cl_khr_icd 2.0.0 loader managed dispatch

.github/workflows/windows.yml

@@ -168,17 +168,29 @@ jobs:
         if /I "${{matrix.BIN}}"=="x64" set REG=reg
         if /I "${{matrix.BIN}}"=="x86" set REG=%systemroot%\Syswow64\reg.exe
         %REG% ADD HKEY_LOCAL_MACHINE\SOFTWARE\Khronos\OpenCL\Vendors /v %GITHUB_WORKSPACE%/build/Release/OpenCLDriverStub.dll /t REG_DWORD /d 0
-        %CTEST_EXE% -C Release --output-on-failure --parallel %NUMBER_OF_PROCESSORS%
+        %CTEST_EXE% -C Release --output-on-failure --parallel %NUMBER_OF_PROCESSORS% -R opencl_icd_loader_test^|cllayerinfo_test
         if errorlevel 1 (
           exit /b %errorlevel%
         )
         %REG% DELETE HKEY_LOCAL_MACHINE\SOFTWARE\Khronos\OpenCL\Vendors /v %GITHUB_WORKSPACE%/build/Release/OpenCLDriverStub.dll /f
+        %REG% ADD HKEY_LOCAL_MACHINE\SOFTWARE\Khronos\OpenCL\Vendors /v %GITHUB_WORKSPACE%/build/Release/OpenCLDriverStubICD2.dll /t REG_DWORD /d 0
+        %CTEST_EXE% -C Release --output-on-failure --parallel %NUMBER_OF_PROCESSORS% -R opencl_icd_loader_icd2_test^|opencl_icd_loader_icd2_instance_test
+        if errorlevel 1 (
+          exit /b %errorlevel%
+        )
+        %REG% DELETE HKEY_LOCAL_MACHINE\SOFTWARE\Khronos\OpenCL\Vendors /v %GITHUB_WORKSPACE%/build/Release/OpenCLDriverStubICD2.dll /f
         %REG% ADD HKEY_LOCAL_MACHINE\SOFTWARE\Khronos\OpenCL\Vendors /v %GITHUB_WORKSPACE%/build/Debug/OpenCLDriverStub.dll /t REG_DWORD /d 0
-        %CTEST_EXE% -C Debug --output-on-failure --parallel %NUMBER_OF_PROCESSORS%
+        %CTEST_EXE% -C Debug --output-on-failure --parallel %NUMBER_OF_PROCESSORS% -R opencl_icd_loader_test^|cllayerinfo_test
         if errorlevel 1 (
           exit /b %errorlevel%
         )
         %REG% DELETE HKEY_LOCAL_MACHINE\SOFTWARE\Khronos\OpenCL\Vendors /v %GITHUB_WORKSPACE%/build/Debug/OpenCLDriverStub.dll /f
+        %REG% ADD HKEY_LOCAL_MACHINE\SOFTWARE\Khronos\OpenCL\Vendors /v %GITHUB_WORKSPACE%/build/Debug/OpenCLDriverStubICD2.dll /t REG_DWORD /d 0
+        %CTEST_EXE% -C Debug --output-on-failure --parallel %NUMBER_OF_PROCESSORS% -R opencl_icd_loader_icd2_test^|opencl_icd_loader_icd2_instance_test
+        if errorlevel 1 (
+          exit /b %errorlevel%
+        )
+        %REG% DELETE HKEY_LOCAL_MACHINE\SOFTWARE\Khronos\OpenCL\Vendors /v %GITHUB_WORKSPACE%/build/Debug/OpenCLDriverStubICD2.dll /f
 
     - name: Install
       shell: cmd

CMakeLists.txt

@@ -52,6 +52,7 @@ endif()
 # It is currently needed default while the specification is being formalized,
 # and to study the performance impact.
 option (ENABLE_OPENCL_LAYERS "Enable OpenCL Layers" ON)
+option (ENABLE_OPENCL_LOADER_MANAGED_DISPATCH "Enable OpenCL Loader managed dispatch" ON)
 include(CMakeDependentOption)
 cmake_dependent_option(ENABLE_OPENCL_LAYERINFO "Enable building cllayerinfo tool" ON ENABLE_OPENCL_LAYERS OFF)
 
@@ -65,6 +66,7 @@ configure_file(${CMAKE_CURRENT_SOURCE_DIR}/loader/icd_cmake_config.h.in
     ${CMAKE_CURRENT_BINARY_DIR}/icd_cmake_config.h)
 
 set (OPENCL_ICD_LOADER_SOURCES
+    include/cl_khr_icd2.h
     loader/icd.c
     loader/icd.h
     loader/icd_version.h
@@ -73,6 +75,7 @@ set (OPENCL_ICD_LOADER_SOURCES
     loader/icd_dispatch_generated.c
     loader/icd_envvars.h
     loader/icd_platform.h)
+include_directories (include)
 
 if (WIN32)
     list (APPEND OPENCL_ICD_LOADER_SOURCES
@@ -158,6 +161,7 @@ set (OPENCL_COMPILE_DEFINITIONS
     OPENCL_ICD_LOADER_VERSION_MINOR=0
     OPENCL_ICD_LOADER_VERSION_REV=6
     $<$<BOOL:${ENABLE_OPENCL_LAYERS}>:CL_ENABLE_LAYERS>
+    $<$<BOOL:${ENABLE_OPENCL_LOADER_MANAGED_DISPATCH}>:CL_ENABLE_LOADER_MANAGED_DISPATCH>
 )
 
 target_compile_definitions (OpenCL

include/cl_khr_icd2.h

@@ -0,0 +1,86 @@
+#if !defined(CL_PLATFORM_UNLOADABLE)
+#define CL_PLATFORM_UNLOADABLE_KHR 0x0921
+#endif
+
+#if defined(CL_ENABLE_LAYERS) && !defined(CL_LAYER_API_VERSION_200)
+#define CL_LAYER_API_VERSION_200 200
+
+typedef cl_int CL_API_CALL
+clDeinitLayer_t(void);
+
+typedef clDeinitLayer_t *pfn_clDeinitLayer;
+#endif //defined(CL_ENABLE_LAYERS) && !defined(CL_LAYER_API_VERSION_200)
+
+#if !defined(CL_ICD2_TAG_KHR)
+#if INTPTR_MAX == INT32_MAX
+#define CL_ICD2_TAG_KHR ((intptr_t)0x434C3331)
+#else
+#define CL_ICD2_TAG_KHR ((intptr_t)0x4F50454E434C3331)
+#endif
+
+typedef void * CL_API_CALL
+clIcdGetFunctionAddressForPlatformKHR_t(
+    cl_platform_id platform,
+    const char* func_name);
+
+typedef clIcdGetFunctionAddressForPlatformKHR_t *
+clIcdGetFunctionAddressForPlatformKHR_fn;
+
+typedef cl_int CL_API_CALL
+clIcdSetPlatformDispatchDataKHR_t(
+    cl_platform_id platform,
+    void *disp_data);
+
+typedef clIcdSetPlatformDispatchDataKHR_t *
+clIcdSetPlatformDispatchDataKHR_fn;
+#endif // !defined(CL_ICD2_TAG_KHR)
+
+#if !defined(CL_INVALID_INSTANCE_KHR)
+#define CL_INVALID_INSTANCE_KHR                         -1154
+
+#define CL_INSTANCE_PROPERTIES_LIST_END_KHR                 0
+
+typedef struct _cl_instance_khr *cl_instance_khr;
+typedef intptr_t cl_instance_properties_khr;
+
+typedef cl_instance_khr CL_API_CALL
+clCreateInstanceKHR_t(
+    const cl_instance_properties_khr *properties,
+    cl_int *errcode_ret);
+
+typedef clCreateInstanceKHR_t *
+clCreateInstanceKHR_fn;
+
+typedef cl_int CL_API_CALL
+clDestroyInstanceKHR_t(
+    cl_instance_khr instance);
+
+typedef clDestroyInstanceKHR_t *
+clDestroyInstanceKHR_fn;
+
+typedef cl_int CL_API_CALL
+clGetPlatformIDsForInstanceKHR_t(
+    cl_instance_khr instance,
+    cl_uint num_entries,
+    cl_platform_id *platforms,
+    cl_uint *num_platforms);
+
+typedef clGetPlatformIDsForInstanceKHR_t *
+clGetPlatformIDsForInstanceKHR_fn;
+
+typedef cl_platform_id CL_API_CALL
+clIcdCreateInstancePlatformKHR_t(
+    cl_platform_id platform,
+    cl_int *errcode_ret);
+
+typedef clIcdCreateInstancePlatformKHR_t *
+clIcdCreateInstancePlatformKHR_fn;
+
+typedef cl_int CL_API_CALL
+clIcdDestroyInstancePlatformKHR_t(
+    cl_platform_id platform);
+
+typedef clIcdDestroyInstancePlatformKHR_t *
+clIcdDestroyInstancePlatformKHR_fn;
+
+#endif // !defined(CL_INVALID_INSTANCE_KHR)

loader/cllayerinfo.c

@@ -19,7 +19,6 @@
 #include "icd.h"
 #include <stdio.h>
 #include <stdlib.h>
-#include <CL/cl_layer.h>
 #if defined(_WIN32)
 #include <io.h>
 #include <share.h>
@@ -90,7 +89,7 @@ static void restore_outputs(void)
 void printLayerInfo(const struct KHRLayer *layer)
 {
     cl_layer_api_version api_version = 0;
-    pfn_clGetLayerInfo p_clGetLayerInfo = (pfn_clGetLayerInfo)(size_t)layer->p_clGetLayerInfo;
+    pfn_clGetLayerInfo p_clGetLayerInfo = layer->p_clGetLayerInfo;
     cl_int result = CL_SUCCESS;
     size_t sz;