!5053 [MS][LITE][Develop] opencl activation function support 2d data

Merge pull request !5053 from liuzhongkai/2d_data

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

@@ -15,6 +15,7 @@
  */
 
 #include <vector>
+#include <map>
 #include <string>
 #include <set>
 
@@ -23,7 +24,6 @@
 #include "src/kernel_registry.h"
 #include "src/runtime/runtime_api.h"
 #include "include/errorcode.h"
-
 #include "src/runtime/kernel/opencl/cl/activation.cl.inc"
 
 using mindspore::kernel::KERNEL_ARCH::kGPU;
@@ -39,61 +39,58 @@ using mindspore::schema::PrimitiveType_Activation;
 namespace mindspore::kernel {
 
 int ActivationOpenClKernel::Init() {
-  const int max_shape_dim = 4;
-  if (in_tensors_[0]->shape().size() != max_shape_dim) {
-    MS_LOG(ERROR) << "Activate fun only support dim=4, but your dim=" << in_tensors_[0]->shape().size();
+  in_size_ = in_tensors_[0]->shape().size();
+  out_size_ = out_tensors_[0]->shape().size();
+  if (in_size_ != 2 && in_size_ != 4) {
+    MS_LOG(ERROR) << "Activate fun only support dim=4 or 2, but your dim=" << in_size_;
     return RET_ERROR;
   }
-  std::string program_name = "";
-  std::string kernel_name = "";
-  std::string source = activation_source;
-  if (type_ == ActivationType_RELU) {
-    program_name = "RELU";
-    kernel_name = "Relu";
-  } else if (type_ == ActivationType_RELU6) {
-    program_name = "RELU6";
-    kernel_name = "Relu6";
-  } else if (type_ == ActivationType_LEAKY_RELU) {
-    program_name = "LEAKY_RELU";
-    kernel_name = "ReluScalar";
-  } else if (type_ == ActivationType_SIGMOID) {
-    program_name = "SIGMOID";
-    kernel_name = "Sigmoid";
-  } else {
-    MS_LOG(ERROR) << "Activation type error";
+  std::map<int, std::vector<std::string>> Program_Kernel{
+    {ActivationType_LEAKY_RELU, std::vector<std::string>{"LEAKY_RELU", "ReluScalar"}},
+    {ActivationType_RELU, std::vector<std::string>{"RELU", "Relu"}},
+    {ActivationType_SIGMOID, std::vector<std::string>{"SIGMOID", "Sigmoid"}},
+    {ActivationType_RELU6, std::vector<std::string>{"RELU6", "Relu6"}}};
+  if (Program_Kernel.count(type_) == 0) {
+    MS_LOG(ERROR) << "schema::ActivationType:" << type_ << "not found";
     return RET_ERROR;
   }
+
+  std::string source = activation_source;
   std::set<std::string> build_options;
   auto ocl_runtime = lite::opencl::OpenCLRuntime::GetInstance();
-  ocl_runtime->LoadSource(program_name, source);
-  ocl_runtime->BuildKernel(kernel_, program_name, kernel_name, build_options);
+  ocl_runtime->LoadSource(Program_Kernel[type_][0], source);
+  ocl_runtime->BuildKernel(kernel_, Program_Kernel[type_][0], Program_Kernel[type_][1], build_options);
+
+  std::map<int, schema::Format> format{{4, schema::Format_NHWC4}, {2, schema::Format_NC4}};
+  if (format.count(out_size_) == 0) {
+    MS_LOG(ERROR) << "Not found output tensor format";
+    return RET_ERROR;
+  }
   in_ori_format_ = in_tensors_[0]->GetFormat();
-  in_tensors_[0]->SetFormat(schema::Format_NHWC4);
   out_ori_format_ = out_tensors_[0]->GetFormat();
-  out_tensors_[0]->SetFormat(schema::Format_NHWC4);
+  in_tensors_[0]->SetFormat(format[in_size_]);
+  out_tensors_[0]->SetFormat(format[out_size_]);
+  if (in_size_ == 2) {
+    in_ori_format_ = schema::Format_NC4;
+    out_ori_format_ = schema::Format_NC4;
+  }
   MS_LOG(DEBUG) << op_parameter_->name_ << " init Done!";
   return RET_OK;
 }
 
 int ActivationOpenClKernel::Run() {
   MS_LOG(DEBUG) << op_parameter_->name_ << " begin running!";
-  int N = in_tensors_[0]->shape()[0];
-  int H = in_tensors_[0]->shape()[1];
-  int W = in_tensors_[0]->shape()[2];
-  int C = in_tensors_[0]->shape()[3];
-  cl_int4 input_shape = {N, H, W, C};
-
+  cl_int4 img2d_shape = GetImg2dShape();
   auto ocl_runtime = lite::opencl::OpenCLRuntime::GetInstance();
   int arg_idx = 0;
   ocl_runtime->SetKernelArg(kernel_, arg_idx++, in_tensors_[0]->Data());
   ocl_runtime->SetKernelArg(kernel_, arg_idx++, out_tensors_[0]->Data());
-  ocl_runtime->SetKernelArg(kernel_, arg_idx++, input_shape);
+  ocl_runtime->SetKernelArg(kernel_, arg_idx++, img2d_shape);
   if (type_ == ActivationType_LEAKY_RELU) {
     ocl_runtime->SetKernelArg(kernel_, arg_idx++, alpha_);
   }
   std::vector<size_t> local = {1, 1};
-  std::vector<size_t> global = {static_cast<size_t>(H), static_cast<size_t>(W)};
-  std::cout << type_ << " " << std::endl;
+  std::vector<size_t> global = {static_cast<size_t>(img2d_shape.s[1]), static_cast<size_t>(img2d_shape.s[2])};
   auto ret = ocl_runtime->RunKernel(kernel_, global, local, nullptr);
   if (ret != RET_OK) {
     MS_LOG(ERROR) << "Run kernel:" << op_parameter_->name_ << " fail.";
@@ -102,11 +99,21 @@ int ActivationOpenClKernel::Run() {
   return RET_OK;
 }
 
-int ActivationOpenClKernel::GetImageSize(size_t idx, std::vector<size_t> *img_size) {
-  int H = in_tensors_[0]->shape()[1];
-  int W = in_tensors_[0]->shape()[2];
-  int C = in_tensors_[0]->shape()[3];
+cl_int4 ActivationOpenClKernel::GetImg2dShape() {
+  cl_int4 img2d_shape = {0, 0, 0, 0};
+  for (int i = 0; i < in_size_; ++i) {
+    img2d_shape.s[i + 4 - in_size_] = in_tensors_[0]->shape()[i];
+  }
+  if (in_size_ == 2) {
+    img2d_shape.s[1] = img2d_shape.s[2];
+    img2d_shape.s[2] = UP_DIV(img2d_shape.s[3], C4NUM);
+    img2d_shape.s[3] = C4NUM;
+  }
+  return img2d_shape;
+}
 
+int ActivationOpenClKernel::GetImageSize(size_t idx, std::vector<size_t> *img_size) {
+  cl_int4 img_shape = GetImg2dShape();
 #ifdef ENABLE_FP16
   size_t img_dtype = CL_HALF_FLOAT;
 #else
@@ -114,8 +121,8 @@ int ActivationOpenClKernel::GetImageSize(size_t idx, std::vector<size_t> *img_si
 #endif
 
   img_size->clear();
-  img_size->push_back(W * UP_DIV(C, C4NUM));
-  img_size->push_back(H);
+  img_size->push_back(img_shape.s[2] * UP_DIV(img_shape.s[3], C4NUM));
+  img_size->push_back(img_shape.s[1]);
   img_size->push_back(img_dtype);
   return RET_OK;
 }
@@ -125,11 +132,11 @@ kernel::LiteKernel *OpenClActivationFp32KernelCreator(const std::vector<lite::te
                                                       OpParameter *opParameter, const lite::Context *ctx,
                                                       const kernel::KernelKey &desc,
                                                       const mindspore::lite::PrimitiveC *primitive) {
-  if (inputs.size() == 0) {
+  if (inputs.empty()) {
     MS_LOG(ERROR) << "Input data size must be greater than 0, but your size is " << inputs.size();
     return nullptr;
   }
-  if (inputs[0]->shape()[0] > 1) {
+  if (inputs[0]->shape().size() > 2 && inputs[0]->shape()[0] > 1) {
     MS_LOG(ERROR) << "Activation kernel:" << opParameter->name_ << " failed: Unsupported multi-batch.";
     return nullptr;
   }

mindspore/lite/src/runtime/kernel/opencl/kernel/activation.h

@@ -38,11 +38,14 @@ class ActivationOpenClKernel : public OpenCLKernel {
   int Init() override;
   int Run() override;
   int GetImageSize(size_t idx, std::vector<size_t> *img_size) override;
+  cl_int4 GetImg2dShape();
 
  private:
   cl::Kernel kernel_;
   int type_;
   float alpha_;
+  int in_size_;
+  int out_size_;
 };
 
 }  // namespace mindspore::kernel

mindspore/lite/test/ut/src/runtime/kernel/opencl/activation_tests.cc

@@ -73,91 +73,230 @@ void printf_tensor(mindspore::lite::tensor::Tensor *in_data) {
   MS_LOG(INFO) << "Print tensor done";
 }
 
-kernel::ActivationOpenClKernel *create_kernel(lite::opencl::OpenCLAllocator *allocator,
-                                              const std::vector<lite::tensor::Tensor *> &inputs,
-                                              const std::vector<lite::tensor::Tensor *> &outputs, std::string test_name,
-                                              int type, std::string in_file, float alpha = 0.2) {
+TEST_F(TestActivationOpenCL, ReluFp32_dim4) {
+  std::string in_file = "/data/local/tmp/in_data.bin";
+  std::string out_file = "/data/local/tmp/relu.bin";
+  MS_LOG(INFO) << "Relu Begin test!";
+  auto ocl_runtime = lite::opencl::OpenCLRuntime::GetInstance();
+  ocl_runtime->Init();
+  auto allocator = ocl_runtime->GetAllocator();
+
+  MS_LOG(INFO) << "Init tensors.";
+  std::vector<int> input_shape = {1, 9};
+  auto data_type = kNumberTypeFloat32;
+  auto tensor_type = schema::NodeType_ValueNode;
+  auto *input_tensor = new (std::nothrow) lite::tensor::Tensor(data_type, input_shape, schema::Format_NC, tensor_type);
+  if (input_tensor == nullptr) {
+    MS_LOG(ERROR) << "new input tensor error!";
+    return;
+  }
+  auto *output_tensor = new (std::nothrow) lite::tensor::Tensor(data_type, input_shape, schema::Format_NC, tensor_type);
+  if (output_tensor == nullptr) {
+    MS_LOG(ERROR) << "new output tensor error!";
+    delete input_tensor;
+    return;
+  }
+  std::vector<lite::tensor::Tensor *> inputs{input_tensor};
+  std::vector<lite::tensor::Tensor *> outputs{output_tensor};
+  inputs[0]->MallocData(allocator);
+  MS_LOG(INFO) << "Initialize input data";
+  LoadActivationData(inputs[0]->Data(), inputs[0]->Size(), in_file);
+  MS_LOG(INFO) << "==================input data================";
+  printf_tensor(inputs[0]);
+
   auto *param = new (std::nothrow) ActivationParameter();
   if (param == nullptr) {
     MS_LOG(ERROR) << "New ActivationParameter fail.";
-    return nullptr;
+    delete input_tensor;
+    delete output_tensor;
+    return;
   }
-  memcpy(param->op_parameter_.name_, test_name.c_str(), test_name.size());
-  param->alpha_ = alpha;
-  param->type_ = type;
+  param->type_ = ActivationType_RELU;
   auto *kernel =
     new (std::nothrow) kernel::ActivationOpenClKernel(reinterpret_cast<OpParameter *>(param), inputs, outputs);
   if (kernel == nullptr) {
+    MS_LOG(ERROR) << "Kernel:Relu create fail.";
     delete param;
-    MS_LOG(ERROR) << "Kernel:" << test_name << " create fail.";
-    delete param;
-    return nullptr;
+    delete input_tensor;
+    delete output_tensor;
+    return;
   }
   auto ret = kernel->Init();
   if (ret != RET_OK) {
     delete param;
     delete kernel;
-    MS_LOG(ERROR) << "Init " << test_name << " fail.";
+    delete input_tensor;
+    delete output_tensor;
+    MS_LOG(ERROR) << "Init relu fail.";
+    return;
+  }
+  MS_LOG(INFO) << "Create kernel SubGraphOpenCLKernel.";
+  std::vector<kernel::LiteKernel *> kernels{kernel};
+  auto *sub_graph = new (std::nothrow) kernel::SubGraphOpenCLKernel(inputs, outputs, kernels, kernels, kernels);
+  if (sub_graph == nullptr) {
+    delete kernel;
+    delete param;
+    delete input_tensor;
+    delete output_tensor;
+    MS_LOG(ERROR) << "Kernel SubGraphOpenCLKernel create fail.";
+    return;
+  }
+
+  MS_LOG(INFO) << "Initialize sub_graph.";
+  ret = sub_graph->Init();
+  if (ret != RET_OK) {
+    MS_LOG(ERROR) << "Init sub_graph error.";
+    delete kernel;
+    delete param;
+    delete input_tensor;
+    delete output_tensor;
+    delete sub_graph;
+    return;
+  }
+  MS_LOG(INFO) << "Run SubGraphOpenCLKernel.";
+  ret = sub_graph->Run();
+  if (ret != RET_OK) {
     delete kernel;
     delete param;
-    return nullptr;
+    delete input_tensor;
+    delete output_tensor;
+    delete sub_graph;
+    MS_LOG(ERROR) << "Run SubGraphOpenCLKernel error.";
+    return;
+  }
+
+  MS_LOG(INFO) << "==================output data================";
+  printf_tensor(outputs[0]);
+  CompareRes(output_tensor, out_file);
+  delete kernel;
+  delete param;
+  delete input_tensor;
+  delete output_tensor;
+  delete sub_graph;
+  lite::opencl::OpenCLRuntime::DeleteInstance();
+}
+
+TEST_F(TestActivationOpenCL, Relu6Fp32_dim4) {
+  std::string in_file = "/data/local/tmp/in_data.bin";
+  std::string out_file = "/data/local/tmp/relu6.bin";
+  MS_LOG(INFO) << "Relu6 Begin test!";
+  auto ocl_runtime = lite::opencl::OpenCLRuntime::GetInstance();
+  ocl_runtime->Init();
+  auto allocator = ocl_runtime->GetAllocator();
+
+  MS_LOG(INFO) << "Init tensors.";
+  std::vector<int> input_shape = {1, 9};
+  auto data_type = kNumberTypeFloat32;
+  auto tensor_type = schema::NodeType_ValueNode;
+  auto *input_tensor = new (std::nothrow) lite::tensor::Tensor(data_type, input_shape, schema::Format_NC, tensor_type);
+  if (input_tensor == nullptr) {
+    MS_LOG(ERROR) << "new input tensor error!";
+    return;
   }
+  auto *output_tensor = new (std::nothrow) lite::tensor::Tensor(data_type, input_shape, schema::Format_NC, tensor_type);
+  if (output_tensor == nullptr) {
+    MS_LOG(ERROR) << "new output tensor error!";
+    delete input_tensor;
+    return;
+  }
+  std::vector<lite::tensor::Tensor *> inputs{input_tensor};
+  std::vector<lite::tensor::Tensor *> outputs{output_tensor};
+  inputs[0]->MallocData(allocator);
   MS_LOG(INFO) << "Initialize input data";
   LoadActivationData(inputs[0]->Data(), inputs[0]->Size(), in_file);
   MS_LOG(INFO) << "==================input data================";
   printf_tensor(inputs[0]);
-  return kernel;
-}
 
-int RunSubGraphOpenCLKernel(const std::vector<lite::tensor::Tensor *> &inputs,
-                            const std::vector<lite::tensor::Tensor *> &outputs,
-                            kernel::ActivationOpenClKernel *kernel) {
+  auto *param = new (std::nothrow) ActivationParameter();
+  if (param == nullptr) {
+    MS_LOG(ERROR) << "New ActivationParameter fail.";
+    delete input_tensor;
+    delete output_tensor;
+    return;
+  }
+  param->type_ = ActivationType_RELU6;
+  auto *kernel =
+    new (std::nothrow) kernel::ActivationOpenClKernel(reinterpret_cast<OpParameter *>(param), inputs, outputs);
+  if (kernel == nullptr) {
+    MS_LOG(ERROR) << "Kernel:Relu6 create fail.";
+    delete param;
+    delete input_tensor;
+    delete output_tensor;
+    return;
+  }
+  auto ret = kernel->Init();
+  if (ret != RET_OK) {
+    delete param;
+    delete kernel;
+    delete input_tensor;
+    delete output_tensor;
+    MS_LOG(ERROR) << "Init relu6 fail.";
+    return;
+  }
   MS_LOG(INFO) << "Create kernel SubGraphOpenCLKernel.";
   std::vector<kernel::LiteKernel *> kernels{kernel};
   auto *sub_graph = new (std::nothrow) kernel::SubGraphOpenCLKernel(inputs, outputs, kernels, kernels, kernels);
   if (sub_graph == nullptr) {
     delete kernel;
+    delete param;
+    delete input_tensor;
+    delete output_tensor;
     MS_LOG(ERROR) << "Kernel SubGraphOpenCLKernel create fail.";
-    return RET_ERROR;
+    return;
   }
+
   MS_LOG(INFO) << "Initialize sub_graph.";
-  auto ret = sub_graph->Init();
+  ret = sub_graph->Init();
   if (ret != RET_OK) {
+    MS_LOG(ERROR) << "Init sub_graph error.";
     delete kernel;
+    delete param;
+    delete input_tensor;
+    delete output_tensor;
     delete sub_graph;
-    MS_LOG(ERROR) << "Init sub_graph error.";
-    return RET_ERROR;
+    return;
   }
   MS_LOG(INFO) << "Run SubGraphOpenCLKernel.";
   ret = sub_graph->Run();
   if (ret != RET_OK) {
+    delete kernel;
+    delete param;
+    delete input_tensor;
+    delete output_tensor;
     delete sub_graph;
     MS_LOG(ERROR) << "Run SubGraphOpenCLKernel error.";
-    return RET_ERROR;
+    return;
   }
+
+  MS_LOG(INFO) << "==================output data================";
+  printf_tensor(outputs[0]);
+  CompareRes(output_tensor, out_file);
+  delete kernel;
+  delete param;
+  delete input_tensor;
+  delete output_tensor;
   delete sub_graph;
-  return RET_OK;
+  lite::opencl::OpenCLRuntime::DeleteInstance();
 }
 
-TEST_F(TestActivationOpenCL, ActivationFp32_dim4) {
-  MS_LOG(INFO) << "Begin test!";
+TEST_F(TestActivationOpenCL, SigmoidFp32_dim4) {
+  std::string in_file = "/data/local/tmp/in_data.bin";
+  std::string out_file = "/data/local/tmp/sigmoid.bin";
+  MS_LOG(INFO) << "Sigmoid Begin test!";
   auto ocl_runtime = lite::opencl::OpenCLRuntime::GetInstance();
   ocl_runtime->Init();
   auto allocator = ocl_runtime->GetAllocator();
 
   MS_LOG(INFO) << "Init tensors.";
-  std::vector<int> input_shape = {1, 4, 3, 8};
-
+  std::vector<int> input_shape = {1, 9};
   auto data_type = kNumberTypeFloat32;
   auto tensor_type = schema::NodeType_ValueNode;
-  auto *input_tensor =
-    new (std::nothrow) lite::tensor::Tensor(data_type, input_shape, schema::Format_NHWC4, tensor_type);
+  auto *input_tensor = new (std::nothrow) lite::tensor::Tensor(data_type, input_shape, schema::Format_NC, tensor_type);
   if (input_tensor == nullptr) {
     MS_LOG(ERROR) << "new input tensor error!";
     return;
   }
-  auto *output_tensor =
-    new (std::nothrow) lite::tensor::Tensor(data_type, input_shape, schema::Format_NHWC4, tensor_type);
+  auto *output_tensor = new (std::nothrow) lite::tensor::Tensor(data_type, input_shape, schema::Format_NC, tensor_type);
   if (output_tensor == nullptr) {
     MS_LOG(ERROR) << "new output tensor error!";
     delete input_tensor;
@@ -166,40 +305,184 @@ TEST_F(TestActivationOpenCL, ActivationFp32_dim4) {
   std::vector<lite::tensor::Tensor *> inputs{input_tensor};
   std::vector<lite::tensor::Tensor *> outputs{output_tensor};
   inputs[0]->MallocData(allocator);
+  MS_LOG(INFO) << "Initialize input data";
+  LoadActivationData(inputs[0]->Data(), inputs[0]->Size(), in_file);
+  MS_LOG(INFO) << "==================input data================";
+  printf_tensor(inputs[0]);
+
+  auto *param = new (std::nothrow) ActivationParameter();
+  if (param == nullptr) {
+    MS_LOG(ERROR) << "New ActivationParameter fail.";
+    delete input_tensor;
+    delete output_tensor;
+    return;
+  }
+  param->type_ = ActivationType_SIGMOID;
+  auto *kernel =
+    new (std::nothrow) kernel::ActivationOpenClKernel(reinterpret_cast<OpParameter *>(param), inputs, outputs);
+  if (kernel == nullptr) {
+    MS_LOG(ERROR) << "Kernel:Sigmoid create fail.";
+    delete param;
+    delete input_tensor;
+    delete output_tensor;
+    return;
+  }
+  auto ret = kernel->Init();
+  if (ret != RET_OK) {
+    delete param;
+    delete kernel;
+    delete input_tensor;
+    delete output_tensor;
+    MS_LOG(ERROR) << "Init sigmoid fail.";
+    return;
+  }
+  MS_LOG(INFO) << "Create kernel SubGraphOpenCLKernel.";
+  std::vector<kernel::LiteKernel *> kernels{kernel};
+  auto *sub_graph = new (std::nothrow) kernel::SubGraphOpenCLKernel(inputs, outputs, kernels, kernels, kernels);
+  if (sub_graph == nullptr) {
+    delete kernel;
+    delete param;
+    delete input_tensor;
+    delete output_tensor;
+    MS_LOG(ERROR) << "Kernel SubGraphOpenCLKernel create fail.";
+    return;
+  }
+
+  MS_LOG(INFO) << "Initialize sub_graph.";
+  ret = sub_graph->Init();
+  if (ret != RET_OK) {
+    MS_LOG(ERROR) << "Init sub_graph error.";
+    delete kernel;
+    delete param;
+    delete input_tensor;
+    delete output_tensor;
+    delete sub_graph;
+    return;
+  }
+  MS_LOG(INFO) << "Run SubGraphOpenCLKernel.";
+  ret = sub_graph->Run();
+  if (ret != RET_OK) {
+    delete kernel;
+    delete param;
+    delete input_tensor;
+    delete output_tensor;
+    delete sub_graph;
+    MS_LOG(ERROR) << "Run SubGraphOpenCLKernel error.";
+    return;
+  }
 
-  std::map<std::string, int> Test_Activation_Type;
-  std::map<std::string, std::string> Test_Res_File;
-  Test_Activation_Type["Relu"] = ActivationType_RELU;
-  Test_Activation_Type["Leaky_Relu"] = ActivationType_LEAKY_RELU;
-  Test_Activation_Type["Relu6"] = ActivationType_RELU6;
-  Test_Activation_Type["Sigmoid"] = ActivationType_SIGMOID;
-  Test_Res_File["Leaky_Relu"] = "/data/local/tmp/leaky_relu.bin";
-  Test_Res_File["Relu"] = "/data/local/tmp/relu.bin";
-  Test_Res_File["Relu6"] = "/data/local/tmp/relu6.bin";
-  Test_Res_File["Sigmoid"] = "/data/local/tmp/sigmoid.bin";
+  MS_LOG(INFO) << "==================output data================";
+  printf_tensor(outputs[0]);
+  CompareRes(output_tensor, out_file);
+  delete kernel;
+  delete param;
+  delete input_tensor;
+  delete output_tensor;
+  delete sub_graph;
+  lite::opencl::OpenCLRuntime::DeleteInstance();
+}
+
+TEST_F(TestActivationOpenCL, LeakyReluFp32_dim4) {
   std::string in_file = "/data/local/tmp/in_data.bin";
+  std::string out_file = "/data/local/tmp/leaky_relu.bin";
+  MS_LOG(INFO) << "Leaky relu Begin test!";
+  auto ocl_runtime = lite::opencl::OpenCLRuntime::GetInstance();
+  ocl_runtime->Init();
+  auto allocator = ocl_runtime->GetAllocator();
 
-  std::map<std::string, int>::iterator it = Test_Activation_Type.begin();
-  while (it != Test_Activation_Type.end()) {
-    auto kernel = create_kernel(allocator, inputs, outputs, it->first, it->second, in_file, 0.3);
-    if (kernel == nullptr) {
-      MS_LOG(ERROR) << "Create kernel:" << it->first << " error.";
-      return;
-    }
+  MS_LOG(INFO) << "Init tensors.";
+  std::vector<int> input_shape = {1, 9};
+  auto data_type = kNumberTypeFloat32;
+  auto tensor_type = schema::NodeType_ValueNode;
+  auto *input_tensor = new (std::nothrow) lite::tensor::Tensor(data_type, input_shape, schema::Format_NC, tensor_type);
+  if (input_tensor == nullptr) {
+    MS_LOG(ERROR) << "new input tensor error!";
+    return;
+  }
+  auto *output_tensor = new (std::nothrow) lite::tensor::Tensor(data_type, input_shape, schema::Format_NC, tensor_type);
+  if (output_tensor == nullptr) {
+    MS_LOG(ERROR) << "new output tensor error!";
+    delete input_tensor;
+    return;
+  }
+  std::vector<lite::tensor::Tensor *> inputs{input_tensor};
+  std::vector<lite::tensor::Tensor *> outputs{output_tensor};
+  inputs[0]->MallocData(allocator);
+  MS_LOG(INFO) << "Initialize input data";
+  LoadActivationData(inputs[0]->Data(), inputs[0]->Size(), in_file);
+  MS_LOG(INFO) << "==================input data================";
+  printf_tensor(inputs[0]);
 
-    auto ret = RunSubGraphOpenCLKernel(inputs, outputs, kernel);
-    if (ret != RET_OK) {
-      MS_LOG(ERROR) << it->first << " RunSubGraphOpenCLKernel error.";
-      return;
-    }
-    MS_LOG(INFO) << "==================output data================";
-    printf_tensor(outputs[0]);
-    CompareRes(output_tensor, Test_Res_File[it->first]);
-    it++;
+  auto *param = new (std::nothrow) ActivationParameter();
+  if (param == nullptr) {
+    MS_LOG(ERROR) << "New ActivationParameter fail.";
+    delete input_tensor;
+    delete output_tensor;
+    return;
+  }
+  param->alpha_ = 0.3;
+  param->type_ = ActivationType_LEAKY_RELU;
+  auto *kernel =
+    new (std::nothrow) kernel::ActivationOpenClKernel(reinterpret_cast<OpParameter *>(param), inputs, outputs);
+  if (kernel == nullptr) {
+    MS_LOG(ERROR) << "Kernel:leaky relu create fail.";
+    delete param;
+    delete input_tensor;
+    delete output_tensor;
+    return;
+  }
+  auto ret = kernel->Init();
+  if (ret != RET_OK) {
+    delete param;
+    delete kernel;
+    delete input_tensor;
+    delete output_tensor;
+    MS_LOG(ERROR) << "Init leaky relu fail.";
+    return;
+  }
+  MS_LOG(INFO) << "Create kernel SubGraphOpenCLKernel.";
+  std::vector<kernel::LiteKernel *> kernels{kernel};
+  auto *sub_graph = new (std::nothrow) kernel::SubGraphOpenCLKernel(inputs, outputs, kernels, kernels, kernels);
+  if (sub_graph == nullptr) {
+    delete kernel;
+    delete param;
+    delete input_tensor;
+    delete output_tensor;
+    MS_LOG(ERROR) << "Kernel SubGraphOpenCLKernel create fail.";
+    return;
   }
 
+  MS_LOG(INFO) << "Initialize sub_graph.";
+  ret = sub_graph->Init();
+  if (ret != RET_OK) {
+    MS_LOG(ERROR) << "Init sub_graph error.";
+    delete kernel;
+    delete param;
+    delete input_tensor;
+    delete output_tensor;
+    delete sub_graph;
+    return;
+  }
+  MS_LOG(INFO) << "Run SubGraphOpenCLKernel.";
+  ret = sub_graph->Run();
+  if (ret != RET_OK) {
+    delete kernel;
+    delete param;
+    delete input_tensor;
+    delete output_tensor;
+    delete sub_graph;
+    MS_LOG(ERROR) << "Run SubGraphOpenCLKernel error.";
+    return;
+  }
+
+  MS_LOG(INFO) << "==================output data================";
+  printf_tensor(outputs[0]);
+  CompareRes(output_tensor, out_file);
+  delete kernel;
+  delete param;
   delete input_tensor;
   delete output_tensor;
+  delete sub_graph;
   lite::opencl::OpenCLRuntime::DeleteInstance();
 }
 }  // namespace mindspore