwise_slidewindow_int8.h"
 #include "src/runtime/runtime_api.h"
 #ifdef ENABLE_ARM64
 #include "src/runtime/kernel/arm/int8/opt_op_handler.h"
 #endif
 
-using mindspore::kernel::KERNEL_ARCH::kCPU;
-using mindspore::lite::KernelRegistrar;
 using mindspore::lite::RET_ERROR;
 using mindspore::lite::RET_OK;
-using mindspore::schema::PrimitiveType_Conv2DFusion;
-using mindspore::schema::Format::Format_NHWC;
 
 namespace mindspore::kernel {
 void ConvolutionInt8CPUKernel::CheckSupportOptimize() {
@@ -243,73 +232,4 @@ int ConvolutionInt8CPUKernel::Run() {
   FreeTmpBuffer();
   return RET_OK;
 }
-
-kernel::LiteKernel *CpuGroupConvInt8KernelCreator(const std::vector<lite::Tensor *> &inputs,
-                                                  const std::vector<lite::Tensor *> &outputs, OpParameter *op_parameter,
-                                                  const InnerContext *ctx, int group) {
-  lite::GroupConvCreator group_conv_creator(inputs, outputs, op_parameter, ctx, true);
-  group_conv_creator.SetShapeOfTensors();
-  if (group_conv_creator.CreatGroupConv() != RET_OK) {
-    MS_LOG(ERROR) << "Create group conv failed.";
-    return nullptr;
-  }
-  return new (std::nothrow)
-    GroupConvolutionInt8CPUKernel(op_parameter, inputs, outputs, ctx, group_conv_creator.get_group_conv(), group);
-}
-
-kernel::LiteKernel *CpuConvDwInt8KernelCreator(const std::vector<lite::Tensor *> &inputs,
-                                               const std::vector<lite::Tensor *> &outputs, OpParameter *op_parameter,
-                                               const InnerContext *ctx, const kernel::KernelKey &desc) {
-  auto conv_param = reinterpret_cast<ConvParameter *>(op_parameter);
-  kernel::LiteKernel *kernel = nullptr;
-
-  auto act_quant_size =
-    MSMAX(inputs.at(kInputIndex)->quant_params().size(), outputs.at(kOutputIndex)->quant_params().size());
-  if (act_quant_size == 1) {  // per tensor
-    if (CheckConvDwUse3X3(conv_param) && conv_param->input_channel_ % C8NUM == 0) {
-#ifdef ENABLE_ARM64
-      kernel = new (std::nothrow) kernel::ConvolutionDepthwise3x3Int8CPUKernel(op_parameter, inputs, outputs, ctx);
-#endif
-    }
-    if (kernel == nullptr) {
-      kernel = new (std::nothrow) kernel::ConvolutionDepthwiseInt8CPUKernel(op_parameter, inputs, outputs, ctx);
-    }
-  } else {  // per channel
-    kernel = new (std::nothrow) kernel::ConvolutionDepthwiseSWInt8CPUKernel(op_parameter, inputs, outputs, ctx);
-  }
-  return kernel;
-}
-
-kernel::LiteKernel *CpuConvInt8KernelCreator(const std::vector<lite::Tensor *> &inputs,
-                                             const std::vector<lite::Tensor *> &outputs, OpParameter *op_parameter,
-                                             const InnerContext *ctx, const kernel::KernelKey &desc) {
-  MS_ASSERT(op_parameter != nullptr);
-  MS_ASSERT(desc.type == schema::PrimitiveType_Conv2DFusion);
-  auto conv_param = reinterpret_cast<ConvParameter *>(op_parameter);
-  kernel::LiteKernel *kernel = nullptr;
-
-  if (conv_param->group_ == 1) {
-    kernel = CpuConvInt8KernelSelect(inputs, outputs, op_parameter, ctx);
-  } else if (conv_param->group_ == conv_param->input_channel_ && conv_param->group_ == conv_param->output_channel_) {
-    kernel = CpuConvDwInt8KernelCreator(inputs, outputs, op_parameter, ctx, desc);
-  } else {
-    MS_ASSERT(conv_param->group_ > 1);
-    kernel = CpuGroupConvInt8KernelCreator(inputs, outputs, op_parameter, ctx, conv_param->group_);
-  }
-  if (kernel == nullptr) {
-    MS_LOG(ERROR) << "kernel is nullptr.";
-    free(op_parameter);
-    return nullptr;
-  }
-  auto ret = kernel->Init();
-  if (ret != RET_OK) {
-    MS_LOG(ERROR) << "Init kernel failed, name: " << op_parameter->name_ << ", type: "
-                  << schema::EnumNamePrimitiveType(static_cast<schema::PrimitiveType>(op_parameter->type_));
-    delete kernel;
-    return nullptr;
-  }
-  return kernel;
-}
-
-REG_KERNEL(kCPU, kNumberTypeInt8, PrimitiveType_Conv2DFusion, CpuConvInt8KernelCreator)
 }  // namespace mindspore::kernel
diff --git a/mindspore/lite/src/runtime/kernel/arm/int8/convolution_int8_creator.cc b/mindspore/lite/src/runtime/kernel/arm/int8/convolution_int8_creator.cc
new file mode 100644
index 0000000000..de91a2910d
--- /dev/null
+++ b/mindspore/lite/src/runtime/kernel/arm/int8/convolution_int8_creator.cc
@@ -0,0 +1,118 @@
+/**
+ * Copyright 2021 Huawei Technologies Co., Ltd
+ *
+ * Licensed under the Apache License, Version 2.0 (the "License");
+ * you may not use this file except in compliance with the License.
+ * You may obtain a copy of the License at
+ *
+ * http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS,
+ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License.
+ */
+
+#include "src/runtime/kernel/arm/int8/convolution_int8_creator.h"
+#include "src/runtime/kernel/arm/int8/convolution_int8.h"
+#include "src/runtime/kernel/arm/int8/convolution_1x1_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/convolution_depthwise_3x3_int8.h"
+#include "src/runtime/kernel/arm/int8/convolution_depthwise_slidewindow_int8.h"
+#include "src/runtime/kernel/arm/base/group_convolution_creator.h"
+#include "schema/model_generated.h"
+#include "src/kernel_registry.h"
+#include "include/errorcode.h"
+#include "src/runtime/runtime_api.h"
+
+using mindspore::kernel::KERNEL_ARCH::kCPU;
+using mindspore::lite::KernelRegistrar;
+using mindspore::lite::RET_ERROR;
+using mindspore::lite::RET_OK;
+using mindspore::schema::PrimitiveType_Conv2DFusion;
+using mindspore::schema::Format::Format_NHWC;
+
+namespace mindspore::kernel {
+kernel::LiteKernel *CpuConvDwInt8KernelCreator(const std::vector<lite::Tensor *> &inputs,
+                                               const std::vector<lite::Tensor *> &outputs, OpParameter *op_parameter,
+                                               const InnerContext *ctx, const kernel::KernelKey &desc) {
+  auto conv_param = reinterpret_cast<ConvParameter *>(op_parameter);
+  kernel::LiteKernel *kernel = nullptr;
+
+  auto act_quant_size =
+    MSMAX(inputs.at(kInputIndex)->quant_params().size(), outputs.at(kOutputIndex)->quant_params().size());
+  if (act_quant_size == 1) {  // per tensor
+    if (CheckConvDwUse3X3(conv_param) && conv_param->input_channel_ % C8NUM == 0) {
+#ifdef ENABLE_ARM64
+      kernel = new (std::nothrow) kernel::ConvolutionDepthwise3x3Int8CPUKernel(op_parameter, inputs, outputs, ctx);
+#endif
+    }
+    if (kernel == nullptr) {
+      kernel = new (std::nothrow) kernel::ConvolutionDepthwiseInt8CPUKernel(op_parameter, inputs, outputs, ctx);
+    }
+  } else {  // per channel
+    kernel = new (std::nothrow) kernel::ConvolutionDepthwiseSWInt8CPUKernel(op_parameter, inputs, outputs, ctx);
+  }
+  return kernel;
+}
+
+/* Kernel creator func part */
+kernel::LiteKernel *CpuConvInt8KernelSelect(const std::vector<lite::Tensor *> &inputs,
+                                            const std::vector<lite::Tensor *> &outputs, OpParameter *op_parameter,
+                                            const InnerContext *ctx) {
+  auto conv_param = reinterpret_cast<ConvParameter *>(op_parameter);
+  kernel::LiteKernel *kernel = nullptr;
+  if (conv_param->kernel_h_ == 3 && conv_param->kernel_w_ == 3 && conv_param->stride_h_ == 1 &&
+      conv_param->stride_w_ == 1 && conv_param->dilation_h_ == 1 && conv_param->dilation_w_ == 1) {
+#ifdef ENABLE_ARM64
+    if (mindspore::lite::IsSupportSDot()) {
+      kernel = new (std::nothrow) ConvolutionInt8CPUKernel(op_parameter, inputs, outputs, ctx);
+    } else {
+      kernel = new (std::nothrow) Convolution3x3Int8CPUKernel(op_parameter, inputs, outputs, ctx);
+    }
+#else
+    kernel = new (std::nothrow) kernel::Convolution3x3Int8CPUKernel(op_parameter, inputs, outputs, ctx);
+#endif
+  } else if (conv_param->kernel_h_ == 1 && conv_param->kernel_w_ == 1) {
+    kernel = new (std::nothrow) Convolution1x1Int8CPUKernel(op_parameter, inputs, outputs, ctx);
+  } else {
+    kernel = new (std::nothrow) ConvolutionInt8CPUKernel(op_parameter, inputs, outputs, ctx);
+  }
+  return kernel;
+}
+
+kernel::LiteKernel *CpuConvInt8KernelCreator(const std::vector<lite::Tensor *> &inputs,
+                                             const std::vector<lite::Tensor *> &outputs, OpParameter *op_parameter,
+                                             const InnerContext *ctx, const kernel::KernelKey &desc) {
+  MS_ASSERT(op_parameter != nullptr);
+  MS_ASSERT(desc.type == schema::PrimitiveType_Conv2DFusion);
+  auto conv_param = reinterpret_cast<ConvParameter *>(op_parameter);
+  kernel::LiteKernel *kernel = nullptr;
+
+  if (conv_param->group_ == 1) {
+    kernel = CpuConvInt8KernelSelect(inputs, outputs, op_parameter, ctx);
+  } else if (conv_param->group_ == conv_param->input_channel_ && conv_param->group_ == conv_param->output_channel_) {
+    kernel = CpuConvDwInt8KernelCreator(inputs, outputs, op_parameter, ctx, desc);
+  } else {
+    MS_ASSERT(conv_param->group_ > 1);
+    kernel = CpuGroupConvInt8KernelCreator(inputs, outputs, op_parameter, ctx, conv_param->group_);
+  }
+  if (kernel == nullptr) {
+    MS_LOG(ERROR) << "kernel is nullptr.";
+    free(op_parameter);
+    return nullptr;
+  }
+  auto ret = kernel->Init();
+  if (ret != RET_OK) {
+    MS_LOG(ERROR) << "Init kernel failed, name: " << op_parameter->name_ << ", type: "
+                  << schema::EnumNamePrimitiveType(static_cast<schema::PrimitiveType>(op_parameter->type_));
+    delete kernel;
+    return nullptr;
+  }
+  return kernel;
+}
+
+REG_KERNEL(kCPU, kNumberTypeInt8, PrimitiveType_Conv2DFusion, CpuConvInt8KernelCreator)
+}  // namespace mindspore::kernel
diff --git a/mindspore/lite/src/runtime/kernel/arm/int8/convolution_int8_creator.h b/mindspore/lite/src/runtime/kernel/arm/int8/convolution_int8_creator.h
new file mode 100644
index 0000000000..4743d320d6
--- /dev/null
+++ b/mindspore/lite/src/runtime/kernel/arm/int8/convolution_int8_creator.h
@@ -0,0 +1,29 @@
+/**
+ * Copyright 2021 Huawei Technologies Co., Ltd
+ *
+ * Licensed under the Apache License, Version 2.0 (the "License");
+ * you may not use this file except in compliance with the License.
+ * You may obtain a copy of the License at
+ *
+ * http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS,
+ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License.
+ */
+
+#ifndef MINDSPORE_LITE_SRC_RUNTIME_KERNEL_ARM_INT8_CONVOLUTION_INT8_CREATOR_H_
+#define MINDSPORE_LITE_SRC_RUNTIME_KERNEL_ARM_INT8_CONVOLUTION_INT8_CREATOR_H_
+
+#include <vector>
+#include "nnacl/op_base.h"
+#include "src/lite_kernel.h"
+
+namespace mindspore::kernel {
+LiteKernel *CpuConvInt8KernelSelect(const std::vector<lite::Tensor *> &inputs,
+                                    const std::vector<lite::Tensor *> &outputs, OpParameter *op_parameter,
+                                    const lite::InnerContext *ctx);
+}
+#endif  // MINDSPORE_LITE_SRC_RUNTIME_KERNEL_ARM_INT8_CONVOLUTION_INT8_CREATOR_H_
diff --git a/mindspore/lite/src/runtime/kernel/arm/int8/group_convolution_int8.cc b/mindspore/lite/src/runtime/kernel/arm/int8/group_convolution_int8.cc
index 080166ea4e..4a423fa7b3 100644
--- a/mindspore/lite/src/runtime/kernel/arm/int8/group_convolution_int8.cc
+++ b/mindspore/lite/src/runtime/kernel/arm/int8/group_convolution_int8.cc
@@ -15,14 +15,6 @@
  */
 
 #include "src/runtime/kernel/arm/int8/group_convolution_int8.h"
-#include "schema/model_generated.h"
-#include "src/kernel_registry.h"
-#include "include/errorcode.h"
-
-using mindspore::kernel::KERNEL_ARCH::kCPU;
-using mindspore::lite::KernelRegistrar;
-using mindspore::lite::RET_ERROR;
-using mindspore::lite::RET_OK;
 
 namespace mindspore::kernel {
 void GroupConvolutionInt8CPUKernel::SeparateInput(int group_id) {
@@ -30,7 +22,7 @@ void GroupConvolutionInt8CPUKernel::SeparateInput(int group_id) {
   int sub_in_channel = conv_param_->input_channel_;
   int ori_in_channel = sub_in_channel * group_num_;
   auto sub_in_data = reinterpret_cast<int8_t *>(group_convs_.at(group_id)->in_tensors().front()->data_c());
-  int8_t *src_ptr = ori_in_data_ + group_id * sub_in_channel;
+  int8_t *src_ptr = reinterpret_cast<int8_t *>(ori_in_data_) + group_id * sub_in_channel;
   int8_t *dst_ptr = sub_in_data;
   for (int i = 0; i < in_plane; ++i) {
     memcpy(dst_ptr, src_ptr, sub_in_channel * sizeof(int8_t));
@@ -45,29 +37,11 @@ void GroupConvolutionInt8CPUKernel::PostConcat(int group_id) {
   int ori_out_channel = sub_out_channel * group_num_;
   auto sub_out_data = reinterpret_cast<int8_t *>(group_convs_.at(group_id)->out_tensors().front()->data_c());
   int8_t *src_ptr = sub_out_data;
-  int8_t *dst_ptr = ori_out_data_ + group_id * sub_out_channel;
+  int8_t *dst_ptr = reinterpret_cast<int8_t *>(ori_out_data_) + group_id * sub_out_channel;
   for (int i = 0; i < out_plane; ++i) {
     memcpy(dst_ptr, src_ptr, sub_out_channel * sizeof(int8_t));
     src_ptr += sub_out_channel;
     dst_ptr += ori_out_channel;
   }
 }
-
-int GroupConvolutionInt8CPUKernel::Run() {
-  ori_in_data_ = reinterpret_cast<int8_t *>(in_tensors().front()->data_c());
-  ori_out_data_ = reinterpret_cast<int8_t *>(out_tensors().front()->data_c());
-  for (int i = 0; i < group_num_; ++i) {
-    // first, separate group conv input into several parts. This step must be in runtime stage.
-    SeparateInput(i);
-    // sun kernels run
-    auto ret = group_convs_.at(i)->Run();
-    if (ret != RET_OK) {
-      MS_LOG(ERROR) << "sub kernel " << i << " execute failed.";
-      return ret;
-    }
-    // post process, concat all outputs of sub-kernels into one output
-    PostConcat(i);
-  }
-  return RET_OK;
-}
 }  // namespace mindspore::kernel
diff --git a/mindspore/lite/src/runtime/kernel/arm/int8/group_convolution_int8.h b/mindspore/lite/src/runtime/kernel/arm/int8/group_convolution_int8.h
index 614536260d..53a49cdf42 100644
--- a/mindspore/lite/src/runtime/kernel/arm/int8/group_convolution_int8.h
+++ b/mindspore/lite/src/runtime/kernel/arm/int8/group_convolution_int8.h
@@ -21,7 +21,7 @@
 #include <vector>
 #include "src/lite_kernel.h"
 #include "nnacl/op_base.h"
-#include "src/runtime/kernel/arm/fp32/group_convolution_fp32.h"
+#include "src/runtime/kernel/arm/base/group_convolution.h"
 
 namespace mindspore::kernel {
 class GroupConvolutionInt8CPUKernel : public GroupConvolutionCPUKernel {
@@ -32,14 +32,8 @@ class GroupConvolutionInt8CPUKernel : public GroupConvolutionCPUKernel {
       : GroupConvolutionCPUKernel(parameter, inputs, outputs, ctx, std::move(group_convs), group_num) {
   }  // opParameter(in channel, out channel) in this kernel has been split to groups, if
   // you want to get real params, multiply in channel / out channel with group num
-
-  int Run() override;
   void SeparateInput(int group_id) override;
   void PostConcat(int group_id) override;
-
- private:
-  int8_t *ori_in_data_ = nullptr;   // do not free
-  int8_t *ori_out_data_ = nullptr;  // do not free
 };
 }  // namespace mindspore::kernel