diff --git a/modules/nonfree/perf/perf_main.cpp b/modules/nonfree/perf/perf_main.cpp index 03f9b71852..2cd609d414 100644 --- a/modules/nonfree/perf/perf_main.cpp +++ b/modules/nonfree/perf/perf_main.cpp @@ -5,7 +5,7 @@ static const char * impls[] = { #ifdef HAVE_CUDA "cuda", #endif -#ifdef HAVE_OPENCL +#ifdef HAVE_OPENCV_OCL "ocl", #endif "plain" diff --git a/modules/nonfree/perf/perf_surf_ocl.cpp b/modules/nonfree/perf/perf_surf_ocl.cpp index 4528b79268..8413ed072c 100644 --- a/modules/nonfree/perf/perf_surf_ocl.cpp +++ b/modules/nonfree/perf/perf_surf_ocl.cpp @@ -59,7 +59,7 @@ typedef perf::TestBaseWithParam OCL_SURF; #define OCL_TEST_CYCLE() for( ; startTimer(), next(); cv::ocl::finish(), stopTimer()) -PERF_TEST_P(OCL_SURF, with_data_transfer, testing::Values(SURF_IMAGES)) +PERF_TEST_P(OCL_SURF, DISABLED_with_data_transfer, testing::Values(SURF_IMAGES)) { string filename = getDataPath(GetParam()); Mat src = imread(filename, IMREAD_GRAYSCALE); @@ -94,7 +94,7 @@ PERF_TEST_P(OCL_SURF, with_data_transfer, testing::Values(SURF_IMAGES)) SANITY_CHECK_NOTHING(); } -PERF_TEST_P(OCL_SURF, without_data_transfer, testing::Values(SURF_IMAGES)) +PERF_TEST_P(OCL_SURF, DISABLED_without_data_transfer, testing::Values(SURF_IMAGES)) { string filename = getDataPath(GetParam()); Mat src = imread(filename, IMREAD_GRAYSCALE); diff --git a/modules/ocl/src/arithm.cpp b/modules/ocl/src/arithm.cpp index 0b316c5ea3..e7638ecc4d 100644 --- a/modules/ocl/src/arithm.cpp +++ b/modules/ocl/src/arithm.cpp @@ -313,32 +313,28 @@ void cv::ocl::compare(const oclMat &src1, const oclMat &src2, oclMat &dst , int enum { SUM = 0, ABS_SUM, SQR_SUM }; -static void arithmetic_sum_buffer_run(const oclMat &src, cl_mem &dst, int groupnum, int type, int ddepth) +static void arithmetic_sum_buffer_run(const oclMat &src, cl_mem &dst, int groupnum, int type, int ddepth, int vlen) { - int ochannels = src.oclchannels(); - int all_cols = src.step / src.elemSize(); - int pre_cols = (src.offset % src.step) / src.elemSize(); - int sec_cols = all_cols - (src.offset % src.step + src.cols * src.elemSize() - 1) / src.elemSize() - 1; - int invalid_cols = pre_cols + sec_cols; - int cols = all_cols - invalid_cols , elemnum = cols * src.rows;; - int offset = src.offset / src.elemSize(); + int vElemSize = vlen * src.elemSize(); + int src_offset = src.offset / vElemSize, src_step = src.step / vElemSize; + int src_cols = src.cols / vlen, total = src.size().area() / vlen; + + vlen *= src.oclchannels(); const char * const typeMap[] = { "uchar", "char", "ushort", "short", "int", "float", "double" }; const char * const funcMap[] = { "FUNC_SUM", "FUNC_ABS_SUM", "FUNC_SQR_SUM" }; - const char * const channelMap[] = { " ", " ", "2", "4", "4" }; + const char * const channelMap[] = { " ", " ", "2", "4", "4", "", "", "", "8" }; string buildOptions = format("-D srcT=%s%s -D dstT=%s%s -D convertToDstT=convert_%s%s -D %s", - typeMap[src.depth()], channelMap[ochannels], - typeMap[ddepth], channelMap[ochannels], - typeMap[ddepth], channelMap[ochannels], - funcMap[type]); + typeMap[src.depth()], channelMap[vlen], typeMap[ddepth], + channelMap[vlen], typeMap[ddepth], channelMap[vlen], funcMap[type]); vector > args; - args.push_back( make_pair( sizeof(cl_int) , (void *)&cols )); - args.push_back( make_pair( sizeof(cl_int) , (void *)&invalid_cols )); - args.push_back( make_pair( sizeof(cl_int) , (void *)&offset)); - args.push_back( make_pair( sizeof(cl_int) , (void *)&elemnum)); - args.push_back( make_pair( sizeof(cl_int) , (void *)&groupnum)); args.push_back( make_pair( sizeof(cl_mem) , (void *)&src.data)); + args.push_back( make_pair( sizeof(cl_int) , (void *)&src_step )); + args.push_back( make_pair( sizeof(cl_int) , (void *)&src_offset )); + args.push_back( make_pair( sizeof(cl_int) , (void *)&src_cols )); + args.push_back( make_pair( sizeof(cl_int) , (void *)&total )); + args.push_back( make_pair( sizeof(cl_int) , (void *)&groupnum )); args.push_back( make_pair( sizeof(cl_mem) , (void *)&dst )); size_t globalThreads[3] = { groupnum * 256, 1, 1 }; @@ -360,7 +356,11 @@ Scalar arithmetic_sum(const oclMat &src, int type, int ddepth) size_t groupnum = src.clCxt->getDeviceInfo().maxComputeUnits; CV_Assert(groupnum != 0); - int dbsize = groupnum * src.oclchannels(); + int vlen = 8 / src.channels(), vElemSize = vlen * src.elemSize1(); + while (src.offset % vElemSize != 0 || src.step % vElemSize != 0 || src.cols % vlen != 0) + vlen >>= 1, vElemSize >>= 1; + + int dbsize = groupnum * src.oclchannels() * vlen; Context *clCxt = src.clCxt; AutoBuffer _buf(dbsize); @@ -368,12 +368,12 @@ Scalar arithmetic_sum(const oclMat &src, int type, int ddepth) memset(p, 0, dbsize * sizeof(T)); cl_mem dstBuffer = openCLCreateBuffer(clCxt, CL_MEM_WRITE_ONLY, dbsize * sizeof(T)); - arithmetic_sum_buffer_run(src, dstBuffer, groupnum, type, ddepth); + arithmetic_sum_buffer_run(src, dstBuffer, groupnum, type, ddepth, vlen); openCLReadBuffer(clCxt, dstBuffer, (void *)p, dbsize * sizeof(T)); openCLFree(dstBuffer); Scalar s = Scalar::all(0.0); - for (int i = 0; i < dbsize;) + for (int i = 0; i < dbsize; ) for (int j = 0; j < src.oclchannels(); j++, i++) s.val[j] += p[i]; @@ -473,20 +473,13 @@ void cv::ocl::meanStdDev(const oclMat &src, Scalar &mean, Scalar &stddev) ////////////////////////////////////////////////////////////////////////////// template -static void arithmetic_minMax_run(const oclMat &src, const oclMat & mask, cl_mem &dst, int groupnum, string kernelName) +static void arithmetic_minMax_run(const oclMat &src, const oclMat & mask, cl_mem &dst, int vlen, int groupnum) { - int all_cols = src.step / src.elemSize(); - int pre_cols = (src.offset % src.step) / src.elemSize(); - int sec_cols = all_cols - (src.offset % src.step + src.cols * src.elemSize() - 1) / src.elemSize() - 1; - int invalid_cols = pre_cols + sec_cols; - int cols = all_cols - invalid_cols , elemnum = cols * src.rows; - int offset = src.offset / src.elemSize(); - const char * const typeMap[] = { "uchar", "char", "ushort", "short", "int", "float", "double" }; - const char * const channelMap[] = { " ", " ", "2", "4", "4" }; + const char * const channelMap[] = { " ", " ", "2", "4", "4", "", "", "", "8" }; ostringstream stream; - stream << "-D T=" << typeMap[src.depth()] << channelMap[src.channels()]; + stream << "-D T=" << typeMap[src.depth()] << channelMap[vlen]; if (numeric_limits::is_integer) { stream << " -D MAX_VAL=" << (WT)numeric_limits::max(); @@ -494,38 +487,38 @@ static void arithmetic_minMax_run(const oclMat &src, const oclMat & mask, cl_mem } else stream << " -D DEPTH_" << src.depth(); + stream << " -D vlen=" << vlen; std::string buildOptions = stream.str(); + int vElemSize = src.elemSize1() * vlen, src_cols = src.cols / vlen; + int src_step = src.step / vElemSize, src_offset = src.offset / vElemSize; + int mask_step = mask.step / vlen, mask_offset = mask.offset / vlen; + int total = src.size().area() / vlen; + vector > args; args.push_back( make_pair( sizeof(cl_mem) , (void *)&src.data)); - args.push_back( make_pair( sizeof(cl_mem) , (void *)&dst )); - args.push_back( make_pair( sizeof(cl_int) , (void *)&cols )); - args.push_back( make_pair( sizeof(cl_int) , (void *)&invalid_cols )); - args.push_back( make_pair( sizeof(cl_int) , (void *)&offset)); - args.push_back( make_pair( sizeof(cl_int) , (void *)&elemnum)); + args.push_back( make_pair( sizeof(cl_int) , (void *)&src_step )); + args.push_back( make_pair( sizeof(cl_int) , (void *)&src_offset )); + args.push_back( make_pair( sizeof(cl_int) , (void *)&src.rows )); + args.push_back( make_pair( sizeof(cl_int) , (void *)&src_cols )); + args.push_back( make_pair( sizeof(cl_int) , (void *)&total)); args.push_back( make_pair( sizeof(cl_int) , (void *)&groupnum)); + args.push_back( make_pair( sizeof(cl_mem) , (void *)&dst)); - int minvalid_cols = 0, moffset = 0; if (!mask.empty()) { - int mall_cols = mask.step / mask.elemSize(); - int mpre_cols = (mask.offset % mask.step) / mask.elemSize(); - int msec_cols = mall_cols - (mask.offset % mask.step + mask.cols * mask.elemSize() - 1) / mask.elemSize() - 1; - minvalid_cols = mpre_cols + msec_cols; - moffset = mask.offset / mask.elemSize(); - args.push_back( make_pair( sizeof(cl_mem) , (void *)&mask.data )); - args.push_back( make_pair( sizeof(cl_int) , (void *)&minvalid_cols )); - args.push_back( make_pair( sizeof(cl_int) , (void *)&moffset )); + args.push_back( make_pair( sizeof(cl_int) , (void *)&mask_step )); + args.push_back( make_pair( sizeof(cl_int) , (void *)&mask_offset )); - kernelName += "_mask"; + buildOptions += " -D WITH_MASK"; } - size_t globalThreads[3] = {groupnum * 256, 1, 1}; - size_t localThreads[3] = {256, 1, 1}; + size_t globalThreads[3] = { groupnum * 256, 1, 1 }; + size_t localThreads[3] = { 256, 1, 1 }; - // kernel use fixed grid size, replace lt on NULL is imposible without kernel changes - openCLExecuteKernel(src.clCxt, &arithm_minMax, kernelName, globalThreads, localThreads, + // kernel use fixed grid size, replace lt on NULL is impossible without kernel changes + openCLExecuteKernel(src.clCxt, &arithm_minMax, "arithm_op_minMax", globalThreads, localThreads, args, -1, -1, buildOptions.c_str()); } @@ -535,25 +528,33 @@ void arithmetic_minMax(const oclMat &src, double *minVal, double *maxVal, const size_t groupnum = src.clCxt->getDeviceInfo().maxComputeUnits; CV_Assert(groupnum != 0); - int dbsize = groupnum * 2 * src.elemSize(); + int vlen = mask.empty() ? 8 : 1, vElemSize = vlen * src.elemSize1(); + while (src.offset % vElemSize != 0 || src.step % vElemSize != 0 || src.cols % vlen != 0) + { + vlen >>= 1; + vElemSize >>= 1; + } + + int dbsize = groupnum * 2 * vElemSize; oclMat buf; ensureSizeIsEnough(1, dbsize, CV_8UC1, buf); cl_mem buf_data = reinterpret_cast(buf.data); - arithmetic_minMax_run(src, mask, buf_data, groupnum, "arithm_op_minMax"); + arithmetic_minMax_run(src, mask, buf_data, vlen, groupnum); Mat matbuf = Mat(buf); T *p = matbuf.ptr(); + if (minVal != NULL) { *minVal = std::numeric_limits::max(); - for (int i = 0, end = src.oclchannels() * (int)groupnum; i < end; i++) + for (int i = 0, end = vlen * (int)groupnum; i < end; i++) *minVal = *minVal < p[i] ? *minVal : p[i]; } if (maxVal != NULL) { *maxVal = -std::numeric_limits::max(); - for (int i = src.oclchannels() * (int)groupnum, end = i << 1; i < end; i++) + for (int i = vlen * (int)groupnum, end = i << 1; i < end; i++) *maxVal = *maxVal > p[i] ? *maxVal : p[i]; } } @@ -564,7 +565,7 @@ void cv::ocl::minMax(const oclMat &src, double *minVal, double *maxVal, const oc { CV_Assert(src.channels() == 1); CV_Assert(src.size() == mask.size() || mask.empty()); - CV_Assert(src.step % src.elemSize() == 0); + CV_Assert(src.step % src.elemSize1() == 0); if (minVal == NULL && maxVal == NULL) return; @@ -1139,7 +1140,7 @@ static void arithmetic_minMaxLoc_run(const oclMat &src, cl_mem &dst, int vlen , sprintf(build_options, "-D DEPTH_%d -D REPEAT_S%d -D REPEAT_E%d", src.depth(), repeat_s, repeat_e); size_t gt[3] = {groupnum * 256, 1, 1}, lt[3] = {256, 1, 1}; - // kernel use fixed grid size, replace lt on NULL is imposible without kernel changes + // kernel use fixed grid size, replace lt on NULL is impossible without kernel changes openCLExecuteKernel(src.clCxt, &arithm_minMaxLoc, "arithm_op_minMaxLoc", gt, lt, args, -1, -1, build_options); } @@ -1169,7 +1170,7 @@ static void arithmetic_minMaxLoc_mask_run(const oclMat &src, const oclMat &mask, args.push_back( make_pair( sizeof(cl_mem) , (void *)&mask.data )); args.push_back( make_pair( sizeof(cl_mem) , (void *)&dst )); - // kernel use fixed grid size, replace lt on NULL is imposible without kernel changes + // kernel use fixed grid size, replace lt on NULL is impossible without kernel changes openCLExecuteKernel(src.clCxt, &arithm_minMaxLoc_mask, "arithm_op_minMaxLoc_mask", gt, lt, args, -1, -1, build_options); } } @@ -1262,38 +1263,35 @@ void cv::ocl::minMaxLoc(const oclMat &src, double *minVal, double *maxVal, ///////////////////////////// countNonZero /////////////////////////////////// ////////////////////////////////////////////////////////////////////////////// -static void arithmetic_countNonZero_run(const oclMat &src, cl_mem &dst, int groupnum, string kernelName) +static void arithmetic_countNonZero_run(const oclMat &src, cl_mem &dst, int groupnum, int vlen) { - int ochannels = src.oclchannels(); - int all_cols = src.step / src.elemSize(); - int pre_cols = (src.offset % src.step) / src.elemSize(); - int sec_cols = all_cols - (src.offset % src.step + src.cols * src.elemSize() - 1) / src.elemSize() - 1; - int invalid_cols = pre_cols + sec_cols; - int cols = all_cols - invalid_cols , elemnum = cols * src.rows;; - int offset = src.offset / src.elemSize(); + int vElemSize = vlen * src.elemSize1(); + int src_step = src.step / vElemSize, src_offset = src.offset / vElemSize; + int src_cols = src.cols / vlen, total = src.size().area() / vlen; const char * const typeMap[] = { "uchar", "char", "ushort", "short", "int", "float", "double" }; - const char * const channelMap[] = { " ", " ", "2", "4", "4" }; - string buildOptions = format("-D srcT=%s%s -D dstT=int%s", typeMap[src.depth()], channelMap[ochannels], - channelMap[ochannels]); + const char * const channelMap[] = { "", "", "2", "4", "4", "", "", "", "8" }; + string buildOptions = format("-D srcT=%s%s -D dstT=int%s -D convertToDstT=convert_int%s", + typeMap[src.depth()], channelMap[vlen], + channelMap[vlen], channelMap[vlen]); vector > args; - args.push_back( make_pair( sizeof(cl_int) , (void *)&cols )); - args.push_back( make_pair( sizeof(cl_int) , (void *)&invalid_cols )); - args.push_back( make_pair( sizeof(cl_int) , (void *)&offset)); - args.push_back( make_pair( sizeof(cl_int) , (void *)&elemnum)); - args.push_back( make_pair( sizeof(cl_int) , (void *)&groupnum)); args.push_back( make_pair( sizeof(cl_mem) , (void *)&src.data)); + args.push_back( make_pair( sizeof(cl_int) , (void *)&src_step )); + args.push_back( make_pair( sizeof(cl_int) , (void *)&src_offset )); + args.push_back( make_pair( sizeof(cl_int) , (void *)&src_cols )); + args.push_back( make_pair( sizeof(cl_int) , (void *)&total )); + args.push_back( make_pair( sizeof(cl_int) , (void *)&groupnum )); args.push_back( make_pair( sizeof(cl_mem) , (void *)&dst )); size_t globalThreads[3] = { groupnum * 256, 1, 1 }; #ifdef ANDROID - openCLExecuteKernel(src.clCxt, &arithm_nonzero, kernelName, globalThreads, NULL, + openCLExecuteKernel(src.clCxt, &arithm_nonzero, "arithm_op_nonzero", globalThreads, NULL, args, -1, -1, buildOptions.c_str()); #else size_t localThreads[3] = { 256, 1, 1 }; - openCLExecuteKernel(src.clCxt, &arithm_nonzero, kernelName, globalThreads, localThreads, + openCLExecuteKernel(src.clCxt, &arithm_nonzero, "arithm_op_nonzero", globalThreads, localThreads, args, -1, -1, buildOptions.c_str()); #endif } @@ -1310,18 +1308,20 @@ int cv::ocl::countNonZero(const oclMat &src) return -1; } + int vlen = 8, vElemSize = src.elemSize1() * vlen; + while (src.offset % vElemSize != 0 || src.step % vElemSize != 0 || src.cols % vlen != 0) + vlen >>= 1, vElemSize >>= 1; + size_t groupnum = src.clCxt->getDeviceInfo().maxComputeUnits; CV_Assert(groupnum != 0); - int dbsize = groupnum; - - string kernelName = "arithm_op_nonzero"; + int dbsize = groupnum * vlen; AutoBuffer _buf(dbsize); int *p = (int*)_buf, nonzero = 0; memset(p, 0, dbsize * sizeof(int)); cl_mem dstBuffer = openCLCreateBuffer(clCxt, CL_MEM_WRITE_ONLY, dbsize * sizeof(int)); - arithmetic_countNonZero_run(src, dstBuffer, groupnum, kernelName); + arithmetic_countNonZero_run(src, dstBuffer, groupnum, vlen); openCLReadBuffer(clCxt, dstBuffer, (void *)p, dbsize * sizeof(int)); for (int i = 0; i < dbsize; i++) @@ -1336,157 +1336,118 @@ int cv::ocl::countNonZero(const oclMat &src) ////////////////////////////////bitwise_op//////////////////////////////////// ////////////////////////////////////////////////////////////////////////////// -static void bitwise_unary_run(const oclMat &src1, oclMat &dst, string kernelName, const cv::ocl::ProgramEntry* source) +enum { AND = 0, OR, XOR, NOT }; + +static void bitwise_run(const oclMat & src1, const oclMat & src2, const Scalar & src3, const oclMat & mask, + oclMat & dst, int operationType) { - dst.create(src1.size(), src1.type()); - - int channels = dst.oclchannels(); - int depth = dst.depth(); - - int vector_lengths[4][7] = {{4, 4, 4, 4, 1, 1, 1}, - {4, 4, 4, 4, 1, 1, 1}, - {4, 4, 4, 4, 1, 1, 1}, - {4, 4, 4, 4, 1, 1, 1} - }; - - size_t vector_length = vector_lengths[channels - 1][depth]; - int offset_cols = (dst.offset / dst.elemSize1()) & (vector_length - 1); - int cols = divUp(dst.cols * channels + offset_cols, vector_length); - -#ifdef ANDROID - size_t localThreads[3] = { 64, 2, 1 }; -#else - size_t localThreads[3] = { 64, 4, 1 }; -#endif - size_t globalThreads[3] = { cols, dst.rows, 1 }; - - int dst_step1 = dst.cols * dst.elemSize(); - vector > args; - args.push_back( make_pair( sizeof(cl_mem), (void *)&src1.data )); - args.push_back( make_pair( sizeof(cl_int), (void *)&src1.step )); - args.push_back( make_pair( sizeof(cl_int), (void *)&src1.offset )); - args.push_back( make_pair( sizeof(cl_mem), (void *)&dst.data )); - args.push_back( make_pair( sizeof(cl_int), (void *)&dst.step )); - args.push_back( make_pair( sizeof(cl_int), (void *)&dst.offset )); - args.push_back( make_pair( sizeof(cl_int), (void *)&src1.rows )); - args.push_back( make_pair( sizeof(cl_int), (void *)&cols )); - args.push_back( make_pair( sizeof(cl_int), (void *)&dst_step1 )); - - openCLExecuteKernel(src1.clCxt, source, kernelName, globalThreads, localThreads, args, -1, depth); -} - -enum { AND = 0, OR, XOR }; - -static void bitwise_binary_run(const oclMat &src1, const oclMat &src2, const Scalar& src3, const oclMat &mask, - oclMat &dst, int operationType) -{ - CV_Assert(operationType >= AND && operationType <= XOR); - CV_Assert(src2.empty() || (!src2.empty() && src1.type() == src2.type() && src1.size() == src2.size())); - CV_Assert(mask.empty() || (!mask.empty() && mask.type() == CV_8UC1 && mask.size() == src1.size())); + CV_Assert(operationType >= AND && operationType <= NOT); + CV_Assert(src2.empty() || (src1.type() == src2.type() && src1.size() == src2.size())); + CV_Assert(mask.empty() || (mask.type() == CV_8UC1 && mask.size() == src1.size())); dst.create(src1.size(), src1.type()); - oclMat m; + double scalar[4]; - const char operationMap[] = { '&', '|', '^' }; - std::string kernelName("arithm_bitwise_binary"); + bool haveScalar = src2.empty() && operationType != NOT, haveMask = !mask.empty(); + int ocn = dst.oclchannels(), depth = dst.depth(); + const char operationMap[] = { '&', '|', '^', '~' }; + const char * const typeMap[] = { "uchar", "uchar", "ushort", "ushort", "int", "int", "ulong" }; + const char * const channelMap[] = { "", "", "2", "4", "4", "", "", "", "8", "", "", "", "", "", "", "", "16" }; + const int preferredVectorWidth[] = { 4, 4, 2, 2, 1, 1, 1 }; + int kercn = haveMask || haveScalar ? ocn : preferredVectorWidth[depth]; - int vlen = std::min(8, src1.elemSize1() * src1.oclchannels()); - std::string vlenstr = vlen > 1 ? format("%d", vlen) : ""; - std::string buildOptions = format("-D Operation=%c -D vloadn=vload%s -D vstoren=vstore%s -D elemSize=%d -D vlen=%d" - " -D ucharv=uchar%s", - operationMap[operationType], vlenstr.c_str(), vlenstr.c_str(), - (int)src1.elemSize(), vlen, vlenstr.c_str()); - -#ifdef ANDROID - size_t localThreads[3] = { 16, 10, 1 }; -#else - size_t localThreads[3] = { 16, 16, 1 }; -#endif - size_t globalThreads[3] = { dst.cols, dst.rows, 1 }; - - vector > args; - args.push_back( make_pair( sizeof(cl_mem), (void *)&src1.data )); - args.push_back( make_pair( sizeof(cl_int), (void *)&src1.step )); - args.push_back( make_pair( sizeof(cl_int), (void *)&src1.offset )); - - if (src2.empty()) + if (!haveScalar && !haveMask) { - m.create(1, 1, dst.type()); - m.setTo(src3); - - args.push_back( make_pair( sizeof(cl_mem), (void *)&m.data )); - - kernelName += "_scalar"; + int velemsize = dst.elemSize1() * kercn; + while (src1.offset % velemsize != 0 || src1.step % velemsize != 0 || src1.cols * ocn % kercn != 0 || + src2.offset % velemsize != 0 || src2.step % velemsize != 0 || src2.cols * ocn % kercn != 0 || + dst.offset % velemsize != 0 || dst.step % velemsize != 0 || dst.cols * ocn % kercn != 0) + kercn >>= 1, velemsize >>= 1; } - else + + int cols = dst.cols * ocn / kercn; + + std::string buildOptions = format("-D Operation=%c -D T=%s%s", operationMap[operationType], + typeMap[depth], channelMap[kercn]); + + vector > args; + args.push_back( make_pair( sizeof(cl_mem), (void *)&src1.data )); + args.push_back( make_pair( sizeof(cl_int), (void *)&src1.step )); + args.push_back( make_pair( sizeof(cl_int), (void *)&src1.offset )); + + if (haveScalar) + { + int sctype = CV_MAKE_TYPE(dst.depth(), ocn); + cv::scalarToRawData(src3, scalar, sctype); + + args.push_back( make_pair( CV_ELEM_SIZE(sctype), (void *)scalar )); + + buildOptions += " -D HAVE_SCALAR"; + } + else if (operationType != NOT) { args.push_back( make_pair( sizeof(cl_mem), (void *)&src2.data )); args.push_back( make_pair( sizeof(cl_int), (void *)&src2.step )); args.push_back( make_pair( sizeof(cl_int), (void *)&src2.offset )); + + buildOptions += " -D OP_BINARY"; } - if (!mask.empty()) + if (haveMask) { args.push_back( make_pair( sizeof(cl_mem), (void *)&mask.data )); args.push_back( make_pair( sizeof(cl_int), (void *)&mask.step )); args.push_back( make_pair( sizeof(cl_int), (void *)&mask.offset )); - kernelName += "_mask"; + buildOptions += " -D HAVE_MASK"; } args.push_back( make_pair( sizeof(cl_mem), (void *)&dst.data )); args.push_back( make_pair( sizeof(cl_int), (void *)&dst.step )); args.push_back( make_pair( sizeof(cl_int), (void *)&dst.offset )); - args.push_back( make_pair( sizeof(cl_int), (void *)&src1.cols )); - args.push_back( make_pair( sizeof(cl_int), (void *)&src1.rows )); + args.push_back( make_pair( sizeof(cl_int), (void *)&dst.rows )); + args.push_back( make_pair( sizeof(cl_int), (void *)&cols )); - openCLExecuteKernel(src1.clCxt, mask.empty() ? (!src2.empty() ? &arithm_bitwise_binary : &arithm_bitwise_binary_scalar) : - (!src2.empty() ? &arithm_bitwise_binary_mask : &arithm_bitwise_binary_scalar_mask), - kernelName, globalThreads, localThreads, + size_t globalsize[3] = { dst.cols * ocn / kercn, dst.rows, 1 }; + globalsize[0] = divUp(globalsize[0], 256) * 256; + openCLExecuteKernel(src1.clCxt, &arithm_bitwise, "arithm_bitwise", globalsize, NULL, args, -1, -1, buildOptions.c_str()); } void cv::ocl::bitwise_not(const oclMat &src, oclMat &dst) { - if (!src.clCxt->supportsFeature(FEATURE_CL_DOUBLE) && src.depth() == CV_64F) - { - CV_Error(CV_OpenCLDoubleNotSupported, "Selected device doesn't support double"); - return; - } - - dst.create(src.size(), src.type()); - bitwise_unary_run(src, dst, "arithm_bitwise_not", &arithm_bitwise_not); + bitwise_run(src, oclMat(), Scalar(), oclMat(), dst, NOT); } void cv::ocl::bitwise_or(const oclMat &src1, const oclMat &src2, oclMat &dst, const oclMat &mask) { - bitwise_binary_run(src1, src2, Scalar(), mask, dst, OR); + bitwise_run(src1, src2, Scalar(), mask, dst, OR); } void cv::ocl::bitwise_or(const oclMat &src1, const Scalar &src2, oclMat &dst, const oclMat &mask) { - bitwise_binary_run(src1, oclMat(), src2, mask, dst, OR); + bitwise_run(src1, oclMat(), src2, mask, dst, OR); } void cv::ocl::bitwise_and(const oclMat &src1, const oclMat &src2, oclMat &dst, const oclMat &mask) { - bitwise_binary_run(src1, src2, Scalar(), mask, dst, AND); + bitwise_run(src1, src2, Scalar(), mask, dst, AND); } void cv::ocl::bitwise_and(const oclMat &src1, const Scalar &src2, oclMat &dst, const oclMat &mask) { - bitwise_binary_run(src1, oclMat(), src2, mask, dst, AND); + bitwise_run(src1, oclMat(), src2, mask, dst, AND); } void cv::ocl::bitwise_xor(const oclMat &src1, const oclMat &src2, oclMat &dst, const oclMat &mask) { - bitwise_binary_run(src1, src2, Scalar(), mask, dst, XOR); + bitwise_run(src1, src2, Scalar(), mask, dst, XOR); } void cv::ocl::bitwise_xor(const oclMat &src1, const Scalar &src2, oclMat &dst, const oclMat &mask) { - bitwise_binary_run(src1, oclMat(), src2, mask, dst, XOR); + bitwise_run(src1, oclMat(), src2, mask, dst, XOR); } oclMat cv::ocl::operator ~ (const oclMat &src) diff --git a/modules/ocl/src/gftt.cpp b/modules/ocl/src/gftt.cpp index 4f24d13588..57ed12f009 100644 --- a/modules/ocl/src/gftt.cpp +++ b/modules/ocl/src/gftt.cpp @@ -146,34 +146,33 @@ static void minMaxEig_caller(const oclMat &src, oclMat &dst, oclMat & tozero) CV_Assert(groupnum != 0); int dbsize = groupnum * 2 * src.elemSize(); - ensureSizeIsEnough(1, dbsize, CV_8UC1, dst); cl_mem dst_data = reinterpret_cast(dst.data); - int all_cols = src.step / src.elemSize(); - int pre_cols = (src.offset % src.step) / src.elemSize(); - int sec_cols = all_cols - (src.offset % src.step + src.cols * src.elemSize() - 1) / src.elemSize() - 1; - int invalid_cols = pre_cols + sec_cols; - int cols = all_cols - invalid_cols , elemnum = cols * src.rows; - int offset = src.offset / src.elemSize(); + int vElemSize = src.elemSize1(); + int src_step = src.step / vElemSize, src_offset = src.offset / vElemSize; + int total = src.size().area(); - {// first parallel pass + { + // first parallel pass vector > args; args.push_back( make_pair( sizeof(cl_mem) , (void *)&src.data)); - args.push_back( make_pair( sizeof(cl_mem) , (void *)&dst_data )); - args.push_back( make_pair( sizeof(cl_int) , (void *)&cols )); - args.push_back( make_pair( sizeof(cl_int) , (void *)&invalid_cols )); - args.push_back( make_pair( sizeof(cl_int) , (void *)&offset)); - args.push_back( make_pair( sizeof(cl_int) , (void *)&elemnum)); + args.push_back( make_pair( sizeof(cl_int) , (void *)&src_step)); + args.push_back( make_pair( sizeof(cl_int) , (void *)&src_offset)); + args.push_back( make_pair( sizeof(cl_int) , (void *)&src.rows )); + args.push_back( make_pair( sizeof(cl_int) , (void *)&src.cols )); + args.push_back( make_pair( sizeof(cl_int) , (void *)&total)); args.push_back( make_pair( sizeof(cl_int) , (void *)&groupnum)); + args.push_back( make_pair( sizeof(cl_mem) , (void *)&dst_data )); size_t globalThreads[3] = {groupnum * 256, 1, 1}; size_t localThreads[3] = {256, 1, 1}; openCLExecuteKernel(src.clCxt, &arithm_minMax, "arithm_op_minMax", globalThreads, localThreads, - args, -1, -1, "-D T=float -D DEPTH_5"); + args, -1, -1, "-D T=float -D DEPTH_5 -D vlen=1"); } - {// run final "serial" kernel to find accumulate results from threads and reset corner counter + { + // run final "serial" kernel to find accumulate results from threads and reset corner counter vector > args; args.push_back( make_pair( sizeof(cl_mem) , (void *)&dst_data )); args.push_back( make_pair( sizeof(cl_int) , (void *)&groupnum )); diff --git a/modules/ocl/src/opencl/arithm_bitwise_binary.cl b/modules/ocl/src/opencl/arithm_bitwise.cl similarity index 67% rename from modules/ocl/src/opencl/arithm_bitwise_binary.cl rename to modules/ocl/src/opencl/arithm_bitwise.cl index 56cd745d29..899d41970f 100644 --- a/modules/ocl/src/opencl/arithm_bitwise_binary.cl +++ b/modules/ocl/src/opencl/arithm_bitwise.cl @@ -48,35 +48,46 @@ /////////////////////////////////////////// bitwise_binary ////////////////////////////////////////// ///////////////////////////////////////////////////////////////////////////////////////////////////// -__kernel void arithm_bitwise_binary(__global uchar * src1, int src1_step, int src1_offset, - __global uchar * src2, int src2_step, int src2_offset, - __global uchar * dst, int dst_step, int dst_offset, - int cols, int rows) +__kernel void arithm_bitwise(__global uchar * src1ptr, int src1_step, int src1_offset, +#ifdef OP_BINARY + __global uchar * src2ptr, int src2_step, int src2_offset, +#elif defined HAVE_SCALAR + T scalar, +#endif +#ifdef HAVE_MASK + __global uchar * mask, int mask_step, int mask_offset, +#endif + __global uchar * dstptr, int dst_step, int dst_offset, int dst_rows, int dst_cols) { int x = get_global_id(0); int y = get_global_id(1); - if (x < cols && y < rows) + if (x < dst_cols && y < dst_rows) { -#if elemSize > 1 - x *= elemSize; +#ifdef HAVE_MASK + mask += mad24(y, mask_step, x + mask_offset); + if (mask[0]) #endif - int src1_index = mad24(y, src1_step, x + src1_offset); - int src2_index = mad24(y, src2_step, x + src2_offset); - int dst_index = mad24(y, dst_step, x + dst_offset); - -#if elemSize > 1 - #pragma unroll - for (int i = 0; i < elemSize; i += vlen) { - ucharv t0 = vloadn(0, src1 + src1_index + i); - ucharv t1 = vloadn(0, src2 + src2_index + i); - ucharv t2 = t0 Operation t1; - - vstoren(t2, 0, dst + dst_index + i); - } -#else - dst[dst_index] = src1[src1_index] Operation src2[src2_index]; + int src1_index = mad24(y, src1_step, mad24(x, (int)sizeof(T), src1_offset)); +#ifdef OP_BINARY + int src2_index = mad24(y, src2_step, mad24(x, (int)sizeof(T), src2_offset)); #endif + int dst_index = mad24(y, dst_step, mad24(x, (int)sizeof(T), dst_offset)); + + __global const T * src1 = (__global const T *)(src1ptr + src1_index); +#ifdef OP_BINARY + __global const T * src2 = (__global const T *)(src2ptr + src2_index); +#endif + __global T * dst = (__global T *)(dstptr + dst_index); + +#ifdef OP_BINARY + dst[0] = src1[0] Operation src2[0]; +#elif defined HAVE_SCALAR + dst[0] = src1[0] Operation scalar; +#else + dst[0] = Operation src1[0]; +#endif + } } } diff --git a/modules/ocl/src/opencl/arithm_bitwise_binary_mask.cl b/modules/ocl/src/opencl/arithm_bitwise_binary_mask.cl deleted file mode 100644 index 328ccd91ae..0000000000 --- a/modules/ocl/src/opencl/arithm_bitwise_binary_mask.cl +++ /dev/null @@ -1,88 +0,0 @@ -/*M/////////////////////////////////////////////////////////////////////////////////////// -// -// IMPORTANT: READ BEFORE DOWNLOADING, COPYING, INSTALLING OR USING. -// -// By downloading, copying, installing or using the software you agree to this license. -// If you do not agree to this license, do not download, install, -// copy or use the software. -// -// -// License Agreement -// For Open Source Computer Vision Library -// -// Copyright (C) 2010-2012, Institute Of Software Chinese Academy Of Science, all rights reserved. -// Copyright (C) 2010-2012, Advanced Micro Devices, Inc., all rights reserved. -// Third party copyrights are property of their respective owners. -// -// @Authors -// Jiang Liyuan, jlyuan001.good@163.com -// Peng Xiao, pengxiao@outlook.com -// -// Redistribution and use in source and binary forms, with or without modification, -// are permitted provided that the following conditions are met: -// -// * Redistribution's of source code must retain the above copyright notice, -// this list of conditions and the following disclaimer. -// -// * Redistribution's in binary form must reproduce the above copyright notice, -// this list of conditions and the following disclaimer in the documentation -// and/or other materials provided with the distribution. -// -// * The name of the copyright holders may not be used to endorse or promote products -// derived from this software without specific prior written permission. -// -// This software is provided by the copyright holders and contributors as is and -// any express or implied warranties, including, but not limited to, the implied -// warranties of merchantability and fitness for a particular purpose are disclaimed. -// In no event shall the Intel Corporation or contributors be liable for any direct, -// indirect, incidental, special, exemplary, or consequential damages -// (including, but not limited to, procurement of substitute goods or services; -// loss of use, data, or profits; or business interruption) however caused -// and on any theory of liability, whether in contract, strict liability, -// or tort (including negligence or otherwise) arising in any way out of -// the use of this software, even if advised of the possibility of such damage. -// -//M*/ - -////////////////////////////////////////////////////////////////////////////////////////////////////// -////////////////////////////////////////////bitwise_binary//////////////////////////////////////////// -////////////////////////////////////////////////////////////////////////////////////////////////////// - -__kernel void arithm_bitwise_binary_mask(__global uchar * src1, int src1_step, int src1_offset, - __global uchar * src2, int src2_step, int src2_offset, - __global uchar * mask, int mask_step, int mask_offset, - __global uchar * dst, int dst_step, int dst_offset, - int cols1, int rows) -{ - int x = get_global_id(0); - int y = get_global_id(1); - - if (x < cols1 && y < rows) - { - int mask_index = mad24(y, mask_step, mask_offset + x); - - if (mask[mask_index]) - { -#if elemSize > 1 - x *= elemSize; -#endif - int src1_index = mad24(y, src1_step, x + src1_offset); - int src2_index = mad24(y, src2_step, x + src2_offset); - int dst_index = mad24(y, dst_step, x + dst_offset); - -#if elemSize > 1 - #pragma unroll - for (int i = 0; i < elemSize; i += vlen) - { - ucharv t0 = vloadn(0, src1 + src1_index + i); - ucharv t1 = vloadn(0, src2 + src2_index + i); - ucharv t2 = t0 Operation t1; - - vstoren(t2, 0, dst + dst_index + i); - } -#else - dst[dst_index] = src1[src1_index] Operation src2[src2_index]; -#endif - } - } -} diff --git a/modules/ocl/src/opencl/arithm_bitwise_binary_scalar.cl b/modules/ocl/src/opencl/arithm_bitwise_binary_scalar.cl deleted file mode 100644 index 434bd5eca8..0000000000 --- a/modules/ocl/src/opencl/arithm_bitwise_binary_scalar.cl +++ /dev/null @@ -1,82 +0,0 @@ -//////////////////////////////////////////////////////////////////////////////////////// -// -// IMPORTANT: READ BEFORE DOWNLOADING, COPYING, INSTALLING OR USING. -// -// By downloading, copying, installing or using the software you agree to this license. -// If you do not agree to this license, do not download, install, -// copy or use the software. -// -// -// License Agreement -// For Open Source Computer Vision Library -// -// Copyright (C) 2010-2012, Institute Of Software Chinese Academy Of Science, all rights reserved. -// Copyright (C) 2010-2012, Advanced Micro Devices, Inc., all rights reserved. -// Third party copyrights are property of their respective owners. -// -// @Authors -// Jiang Liyuan, jlyuan001.good@163.com -// Peng Xiao, pengxiao@outlook.com -// -// Redistribution and use in source and binary forms, with or without modification, -// are permitted provided that the following conditions are met: -// -// * Redistribution's of source code must retain the above copyright notice, -// this list of conditions and the following disclaimer. -// -// * Redistribution's in binary form must reproduce the above copyright notice, -// this list of conditions and the following disclaimer in the documentation -// and/or other materials provided with the distribution. -// -// * The name of the copyright holders may not be used to endorse or promote products -// derived from this software without specific prior written permission. -// -// This software is provided by the copyright holders and contributors as is and -// any express or implied warranties, including, but not limited to, the implied -// warranties of merchantability and fitness for a particular purpose are disclaimed. -// In no event shall the Intel Corporation or contributors be liable for any direct, -// indirect, incidental, special, exemplary, or consequential damages -// (including, but not limited to, procurement of substitute goods or services; -// loss of use, data, or profits; or business interruption) however caused -// and on any theory of liability, whether in contract, strict liability, -// or tort (including negligence or otherwise) arising in any way out of -// the use of this software, even if advised of the possibility of such damage. -// -// - -/////////////////////////////////////////////////////////////////////////////////////////////////////// -////////////////////////////////////////////bitwise_binary///////////////////////////////////////////// -/////////////////////////////////////////////////////////////////////////////////////////////////////// - -__kernel void arithm_bitwise_binary_scalar( - __global uchar *src1, int src1_step, int src1_offset, - __global uchar *src2, - __global uchar *dst, int dst_step, int dst_offset, - int cols, int rows) -{ - int x = get_global_id(0); - int y = get_global_id(1); - - if (x < cols && y < rows) - { -#if elemSize > 1 - x *= elemSize; -#endif - int src1_index = mad24(y, src1_step, src1_offset + x); - int dst_index = mad24(y, dst_step, dst_offset + x); - -#if elemSize > 1 - #pragma unroll - for (int i = 0; i < elemSize; i += vlen) - { - ucharv t0 = vloadn(0, src1 + src1_index + i); - ucharv t1 = vloadn(0, src2 + i); - ucharv t2 = t0 Operation t1; - - vstoren(t2, 0, dst + dst_index + i); - } -#else - dst[dst_index] = src1[src1_index] Operation src2[0]; -#endif - } -} diff --git a/modules/ocl/src/opencl/arithm_bitwise_binary_scalar_mask.cl b/modules/ocl/src/opencl/arithm_bitwise_binary_scalar_mask.cl deleted file mode 100644 index 756f20165b..0000000000 --- a/modules/ocl/src/opencl/arithm_bitwise_binary_scalar_mask.cl +++ /dev/null @@ -1,86 +0,0 @@ -/*M/////////////////////////////////////////////////////////////////////////////////////// -// -// IMPORTANT: READ BEFORE DOWNLOADING, COPYING, INSTALLING OR USING. -// -// By downloading, copying, installing or using the software you agree to this license. -// If you do not agree to this license, do not download, install, -// copy or use the software. -// -// -// License Agreement -// For Open Source Computer Vision Library -// -// Copyright (C) 2010-2012, Institute Of Software Chinese Academy Of Science, all rights reserved. -// Copyright (C) 2010-2012, Advanced Micro Devices, Inc., all rights reserved. -// Third party copyrights are property of their respective owners. -// -// @Authors -// Jiang Liyuan, jlyuan001.good@163.com -// -// Redistribution and use in source and binary forms, with or without modification, -// are permitted provided that the following conditions are met: -// -// * Redistribution's of source code must retain the above copyright notice, -// this list of conditions and the following disclaimer. -// -// * Redistribution's in binary form must reproduce the above copyright notice, -// this list of conditions and the following disclaimer in the documentation -// and/or other materials provided with the distribution. -// -// * The name of the copyright holders may not be used to endorse or promote products -// derived from this software without specific prior written permission. -// -// This software is provided by the copyright holders and contributors as is and -// any express or implied warranties, including, but not limited to, the implied -// warranties of merchantability and fitness for a particular purpose are disclaimed. -// In no event shall the Intel Corporation or contributors be liable for any direct, -// indirect, incidental, special, exemplary, or consequential damages -// (including, but not limited to, procurement of substitute goods or services; -// loss of use, data, or profits; or business interruption) however caused -// and on any theory of liability, whether in contract, strict liability, -// or tort (including negligence or otherwise) arising in any way out of -// the use of this software, even if advised of the possibility of such damage. -// -//M*/ - -////////////////////////////////////////////////////////////////////////////////////////////////////// -////////////////////////////////////////////bitwise_binary//////////////////////////////////////////// -////////////////////////////////////////////////////////////////////////////////////////////////////// - -__kernel void arithm_bitwise_binary_scalar_mask(__global uchar *src1, int src1_step, int src1_offset, - __global uchar *src2, - __global uchar *mask, int mask_step, int mask_offset, - __global uchar *dst, int dst_step, int dst_offset, - int cols, int rows) -{ - int x = get_global_id(0); - int y = get_global_id(1); - - if (x < cols && y < rows) - { - int mask_index = mad24(y, mask_step, x + mask_offset); - - if (mask[mask_index]) - { -#if elemSize > 1 - x *= elemSize; -#endif - int src1_index = mad24(y, src1_step, x + src1_offset); - int dst_index = mad24(y, dst_step, x + dst_offset); - -#if elemSize > 1 - #pragma unroll - for (int i = 0; i < elemSize; i += vlen) - { - ucharv t0 = vloadn(0, src1 + src1_index + i); - ucharv t1 = vloadn(0, src2 + i); - ucharv t2 = t0 Operation t1; - - vstoren(t2, 0, dst + dst_index + i); - } -#else - dst[dst_index] = src1[src1_index] Operation src2[0]; -#endif - } - } -} diff --git a/modules/ocl/src/opencl/arithm_bitwise_not.cl b/modules/ocl/src/opencl/arithm_bitwise_not.cl deleted file mode 100644 index b6f76d6065..0000000000 --- a/modules/ocl/src/opencl/arithm_bitwise_not.cl +++ /dev/null @@ -1,253 +0,0 @@ -/*M/////////////////////////////////////////////////////////////////////////////////////// -// -// IMPORTANT: READ BEFORE DOWNLOADING, COPYING, INSTALLING OR USING. -// -// By downloading, copying, installing or using the software you agree to this license. -// If you do not agree to this license, do not download, install, -// copy or use the software. -// -// -// License Agreement -// For Open Source Computer Vision Library -// -// Copyright (C) 2010-2012, Institute Of Software Chinese Academy Of Science, all rights reserved. -// Copyright (C) 2010-2012, Advanced Micro Devices, Inc., all rights reserved. -// Third party copyrights are property of their respective owners. -// -// @Authors -// Jiang Liyuan, jlyuan001.good@163.com -// -// Redistribution and use in source and binary forms, with or without modification, -// are permitted provided that the following conditions are met: -// -// * Redistribution's of source code must retain the above copyright notice, -// this list of conditions and the following disclaimer. -// -// * Redistribution's in binary form must reproduce the above copyright notice, -// this list of conditions and the following disclaimer in the documentation -// and/or other materials provided with the distribution. -// -// * The name of the copyright holders may not be used to endorse or promote products -// derived from this software without specific prior written permission. -// -// This software is provided by the copyright holders and contributors as is and -// any express or implied warranties, including, but not limited to, the implied -// warranties of merchantability and fitness for a particular purpose are disclaimed. -// In no event shall the Intel Corporation or contributors be liable for any direct, -// indirect, incidental, special, exemplary, or consequential damages -// (including, but not limited to, procurement of substitute goods or services; -// loss of use, data, or profits; or business interruption) however caused -// and on any theory of liability, whether in contract, strict liability, -// or tort (including negligence or otherwise) arising in any way out of -// the use of this software, even if advised of the possibility of such damage. -// -//M*/ - -#ifdef DOUBLE_SUPPORT -#ifdef cl_amd_fp64 -#pragma OPENCL EXTENSION cl_amd_fp64:enable -#elif defined (cl_khr_fp64) -#pragma OPENCL EXTENSION cl_khr_fp64:enable -#endif -#endif - -/////////////////////////////////////////////////////////////////////////////////////////////////////// -////////////////////////////////////////////BITWISE_NOT//////////////////////////////////////////////// -/////////////////////////////////////////////////////////////////////////////////////////////////////// - -__kernel void arithm_bitwise_not_D0 (__global uchar *src1, int src1_step, int src1_offset, - __global uchar *dst, int dst_step, int dst_offset, - int rows, int cols, int dst_step1) -{ - int x = get_global_id(0); - int y = get_global_id(1); - - if (x < cols && y < rows) - { - x = x << 2; - int src1_index = mad24(y, src1_step, x + src1_offset); - - int dst_end = mad24(y, dst_step, dst_offset + dst_step1); - int dst_index = mad24(y, dst_step, dst_offset + x); - - uchar4 src1_data = vload4(0, src1 + src1_index); - uchar4 dst_data = vload4(0, dst + dst_index); - uchar4 tmp_data = ~src1_data; - - dst_data.x = dst_index + 0 < dst_end ? tmp_data.x : dst_data.x; - dst_data.y = dst_index + 1 < dst_end ? tmp_data.y : dst_data.y; - dst_data.z = dst_index + 2 < dst_end ? tmp_data.z : dst_data.z; - dst_data.w = dst_index + 3 < dst_end ? tmp_data.w : dst_data.w; - - vstore4(dst_data, 0, dst + dst_index); - } -} - - -__kernel void arithm_bitwise_not_D1 (__global char *src1, int src1_step, int src1_offset, - __global char *dst, int dst_step, int dst_offset, - int rows, int cols, int dst_step1) -{ - int x = get_global_id(0); - int y = get_global_id(1); - - if (x < cols && y < rows) - { - x = x << 2; - int src1_index = mad24(y, src1_step, x + src1_offset); - - int dst_end = mad24(y, dst_step, dst_offset + dst_step1); - int dst_index = mad24(y, dst_step, dst_offset + x); - - char4 src1_data = vload4(0, src1 + src1_index); - char4 dst_data = vload4(0, dst + dst_index); - char4 tmp_data = ~src1_data; - - dst_data.x = dst_index + 0 < dst_end ? tmp_data.x : dst_data.x; - dst_data.y = dst_index + 1 < dst_end ? tmp_data.y : dst_data.y; - dst_data.z = dst_index + 2 < dst_end ? tmp_data.z : dst_data.z; - dst_data.w = dst_index + 3 < dst_end ? tmp_data.w : dst_data.w; - - vstore4(dst_data, 0, dst + dst_index); - } -} - - -__kernel void arithm_bitwise_not_D2 (__global ushort *src1, int src1_step, int src1_offset, - __global ushort *dst, int dst_step, int dst_offset, - int rows, int cols, int dst_step1) - -{ - int x = get_global_id(0); - int y = get_global_id(1); - - if (x < cols && y < rows) - { - x = x << 2; - -#ifdef dst_align -#undef dst_align -#endif -#define dst_align ((dst_offset >> 1) & 3) - int src1_index = mad24(y, src1_step, (x << 1) + src1_offset - (dst_align << 1)); - - int dst_start = mad24(y, dst_step, dst_offset); - int dst_end = mad24(y, dst_step, dst_offset + dst_step1); - int dst_index = mad24(y, dst_step, dst_offset + (x << 1) & (int)0xfffffff8); - - ushort4 src1_data = vload4(0, (__global ushort *)((__global char *)src1 + src1_index)); - - ushort4 dst_data = *((__global ushort4 *)((__global char *)dst + dst_index)); - ushort4 tmp_data = ~ src1_data; - - dst_data.x = ((dst_index + 0 >= dst_start) && (dst_index + 0 < dst_end)) ? tmp_data.x : dst_data.x; - dst_data.y = ((dst_index + 2 >= dst_start) && (dst_index + 2 < dst_end)) ? tmp_data.y : dst_data.y; - dst_data.z = ((dst_index + 4 >= dst_start) && (dst_index + 4 < dst_end)) ? tmp_data.z : dst_data.z; - dst_data.w = ((dst_index + 6 >= dst_start) && (dst_index + 6 < dst_end)) ? tmp_data.w : dst_data.w; - - *((__global ushort4 *)((__global char *)dst + dst_index)) = dst_data; - } -} - - - -__kernel void arithm_bitwise_not_D3 (__global short *src1, int src1_step, int src1_offset, - __global short *dst, int dst_step, int dst_offset, - int rows, int cols, int dst_step1) - -{ - int x = get_global_id(0); - int y = get_global_id(1); - - if (x < cols && y < rows) - { - x = x << 2; - -#ifdef dst_align -#undef dst_align -#endif -#define dst_align ((dst_offset >> 1) & 3) - int src1_index = mad24(y, src1_step, (x << 1) + src1_offset - (dst_align << 1)); - - int dst_start = mad24(y, dst_step, dst_offset); - int dst_end = mad24(y, dst_step, dst_offset + dst_step1); - int dst_index = mad24(y, dst_step, dst_offset + (x << 1) & (int)0xfffffff8); - - short4 src1_data = vload4(0, (__global short *)((__global char *)src1 + src1_index)); - - short4 dst_data = *((__global short4 *)((__global char *)dst + dst_index)); - short4 tmp_data = ~ src1_data; - - dst_data.x = ((dst_index + 0 >= dst_start) && (dst_index + 0 < dst_end)) ? tmp_data.x : dst_data.x; - dst_data.y = ((dst_index + 2 >= dst_start) && (dst_index + 2 < dst_end)) ? tmp_data.y : dst_data.y; - dst_data.z = ((dst_index + 4 >= dst_start) && (dst_index + 4 < dst_end)) ? tmp_data.z : dst_data.z; - dst_data.w = ((dst_index + 6 >= dst_start) && (dst_index + 6 < dst_end)) ? tmp_data.w : dst_data.w; - - *((__global short4 *)((__global char *)dst + dst_index)) = dst_data; - } -} - - - -__kernel void arithm_bitwise_not_D4 (__global int *src1, int src1_step, int src1_offset, - __global int *dst, int dst_step, int dst_offset, - int rows, int cols, int dst_step1) -{ - int x = get_global_id(0); - int y = get_global_id(1); - - if (x < cols && y < rows) - { - int src1_index = mad24(y, src1_step, (x << 2) + src1_offset); - int dst_index = mad24(y, dst_step, (x << 2) + dst_offset); - - int data1 = *((__global int *)((__global char *)src1 + src1_index)); - int tmp = ~ data1; - - *((__global int *)((__global char *)dst + dst_index)) = tmp; - } -} - -__kernel void arithm_bitwise_not_D5 (__global char *src, int src_step, int src_offset, - __global char *dst, int dst_step, int dst_offset, - int rows, int cols, int dst_step1) -{ - int x = get_global_id(0); - int y = get_global_id(1); - - if (x < cols && y < rows) - { - int src_index = mad24(y, src_step, (x << 2) + src_offset); - int dst_index = mad24(y, dst_step, (x << 2) + dst_offset); - - char4 data; - - data = *((__global char4 *)((__global char *)src + src_index)); - data = ~ data; - - *((__global char4 *)((__global char *)dst + dst_index)) = data; - } -} - -#if defined (DOUBLE_SUPPORT) -__kernel void arithm_bitwise_not_D6 (__global char *src, int src_step, int src_offset, - __global char *dst, int dst_step, int dst_offset, - int rows, int cols, int dst_step1) -{ - int x = get_global_id(0); - int y = get_global_id(1); - - if (x < cols && y < rows) - { - int src_index = mad24(y, src_step, (x << 3) + src_offset); - int dst_index = mad24(y, dst_step, (x << 3) + dst_offset); - - char8 data; - - data = *((__global char8 *)((__global char *)src + src_index)); - data = ~ data; - - *((__global char8 *)((__global char *)dst + dst_index)) = data; - } -} -#endif diff --git a/modules/ocl/src/opencl/arithm_minMax.cl b/modules/ocl/src/opencl/arithm_minMax.cl index 01db7d0646..b0cd1c8b68 100644 --- a/modules/ocl/src/opencl/arithm_minMax.cl +++ b/modules/ocl/src/opencl/arithm_minMax.cl @@ -63,81 +63,31 @@ /**************************************Array minMax**************************************/ -__kernel void arithm_op_minMax(__global const T * src, __global T * dst, - int cols, int invalid_cols, int offset, int elemnum, int groupnum) +__kernel void arithm_op_minMax(__global const T * src, int src_step, int src_offset, int src_rows, int src_cols, + int total, int groupnum, __global T * dst +#ifdef WITH_MASK + , __global const uchar * mask, int mask_step, int mask_offset +#endif +) { int lid = get_local_id(0); int gid = get_group_id(0); int id = get_global_id(0); - int idx = offset + id + (id / cols) * invalid_cols; - __local T localmem_max[128], localmem_min[128]; T minval = (T)(MAX_VAL), maxval = (T)(MIN_VAL), temp; + int y, x; - for (int grainSize = groupnum << 8; id < elemnum; id += grainSize) + for (int grainSize = groupnum << 8; id < total; id += grainSize) { - idx = offset + id + (id / cols) * invalid_cols; - temp = src[idx]; - minval = min(minval, temp); - maxval = max(maxval, temp); - } + y = id / src_cols; + x = id % src_cols; - if (lid > 127) - { - localmem_min[lid - 128] = minval; - localmem_max[lid - 128] = maxval; - } - barrier(CLK_LOCAL_MEM_FENCE); - - if (lid < 128) - { - localmem_min[lid] = min(minval, localmem_min[lid]); - localmem_max[lid] = max(maxval, localmem_max[lid]); - } - barrier(CLK_LOCAL_MEM_FENCE); - - for (int lsize = 64; lsize > 0; lsize >>= 1) - { - if (lid < lsize) +#ifdef WITH_MASK + if (mask[mad24(y, mask_step, x + mask_offset)]) +#endif { - int lid2 = lsize + lid; - localmem_min[lid] = min(localmem_min[lid], localmem_min[lid2]); - localmem_max[lid] = max(localmem_max[lid], localmem_max[lid2]); - } - barrier(CLK_LOCAL_MEM_FENCE); - } - - if (lid == 0) - { - dst[gid] = localmem_min[0]; - dst[gid + groupnum] = localmem_max[0]; - } -} - -__kernel void arithm_op_minMax_mask(__global const T * src, __global T * dst, - int cols, int invalid_cols, int offset, - int elemnum, int groupnum, - const __global uchar * mask, int minvalid_cols, int moffset) -{ - int lid = get_local_id(0); - int gid = get_group_id(0); - int id = get_global_id(0); - - int idx = offset + id + (id / cols) * invalid_cols; - int midx = moffset + id + (id / cols) * minvalid_cols; - - __local T localmem_max[128], localmem_min[128]; - T minval = (T)(MAX_VAL), maxval = (T)(MIN_VAL), temp; - - for (int grainSize = groupnum << 8; id < elemnum; id += grainSize) - { - idx = offset + id + (id / cols) * invalid_cols; - midx = moffset + id + (id / cols) * minvalid_cols; - - if (mask[midx]) - { - temp = src[idx]; + temp = src[mad24(y, src_step, x + src_offset)]; minval = min(minval, temp); maxval = max(maxval, temp); } diff --git a/modules/ocl/src/opencl/arithm_nonzero.cl b/modules/ocl/src/opencl/arithm_nonzero.cl index 3180c26e86..6a21f27e0e 100644 --- a/modules/ocl/src/opencl/arithm_nonzero.cl +++ b/modules/ocl/src/opencl/arithm_nonzero.cl @@ -52,23 +52,18 @@ /**************************************Count NonZero**************************************/ -__kernel void arithm_op_nonzero(int cols, int invalid_cols, int offset, int elemnum, int groupnum, - __global srcT *src, __global dstT *dst) +__kernel void arithm_op_nonzero(__global srcT * src, int src_step, int src_offset, int src_cols, + int total, int groupnum, __global dstT * dst) { int lid = get_local_id(0); int gid = get_group_id(0); int id = get_global_id(0); - int idx = offset + id + (id / cols) * invalid_cols; __local dstT localmem_nonzero[128]; dstT nonzero = (dstT)(0); - srcT zero = (srcT)(0), one = (srcT)(1); - for (int grain = groupnum << 8; id < elemnum; id += grain) - { - idx = offset + id + (id / cols) * invalid_cols; - nonzero += src[idx] == zero ? zero : one; - } + for (int grain = groupnum << 8; id < total; id += grain) + nonzero += convertToDstT(src[mad24(id / src_cols, src_step, id % src_cols + src_offset)] == (srcT)(0)) ? (dstT)(0) : (dstT)(1); if (lid > 127) localmem_nonzero[lid - 128] = nonzero; diff --git a/modules/ocl/src/opencl/arithm_sum.cl b/modules/ocl/src/opencl/arithm_sum.cl index 514cf2a7f1..86015cfe93 100644 --- a/modules/ocl/src/opencl/arithm_sum.cl +++ b/modules/ocl/src/opencl/arithm_sum.cl @@ -63,21 +63,19 @@ /**************************************Array buffer SUM**************************************/ -__kernel void arithm_op_sum(int cols,int invalid_cols,int offset,int elemnum,int groupnum, - __global srcT *src, __global dstT *dst) +__kernel void arithm_op_sum(__global srcT * src, int src_step, int src_offset, int src_cols, + int total, int groupnum, __global dstT * dst) { int lid = get_local_id(0); int gid = get_group_id(0); int id = get_global_id(0); - int idx = offset + id + (id / cols) * invalid_cols; __local dstT localmem_sum[128]; dstT sum = (dstT)(0), temp; - for (int grainSize = groupnum << 8; id < elemnum; id += grainSize) + for (int grainSize = groupnum << 8; id < total; id += grainSize) { - idx = offset + id + (id / cols) * invalid_cols; - temp = convertToDstT(src[idx]); + temp = convertToDstT(src[mad24(id / src_cols, src_step, id % src_cols + src_offset)]); FUNC(temp, sum); } diff --git a/modules/ocl/test/test_arithm.cpp b/modules/ocl/test/test_arithm.cpp index 17260580de..29976b6e4e 100644 --- a/modules/ocl/test/test_arithm.cpp +++ b/modules/ocl/test/test_arithm.cpp @@ -198,7 +198,7 @@ PARAM_TEST_CASE(ArithmTestBase, MatDepth, Channels, bool) Size roiSize = randomSize(1, MAX_VALUE); Border src1Border = randomBorder(0, use_roi ? MAX_VALUE : 0); - randomSubMat(src1, src1_roi, roiSize, src1Border, type, 2, 11); + randomSubMat(src1, src1_roi, roiSize, src1Border, type, -11, 11); Border src2Border = randomBorder(0, use_roi ? MAX_VALUE : 0); randomSubMat(src2, src2_roi, roiSize, src2Border, type, -1540, 1740); @@ -1163,7 +1163,7 @@ OCL_TEST_P(CountNonZero, MAT) int cpures = cv::countNonZero(src1_roi); int gpures = cv::ocl::countNonZero(gsrc1_roi); - EXPECT_DOUBLE_EQ((double)cpures, (double)gpures); + EXPECT_EQ(cpures, gpures); } }