train on device

mindspore/lite/schema/model.fbs

@@ -174,6 +174,19 @@ union PrimitiveType {
     Where,
     OneHot,
     Lstm,
+    Conv2DGradFilter,
+    Conv2DGradInput,
+    PoolingGrad,
+    BNGradInput,
+    OptMomentum,
+    BiasGrad,
+    SoftmaxCrossEntropy,
+    AddGrad,
+    SubGrad,
+    MulGrad,
+    DivGrad,
+    PowerGrad,
+    ActivationGrad,
     PriorBox
 }
 

mindspore/lite/schema/ops.fbs

@@ -55,7 +55,25 @@ enum ActivationType : byte {
     LINEAR = 15,
     UNKNOW = 16
 }
-
+enum ActivationGradType : byte {
+    NO_ACTIVATION = 0,
+    RELU = 1,
+    SIGMOID = 2,
+    RELU6 = 3,
+    ELU = 4,
+    LEAKY_RELU = 5,
+    ABS = 6,
+    RELU1 = 7,
+    SOFTSIGN = 8,
+    SOFTPLUS = 9,
+    TANH = 10,
+    SELU = 11,
+    HSWISH = 12,
+    HSIGMOID = 13,
+    THRESHOLDRELU = 14,
+    LINEAR = 15,
+    UNKNOW = 16
+}
 enum ReduceType : byte {
     REDUCE_MAX = 0,
     REDUCE_MEAN = 1,
@@ -125,6 +143,10 @@ table SoftMax {
 table Activation {
     type: ActivationType = 0;
 }
+table ActivationGrad {
+    type: ActivationGradType = 0;
+}
+
 
 table Conv2D {
     format: Format = 0;
@@ -146,7 +168,45 @@ table Conv2D {
     activationType: ActivationType = 0;
 }
 
-table FusedBatchNorm {
+table Conv2DGradFilter {
+    format: Format = 0;
+    group: int;
+    channelIn: int;
+    channelOut: int;
+    kernelW: int;
+    kernelH: int;
+    strideW: int;
+    strideH: int;
+    padMode: PadMode;
+    padUp: int;
+    padDown: int;
+    padLeft: int;
+    padRight: int;
+    dilateW: int;
+    dilateH: int;
+    hasBias: bool = false;
+    activationType: ActivationType = 0;
+}
+
+table Conv2DGradInput {
+    format: Format = 0;
+    group: int;
+    channelIn: int;
+    channelOut: int;
+    kernelW: int;
+    kernelH: int;
+    strideW: int;
+    strideH: int;
+    padMode: PadMode;
+    padUp: int;
+    padDown: int;
+    padLeft: int;
+    padRight: int;
+    dilateW: int;
+    dilateH: int;
+    hasBias: bool = false;
+    activationType: ActivationType = 0;
+}table FusedBatchNorm {
     epsilon: float = 0.00001;   // eg. epsilon=0.001
     momentum: float = 0.9;
     spatial: int = 1;
@@ -156,6 +216,31 @@ table CaffeBatchNorm {
     epsilon: float;   // eg. epsilon=0.001
 }
 
+table BiasGrad {
+    axis: [int];
+}
+
+
+table SoftmaxCrossEntropy {
+    axis: [int];
+}
+
+
+table PoolingGrad {
+    format: Format = 0;
+    poolingMode: PoolMode;
+    global: bool = false;
+    windowW: int;
+    windowH: int;
+    strideW: int;
+    strideH: int;
+    padMode: PadMode;
+    padUp: int;
+    padDown: int;
+    padLeft: int;
+    padRight: int;
+    roundMode: RoundMode;
+}
 table Shape {
 }
 
@@ -286,7 +371,10 @@ table DeConv2D {
     hasBias: bool = false;
     activationType: ActivationType = 0;
 }
-
+table BNGradInput {
+    eps : float;
+    channels: int;
+}
 table Scale {
     format: Format = 0;
 }
@@ -307,6 +395,17 @@ table Mul {
 table Div {
 }
 
+table AddGrad {
+}
+
+table SubGrad {
+}
+
+table MulGrad {
+}
+
+table DivGrad {
+}
 table RealDiv {
 }
 
@@ -389,7 +488,11 @@ table Power {
     scale: float;
     shift: float;
 }
-
+table PowerGrad {
+    power: float;
+    scale: float;
+    shift: float;
+}
 table ArgMax {
     axis: int;
     outMaxValue: bool;
@@ -712,6 +815,10 @@ table SquaredDifference {
 table TupleGetItem {
 }
 
+table OptMomentum {
+}
+
+
 table Where{
 }
 

mindspore/lite/src/common/file_utils_ext.cc

@@ -0,0 +1,53 @@
+/**
+ * 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 <cmath>
+#include <cstddef>
+#include <iostream>
+#include "src/common/file_utils.h"
+#include "src/common/file_utils_ext.h"
+
+namespace mindspore {
+namespace lite {
+static int CompareOutputRelativeData(float *output_data, float *correct_data, int data_size) {
+  float error = 0;
+
+  // relative error
+  float diffSum = 0.0f;
+  float sum = 0.0f;
+  for (int i = 0; i < data_size; i++) {
+    sum += std::abs(correct_data[i]);
+  }
+  for (int i = 0; i < data_size; i++) {
+    float diff = std::abs(output_data[i] - correct_data[i]);
+    diffSum += diff;
+  }
+  error = diffSum / sum;
+  if (error > 1e-4) {
+    std::cout << "has accuracy error!\n" << error << "\n";
+    return 1;
+  }
+  return 0;
+}
+
+int CompareRelativeOutput(float *output_data, std::string file_path) {
+  size_t output_size;
+  auto ground_truth = reinterpret_cast<float *>(mindspore::lite::ReadFile(file_path.c_str(), &output_size));
+  size_t output_num = output_size / sizeof(float);
+  std::cout << "output num : " << output_num << "\n";
+  return CompareOutputRelativeData(output_data, ground_truth, output_num);
+}
+}  // namespace lite
+}  // namespace mindspore

mindspore/lite/src/common/file_utils_ext.h

@@ -0,0 +1,28 @@
+/**
+ * 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_COMMON_FILE_UTILS_EXT_H_
+#define MINDSPORE_LITE_COMMON_FILE_UTILS_EXT_H_
+#include <string>
+
+
+namespace mindspore {
+namespace lite {
+int CompareRelativeOutput(float *output_data, std::string file_path);
+
+}
+}  // namespace mindspore
+#endif  // MINDSPORE_LITE_COMMON_FILE_UTILS_EXT_H_

mindspore/lite/src/lite_kernel.h

@@ -64,7 +64,7 @@ class LiteKernel {
   LiteKernel() = default;
   explicit LiteKernel(OpParameter *parameter, const std::vector<lite::tensor::Tensor *> &inputs,
                       const std::vector<lite::tensor::Tensor *> &outputs)
-      : opParameter(parameter), inputs_(inputs), outputs_(outputs) {
+      : opParameter(parameter), inputs_(inputs), outputs_(outputs), train_mode(false) {
     this->in_kernel_.clear();
     this->out_kernel_.clear();
   }
@@ -77,7 +77,10 @@ class LiteKernel {
   virtual int Run() { return -1; }
 
   std::string Name() { return this->name; }
-
+  virtual void train() { train_mode = true; }
+  virtual bool is_train() { return train_mode == true; }
+  virtual void eval() { train_mode = false; }
+  virtual bool is_eval() { return train_mode == false; }
   void set_name(const std::string &name) { this->name = name; }
 
   schema::PrimitiveType type() { return (schema::PrimitiveType)this->opParameter->type_; }
@@ -117,6 +120,7 @@ class LiteKernel {
   std::vector<lite::tensor::Tensor *> outputs_;
   std::vector<LiteKernel *> in_kernel_;
   std::vector<LiteKernel *> out_kernel_;
+  bool train_mode;
 };
 
 class SubGraphKernel : public LiteKernel {

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

@@ -0,0 +1,110 @@
+/**
+ * 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/fp32/activation_grad.h"
+#include "schema/model_generated.h"
+#include "src/kernel_registry.h"
+#include "src/runtime/runtime_api.h"
+#include "include/errorcode.h"
+
+using mindspore::lite::KernelRegistrar;
+using mindspore::kernel::KERNEL_ARCH::kCPU;
+using mindspore::lite::RET_ERROR;
+using mindspore::lite::RET_OK;
+using mindspore::schema::ActivationGradType_HSWISH;
+using mindspore::schema::ActivationGradType_LEAKY_RELU;
+using mindspore::schema::ActivationGradType_RELU;
+using mindspore::schema::ActivationGradType_RELU6;
+using mindspore::schema::PrimitiveType_ActivationGrad;
+
+namespace mindspore::kernel {
+int ActivationGradCPUKernel::Init() {
+    outputs_[0]->set_shape(inputs_[0]->shape());
+    return RET_OK;
+}
+
+int ActivationGradCPUKernel::ReSize() { return RET_OK; }
+
+int ActivationGradCPUKernel::DoActivation(int task_id) {
+  auto yt_addr = reinterpret_cast<float *>(inputs_.at(0)->Data());
+  auto input_addr = reinterpret_cast<float *>(inputs_.at(1)->Data());
+  auto output_addr = reinterpret_cast<float *>(outputs_.at(0)->Data());
+  auto length = inputs_.at(0)->ElementsNum();
+
+  auto error_code = RET_OK;
+
+  if (type_ == schema::ActivationGradType_RELU) {
+    error_code = ReluGrad(yt_addr, input_addr, length, output_addr);
+  } else if (type_ == schema::ActivationGradType_RELU6) {
+    error_code = Relu6Grad(yt_addr, input_addr, length, output_addr);
+  } else if (type_ == schema::ActivationGradType_LEAKY_RELU) {
+    error_code = LReluGrad(yt_addr, input_addr, length, output_addr, alpha_);
+  } else if (type_ == schema::ActivationGradType_SIGMOID) {
+    error_code = SigmoidGrad(yt_addr, input_addr, length, output_addr);
+  } else if (type_ == schema::ActivationGradType_TANH) {
+    error_code = TanhGrad(yt_addr, input_addr, length, output_addr);
+  } else if (type_ == schema::ActivationGradType_HSWISH) {
+    error_code = HSwishGrad(yt_addr, input_addr, length, output_addr);
+    } else if (type_ == schema::ActivationGradType_HSIGMOID) {
+    error_code = HSigmoidGrad(yt_addr, input_addr, length, output_addr);
+  } else {
+    MS_LOG(ERROR) << "Activation type error";
+    return RET_ERROR;
+  }
+  if (error_code != RET_OK) {
+    return RET_ERROR;
+  }
+  return RET_OK;
+}
+
+int ActivationGradRun(int task_id, LiteParallelGroupEnv *penv, void *cdata) {
+  auto activationGrad_kernel = reinterpret_cast<ActivationGradCPUKernel *>(cdata);
+  auto error_code = activationGrad_kernel->DoActivation(task_id);
+  if (error_code != RET_OK) {
+    MS_LOG(ERROR) << "ActivationGradRun error task_id[" << task_id << "] error_code[" << error_code << "]";
+    return RET_ERROR;
+  }
+  return RET_OK;
+}
+
+int ActivationGradCPUKernel::Run() {
+  int error_code = LiteBackendParallelLaunch(ActivationGradRun, this, thread_count_);
+  if (error_code != RET_OK) {
+    MS_LOG(ERROR) << "Activation function error error_code[" << error_code << "]";
+    return RET_ERROR;
+  }
+  return RET_OK;
+}
+
+kernel::LiteKernel *CpuActivationGradFp32KernelCreator(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_ActivationGrad);
+  auto *kernel = new (std::nothrow) ActivationGradCPUKernel(opParameter, inputs, outputs);
+  MS_ASSERT(kernel != nullptr);
+  auto ret = kernel->Init();
+  if (ret != RET_OK) {
+    MS_LOG(ERROR) << "InferShape kernel failed, name: " << opParameter->name_
+                  << ", type: "
+                  << schema::EnumNamePrimitiveType(static_cast<schema::PrimitiveType>(opParameter->type_));
+  }
+  return kernel;
+}
+
+REG_KERNEL(kCPU, kNumberTypeFloat32, PrimitiveType_ActivationGrad, CpuActivationGradFp32KernelCreator)
+}  // namespace mindspore::kernel

mindspore/lite/src/runtime/kernel/arm/fp32/activation_grad.h

@@ -0,0 +1,50 @@
+/**
+ * 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_FP32_ACTIVATION_GRAD_H_
+#define MINDSPORE_LITE_SRC_RUNTIME_KERNEL_ARM_FP32_ACTIVATION_GRAD_H_
+
+#include <vector>
+#include "src/lite_kernel.h"
+#include "ir/anf.h"
+
+#include "src/runtime/kernel/arm/opclib/activation_grad.h"
+
+namespace mindspore::kernel {
+class ActivationGradCPUKernel : public LiteKernel {
+ public:
+  explicit ActivationGradCPUKernel(OpParameter *param, const std::vector<lite::tensor::Tensor *> &inputs,
+                                   const std::vector<lite::tensor::Tensor *> &outputs)
+      : LiteKernel(param, inputs, outputs) {
+    ActivationGradParameter *param_act_grad = reinterpret_cast<ActivationGradParameter *>(param);
+    type_ = param_act_grad->type_;
+    alpha_ = param_act_grad->alpha_;
+  }
+  ~ActivationGradCPUKernel() override = default;
+
+  int Init() override;
+  int ReSize() override;
+  int Run() override;
+  int DoActivation(int task_id);
+
+ private:
+  int thread_count_;
+  int type_;
+  float alpha_;
+};
+}  // namespace mindspore::kernel
+
+#endif  // MINDSPORE_LITE_SRC_RUNTIME_KERNEL_ARM_FP32_ACTIVATION_GRAD_H_

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

@@ -0,0 +1,285 @@
+/**
+ * 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 "schema/model_generated.h"
+#include "src/kernel_registry.h"
+#include "src/runtime/kernel/arm/opclib/fp32/reduce_grad.h"
+#include "src/runtime/kernel/arm/fp32/arithmetic_grad.h"
+#include "src/runtime/kernel/arm/opclib/fp32/arithmetic_grad.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 {
+namespace {
+constexpr int kArithGradOpInputNum = 3;
+constexpr int kArithGradOpOutputNum = 2;
+}  // namespace
+
+int ArithmeticGradCPUKernel::Init() {
+  auto ret = InferShape();
+  return ret;
+}
+
+int ArithmeticGradCPUKernel::InferShape() {
+  if (inputs_.size() != kArithGradOpInputNum) {
+    MS_LOG(ERROR) << "The number of input must be " << kArithGradOpInputNum;
+    return RET_ERROR;
+  }
+  if (outputs_.size() != kArithGradOpOutputNum) {
+    MS_LOG(ERROR) << "The number of output must be " << kArithGradOpOutputNum;
+    return RET_ERROR;
+  }
+  auto dy = inputs_[0];
+  auto x1 = inputs_[1];
+  auto x2 = inputs_[2];
+  auto dx1 = outputs_[0];
+  auto dx2 = outputs_[1];
+
+  MS_ASSERT(dy != nullptr);
+  MS_ASSERT(x1 != nullptr);
+  MS_ASSERT(x2 != nullptr);
+  MS_ASSERT(dx1 != nullptr);
+  MS_ASSERT(dx2 != nullptr);
+
+  auto inShape0 = x1->shape();
+  auto inShape1 = x2->shape();
+  auto outShape = dy->shape();
+
+  if ((type() == PrimitiveType_AddGrad) || (type() == PrimitiveType_SubGrad)) {
+    arithmeticParameter_->ndim_ = outShape.size();
+    auto fillDimNum0 = outShape.size() - inShape0.size();
+    auto fillDimNum1 = outShape.size() - inShape1.size();
+    int j0 = 0;
+    int j1 = 0;
+    for (unsigned int i = 0; i < outShape.size(); i++) {
+      arithmeticParameter_->in_shape0_[i] = (i < fillDimNum0) ? 1 : inShape0[j0++];
+      arithmeticParameter_->in_shape1_[i] = (i < fillDimNum1) ? 1 : inShape1[j1++];
+      arithmeticParameter_->out_shape_[i] = outShape[i];
+    }
+  } else {
+    // if (inShape0.size() < inShape1.size())
+    if (dx1->ElementsNum() < dx2->ElementsNum()) {
+      arithmeticParameter_->ndim_ = inShape1.size();
+      if (type() == PrimitiveType_MulGrad)
+        arithmetic_grad_ = &ArithmeticGradCPUKernel::ArithmeticGradMul2L;
+      else if (type() == PrimitiveType_DivGrad)
+        arithmetic_grad_ = &ArithmeticGradCPUKernel::ArithmeticGradDiv2L;
+
+      auto fillDimNum = inShape1.size() - inShape0.size();  // This will not work for batch!
+      int j = 0;
+      for (unsigned int i = 0; i < inShape1.size(); i++) {
+        if (i < fillDimNum) {
+          arithmeticParameter_->in_shape1_[i] = 1;
+        } else {
+          arithmeticParameter_->in_shape1_[i] = inShape0[j++];
+        }
+        arithmeticParameter_->in_shape0_[i] = inShape1[i];
+        arithmeticParameter_->out_shape_[i] = outShape[i];
+      }
+    } else if (dx2->ElementsNum() < dx1->ElementsNum()) {  // if (inShape0.size() > inShape1.size())
+      arithmeticParameter_->ndim_ = inShape0.size();
+      if (type() == PrimitiveType_MulGrad)
+        arithmetic_grad_ = &ArithmeticGradCPUKernel::ArithmeticGradMul1L;
+      else if (type() == PrimitiveType_DivGrad)
+        arithmetic_grad_ = &ArithmeticGradCPUKernel::ArithmeticGradDiv1L;
+      arithmeticParameter_->broadcasting_ = true;
+      arithmeticParameter_->ndim_ = inShape0.size();
+      int j = 0;
+      auto fillDimNum = inShape0.size() - inShape1.size();
+      for (unsigned int i = 0; i < inShape0.size(); i++) {
+        if (i < fillDimNum) {
+          arithmeticParameter_->in_shape1_[i] = 1;
+        } else {
+          arithmeticParameter_->in_shape1_[i] = inShape1[j++];
+        }
+        arithmeticParameter_->in_shape0_[i] = inShape0[i];
+        arithmeticParameter_->out_shape_[i] = outShape[i];
+      }
+    } else {
+      arithmeticParameter_->broadcasting_ = false;
+      for (unsigned int i = 0; i < inShape0.size(); i++) {
+        arithmeticParameter_->in_shape1_[i] = inShape1[i];
+        arithmeticParameter_->in_shape0_[i] = inShape0[i];
+        arithmeticParameter_->out_shape_[i] = outShape[i];
+      }
+    }
+    tile_data0 = new (std::nothrow) float[inputs_.at(0)->ElementsNum()];
+    MS_ASSERT(tile_data0 != nullptr);
+    tile_data1 = new (std::nothrow) float[inputs_.at(0)->ElementsNum()];
+    MS_ASSERT(tile_data1 != nullptr);
+    if (type() == PrimitiveType_DivGrad) {
+      tile_data2 = new (std::nothrow) float[inputs_.at(0)->ElementsNum()];
+      MS_ASSERT(tile_data2 != nullptr);
+    }
+  }
+
+  dx1->set_shape(x1->shape());
+  dx2->set_shape(x2->shape());
+  // outTensor->set_shape(out_shape);
+  dx1->set_data_type(dy->data_type());
+  dx2->set_data_type(dy->data_type());
+  return RET_OK;
+}
+
+void ArithmeticGradCPUKernel::ArithmeticGradAdd(float *dy, int dy_size, float *dx1, int dx1_size, float *dx2,
+                                                int dx2_size) {
+  if (dx1_size == dy_size)
+    memcpy(dx1, dy, dy_size * sizeof(float));
+  else
+    ReduceSumByAxes(dy, arithmeticParameter_->out_shape_, dx1, arithmeticParameter_->in_shape0_,
+                    arithmeticParameter_->ndim_);
+  if (dx2_size == dy_size)
+    memcpy(dx2, dy, dy_size * sizeof(float));
+  else
+    ReduceSumByAxes(dy, arithmeticParameter_->out_shape_, dx2, arithmeticParameter_->in_shape1_,
+                    arithmeticParameter_->ndim_);
+}
+
+void ArithmeticGradCPUKernel::ArithmeticGradSub(float *dy, int dy_size, float *dx1, int dx1_size, float *dx2,
+                                                int dx2_size) {
+  if (dx1_size == dy_size)
+    memcpy(dx1, dy, dy_size * sizeof(float));
+  else
+    ReduceSumByAxes(dy, arithmeticParameter_->out_shape_, dx1, arithmeticParameter_->in_shape0_,
+                    arithmeticParameter_->ndim_);
+  if (dx2_size == dy_size) {
+    for (int i = 0; i < dx2_size; i++) {
+      dx2[i] = -dy[i];
+    }
+  } else {
+    ReduceSumByAxes(dy, arithmeticParameter_->out_shape_, dx2, arithmeticParameter_->in_shape1_,
+                    arithmeticParameter_->ndim_);
+    for (int i = 0; i < dx2_size; i++) {
+      dx2[i] = -dx2[i];
+    }
+  }
+}
+
+void ArithmeticGradCPUKernel::ArithmeticGradMul(float *dy, int dy_size, float *dx1, int dx1_size, float *dx2,
+                                                int dx2_size) {
+  auto x1_data = reinterpret_cast<float *>(inputs_[1]->Data());
+  auto x2_data = reinterpret_cast<float *>(inputs_[2]->Data());
+  ElementMul(dy, x1_data, dx2, dy_size);
+  ElementMul(dy, x2_data, dx1, dy_size);
+}
+
+void ArithmeticGradCPUKernel::ArithmeticGradMul1L(float *dy, int dy_size, float *dx1, int dx1_size, float *dx2,
+                                                  int dx2_size) {
+  auto x1_data = reinterpret_cast<float *>(inputs_[1]->Data());
+  auto x2_data = reinterpret_cast<float *>(inputs_[2]->Data());
+  ElementMul(dy, x1_data, tile_data0, dy_size);
+  ReduceSumByAxes(tile_data0, arithmeticParameter_->in_shape0_, dx2, arithmeticParameter_->in_shape1_,
+                  arithmeticParameter_->ndim_);
+
+  BroadcastMul(dy, x2_data, tile_data0, tile_data1, dx1, dy_size, arithmeticParameter_);  // broadcast directly to dx1
+}
+
+void ArithmeticGradCPUKernel::ArithmeticGradMul2L(float *dy, int dy_size, float *dx1, int dx1_size, float *dx2,
+                                                  int dx2_size) {
+  auto x1_data = reinterpret_cast<float *>(inputs_[1]->Data());
+  auto x2_data = reinterpret_cast<float *>(inputs_[2]->Data());
+  ElementMul(dy, x2_data, tile_data0, dy_size);
+  ReduceSumByAxes(tile_data0, arithmeticParameter_->in_shape0_, dx1, arithmeticParameter_->in_shape1_,
+                  arithmeticParameter_->ndim_);
+
+  BroadcastMul(dy, x1_data, tile_data0, tile_data1, dx2, dy_size, arithmeticParameter_);  // broadcast directly to dx2
+}
+
+void ArithmeticGradCPUKernel::ArithmeticGradDiv(float *dy, int dy_size, float *dx1, int dx1_size, float *dx2,
+                                                int dx2_size) {
+  auto x1 = reinterpret_cast<float *>(inputs_[1]->Data());
+  auto x2 = reinterpret_cast<float *>(inputs_[2]->Data());
+  ElementDiv(dy, x2, dx1, dy_size);
+  ElementMulAndDivNegSquare(dy, x1, x2, dx2, dy_size);
+}
+
+void ArithmeticGradCPUKernel::ArithmeticGradDiv1L(float *dy, int dy_size, float *dx1, int dx1_size, float *dx2,
+                                                  int dx2_size) {
+  auto x1_data = reinterpret_cast<float *>(inputs_[1]->Data());
+  auto x2_data = reinterpret_cast<float *>(inputs_[2]->Data());
+
+  ElementMul(x2_data, x2_data, dx2, dx2_size);
+  ElementMul(x1_data, dy, dx1, dy_size);  // use dx1 buffer
+  BroadcastDiv(dx1, dx2, tile_data0, tile_data1, tile_data2, dy_size,
+               arithmeticParameter_);  // broadcast directly to dx1
+  ReduceSumByAxes(tile_data2, arithmeticParameter_->in_shape0_, dx2, arithmeticParameter_->in_shape1_,
+                  arithmeticParameter_->ndim_);
+  for (int i = 0; i < dx2_size; i++) dx2[i] = -dx2[i];
+  // ReduceNegSumPrefix(tile_data2, dy_size, dx2, dx2_size); //then reduce into dx2
+
+  // broadcasting x2
+  BroadcastDiv(dy, x2_data, tile_data0, tile_data1, dx1, dy_size, arithmeticParameter_);  // broadcast directly to dx1
+}
+
+void ArithmeticGradCPUKernel::ArithmeticGradDiv2L(float *dy, int dy_size, float *dx1, int dx1_size, float *dx2,
+                                                  int dx2_size) {
+  auto x1_data = reinterpret_cast<float *>(inputs_[1]->Data());
+  auto x2_data = reinterpret_cast<float *>(inputs_[2]->Data());
+
+  // dx1 = dy/x2
+  ElementDiv(dy, x2_data, tile_data0, dy_size);  // first multiply into temp
+  ReduceSumByAxes(tile_data0, arithmeticParameter_->in_shape0_, dx1, arithmeticParameter_->in_shape1_,
+                  arithmeticParameter_->ndim_);
+
+  // dx2 = -dy*x1/(x2*x2)
+  BroadcastMul(dy, x1_data, tile_data0, tile_data1, tile_data2, dy_size, arithmeticParameter_);  // broadcast numerator
+  ElementDivNegSquare(tile_data2, x2_data, dx2, dy_size);
+}
+
+int ArithmeticGradCPUKernel::ReSize() { return RET_OK; }
+
+int ArithmeticGradCPUKernel::Run() {
+  auto dy = reinterpret_cast<float *>(inputs_[0]->Data());
+  // auto input1_data1 = reinterpret_cast<float *>(inputs_[1]->Data());
+  auto dx1 = reinterpret_cast<float *>(outputs_[0]->Data());
+  auto dx2 = reinterpret_cast<float *>(outputs_[1]->Data());
+
+  size_t dy_size = inputs_.at(0)->ElementsNum();
+  size_t dx1_size = outputs_.at(0)->ElementsNum();
+  size_t dx2_size = outputs_[1]->ElementsNum();
+  (this->*arithmetic_grad_)(dy, dy_size, dx1, dx1_size, dx2, dx2_size);
+  return RET_OK;
+}
+
+kernel::LiteKernel *CpuArithmeticGradFp32KernelCreator(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_EXCEPTION_IF_NULL(opParameter);
+  if (opParameter == nullptr) {
+    return nullptr;
+  }
+  auto *kernel = new (std::nothrow) ArithmeticGradCPUKernel(opParameter, inputs, outputs);
+  MS_ASSERT(kernel != 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_));
+    delete kernel;
+    return nullptr;
+  }
+  return kernel;
+}
+
+REG_KERNEL(kCPU, kNumberTypeFloat32, PrimitiveType_MulGrad, CpuArithmeticGradFp32KernelCreator)
+REG_KERNEL(kCPU, kNumberTypeFloat32, PrimitiveType_AddGrad, CpuArithmeticGradFp32KernelCreator)
+REG_KERNEL(kCPU, kNumberTypeFloat32, PrimitiveType_SubGrad, CpuArithmeticGradFp32KernelCreator)
+REG_KERNEL(kCPU, kNumberTypeFloat32, PrimitiveType_DivGrad, CpuArithmeticGradFp32KernelCreator)
+}  // namespace mindspore::kernel

mindspore/lite/src/runtime/kernel/arm/fp32/arithmetic_grad.h

@@ -0,0 +1,90 @@
+/**
+ * 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_FP32_ARITHMETIC_GRAD_H_
+#define MINDSPORE_LITE_SRC_RUNTIME_KERNEL_ARM_FP32_ARITHMETIC_GRAD_H_
+
+#include <vector>
+#include "src/lite_kernel.h"
+#include "src/runtime/kernel/arm/opclib/fp32/arithmetic.h"
+#include "schema/model_generated.h"
+#include "ir/anf.h"
+
+using mindspore::schema::PrimitiveType_AddGrad;
+using mindspore::schema::PrimitiveType_DivGrad;
+using mindspore::schema::PrimitiveType_MulGrad;
+using mindspore::schema::PrimitiveType_SubGrad;
+
+namespace mindspore::kernel {
+
+class ArithmeticGradCPUKernel;
+
+class ArithmeticGradCPUKernel : public LiteKernel {
+  typedef void (ArithmeticGradCPUKernel::*ArithmeticGradOperation)(float *, int, float *, int, float *, int);
+
+ public:
+  explicit ArithmeticGradCPUKernel(OpParameter *parameter, const std::vector<lite::tensor::Tensor *> &inputs,
+                                   const std::vector<lite::tensor::Tensor *> &outputs)
+      : LiteKernel(parameter, inputs, outputs), tile_data0(NULL), tile_data1(NULL), tile_data2(NULL) {
+    switch (type()) {
+      case PrimitiveType_MulGrad:
+        arithmetic_grad_ = &ArithmeticGradCPUKernel::ArithmeticGradMul;  // this will be adjusted in InferShape
+        break;
+      case PrimitiveType_AddGrad:
+        arithmetic_grad_ = &ArithmeticGradCPUKernel::ArithmeticGradAdd;
+        break;
+      case PrimitiveType_SubGrad:
+        arithmetic_grad_ = &ArithmeticGradCPUKernel::ArithmeticGradSub;
+        break;
+      case PrimitiveType_DivGrad:
+        arithmetic_grad_ = &ArithmeticGradCPUKernel::ArithmeticGradDiv;  // this will be adjusted in InferShape
+        break;
+      default:
+        MS_LOG(ERROR) << "Error Operator type " << parameter->type_;
+        break;
+    }
+    arithmeticParameter_ = reinterpret_cast<ArithmeticParameter *>(parameter);
+  }
+  ~ArithmeticGradCPUKernel() override {
+    if (tile_data0) delete[] tile_data0;
+    if (tile_data1) delete[] tile_data1;
+    if (tile_data2) delete[] tile_data2;
+  }
+  void InitKernel(const CNodePtr &kernel_node);
+
+  int Init() override;
+  int InferShape();
+  int ReSize() override;
+  int Run() override;
+
+ private:
+  void ArithmeticGradAdd(float *dy, int dy_size, float *dx1, int dx1_size, float *dx2, int dx2_size);
+  void ArithmeticGradSub(float *dy, int dy_size, float *dx1, int dx1_size, float *dx2, int dx2_size);
+  void ArithmeticGradMul(float *dy, int dy_size, float *dx1, int dx1_size, float *dx2, int dx2_size);
+  void ArithmeticGradMul1L(float *dy, int dy_size, float *dx1, int dx1_size, float *dx2, int dx2_size);
+  void ArithmeticGradMul2L(float *dy, int dy_size, float *dx1, int dx1_size, float *dx2, int dx2_size);
+  void ArithmeticGradDiv(float *dy, int dy_size, float *dx1, int dx1_size, float *dx2, int dx2_size);
+  void ArithmeticGradDiv1L(float *dy, int dy_size, float *dx1, int dx1_size, float *dx2, int dx2_size);
+  void ArithmeticGradDiv2L(float *dy, int dy_size, float *dx1, int dx1_size, float *dx2, int dx2_size);
+  ArithmeticParameter *arithmeticParameter_;
+  ArithmeticGradOperation arithmetic_grad_;
+  float *tile_data0;
+  float *tile_data1;
+  float *tile_data2;
+};
+}  // namespace mindspore::kernel
+
+#endif  // MINDSPORE_LITE_SRC_RUNTIME_KERNEL_ARM_FP32_ARITHMETIC_GRAD_H_

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

@@ -0,0 +1,115 @@
+/**
+ * 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 <vector>
+#include "src/runtime/kernel/arm/fp32/bias_grad.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::schema::PrimitiveType_BiasGrad;
+using mindspore::lite::RET_ERROR;
+using mindspore::lite::RET_OK;
+
+namespace mindspore::kernel {
+int BiasGradCPUKernel::InferShape() {
+  if (1 != this->inputs_.size()) {
+    MS_LOG(ERROR) << "BiasGrad should have one input";
+    return RET_ERROR;
+  }
+  if (1 != this->outputs_.size()) {
+    MS_LOG(ERROR) << "BiasGrad should have one output";
+    return RET_ERROR;
+  }
+  auto *in0 = inputs_.front();
+  auto *out = outputs_.front();
+  MS_ASSERT(in0 != nullptr);
+  MS_ASSERT(out != nullptr);
+  auto inshape = in0->shape();
+  int ndim = inshape.size();
+  for (int i = 0; i < ndim - 1; i++) {
+    inshape[i] = 1;
+  }
+  out->set_shape(inshape);
+  out->set_data_type(in0->data_type());
+  return RET_OK;
+}
+
+int BiasGradCPUKernel::Init() {
+  MS_ASSERT(InferShape() == RET_OK);
+
+  auto dims = inputs_[0]->shape();
+  bias_param->ndim_ = dims.size();
+  for (unsigned int i = 0; i < bias_param->ndim_; i++) {
+    bias_param->in_shape0_[i] = dims[i];
+    bias_param->out_shape_[i] = 1;  // 1 dimension for N,H,W,
+  }
+  bias_param->out_shape_[bias_param->ndim_ - 1] = dims[bias_param->ndim_ - 1];
+  for (int i = bias_param->ndim_; i < 4; i++) {
+    bias_param->in_shape0_[i] = 0;
+    bias_param->out_shape_[i] = 0;
+  }
+  return RET_OK;
+}
+
+
+int BiasGradCPUKernel::ReSize() { return 0; }
+
+int BiasGradCPUKernel::Run() {
+  auto in = reinterpret_cast<float *>(inputs_.at(0)->Data());
+  auto out = reinterpret_cast<float *>(outputs_.at(0)->Data());
+  // size_t data_size = inputs_.at(0)->ElementsNum();
+
+  size_t nhw_size = 1;
+  size_t channels = bias_param->in_shape0_[bias_param->ndim_ - 1];  // C in NHWC
+  for (unsigned int i = 0; i < bias_param->ndim_ - 1; i++) nhw_size *= bias_param->in_shape0_[i];
+
+  size_t total_size = channels * nhw_size;
+  for (size_t c = 0; c < channels; ++c) {
+    out[c] = 0;
+    for (size_t offset = 0; offset < total_size; offset += channels) {
+      out[c] += in[offset + c];
+    }
+  }
+
+  return RET_OK;
+}
+
+
+kernel::LiteKernel *CpuBiasGradFp32KernelCreator(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_BiasGrad);
+  auto *kernel = new  (std::nothrow) BiasGradCPUKernel(reinterpret_cast<OpParameter *>(opParameter), inputs, outputs);
+  MS_ASSERT(kernel != nullptr);
+
+  auto ret = kernel->Init();
+  if (RET_OK != ret) {
+    MS_LOG(ERROR) << "Init kernel failed, name: " << opParameter->name_ << ", type: "
+                  << schema::EnumNamePrimitiveType(static_cast<schema::PrimitiveType>(opParameter->type_));
+    delete kernel;
+    return nullptr;
+  }
+  return kernel;
+}
+
+REG_KERNEL(kCPU, kNumberTypeFloat32, PrimitiveType_BiasGrad, CpuBiasGradFp32KernelCreator)
+}  // namespace mindspore::kernel

mindspore/lite/src/runtime/kernel/arm/fp32/bias_grad.h

@@ -0,0 +1,46 @@
+/**
+ * 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_FP32_BIAS_GRAD_H_
+#define MINDSPORE_LITE_SRC_RUNTIME_KERNEL_ARM_FP32_BIAS_GRAD_H_
+
+#include <vector>
+#include "src/lite_kernel.h"
+#include "ir/anf.h"
+
+#include "src/runtime/kernel/arm/opclib/fp32/arithmetic.h"
+
+namespace mindspore::kernel {
+class BiasGradCPUKernel : public LiteKernel {
+ public:
+  explicit BiasGradCPUKernel(OpParameter *parameter, const std::vector<lite::tensor::Tensor *> &inputs,
+                             const std::vector<lite::tensor::Tensor *> &outputs)
+      : LiteKernel(parameter, inputs, outputs) {
+    bias_param = reinterpret_cast<ArithmeticParameter *>(parameter);
+  }
+  ~BiasGradCPUKernel() override = default;
+
+  int Init() override;
+  int InferShape();
+  int ReSize() override;
+  int Run() override;
+
+ private:
+  ArithmeticParameter *bias_param;
+};
+}  // namespace mindspore::kernel
+
+#endif  // MINDSPORE_LITE_SRC_RUNTIME_KERNEL_ARM_FP32_BIAS_GRAD_H_

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

@@ -0,0 +1,115 @@
+/**
+ * Copyright 2019 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 <algorithm>
+#include <vector>
+#include "schema/model_generated.h"
+#include "src/kernel_factory.h"
+#include "src/runtime/kernel/arm/fp32/bngrad_input.h"
+#include "src/runtime//kernel/arm/opclib/batch_norm.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;
+// using mindspore::lite::REG_OP;
+using mindspore::schema::PrimitiveType_BNGradInput;
+
+namespace mindspore::kernel {
+int BNGradInputCPUKernel::Init() {
+  auto bn_param = reinterpret_cast<bnParameter *>(opParameter);
+  workspace_size = 5 * bn_param->channels;
+  workspace = new float[workspace_size];
+
+  if (2 != this->inputs_.size()) {
+    MS_LOG(ERROR) << "Conv2d Grad should has 2 inputs";
+    return RET_ERROR;
+  }
+  if (1 != this->outputs_.size()) {
+    MS_LOG(ERROR) << "Conv2d Grad should has one output";
+    return RET_ERROR;
+  }
+  auto *input_tensor = inputs_.at(0);
+  // auto *weight_tensor = inputs_.at(1);
+  auto *out_tensor = outputs_.at(0);
+  auto in_shape = input_tensor->shape();
+  out_tensor->set_shape(in_shape);
+  out_tensor->set_data_type(input_tensor->data_type());
+  return RET_OK;
+}
+
+int BNGradInputCPUKernel::ReSize() { return RET_OK; }
+
+/*
+according to https://wiseodd.github.io/techblog/2016/07/04/batchnorm
+*/
+
+int BNGradInputCPUKernel::Run() {
+  // std::cout << "run succ" << std::endl;
+  auto *input_x = inputs_.at(0);
+  auto *input_yt = inputs_.at(1);
+  auto *input_scale = inputs_.at(2);
+  auto *output_grad = outputs_.at(0);
+  // Tensor *bias = input[5];
+  auto bn_param = reinterpret_cast<bnParameter *>(opParameter);
+  int batch = bn_param->batch;
+  int channels = bn_param->channels;
+  int spatial = bn_param->spatial;
+  float eps = bn_param->eps;
+  std::fill(workspace, workspace + workspace_size, 0.f);
+
+  float *mean = workspace;
+  float *variance = mean + channels;
+  float *mean_delta = variance + channels;
+  float *variance_delta = mean_delta + channels;
+  float *mean_add_delta = variance_delta + channels;
+
+  float *x = reinterpret_cast<float *>(input_x->Data());
+  float *yt = reinterpret_cast<float *>(input_yt->Data());
+  float *scale = reinterpret_cast<float *>(input_scale->Data());
+  float *out = reinterpret_cast<float *>(output_grad->Data());
+
+  std::copy(yt, yt + batch * channels * spatial, out);
+  meanVar(x, batch, spatial, channels, mean, variance);
+  scaleBias(scale, batch, channels, spatial, out);
+  meanDelta(out, spatial, channels, eps, variance, mean_delta);
+  varianceDelta(x, out, mean, variance, batch, channels, spatial, eps, variance_delta);
+  meanAdd(x, mean, variance_delta, batch, channels, spatial, mean_add_delta, mean_delta);
+  NormalizeDelta(x, mean, variance, mean_delta, variance_delta, batch, channels, eps, spatial, out);
+  return RET_OK;
+}
+
+kernel::LiteKernel *CpuBNGradInputFp32KernelCreator(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_BNGradInput);
+  //  parameter->name = opDef.name()->str().data();
+  //  parameter->type = opDef.attr_type();
+  auto *kernel = new (std::nothrow) BNGradInputCPUKernel(opParameter, inputs, outputs);
+  MS_ASSERT(kernel != nullptr);
+  auto ret = kernel->Init();
+  if (RET_OK != ret) {
+    MS_LOG(ERROR) << "Init kernel failed, name: " << opParameter->name_ << ", type: "
+                  << schema::EnumNamePrimitiveType(static_cast<schema::PrimitiveType>(opParameter->type_));
+  }
+  return kernel;
+}
+
+REG_KERNEL(kCPU, kNumberTypeFloat32, PrimitiveType_BNGradInput, CpuBNGradInputFp32KernelCreator)
+}  // namespace mindspore::kernel

mindspore/lite/src/runtime/kernel/arm/fp32/bngrad_input.h

@@ -0,0 +1,41 @@
+/**
+ * Copyright 2019 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_BNGRAD_INPUT_H_
+#define MINDSPORE_LITE_SRC_RUNTIME_KERNEL_ARM_FP32_BNGRAD_INPUT_H_
+
+#include <vector>
+#include "src/lite_kernel.h"
+#include "ir/anf.h"
+
+namespace mindspore::kernel {
+class BNGradInputCPUKernel : public LiteKernel {
+ public:
+  explicit BNGradInputCPUKernel(OpParameter *parameter, const std::vector<lite::tensor::Tensor *> &inputs,
+                                const std::vector<lite::tensor::Tensor *> &outputs)
+      : LiteKernel(parameter, inputs, outputs) {}
+  ~BNGradInputCPUKernel() override { delete workspace; }
+
+  int Init() override;
+  int ReSize() override;
+  int Run() override;
+
+ private:
+  float *workspace;
+  int workspace_size;
+};
+}  // namespace mindspore::kernel
+#endif  // MINDSPORE_LITE_SRC_RUNTIME_KERNEL_ARM_FP32_BNGRAD_INPUT_H_

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

@@ -0,0 +1,156 @@
+/**
+ * Copyright 2019 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/fp32/convolution_grad_filter.h"
+#include "src/kernel_registry.h"
+#include "src/runtime/kernel/arm/opclib/pack.h"
+#include "src/runtime/kernel/arm/opclib/pack_ext.h"
+#include "src/runtime/kernel/arm/opclib/fp32/gemm.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;
+using mindspore::schema::PrimitiveType_Conv2DGradFilter;
+
+namespace mindspore::kernel {
+int ConvolutionGradFilterCPUKernel::Init() {
+  // dy is in input 0
+  // x is in input 1
+  // dw is output 0
+
+  if (2 != this->inputs_.size()) {
+    MS_LOG(ERROR) << "Conv2d Grad should has 2 inputs";
+    return RET_ERROR;
+  }
+  if (1 != this->outputs_.size()) {
+    MS_LOG(ERROR) << "Conv2d Grad should has one output";
+    return RET_ERROR;
+  }
+
+  auto *input_tensor = inputs_.at(1);
+  MS_ASSERT(input_tensor != nullptr);
+  auto *dy = inputs_.at(0);
+  MS_ASSERT(dy != nullptr);
+  auto *weight_tensor = outputs_.at(0);
+  MS_ASSERT(weight_tensor != nullptr);
+
+  auto conv_param = reinterpret_cast<ConvParameter *>(opParameter);
+  conv_param->output_batch_ = this->inputs_.at(0)->shape().at(kNHWC_N);
+  conv_param->input_batch_ = this->inputs_.at(1)->shape().at(kNHWC_N);
+  conv_param->input_h_ = this->inputs_.at(1)->shape().at(kNHWC_H);
+  conv_param->input_w_ = this->inputs_.at(1)->shape().at(kNHWC_W);
+  // assume OutCh|kh|kw|In
+  conv_param->input_channel_ = this->inputs_.at(1)->shape().at(kNHWC_C);
+  conv_param->output_channel_ = this->outputs_.at(0)->shape().at(kNHWC_N);
+
+  int ws_size = conv_param->output_h_ * conv_param->output_w_ * conv_param->kernel_h_ * conv_param->kernel_w_ *
+                conv_param->input_channel_ / conv_param->group_;
+
+  workspace = new float[ws_size];
+
+  int output_w = 0;
+  int output_h = 0;
+  output_h = dy->shape()[kNHWC_H];
+  output_w = dy->shape()[kNHWC_W];
+
+  std::vector<int> out_shape(4);
+  out_shape.at(0) = conv_param->output_channel_;
+  out_shape.at(1) = conv_param->kernel_h_;
+  out_shape.at(2) = conv_param->kernel_w_;
+  out_shape.at(3) = conv_param->input_channel_ / conv_param->group_;
+
+  // weight is output
+  weight_tensor->set_shape(out_shape);
+  weight_tensor->set_data_type(input_tensor->data_type());
+
+  conv_param->output_h_ = output_h;
+  conv_param->output_w_ = output_w;
+
+  return RET_OK;
+}
+
+int ConvolutionGradFilterCPUKernel::ReSize() { return 0; }
+
+int ConvolutionGradFilterCPUKernel::Run() {
+  auto conv_param = reinterpret_cast<ConvParameter *>(opParameter);
+  auto *input_dy = inputs_.at(0);
+  auto *input_x = inputs_.at(1);
+  auto *out_dw = outputs_.at(0);
+
+  auto x_addr = reinterpret_cast<float *>(input_x->Data());
+  auto dy_addr = reinterpret_cast<float *>(input_dy->Data());
+  auto dw_addr = reinterpret_cast<float *>(out_dw->Data());
+
+  int i, j;
+  int nweights = out_dw->ElementsNum();
+  int in_ch = conv_param->input_channel_;
+  int in_h = conv_param->input_h_;
+  int in_w = conv_param->input_w_;
+  int k_h = conv_param->kernel_h_;  // out_dw->shape()[1];
+  int k_w = conv_param->kernel_w_;  // out_dw->shape()[2];
+  int batch = conv_param->output_batch_;
+  int out_ch = conv_param->output_channel_;
+  int groups = conv_param->group_;
+  int out_h = conv_param->output_h_;
+  int out_w = conv_param->output_w_;
+
+  int m = out_h * out_w;
+  int n = k_h * k_w * in_ch / groups;
+  int k = out_ch / groups;
+
+  // zero out pointer
+  memset(dw_addr, 0, out_dw->Size());
+
+  for (i = 0; i < batch; ++i) {
+    for (j = 0; j < groups; ++j) {
+      float *mat_a = dy_addr + (i * groups) * m * k + j * (out_ch / groups);
+      float *mat_b = workspace;
+      float *mat_c = dw_addr + j * nweights / groups;
+      float *im = x_addr + (i * groups) * (in_ch / groups) * in_h * in_w + j * (in_ch / groups);
+
+      im2row_hwc(im, mat_b, conv_param);
+      gemm(1, 1, k, n, m, 1, mat_a, out_ch, mat_b, m, 1, mat_c, n);
+    }
+  }
+
+  // std::cout << "run succ" << std::endl;
+  return RET_OK;
+}
+
+kernel::LiteKernel *CpuConvGradFilterFp32KernelCreator(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_Conv2DGradFilter);
+
+  auto *kernel = new (std::nothrow) ConvolutionGradFilterCPUKernel(opParameter, inputs, outputs);
+  MS_ASSERT(kernel != nullptr);
+
+  auto ret = kernel->Init();
+  if (RET_OK != ret) {
+    MS_LOG(ERROR) << "Init kernel failed, name: " << opParameter->name_ << ", type: "
+                  << schema::EnumNamePrimitiveType(static_cast<schema::PrimitiveType>(opParameter->type_));
+    delete kernel;
+    return nullptr;
+  }
+  return kernel;
+}
+
+REG_KERNEL(kCPU, kNumberTypeFloat32, PrimitiveType_Conv2DGradFilter, CpuConvGradFilterFp32KernelCreator)
+}  // namespace mindspore::kernel

mindspore/lite/src/runtime/kernel/arm/fp32/convolution_grad_filter.h

@@ -0,0 +1,41 @@
+/**
+ * Copyright 2019 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_GRAD_FILTER_H_
+#define MINDSPORE_LITE_SRC_RUNTIME_KERNEL_ARM_FP32_CONVOLUTION_GRAD_FILTER_H_
+
+#include <vector>
+#include "src/lite_kernel.h"
+#include "ir/anf.h"
+
+namespace mindspore::kernel {
+class ConvolutionGradFilterCPUKernel : public LiteKernel {
+ public:
+  explicit ConvolutionGradFilterCPUKernel(OpParameter *parameter, const std::vector<lite::tensor::Tensor *> &inputs,
+                                          const std::vector<lite::tensor::Tensor *> &outputs)
+      : LiteKernel(parameter, inputs, outputs) {}
+  ~ConvolutionGradFilterCPUKernel() override { delete workspace; }
+
+  int Init() override;
+  int ReSize() override;
+  int Run() override;
+
+ private:
+  float *workspace;
+};
+}  // namespace mindspore::kernel
+
+#endif  // MINDSPORE_LITE_SRC_RUNTIME_KERNEL_ARM_FP32_CONVOLUTION_GRAD_FILTER_H_

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

@@ -0,0 +1,136 @@
+/**
+ * Copyright 2019 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/fp32/convolution_grad_input.h"
+#include "src/kernel_registry.h"
+#include "src/runtime/kernel/arm/opclib/pack.h"
+#include "src/runtime/kernel/arm/opclib/pack_ext.h"
+#include "src/runtime/kernel/arm/opclib/fp32/gemm.h"
+#include "include/errorcode.h"
+
+using mindspore::kernel::KERNEL_ARCH::kCPU;
+using mindspore::lite::KernelRegistrar;
+using mindspore::schema::PrimitiveType_Conv2DGradInput;
+using mindspore::lite::RET_ERROR;
+using mindspore::lite::RET_OK;
+
+namespace mindspore::kernel {
+int ConvolutionGradInputCPUKernel::Init() {
+  if (2 != this->inputs_.size()) {
+    MS_LOG(ERROR) << "Conv2d Grad should has 2 inputs";
+    return RET_ERROR;
+  }
+  if (1 != this->outputs_.size()) {
+    MS_LOG(ERROR) << "Conv2d Grad should has one output";
+    return RET_ERROR;
+  }
+
+  auto *dy_tensor = inputs_.at(kInputIndex);
+  MS_ASSERT(dy_tensor != nullptr);
+  auto *weight_tensor = inputs_.at(kWeightIndex);
+  MS_ASSERT(weight_tensor != nullptr);
+  auto *dx_tensor = outputs_.at(kOutputIndex);
+  MS_ASSERT(dx_tensor != nullptr);
+
+  auto conv_param = reinterpret_cast<ConvParameter *>(opParameter);
+  conv_param->output_batch_ = dx_tensor->shape()[(kNHWC_N)];
+  conv_param->input_batch_ = dy_tensor->shape()[(kNHWC_N)];
+
+  conv_param->input_h_ = dx_tensor->shape()[(kNHWC_H)];
+  conv_param->input_w_ = dx_tensor->shape()[(kNHWC_W)];
+
+  // assume OutCh|kh|kw|In
+  conv_param->input_channel_ = dx_tensor->shape()[(kNHWC_C)];
+  conv_param->output_channel_ = weight_tensor->shape()[(kNHWC_N)];
+
+  // TBD
+  conv_param->output_h_ = dy_tensor->shape()[kNHWC_H];
+  conv_param->output_w_ = dy_tensor->shape()[kNHWC_W];
+
+  int ws_size = conv_param->output_h_ * conv_param->output_w_ * conv_param->kernel_h_ * conv_param->kernel_w_ *
+                conv_param->input_channel_ / conv_param->group_;
+
+  workspace = new float[ws_size];
+  return 0;
+}
+
+int ConvolutionGradInputCPUKernel::ReSize() { return 0; }
+
+int ConvolutionGradInputCPUKernel::Run() {
+  auto conv_param = reinterpret_cast<ConvParameter *>(opParameter);
+  auto *input_dy = inputs_.at(0);
+  auto *input_w = inputs_.at(1);
+  auto *out_dx = outputs_.at(0);
+
+  auto dy_addr = reinterpret_cast<float *>(input_dy->Data());
+  auto w_addr = reinterpret_cast<float *>(input_w->Data());
+  auto dx_addr = reinterpret_cast<float *>(out_dx->Data());
+
+  int i, j;
+  int nweights = input_w->ElementsNum();
+  int in_ch = conv_param->input_channel_;
+  int in_h = conv_param->input_h_;
+  int in_w = conv_param->input_w_;
+  int k_h = conv_param->kernel_h_;  // out_dw->shape()[1];
+  int k_w = conv_param->kernel_w_;  // out_dw->shape()[2];
+  int batch = conv_param->output_batch_;
+  int out_ch = conv_param->output_channel_;
+  int groups = conv_param->group_;
+  int out_h = conv_param->output_h_;
+  int out_w = conv_param->output_w_;
+
+  int m = out_h * out_w;
+  int n = k_w * k_h * in_ch / groups;
+  int k = out_ch / groups;
+
+  memset(dx_addr, 0, sizeof(float) * batch * in_ch * in_h * in_w);
+
+  for (i = 0; i < batch; ++i) {
+    for (j = 0; j < groups; ++j) {
+      float *mat_a = dy_addr + (i * groups) * m * k + j * (out_ch / groups);
+      float *mat_b = w_addr + j * nweights / groups;
+      float *mat_c = workspace;
+      gemm(0, 0, m, n, k, 1, mat_a, out_ch, mat_b, n, 0, mat_c, n);
+      col2im_hwc(mat_c, dx_addr + (i * groups) * (in_ch / groups) * in_h * in_w + j * (in_ch / groups), conv_param);
+    }
+  }
+
+  // std::cout << "run succ" << std::endl;
+  return 0;
+}
+
+kernel::LiteKernel *CpuConvGradInputFp32KernelCreator(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_Conv2DGradInput);
+
+  auto *kernel = new (std::nothrow) ConvolutionGradInputCPUKernel(opParameter, inputs, outputs);
+  MS_ASSERT(kernel != nullptr);
+
+  auto ret = kernel->Init();
+  if (0 != ret) {
+    MS_LOG(ERROR) << "Init kernel failed, name: " << opParameter->name_ << ", type: "
+                  << schema::EnumNamePrimitiveType(static_cast<schema::PrimitiveType>(opParameter->type_));
+    delete kernel;
+    return nullptr;
+  }
+  return kernel;
+}
+
+REG_KERNEL(kCPU, kNumberTypeFloat32, PrimitiveType_Conv2DGradInput, CpuConvGradInputFp32KernelCreator)
+}  // namespace mindspore::kernel

mindspore/lite/src/runtime/kernel/arm/fp32/convolution_grad_input.h

@@ -0,0 +1,41 @@
+/**
+ * Copyright 2019 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_GRAD_INPUT_H_
+#define MINDSPORE_LITE_SRC_RUNTIME_KERNEL_ARM_FP32_CONVOLUTION_GRAD_INPUT_H_
+
+#include <vector>
+#include "src/lite_kernel.h"
+#include "ir/anf.h"
+
+namespace mindspore::kernel {
+class ConvolutionGradInputCPUKernel : public LiteKernel {
+ public:
+  explicit ConvolutionGradInputCPUKernel(OpParameter *parameter, const std::vector<lite::tensor::Tensor *> &inputs,
+                                         const std::vector<lite::tensor::Tensor *> &outputs)
+      : LiteKernel(parameter, inputs, outputs) {}
+  ~ConvolutionGradInputCPUKernel() override { delete workspace; }
+
+  int Init() override;
+  int ReSize() override;
+  int Run() override;
+
+ private:
+  float *workspace;
+};
+}  // namespace mindspore::kernel
+
+#endif  // MINDSPORE_LITE_SRC_RUNTIME_KERNEL_ARM_FP32_CONVOLUTION_GRAD_INPUT_H_

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

@@ -0,0 +1,78 @@
+
+/**
+ * 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 "schema/model_generated.h"
+#include "src/kernel_registry.h"
+#include "src/runtime/kernel/arm/fp32/opt_momentum.h"
+#include "include/errorcode.h"
+
+using mindspore::kernel::KERNEL_ARCH::kCPU;
+using mindspore::lite::KernelRegistrar;
+using mindspore::schema::PrimitiveType_OptMomentum;
+using mindspore::lite::RET_ERROR;
+using mindspore::lite::RET_OK;
+
+namespace mindspore::kernel {
+
+int OptMomentumCPUKernel::ReSize() { return 0; }
+
+int OptMomentumCPUKernel::Run() {
+  if (inputs_.size() != 5 || !outputs_.empty()) {
+    MS_LOG(ERROR) << "OptMomentumCPUKernel error input output size!";
+    return RET_ERROR;
+  }
+
+  if (inputs_[0]->ElementsNum() != inputs_[1]->ElementsNum() ||
+      inputs_[0]->ElementsNum() != inputs_[3]->ElementsNum()) {
+    MS_LOG(ERROR) << "error input data size!";
+    return RET_ERROR;
+  }
+  auto weight = reinterpret_cast<float *>(inputs_[0]->Data());
+  auto accumulate = reinterpret_cast<float *>(inputs_[1]->Data());
+  float learning_rate = reinterpret_cast<float *>(inputs_[2]->Data())[0];
+  auto gradient = reinterpret_cast<float *>(inputs_[3]->Data());
+  float moment = reinterpret_cast<float *>(inputs_[4]->Data())[0];
+  size_t elem_num = inputs_[0]->ElementsNum();
+  for (size_t i = 0; i < elem_num; ++i) {
+    accumulate[i] = accumulate[i] * moment + gradient[i];
+    weight[i] -= accumulate[i] * learning_rate;
+  }
+  return RET_OK;
+}
+
+int OptMomentumCPUKernel::Init() { return 0; }
+
+kernel::LiteKernel *CpuOptMomentumFp32KernelCreator(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(desc.type == schema::PrimitiveType_OptMomentum);
+  auto *kernel = new (std::nothrow) OptMomentumCPUKernel(opParameter, inputs, outputs);
+  MS_ASSERT(kernel != nullptr);
+
+  auto ret = kernel->Init();
+  if (0 != ret) {
+    MS_LOG(ERROR) << "Init kernel failed, name: " << opParameter->name_ << ", type: "
+                  << schema::EnumNamePrimitiveType(static_cast<schema::PrimitiveType>(opParameter->type_));
+    delete kernel;
+    return nullptr;
+  }
+  return kernel;
+}
+
+REG_KERNEL(kCPU, kNumberTypeFloat32, PrimitiveType_OptMomentum, CpuOptMomentumFp32KernelCreator)
+}  // namespace mindspore::kernel

mindspore/lite/src/runtime/kernel/arm/fp32/opt_momentum.h

@@ -0,0 +1,40 @@
+/**
+ * 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_BACKEND_ARM_FP32_OPT_MOMENTUM_H_
+#define MINDSPORE_LITE_SRC_RUNTIME_KERNEL_ARM_OPT_MOMENTUM_H_
+
+#include <vector>
+#include "src/lite_kernel.h"
+#include "ir/anf.h"
+
+namespace mindspore::kernel {
+class OptMomentumCPUKernel : public LiteKernel {
+ public:
+  explicit OptMomentumCPUKernel(OpParameter *parameter, const std::vector<lite::tensor::Tensor *> &inputs,
+                                   const std::vector<lite::tensor::Tensor *> &outputs)
+      : LiteKernel(parameter, inputs, outputs) {}
+  ~OptMomentumCPUKernel() override {}
+
+  int Init() override;
+  int ReSize() override;
+  int Run() override;
+
+ private:
+};
+}  // namespace mindspore::kernel
+
+#endif  // MINDSPORE_LITE_SRC_RUNTIME_KERNEL_ARM_OPT_MOMENTUM_H_

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

@@ -0,0 +1,195 @@
+/**
+ * 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/fp32/pooling_grad.h"
+#include "schema/model_generated.h"
+#include "src/kernel_registry.h"
+#include "src/runtime/kernel/arm/opclib/fp32/pooling.h"
+#include "src/runtime/kernel/arm/opclib/fp32/pooling_grad.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;
+using mindspore::schema::PrimitiveType_PoolingGrad;
+
+namespace mindspore::kernel {
+#if 0
+int PoolingGradCPUKernel::TfPadding(int input_w, int input_h, int &output_w, int &output_h) {
+  PoolingParameter *pool_param = reinterpret_cast<PoolingParameter *> (opParameter);
+
+  auto stride_w = pool_param->stride_w_;
+  auto stride_h = pool_param->stride_h_;
+  auto window_w = pool_param->window_w_;
+  auto window_h = pool_param->window_h_;
+  auto pad_up = pool_param->pad_u_;
+  auto pad_down = pool_param->pad_d_;
+  auto pad_left = pool_param->pad_l_;
+  auto pad_right = pool_param->pad_r_;
+  if (pool_param->pad_mode_ == PADMODE_SAME) {
+    output_w = ceil(input_w / stride_w);
+    output_h = ceil(input_h / stride_h);
+  } else {
+    output_w = ceil((input_w + pad_left + pad_right - window_w + 1) / stride_w);
+    output_h = ceil((input_h + pad_up + pad_down - window_h + 1) / stride_h);
+  }
+  return RET_OK;
+}
+
+int PoolingGradCPUKernel::CaffePadding(int input_w, int input_h, int &output_w, int &output_h) {
+  PoolingParameter *pool_param = reinterpret_cast<PoolingParameter *> (opParameter);
+
+  auto round_mode = pool_param->round_mode_;
+  auto stride_w = pool_param->stride_w_;
+  auto stride_h = pool_param->stride_h_;
+  auto window_w = pool_param->window_w_;
+  auto window_h = pool_param->window_h_;
+  auto pad_up = pool_param->pad_u_;
+  auto pad_down = pool_param->pad_d_;
+  auto pad_left = pool_param->pad_l_;
+  auto pad_right = pool_param->pad_r_;
+  if (round_mode == ROUNDMODE_FLOOR && false) {
+    output_w = floor((input_w + pad_left + pad_right - window_w) / stride_w + 1);
+    output_h = floor((input_h + pad_up + pad_down - window_h) / stride_h + 1);
+  } else if (round_mode == ROUNDMODE_CEIL || true) {
+    output_w = ceil((input_w + pad_left + pad_right - window_w) / stride_w + 1);
+    output_h = ceil((input_h + pad_up + pad_down - window_h) / stride_h + 1);
+  } else {
+    MS_LOG(ERROR) << "round mode not support.";
+  }
+
+  if (pad_left > 0 || pad_up > 0) {
+    if ((output_w - 1) * stride_w >= input_w + pad_left) {
+      --output_w;
+    }
+    if ((output_h - 1) * stride_h >= input_h + pad_up) {
+      --output_h;
+    }
+  }
+  return RET_OK;
+}
+
+int PoolingGradCPUKernel::OnnxPadding(int input_w, int input_h, int &output_w, int &output_h) {
+  PoolingParameter *pool_param = reinterpret_cast<PoolingParameter *> (opParameter);
+
+  auto round_mode = pool_param->round_mode_;
+  auto stride_w = pool_param->stride_w_;
+  auto stride_h = pool_param->stride_h_;
+  auto window_w = pool_param->window_w_;
+  auto window_h = pool_param->window_h_;
+  auto pad_up = pool_param->pad_u_;
+  auto pad_down = pool_param->pad_d_;
+  auto pad_left = pool_param->pad_l_;
+  auto pad_right = pool_param->pad_r_;
+  if (round_mode == ROUNDMODE_FLOOR) {
+    output_w = floor((input_w + pad_left + pad_right - window_w) / stride_w + 1);
+    output_h = floor((input_h + pad_up + pad_down - window_h) / stride_h + 1);
+  } else if (round_mode == ROUNDMODE_CEIL) {
+    MS_LOG(ERROR) << "RoundMode_CEIL mode not support.";
+  } else {
+    MS_LOG(ERROR) << "OnnxPadding round mode not support.";
+  }
+  return RET_OK;
+}
+#endif
+
+int PoolingGradCPUKernel::Init() {
+  // InferShape():
+  // auto *in_tensor = reinterpret_cast<float *>(inputs_.at(0)->Data());
+  // auto *x_tensor = reinterpret_cast<float *>(inputs_.at(1)->Data());
+
+  PoolingParameter *pool_param = reinterpret_cast<PoolingParameter *>(opParameter);
+
+  auto in_shape = inputs_.at(0)->shape();
+  int input_h = in_shape.at(1);
+  int input_w = in_shape.at(2);
+
+  if (pool_param->global_) {
+    pool_param->window_w_ = input_w;
+    pool_param->window_h_ = input_h;
+  }
+
+  // Emir -- here I assume we get the outputshape in the output tensor
+  auto *out_tensor = outputs_.front();
+  auto out_shape = out_tensor->shape();
+
+#if 0
+  int output_w = 0, output_h = 0;
+  auto fmk_type = pool_param->fmk_type_;
+  switch (fmk_type) {
+    case lite::FmkType_TF:
+      break;
+    case lite::FmkType_CAFFE:
+      CaffePadding(input_w, input_h, output_w, output_h);
+      break;
+    case lite::FmkType_ONNX:
+      OnnxPadding(input_w, input_h, output_w, output_h);
+      break;
+    case lite::FmkType_MS:
+      break;
+    case lite::FmkType_TFLITE:
+      TfPadding(input_w, input_h, output_w, output_h);
+      break;
+    default:
+      MS_LOG(ERROR) << "Not support this framework.";
+  }
+  std::vector<int> out_shape{in_tensor->shape()};
+  out_shape.at(1) = output_h;
+  out_shape.at(2) = output_w;
+#endif
+  out_tensor->set_shape(out_shape);
+  out_tensor->set_data_type(inputs_.at(0)->data_type());
+  return RET_OK;
+}
+
+int PoolingGradCPUKernel::ReSize() { return RET_OK; }
+
+int PoolingGradCPUKernel::Run() {
+  PoolingParameter *pool_param = reinterpret_cast<PoolingParameter *>(opParameter);
+  auto input_ptr = reinterpret_cast<float *>(inputs_.at(0)->Data());
+  auto output_ptr = reinterpret_cast<float *>(outputs_.at(0)->Data());
+
+  if (pool_param->max_pooling_) {
+    auto ind = reinterpret_cast<int *>(inputs_.at(1)->Data());
+    MaxPoolingGrad(input_ptr, ind, output_ptr, pool_param);
+  } else {
+    AvgPoolingGrad(input_ptr, output_ptr, pool_param);
+  }
+  return RET_OK;
+}
+
+kernel::LiteKernel *CpuPoolingGradFp32KernelCreator(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_PoolingGrad);
+
+  auto *kernel = new (std::nothrow) PoolingGradCPUKernel(opParameter, inputs, outputs);
+  MS_ASSERT(kernel != nullptr);
+  auto ret = kernel->Init();
+  if (RET_OK != ret) {
+    MS_LOG(ERROR) << "Init kernel failed, name: " << opParameter->name_ << ", type: "
+                  << schema::EnumNamePrimitiveType(static_cast<schema::PrimitiveType>(opParameter->type_));
+    delete kernel;
+    return nullptr;
+  }
+  return kernel;
+}
+
+REG_KERNEL(kCPU, kNumberTypeFloat32, PrimitiveType_PoolingGrad, CpuPoolingGradFp32KernelCreator)
+}  // namespace mindspore::kernel

mindspore/lite/src/runtime/kernel/arm/fp32/pooling_grad.h

@@ -0,0 +1,50 @@
+/**
+ * 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_FP32_POOLING_GRAD_H_
+#define MINDSPORE_LITE_SRC_RUNTIME_KERNEL_ARM_FP32_POOLING_GRAD_H_
+
+#include <vector>
+#include "src/lite_kernel.h"
+#include "ir/anf.h"
+
+namespace mindspore::kernel {
+using mindspore::schema::PadMode;
+using mindspore::schema::PoolMode;
+using mindspore::schema::QuantType;
+using mindspore::schema::RoundMode;
+
+class PoolingGradCPUKernel : public LiteKernel {
+ public:
+  explicit PoolingGradCPUKernel(OpParameter *parameter, const std::vector<lite::tensor::Tensor *> &inputs,
+                                const std::vector<lite::tensor::Tensor *> &outputs)
+      : LiteKernel(parameter, inputs, outputs) {}
+  ~PoolingGradCPUKernel() override = default;
+
+  // int TfPadding(int input_w, int input_h, int &output_w, int &output_h);
+  // int CaffePadding(int input_w, int input_h, int &output_w, int &output_h);
+  // int OnnxPadding(int input_w, int input_h, int &output_w, int &output_h);
+
+  int Init() override;
+  int ReSize() override;
+  int Run() override;
+
+ private:
+  uint8_t data_shape_{0};
+};
+}  // namespace mindspore::kernel
+
+#endif  // MINDSPORE_LITE_SRC_RUNTIME_KERNEL_ARM_FP32_POOLING_GRAD_H_

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

@@ -0,0 +1,67 @@
+/**
+ * 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/fp32/power_grad.h"
+#include "schema/model_generated.h"
+#include "src/kernel_registry.h"
+#include "include/errorcode.h"
+#include "src/runtime/kernel/arm/opclib/fp32/arithmetic.h"
+
+using mindspore::lite::KernelRegistrar;
+using mindspore::lite::RET_ERROR;
+using mindspore::lite::RET_OK;
+using mindspore::schema::PrimitiveType_PowerGrad;
+
+namespace mindspore::kernel {
+int PowerGradCPUKernel::Init() { return RET_OK; }
+
+int PowerGradCPUKernel::ReSize() { return RET_OK; }
+
+int PowerGradCPUKernel::Run() {
+  auto dy_addr = reinterpret_cast<float *>(inputs_.at(0)->Data());
+  auto x_addr = reinterpret_cast<float *>(inputs_.at(1)->Data());
+  auto dx_addr = reinterpret_cast<float *>(outputs_.at(0)->Data());
+  auto size = inputs_.at(0)->ElementsNum();
+
+  Power(x_addr, dx_addr, size, power_ - 1, scale_, shift_);
+  ElementMul(dx_addr, dy_addr, dx_addr, size);
+  float scale = scale_ * power_;
+  for (int i = 0; i < size; i++) {
+    dx_addr[i] *= scale;
+  }
+
+  return RET_OK;
+}
+
+kernel::LiteKernel *CpuPowerGradFp32KernelCreator(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_PowerGrad);
+  auto *kernel = new (std::nothrow) PowerGradCPUKernel(opParameter, inputs, outputs);
+  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_));
+    delete kernel;
+    return nullptr;
+  }
+  return kernel;
+}
+
+REG_KERNEL(kCPU, kNumberTypeFloat32, PrimitiveType_PowerGrad, CpuPowerGradFp32KernelCreator)
+}  // namespace mindspore::kernel

mindspore/lite/src/runtime/kernel/arm/fp32/power_grad.h

@@ -0,0 +1,49 @@
+/**
+ * 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_FP32_POWER_GRAD_H_
+#define MINDSPORE_LITE_SRC_RUNTIME_KERNEL_ARM_FP32_POWER_GRAD_H_
+
+#include <vector>
+#include "src/lite_kernel.h"
+#include "ir/anf.h"
+#include "src/runtime/kernel/arm/opclib/power.h"
+
+namespace mindspore::kernel {
+class PowerGradCPUKernel : public LiteKernel {
+ public:
+  PowerGradCPUKernel(OpParameter *param, const std::vector<lite::tensor::Tensor *> &inputs,
+                          const std::vector<lite::tensor::Tensor *> &outputs)
+      : LiteKernel(param, inputs, outputs) {
+        PowerParameter *power_param = reinterpret_cast<PowerParameter *>(param);
+        power_ = power_param->power_;
+        scale_ = power_param->scale_;
+        shift_ = power_param->shift_;
+  }
+  ~PowerGradCPUKernel() override = default;
+
+  int Init() override;
+  int ReSize() override;
+  int Run() override;
+
+ private:
+  float power_;
+  float scale_;
+  float shift_;
+};
+}  // namespace mindspore::kernel
+
+#endif  // MINDSPORE_LITE_SRC_RUNTIME_KERNEL_ARM_FP32_POWER_GRAD_H_

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

@@ -0,0 +1,145 @@
+/**
+ * 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/fp32/sparse_softmax_cross_entropy_with_logits.h"
+#include "src/runtime/kernel/arm/opclib/fp32/softmax.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;
+using mindspore::schema::PrimitiveType_SoftmaxCrossEntropy;
+
+namespace mindspore::kernel {
+
+int SparseSoftmaxCrossEntropyWithLogitsCPUKernel::ReSize() { return RET_OK; }
+
+void SparseSoftmaxCrossEntropyWithLogitsCPUKernel::ForwardPostExecute(const int *labels, const float *losses,
+                                                                      float *output) const {
+  float total_loss = 0;
+  for (int i = 0; i < param->batch_size_; ++i) {
+    if (labels[i] < 0) {
+      MS_LOG(EXCEPTION) << "label value must >= 0";
+    }
+    size_t label = labels[i];
+    if (label > param->number_of_classes_) {
+      MS_LOG(EXCEPTION) << "error label input!";
+    } else {
+      total_loss -= logf(losses[i * param->number_of_classes_ + label]);
+    }
+  }
+  output[0] = total_loss / param->batch_size_;
+}
+
+void SparseSoftmaxCrossEntropyWithLogitsCPUKernel::GradPostExecute(const int *labels, const float *losses,
+                                                                   float *output) const {
+  size_t row_start = 0;
+  for (int i = 0; i < param->batch_size_; ++i) {
+    if (labels[i] < 0) {
+      MS_LOG(EXCEPTION) << "label value must >= 0";
+    }
+    size_t label = labels[i];
+    if (label > param->number_of_classes_) {
+      MS_LOG(EXCEPTION) << "error label input!";
+    }
+    for (size_t j = 0; j < param->number_of_classes_; ++j) {
+      size_t index = row_start + j;
+      if (j == label) {
+        output[index] = (losses[index] - 1) / param->batch_size_;
+      } else {
+        output[index] = losses[index] / param->batch_size_;
+      }
+    }
+    row_start += param->number_of_classes_;
+  }
+}
+
+int SparseSoftmaxCrossEntropyWithLogitsCPUKernel::Run() {
+  auto ins = reinterpret_cast<float *>(inputs_.at(0)->Data());
+  auto labels = reinterpret_cast<int *>(inputs_.at(1)->Data());
+  auto out = reinterpret_cast<float *>(outputs_.at(0)->Data());
+  float *grads = NULL;
+  if (is_train()) {  // outputs_.size() > 1)
+    grads = reinterpret_cast<float *>(outputs_.at(0)->Data());
+  }
+  size_t data_size = inputs_.at(0)->ElementsNum();
+  float *losses = new (std::nothrow) float[data_size];
+  MS_ASSERT(losses != nullptr);
+  std::fill(losses, losses + data_size, 0);
+
+  MS_ASSERT(out != nullptr);
+  MS_ASSERT(labels != nullptr);
+  MS_ASSERT(ins != nullptr);
+
+  SoftmaxParameter sm_params;
+  sm_params.n_dim_ = param->n_dim_;
+  sm_params.element_size_ = data_size;
+  sm_params.axis_ = 1;
+  for (int i = 0; i < 4; i++)  // softmax has only 4 params in shape
+    sm_params.input_shape_[i] = param->input_shape_[i];
+  float sum_data[sm_params.input_shape_[sm_params.axis_]];
+  Softmax(ins, losses, sum_data, &sm_params);
+
+  if (is_train()) {
+    GradPostExecute(labels, losses, grads);
+  } else {
+    ForwardPostExecute(labels, losses, out);
+  }
+  return RET_OK;
+}
+
+int SparseSoftmaxCrossEntropyWithLogitsCPUKernel::Init() {
+  auto dims = inputs_[0]->shape();
+  param->n_dim_ = 2;
+  param->number_of_classes_ = dims[1];
+  param->batch_size_ = dims[0];
+  for (unsigned int i = 0; i < dims.size(); i++) param->input_shape_[i] = dims[i];
+  if (2 != this->inputs_.size()) {
+    MS_LOG(ERROR) << "softmax entropy loss should have two inputs";
+    return RET_ERROR;
+  }
+  auto *in0 = inputs_.front();
+  if (in0 == nullptr) {
+    MS_LOG(ERROR) << "softmax etropy loss in0 have no data";
+    return RET_ERROR;
+  }
+
+  return RET_OK;
+}
+
+kernel::LiteKernel *CpuSoftmaxCrossEntropyFp32KernelCreator(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_SoftmaxCrossEntropy);
+  auto *kernel = new (std::nothrow) SparseSoftmaxCrossEntropyWithLogitsCPUKernel(opParameter, inputs, outputs);
+  MS_ASSERT(kernel != nullptr);
+  auto ret = kernel->Init();
+  if (RET_OK != ret) {
+    MS_LOG(ERROR) << "Init kernel failed, name: " << opParameter->name_ << ", type: "
+                  << schema::EnumNamePrimitiveType(static_cast<schema::PrimitiveType>(opParameter->type_));
+    delete kernel;
+    return nullptr;
+  }
+  return kernel;
+}
+
+REG_KERNEL(kCPU, kNumberTypeFloat32, PrimitiveType_SoftmaxCrossEntropy, CpuSoftmaxCrossEntropyFp32KernelCreator)
+}  // namespace mindspore::kernel

mindspore/lite/src/runtime/kernel/arm/fp32/sparse_softmax_cross_entropy_with_logits.h

@@ -0,0 +1,50 @@
+/**
+ * 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_FP32_SPARSE_SOFTMAX_CROSS_ENTROPY_WITH_LOGITS_H_
+#define MINDSPORE_LITE_SRC_RUNTIME_KERNEL_ARM_FP32_SPARSE_SOFTMAX_CROSS_ENTROPY_WITH_LOGITS_H_
+
+#include <vector>
+#include "src/lite_kernel.h"
+#include "ir/anf.h"
+#include "src/runtime/kernel/arm/opclib/fp32/softmax_grad.h"
+#include "src/runtime/kernel/arm/opclib/fp32/arithmetic.h"
+
+namespace mindspore::kernel {
+
+class SparseSoftmaxCrossEntropyWithLogitsCPUKernel : public LiteKernel {
+ public:
+  explicit SparseSoftmaxCrossEntropyWithLogitsCPUKernel(OpParameter *parameter,
+                                                        const std::vector<lite::tensor::Tensor *> &inputs,
+                                                        const std::vector<lite::tensor::Tensor *> &outputs)
+      : LiteKernel(parameter, inputs, outputs) {
+    param = reinterpret_cast<SoftmaxCrossEntropyParameter *>(parameter);
+  }
+  ~SparseSoftmaxCrossEntropyWithLogitsCPUKernel() override = default;
+
+  void ForwardPostExecute(const int *labels, const float *losses, float *output) const;
+  void GradPostExecute(const int *labels, const float *losses, float *output) const;
+
+  int Init() override;
+  int ReSize() override;
+  int Run() override;
+
+ private:
+  SoftmaxCrossEntropyParameter *param;
+};
+}  // namespace mindspore::kernel
+
+#endif  // MINDSPORE_LITE_SRC_RUNTIME_KERNEL_ARM_FP32_SPARSE_SOFTMAX_CROSS_ENTROPY_WITH_LOGITS_H_

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

@@ -0,0 +1,88 @@
+/**
+ * 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_ACTIVATION_GRAD_H_
+#define MINDSPORE_LITE_SRC_RUNTIME_KERNEL_ARM_OPCLIB_ACTIVATION_GRAD_H_
+
+#include <math.h>
+#include "src/runtime/kernel/arm/opclib/op_base.h"
+#include "src/runtime/kernel/arm/opclib/fp32/arithmetic.h"
+#include "src/runtime/kernel/arm/opclib/errorcode.h"
+
+struct ActivationGradParameter {
+  OpParameter op_parameter{};
+  int type_;
+  float alpha_{0.01};
+};
+
+inline int ReluGrad(float *src0, float *src1, int length, float *dst) {
+  for (int i = 0; i < length; ++i) {
+    dst[i] = src1[i] > 0 ? 1.0f : 0.0f;
+  }
+  ElementMul(src0, dst, dst, length);
+  return OPCLIB_OK;
+}
+
+inline int Relu6Grad(float *src0, float *src1, int length, float *dst) {
+  for (int i = 0; i < length; ++i) {
+    if (src1[i] < 0) {
+      dst[i] = 0;
+    } else {
+      dst[i] = src1[i] > 6.0f ? 0.0f : 1.0f;
+    }
+  }
+  ElementMul(src0, dst, dst, length);
+  return OPCLIB_OK;
+}
+
+inline int LReluGrad(float *src0, float *src1, int length, float *dst, float alpha) {
+  for (int i = 0; i < length; ++i) {
+    dst[i] = src1[i] > 0.0f ? 1.0f : alpha;
+  }
+  ElementMul(src0, dst, dst, length);
+  return OPCLIB_OK;
+}
+
+inline int SigmoidGrad(float *src0, float *src1, int length, float *dst) {
+  for (int i = 0; i < length; ++i) {
+    dst[i] = src0[i] * (src1[i] * (1.0f - src1[i]));
+  }
+  return OPCLIB_OK;
+}
+
+inline int TanhGrad(float *src0, float *src1, int length, float *dst) {
+  for (int i = 0; i < length; ++i) {
+    dst[i] = (1.0f - (src1[i] * src1[i])) * src0[i];
+  }
+  return OPCLIB_OK;
+}
+
+inline int HSwishGrad(float *src0, float *src1, int length, float *dst) {
+  for (int i = 0; i < length; ++i) {
+    float tmp = (src1[i] > 3.0f ? 1.0f : (src1[i] < -3.0f ? 0.0f : (2.0f * src1[i] + 3.0f) / 6.0f));
+    dst[i] = tmp * src0[i];
+  }
+  return OPCLIB_OK;
+}
+
+inline int HSigmoidGrad(float *src0, float *src1, int length, float *dst) {
+  for (int i = 0; i < length; ++i) {
+    float tmp = (src1[i] > 3.0f ? 1.0f : (src1[i] < -3.0f ? 0.0f : 1.0f / 6.0f));
+    dst[i] = tmp * src0[i];
+  }
+  return OPCLIB_OK;
+}
+
+#endif  // MINDSPORE_LITE_SRC_RUNTIME_KERNEL_ARM_OPCLIB_ACTIVATION_GRAD_H_

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

@@ -0,0 +1,120 @@
+/**
+ * Copyright 2019 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 <algorithm>
+#include <cmath>
+#include "src/runtime/kernel/arm/opclib/batch_norm.h"
+
+static void sumSpatialBatch(const float *in, int size, int ch, float *out) {
+  std::fill(out, out + ch, 0.f);
+  for (int i = 0; i < size; i++) {
+    const float *ptr = in + i * ch;
+    for (int c = 0; c < ch; c++) {
+      out[c] += ptr[c];
+    }
+  }
+}
+
+void scaleBias(const float *scales, int batch, int n, int size, float *output) {
+  for (int i = 0; i < batch * size; i++)
+    for (int c = 0; c < n; c++) output[i * n + c] *= scales[c];
+}
+
+void normalize(const float *x, const float *mean, const float *variance, float eps, int batch, int filters, int spatial,
+               float *out) {
+  int b, f, i;
+  for (b = 0; b < batch; ++b) {
+    for (i = 0; i < spatial; ++i) {
+      for (f = 0; f < filters; ++f) {
+        int index = b * filters * spatial + i * filters + f;
+        out[index] = (x[index] - mean[f]) / (std::sqrt(variance[f]) + eps);
+      }
+    }
+  }
+}
+
+void backwardScale(const float *x_norm, const float *delta, int batch, int n, int size, float *scale_updates) {
+  int i, b, f;
+  std::fill(scale_updates, scale_updates + n, 0.f);
+  for (b = 0; b < batch; ++b) {
+    for (i = 0; i < size; ++i) {
+      for (f = 0; f < n; ++f) {
+        int index = (b * size + i) * n + f;
+        scale_updates[f] += delta[index] * x_norm[index];
+      }
+    }
+  }
+}
+
+void meanVar(const float *in, int batch, int spatial, int ch, float *mean, float *var) {
+  float N = batch * spatial;
+  sumSpatialBatch(in, N, ch, mean);
+  for (int f = 0; f < ch; ++f) mean[f] /= N;
+  std::fill(var, var + ch, 0.f);
+  for (int i = 0; i < N; i++) {
+    for (int f = 0; f < ch; f++) {
+      float x = in[i * ch + f];
+      var[f] += (x - mean[f]) * (x - mean[f]);
+    }
+  }
+  for (int f = 0; f < ch; f++) var[f] /= N;
+}
+
+void meanDelta(float *yt, int size, int ch, float eps, float *variance, float *mean_delta) {
+  sumSpatialBatch(yt, size, ch, mean_delta);
+  for (int i = 0; i < ch; i++) mean_delta[i] *= -1.f / std::sqrt((variance[i] + eps));
+}
+
+void meanAdd(const float *x, const float *mean, const float *variance_delta, int batch, int filters, int spatial,
+             float *mean_add, float *mean_delta) {
+  int i, k;
+  std::fill(mean_add, mean_add + filters, 0.f);
+  for (k = 0; k < spatial * batch; ++k) {
+    for (i = 0; i < filters; ++i) {
+      int index = k * filters + i;
+      mean_add[i] += x[index] - mean[i];
+    }
+  }
+  for (i = 0; i < filters; ++i) {
+    mean_add[i] *= variance_delta[i] * (-2.f / (spatial * batch));
+    mean_delta[i] += mean_add[i];
+  }
+}
+
+void varianceDelta(const float *x, const float *delta, const float *mean, const float *variance, int batch, int filters,
+                   int spatial, float eps, float *variance_delta) {
+  int i, k;
+  std::fill(variance_delta, variance_delta + filters, 0.f);
+  for (k = 0; k < batch * spatial; k++) {
+    for (i = 0; i < filters; i++) {
+      int index = k * filters + i;
+      variance_delta[i] += delta[index] * (x[index] - mean[i]);
+    }
+  }
+  for (i = 0; i < filters; i++) variance_delta[i] *= -.5 * pow(variance[i] + eps, (-3.f / 2.f));
+}
+
+void NormalizeDelta(const float *x, const float *mean, const float *variance, const float *mean_delta,
+                    const float *variance_delta, int batch, int filters, int spatial, float eps, float *delta) {
+  int f, k;
+  for (k = 0; k < batch * spatial; k++) {
+    for (f = 0; f < filters; f++) {
+      int index = k * filters + f;
+      delta[index] = delta[index] * 1. / (std::sqrt(variance[f] + eps)) +
+                     variance_delta[f] * 2. * (x[index] - mean[f]) / (spatial * batch) +
+                     mean_delta[f] / (spatial * batch);
+    }
+  }
+}

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

@@ -0,0 +1,39 @@
+/**
+ * Copyright 2019 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_BACKEND_ARM_BATCH_NORM_H_
+#define MINDSPORE_LITE_SRC_BACKEND_ARM_BATCH_NORM_H_
+
+struct bnParameter {
+  int batch;
+  int channels;
+  int spatial;
+  float eps;
+};
+void scaleBias(const float *scales, int batch, int n, int size, float *output);
+void normalize(const float *x, const float *mean, const float *variance, float eps, int batch, int filters, int spatial,
+               float *out);
+void backwardScale(const float *x_norm, const float *delta, int batch, int n, int size, float *scale_updates);
+void meanVar(const float *in, int batch, int size, int ch, float *mean, float *var);
+void meanDelta(float *yt, int size, int ch, float eps, float *variance, float *mean_delta);
+void varianceDelta(const float *x, const float *delta, const float *mean, const float *variance, int batch, int ch,
+                   int spatial, float eps, float *variance_delta);
+void meanAdd(const float *x, const float *mean, const float *variance_delta, int batch, int filters, int spatial,
+             float *mean_add, float *mean_delta);
+void NormalizeDelta(const float *x, const float *mean, const float *variance, const float *mean_delta,
+                    const float *variance_delta, int batch, int filters, int spatial, float eps, float *delta);
+
+#endif

mindspore/lite/src/runtime/kernel/arm/opclib/fp32/arithmetic_grad.cc

@@ -0,0 +1,29 @@
+/**
+ * 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/fp32/arithmetic_grad.h"
+
+void ElementDivNegSquare(const float *nom, const float *denom, float *output, int element_size) {
+  for (int i = 0; i < element_size; i++) {
+    output[i] = -nom[i] / (denom[i] * denom[i]);
+  }
+}
+
+void ElementMulAndDivNegSquare(const float *a, const float *b, const float *denom, float *output, int element_size) {
+  for (int i = 0; i < element_size; i++) {
+    output[i] = -a[i] * b[i] / (denom[i] * denom[i]);
+  }
+}

mindspore/lite/src/runtime/kernel/arm/opclib/fp32/arithmetic_grad.h

@@ -0,0 +1,22 @@
+/**
+ * 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_FP32_ARITHMETIC_GRAD_H_
+#define MINDSPORE_LITE_SRC_RUNTIME_KERNEL_ARM_OPCLIB_FP32_ARITHMETIC_GRAD_H_
+
+void ElementDivNegSquare(const float *nom, const float *denom, float *output, int element_size);
+void ElementMulAndDivNegSquare(const float *a, const float *b, const float *denom, float *output, int element_size);
+
+#endif  // MINDSPORE_LITE_SRC_RUNTIME_KERNEL_ARM_OPCLIB_FP32_ARITHMETIC_GRAD_H_

mindspore/lite/src/runtime/kernel/arm/opclib/fp32/gemm.cc

@@ -0,0 +1,108 @@
+/**
+ * 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/fp32/gemm.h"
+
+static void gemm_nn(int M, int N, int K, float alpha, float *mat_a, int lda, float *mat_B, int ldb, float *mat_c,
+                    int ldc) {
+  int i, j, k;
+  for (i = 0; i < M; ++i) {
+    for (k = 0; k < K; ++k) {
+      float a = alpha * mat_a[i * lda + k];
+      for (j = 0; j < N; ++j) {
+        mat_c[i * ldc + j] += a * mat_B[k * ldb + j];
+      }
+    }
+  }
+}
+
+static void gemm_nt(int M, int N, int K, float alpha, float *mat_a, int lda, float *mat_b, int ldb, float *mat_c,
+                    int ldc) {
+  int i, j, k;
+  for (i = 0; i < M; ++i) {
+    for (j = 0; j < N; ++j) {
+      float sum = 0;
+      for (k = 0; k < K; ++k) {
+        sum += alpha * mat_a[i * lda + k] * mat_b[j * ldb + k];
+      }
+      mat_c[i * ldc + j] += sum;
+    }
+  }
+}
+
+static void gemm_tn(int M, int N, int K, float alpha, float *mat_a, int lda, float *mat_b, int ldb, float *mat_c,
+                    int ldc) {
+  int i, j, k;
+  for (i = 0; i < M; ++i) {
+    for (k = 0; k < K; ++k) {
+      float a = alpha * mat_a[k * lda + i];
+      for (j = 0; j < N; ++j) {
+        mat_c[i * ldc + j] += a * mat_b[k * ldb + j];
+      }
+    }
+  }
+}
+
+static void gemm_tt(int M, int N, int K, float alpha, float *mat_a, int lda, float *mat_b, int ldb, float *mat_c,
+                    int ldc) {
+  int i, j, k;
+  for (i = 0; i < M; ++i) {
+    for (j = 0; j < N; ++j) {
+      float sum = 0;
+      for (k = 0; k < K; ++k) {
+        sum += alpha * mat_a[i + k * lda] * mat_b[k + j * ldb];
+      }
+      mat_c[i * ldc + j] += sum;
+    }
+  }
+}
+
+// mat_c = alpha*op( mat_a )*op( mat_b ) + beta*C
+// M - number of rows of matrix a
+// N - number of cols of matrix b
+// K - number of cols of matrix a
+
+void gemm(int transpose_a, int transpose_b, int M, int N, int K, float alpha, float *mat_a, int lda, float *mat_b,
+          int ldb, float beta, float *mat_c, int ldc) {
+  // printf("cpu: %d %d %d %d %d %f %d %d %f %d\n",TA, TB, M, N, K, ALPHA, lda, ldb, BETA, ldc);
+  if (beta >= 0.f && beta <= 0.f) {
+    for (int i = 0; i < M; ++i) {
+      for (int j = 0; j < N; ++j) {
+        mat_c[i * ldc + j] = 0;
+      }
+    }
+  } else if (beta < 1.f || beta > 1.f) {
+    for (int i = 0; i < M; ++i) {
+      for (int j = 0; j < N; ++j) {
+        mat_c[i * ldc + j] *= beta;
+      }
+    }
+  }
+
+  int t;
+
+  for (t = 0; t < M; ++t) {
+    if (!transpose_a && !transpose_b) {
+      gemm_nn(1, N, K, alpha, mat_a + t * lda, lda, mat_b, ldb, mat_c + t * ldc, ldc);
+    } else if (transpose_a && !transpose_b) {
+      gemm_tn(1, N, K, alpha, mat_a + t, lda, mat_b, ldb, mat_c + t * ldc, ldc);
+    } else if (!transpose_a && transpose_b) {
+      gemm_nt(1, N, K, alpha, mat_a + t * lda, lda, mat_b, ldb, mat_c + t * ldc, ldc);
+    } else {
+      gemm_tt(1, N, K, alpha, mat_a + t, lda, mat_b, ldb, mat_c + t * ldc, ldc);
+    }
+  }
+}

mindspore/lite/src/runtime/kernel/arm/opclib/fp32/gemm.h

@@ -0,0 +1,23 @@
+/**
+ * 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_FP32_GEMM_H_
+#define MINDSPORE_LITE_SRC_RUNTIME_KERNEL_ARM_OPCLIB_FP32_GEMM_H_
+
+void gemm(int transpose_a, int transpose_b, int M, int N, int K, float alpha, float *mat_a, int lda, float *mat_b,
+          int ldb, float beta, float *mat_c, int ldc);
+
+#endif  // MINDSPORE_LITE_SRC_RUNTIME_KERNEL_ARM_OPCLIB_FP32_GEMM_H_

mindspore/lite/src/runtime/kernel/arm/opclib/fp32/pooling_grad.cc

@@ -0,0 +1,149 @@
+/**
+ * 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 <cstdint>
+#include "src/runtime/kernel/arm/opclib/fp32/pooling_grad.h"
+
+void AvgPoolingGrad(const float *input_ptr, float *output_ptr, PoolingParameter *pooling_param) {
+  int stride_w = pooling_param->stride_w_;
+  int stride_h = pooling_param->stride_h_;
+  int pad_w = pooling_param->pad_l_;
+  int pad_h = pooling_param->pad_u_;
+  int win_w = pooling_param->window_w_;
+  int win_h = pooling_param->window_h_;
+  int channel = pooling_param->input_channel_;
+  int in_w = pooling_param->input_w_;
+  int in_h = pooling_param->input_h_;
+  int output_w = pooling_param->output_w_;
+  int output_h = pooling_param->output_h_;
+  int output_batch = pooling_param->output_batch_;
+
+  const float *inPtr;
+  for (int i = 0; i < output_h * output_w * channel * output_batch; i++) output_ptr[i] = 0.0;
+
+  // int pad_top = padding[2];
+
+  float kk = static_cast<float>(win_h * win_w);
+
+  for (uint16_t ib = 0; ib < output_batch; ib++) {
+    // int in_batch_offset = batch * in_h * in_w * channel;
+    // int out_batch_offset = batch * output_h * output_w * channel;
+    // out = grads->getData(ib*grads->imgSize());
+    // inPtr = in->getData(ib*in->imgSize());
+    float *out;
+    out = &output_ptr[(ib * output_h * output_w)];
+    inPtr = reinterpret_cast<const float *>(&input_ptr[(ib * in_h * in_w)]);
+    if (1) {  // in->layout() == Tensor::nhwc)
+      // iterate over yt
+      for (uint16_t yh = 0; yh < in_h; yh++) {
+        for (uint16_t yw = 0; yw < in_w; yw++) {
+          for (uint16_t ic = 0; ic < channel; ic++) {
+            int idx = (yw + yh * in_w) * channel + ic;  // (ic*in_h*in_w) + (in_w*yh) + yw;
+            float delta = inPtr[idx] / kk;
+            for (int32_t kh = 0; kh < win_h; kh++) {
+              int xh = yh * stride_h + kh - pad_h;
+              if ((xh < 0) || (xh >= output_h)) {
+                continue;
+              }
+              for (int32_t kw = 0; kw < win_w; kw++) {
+                int xw = yw * stride_w + kw - pad_w;
+                if ((xw < 0) || (xw >= output_w)) {
+                  continue;
+                }
+                // out[(ic*output_h*output_w) + (xh*output_w) + xw] += delta;
+                out[(xw + output_w * xh) * channel + ic] += delta;
+              }
+            }
+          }
+        }
+      }
+    } else {  // nchw
+      for (uint16_t ic = 0; ic < channel; ic++) {
+        // iterate over yt
+        for (uint16_t yh = 0; yh < in_h; yh++) {
+          for (uint16_t yw = 0; yw < in_w; yw++) {
+            int idx = (ic * in_h * in_w) + (in_w * yh) + yw;
+            float delta = inPtr[idx] / kk;
+            for (int32_t kh = 0; kh < win_h; kh++) {
+              int xh = yh * stride_h + kh - pad_h;
+              if ((xh < 0) || (xh >= output_h)) {
+                continue;
+              }
+              for (int32_t kw = 0; kw < win_w; kw++) {
+                int xw = yw * stride_w + kw - pad_w;
+                if ((xw < 0) || (xw >= output_w)) {
+                  continue;
+                }
+                out[(ic * output_h * output_w) + (xh * output_w) + xw] += delta;
+              }
+            }
+          }
+        }
+      }
+    }
+  }
+}
+
+void MaxPoolingGrad(const float *dy, const int *indices, float *output_ptr, PoolingParameter *pooling_param) {
+  // int stride_w = pooling_param->stride_w_;
+  // int stride_h = pooling_param->stride_h_;
+  // int pad_w = pooling_param->pad_l_;
+  // int pad_h = pooling_param->pad_u_;
+  // int win_w = pooling_param->window_w_;
+  // int win_h = pooling_param->window_h_;
+  int channel = pooling_param->input_channel_;
+  int in_w = pooling_param->input_w_;
+  int in_h = pooling_param->input_h_;
+  int output_w = pooling_param->output_w_;
+  int output_h = pooling_param->output_h_;
+  int output_batch = pooling_param->output_batch_;
+
+  int out_img_size =
+    output_h * output_w;  // Emir -- in original code this varible is calculated according to input size ??
+  int ind_img_size = in_h * in_w;
+  // const int w_pad = (output_w + pad_w + pad_w);
+
+  for (int i = 0; i < output_h * output_w * channel * output_batch; i++) output_ptr[i] = 0.0;
+
+  const float *yt = reinterpret_cast<const float *>(dy);
+  const int *pos = reinterpret_cast<const int *>(indices);
+  float *out;
+
+  if (1) {  // grads->layout() == Tensor::nhwc)
+    for (int ib = 0; ib < output_batch; ib++) {
+      out = &(output_ptr[ib * output_w * output_w * channel]);
+      for (int ix = 0; ix < ind_img_size; ix++) {
+        for (int cix = 0; cix < channel; cix++) {
+          int idx = (*pos) * channel + cix;
+          out[idx] += *yt;
+          pos++;
+          yt++;
+        }
+      }
+    }
+  } else {
+    for (int ib = 0; ib < output_batch; ib++) {
+      out = &output_ptr[(ib * out_img_size)];
+      for (int cix = 0; cix < channel; cix++) {
+        for (int ix = 0; ix < ind_img_size; ix++) {
+          int idx = cix * output_h * output_w + *pos;  // cord_y*output_w + cord_x;
+          out[idx] += *yt;
+          pos++;
+          yt++;
+        }
+      }
+    }
+  }
+}

mindspore/lite/src/runtime/kernel/arm/opclib/fp32/pooling_grad.h

@@ -0,0 +1,25 @@
+/**
+ * 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_FP32_POOLING_GRAD_H_
+#define MINDSPORE_LITE_SRC_RUNTIME_KERNEL_ARM_OPCLIB_FP32_POOLING_GRAD_H_
+
+#include "src/runtime/kernel/arm/opclib/fp32/pooling.h"
+
+void AvgPoolingGrad(const float *input_ptr, float *output_ptr, PoolingParameter *pooling_param);
+void MaxPoolingGrad(const float *dy, const int *indices_ptr, float *output_ptr, PoolingParameter *pooling_param);
+
+#endif  // MINDSPORE_LITE_SRC_RUNTIME_KERNEL_ARM_OPCLIB_FP32_POOLING_GRAD_H_

mindspore/lite/src/runtime/kernel/arm/opclib/fp32/reduce_grad.cc

@@ -0,0 +1,130 @@
+/**
+ * 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 <cstddef>
+#include <algorithm>
+#include "mindspore/lite/src/runtime/kernel/arm/opclib/fp32/reduce_grad.h"
+
+static inline bool NextIndex(const int num_dims, const int *dims, int *current) {
+  int carry = 1;
+  for (int idx = num_dims - 1; idx >= 0; --idx) {
+    int current_val = current[idx] + carry;
+    if (dims[idx] == current_val) {
+      current[idx] = 0;
+    } else {
+      current[idx] = current_val;
+      carry = 0;
+      break;
+    }
+  }
+  return (carry == 0);
+}
+
+static inline size_t GetInputOffset(const int num_dims, const int *dims, const int *iter) {
+  size_t offset = 0;
+  for (int idx = 0; idx < num_dims; ++idx) {
+    offset = offset * (size_t)(dims[idx]) + (size_t)(iter[idx]);
+  }
+
+  return offset;
+}
+
+static inline size_t GetOutputOffset(const int num_dims, const int *dims, const int *iter, const int num_axis,
+                                     const int *axes) {
+  size_t offset = 0;
+  for (int idx = 0; idx < num_dims; ++idx) {
+    // if we need to skip this axis
+    bool is_axis = false;
+    for (int axis_idx = 0; axis_idx < num_axis; ++axis_idx) {
+      if (idx == axes[axis_idx]) {
+        is_axis = true;
+        break;
+      }
+    }
+
+    if (!is_axis) {
+      offset = offset * (size_t)(dims[idx]) + (size_t)(iter[idx]);
+    }
+  }
+  return offset;
+}
+
+void ReduceMeanByAxes(const float *input_data, int *input_iter, const int *input_dims, int input_num_dims,
+                      const int *axes, int num_axes, float *output_data, const int *output_dims, int output_num_dims) {
+  size_t num_outputs = 1;
+  for (int idx = 0; idx < output_num_dims; ++idx) {
+    size_t current = (size_t)(output_dims[idx]);
+    num_outputs *= current;
+  }
+
+  // Reset input iterator.
+  for (int idx = 0; idx < input_num_dims; ++idx) {
+    input_iter[idx] = 0;
+  }
+  // Iterate through input_data.
+  do {
+    size_t input_offset = GetInputOffset(input_num_dims, input_dims, input_iter);
+    size_t output_offset = GetOutputOffset(input_num_dims, input_dims, input_iter, num_axes, axes);
+    output_data[output_offset] += input_data[input_offset];
+  } while (NextIndex(input_num_dims, input_dims, input_iter));
+
+  // Calculate mean by dividing output_data by num of aggregated element.
+  size_t num_elements_in_axis = 1;
+  for (int idx = 0; idx < num_axes; ++idx) {
+    size_t current = (size_t)(input_dims[axes[idx]]);
+    num_elements_in_axis *= current;
+  }
+
+  for (size_t idx = 0; idx < num_outputs; ++idx) {
+    output_data[idx] = output_data[idx] / static_cast<float>(num_elements_in_axis);
+  }
+}
+
+float ReduceMeanAll(const float *src, int size) {
+  float sum = 0;
+  for (int i = 0; i < size; ++i) {
+    sum += src[i];
+  }
+  return sum / size;
+}
+
+void ReduceSumByAxes(const float *input, const int *input_dims, float *output, const int *output_dims, int num_dims) {
+  int num_outputs = 1;
+  int same_shape = true;
+  for (int idx = 0; idx < num_dims; ++idx) {
+    num_outputs *= output_dims[idx];
+    if (output_dims[idx] != input_dims[idx]) same_shape = false;
+  }
+  if (same_shape) {
+    std::copy(input, input + num_outputs * sizeof(float), output);
+    // memcpy(output, input, num_outputs*sizeof(float));
+    return;
+  }
+
+  for (int idx = 0; idx < num_outputs; ++idx) output[idx] = 0;  // zero output
+
+  int input_iter[8] = {0};
+  int axes[5] = {0};
+  int num_axes = 0;
+  for (int i = 0; i < num_dims; i++)
+    if (output_dims[i] == 1) axes[num_axes++] = i;
+
+  // Iterate through input_data.
+  do {
+    size_t input_offset = GetInputOffset(num_dims, input_dims, input_iter);
+    size_t output_offset = GetOutputOffset(num_dims, input_dims, input_iter, num_axes, axes);
+    output[output_offset] += input[input_offset];
+  } while (NextIndex(num_dims, input_dims, input_iter));
+}

mindspore/lite/src/runtime/kernel/arm/opclib/fp32/reduce_grad.h

@@ -0,0 +1,24 @@
+/**
+ * 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_FP32_REDUCE_GRAD_H_
+#define MINDSPORE_LITE_SRC_RUNTIME_KERNEL_ARM_OPCLIB_FP32_REDUCE_GRAD_H_
+
+float ReduceMeanAll(const float *src, int size);
+void ReduceSumByAxes(const float *input, const int *input_dims, float *output, const int *output_dims, int num_dims);
+
+#endif  // MINDSPORE_LITE_SRC_RUNTIME_KERNEL_ARM_OPCLIB_FP32_REDUCE_GRAD_H_
+

mindspore/lite/src/runtime/kernel/arm/opclib/fp32/softmax_grad.h

@@ -0,0 +1,29 @@
+/**
+ * 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_FP32_SOFTMAX_GRAD_H_
+#define MINDSPORE_LITE_SRC_RUNTIME_KERNEL_ARM_OPCLIB_FP32_SOFTMAX_GRAD_H_
+
+#include "src/runtime/kernel/arm/opclib/op_base.h"
+
+struct SoftmaxCrossEntropyParameter {
+    OpParameter op_parameter;
+    int32_t batch_size_;
+    unsigned int number_of_classes_;
+    int n_dim_;
+    int input_shape_[5];
+};
+#endif  // MINDSPORE_LITE_SRC_RUNTIME_KERNEL_ARM_OPCLIB_FP32_SOFTMAX_GRAD_H_

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

@@ -0,0 +1,176 @@
+/**
+ * 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 <string.h>
+#include "src/runtime/kernel/arm/opclib/pack_ext.h"
+
+static int is_a_ge_zero_and_a_lt_b(int a, int b) { return (unsigned)(a) < (unsigned)(b); }
+
+void im2col_hwc(const float *in_data, float *data_col, ConvParameter *conv_param) {
+  const int pad_left = /*conv_param->pad_l_*/ conv_param->pad_w_;
+  // const int pad_right =  /*conv_param->pad_r_*/conv_param->pad_w_;
+  const int pad_up = /*conv_param->pad_u_*/ conv_param->pad_h_;
+  // const int pad_down =   /*conv_param->pad_d/*/conv_param->pad_h_;
+
+  const int stride_h = conv_param->stride_h_;
+  const int stride_w = conv_param->stride_w_;
+
+  const int dilation_h = conv_param->dilation_h_;
+  const int dilation_w = conv_param->dilation_w_;
+
+  const int kernel_h = conv_param->kernel_h_;
+  const int kernel_w = conv_param->kernel_w_;
+
+  const int in_height = conv_param->input_h_;
+  const int in_width = conv_param->input_w_;
+
+  const int output_h = conv_param->output_h_;
+  const int output_w = conv_param->output_w_;
+  const int channels = conv_param->input_channel_ / conv_param->group_;
+  const int tot_channels = conv_param->input_channel_;
+
+  int /*channel,*/ kernel_row, kernel_col, output_rows, output_col;
+
+  int row_stride_offset = 0;
+
+  for (output_rows = output_h; output_rows; output_rows--) {
+    int col_stride_offset = 0;
+    for (output_col = output_w; output_col; output_col--) {
+      for (kernel_row = 0; kernel_row < kernel_h; kernel_row++) {
+        int input_row = -pad_up + kernel_row * dilation_h + row_stride_offset;
+        for (kernel_col = 0; kernel_col < kernel_w; kernel_col++) {
+          int input_col = -pad_left + kernel_col * dilation_w + col_stride_offset;
+
+          if (is_a_ge_zero_and_a_lt_b(input_row, in_height) && is_a_ge_zero_and_a_lt_b(input_col, in_width)) {
+            const int offset = (input_row * in_width + input_col) * tot_channels;
+            memcpy(data_col, in_data + offset, sizeof(float) * channels);
+            data_col += channels;
+          } else {
+            memset(data_col, 0, sizeof(float) * channels);
+            data_col += channels;
+          }
+        }
+      }
+      col_stride_offset += stride_w;
+    }
+    row_stride_offset += stride_h;
+  }
+}
+
+// output matrix is (kernel_h*kernel_w*channels)X(output_h*output_w)
+void im2row_hwc(const float *in_data, float *data_row, ConvParameter *conv_param) {
+  const int pad_left = /*conv_param->pad_l_*/ conv_param->pad_w_;
+  // const int pad_right =  /*conv_param->pad_r_*/conv_param->pad_w_;
+  const int pad_up = /*conv_param->pad_u_*/ conv_param->pad_h_;
+  // const int pad_down =   /*conv_param->pad_d/*/conv_param->pad_h_;
+
+  const int stride_h = conv_param->stride_h_;
+  const int stride_w = conv_param->stride_w_;
+
+  const int dilation_h = conv_param->dilation_h_;
+  const int dilation_w = conv_param->dilation_w_;
+
+  const int kernel_h = conv_param->kernel_h_;
+  const int kernel_w = conv_param->kernel_w_;
+
+  const int in_height = conv_param->input_h_;
+  const int in_width = conv_param->input_w_;
+
+  const int output_h = conv_param->output_h_;
+  const int output_w = conv_param->output_w_;
+  const int channels = conv_param->input_channel_ / conv_param->group_;
+  const int tot_channels = conv_param->input_channel_;
+
+  int channel, kernel_row, kernel_col, output_rows, output_col;
+
+  for (kernel_row = 0; kernel_row < kernel_h; kernel_row++) {
+    for (kernel_col = 0; kernel_col < kernel_w; kernel_col++) {
+      for (channel = 0; channel < channels; channel++) {
+        int input_row = -pad_up + kernel_row * dilation_h;
+        for (output_rows = output_h; output_rows; output_rows--) {
+          if (!is_a_ge_zero_and_a_lt_b(input_row, in_height)) {
+            for (output_col = output_w; output_col; output_col--) {
+              *(data_row++) = 0;
+            }
+          } else {
+            int input_col = -pad_left + kernel_col * dilation_w;
+            for (output_col = output_w; output_col; output_col--) {
+              if (is_a_ge_zero_and_a_lt_b(input_col, in_width)) {
+                const int offset = (input_row * in_width + input_col) * tot_channels + channel;
+                *(data_row++) = in_data[offset];
+              } else {
+                *(data_row++) = 0;
+              }
+              input_col += stride_w;
+            }
+          }
+          input_row += stride_h;
+        }
+      }
+    }
+  }
+}
+
+void col2im_hwc(const float *data_col, float *data_im, ConvParameter *conv_param) {
+  const int pad_left = /*conv_param->pad_l_*/ conv_param->pad_w_;
+  // const int pad_right =  /*conv_param->pad_r_*/conv_param->pad_w_;
+  const int pad_up = /*conv_param->pad_u_*/ conv_param->pad_h_;
+  // const int pad_down =   /*conv_param->pad_d/*/conv_param->pad_h_;
+
+  const int stride_h = conv_param->stride_h_;
+  const int stride_w = conv_param->stride_w_;
+
+  const int dilation_h = conv_param->dilation_h_;
+  const int dilation_w = conv_param->dilation_w_;
+
+  const int kernel_h = conv_param->kernel_h_;
+  const int kernel_w = conv_param->kernel_w_;
+
+  const int in_height = conv_param->input_h_;
+  const int in_width = conv_param->input_w_;
+
+  const int output_h = conv_param->output_h_;
+  const int output_w = conv_param->output_w_;
+  const int channels = conv_param->input_channel_ / conv_param->group_;
+  const int tot_channels = conv_param->input_channel_;
+
+  int kernel_row, kernel_col, output_rows, output_col;
+
+  int row_stride_offset = 0;
+
+  for (output_rows = output_h; output_rows; output_rows--) {
+    int col_stride_offset = 0;
+    for (output_col = output_w; output_col; output_col--) {
+      for (kernel_row = 0; kernel_row < kernel_h; kernel_row++) {
+        int input_row = -pad_up + kernel_row * dilation_h + row_stride_offset;
+        for (kernel_col = 0; kernel_col < kernel_w; kernel_col++) {
+          int input_col = -pad_left + kernel_col * dilation_w + col_stride_offset;
+
+          if (is_a_ge_zero_and_a_lt_b(input_row, in_height) && is_a_ge_zero_and_a_lt_b(input_col, in_width)) {
+            int offset = (input_row * in_width + input_col) * tot_channels;
+            float *data_im_ptr = &data_im[offset];
+            for (int i = 0; i < channels; i++) {
+              data_im_ptr[i] += data_col[i];
+            }
+          }
+          data_col += channels;
+        }
+      }
+      col_stride_offset += stride_w;
+    }
+    row_stride_offset += stride_h;
+  }
+}

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

@@ -0,0 +1,26 @@
+/**
+ * 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_PACK_EXT_H_
+#define MINDSPORE_LITE_SRC_RUNTIME_KERNEL_ARM_OPCLIB_PACK_EXT_H_
+
+#include "src/runtime/kernel/arm/opclib/conv_parameter.h"
+
+void im2col_hwc(const float *in_data, float *data_col, ConvParameter *conv_param);
+void im2row_hwc(const float *in_data, float *data_row, ConvParameter *conv_param);
+void col2im_hwc(const float *data_col, float *data_im, ConvParameter *conv_param);
+
+#endif  // MINDSPORE_LITE_SRC_RUNTIME_KERNEL_ARM_OPCLIB_PACK_EXT_H

mindspore/lite/test/CMakeLists.txt

@@ -152,6 +152,7 @@ set(TEST_LITE_SRC
         ${LITE_DIR}/src/scheduler.cc
         ${LITE_DIR}/src/common/graph_util.cc
         ${LITE_DIR}/src/common/file_utils.cc
+        ${LITE_DIR}/src/common/file_utils_ext.cc
         ${LITE_DIR}/src/common/utils.cc
         ${LITE_DIR}/tools/common/graph_util.cc
         ${LITE_DIR}/tools/common/tensor_util.cc

mindspore/lite/test/ut/src/runtime/kernel/arm/fp32/activation_grad_fp32_tests.cc

@@ -0,0 +1,312 @@
+/**
+ * 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 <iostream>
+#include <memory>
+#include <vector>
+
+#include "utils/log_adapter.h"
+#include "common/common_test.h"
+#include "src/common/file_utils.h"
+#include "src/common/file_utils_ext.h"
+#include "mindspore/lite/src/kernel_registry.h"
+#include "mindspore/lite/src/ir/tensor.h"
+#include "mindspore/lite/src/lite_kernel.h"
+#include "mindspore/lite/src/runtime/kernel/arm/fp32/activation_grad.h"
+
+namespace mindspore {
+class TestActGradFp32 :  public mindspore::Common {
+ public:
+  TestActGradFp32() {}
+};
+
+TEST_F(TestActGradFp32, ReluGradFp32) {
+  // runtime part
+  printf("Calculating runtime cost...\n");
+  uint64_t time_avg = 0;
+  size_t output_data_size = 50;
+
+  size_t input_size;
+  std::string input_path = "./test_data/activationGrad/relu_y_50.bin";
+  auto input_data = reinterpret_cast<float *>(mindspore::lite::ReadFile(input_path.c_str(), &input_size));
+  std::string yt_path = "./test_data/activationGrad/relu_yt_50.bin";
+  auto yt_data = reinterpret_cast<float *>(mindspore::lite::ReadFile(yt_path.c_str(), &input_size));
+
+  auto output_data = new float[output_data_size];
+  // warm up loop
+  for (int i = 0; i < 3; i++) {
+    ReluGrad(yt_data, input_data, 50, output_data);
+  }
+
+  int loop_count = 100;
+  auto time_start = mindspore::lite::GetTimeUs();
+  for (int i = 0; i < loop_count; i++) {
+    ReluGrad(yt_data, input_data, 50, output_data);
+  }
+  auto time_end = mindspore::lite::GetTimeUs();
+  auto cost = time_end - time_start;
+  time_avg = cost / loop_count;
+  printf("single thread running time : %f ms\n", time_avg / 1000.0f);
+
+  printf("==================output data=================\n");
+  for (int i = 0; i < 20; i++) {
+    std::cout << output_data[i] << " ,";
+  }
+  std::cout << std::endl;
+
+  std::string output_path = "./test_data/activationGrad/relu_out_50.bin";
+
+  int res = lite::CompareRelativeOutput(output_data, output_path);
+
+  EXPECT_EQ(res, 0);
+
+  delete input_data;
+  delete[] output_data;
+  delete yt_data;
+
+  MS_LOG(INFO) << "ReluGradFp32 passed";
+}
+
+TEST_F(TestActGradFp32, Relu6GradFp32) {
+  // runtime part
+  printf("Calculating runtime cost...\n");
+  uint64_t time_avg = 0;
+  size_t output_data_size = 50;
+
+  size_t input_size;
+  std::string input_path = "./test_data/activationGrad/relu6_y_50.bin";
+  auto input_data = reinterpret_cast<float *>(mindspore::lite::ReadFile(input_path.c_str(), &input_size));
+  std::string yt_path = "./test_data/activationGrad/relu6_yt_50.bin";
+  auto yt_data = reinterpret_cast<float *>(mindspore::lite::ReadFile(yt_path.c_str(), &input_size));
+
+  auto output_data = new float[output_data_size];
+  // warm up loop
+  for (int i = 0; i < 3; i++) {
+    Relu6Grad(yt_data, input_data, 50, output_data);
+  }
+
+  int loop_count = 100;
+  auto time_start = mindspore::lite::GetTimeUs();
+  for (int i = 0; i < loop_count; i++) {
+    Relu6Grad(yt_data, input_data, 50, output_data);
+  }
+  auto time_end = mindspore::lite::GetTimeUs();
+  auto cost = time_end - time_start;
+  time_avg = cost / loop_count;
+  printf("single thread running time : %f ms\n", time_avg / 1000.0f);
+
+  printf("==================output data=================\n");
+  for (int i = 0; i < 20; i++) {
+    std::cout << output_data[i] << " ,";
+  }
+  std::cout << std::endl;
+
+  std::string output_path = "./test_data/activationGrad/relu6_out_50.bin";
+  int res = lite::CompareRelativeOutput(output_data, output_path);
+
+  EXPECT_EQ(res, 0);
+
+  delete input_data;
+  delete[] output_data;
+  delete yt_data;
+
+  MS_LOG(INFO) << "Relu6GradFp32 passed";
+}
+
+TEST_F(TestActGradFp32, LReluGradFp32) {
+  // runtime part
+  printf("Calculating runtime cost...\n");
+  uint64_t time_avg = 0;
+  size_t output_data_size = 50;
+
+  size_t input_size;
+  std::string input_path = "./test_data/activationGrad/lrelu_y_50.bin";
+  auto input_data = reinterpret_cast<float *>(mindspore::lite::ReadFile(input_path.c_str(), &input_size));
+  std::string yt_path = "./test_data/activationGrad/lrelu_yt_50.bin";
+  auto yt_data = reinterpret_cast<float *>(mindspore::lite::ReadFile(yt_path.c_str(), &input_size));
+
+  auto output_data = new float[output_data_size];
+  // warm up loop
+  for (int i = 0; i < 3; i++) {
+    LReluGrad(yt_data, input_data, 50, output_data, 0.1);
+  }
+
+  int loop_count = 100;
+  auto time_start = mindspore::lite::GetTimeUs();
+  for (int i = 0; i < loop_count; i++) {
+    LReluGrad(yt_data, input_data, 50, output_data, 0.1);
+  }
+  auto time_end = mindspore::lite::GetTimeUs();
+  auto cost = time_end - time_start;
+  time_avg = cost / loop_count;
+  printf("single thread running time : %f ms\n", time_avg / 1000.0f);
+
+  printf("==================output data=================\n");
+  for (int i = 0; i < 20; i++) {
+    std::cout << output_data[i] << " ,";
+  }
+  std::cout << std::endl;
+
+  std::string output_path = "./test_data/activationGrad/lrelu_out_50.bin";
+  int res = lite::CompareRelativeOutput(output_data, output_path);
+
+  EXPECT_EQ(res, 0);
+
+  delete input_data;
+  delete[] output_data;
+  delete yt_data;
+
+  MS_LOG(INFO) << "LReluGradFp32 passed";
+}
+
+TEST_F(TestActGradFp32, SigmoidGradFp32) {
+  // runtime part
+  printf("Calculating runtime cost...\n");
+  uint64_t time_avg = 0;
+  size_t output_data_size = 50;
+
+  size_t input_size;
+  std::string input_path = "./test_data/activationGrad/sigmoid_y_50.bin";
+  auto input_data = reinterpret_cast<float *>(mindspore::lite::ReadFile(input_path.c_str(), &input_size));
+  std::string yt_path = "./test_data/activationGrad/sigmoid_yt_50.bin";
+  auto yt_data = reinterpret_cast<float *>(mindspore::lite::ReadFile(yt_path.c_str(), &input_size));
+
+  auto output_data = new float[output_data_size];
+  // warm up loop
+  for (int i = 0; i < 3; i++) {
+    SigmoidGrad(yt_data, input_data, 50, output_data);
+  }
+
+  int loop_count = 100;
+  auto time_start = mindspore::lite::GetTimeUs();
+  for (int i = 0; i < loop_count; i++) {
+    SigmoidGrad(yt_data, input_data, 50, output_data);
+  }
+  auto time_end = mindspore::lite::GetTimeUs();
+  auto cost = time_end - time_start;
+  time_avg = cost / loop_count;
+  printf("single thread running time : %f ms\n", time_avg / 1000.0f);
+
+  printf("==================output data=================\n");
+  for (int i = 0; i < 20; i++) {
+    std::cout << output_data[i] << " ,";
+  }
+  std::cout << std::endl;
+
+  std::string output_path = "./test_data/activationGrad/sigmoid_out_50.bin";
+  int res = lite::CompareRelativeOutput(output_data, output_path);
+
+  EXPECT_EQ(res, 0);
+  // lite::CompareOutput(output_data, output_path);
+
+  delete input_data;
+  delete[] output_data;
+  delete yt_data;
+
+  MS_LOG(INFO) << "SigmoidGradFp32 passed";
+}
+
+TEST_F(TestActGradFp32, tanhGradFp32) {
+  // runtime part
+  printf("Calculating runtime cost...\n");
+  uint64_t time_avg = 0;
+  size_t output_data_size = 50;
+
+  size_t input_size;
+  std::string input_path = "./test_data/activationGrad/tanh_y_50.bin";
+  auto input_data = reinterpret_cast<float *>(mindspore::lite::ReadFile(input_path.c_str(), &input_size));
+  std::string yt_path = "./test_data/activationGrad/tanh_yt_50.bin";
+  auto yt_data = reinterpret_cast<float *>(mindspore::lite::ReadFile(yt_path.c_str(), &input_size));
+
+  auto output_data = new float[output_data_size];
+  // warm up loop
+  for (int i = 0; i < 3; i++) {
+    TanhGrad(yt_data, input_data, 50, output_data);
+  }
+
+  int loop_count = 100;
+  auto time_start = mindspore::lite::GetTimeUs();
+  for (int i = 0; i < loop_count; i++) {
+    TanhGrad(yt_data, input_data, 50, output_data);
+  }
+  auto time_end = mindspore::lite::GetTimeUs();
+  auto cost = time_end - time_start;
+  time_avg = cost / loop_count;
+  printf("single thread running time : %f ms\n", time_avg / 1000.0f);
+
+  printf("==================output data=================\n");
+  for (int i = 0; i < 20; i++) {
+    std::cout << output_data[i] << " ,";
+  }
+  std::cout << std::endl;
+
+  std::string output_path = "./test_data/activationGrad/tanh_out_50.bin";
+  int res = lite::CompareRelativeOutput(output_data, output_path);
+
+  EXPECT_EQ(res, 0);
+
+  delete input_data;
+  delete[] output_data;
+  delete yt_data;
+  MS_LOG(INFO) << "TanhGradFp32 passed";
+}
+
+TEST_F(TestActGradFp32, hswishGradFp32) {
+  // runtime part
+  printf("Calculating runtime cost...\n");
+  uint64_t time_avg = 0;
+  size_t output_data_size = 50;
+
+  size_t input_size;
+  std::string input_path = "./test_data/activationGrad/hswish_x_50.bin";
+  auto input_data = reinterpret_cast<float *>(mindspore::lite::ReadFile(input_path.c_str(), &input_size));
+  std::string yt_path = "./test_data/activationGrad/hswish_yt_50.bin";
+  auto yt_data = reinterpret_cast<float *>(mindspore::lite::ReadFile(yt_path.c_str(), &input_size));
+
+  auto output_data = new float[output_data_size];
+  // warm up loop
+  for (int i = 0; i < 3; i++) {
+    HSwishGrad(yt_data, input_data, 50, output_data);
+  }
+
+  int loop_count = 100;
+  auto time_start = mindspore::lite::GetTimeUs();
+  for (int i = 0; i < loop_count; i++) {
+    HSwishGrad(yt_data, input_data, 50, output_data);
+  }
+  auto time_end = mindspore::lite::GetTimeUs();
+  auto cost = time_end - time_start;
+  time_avg = cost / loop_count;
+  printf("single thread running time : %f ms\n", time_avg / 1000.0f);
+
+  printf("==================output data=================\n");
+  for (int i = 0; i < 20; i++) {
+    std::cout << output_data[i] << " ,";
+  }
+  std::cout << std::endl;
+
+  std::string output_path = "./test_data/activationGrad/hswish_out_50.bin";
+  int res = lite::CompareRelativeOutput(output_data, output_path);
+
+  EXPECT_EQ(res, 0);
+
+  delete input_data;
+  delete[] output_data;
+  delete yt_data;
+  MS_LOG(INFO) << "hswishGradFp32 passed";
+}
+
+}  // namespace mindspore

mindspore/lite/test/ut/src/runtime/kernel/arm/fp32/arithmetic_grad_fp32_tests.cc

@@ -0,0 +1,497 @@
+/**
+ * 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 <iostream>
+#include <memory>
+#include <vector>
+#include "utils/log_adapter.h"
+#include "common/common_test.h"
+#include "src/common/file_utils.h"
+#include "src/common/file_utils_ext.h"
+#include "mindspore/lite/src/runtime/kernel/arm/opclib/fp32/reduce.h"
+#include "mindspore/lite/src/runtime/kernel/arm/fp32/arithmetic_grad.h"
+#include "mindspore/lite/src/kernel_registry.h"
+
+namespace mindspore {
+
+class TestArithmeticGradFp32 : public mindspore::Common {
+ public:
+  TestArithmeticGradFp32() {}
+};
+
+std::vector<lite::tensor::Tensor *> GenerateTensorsForTest(const char *test, int test_id) {
+  size_t input_size;
+  std::vector<int> large_dim({4, 6});
+  std::vector<int> small_dim({6});
+  int large_size = (4 * 6);
+  int small_size = (1 * 6);
+  char *dx1_file = const_cast<char *>("./test_data/operators/arithmetic_fp32_1_x1_4_6.bin");
+  char *dx2_file = const_cast<char *>("./test_data/operators/arithmetic_fp32_1_x2_1_6.bin");
+
+  if (test_id == 7) {
+    large_dim = std::vector<int>({4, 5, 6});
+    small_dim = std::vector<int>({6});
+    large_size = (4 * 5 * 6);
+    small_size = (6);
+    dx1_file = const_cast<char *>("./test_data/operators/arithmetic_fp32_7_x1_4_5_6.bin");
+    dx2_file = const_cast<char *>("./test_data/operators/arithmetic_fp32_7_x2_1_1_6.bin");
+  }
+  if (test_id >= 8) {
+    large_dim = std::vector<int>({5, 4, 6});
+    small_dim = std::vector<int>({5, 1, 6});
+    large_size = (4 * 5 * 6);
+    small_size = (5 * 6);
+    dx1_file = const_cast<char *>("./test_data/operators/arithmetic_fp32_8_x1_5_4_6.bin");
+    dx2_file = const_cast<char *>("./test_data/operators/arithmetic_fp32_8_x2_5_1_6.bin");
+  }
+
+  auto dy_data = reinterpret_cast<float *>(mindspore::lite::ReadFile(test, &input_size));
+  lite::tensor::Tensor *dy_tensor = new lite::tensor::Tensor(TypeId::kNumberTypeFloat32, large_dim);
+  dy_tensor->SetData(dy_data);
+
+  auto x1_data = reinterpret_cast<float *>(mindspore::lite::ReadFile(dx1_file, &input_size));
+  lite::tensor::Tensor *x1_tensor = new lite::tensor::Tensor(TypeId::kNumberTypeFloat32, large_dim);
+  x1_tensor->SetData(x1_data);
+
+  auto x2_data = reinterpret_cast<float *>(mindspore::lite::ReadFile(dx2_file, &input_size));
+  lite::tensor::Tensor *x2_tensor = new lite::tensor::Tensor(TypeId::kNumberTypeFloat32, small_dim);
+  x2_tensor->SetData(x2_data);
+
+  auto dx1_data = new float[large_size];
+  lite::tensor::Tensor *dx1_tensor = new lite::tensor::Tensor(TypeId::kNumberTypeFloat32, large_dim);
+  dx1_tensor->SetData(dx1_data);
+
+  auto dx2_data = new float[small_size];
+  lite::tensor::Tensor *dx2_tensor = new lite::tensor::Tensor(TypeId::kNumberTypeFloat32, small_dim);
+  dx2_tensor->SetData(dx2_data);
+
+  std::vector<lite::tensor::Tensor *> ret_vector = {dy_tensor, x1_tensor, x2_tensor, dx1_tensor, dx2_tensor};
+  return ret_vector;
+}
+
+TEST_F(TestArithmeticGradFp32, TestAddGradFp32) {
+  auto param = new ArithmeticParameter();
+  param->op_parameter_.type_ = PrimitiveType_AddGrad;
+  std::vector<lite::tensor::Tensor *> all_tensors =
+    GenerateTensorsForTest("./test_data/operators/arithmetic_fp32_1_dy_4_6.bin", 1);
+
+  std::vector<lite::tensor::Tensor *> inputs = {all_tensors[0], all_tensors[1], all_tensors[2]};
+  std::vector<lite::tensor::Tensor *> outputs = {all_tensors[3], all_tensors[4]};
+  kernel::KernelKey desc = {kernel::kCPU, TypeId::kNumberTypeFloat32, schema::PrimitiveType_AddGrad};
+  auto creator = lite::KernelRegistry::GetInstance()->GetCreator(desc);
+  auto kernel_obj = creator(inputs, outputs, reinterpret_cast<OpParameter *>(param), NULL, desc);
+  kernel_obj->Run();
+
+  float *output_ptr = reinterpret_cast<float *>(outputs[1]->Data());
+  printf("==================output data=================\n");
+  for (int i = 0; i < 6; i++) {
+    std::cout << output_ptr[i] << " ,";
+  }
+  std::cout << std::endl;
+
+  std::string output_path = "./test_data/operators/arithmetic_fp32_1_dx1_4_6.bin";
+  EXPECT_EQ(0, lite::CompareRelativeOutput(reinterpret_cast<float *>(outputs[0]->Data()), output_path));
+
+  std::string dx2_path = "./test_data/operators/arithmetic_fp32_1_dx2_1_6.bin";
+  EXPECT_EQ(0, lite::CompareRelativeOutput(output_ptr, dx2_path));
+
+  for (int i = 0; i < 5; i++) delete all_tensors[i];
+  delete param;
+  MS_LOG(INFO) << "TestAddGradFp32 passed";
+}
+
+TEST_F(TestArithmeticGradFp32, TestAddGrad2Fp32) {
+  auto param = new ArithmeticParameter();
+  param->op_parameter_.type_ = PrimitiveType_AddGrad;
+  std::vector<lite::tensor::Tensor *> all_tensors =
+    GenerateTensorsForTest("./test_data/operators/arithmetic_fp32_1_dy_4_6.bin", 1);
+
+  std::vector<lite::tensor::Tensor *> inputs = {all_tensors[0], all_tensors[2], all_tensors[1]};
+  std::vector<lite::tensor::Tensor *> outputs = {all_tensors[4], all_tensors[3]};
+  kernel::KernelKey desc = {kernel::kCPU, TypeId::kNumberTypeFloat32, schema::PrimitiveType_AddGrad};
+  auto creator = lite::KernelRegistry::GetInstance()->GetCreator(desc);
+  auto kernel_obj = creator(inputs, outputs, reinterpret_cast<OpParameter *>(param), NULL, desc);
+  kernel_obj->Run();
+
+  float *output_ptr = reinterpret_cast<float *>(outputs[0]->Data());
+  printf("==================output data=================\n");
+  for (int i = 0; i < 6; i++) {
+    std::cout << output_ptr[i] << " ,";
+  }
+  std::cout << std::endl;
+
+  std::string output_path = "./test_data/operators/arithmetic_fp32_1_dx1_4_6.bin";
+  EXPECT_EQ(0, lite::CompareRelativeOutput(reinterpret_cast<float *>(outputs[1]->Data()), output_path));
+
+  std::string dx2_path = "./test_data/operators/arithmetic_fp32_1_dx2_1_6.bin";
+  EXPECT_EQ(0, lite::CompareRelativeOutput(output_ptr, dx2_path));
+
+  for (int i = 0; i < 5; i++) delete all_tensors[i];
+  delete param;
+  MS_LOG(INFO) << "TestAddGrad2Fp32 passed";
+}
+
+TEST_F(TestArithmeticGradFp32, TestAddGrad3Fp32) {
+  auto param = new ArithmeticParameter();
+  param->op_parameter_.type_ = PrimitiveType_AddGrad;
+  std::vector<lite::tensor::Tensor *> all_tensors =
+    GenerateTensorsForTest("./test_data/operators/arithmetic_fp32_8_dy_5_4_6.bin", 8);
+
+  std::vector<lite::tensor::Tensor *> inputs = {all_tensors[0], all_tensors[1], all_tensors[2]};
+  std::vector<lite::tensor::Tensor *> outputs = {all_tensors[3], all_tensors[4]};
+  kernel::KernelKey desc = {kernel::kCPU, TypeId::kNumberTypeFloat32, schema::PrimitiveType_AddGrad};
+  auto creator = lite::KernelRegistry::GetInstance()->GetCreator(desc);
+  auto kernel_obj = creator(inputs, outputs, reinterpret_cast<OpParameter *>(param), NULL, desc);
+  kernel_obj->Run();
+
+  float *output_ptr = reinterpret_cast<float *>(outputs[0]->Data());
+  printf("==================output data=================\n");
+  for (int i = 0; i < 6; i++) {
+    std::cout << output_ptr[i] << " ,";
+  }
+  std::cout << std::endl;
+
+  std::string output_path = "./test_data/operators/arithmetic_fp32_8_dx2_5_1_6.bin";
+  EXPECT_EQ(0, lite::CompareRelativeOutput(reinterpret_cast<float *>(outputs[1]->Data()), output_path));
+
+  std::string dx2_path = "./test_data/operators/arithmetic_fp32_8_dx1_5_4_6.bin";
+  EXPECT_EQ(0, lite::CompareRelativeOutput(output_ptr, dx2_path));
+
+  for (int i = 0; i < 5; i++) delete all_tensors[i];
+  delete param;
+  MS_LOG(INFO) << "TestAddGrad3Fp32 passed";
+}
+
+TEST_F(TestArithmeticGradFp32, TestSubGradFp32) {
+  auto param = new ArithmeticParameter();
+  param->op_parameter_.type_ = PrimitiveType_SubGrad;
+  std::vector<lite::tensor::Tensor *> all_tensors =
+    GenerateTensorsForTest("./test_data/operators/arithmetic_fp32_2_dy_4_6.bin", 2);
+
+  std::vector<lite::tensor::Tensor *> inputs = {all_tensors[0], all_tensors[1], all_tensors[2]};
+  std::vector<lite::tensor::Tensor *> outputs = {all_tensors[3], all_tensors[4]};
+  kernel::KernelKey desc = {kernel::kCPU, TypeId::kNumberTypeFloat32, schema::PrimitiveType_SubGrad};
+  auto creator = lite::KernelRegistry::GetInstance()->GetCreator(desc);
+  auto kernel_obj = creator(inputs, outputs, reinterpret_cast<OpParameter *>(param), NULL, desc);
+  kernel_obj->Run();
+
+  float *output_ptr = reinterpret_cast<float *>(outputs[1]->Data());
+  printf("==================output data=================\n");
+  for (int i = 0; i < 6; i++) {
+    std::cout << output_ptr[i] << " ,";
+  }
+  std::cout << std::endl;
+
+  std::string output_path = "./test_data/operators/arithmetic_fp32_2_dx1_4_6.bin";
+  EXPECT_EQ(0, lite::CompareRelativeOutput(reinterpret_cast<float *>(outputs[0]->Data()), output_path));
+
+  std::string dx2_path = "./test_data/operators/arithmetic_fp32_2_dx2_1_6.bin";
+  EXPECT_EQ(0, lite::CompareRelativeOutput(output_ptr, dx2_path));
+
+  for (int i = 0; i < 5; i++) delete all_tensors[i];
+  delete param;
+  MS_LOG(INFO) << "TestSubGradFp32 passed";
+}
+
+TEST_F(TestArithmeticGradFp32, TestSubGrad2Fp32) {
+  auto param = new ArithmeticParameter();
+  param->op_parameter_.type_ = PrimitiveType_SubGrad;
+  std::vector<lite::tensor::Tensor *> all_tensors =
+    GenerateTensorsForTest("./test_data/operators/arithmetic_fp32_3_dy_4_6.bin", 3);
+
+  std::vector<lite::tensor::Tensor *> inputs = {all_tensors[0], all_tensors[2], all_tensors[1]};
+  std::vector<lite::tensor::Tensor *> outputs = {all_tensors[4], all_tensors[3]};
+  kernel::KernelKey desc = {kernel::kCPU, TypeId::kNumberTypeFloat32, schema::PrimitiveType_SubGrad};
+  auto creator = lite::KernelRegistry::GetInstance()->GetCreator(desc);
+  auto kernel_obj = creator(inputs, outputs, reinterpret_cast<OpParameter *>(param), NULL, desc);
+  kernel_obj->Run();
+
+  float *output_ptr = reinterpret_cast<float *>(outputs[0]->Data());
+  printf("==================output data=================\n");
+  for (int i = 0; i < 6; i++) {
+    std::cout << output_ptr[i] << " ,";
+  }
+  std::cout << std::endl;
+
+  std::string output_path = "./test_data/operators/arithmetic_fp32_3_dx1_4_6.bin";
+  EXPECT_EQ(0, lite::CompareRelativeOutput(reinterpret_cast<float *>(outputs[1]->Data()), output_path));
+
+  std::string dx2_path = "./test_data/operators/arithmetic_fp32_3_dx2_1_6.bin";
+  EXPECT_EQ(0, lite::CompareRelativeOutput(output_ptr, dx2_path));
+
+  for (int i = 0; i < 5; i++) delete all_tensors[i];
+  delete param;
+  MS_LOG(INFO) << "TestSubGrad2Fp32 passed";
+}
+
+TEST_F(TestArithmeticGradFp32, TestMulGradFp32) {
+  auto param = new ArithmeticParameter();
+  param->op_parameter_.type_ = PrimitiveType_MulGrad;
+  std::vector<lite::tensor::Tensor *> all_tensors =
+    GenerateTensorsForTest("./test_data/operators/arithmetic_fp32_4_dy_4_6.bin", 4);
+
+  std::vector<lite::tensor::Tensor *> inputs = {all_tensors[0], all_tensors[1], all_tensors[2]};
+  std::vector<lite::tensor::Tensor *> outputs = {all_tensors[3], all_tensors[4]};
+  kernel::KernelKey desc = {kernel::kCPU, TypeId::kNumberTypeFloat32, schema::PrimitiveType_MulGrad};
+  auto creator = lite::KernelRegistry::GetInstance()->GetCreator(desc);
+  auto kernel_obj = creator(inputs, outputs, reinterpret_cast<OpParameter *>(param), NULL, desc);
+
+  int loop_count = 1000;
+  auto time_start = mindspore::lite::GetTimeUs();
+  for (int i = 0; i < loop_count; i++) {
+    kernel_obj->Run();
+  }
+  auto time_end = mindspore::lite::GetTimeUs();
+  auto cost = time_end - time_start;
+  printf("total cost (for %d loops): %lu us\n", loop_count, cost);
+  // auto time_avg = cost / loop_count;
+  // printf("single thread running time : %f ms\n", time_avg / 1000.0f);
+
+  float *output_ptr = reinterpret_cast<float *>(outputs[1]->Data());
+  printf("==================output data=================\n");
+  for (int i = 0; i < 6; i++) {
+    std::cout << output_ptr[i] << " ,";
+  }
+  std::cout << std::endl;
+
+  std::string output_path = "./test_data/operators/arithmetic_fp32_4_dx1_4_6.bin";
+  EXPECT_EQ(0, lite::CompareRelativeOutput(reinterpret_cast<float *>(outputs[0]->Data()), output_path));
+
+  std::string dx2_path = "./test_data/operators/arithmetic_fp32_4_dx2_1_6.bin";
+  EXPECT_EQ(0, lite::CompareRelativeOutput(output_ptr, dx2_path));
+
+  for (int i = 0; i < 5; i++) delete all_tensors[i];
+  delete param;
+  MS_LOG(INFO) << "TestMulGradFp32 passed";
+}
+
+TEST_F(TestArithmeticGradFp32, TestMulGrad2Fp32) {
+  auto param = new ArithmeticParameter();
+  param->op_parameter_.type_ = PrimitiveType_MulGrad;
+  std::vector<lite::tensor::Tensor *> all_tensors =
+    GenerateTensorsForTest("./test_data/operators/arithmetic_fp32_4_dy_4_6.bin", 4);
+
+  std::vector<lite::tensor::Tensor *> inputs = {all_tensors[0], all_tensors[2], all_tensors[1]};
+  std::vector<lite::tensor::Tensor *> outputs = {all_tensors[4], all_tensors[3]};
+  kernel::KernelKey desc = {kernel::kCPU, TypeId::kNumberTypeFloat32, schema::PrimitiveType_MulGrad};
+  auto creator = lite::KernelRegistry::GetInstance()->GetCreator(desc);
+  auto kernel_obj = creator(inputs, outputs, reinterpret_cast<OpParameter *>(param), NULL, desc);
+  kernel_obj->Run();
+
+  float *output_ptr = reinterpret_cast<float *>(outputs[0]->Data());
+  printf("==================output data=================\n");
+  for (int i = 0; i < 6; i++) {
+    std::cout << output_ptr[i] << " ,";
+  }
+  std::cout << std::endl;
+
+  std::string output_path = "./test_data/operators/arithmetic_fp32_4_dx1_4_6.bin";
+  EXPECT_EQ(0, lite::CompareRelativeOutput(reinterpret_cast<float *>(outputs[1]->Data()), output_path));
+
+  std::string dx2_path = "./test_data/operators/arithmetic_fp32_4_dx2_1_6.bin";
+  EXPECT_EQ(0, lite::CompareRelativeOutput(output_ptr, dx2_path));
+
+  for (int i = 0; i < 5; i++) delete all_tensors[i];
+  delete param;
+  MS_LOG(INFO) << "TestMulGrad2Fp32 passed";
+}
+
+TEST_F(TestArithmeticGradFp32, TestMulGrad3Fp32) {
+  auto param = new ArithmeticParameter();
+  param->op_parameter_.type_ = PrimitiveType_MulGrad;
+  std::vector<lite::tensor::Tensor *> all_tensors =
+    GenerateTensorsForTest("./test_data/operators/arithmetic_fp32_9_dy_5_4_6.bin", 9);
+
+  std::vector<lite::tensor::Tensor *> inputs = {all_tensors[0], all_tensors[1], all_tensors[2]};
+  std::vector<lite::tensor::Tensor *> outputs = {all_tensors[3], all_tensors[4]};
+  kernel::KernelKey desc = {kernel::kCPU, TypeId::kNumberTypeFloat32, schema::PrimitiveType_MulGrad};
+  auto creator = lite::KernelRegistry::GetInstance()->GetCreator(desc);
+  auto kernel_obj = creator(inputs, outputs, reinterpret_cast<OpParameter *>(param), NULL, desc);
+  kernel_obj->Run();
+
+  float *output_ptr = reinterpret_cast<float *>(outputs[1]->Data());
+  printf("==================output data=================\n");
+  for (int i = 0; i < 6; i++) {
+    std::cout << output_ptr[i] << " ,";
+  }
+  std::cout << std::endl;
+
+  std::string output_path = "./test_data/operators/arithmetic_fp32_9_dx1_5_4_6.bin";
+  EXPECT_EQ(0, lite::CompareRelativeOutput(reinterpret_cast<float *>(outputs[0]->Data()), output_path));
+
+  std::string dx2_path = "./test_data/operators/arithmetic_fp32_9_dx2_5_1_6.bin";
+  EXPECT_EQ(0, lite::CompareRelativeOutput(output_ptr, dx2_path));
+
+  for (int i = 0; i < 5; i++) delete all_tensors[i];
+  delete param;
+  MS_LOG(INFO) << "TestMulGrad3Fp32 passed";
+}
+
+TEST_F(TestArithmeticGradFp32, TestMulGrad4Fp32) {
+  auto param = new ArithmeticParameter();
+  param->op_parameter_.type_ = PrimitiveType_MulGrad;
+  std::vector<lite::tensor::Tensor *> all_tensors =
+    GenerateTensorsForTest("./test_data/operators/arithmetic_fp32_9_dy_5_4_6.bin", 9);
+
+  std::vector<lite::tensor::Tensor *> inputs = {all_tensors[0], all_tensors[2], all_tensors[1]};
+  std::vector<lite::tensor::Tensor *> outputs = {all_tensors[4], all_tensors[3]};
+  kernel::KernelKey desc = {kernel::kCPU, TypeId::kNumberTypeFloat32, schema::PrimitiveType_MulGrad};
+  auto creator = lite::KernelRegistry::GetInstance()->GetCreator(desc);
+  auto kernel_obj = creator(inputs, outputs, reinterpret_cast<OpParameter *>(param), NULL, desc);
+  kernel_obj->Run();
+
+  float *output_ptr = reinterpret_cast<float *>(outputs[0]->Data());
+  printf("==================output data=================\n");
+  for (int i = 0; i < 6; i++) {
+    std::cout << output_ptr[i] << " ,";
+  }
+  std::cout << std::endl;
+
+  std::string output_path = "./test_data/operators/arithmetic_fp32_9_dx1_5_4_6.bin";
+  EXPECT_EQ(0, lite::CompareRelativeOutput(reinterpret_cast<float *>(outputs[1]->Data()), output_path));
+
+  std::string dx2_path = "./test_data/operators/arithmetic_fp32_9_dx2_5_1_6.bin";
+  EXPECT_EQ(0, lite::CompareRelativeOutput(output_ptr, dx2_path));
+
+  for (int i = 0; i < 5; i++) delete all_tensors[i];
+  delete param;
+  MS_LOG(INFO) << "TestMulGrad4Fp32 passed";
+}
+
+TEST_F(TestArithmeticGradFp32, TestDivGradFp32) {
+  auto param = new ArithmeticParameter();
+  param->op_parameter_.type_ = PrimitiveType_DivGrad;
+  std::vector<lite::tensor::Tensor *> all_tensors =
+    GenerateTensorsForTest("./test_data/operators/arithmetic_fp32_5_dy_4_6.bin", 5);
+
+  std::vector<lite::tensor::Tensor *> inputs = {all_tensors[0], all_tensors[1], all_tensors[2]};
+  std::vector<lite::tensor::Tensor *> outputs = {all_tensors[3], all_tensors[4]};
+  kernel::KernelKey desc = {kernel::kCPU, TypeId::kNumberTypeFloat32, schema::PrimitiveType_DivGrad};
+  auto creator = lite::KernelRegistry::GetInstance()->GetCreator(desc);
+  auto kernel_obj = creator(inputs, outputs, reinterpret_cast<OpParameter *>(param), NULL, desc);
+  kernel_obj->Run();
+
+  float *output_ptr = reinterpret_cast<float *>(outputs[1]->Data());
+  printf("==================output data=================\n");
+  for (int i = 0; i < 6; i++) {
+    std::cout << output_ptr[i] << " ,";
+  }
+  std::cout << std::endl;
+
+  std::string output_path = "./test_data/operators/arithmetic_fp32_5_dx1_4_6.bin";
+  EXPECT_EQ(0, lite::CompareRelativeOutput(reinterpret_cast<float *>(outputs[0]->Data()), output_path));
+
+  std::string dx2_path = "./test_data/operators/arithmetic_fp32_5_dx2_1_6.bin";
+  EXPECT_EQ(0, lite::CompareRelativeOutput(output_ptr, dx2_path));
+
+  for (int i = 0; i < 5; i++) delete all_tensors[i];
+  delete param;
+  MS_LOG(INFO) << "TestDivGradFp32 passed";
+}
+
+TEST_F(TestArithmeticGradFp32, TestDivGrad2Fp32) {
+  auto param = new ArithmeticParameter();
+  param->op_parameter_.type_ = PrimitiveType_DivGrad;
+  std::vector<lite::tensor::Tensor *> all_tensors =
+    GenerateTensorsForTest("./test_data/operators/arithmetic_fp32_6_dy_4_6.bin", 6);
+
+  std::vector<lite::tensor::Tensor *> inputs = {all_tensors[0], all_tensors[2], all_tensors[1]};
+  std::vector<lite::tensor::Tensor *> outputs = {all_tensors[4], all_tensors[3]};
+  kernel::KernelKey desc = {kernel::kCPU, TypeId::kNumberTypeFloat32, schema::PrimitiveType_DivGrad};
+  auto creator = lite::KernelRegistry::GetInstance()->GetCreator(desc);
+  auto kernel_obj = creator(inputs, outputs, reinterpret_cast<OpParameter *>(param), NULL, desc);
+  kernel_obj->Run();
+
+  float *output_ptr = reinterpret_cast<float *>(outputs[0]->Data());
+  printf("==================output data=================\n");
+  for (int i = 0; i < 6; i++) {
+    std::cout << output_ptr[i] << " ,";
+  }
+  std::cout << std::endl;
+
+  std::string dx2_path = "./test_data/operators/arithmetic_fp32_6_dx2_4_6.bin";
+  EXPECT_EQ(0, lite::CompareRelativeOutput(reinterpret_cast<float *>(outputs[1]->Data()), dx2_path));
+
+  std::string output_path = "./test_data/operators/arithmetic_fp32_6_dx1_1_6.bin";
+  EXPECT_EQ(0, lite::CompareRelativeOutput(output_ptr, output_path));
+
+  for (int i = 0; i < 5; i++) delete all_tensors[i];
+  delete param;
+  MS_LOG(INFO) << "TestDivGrad2Fp32 passed";
+}
+
+TEST_F(TestArithmeticGradFp32, TestDivGrad3Fp32) {
+  auto param = new ArithmeticParameter();
+  param->op_parameter_.type_ = PrimitiveType_DivGrad;
+  std::vector<lite::tensor::Tensor *> all_tensors =
+    GenerateTensorsForTest("./test_data/operators/arithmetic_fp32_10_dy_5_4_6.bin", 10);
+
+  std::vector<lite::tensor::Tensor *> inputs = {all_tensors[0], all_tensors[1], all_tensors[2]};
+  std::vector<lite::tensor::Tensor *> outputs = {all_tensors[3], all_tensors[4]};
+  kernel::KernelKey desc = {kernel::kCPU, TypeId::kNumberTypeFloat32, schema::PrimitiveType_DivGrad};
+  auto creator = lite::KernelRegistry::GetInstance()->GetCreator(desc);
+  auto kernel_obj = creator(inputs, outputs, reinterpret_cast<OpParameter *>(param), NULL, desc);
+  kernel_obj->Run();
+
+  float *output_ptr = reinterpret_cast<float *>(outputs[1]->Data());
+  printf("==================output data=================\n");
+  for (int i = 0; i < 6; i++) {
+    std::cout << output_ptr[i] << " ,";
+  }
+  std::cout << std::endl;
+
+  std::string dx1_path = "./test_data/operators/arithmetic_fp32_10_dx1_5_4_6.bin";
+  EXPECT_EQ(0, lite::CompareRelativeOutput(reinterpret_cast<float *>(outputs[0]->Data()), dx1_path));
+
+  std::string output_path = "./test_data/operators/arithmetic_fp32_10_dx2_5_1_6.bin";
+  EXPECT_EQ(0, lite::CompareRelativeOutput(output_ptr, output_path));
+
+  for (int i = 0; i < 5; i++) delete all_tensors[i];
+  delete param;
+  MS_LOG(INFO) << "TestDivGrad3Fp32 passed";
+}
+
+TEST_F(TestArithmeticGradFp32, Test3DDivGrad2Fp32) {
+  auto param = new ArithmeticParameter();
+  param->op_parameter_.type_ = PrimitiveType_DivGrad;
+  std::vector<lite::tensor::Tensor *> all_tensors =
+    GenerateTensorsForTest("./test_data/operators/arithmetic_fp32_7_dy_4_5_6.bin", 7);
+
+  std::vector<lite::tensor::Tensor *> inputs = {all_tensors[0], all_tensors[1], all_tensors[2]};
+  std::vector<lite::tensor::Tensor *> outputs = {all_tensors[3], all_tensors[4]};
+  kernel::KernelKey desc = {kernel::kCPU, TypeId::kNumberTypeFloat32, schema::PrimitiveType_DivGrad};
+  auto creator = lite::KernelRegistry::GetInstance()->GetCreator(desc);
+  auto kernel_obj = creator(inputs, outputs, reinterpret_cast<OpParameter *>(param), NULL, desc);
+  kernel_obj->Run();
+
+  float *output_ptr = reinterpret_cast<float *>(outputs[1]->Data());
+  printf("==================output data=================\n");
+  for (int i = 0; i < 6; i++) {
+    std::cout << output_ptr[i] << " ,";
+  }
+  std::cout << std::endl;
+
+  std::string dx1_path = "./test_data/operators/arithmetic_fp32_7_dx1_4_5_6.bin";
+  EXPECT_EQ(0, lite::CompareRelativeOutput(reinterpret_cast<float *>(outputs[0]->Data()), dx1_path));
+
+  std::string output_path = "./test_data/operators/arithmetic_fp32_7_dx2_1_1_6.bin";
+  EXPECT_EQ(0, lite::CompareRelativeOutput(output_ptr, output_path));
+
+  for (int i = 0; i < 5; i++) delete all_tensors[i];
+  delete param;
+  MS_LOG(INFO) << "TestDivGrad2Fp32 passed";
+}
+
+}  // namespace mindspore

mindspore/lite/test/ut/src/runtime/kernel/arm/fp32/bias_grad_fp32_tests.cc

@@ -0,0 +1,71 @@
+/**
+ * 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 <iostream>
+#include <memory>
+#include "utils/log_adapter.h"
+#include "common/common_test.h"
+#include "src/common/file_utils.h"
+#include "mindspore/lite/src/runtime/kernel/arm/fp32/bias_grad.h"
+#include "mindspore/lite/src/kernel_registry.h"
+
+namespace mindspore {
+
+class TestBiasGradFp32 : public mindspore::Common {
+ public:
+  TestBiasGradFp32() {}
+};
+
+TEST_F(TestBiasGradFp32, BiasGradFp32) {
+  // prepare stage
+  auto bias_param = new ArithmeticParameter();
+
+  size_t input_size;
+  std::string input_path = "./test_data/operators/biasgradfp32_1_dy_10_28_28_7.bin";
+  auto input_data = reinterpret_cast<float *>(mindspore::lite::ReadFile(input_path.c_str(), &input_size));
+  std::vector<int> dim_dy({10, 28, 28, 7});
+  lite::tensor::Tensor dy_tensor(TypeId::kNumberTypeFloat32, dim_dy);
+  dy_tensor.SetData(input_data);
+
+  std::vector<lite::tensor::Tensor *> inputs = {&dy_tensor};
+
+  auto output_data = new float[7];
+  std::vector<int> dim_dw({7});
+  lite::tensor::Tensor dw_tensor(TypeId::kNumberTypeFloat32, dim_dw);
+  dw_tensor.SetData(output_data);
+  std::vector<lite::tensor::Tensor *> outputs = {&dw_tensor};
+
+  kernel::KernelKey desc = {kernel::kCPU, TypeId::kNumberTypeFloat32, schema::PrimitiveType_BiasGrad};
+
+  auto creator = lite::KernelRegistry::GetInstance()->GetCreator(desc);
+  auto kernel_obj = creator(inputs, outputs, reinterpret_cast<OpParameter *>(bias_param), NULL, desc);
+
+  kernel_obj->Run();
+
+  printf("==================output data=================\n");
+  for (int i = 0; i < 7; i++) {
+    std::cout << output_data[i] << " ,";
+  }
+  std::cout << std::endl;
+  std::string output_path = "./test_data/operators/biasgradfp32_1_db_7.bin";
+  lite::CompareOutput(output_data, output_path);
+
+  // delete input_data;
+  // delete[] output_data;
+  delete bias_param;
+  MS_LOG(INFO) << "BiasGradFp32 passed";
+}
+
+}  // namespace mindspore

mindspore/lite/test/ut/src/runtime/kernel/arm/fp32/convolution_grad_fp32_tests.cc

@@ -0,0 +1,521 @@
+/**
+ * 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 <iostream>
+#include <memory>
+#include <vector>
+#include "utils/log_adapter.h"
+#include "common/common_test.h"
+#include "src/common/file_utils.h"
+#include "src/common/file_utils_ext.h"
+#include "mindspore/lite/src/runtime/kernel/arm/fp32/convolution_grad_filter.h"
+#include "mindspore/lite/src/runtime/kernel/arm/fp32/convolution_grad_input.h"
+#include "mindspore/lite/src/runtime/kernel/arm/opclib/conv_parameter.h"
+#include "mindspore/lite/src/kernel_registry.h"
+
+namespace mindspore {
+class TestConvolutionGradFp32 :  public mindspore::Common {
+ public:
+  TestConvolutionGradFp32() {}
+};
+
+void InitConvParamGroup1FP32(ConvParameter *conv_param) {
+  conv_param->input_batch_ = 1;
+  conv_param->input_h_ = 28;
+  conv_param->input_w_ = 28;
+  conv_param->input_channel_ = 3;
+
+  conv_param->output_batch_ = 1;
+  conv_param->output_h_ = 28;
+  conv_param->output_w_ = 28;
+  conv_param->output_channel_ = 32;
+
+  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;
+
+  conv_param->pad_h_ = 1;
+  conv_param->pad_w_ = 1;
+
+  conv_param->group_ = 1;
+  conv_param->is_relu_ = false;
+  conv_param->is_relu6_ = false;
+  conv_param->thread_num_ = 1;
+}
+
+void InitConvParamGroup3FP32(ConvParameter *conv_param) {
+  InitConvParamGroup1FP32(conv_param);
+  conv_param->group_ = 3;
+  conv_param->output_channel_ = 18;
+}
+
+void InitConvParamGroup3Dilation2FP32(ConvParameter *conv_param) {
+  InitConvParamGroup3FP32(conv_param);
+  conv_param->dilation_h_ = 2;
+  conv_param->dilation_w_ = 2;
+  conv_param->output_h_ = 26;
+  conv_param->output_w_ = 26;
+}
+
+TEST_F(TestConvolutionGradFp32, ConvFp32FilterGrad) {
+  // prepare stage
+  auto conv_param = new ConvParameter();
+  InitConvParamGroup1FP32(conv_param);
+
+  size_t dy_size;
+  std::string dy_path = "./test_data/conv/convfp32_dy_1_28_28_32.bin";
+  auto dy_data = reinterpret_cast<float *>(mindspore::lite::ReadFile(dy_path.c_str(), &dy_size));
+  std::vector<int> dim_dy({1, 28, 28, 32});
+  lite::tensor::Tensor dy_tensor(TypeId::kNumberTypeFloat32, dim_dy);
+  dy_tensor.SetData(dy_data);
+
+  // runtime part
+  printf("Calculating runtime cost...\n");
+  uint64_t time_avg = 0;
+  size_t output_data_size =
+    conv_param->output_channel_ * conv_param->kernel_h_ * conv_param->kernel_w_ * conv_param->input_channel_;
+
+  size_t input_size;
+  std::string input_path = "./test_data/conv/convfp32_x_1_28_28_3.bin";
+  auto input_data = reinterpret_cast<float *>(mindspore::lite::ReadFile(input_path.c_str(), &input_size));
+  std::vector<int> dim_x({1, 28, 28, 3});
+  lite::tensor::Tensor x_tensor(TypeId::kNumberTypeFloat32, dim_x);
+  x_tensor.SetData(input_data);
+
+  auto dw_data = new float[output_data_size];
+  std::vector<int> dim_dw({32, 3, 3, 3});
+  lite::tensor::Tensor dw_tensor(TypeId::kNumberTypeFloat32, dim_dw);
+  dw_tensor.SetData(dw_data);
+  std::vector<lite::tensor::Tensor *> inputs = {&dy_tensor, &x_tensor};
+  std::vector<lite::tensor::Tensor *> outputs = {&dw_tensor};
+
+  kernel::KernelKey desc = {kernel::kCPU, TypeId::kNumberTypeFloat32, schema::PrimitiveType_Conv2DGradFilter};
+  auto creator = lite::KernelRegistry::GetInstance()->GetCreator(desc);
+  auto kernel = creator(inputs, outputs, reinterpret_cast<OpParameter *>(conv_param), NULL, desc);
+
+  // warm up loop
+  for (int i = 0; i < 3; i++) {
+    kernel->Run();
+  }
+
+  int loop_count = 100;
+  auto time_start = mindspore::lite::GetTimeUs();
+  for (int i = 0; i < loop_count; i++) {
+    kernel->Run();
+  }
+  auto time_end = mindspore::lite::GetTimeUs();
+  auto cost = time_end - time_start;
+  time_avg = cost / loop_count;
+  printf("single thread running time : %f ms\n", time_avg / 1000.0f);
+
+  std::string output_path = "./test_data/conv/convfp32_dw_32_3_3_3.bin";
+  auto res = lite::CompareRelativeOutput(dw_data, output_path);
+
+  EXPECT_EQ(res, 0);
+
+  // delete input_data;
+  // delete dy_data;
+  // delete [] dw_data;
+  delete kernel;
+  delete conv_param;
+  MS_LOG(INFO) << "TestConvolutionGradFp32 Filter Grad passed";
+}
+
+TEST_F(TestConvolutionGradFp32, ConvFp32InputGrad) {
+  // prepare stage
+  auto conv_param = new ConvParameter();
+  InitConvParamGroup1FP32(conv_param);
+
+  size_t dy_size;
+  std::string dy_path = "./test_data/conv/convfp32_dy_1_28_28_32.bin";
+  auto dy_data = reinterpret_cast<float *>(mindspore::lite::ReadFile(dy_path.c_str(), &dy_size));
+  std::vector<int> dim_dy({1, 28, 28, 32});
+  lite::tensor::Tensor dy_tensor(TypeId::kNumberTypeFloat32, dim_dy);
+  dy_tensor.SetData(dy_data);
+
+  size_t w_size;
+  std::string w_path = "./test_data/conv/convfp32_w_32_3_3_3.bin";
+  auto w_data = reinterpret_cast<float *>(mindspore::lite::ReadFile(w_path.c_str(), &w_size));
+  std::vector<int> dim_dw({32, 3, 3, 3});
+  lite::tensor::Tensor w_tensor(TypeId::kNumberTypeFloat32, dim_dw);
+  w_tensor.SetData(w_data);
+
+  size_t output_data_size =
+    conv_param->input_batch_ * conv_param->input_h_ * conv_param->input_w_ * conv_param->input_channel_;
+  auto dx_data = new float[output_data_size];
+  std::vector<int> dim_dx({1, 28, 28, 3});
+  lite::tensor::Tensor dx_tensor(TypeId::kNumberTypeFloat32, dim_dx);
+  dx_tensor.SetData(dx_data);
+
+  std::vector<lite::tensor::Tensor *> inputs = {&dy_tensor, &w_tensor};
+  std::vector<lite::tensor::Tensor *> outputs = {&dx_tensor};
+  // runtime part
+
+  printf("Calculating runtime cost...\n");
+  uint64_t time_avg = 0;
+
+  kernel::KernelKey desc = {kernel::kCPU, TypeId::kNumberTypeFloat32, schema::PrimitiveType_Conv2DGradInput};
+  auto creator = lite::KernelRegistry::GetInstance()->GetCreator(desc);
+  auto kernel = creator(inputs, outputs, reinterpret_cast<OpParameter *>(conv_param), NULL, desc);
+
+  // warm up loop
+  for (int i = 0; i < 3; i++) {
+    kernel->Run();
+  }
+
+  int loop_count = 100;
+  auto time_start = mindspore::lite::GetTimeUs();
+  for (int i = 0; i < loop_count; i++) {
+    kernel->Run();
+  }
+  auto time_end = mindspore::lite::GetTimeUs();
+  auto cost = time_end - time_start;
+  time_avg = cost / loop_count;
+  printf("single thread running time : %f ms\n", time_avg / 1000.0f);
+
+  std::string output_path = "./test_data/conv/convfp32_dx_1_28_28_3.bin";
+  auto res = lite::CompareRelativeOutput(dx_data, output_path);
+  EXPECT_EQ(res, 0);
+
+  delete kernel;
+  delete conv_param;
+  MS_LOG(INFO) << "TestConvolutionGradFp32 Filter Grad passed";
+}
+
+TEST_F(TestConvolutionGradFp32, ConvFp32GroupFilterGrad) {
+  // prepare stage
+  auto conv_param = new ConvParameter();
+  InitConvParamGroup3FP32(conv_param);
+
+  size_t dy_size;
+  std::string dy_path = "./test_data/conv/convfp32_dy_g3_1_28_28_18.bin";
+  auto dy_data = reinterpret_cast<float *>(mindspore::lite::ReadFile(dy_path.c_str(), &dy_size));
+  std::vector<int> dim_dy({1, 28, 28, 18});
+  lite::tensor::Tensor dy_tensor(TypeId::kNumberTypeFloat32, dim_dy);
+  dy_tensor.SetData(dy_data);
+
+  // runtime part
+  printf("Calculating runtime cost...\n");
+  uint64_t time_avg = 0;
+  size_t output_data_size = conv_param->output_channel_ * conv_param->kernel_h_ * conv_param->kernel_w_ *
+                            conv_param->input_channel_ / conv_param->group_;
+
+  size_t input_size;
+  std::string input_path = "./test_data/conv/convfp32_x_g3_1_28_28_3.bin";
+  auto input_data = reinterpret_cast<float *>(mindspore::lite::ReadFile(input_path.c_str(), &input_size));
+  std::vector<int> dim_x({1, 28, 28, 3});
+  lite::tensor::Tensor x_tensor(TypeId::kNumberTypeFloat32, dim_x);
+  x_tensor.SetData(input_data);
+
+  auto dw_data = new float[output_data_size];
+  std::vector<int> dim_dw({18, 3, 3, 1});
+  lite::tensor::Tensor dw_tensor(TypeId::kNumberTypeFloat32, dim_dw);
+  dw_tensor.SetData(dw_data);
+  std::vector<lite::tensor::Tensor *> inputs = {&dy_tensor, &x_tensor};
+  std::vector<lite::tensor::Tensor *> outputs = {&dw_tensor};
+
+  kernel::KernelKey desc = {kernel::kCPU, TypeId::kNumberTypeFloat32, schema::PrimitiveType_Conv2DGradFilter};
+  auto creator = lite::KernelRegistry::GetInstance()->GetCreator(desc);
+  auto kernel = creator(inputs, outputs, reinterpret_cast<OpParameter *>(conv_param), NULL, desc);
+
+  // warm up loop
+  for (int i = 0; i < 3; i++) {
+    kernel->Run();
+  }
+
+  int loop_count = 100;
+  auto time_start = mindspore::lite::GetTimeUs();
+  for (int i = 0; i < loop_count; i++) {
+    kernel->Run();
+  }
+  auto time_end = mindspore::lite::GetTimeUs();
+  auto cost = time_end - time_start;
+  time_avg = cost / loop_count;
+  printf("single thread running time : %f ms\n", time_avg / 1000.0f);
+
+  std::string output_path = "./test_data/conv/convfp32_dw_g3_18_3_3_3.bin";
+  auto res = lite::CompareRelativeOutput(dw_data, output_path);
+  EXPECT_EQ(res, 0);
+
+  // delete input_data;
+  // delete dy_data;
+  // delete [] dw_data;
+  delete kernel;
+  delete conv_param;
+  MS_LOG(INFO) << "TestConvolutionGradFp32 Filter Grad passed";
+}
+
+TEST_F(TestConvolutionGradFp32, ConvFp32GroupInputGrad) {
+  // prepare stage
+  auto conv_param = new ConvParameter();
+  InitConvParamGroup3FP32(conv_param);
+
+  size_t dy_size;
+  std::string dy_path = "./test_data/conv/convfp32_dy_g3_1_28_28_18.bin";
+  auto dy_data = reinterpret_cast<float *>(mindspore::lite::ReadFile(dy_path.c_str(), &dy_size));
+  std::vector<int> dim_dy({1, 28, 28, 18});
+  lite::tensor::Tensor dy_tensor(TypeId::kNumberTypeFloat32, dim_dy);
+  dy_tensor.SetData(dy_data);
+
+  size_t w_size;
+  std::string w_path = "./test_data/conv/convfp32_w_g3_18_3_3_3.bin";
+  auto w_data = reinterpret_cast<float *>(mindspore::lite::ReadFile(w_path.c_str(), &w_size));
+  std::vector<int> dim_dw({18, 3, 3, 1});
+  lite::tensor::Tensor w_tensor(TypeId::kNumberTypeFloat32, dim_dw);
+  w_tensor.SetData(w_data);
+
+  size_t output_data_size =
+    conv_param->input_batch_ * conv_param->input_h_ * conv_param->input_w_ * conv_param->input_channel_;
+  auto dx_data = new float[output_data_size];
+  std::vector<int> dim_dx({1, 28, 28, 3});
+  lite::tensor::Tensor dx_tensor(TypeId::kNumberTypeFloat32, dim_dx);
+  dx_tensor.SetData(dx_data);
+
+  std::vector<lite::tensor::Tensor *> inputs = {&dy_tensor, &w_tensor};
+  std::vector<lite::tensor::Tensor *> outputs = {&dx_tensor};
+  // runtime part
+
+  printf("Calculating runtime cost...\n");
+  uint64_t time_avg = 0;
+
+  kernel::KernelKey desc = {kernel::kCPU, TypeId::kNumberTypeFloat32, schema::PrimitiveType_Conv2DGradInput};
+  auto creator = lite::KernelRegistry::GetInstance()->GetCreator(desc);
+  auto kernel = creator(inputs, outputs, reinterpret_cast<OpParameter *>(conv_param), NULL, desc);
+
+  // warm up loop
+  for (int i = 0; i < 3; i++) {
+    kernel->Run();
+  }
+
+  int loop_count = 100;
+  auto time_start = mindspore::lite::GetTimeUs();
+  for (int i = 0; i < loop_count; i++) {
+    kernel->Run();
+  }
+  auto time_end = mindspore::lite::GetTimeUs();
+  auto cost = time_end - time_start;
+  time_avg = cost / loop_count;
+  printf("single thread running time : %f ms\n", time_avg / 1000.0f);
+
+  std::string output_path = "./test_data/conv/convfp32_dx_g3_1_28_28_3.bin";
+  auto res = lite::CompareRelativeOutput(dx_data, output_path);
+  EXPECT_EQ(res, 0);
+
+  delete kernel;
+  delete conv_param;
+  MS_LOG(INFO) << "TestConvolutionGradFp32 Filter Grad passed";
+}
+
+TEST_F(TestConvolutionGradFp32, ConvFp32GroupDilationFilterGrad) {
+  // prepare stage
+  auto conv_param = new ConvParameter();
+
+  InitConvParamGroup3Dilation2FP32(conv_param);
+
+  size_t dy_size;
+  std::string dy_path = "./test_data/conv/convfp32_dy_g3_d2_1_26_26_18.bin";
+  auto dy_data = reinterpret_cast<float *>(mindspore::lite::ReadFile(dy_path.c_str(), &dy_size));
+  std::vector<int> dim_dy({1, 26, 26, 18});
+  lite::tensor::Tensor dy_tensor(TypeId::kNumberTypeFloat32, dim_dy);
+  dy_tensor.SetData(dy_data);
+
+  // runtime part
+  printf("Calculating runtime cost...\n");
+  uint64_t time_avg = 0;
+  size_t output_data_size = conv_param->output_channel_ * conv_param->kernel_h_ * conv_param->kernel_w_ *
+                            conv_param->input_channel_ / conv_param->group_;
+
+  size_t input_size;
+  std::string input_path = "./test_data/conv/convfp32_x_g3_d2_1_28_28_3.bin";
+  auto input_data = reinterpret_cast<float *>(mindspore::lite::ReadFile(input_path.c_str(), &input_size));
+  std::vector<int> dim_x({1, 28, 28, 3});
+  lite::tensor::Tensor x_tensor(TypeId::kNumberTypeFloat32, dim_x);
+  x_tensor.SetData(input_data);
+
+  auto dw_data = new float[output_data_size];
+  std::vector<int> dim_dw({18, 3, 3, 1});
+  lite::tensor::Tensor dw_tensor(TypeId::kNumberTypeFloat32, dim_dw);
+  dw_tensor.SetData(dw_data);
+  std::vector<lite::tensor::Tensor *> inputs = {&dy_tensor, &x_tensor};
+  std::vector<lite::tensor::Tensor *> outputs = {&dw_tensor};
+
+  kernel::KernelKey desc = {kernel::kCPU, TypeId::kNumberTypeFloat32, schema::PrimitiveType_Conv2DGradFilter};
+  auto creator = lite::KernelRegistry::GetInstance()->GetCreator(desc);
+  auto kernel = creator(inputs, outputs, reinterpret_cast<OpParameter *>(conv_param), NULL, desc);
+
+  // warm up loop
+  for (int i = 0; i < 3; i++) {
+    kernel->Run();
+  }
+
+  int loop_count = 100;
+  auto time_start = mindspore::lite::GetTimeUs();
+  for (int i = 0; i < loop_count; i++) {
+    kernel->Run();
+  }
+  auto time_end = mindspore::lite::GetTimeUs();
+  auto cost = time_end - time_start;
+  time_avg = cost / loop_count;
+  printf("single thread running time : %f ms\n", time_avg / 1000.0f);
+
+  std::string output_path = "./test_data/conv/convfp32_dw_g3_d2_18_3_3_3.bin";
+  auto res = lite::CompareRelativeOutput(dw_data, output_path);
+  EXPECT_EQ(res, 0);
+  // delete input_data;
+  // delete dy_data;
+  // delete [] dw_data;
+  delete kernel;
+  delete conv_param;
+  MS_LOG(INFO) << "TestConvolutionGradFp32 Filter Grad passed";
+}
+
+TEST_F(TestConvolutionGradFp32, ConvFp32GroupDilationInputGrad) {
+  // prepare stage
+  auto conv_param = new ConvParameter();
+  InitConvParamGroup3Dilation2FP32(conv_param);
+
+  size_t dy_size;
+  std::string dy_path = "./test_data/conv/convfp32_dy_g3_d2_1_26_26_18.bin";
+  auto dy_data = reinterpret_cast<float *>(mindspore::lite::ReadFile(dy_path.c_str(), &dy_size));
+  std::vector<int> dim_dy({1, 26, 26, 18});
+  lite::tensor::Tensor dy_tensor(TypeId::kNumberTypeFloat32, dim_dy);
+  dy_tensor.SetData(dy_data);
+
+  size_t w_size;
+  std::string w_path = "./test_data/conv/convfp32_w_g3_d2_18_3_3_3.bin";
+  auto w_data = reinterpret_cast<float *>(mindspore::lite::ReadFile(w_path.c_str(), &w_size));
+  std::vector<int> dim_w({18, 3, 3, 1});
+  lite::tensor::Tensor w_tensor(TypeId::kNumberTypeFloat32, dim_w);
+  w_tensor.SetData(w_data);
+
+  size_t output_data_size =
+    conv_param->input_batch_ * conv_param->input_h_ * conv_param->input_w_ * conv_param->input_channel_;
+  auto dx_data = new float[output_data_size];
+  std::vector<int> dim_dx({1, 28, 28, 3});
+  lite::tensor::Tensor dx_tensor(TypeId::kNumberTypeFloat32, dim_dx);
+  dx_tensor.SetData(dx_data);
+
+  std::vector<lite::tensor::Tensor *> inputs = {&dy_tensor, &w_tensor};
+  std::vector<lite::tensor::Tensor *> outputs = {&dx_tensor};
+  // runtime part
+
+  printf("Calculating runtime cost...\n");
+  uint64_t time_avg = 0;
+
+  kernel::KernelKey desc = {kernel::kCPU, TypeId::kNumberTypeFloat32, schema::PrimitiveType_Conv2DGradInput};
+  auto creator = lite::KernelRegistry::GetInstance()->GetCreator(desc);
+  auto kernel = creator(inputs, outputs, reinterpret_cast<OpParameter *>(conv_param), NULL, desc);
+
+  // warm up loop
+  for (int i = 0; i < 3; i++) {
+    kernel->Run();
+  }
+
+  int loop_count = 100;
+  auto time_start = mindspore::lite::GetTimeUs();
+  for (int i = 0; i < loop_count; i++) {
+    kernel->Run();
+  }
+  auto time_end = mindspore::lite::GetTimeUs();
+  auto cost = time_end - time_start;
+  time_avg = cost / loop_count;
+  printf("single thread running time : %f ms\n", time_avg / 1000.0f);
+
+  std::string output_path = "./test_data/conv/convfp32_dx_g3_d2_1_28_28_3.bin";
+  auto res = lite::CompareRelativeOutput(dx_data, output_path);
+  EXPECT_EQ(res, 0);
+
+  delete kernel;
+  delete conv_param;
+  MS_LOG(INFO) << "TestConvolutionGradFp32 Filter Grad passed";
+}
+
+// TEST_F(TestConvolutionGradFp32, ConvGroupDilation) {
+//   // prepare stage
+//   auto conv_param = new ConvParameter();
+//   InitConvParamGroup3Dilation2FP32(conv_param);
+
+//   size_t x_size;
+//   std::string x_path = "./test_data/conv/convfp32_x_g3_d2_1_28_28_3.bin";
+//   auto x_data = reinterpret_cast<float *>(mindspore::lite::ReadFile(x_path.c_str(), &x_size));
+//   std::vector<int> dim_x({1, 28, 28, 3});
+//   tensor::Tensor x_tensor(TypeId::kNumberTypeFloat32, dim_x);
+//   x_tensor.SetData(x_data);
+
+//   size_t w_size;
+//   std::string w_path = "./test_data/conv/convfp32_w_g3_d2_18_3_3_3.bin";
+//   auto w_data = reinterpret_cast<float *>(mindspore::lite::ReadFile(w_path.c_str(), &w_size));
+//   std::vector<int> dim_w({18, 3, 3, 1});
+//   tensor::Tensor w_tensor(TypeId::kNumberTypeFloat32, dim_w);
+//   w_tensor.SetData(w_data);
+
+//   size_t output_data_size =
+//     conv_param->output_batch_ * conv_param->output_h_ * conv_param->output_w_ * conv_param->output_channel_;
+//   auto y_data = new float[output_data_size];
+//   std::vector<int> dim_y({1, 26, 26, 18});
+//   tensor::Tensor y_tensor(TypeId::kNumberTypeFloat32, dim_y);
+//   y_tensor.SetData(y_data);
+
+//   std::vector<tensor::Tensor *> inputs = {&x_tensor, &w_tensor};
+//   std::vector<tensor::Tensor *> outputs = {&y_tensor};
+//   // runtime part
+
+//   printf("Calculating runtime cost...\n");
+//   uint64_t time_avg = 0;
+
+//   lite::Context context;
+//   ;
+//   context.deviceCtx.type = lite::DT_CPU;
+//   context.threadNum = 1;
+
+//   kernel::KernelKey desc = {kernel::kCPU, kNumberTypeFloat32, schema::PrimitiveType_Conv2D};
+//   auto creator = lite::KernelRegistry::GetInstance()->GetKernelCreator(desc);
+//   auto kernel = creator(inputs, outputs, (OpParameter *)conv_param, &context, desc);
+
+//   kernel->train();
+//   EXPECT_EQ(kernel->is_train(), 1);
+
+//   // warm up loop
+//   for (int i = 0; i < 3; i++) {
+//     kernel->Run();
+//   }
+
+//   int loop_count = 100;
+//   auto time_start = mindspore::lite::GetTimeUs();
+//   for (int i = 0; i < loop_count; i++) {
+//     kernel->Run();
+//   }
+//   auto time_end = mindspore::lite::GetTimeUs();
+//   auto cost = time_end - time_start;
+//   time_avg = cost / loop_count;
+//   printf("single thread running time : %f ms\n", time_avg / 1000.0f);
+
+//   std::string output_path = "./test_data/conv/convfp32_y_g3_d2_1_26_26_18.bin";
+//   auto res = lite::CompareRelativeOutput(y_data, output_path);
+//   EXPECT_EQ(res, 0);
+
+//   delete kernel;
+//   delete conv_param;
+
+//   MS_LOG(INFO) << "TestConvolutionFp32 Filter Grad passed";
+// }
+
+}  // namespace mindspore

mindspore/lite/test/ut/src/runtime/kernel/arm/fp32/pooling_grad_fp32_tests.cc

@@ -0,0 +1,332 @@
+/**
+ * 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 <iostream>
+#include <memory>
+#include "mindspore/lite/include/context.h"
+#include "utils/log_adapter.h"
+#include "common/common_test.h"
+#include "mindspore/lite/src/kernel_registry.h"
+#include "src/common/utils.h"
+#include "src/common/file_utils.h"
+#include "src/runtime/kernel/arm/fp32/pooling_grad.h"
+#include "src/runtime/kernel/arm/opclib/fp32/pooling_grad.h"
+
+namespace mindspore {
+class TestPoolingGradFp32 :  public mindspore::Common {
+ public:
+  TestPoolingGradFp32() {}
+};
+
+void InitPoolingParamFP32(PoolingParameter *pooling_param) {
+  pooling_param->input_batch_ = 1;
+  pooling_param->input_h_ = 28;
+  pooling_param->input_w_ = 28;
+  pooling_param->input_channel_ = 3;
+
+  pooling_param->output_batch_ = 1;
+  pooling_param->output_h_ = 28;
+  pooling_param->output_w_ = 28;
+  pooling_param->output_channel_ = 32;
+
+  pooling_param->window_h_ = 3;
+  pooling_param->window_w_ = 3;
+
+  pooling_param->stride_h_ = 1;
+  pooling_param->stride_w_ = 1;
+
+  pooling_param->pad_u_ = 1;
+  pooling_param->pad_d_ = 1;
+  pooling_param->pad_l_ = 1;
+  pooling_param->pad_r_ = 1;
+  pooling_param->thread_num_ = 1;
+}
+
+TEST_F(TestPoolingGradFp32, AvgPoolingGradFp32) {
+  // prepare stage
+  auto pooling_param = new PoolingParameter();
+  InitPoolingParamFP32(pooling_param);
+  pooling_param->output_channel_ = 3;
+
+  // runtime part
+  printf("Calculating runtime cost...\n");
+  uint64_t time_avg = 0;
+  size_t output_data_size =
+    pooling_param->output_batch_ * pooling_param->output_channel_ * pooling_param->output_h_ * pooling_param->output_w_;
+
+  size_t input_size;
+  std::string input_path = "./test_data/pooling/avgpoolgradfp32_1_dy_1_28_28_3.bin";
+  auto input_data = reinterpret_cast<float *>(mindspore::lite::ReadFile(input_path.c_str(), &input_size));
+
+  auto output_data = new float[output_data_size];
+  // warm up loop
+  for (int i = 0; i < 3; i++) {
+    AvgPoolingGrad(input_data, output_data, pooling_param);
+  }
+
+  int loop_count = 100;
+  auto time_start = mindspore::lite::GetTimeUs();
+  for (int i = 0; i < loop_count; i++) {
+    AvgPoolingGrad(input_data, output_data, pooling_param);
+  }
+  auto time_end = mindspore::lite::GetTimeUs();
+  auto cost = time_end - time_start;
+  time_avg = cost / loop_count;
+  printf("single thread running time : %f ms\n", time_avg / 1000.0f);
+
+  printf("==================output data=================\n");
+  for (int i = 0; i < 20; i++) {
+    std::cout << output_data[i] << " ,";
+  }
+  std::cout << std::endl;
+  std::string output_path = "./test_data/pooling/avgpoolgradfp32_1_dx_1_28_28_3.bin";
+  lite::CompareOutput(output_data, output_path);
+
+  delete input_data;
+  delete[] output_data;
+  delete pooling_param;
+  MS_LOG(INFO) << "TestAvgPoolingGradFp32 passed";
+}
+
+TEST_F(TestPoolingGradFp32, AvgPoolingKernelGradFp32) {
+  // prepare stage
+  auto pooling_param = new PoolingParameter();
+  InitPoolingParamFP32(pooling_param);
+
+  pooling_param->output_channel_ = 3;
+
+  // runtime part
+  printf("Calculating runtime cost...\n");
+  // uint64_t time_avg = 0;
+  size_t output_data_size =
+    pooling_param->output_batch_ * pooling_param->output_channel_ * pooling_param->output_h_ * pooling_param->output_w_;
+
+  size_t input_size;
+  std::string input_path = "./test_data/pooling/avgpoolgradfp32_1_dy_1_28_28_3.bin";
+  auto input_data = reinterpret_cast<float *>(mindspore::lite::ReadFile(input_path.c_str(), &input_size));
+  std::vector<int> dim_dy({1, 28, 28, 3});
+  lite::tensor::Tensor dy_tensor(TypeId::kNumberTypeFloat32, dim_dy);
+  dy_tensor.SetData(input_data);
+
+  std::string input1_path = "./test_data/pooling/avgpoolgradfp32_1_x_1_28_28_3.bin";
+  input_data = reinterpret_cast<float *>(mindspore::lite::ReadFile(input1_path.c_str(), &input_size));
+  std::vector<int> dim_x({1, 28, 28, 3});
+  lite::tensor::Tensor x_tensor(TypeId::kNumberTypeFloat32, dim_x);
+  x_tensor.SetData(input_data);
+
+  std::vector<lite::tensor::Tensor *> inputs = {&dy_tensor, &x_tensor};
+
+  auto output_data = new float[output_data_size];
+  std::vector<int> dim_dx({1, 28, 28, 3});
+  lite::tensor::Tensor dx_tensor(TypeId::kNumberTypeFloat32, dim_dx);
+  dx_tensor.SetData(output_data);
+  std::vector<lite::tensor::Tensor *> outputs = {&dx_tensor};
+
+  kernel::KernelKey desc = {kernel::kCPU, TypeId::kNumberTypeFloat32, schema::PrimitiveType_PoolingGrad};
+
+  auto creator = lite::KernelRegistry::GetInstance()->GetCreator(desc);
+  auto kernel_obj = creator(inputs, outputs, reinterpret_cast<OpParameter *>(pooling_param), NULL, desc);
+
+  kernel_obj->Run();
+
+  printf("==================output data=================\n");
+  for (int i = 0; i < 20; i++) {
+    std::cout << output_data[i] << " ,";
+  }
+  std::cout << std::endl;
+  std::string output_path = "./test_data/pooling/avgpoolgradfp32_1_dx_1_28_28_3.bin";
+  lite::CompareOutput(output_data, output_path);
+
+  // delete input_data;
+  // delete[] output_data;
+  delete pooling_param;
+  MS_LOG(INFO) << "TestAvgPoolingGradFp32 passed";
+}
+
+TEST_F(TestPoolingGradFp32, MaxPoolingGradFp32) {
+  // prepare stage
+  auto pooling_param = new PoolingParameter();
+  InitPoolingParamFP32(pooling_param);
+  pooling_param->output_channel_ = 3;
+  pooling_param->avg_pooling_ = false;
+  pooling_param->max_pooling_ = true;
+  // runtime part
+  printf("Calculating runtime cost...\n");
+  uint64_t time_avg = 0;
+  size_t output_data_size =
+    pooling_param->output_batch_ * pooling_param->output_channel_ * pooling_param->output_h_ * pooling_param->output_w_;
+
+  size_t input_size;
+  std::string i_path = "./test_data/pooling/maxpoolgradfp32_1_i_1_28_28_3.bin";
+  auto ill_data = reinterpret_cast<int64_t *>(mindspore::lite::ReadFile(i_path.c_str(), &input_size));
+  auto i_data = new int[output_data_size];
+  for (uint32_t i = 0; i < output_data_size; i++) {
+    i_data[i] = static_cast<int>(ill_data[i]);
+  }
+
+  std::string dy_path = "./test_data/pooling/maxpoolgradfp32_1_dy_1_28_28_3.bin";
+  auto dy_data = reinterpret_cast<float *>(mindspore::lite::ReadFile(dy_path.c_str(), &input_size));
+
+  auto output_data = new float[output_data_size];
+  // warm up loop
+  for (int i = 0; i < 3; i++) {
+    MaxPoolingGrad(dy_data, i_data, output_data, pooling_param);
+  }
+
+  int loop_count = 100;
+  auto time_start = mindspore::lite::GetTimeUs();
+  for (int i = 0; i < loop_count; i++) {
+    MaxPoolingGrad(dy_data, i_data, output_data, pooling_param);
+  }
+  auto time_end = mindspore::lite::GetTimeUs();
+  auto cost = time_end - time_start;
+  time_avg = cost / loop_count;
+  printf("single thread running time : %f ms\n", time_avg / 1000.0f);
+
+  printf("==================output data=================\n");
+  for (int i = 0; i < 20; i++) {
+    std::cout << output_data[i] << " ,";
+  }
+  std::cout << std::endl;
+  std::string output_path = "./test_data/pooling/maxpoolgradfp32_1_dx_1_28_28_3.bin";
+  lite::CompareOutput(output_data, output_path);
+
+  // delete input_data;
+  delete pooling_param;
+  delete[] output_data;
+  MS_LOG(INFO) << "TestMaxPoolingGradFp32 passed";
+}
+
+#if 0
+TEST_F(TestPoolingGradFp32, MaxPoolingKernelGradFp32) {
+  // prepare stage
+  auto maxpool = new PoolingParameter();
+  InitPoolingParamFP32(maxpool);
+  maxpool->avg_pooling_ = false;
+  maxpool->max_pooling_ = true;
+  maxpool->input_h_ = 30;
+  maxpool->input_w_ = 30;
+  maxpool->input_channel_ = 3;
+
+  maxpool->output_batch_ = 1;
+  maxpool->output_h_ = 10;
+  maxpool->output_w_ = 10;
+  maxpool->output_channel_ = 3;
+  maxpool->stride_h_ = 3;
+  maxpool->stride_w_ = 3;
+
+  maxpool->pad_u_ = 0;
+  maxpool->pad_d_ = 0;
+  maxpool->pad_l_ = 0;
+  maxpool->pad_r_ = 0;
+
+  size_t input_size;
+  size_t y_data_size = maxpool->output_batch_ * maxpool->output_channel_ * maxpool->output_h_ * maxpool->output_w_;
+
+  auto x_data = reinterpret_cast<float *>(
+    mindspore::lite::ReadFile("./test_data/pooling/maxpoolgradfp32_2_x_1_30_30_3.bin", &input_size));
+  std::vector<int> dim_x({1, 30, 30, 3});
+  lite::tensor::Tensor x_tensor(TypeId::kNumberTypeFloat32, dim_x);
+  x_tensor.SetData(x_data);
+  std::vector<lite::tensor::Tensor *> maxpool_inputs = {&x_tensor};
+
+  auto y_data = new float[y_data_size];
+  std::vector<int> dim_y({1, 10, 10, 3});
+  lite::tensor::Tensor y_tensor(TypeId::kNumberTypeFloat32, dim_y);
+  y_tensor.SetData(y_data);
+
+  auto ind_data = new int[y_data_size];
+  lite::tensor::Tensor ind_tensor(TypeId::kNumberTypeInt32, dim_y);
+  ind_tensor.SetData(ind_data);
+
+  std::vector<lite::tensor::Tensor *> maxpool_outputs = {&y_tensor, &ind_tensor};
+
+  kernel::KernelKey maxpool_desc = {kernel::kCPU, TypeId::kNumberTypeFloat32, schema::PrimitiveType_Pooling};
+  auto maxpool_creator = lite::KernelRegistry::GetInstance()->GetCreator(maxpool_desc);
+  auto maxpoolobj = maxpool_creator(maxpool_inputs, maxpool_outputs, reinterpret_cast<OpParameter *>(maxpool),
+                                    NULL, maxpool_desc);
+  maxpoolobj->Run();
+
+  printf("==================indices data=================\n");
+  for (int i = 0; i < 10; i++) {
+    std::cout << ind_data[i] << " ,";
+  }
+  std::cout << std::endl;
+
+  auto pooling_param = new PoolingParameter();
+  InitPoolingParamFP32(pooling_param);
+  pooling_param->avg_pooling_ = false;
+  pooling_param->max_pooling_ = true;
+  pooling_param->input_h_ = 10;
+  pooling_param->input_w_ = 10;
+  pooling_param->input_channel_ = 3;
+
+  pooling_param->output_batch_ = 1;
+  pooling_param->output_h_ = 30;
+  pooling_param->output_w_ = 30;
+  pooling_param->output_channel_ = 3;
+
+  // runtime part
+  printf("Calculating runtime cost...\n");
+  // uint64_t time_avg = 0;
+  size_t output_data_size =
+    pooling_param->output_batch_ * pooling_param->output_channel_ * pooling_param->output_h_ * pooling_param->output_w_;
+
+  auto dy_data = reinterpret_cast<float *>(
+    mindspore::lite::ReadFile("./test_data/pooling/maxpoolgradfp32_2_dy_1_10_10_3.bin", &input_size));
+  std::vector<int> dim_dy({1, 3, 10, 10});
+  lite::tensor::Tensor dy_tensor(TypeId::kNumberTypeFloat32, dim_dy);
+  dy_tensor.SetData(dy_data);
+
+#if 0
+  std::string i_path = "./test_data/pooling/maxpoolgradfp32_2_i_1_3_10_10.bin";
+  auto ill_data = reinterpret_cast<int64_t*>(mindspore::lite::ReadFile(i_path.c_str(), &input_size));
+  auto i_data = new int[output_data_size];
+  for (int i=0; i < output_data_size; i++)
+    i_data[i] = static_cast<int>(ill_data[i]);
+  std::vector<int> dim_ind({1, 3, 10, 10});
+  lite::tensor::Tensor ind_tensor(TypeId::kNumberTypeInt32, dim_ind);
+  ind_tensor.SetData(i_data);
+#endif
+
+  std::vector<lite::tensor::Tensor *> inputs = {&dy_tensor, &ind_tensor};
+
+  auto output_data = new float[output_data_size];
+  std::vector<int> dim_dx({1, 3, 30, 30});
+  lite::tensor::Tensor dx_tensor(TypeId::kNumberTypeFloat32, dim_dx);
+  dx_tensor.SetData(output_data);
+  std::vector<lite::tensor::Tensor *> outputs = {&dx_tensor};
+
+  kernel::KernelKey desc = {kernel::kCPU, TypeId::kNumberTypeFloat32, schema::PrimitiveType_PoolingGrad};
+  auto creator = lite::KernelRegistry::GetInstance()->GetCreator(desc);
+  auto kernel_obj = creator(inputs, outputs, reinterpret_cast<OpParameter *>(pooling_param), NULL, desc);
+  kernel_obj->Run();
+
+  printf("==================output data=================\n");
+  for (int i = 0; i < 20; i++) {
+    std::cout << output_data[i] << " ,";
+  }
+  std::cout << std::endl;
+  std::string output_path = "./test_data/pooling/maxpoolgradfp32_2_dx_1_30_30_3.bin";
+  lite::CompareOutput(output_data, output_path);
+
+  // delete input_data;
+  // delete[] output_data;
+  delete pooling_param;
+  MS_LOG(INFO) << "TestMaxPoolingKernelGradFp32 passed";
+}
+#endif  // if 0 before MaxPoolingKernelGradFp32
+}  // namespace mindspore

mindspore/lite/test/ut/src/runtime/kernel/arm/test_data/activationGrad/hsigmoid_out_50.bin

@@ -0,0 +1 @@
+ "x>#Ď>K9�>pR
>)J	>¤á4>™K>¤ĹZ>ńß>ŢÝ>ńńL>‚µ=ËQ>Ń*^>MÖ>&¶>6>Sş>đ*�>ÉN>Ë-ý=Ó+L>ÜvK>+A}>wě^>$ďQ>´Ús>ł/W>ó×Ď=Mţ'>9[*>#%†<#�>CÖ>>ˇ‚>$ÁŻ=Gţj>ňě>Ă7*>´Ă2>łĆ6>•™>ń1p>ős#>Y)>çôk>9ď÷=´ŘŔ=lQ0>ű—w>

mindspore/lite/test/ut/src/runtime/kernel/arm/test_data/activationGrad/hsigmoid_x_50.bin

@@ -0,0 +1 @@
+èM®?Ç·ú¾Ô
å?	
¿H2¿|Ý7>0á?dyX?C�.¿\fT¾¼@?ªÍ³¿Öö¾­Àg?Lwñ¾«˜Å¾E�¾Š9&¿7AÎ?†T?öXF¿4Å?â–?žÒ¹?(k?´0?¬¤?¤VH?-–¿Tz@½&À²»Ç"-ÀÞ1¿£wñ¾šÕË?·Fº¿¼�?�«ç¾¢D
¼¶Â>’øY>ãÌ¿_pœ?ÄØ
¾í]ç¼	ç’?À%R¿5§¿KsË=ó­?

mindspore/lite/test/ut/src/runtime/kernel/arm/test_data/activationGrad/hsigmoid_yt_50.bin

@@ -0,0 +1 @@
+Ь╨?╢6V?ЯUл?╗ШS?=ОM?;╘┤?3P≤?;╓?ИоE?мLn?u╣≥?▐!?╠ЗV?═╕?sаW?9_?яe?}≈H?hюд?▌ ?XБ=?ч ≥?%≥≤?ЮП╫?Y1╖?[s²?Д╤?фc║?ЖА?tЩ{?у┬?╣7и=╣DK?eаW?щЯц?шп?╣>╟?kcY?╓S?г├?┴?_fQ?u%╢?П-u?≈}?╜В╟?kС9?┤╒?=└?Э╠╧?

mindspore/lite/test/ut/src/runtime/kernel/arm/test_data/activationGrad/hswish_out_50.bin

@@ -0,0 +1 @@
+v╚≈=qиы╫ьBs>Эл╬лй╬─вQ=@ъ<U█и?P2╪4[Ф?

mindspore/lite/test/ut/src/runtime/kernel/arm/test_data/activationGrad/hswish_x_50.bin

@@ -0,0 +1 @@
+/�ηΏΦ�ƒΏ™5±>Β"Ώ†¥\Ώ��Ώ=`ΞΏώ;�?σε»Ώ�¥©?

mindspore/lite/test/ut/src/runtime/kernel/arm/test_data/activationGrad/lrelu_y_50.bin

@@ -0,0 +1 @@
+грНїЖ&Н5SaНГЏА?tЩ?Џ�@WЕ,М�СО2еН&�8?;VММ�В?цЉЁ?$х�?5ЮљМpNђМъF7О:�Ё?�5VН:Ю�?їm
Нѕ,!@фљО`|в>VўЕНи�№М
?_бB?0дНвУ"?уЬяН�!>%џ=Ћ�,?��Мѓ>Йа�?�;?ъqЙНGh�?7џЩ<бНР�U>=дх?р-Нaыp?Ђеg?��й>Цr@XА>

mindspore/lite/test/ut/src/runtime/kernel/arm/test_data/activationGrad/sigmoid_out_50.bin

@@ -0,0 +1 @@
+âË]>Eí>òJn>bK>Œ8£=
Š<¾P—>&”“>óg>]±Ð<WBX=S¯t>ä;�>ã¤Q>I�>¢¦º=\ƒ>ºéS>	Å€=äC*>šK=
ën>IyŠ>„Š“>¾l‚>/=—>rp>Ÿ‹”>ü«>Ž(	>ûÁ[>-ï–>{ëj=Ç4’>C¾”>«eŽ>D”>B“=ü£=x”>/m�>v¾j>P–>Ävï=PÊ•>“·=�3>vN=œ �>Ó‚—>

mindspore/lite/test/ut/src/runtime/kernel/arm/test_data/activationGrad/sigmoid_y_50.bin

@@ -0,0 +1 @@
+wá>õ“¢>XOï>?áÑ>ƒ�h>é¿Å=¸¡%?±o:?9¸é>"qö=ù¨7>‡ õ>?
?µ ×>9Œ?Á´{>št?D2\?½J>nž¸>°�1>ÈÞï>�OF?/Ç?7y?J0?ÍeT?A?F$¦>'¾Ÿ>Abß>¹Œ#?"m@>Ë<?›8?ï?•?ýÑZ>—$i>­8?à*C?�)ì>rï3?ᆒ>óX?9y>¿ª¿>^ó2>SÔ??w!'?

mindspore/lite/test/ut/src/runtime/kernel/arm/test_data/activationGrad/sigmoid_yt_50.bin

@@ -0,0 +1 @@
+wa?ő“"?XOo??áQ?��č>éżE>¸ˇĄ?±oş?9¸i?"qv>ů¨·>‡ u??�?µ W?9Ś“?Á´ű>št�?D2Ü?˝Ę>nž8?°Ź±>ČŢo?ŤOĆ?/Ç™?7y‚?J°?ÍeÔ?Aś?F$&?'ľ?Ab_?ąŚŁ?"mŔ>ËĽ?›¸?ď�?•›?ýŃÚ>—$é>­¸?ŕ*Ă?�)l?rďł?á†?óXź?9ů>żŞ??^ó˛>SÔż?w!§?

mindspore/lite/test/ut/src/runtime/kernel/arm/test_data/activationGrad/tanh_out_50.bin

@@ -0,0 +1,2 @@
+„ù@?^û*¿Su>?‹“¿(1?YÙ?O]Í>8�©>yåͽâ³h>Y·:¿¬×Ÿ<e
+?@?C?È‘¾®�C?6GU¾páž>¶_=¿I³¿0`Í>­0¾>ÝŽ9?Úÿ;?Gs*>e3>£”?¯Ê‘>œ»;?(ô,?õ&¿3*ˆ¾©Ü?ŠŸC?çC¿w<2?š�ð=ôKý>%HC¿ß¾ñ%8?òMâ>£¥œ¾ºñ~>'uû>Jß¿ÙI>^4Y¾uZ?ó¿

mindspore/lite/test/ut/src/runtime/kernel/arm/test_data/conv/convfp32_w_g3_18_3_3_3.bin

@@ -0,0 +1,4 @@
+‹F
½.¸¿NÓ2¾œ³Õ?󻾩`?°�ͽÕåØ¿S¿Ä¯
¿”2x¿¥R=}è%À–Tá?9¢>¾Ró
?E„ÿ†Ö?ÇÕÜ>©´?®Á@< ƒ¿ˆ�*¿Fü‰¿sÑ?¹Ýh¿ˆêý>ðÛ¾i) ¿�W>+
Ò;ôÎä=y«
@\
ô?ð¿V=~?ú)о‡Ï¬?HF}?åžÕ¿˜Õ«?Fê“¿±E
+?ŽG¿¾»ÿ#μ¿
>P¨D>‘¼È>J‹Å?gNð¾Y,	<¹Öˆ¿�–u?Y_À"¶�¾ ñ4?À¢f¾ôœ¿x‹¾	YÀ7ü;¡̾…ÑÚ?Ö)™?£°©?€ì…?Ö]@-Ç/¼z²b?Áäï¿Y¸¾ñ e?MÖý¾	/6?¦"¿ë‡?œt«?ŽØT?;	-¾½Ø1?,6?¿•¨¿.ª>nÉÞ>8«D?�Ǿ
+ãF¿Ö+j?~B?
+ê¿»¤P¾æ›Ç?Šœ?t ¾½ªek?›ûI?W³J>®ó&?æÑ ?;ñ;ƒéš¿Êjä¿¾=¾í±¿È¾sg»?¡ÝÀ¿[kÄ?‚âr?Ý–c>.þ¾äÏl¼žjy?¥DÊ>S«î¼“¬Â¿º‡?lìÒ?rS¾Ùq´¾Åä?#m(@±_?>±¿l ‚¿Ž%6À˜¢?’<j¾¹­z>h%¿Ðké¾>Þ=4 ù¿¨Œö?ªÅ‹¼o´J¾—û¾ �¿¥§s>•¾fW—?8c;?‚ä�?Æk»¾Š:º?bQ1¿ƒ¤Y>ypþ>½nW=úz¿|S:?P‚’?çrð?€KŠ?嘿k†µ¿wØâ>‘„-¼æ«>³Ví?Å~>

mindspore/lite/test/ut/src/runtime/kernel/arm/test_data/matmul/matmulfp32_c_4x5.bin

@@ -0,0 +1 @@
+.Àø@8�|À¸A-=,Ëú¿fAQÀ>2@du¤Ài}?ÿÃÀtÏ4@œˆš@2ªÊÀüzN@Ò¥¿Ô£x@Þ&ó½(Ó‚½‚¸e@�g�¼

mindspore/lite/test/ut/src/runtime/kernel/arm/test_data/operators/arithmetic_fp32_10_dy_5_4_6.bin

@@ -0,0 +1 @@
+Ñ&J¿vÿ�ÀB×AoLÝ?ˆÍƒÀI?ïÁ…óÌ¿ÜÐA6µÔ½ÏS¹¾?Á…—¸Áãþ�Àd>‘À5²®?±Çi¾’’Ó¿`ó@Úý ¾¿uº@G×Ý@`M>Aü¹À
v>B)¦ä>c”¾ÉÁ@$Ì/AwèŒA‡àÁŒˆË¿ß^ ¿ñ0ÁkܾÈ�£AfûÁèrñ>0xË¿€c¬ÀR†?ÂvuÁÊ=ˆ¿`,>pÅ”?aKÇÀªû@ðóîÁÓÀy?¼×¾ÿdb¿3ž?¹@Ú¤
À¬�’¿eK?üÑÀЧ9¿)ÖÓ?u£u@ýþ?¡"á=ûP’>bæë>vá°@þNl@ÅÃ�ÀUî>otÀ
’ð?Æ*�@y	;³õ	@´ÖAv‘©¿_„¿×
俶�î¿\�¿™qØ?˜w@–0ÀÝ»¶¾©j€?òï÷ÀAq;Àb½oÀ¡Ž,¿Ò3@ûI`?3sfÀlø‡À€Ê@@IÚ?˜£§¿
Ž·?
¡AC_>=ƒL�¿†Äó@„œÏ?	.–¿ïà¢@xy`? ´á¿¿�ŒA3
3¿³Ë‘?Ãn�?.À¾ûžÛ=é\B@/õÀõAç>B_‚¿»�©¾Kø¿–Ç<wÀWcz=m9=

mindspore/lite/test/ut/src/runtime/kernel/arm/test_data/operators/arithmetic_fp32_1_dx1_4_6.bin

@@ -0,0 +1 @@
+œwÓ¿Öè�?8ðÖ>GF¼?-¢^À�.;?º£÷¿V]«?K“@zè:À} QÀ’Põ¾qÂH¿Í Å¿‚‘ã?¯Õ,@~Ð?LP§>7>ÄP«¿¥¼q@Ÿå¿P«š>ƒ¬·@

mindspore/lite/test/ut/src/runtime/kernel/arm/test_data/operators/arithmetic_fp32_1_dx2_1_6.bin

@@ -0,0 +1 @@
+L&†À”­œ¾h>)AÃ[7?Á.šÀO2Ê@

mindspore/lite/test/ut/src/runtime/kernel/arm/test_data/operators/arithmetic_fp32_1_dy_4_6.bin

@@ -0,0 +1 @@
+œwÓ¿Öè�?8ðÖ>GF¼?-¢^À�.;?º£÷¿V]«?K“@zè:À} QÀ’Põ¾qÂH¿Í Å¿‚‘ã?¯Õ,@~Ð?LP§>7>ÄP«¿¥¼q@Ÿå¿P«š>ƒ¬·@

mindspore/lite/test/ut/src/runtime/kernel/arm/test_data/operators/arithmetic_fp32_1_x1_4_6.bin

@@ -0,0 +1 @@
+ˆÁŠ½.%Ž?ý•Ž¿.s?d¤“¿-f×=<V¾nß”?`Dz?�tŸ¿´"†¿?ýÿ¾e|»>�~�¾­Þ¾šIÈ?ˆ
³??³Æ½ÌY?ü¼3¾B‘?Ž¯X½f¦<?ä
'@

mindspore/lite/test/ut/src/runtime/kernel/arm/test_data/operators/arithmetic_fp32_1_x2_1_6.bin

@@ -0,0 +1 @@
+kB����t�?[�[�����T�>

mindspore/lite/test/ut/src/runtime/kernel/arm/test_data/operators/arithmetic_fp32_2_dx2_1_6.bin

@@ -0,0 +1 @@
+‹þÀ2øòÀ ©)A%»„ÀdÈ‘À°º	À

mindspore/lite/test/ut/src/runtime/kernel/arm/test_data/operators/arithmetic_fp32_3_dx2_1_6.bin

@@ -0,0 +1 @@
+‹þÀ2øòÀ ©)A%»„ÀdÈ‘À°º	À