From bfa0f02303457acb69269ed11079b1225809564d Mon Sep 17 00:00:00 2001 From: peng xiao Date: Thu, 9 May 2013 17:15:26 +0800 Subject: [PATCH 1/2] Rewrite queryDeviceInfo interface. Previously the function may cause some unsafe issue. It is fixed now by introducing a template parameter. --- .../ocl/include/opencv2/ocl/private/util.hpp | 17 ++-- modules/ocl/src/initialization.cpp | 89 +++++++------------ 2 files changed, 45 insertions(+), 61 deletions(-) diff --git a/modules/ocl/include/opencv2/ocl/private/util.hpp b/modules/ocl/include/opencv2/ocl/private/util.hpp index 77d52bfb7..76d3faf5d 100644 --- a/modules/ocl/include/opencv2/ocl/private/util.hpp +++ b/modules/ocl/include/opencv2/ocl/private/util.hpp @@ -125,16 +125,21 @@ namespace cv bool CV_EXPORTS support_image2d(Context *clCxt = Context::getContext()); // the enums are used to query device information - // currently only support wavefront size queries enum DEVICE_INFO { - WAVEFRONT_SIZE, //in AMD speak - WARP_SIZE = WAVEFRONT_SIZE, //in nvidia speak - IS_CPU_DEVICE //check if the device is CPU + WAVEFRONT_SIZE, + IS_CPU_DEVICE //supports }; - //info should have been pre-allocated - void CV_EXPORTS queryDeviceInfo(DEVICE_INFO info_type, void* info); + template + _ty queryDeviceInfo(cl_kernel kernel = NULL); + //only these three specializations are implemented at the moment + template<> + int CV_EXPORTS queryDeviceInfo(cl_kernel kernel); + template<> + size_t CV_EXPORTS queryDeviceInfo(cl_kernel kernel); + template<> + bool CV_EXPORTS queryDeviceInfo(cl_kernel kernel); }//namespace ocl }//namespace cv diff --git a/modules/ocl/src/initialization.cpp b/modules/ocl/src/initialization.cpp index 7dc8d5718..55d52c454 100644 --- a/modules/ocl/src/initialization.cpp +++ b/modules/ocl/src/initialization.cpp @@ -355,64 +355,43 @@ namespace cv clFinish(Context::getContext()->impl->clCmdQueue); } - void queryDeviceInfo(DEVICE_INFO info_type, void* info) + //template specializations of queryDeviceInfo + template<> + bool queryDeviceInfo(cl_kernel) { - static Info::Impl* impl = Context::getContext()->impl; - switch(info_type) - { - case WAVEFRONT_SIZE: - { - bool is_cpu = false; - queryDeviceInfo(IS_CPU_DEVICE, &is_cpu); - if(is_cpu) - { - *(int*)info = 1; - return; - } -#ifdef CL_DEVICE_WAVEFRONT_WIDTH_AMD - try - { - openCLSafeCall(clGetDeviceInfo(Context::getContext()->impl->devices[0], - CL_DEVICE_WAVEFRONT_WIDTH_AMD, sizeof(size_t), info, 0)); - } - catch(const cv::Exception&) -#elif defined (CL_DEVICE_WARP_SIZE_NV) - const int EXT_LEN = 4096 + 1 ; - char extends_set[EXT_LEN]; - size_t extends_size; - openCLSafeCall(clGetDeviceInfo(impl->devices[impl->devnum], CL_DEVICE_EXTENSIONS, EXT_LEN, (void *)extends_set, &extends_size)); - extends_set[EXT_LEN - 1] = 0; - if(std::string(extends_set).find("cl_nv_device_attribute_query") != std::string::npos) - { - openCLSafeCall(clGetDeviceInfo(Context::getContext()->impl->devices[0], - CL_DEVICE_WARP_SIZE_NV, sizeof(size_t), info, 0)); - } - else -#endif - { - // if no way left for us to query the warp size, we can get it from kernel group info - static const char * _kernel_string = "__kernel void test_func() {}"; - cl_kernel kernel; - kernel = openCLGetKernelFromSource(Context::getContext(), &_kernel_string, "test_func"); - openCLSafeCall(clGetKernelWorkGroupInfo(kernel, impl->devices[impl->devnum], - CL_KERNEL_PREFERRED_WORK_GROUP_SIZE_MULTIPLE, sizeof(size_t), info, NULL)); - } + Info::Impl* impl = Context::getContext()->impl; + cl_device_type devicetype; + openCLSafeCall(clGetDeviceInfo(impl->devices[impl->devnum], + CL_DEVICE_TYPE, sizeof(cl_device_type), + &devicetype, NULL)); + return (devicetype == CVCL_DEVICE_TYPE_CPU); + } - } - break; - case IS_CPU_DEVICE: - { - cl_device_type devicetype; - openCLSafeCall(clGetDeviceInfo(impl->devices[impl->devnum], - CL_DEVICE_TYPE, sizeof(cl_device_type), - &devicetype, NULL)); - *(bool*)info = (devicetype == CVCL_DEVICE_TYPE_CPU); - } - break; - default: - CV_Error(-1, "Invalid device info type"); - break; + template + static _ty queryWavesize(cl_kernel kernel) + { + size_t info = 0; + Info::Impl* impl = Context::getContext()->impl; + bool is_cpu = queryDeviceInfo(); + if(is_cpu) + { + return 1; } + CV_Assert(kernel != NULL); + openCLSafeCall(clGetKernelWorkGroupInfo(kernel, impl->devices[impl->devnum], + CL_KERNEL_PREFERRED_WORK_GROUP_SIZE_MULTIPLE, sizeof(size_t), &info, NULL)); + return static_cast<_ty>(info); + } + + template<> + size_t queryDeviceInfo(cl_kernel kernel) + { + return queryWavesize(kernel); + } + template<> + int queryDeviceInfo(cl_kernel kernel) + { + return queryWavesize(kernel); } void openCLReadBuffer(Context *clCxt, cl_mem dst_buffer, void *host_buffer, size_t size) From e5ea0181855c05f3aa7f82274fe9b24fccb9e771 Mon Sep 17 00:00:00 2001 From: peng xiao Date: Thu, 9 May 2013 17:19:51 +0800 Subject: [PATCH 2/2] Let surf.ocl use the new queryDeviceInfo interface. --- modules/nonfree/src/surf.ocl.cpp | 34 ++++++++----------- .../ocl/include/opencv2/ocl/private/util.hpp | 2 +- 2 files changed, 16 insertions(+), 20 deletions(-) diff --git a/modules/nonfree/src/surf.ocl.cpp b/modules/nonfree/src/surf.ocl.cpp index 508fb3f12..6839294a2 100644 --- a/modules/nonfree/src/surf.ocl.cpp +++ b/modules/nonfree/src/surf.ocl.cpp @@ -59,32 +59,28 @@ namespace cv const char noImage2dOption [] = "-D DISABLE_IMAGE2D"; - static char SURF_OPTIONS [1024] = ""; - static bool USE_IMAGE2d = false; + static bool use_image2d = false; + static void openCLExecuteKernelSURF(Context *clCxt , const char **source, String kernelName, size_t globalThreads[3], size_t localThreads[3], std::vector< std::pair > &args, int channels, int depth) { - char * pSURF_OPTIONS = SURF_OPTIONS; - static bool OPTION_INIT = false; - if(!OPTION_INIT) + char optBuf [100] = {0}; + char * optBufPtr = optBuf; + if( !use_image2d ) { - if( !USE_IMAGE2d ) - { - strcat(pSURF_OPTIONS, noImage2dOption); - pSURF_OPTIONS += strlen(noImage2dOption); - } - - size_t wave_size = 0; - queryDeviceInfo(WAVEFRONT_SIZE, &wave_size); - std::sprintf(pSURF_OPTIONS, "-D WAVE_SIZE=%d", static_cast(wave_size)); - OPTION_INIT = true; + strcat(optBufPtr, noImage2dOption); + optBufPtr += strlen(noImage2dOption); } - openCLExecuteKernel(clCxt, source, kernelName, globalThreads, localThreads, args, channels, depth, SURF_OPTIONS); + cl_kernel kernel; + kernel = openCLGetKernelFromSource(clCxt, source, kernelName, optBufPtr); + size_t wave_size = queryDeviceInfo(kernel); + CV_Assert(clReleaseKernel(kernel) == CL_SUCCESS); + sprintf(optBufPtr, "-D WAVE_SIZE=%d", static_cast(wave_size)); + openCLExecuteKernel(clCxt, source, kernelName, globalThreads, localThreads, args, channels, depth, optBufPtr); } } } - static inline size_t divUp(size_t total, size_t grain) { return (total + grain - 1) / grain; @@ -166,11 +162,11 @@ public: { bindImgTex(img, imgTex); bindImgTex(surf_.sum, sumTex); - USE_IMAGE2d = true; + use_image2d = true; } catch (const cv::Exception& e) { - USE_IMAGE2d = false; + use_image2d = false; if(e.code != CL_IMAGE_FORMAT_NOT_SUPPORTED && e.code != -217) { throw e; diff --git a/modules/ocl/include/opencv2/ocl/private/util.hpp b/modules/ocl/include/opencv2/ocl/private/util.hpp index 76d3faf5d..2463fc279 100644 --- a/modules/ocl/include/opencv2/ocl/private/util.hpp +++ b/modules/ocl/include/opencv2/ocl/private/util.hpp @@ -128,7 +128,7 @@ namespace cv enum DEVICE_INFO { WAVEFRONT_SIZE, - IS_CPU_DEVICE //supports + IS_CPU_DEVICE }; template _ty queryDeviceInfo(cl_kernel kernel = NULL);