!11276 【MS】【LITE】【GPU】 reduce opencl so size 0.5M

From: @wangdongxu6
Reviewed-by: @ddwsky
Signed-off-by: @ddwsky

mindspore/lite/src/CMakeLists.txt

@@ -1,23 +1,20 @@
 add_compile_definitions(USE_ANDROID_LOG)
-if (ENABLE_V0)
-  add_definitions(-DENABLE_V0)
+if(ENABLE_V0)
+    add_definitions(-DENABLE_V0)
 endif()
 set(LITE_DIR ${CMAKE_CURRENT_SOURCE_DIR}/..)
 include_directories(${LITE_DIR}/nnacl/)
 include_directories(${LITE_DIR}/nnacl/optimize)
 
-if (PLATFORM_ARM32 OR PLATFORM_ARM64)
+if(PLATFORM_ARM32 OR PLATFORM_ARM64)
     #for performance
-    if ("${CMAKE_BUILD_TYPE}" STREQUAL "Release")
-        if (SUPPORT_GPU)
-            set(CMAKE_C_FLAGS "${CMAKE_C_FLAGS}  -fomit-frame-pointer -fstrict-aliasing -ffunction-sections -fdata-sections -ffast-math -fno-rtti")
-            set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS}  -fomit-frame-pointer -fstrict-aliasing -ffunction-sections -fdata-sections -ffast-math -fno-rtti")
-        else ()
-            set(CMAKE_C_FLAGS "${CMAKE_C_FLAGS}  -fomit-frame-pointer -fstrict-aliasing -ffunction-sections -fdata-sections -ffast-math -fno-rtti -fno-exceptions")
-            set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS}  -fomit-frame-pointer -fstrict-aliasing -ffunction-sections -fdata-sections -ffast-math -fno-rtti -fno-exceptions")
-        endif ()
-    endif ()
-endif ()
+    if("${CMAKE_BUILD_TYPE}" STREQUAL "Release")
+        set(CMAKE_C_FLAGS "${CMAKE_C_FLAGS}  -fomit-frame-pointer -fstrict-aliasing -ffunction-sections \
+        -fdata-sections -ffast-math -fno-rtti -fno-exceptions")
+        set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS}  -fomit-frame-pointer -fstrict-aliasing -ffunction-sections \
+        -fdata-sections -ffast-math -fno-rtti -fno-exceptions")
+    endif()
+endif()
 
 set(LITE_SRC
         ${CMAKE_CURRENT_SOURCE_DIR}/common/file_utils.cc
@@ -42,7 +39,7 @@ set(LITE_SRC
         ${CMAKE_CURRENT_SOURCE_DIR}/dequant.cc
         )
 
-if (SUPPORT_GPU)
+if(SUPPORT_GPU)
     set(LITE_SRC
             ${LITE_SRC}
             ${CMAKE_CURRENT_SOURCE_DIR}/runtime/kernel/opencl/opencl_kernel.cc
@@ -54,10 +51,10 @@ if (SUPPORT_GPU)
             ${CMAKE_CURRENT_SOURCE_DIR}/runtime/opencl/opencl_runtime.cc
             ${CMAKE_CURRENT_SOURCE_DIR}/runtime/opencl/opencl_wrapper.cc
             )
-endif ()
+endif()
 
 
-if (SUPPORT_TRAIN)
+if(SUPPORT_TRAIN)
     set(ANF_SRC
             ${ANF_SRC}
             )
@@ -70,7 +67,7 @@ if (SUPPORT_TRAIN)
             ${CMAKE_CURRENT_SOURCE_DIR}/train/train_model.cc
             ${CMAKE_CURRENT_SOURCE_DIR}/lite_session.cc
             )
-endif ()
+endif()
 
 add_subdirectory(ops)
 add_subdirectory(runtime/kernel/arm)
@@ -85,53 +82,54 @@ set_target_properties(mindspore-lite_static PROPERTIES OUTPUT_NAME "mindspore-li
 set_target_properties(mindspore-lite_static PROPERTIES CLEAN_DIRECT_OUTPUT 1)
 
 set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -Wno-unused-private-field")
-if (SUPPORT_GPU)
+if(SUPPORT_GPU)
     add_subdirectory(runtime/kernel/opencl)
     target_link_libraries(mindspore-lite cpu_kernel_mid opencl_kernel_mid nnacl cpu_ops_mid)
     target_link_libraries(mindspore-lite_static cpu_kernel_mid opencl_kernel_mid nnacl_mid cpu_ops_mid)
-else ()
+else()
     target_link_libraries(mindspore-lite cpu_kernel_mid nnacl cpu_ops_mid)
     target_link_libraries(mindspore-lite_static cpu_kernel_mid nnacl_mid cpu_ops_mid)
-endif ()
-if (SUPPORT_NPU)
+endif()
+if(SUPPORT_NPU)
     add_subdirectory(runtime/agent/npu)
     include_directories(${DDK_PATH})
     target_link_libraries(mindspore-lite npu_kernel_mid)
     target_link_libraries(mindspore-lite_static npu_kernel_mid)
-endif ()
-if (PLATFORM_ARM32 OR PLATFORM_ARM64)
+endif()
+if(PLATFORM_ARM32 OR PLATFORM_ARM64)
     target_link_libraries(mindspore-lite log)
     target_link_libraries(mindspore-lite_static log)
-endif ()
-if (BUILD_MINDDATA STREQUAL "lite")
+endif()
+if(BUILD_MINDDATA STREQUAL "lite")
     target_link_libraries(mindspore-lite minddata_eager_mid minddata-lite)
     target_link_libraries(mindspore-lite_static minddata_eager_mid)
-endif ()
+endif()
 
-if ("${CMAKE_BUILD_TYPE}" STREQUAL "Release" AND PLATFORM_ARM)
-    add_custom_command(TARGET mindspore-lite POST_BUILD
-            COMMAND ${ANDROID_NDK}/toolchains/aarch64-linux-android-4.9/prebuilt/linux-x86_64/aarch64-linux-android/bin/strip
+if(PLATFORM_ARM)
+    set(NDK_STRIP
+            "${ANDROID_NDK}/toolchains/aarch64-linux-android-4.9/prebuilt/linux-x86_64/aarch64-linux-android/bin/strip")
+endif()
+
+if("${CMAKE_BUILD_TYPE}" STREQUAL "Release" AND PLATFORM_ARM)
+    add_custom_command(TARGET mindspore-lite POST_BUILD COMMAND ${NDK_STRIP}
             ${CMAKE_BINARY_DIR}/src/libmindspore-lite.so)
-endif ()
+endif()
 
-if ("${CMAKE_BUILD_TYPE}" STREQUAL "Release")
-    if (PLATFORM_ARM)
-        add_custom_command(TARGET mindspore-lite POST_BUILD
-                COMMAND ${ANDROID_NDK}/toolchains/aarch64-linux-android-4.9/prebuilt/linux-x86_64/aarch64-linux-android/bin/strip
+if("${CMAKE_BUILD_TYPE}" STREQUAL "Release")
+    if(PLATFORM_ARM)
+        add_custom_command(TARGET mindspore-lite POST_BUILD COMMAND ${NDK_STRIP}
                 ${CMAKE_BINARY_DIR}/src/libmindspore-lite.so)
-    elseif (NOT WIN32)
-        add_custom_command(TARGET mindspore-lite POST_BUILD
-                COMMAND strip ${CMAKE_BINARY_DIR}/src/libmindspore-lite.so)
-    endif ()
-endif ()
+    elseif(NOT WIN32)
+        add_custom_command(TARGET mindspore-lite POST_BUILD COMMAND strip ${CMAKE_BINARY_DIR}/src/libmindspore-lite.so)
+    endif()
+endif()
 
-########################## build optimize and float16 library #################################3
-if (PLATFORM_ARM64)
+########################## build optimize and float16 library #################################
+if(PLATFORM_ARM64)
     target_link_libraries(mindspore-lite cpu_opt_kernel_mid nnacl_optimize_mid)
     target_link_libraries(mindspore-lite_static cpu_opt_kernel_mid nnacl_optimize_mid)
-    if (ENABLE_FP16)
+    if(ENABLE_FP16)
         target_link_libraries(mindspore-lite cpu_fp16_kernel_mid nnacl_fp16_mid)
         target_link_libraries(mindspore-lite_static cpu_fp16_kernel_mid nnacl_fp16_mid)
-    endif ()
-endif ()
-
+    endif()
+endif()

mindspore/lite/src/runtime/opencl/opencl_allocator.cc

@@ -94,11 +94,14 @@ void *OpenCLAllocator::CreateImage2D(size_t size, const std::vector<size_t> &img
   MS_ASSERT(buffer);
   MS_ASSERT(image);
   MS_ASSERT(img_size.size() == 3);
-  cl::ImageFormat image_format(CL_RGBA, img_size[2]);
   if (data == nullptr) {
-    *image = new (std::nothrow)
-      cl::Image2D(*ocl_runtime_->Context(), image_format, **buffer, img_size[0], img_size[1], 0, &ret);
+    // copy from cl2.hpp
+    cl_image_desc desc = {CL_MEM_OBJECT_IMAGE2D, img_size[0], img_size[1], 0, 0, 0, 0, 0, 0, (**buffer).get()};
+    const cl::Context &context = *ocl_runtime_->Context();
+    cl_image_format image_format{CL_RGBA, static_cast<uint32_t>(img_size[2])};
+    *image = new (std::nothrow) cl::Image2D(clCreateImage(context.get(), 0, &image_format, &desc, nullptr, &ret));
   } else {
+    cl::ImageFormat image_format(CL_RGBA, img_size[2]);
     *image = new (std::nothrow) cl::Image2D(*ocl_runtime_->Context(), CL_MEM_READ_ONLY | CL_MEM_COPY_HOST_PTR,
                                             image_format, img_size[0], img_size[1], 0, data, &ret);
   }

mindspore/lite/src/runtime/opencl/opencl_runtime.cc

@@ -562,7 +562,7 @@ int OpenCLRuntime::MapBuffer(void *host_ptr, int flags, size_t size, cl::Command
   if (command_queue == nullptr) {
     command_queue = default_command_queue_;
   }
-  return command_queue->enqueueMapSVM(host_ptr, sync, flags, size);
+  return clEnqueueSVMMap(command_queue->get(), sync, flags, host_ptr, size, 0, nullptr, nullptr);
 }
 
 void *OpenCLRuntime::MapBuffer(const cl::Image2D &buffer, bool sync, int flags, const std::vector<size_t> &region,
@@ -591,7 +591,7 @@ int OpenCLRuntime::UnmapBuffer(void *host_ptr, cl::CommandQueue *command_queue)
   if (command_queue == nullptr) {
     command_queue = default_command_queue_;
   }
-  return command_queue->enqueueUnmapSVM(host_ptr);
+  return clEnqueueSVMUnmap(command_queue->get(), host_ptr, 0, nullptr, nullptr);
 }
 
 bool OpenCLRuntime::SyncCommandQueue(cl::CommandQueue *command_queue) {

mindspore/lite/src/runtime/opencl/opencl_runtime.h

@@ -83,7 +83,7 @@ class OpenCLRuntime {
         auto svm_capabilities = GetSVMCapabilities();
         if (svm_capabilities) {
           MS_LOG(DEBUG) << "Set kernel arg[" << index << "] SVM pointer " << value;
-          return kernel.setArg(index, value);
+          return clSetKernelArgSVMPointer(kernel.get(), index, value);
         }
         cl::Buffer *buffer = reinterpret_cast<cl::Buffer *>(allocator_->GetBuffer(value));
         MS_LOG(DEBUG) << "Set kernel arg[" << index << "] OpenCL Buffer " << buffer << ", host_ptr: " << value;

mindspore/lite/src/runtime/opencl/opencl_wrapper.cc

@@ -142,13 +142,13 @@ bool LoadLibraryFromPath(const std::string &library_path, void **handle_ptr) {
   LOAD_OPENCL_FUNCTION_PTR(clEnqueueCopyImage);
   LOAD_OPENCL_FUNCTION_PTR(clEnqueueCopyBufferToImage);
   LOAD_OPENCL_FUNCTION_PTR(clEnqueueCopyImageToBuffer);
-#if CL_HPP_TARGET_OPENCL_VERSION >= 120
+#if CL_TARGET_OPENCL_VERSION >= 120
   LOAD_OPENCL_FUNCTION_PTR(clRetainDevice);
   LOAD_OPENCL_FUNCTION_PTR(clReleaseDevice);
   LOAD_OPENCL_FUNCTION_PTR(clCreateImage);
   LOAD_OPENCL_FUNCTION_PTR(clEnqueueFillImage);
 #endif
-#if CL_HPP_TARGET_OPENCL_VERSION >= 200
+#if CL_TARGET_OPENCL_VERSION >= 200
   LOAD_OPENCL_FUNCTION_PTR(clCreateCommandQueueWithProperties);
   LOAD_OPENCL_FUNCTION_PTR(clGetExtensionFunctionAddress);
   LOAD_OPENCL_FUNCTION_PTR(clSVMAlloc);
@@ -232,13 +232,13 @@ CL_DEFINE_FUNC_PTR(clGetEventProfilingInfo);
 CL_DEFINE_FUNC_PTR(clGetImageInfo);
 CL_DEFINE_FUNC_PTR(clEnqueueCopyBufferToImage);
 CL_DEFINE_FUNC_PTR(clEnqueueCopyImageToBuffer);
-#if CL_HPP_TARGET_OPENCL_VERSION >= 120
+#if CL_TARGET_OPENCL_VERSION >= 120
 CL_DEFINE_FUNC_PTR(clRetainDevice);
 CL_DEFINE_FUNC_PTR(clReleaseDevice);
 CL_DEFINE_FUNC_PTR(clCreateImage);
 CL_DEFINE_FUNC_PTR(clEnqueueFillImage);
 #endif
-#if CL_HPP_TARGET_OPENCL_VERSION >= 200
+#if CL_TARGET_OPENCL_VERSION >= 200
 CL_DEFINE_FUNC_PTR(clGetKernelSubGroupInfoKHR);
 CL_DEFINE_FUNC_PTR(clCreateCommandQueueWithProperties);
 CL_DEFINE_FUNC_PTR(clGetExtensionFunctionAddress);
@@ -651,7 +651,7 @@ cl_int clEnqueueCopyImageToBuffer(cl_command_queue command_queue, cl_mem src_ima
               event_wait_list, event);
 }
 
-#if CL_HPP_TARGET_OPENCL_VERSION >= 120
+#if CL_TARGET_OPENCL_VERSION >= 120
 
 // clRetainDevice wrapper, use OpenCLWrapper function.
 cl_int clRetainDevice(cl_device_id device) {
@@ -685,7 +685,7 @@ cl_int clEnqueueFillImage(cl_command_queue command_queue, cl_mem image, const vo
 
 #endif
 
-#if CL_HPP_TARGET_OPENCL_VERSION >= 200
+#if CL_TARGET_OPENCL_VERSION >= 200
 
 // clCreateCommandQueueWithProperties wrapper, use OpenCLWrapper function.
 cl_command_queue clCreateCommandQueueWithProperties(cl_context context, cl_device_id device,

mindspore/lite/src/runtime/opencl/opencl_wrapper.h

@@ -110,7 +110,7 @@ using clEnqueueCopyBufferToImageFunc = cl_int(CL_API_CALL *)(cl_command_queue, c
 using clEnqueueCopyImageToBufferFunc = cl_int(CL_API_CALL *)(cl_command_queue, cl_mem, cl_mem, const size_t *,
                                                              const size_t *, size_t, cl_uint, const cl_event *,
                                                              cl_event *);
-#if CL_HPP_TARGET_OPENCL_VERSION >= 120
+#if CL_TARGET_OPENCL_VERSION >= 120
 using clRetainDeviceFunc = cl_int (*)(cl_device_id);
 using clReleaseDeviceFunc = cl_int (*)(cl_device_id);
 using clCreateImageFunc = cl_mem (*)(cl_context, cl_mem_flags, const cl_image_format *, const cl_image_desc *, void *,
@@ -118,7 +118,7 @@ using clCreateImageFunc = cl_mem (*)(cl_context, cl_mem_flags, const cl_image_fo
 using clEnqueueFillImageFunc = cl_int (*)(cl_command_queue, cl_mem, const void *, const size_t *, const size_t *,
                                           cl_uint, const cl_event *, cl_event *);
 #endif
-#if CL_HPP_TARGET_OPENCL_VERSION >= 200
+#if CL_TARGET_OPENCL_VERSION >= 200
 using clCreateProgramWithILFunc = cl_program (*)(cl_context, const void *, size_t, cl_int *);
 using clSVMAllocFunc = void *(*)(cl_context, cl_mem_flags, size_t size, cl_uint);
 using clSVMFreeFunc = void (*)(cl_context, void *);
@@ -185,13 +185,13 @@ CL_DECLARE_FUNC_PTR(clGetEventProfilingInfo);
 CL_DECLARE_FUNC_PTR(clGetImageInfo);
 CL_DECLARE_FUNC_PTR(clEnqueueCopyBufferToImage);
 CL_DECLARE_FUNC_PTR(clEnqueueCopyImageToBuffer);
-#if CL_HPP_TARGET_OPENCL_VERSION >= 120
+#if CL_TARGET_OPENCL_VERSION >= 120
 CL_DECLARE_FUNC_PTR(clRetainDevice);
 CL_DECLARE_FUNC_PTR(clReleaseDevice);
 CL_DECLARE_FUNC_PTR(clCreateImage);
 CL_DECLARE_FUNC_PTR(clEnqueueFillImage);
 #endif
-#if CL_HPP_TARGET_OPENCL_VERSION >= 200
+#if CL_TARGET_OPENCL_VERSION >= 200
 CL_DECLARE_FUNC_PTR(clGetKernelSubGroupInfoKHR);
 CL_DECLARE_FUNC_PTR(clCreateCommandQueueWithProperties);
 CL_DECLARE_FUNC_PTR(clGetExtensionFunctionAddress);