add opencl kernels to libmindspore-lite

mindspore/lite/src/CMakeLists.txt

@@ -36,11 +36,19 @@ file(GLOB_RECURSE C_OPS_SRC ${CMAKE_CURRENT_SOURCE_DIR}/ops/*.cc)
 add_library(mindspore-lite SHARED ${LITE_SRC} ${C_OPS_SRC})
 set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -Wno-unused-private-field")
 add_library(anf OBJECT ${ANF_SRC})
-target_link_libraries(mindspore-lite
-    anf
-    cpu_kernel_mid_
-    )
-
+if (SUPPORT_GPU)
+  add_subdirectory(runtime/kernel/opencl)
+  target_link_libraries(mindspore-lite
+      anf
+      cpu_kernel_mid_
+      opencl_kernel_lib_
+      )
+else ()
+  target_link_libraries(mindspore-lite
+      anf
+      cpu_kernel_mid_
+      )
+endif ()
 add_subdirectory(runtime/kernel/arm)
 if (PLATFORM_ARM32 OR PLATFORM_ARM64)
   target_link_libraries(mindspore-lite log)

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

@@ -1,16 +1,2 @@
-set(OPENCL_KERNEL_SRC
-    ${CMAKE_CURRENT_SOURCE_DIR}/subgraph_opencl_kernel.cc
-    ${CMAKE_CURRENT_SOURCE_DIR}/utils.cc
-    ${CMAKE_CURRENT_SOURCE_DIR}/kernel/arithmetic.cc
-    ${CMAKE_CURRENT_SOURCE_DIR}/kernel/convolution.cc
-    ${CMAKE_CURRENT_SOURCE_DIR}/kernel/depthwise_conv2d.cc
-    ${CMAKE_CURRENT_SOURCE_DIR}/kernel/pooling2d.cc
-    ${CMAKE_CURRENT_SOURCE_DIR}/kernel/matmul.cc
-    ${CMAKE_CURRENT_SOURCE_DIR}/kernel/softmax.cc
-    ${CMAKE_CURRENT_SOURCE_DIR}/kernel/concat.cc
-    ${CMAKE_CURRENT_SOURCE_DIR}/kernel/conv2d_transpose.cc
-    ${CMAKE_CURRENT_SOURCE_DIR}/kernel/transpose.cc
-    ${CMAKE_CURRENT_SOURCE_DIR}/kernel/reshape.cc
-    ${CMAKE_CURRENT_SOURCE_DIR}/kernel/activation.cc
-    ${CMAKE_CURRENT_SOURCE_DIR}/kernel/caffe_prelu.cc
-    )
+file(GLOB_RECURSE OPENCL_KERNEL_SRC ${CMAKE_CURRENT_SOURCE_DIR}/kernel/*.cc)
+add_library(opencl_kernel_lib_ OBJECT ${KERNEL_SRC} ${OPENCL_KERNEL_SRC})

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

@@ -131,7 +131,6 @@ int ArithmeticOpenCLKernel::Run() {
   MS_LOG(DEBUG) << this->name() << " Running!";
   auto ocl_runtime = lite::opencl::OpenCLRuntime::GetInstance();
 
-  uint32_t element_num = out_tensors_[0]->ElementsC4Num();
   int arg_idx = 0;
 
   ocl_runtime->SetKernelArg(kernel_, arg_idx++, in_tensors_[0]->Data());

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

@@ -40,7 +40,6 @@ class Conv2dTransposeOpenCLKernel : public OpenCLKernel {
   int GetImageSize(size_t idx, std::vector<size_t> *img_size) override;
 
  private:
-  ConvParameter *parameter_;
   cl::Kernel kernel_;
   void *padWeight_;
   void *bias_;

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

@@ -126,6 +126,7 @@ int ToFormatOpenCLKernel::GetImageSize(size_t idx, std::vector<size_t> *img_size
     im_dst_y = h;
   } else {
     MS_LOG(ERROR) << "Unsupported format. " << out_tensors_[0]->GetFormat();
+    return RET_ERROR;
   }
   img_size->clear();
   auto enable_fp16_ = lite::opencl::OpenCLRuntime::GetInstance()->GetFp16Enable();

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

@@ -129,7 +129,6 @@ TEST_F(TestAvgPoolingOpenCL, AvgPoolFp32) {
     printf("%.3f ", output_data[i]);
   }
   printf("\n");
-  size_t output_size = tensor_out->Size();
   float expect[4] = {2.0f, 3.0f, 4.0f, 5.0f};
 
   for (int i = 0; i < tensor_out->ElementsNum(); ++i)

mindspore/lite/test/ut/src/runtime/kernel/opencl/depthwise_conv2d_tests.cc

@@ -54,7 +54,6 @@ void DepthWiseTestMain(ConvParameter *conv_param, T2 *input_data, T1 *weight_dat
   }
 
   // pack weight
-  int OC4 = UP_DIV(conv_param->output_channel_, C4NUM);
   int pack_weight_size = conv_param->output_channel_ * conv_param->kernel_h_ * conv_param->kernel_w_;
   T1 *packed_weight = weight_data;
 
@@ -103,7 +102,6 @@ void DepthWiseTestMain(ConvParameter *conv_param, T2 *input_data, T1 *weight_dat
 
   std::vector<kernel::LiteKernel *> kernels{pKernel.get()};
   std::vector<lite::tensor::Tensor *> inputs_{&tensor_a};
-  size_t C4 = UP_DIV(inputs[0]->Channel(), C4NUM);
   auto pGraph = std::make_unique<kernel::SubGraphOpenCLKernel>(inputs_, outputs, kernels, kernels, kernels);
   if (pGraph.get() == nullptr) {
     delete[] packed_input;