!4553 add ori_format and ocr2/mv2 opencl perf testcase

Merge pull request !4553 from wandongdong/up

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

@@ -44,6 +44,7 @@ int ActivationOpenClKernel::Init() {
     MS_LOG(ERROR) << "Activate fun only support dim=4, but your dim=" << in_tensors_[0]->shape().size();
     return RET_ERROR;
   }
+  ori_format_ = out_tensors_[0]->GetFormat();
   std::string program_name = "";
   std::string kernel_name = "";
   std::string source = activation_source_fp32;

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

@@ -111,6 +111,7 @@ int ArithmeticOpenCLKernel::Init() {
   runtime_->LoadSource(program_name, source);
   runtime_->BuildKernel(kernel_, program_name, kernel_name, build_options);
 #endif
+  ori_format_ = out_tensors_[0]->GetFormat();
   out_tensors_[0]->SetFormat(schema::Format_NHWC4);
   Image2dGetWorkGroupSize();
   return 0;

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

@@ -56,6 +56,7 @@ int Conv2dTransposeOpenCLKernel::Init() {
   ocl_runtime->BuildKernel(kernel_, program_name, kernel_name, build_options);
 #endif
   PadWeight();
+  ori_format_ = out_tensors_[0]->GetFormat();
   out_tensors_[0]->SetFormat(schema::Format_NHWC4);
   MS_LOG(DEBUG) << kernel_name << " Init Done!";
   return RET_OK;

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

@@ -40,6 +40,7 @@ int ConvolutionOpenCLKernel::Init() {
   ocl_runtime->LoadSource(program_name, source);
   ocl_runtime->BuildKernel(kernel_, program_name, kernel_name, build_options);
   this->InitBuffer();
+  ori_format_ = out_tensors_[0]->GetFormat();
   out_tensors_[0]->SetFormat(schema::Format_NHWC4);
   MS_LOG(DEBUG) << kernel_name << " Init Done!";
   return RET_OK;

mindspore/lite/src/runtime/kernel/opencl/kernel/depthwise_conv2d.cc

@@ -43,6 +43,7 @@ int DepthwiseConv2dOpenCLKernel::Init() {
   auto ocl_runtime = lite::opencl::OpenCLRuntime::GetInstance();
   std::string kernel_name = "DepthwiseConv2d";
   auto in_format = in_tensors_[0]->GetFormat();
+  ori_format_ = out_tensors_[0]->GetFormat();
   out_tensors_[0]->SetFormat(in_format);
   if (in_format != schema::Format_NHWC4 && in_format != schema::Format_NC4HW4) {
     MS_LOG(ERROR) << "input format(" << in_format << ") "

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

@@ -68,6 +68,7 @@ int MatMulOpenCLKernel::Init() {
   PadWeight();
   allocator->UnmapBuffer(padWeight_);
   allocator->UnmapBuffer(bias_);
+  ori_format_ = out_tensors_[0]->GetFormat();
   out_tensors_[0]->SetFormat(schema::Format_NHWC4);
   MS_LOG(DEBUG) << kernel_name << " Init Done!";
   return 0;

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

@@ -19,7 +19,6 @@
 
 #include <vector>
 
-
 #include "src/runtime/kernel/opencl/opencl_kernel.h"
 #include "src/runtime/kernel/arm/nnacl/conv_parameter.h"
 #include "src/runtime/opencl/opencl_runtime.h"
@@ -32,11 +31,11 @@ using FLOAT_T = float;
 
 namespace mindspore::kernel {
 
-class MatMulOpenCLKernel : public LiteKernel {
+class MatMulOpenCLKernel : public OpenCLKernel {
  public:
   explicit MatMulOpenCLKernel(OpParameter *parameter, const std::vector<lite::tensor::Tensor *> &inputs,
                               const std::vector<lite::tensor::Tensor *> &outputs, bool hasBias)
-      : LiteKernel(parameter, inputs, outputs, nullptr, nullptr) {
+      : OpenCLKernel(parameter, inputs, outputs) {
     hasBias_ = hasBias;
   }
   ~MatMulOpenCLKernel() override{};
@@ -57,4 +56,3 @@ class MatMulOpenCLKernel : public LiteKernel {
 }  // namespace mindspore::kernel
 
 #endif  // MINDSPORE_LITE_SRC_BACKEND_OPENCL_MATMUL_H_
-

mindspore/lite/src/runtime/kernel/opencl/kernel/pooling2d.cc

@@ -73,6 +73,7 @@ int PoolingOpenCLKernel::Init() {
   ocl_runtime->LoadSource(program_name, source);
   ocl_runtime->BuildKernel(kernel_, program_name, kernel_name, build_options);
 #endif
+  ori_format_ = out_tensors_[0]->GetFormat();
   out_tensors_[0]->SetFormat(schema::Format_NHWC4);
   MS_LOG(DEBUG) << kernel_name << " Init Done!";
 

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

@@ -48,6 +48,7 @@ int ReshapeOpenCLKernel::Init() {
   ocl_runtime->LoadSource(program_name, source);
   ocl_runtime->BuildKernel(kernel_, program_name, kernel_name, build_options);
 #endif
+  ori_format_ = out_tensors_[0]->GetFormat();
   out_tensors_[0]->SetFormat(schema::Format_NHWC);
   MS_LOG(DEBUG) << kernel_name << " Init Done!";
   return RET_OK;

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

@@ -115,9 +115,10 @@ int SoftmaxOpenCLKernel::Init() {
   }
   std::set<std::string> build_options;
   runtime_->LoadSource(program_name, source);
-  out_tensors_[0]->SetFormat(schema::Format_NHWC4);
   runtime_->BuildKernel(kernel_, program_name, kernel_name, build_options);
 #endif
+  ori_format_ = out_tensors_[0]->GetFormat();
+  out_tensors_[0]->SetFormat(schema::Format_NHWC4);
   MS_LOG(DEBUG) << kernel_name << " Init Done!";
   return lite::RET_OK;
 }

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

@@ -60,6 +60,7 @@ int TransposeOpenCLKernel::Init() {
     MS_LOG(ERROR) << "input H * W % 4 != 0 not support!";
     return RET_ERROR;
   }
+  ori_format_ = out_tensors_[0]->GetFormat();
   out_tensors_[0]->SetFormat(schema::Format_NCHW);
   MS_LOG(DEBUG) << kernel_name << " Init Done!";
   return RET_OK;

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

@@ -49,9 +49,11 @@ class OpenCLKernel : public LiteKernel {
   }
   OpenCLMemType GetMemType() { return out_mem_type_; }
   void SetMemType(OpenCLMemType mem_type) { out_mem_type_ = mem_type; }
+  schema::Format GetOriFormat() { return ori_format_;}
 
  protected:
   OpenCLMemType out_mem_type_{OpenCLMemType::IMG};
+  schema::Format ori_format_{schema::Format_NHWC4};
 };
 }  // namespace mindspore::kernel
 

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

@@ -34,6 +34,15 @@ int SubGraphOpenCLKernel::GenToFormatOp(const std::vector<lite::tensor::Tensor *
   out_parameters->clear();
   out_convert_ops->clear();
   for (size_t i = 0; i < in_tensors.size(); ++i) {
+    OpenCLKernel* cur_opencl_op = reinterpret_cast<OpenCLKernel*>(in_kernels[i]);
+    schema::Format ori_format = cur_opencl_op->GetOriFormat();
+    if (mem_type == cur_opencl_op->GetMemType() && in_tensors[i]->GetFormat() == ori_format) {
+      continue;
+    }
+    auto dst_format =
+      (mem_type == OpenCLMemType::IMG) ? in_kernels[i]->out_tensors()[0]->GetFormat() : ori_format;
+    auto src_format =
+      (mem_type == OpenCLMemType::IMG) ? in_tensors[i]->GetFormat() : in_kernels[i]->out_tensors()[0]->GetFormat();
     lite::tensor::Tensor *new_tensor = new (std::nothrow) lite::tensor::Tensor();
     MS_ASSERT(new_tensor);
     if (new_tensor == nullptr) {
@@ -41,10 +50,6 @@ int SubGraphOpenCLKernel::GenToFormatOp(const std::vector<lite::tensor::Tensor *
       return RET_ERROR;
     }
     new_tensor->CopyTensor(*in_tensors[i]);
-    auto dst_format =
-      (mem_type == OpenCLMemType::IMG) ? in_kernels.back()->out_tensors()[0]->GetFormat() : in_tensors[i]->GetFormat();
-    auto src_format =
-      (mem_type == OpenCLMemType::IMG) ? in_tensors[i]->GetFormat() : in_kernels.front()->out_tensors()[0]->GetFormat();
     if ((dst_format == schema::Format_NCHW || dst_format == schema::Format_NC4HW4) &&
         (src_format == schema::Format_NHWC || src_format == schema::Format_NHWC4)) {
       auto &shape = new_tensor->shape();

mindspore/lite/test/st/benchmark_test.cc

@@ -49,6 +49,13 @@ const char *argv[] = {"./benchmark", "--modelPath=./hiai/model_02.ms",
 auto status = RunBenchmark(5, argv);
 ASSERT_EQ(status, RET_OK);
 }
+TEST_F(BenchmarkTest, TestOCR_02_GPU_PERF) {
+const char *argv[] = {"./benchmark", "--modelPath=./hiai/model_02.ms",
+                                     "--inDataPath=./hiai/model_02_in.bin",
+                                     "--device=GPU"};
+auto status = RunBenchmark(4, argv);
+ASSERT_EQ(status, RET_OK);
+}
 
 TEST_F(BenchmarkTest, Test_MV2_GPU) {
 const char *argv[] = {"./benchmark", "--modelPath=./hiai/mobilenet_v2.ms",
@@ -59,6 +66,14 @@ auto status = RunBenchmark(5, argv);
 ASSERT_EQ(status, RET_OK);
 }
 
+TEST_F(BenchmarkTest, Test_MV2_GPU_PERF) {
+  const char *argv[] = {"./benchmark", "--modelPath=./hiai/mobilenet_v2.ms",
+                        "--inDataPath=./hiai/mobilenet_v2_in.bin",
+                        "--device=GPU"};
+  auto status = RunBenchmark(4, argv);
+  ASSERT_EQ(status, RET_OK);
+}
+
 TEST_F(BenchmarkTest, TestHebing) {
   const char *argv[] = {"./benchmark", "--modelPath=./hiai/model_hebing_3branch.ms",
                                        "--inDataPath=./hiai/model_hebing_3branch.bin",