From 0f95f0d8b3f4a11d94d1d34326953d89a047fbd5 Mon Sep 17 00:00:00 2001 From: Alexander Alekhin Date: Sat, 26 Oct 2013 05:06:22 +0400 Subject: [PATCH] ocl: rewrite filter2D --- modules/ocl/include/opencv2/ocl/ocl.hpp | 12 +- modules/ocl/src/filtering.cpp | 275 ++++++++----- modules/ocl/src/opencl/filtering_filter2D.cl | 370 +++++++++++++++++ modules/ocl/src/opencl/filtering_laplacian.cl | 381 ------------------ modules/ocl/test/test_filters.cpp | 6 +- 5 files changed, 544 insertions(+), 500 deletions(-) create mode 100644 modules/ocl/src/opencl/filtering_filter2D.cl delete mode 100644 modules/ocl/src/opencl/filtering_laplacian.cl diff --git a/modules/ocl/include/opencv2/ocl/ocl.hpp b/modules/ocl/include/opencv2/ocl/ocl.hpp index 05bd061ca..db386952a 100644 --- a/modules/ocl/include/opencv2/ocl/ocl.hpp +++ b/modules/ocl/include/opencv2/ocl/ocl.hpp @@ -718,8 +718,9 @@ namespace cv CV_EXPORTS Ptr createDerivFilter_GPU( int srcType, int dstType, int dx, int dy, int ksize, int borderType = BORDER_DEFAULT ); //! applies Laplacian operator to the image - // supports only ksize = 1 and ksize = 3 8UC1 8UC4 32FC1 32FC4 data type - CV_EXPORTS void Laplacian(const oclMat &src, oclMat &dst, int ddepth, int ksize = 1, double scale = 1); + // supports only ksize = 1 and ksize = 3 + CV_EXPORTS void Laplacian(const oclMat &src, oclMat &dst, int ddepth, int ksize = 1, double scale = 1, + double delta=0, int borderType=BORDER_DEFAULT); //! returns 2D box filter // dst type must be the same as source type @@ -731,11 +732,12 @@ namespace cv const Point &anchor = Point(-1, -1), int borderType = BORDER_DEFAULT); //! returns 2D filter with the specified kernel - // supports CV_8UC1 and CV_8UC4 types + // supports: dst type must be the same as source type CV_EXPORTS Ptr getLinearFilter_GPU(int srcType, int dstType, const Mat &kernel, const Size &ksize, const Point &anchor = Point(-1, -1), int borderType = BORDER_DEFAULT); //! returns the non-separable linear filter engine + // supports: dst type must be the same as source type CV_EXPORTS Ptr createLinearFilter_GPU(int srcType, int dstType, const Mat &kernel, const Point &anchor = Point(-1, -1), int borderType = BORDER_DEFAULT); @@ -762,10 +764,8 @@ namespace cv } //! applies non-separable 2D linear filter to the image - // Note, at the moment this function only works when anchor point is in the kernel center - // and kernel size supported is either 3x3 or 5x5; otherwise the function will fail to output valid result CV_EXPORTS void filter2D(const oclMat &src, oclMat &dst, int ddepth, const Mat &kernel, - Point anchor = Point(-1, -1), int borderType = BORDER_DEFAULT); + Point anchor = Point(-1, -1), double delta = 0.0, int borderType = BORDER_DEFAULT); //! applies separable 2D linear filter to the image CV_EXPORTS void sepFilter2D(const oclMat &src, oclMat &dst, int ddepth, const Mat &kernelX, const Mat &kernelY, diff --git a/modules/ocl/src/filtering.cpp b/modules/ocl/src/filtering.cpp index fdddc1674..e1255197f 100644 --- a/modules/ocl/src/filtering.cpp +++ b/modules/ocl/src/filtering.cpp @@ -69,37 +69,14 @@ inline void normalizeAnchor(Point &anchor, const Size &ksize) normalizeAnchor(anchor.y, ksize.height); } -inline void normalizeROI(Rect &roi, const Size &ksize, const Point &anchor, const Size &src_size) +inline void normalizeROI(Rect &roi, const Size &ksize, const Point &/*anchor*/, const Size &src_size) { if (roi == Rect(0, 0, -1, -1)) roi = Rect(0, 0, src_size.width, src_size.height); CV_Assert(ksize.height > 0 && ksize.width > 0 && ((ksize.height & 1) == 1) && ((ksize.width & 1) == 1)); - CV_Assert((anchor.x == -1 && anchor.y == -1) || (anchor.x == ksize.width >> 1 && anchor.y == ksize.height >> 1)); CV_Assert(roi.x >= 0 && roi.y >= 0 && roi.width <= src_size.width && roi.height <= src_size.height); } - - -inline void normalizeKernel(const Mat &kernel, oclMat &gpu_krnl, int type = CV_8U, int *nDivisor = 0, bool reverse = false) -{ - int scale = nDivisor && (kernel.depth() == CV_32F || kernel.depth() == CV_64F) ? 256 : 1; - - if (nDivisor) - *nDivisor = scale; - Mat temp(kernel.size(), type); - kernel.convertTo(temp, type, scale); - Mat cont_krnl = temp.reshape(1, 1); - - if (reverse) - { - int count = cont_krnl.cols >> 1; - - for (int i = 0; i < count; ++i) - std::swap(cont_krnl.at(0, i), cont_krnl.at(0, cont_krnl.cols - 1 - i)); - } - - gpu_krnl.upload(cont_krnl); -} } //////////////////////////////////////////////////////////////////////////////////////////////////// @@ -168,7 +145,7 @@ typedef void (*GPUMorfFilter_t)(const oclMat & , oclMat & , oclMat & , Size &, c class MorphFilter_GPU : public BaseFilter_GPU { public: - MorphFilter_GPU(const Size &ksize_, const Point &anchor_, const oclMat &kernel_, GPUMorfFilter_t func_) : + MorphFilter_GPU(const Size &ksize_, const Point &anchor_, const Mat &kernel_, GPUMorfFilter_t func_) : BaseFilter_GPU(ksize_, anchor_, BORDER_CONSTANT), kernel(kernel_), func(func_), rectKernel(false) {} virtual void operator()(const oclMat &src, oclMat &dst) @@ -355,16 +332,17 @@ Ptr cv::ocl::getMorphologyFilter_GPU(int op, int type, const Mat CV_Assert(op == MORPH_ERODE || op == MORPH_DILATE); CV_Assert(type == CV_8UC1 || type == CV_8UC3 || type == CV_8UC4 || type == CV_32FC1 || type == CV_32FC3 || type == CV_32FC4); - oclMat gpu_krnl; - normalizeKernel(kernel, gpu_krnl); normalizeAnchor(anchor, ksize); + Mat kernel8U; + kernel.convertTo(kernel8U, CV_8U); + Mat cont_krnl = kernel8U.reshape(1, 1); bool noZero = true; for(int i = 0; i < kernel.rows * kernel.cols; ++i) if(kernel.data[i] != 1) noZero = false; - MorphFilter_GPU* mfgpu = new MorphFilter_GPU(ksize, anchor, gpu_krnl, GPUMorfFilter_callers[op][CV_MAT_CN(type)]); + MorphFilter_GPU* mfgpu = new MorphFilter_GPU(ksize, anchor, cont_krnl, GPUMorfFilter_callers[op][CV_MAT_CN(type)]); if(noZero) mfgpu->rectKernel = true; @@ -524,12 +502,12 @@ void cv::ocl::morphologyEx(const oclMat &src, oclMat &dst, int op, const Mat &ke namespace { -typedef void (*GPUFilter2D_t)(const oclMat & , oclMat & , const oclMat & , const Size &, const Point&, const int); +typedef void (*GPUFilter2D_t)(const oclMat & , oclMat & , const Mat & , const Size &, const Point&, const int); class LinearFilter_GPU : public BaseFilter_GPU { public: - LinearFilter_GPU(const Size &ksize_, const Point &anchor_, const oclMat &kernel_, GPUFilter2D_t func_, + LinearFilter_GPU(const Size &ksize_, const Point &anchor_, const Mat &kernel_, GPUFilter2D_t func_, int borderType_) : BaseFilter_GPU(ksize_, anchor_, borderType_), kernel(kernel_), func(func_) {} @@ -543,118 +521,192 @@ public: }; } -static void GPUFilter2D(const oclMat &src, oclMat &dst, const oclMat &mat_kernel, +// prepare kernel: transpose and make double rows (+align). Returns size of aligned row +// Samples: +// a b c +// Input: d e f +// g h i +// Output, last two zeros is the alignment: +// a d g a d g 0 0 +// b e h b e h 0 0 +// c f i c f i 0 0 +template +static int _prepareKernelFilter2D(std::vector& data, const Mat &kernel) +{ + Mat _kernel; kernel.convertTo(_kernel, DataDepth::value); + int size_y_aligned = roundUp(kernel.rows * 2, 4); + data.clear(); data.resize(size_y_aligned * kernel.cols, 0); + for (int x = 0; x < kernel.cols; x++) + { + for (int y = 0; y < kernel.rows; y++) + { + data[x * size_y_aligned + y] = _kernel.at(y, x); + data[x * size_y_aligned + y + kernel.rows] = _kernel.at(y, x); + } + } + return size_y_aligned; +} + +static void GPUFilter2D(const oclMat &src, oclMat &dst, const Mat &kernel, const Size &ksize, const Point& anchor, const int borderType) { CV_Assert(src.clCxt == dst.clCxt); CV_Assert((src.cols == dst.cols) && (src.rows == dst.rows)); - CV_Assert((src.oclchannels() == dst.oclchannels())); - CV_Assert(ksize.height > 0 && ksize.width > 0 && ((ksize.height & 1) == 1) && ((ksize.width & 1) == 1)); - CV_Assert((anchor.x == -1 && anchor.y == -1) || (anchor.x == ksize.width >> 1 && anchor.y == ksize.height >> 1)); - CV_Assert(ksize.width == ksize.height); - Context *clCxt = src.clCxt; + CV_Assert(src.oclchannels() == dst.oclchannels()); - int filterWidth = ksize.width; - bool ksize_3x3 = filterWidth == 3 && src.type() != CV_32FC4 && src.type() != CV_32FC3; // CV_32FC4 is not tuned up with filter2d_3x3 kernel + CV_Assert(kernel.cols == ksize.width && kernel.rows == ksize.height); + CV_Assert(kernel.channels() == 1); - string kernelName = ksize_3x3 ? "filter2D_3x3" : "filter2D"; + CV_Assert(anchor.x >= 0 && anchor.x < kernel.cols); + CV_Assert(anchor.y >= 0 && anchor.y < kernel.rows); - size_t src_offset_x = (src.offset % src.step) / src.elemSize(); - size_t src_offset_y = src.offset / src.step; + bool useDouble = src.depth() == CV_64F; - size_t dst_offset_x = (dst.offset % dst.step) / dst.elemSize(); - size_t dst_offset_y = dst.offset / dst.step; - - int paddingPixels = filterWidth & (-2); - - size_t localThreads[3] = {ksize_3x3 ? 256 : 16, ksize_3x3 ? 1 : 16, 1}; - size_t globalThreads[3] = {src.wholecols, src.wholerows, 1}; - - int cn = src.oclchannels(); - int src_step = (int)(src.step/src.elemSize()); - int dst_step = (int)(dst.step/src.elemSize()); - - int localWidth = localThreads[0] + paddingPixels; - int localHeight = localThreads[1] + paddingPixels; - - size_t localMemSize = ksize_3x3 ? 260 * 6 * src.elemSize() : (localWidth * localHeight) * src.elemSize(); - - int vector_lengths[4][7] = {{4, 4, 4, 4, 4, 4, 4}, - {4, 4, 1, 1, 1, 1, 1}, - {1, 1, 1, 1, 1, 1, 1}, - {4, 4, 4, 4, 1, 1, 4} - }; - int cols = dst.cols + ((dst_offset_x) & (vector_lengths[cn - 1][src.depth()] - 1)); - - vector< pair > 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 *)&src_step)); - args.push_back(make_pair(sizeof(cl_int), (void *)&dst_step)); - args.push_back(make_pair(sizeof(cl_mem), (void *)&mat_kernel.data)); - args.push_back(make_pair(localMemSize, (void *)NULL)); - args.push_back(make_pair(sizeof(cl_int), (void *)&src.wholerows)); - args.push_back(make_pair(sizeof(cl_int), (void *)&src.wholecols)); - args.push_back(make_pair(sizeof(cl_int), (void *)&src_offset_x)); - args.push_back(make_pair(sizeof(cl_int), (void *)&src_offset_y)); - args.push_back(make_pair(sizeof(cl_int), (void *)&dst_offset_x)); - args.push_back(make_pair(sizeof(cl_int), (void *)&dst_offset_y)); - args.push_back(make_pair(sizeof(cl_int), (void *)&src.cols)); - args.push_back(make_pair(sizeof(cl_int), (void *)&src.rows)); - args.push_back(make_pair(sizeof(cl_int), (void *)&cols)); - char btype[30]; - switch (borderType) + std::vector kernelDataFloat; + std::vector kernelDataDouble; + int kernel_size_y2_aligned = useDouble ? + _prepareKernelFilter2D(kernelDataDouble, kernel) + : _prepareKernelFilter2D(kernelDataFloat, kernel); + oclMat oclKernelParameter; + if (useDouble) { - case 0: - sprintf(btype, "BORDER_CONSTANT"); + oclKernelParameter.createEx(1, kernelDataDouble.size(), CV_64FC1, DEVICE_MEM_R_ONLY, DEVICE_MEM_DEFAULT); + openCLMemcpy2D(src.clCxt, oclKernelParameter.data, kernelDataDouble.size()*sizeof(double), + &kernelDataDouble[0], kernelDataDouble.size()*sizeof(double), + kernelDataDouble.size()*sizeof(double), 1, clMemcpyHostToDevice); + } + else + { + oclKernelParameter.createEx(1, kernelDataFloat.size(), CV_32FC1, DEVICE_MEM_R_ONLY, DEVICE_MEM_DEFAULT); + openCLMemcpy2D(src.clCxt, oclKernelParameter.data, kernelDataFloat.size()*sizeof(float), + &kernelDataFloat[0], kernelDataFloat.size()*sizeof(float), + kernelDataFloat.size()*sizeof(float), 1, clMemcpyHostToDevice); + } + + size_t BLOCK_SIZE = src.clCxt->getDeviceInfo().maxWorkItemSizes[0]; +#if 1 // TODO Mode with several blocks requires a much more VGPRs, so this optimization is not actual for the current devices + size_t BLOCK_SIZE_Y = 1; +#else + size_t BLOCK_SIZE_Y = 8; // TODO Check heuristic value on devices + while (BLOCK_SIZE_Y < BLOCK_SIZE / 8 && BLOCK_SIZE_Y * src.clCxt->getDeviceInfo().maxComputeUnits * 32 < (size_t)src.rows) + BLOCK_SIZE_Y *= 2; +#endif + + CV_Assert((size_t)ksize.width <= BLOCK_SIZE); + + bool isIsolatedBorder = (borderType & BORDER_ISOLATED) != 0; + + vector > args; + + args.push_back( make_pair( sizeof(cl_mem), (void *)&src.data)); + cl_uint stepBytes = src.step; + args.push_back( make_pair( sizeof(cl_uint), (void *)&stepBytes)); + int offsetXBytes = src.offset % src.step; + int offsetX = offsetXBytes / src.elemSize(); + CV_Assert((int)(offsetX * src.elemSize()) == offsetXBytes); + int offsetY = src.offset / src.step; + int endX = (offsetX + src.cols); + int endY = (offsetY + src.rows); + cl_int rect[4] = {offsetX, offsetY, endX, endY}; + if (!isIsolatedBorder) + { + rect[2] = src.wholecols; + rect[3] = src.wholerows; + } + args.push_back( make_pair( sizeof(cl_int)*4, (void *)&rect[0])); + + args.push_back( make_pair( sizeof(cl_mem), (void *)&dst.data)); + cl_uint _stepBytes = dst.step; + args.push_back( make_pair( sizeof(cl_uint), (void *)&_stepBytes)); + int _offsetXBytes = dst.offset % dst.step; + int _offsetX = _offsetXBytes / dst.elemSize(); + CV_Assert((int)(_offsetX * dst.elemSize()) == _offsetXBytes); + int _offsetY = dst.offset / dst.step; + int _endX = (_offsetX + dst.cols); + int _endY = (_offsetY + dst.rows); + cl_int _rect[4] = {_offsetX, _offsetY, _endX, _endY}; + args.push_back( make_pair( sizeof(cl_int)*4, (void *)&_rect[0])); + + float borderValue[4] = {0, 0, 0, 0}; // DON'T move into 'if' body + double borderValueDouble[4] = {0, 0, 0, 0}; // DON'T move into 'if' body + if ((borderType & ~BORDER_ISOLATED) == BORDER_CONSTANT) + { + if (useDouble) + args.push_back( make_pair( sizeof(double) * src.oclchannels(), (void *)&borderValue[0])); + else + args.push_back( make_pair( sizeof(float) * src.oclchannels(), (void *)&borderValueDouble[0])); + } + + args.push_back( make_pair( sizeof(cl_mem), (void *)&oclKernelParameter.data)); + + const char* btype = NULL; + + switch (borderType & ~BORDER_ISOLATED) + { + case BORDER_CONSTANT: + btype = "BORDER_CONSTANT"; break; - case 1: - sprintf(btype, "BORDER_REPLICATE"); + case BORDER_REPLICATE: + btype = "BORDER_REPLICATE"; break; - case 2: - sprintf(btype, "BORDER_REFLECT"); + case BORDER_REFLECT: + btype = "BORDER_REFLECT"; break; - case 3: + case BORDER_WRAP: CV_Error(CV_StsUnsupportedFormat, "BORDER_WRAP is not supported!"); return; - case 4: - sprintf(btype, "BORDER_REFLECT_101"); + case BORDER_REFLECT101: + btype = "BORDER_REFLECT_101"; break; } - int type = src.depth(); - char build_options[150]; - sprintf(build_options, "-D %s -D IMG_C_%d_%d -D CN=%d -D FILTER_SIZE=%d", btype, cn, type, cn, ksize.width); - openCLExecuteKernel(clCxt, &filtering_laplacian, kernelName, globalThreads, localThreads, args, -1, -1, build_options); + + int requiredTop = anchor.y; + int requiredLeft = BLOCK_SIZE; // not this: anchor.x; + int requiredBottom = ksize.height - 1 - anchor.y; + int requiredRight = BLOCK_SIZE; // not this: ksize.width - 1 - anchor.x; + int h = isIsolatedBorder ? src.rows : src.wholerows; + int w = isIsolatedBorder ? src.cols : src.wholecols; + bool extra_extrapolation = h < requiredTop || h < requiredBottom || w < requiredLeft || w < requiredRight; + + char build_options[1024]; + sprintf(build_options, "-D LOCAL_SIZE=%d -D BLOCK_SIZE_Y=%d -D DATA_DEPTH=%d -D DATA_CHAN=%d -D USE_DOUBLE=%d " + "-D ANCHOR_X=%d -D ANCHOR_Y=%d -D KERNEL_SIZE_X=%d -D KERNEL_SIZE_Y=%d -D KERNEL_SIZE_Y2_ALIGNED=%d " + "-D %s -D %s -D %s", + (int)BLOCK_SIZE, (int)BLOCK_SIZE_Y, + src.depth(), src.oclchannels(), useDouble ? 1 : 0, + anchor.x, anchor.y, ksize.width, ksize.height, kernel_size_y2_aligned, + btype, + extra_extrapolation ? "EXTRA_EXTRAPOLATION" : "NO_EXTRA_EXTRAPOLATION", + isIsolatedBorder ? "BORDER_ISOLATED" : "NO_BORDER_ISOLATED"); + + size_t gt[3] = {divUp(dst.cols, BLOCK_SIZE - (ksize.width - 1)) * BLOCK_SIZE, divUp(dst.rows, BLOCK_SIZE_Y), 1}, lt[3] = {BLOCK_SIZE, 1, 1}; + openCLExecuteKernel(src.clCxt, &filtering_filter2D, "filter2D", gt, lt, args, -1, -1, build_options); } -Ptr cv::ocl::getLinearFilter_GPU(int srcType, int dstType, const Mat &kernel, const Size &ksize, +Ptr cv::ocl::getLinearFilter_GPU(int /*srcType*/, int /*dstType*/, const Mat &kernel, const Size &ksize, const Point &anchor, int borderType) { - static const GPUFilter2D_t GPUFilter2D_callers[] = {0, GPUFilter2D, 0, GPUFilter2D, GPUFilter2D}; - - CV_Assert((srcType == CV_8UC1 || srcType == CV_8UC3 || srcType == CV_8UC4 || srcType == CV_32FC1 || srcType == CV_32FC3 || srcType == CV_32FC4) && dstType == srcType); - - oclMat gpu_krnl; Point norm_archor = anchor; - normalizeKernel(kernel, gpu_krnl, CV_32FC1); normalizeAnchor(norm_archor, ksize); - return Ptr(new LinearFilter_GPU(ksize, anchor, gpu_krnl, GPUFilter2D_callers[CV_MAT_CN(srcType)], + return Ptr(new LinearFilter_GPU(ksize, norm_archor, kernel, GPUFilter2D, borderType)); } Ptr cv::ocl::createLinearFilter_GPU(int srcType, int dstType, const Mat &kernel, const Point &anchor, int borderType) { - Size ksize = kernel.size(); + Size ksize = kernel.size(); // TODO remove duplicated parameter Ptr linearFilter = getLinearFilter_GPU(srcType, dstType, kernel, ksize, anchor, borderType); return createFilter2D_GPU(linearFilter); } -void cv::ocl::filter2D(const oclMat &src, oclMat &dst, int ddepth, const Mat &kernel, Point anchor, int borderType) +void cv::ocl::filter2D(const oclMat &src, oclMat &dst, int ddepth, const Mat &kernel, Point anchor, double delta, int borderType) { + CV_Assert(delta == 0); + if (ddepth < 0) ddepth = src.depth(); @@ -1222,8 +1274,11 @@ void cv::ocl::Scharr(const oclMat &src, oclMat &dst, int ddepth, int dx, int dy, sepFilter2D(src, dst, ddepth, kx, ky, Point(-1, -1), delta, bordertype); } -void cv::ocl::Laplacian(const oclMat &src, oclMat &dst, int ddepth, int ksize, double scale) +void cv::ocl::Laplacian(const oclMat &src, oclMat &dst, int ddepth, int ksize, double scale, + double delta, int borderType) { + CV_Assert(delta == 0); + if (!src.clCxt->supportsFeature(FEATURE_CL_DOUBLE) && src.type() == CV_64F) { CV_Error(CV_OpenCLDoubleNotSupported, "Selected device doesn't support double"); @@ -1232,17 +1287,17 @@ void cv::ocl::Laplacian(const oclMat &src, oclMat &dst, int ddepth, int ksize, d CV_Assert(ksize == 1 || ksize == 3); - int K[2][9] = + double K[2][9] = { {0, 1, 0, 1, -4, 1, 0, 1, 0}, {2, 0, 2, 0, -8, 0, 2, 0, 2} }; - Mat kernel(3, 3, CV_32S, (void *)K[ksize == 3]); + Mat kernel(3, 3, CV_64F, (void *)K[ksize == 3 ? 1 : 0]); if (scale != 1) kernel *= scale; - filter2D(src, dst, ddepth, kernel, Point(-1, -1)); + filter2D(src, dst, ddepth, kernel, Point(-1, -1), 0, borderType); } //////////////////////////////////////////////////////////////////////////////////////////////////// diff --git a/modules/ocl/src/opencl/filtering_filter2D.cl b/modules/ocl/src/opencl/filtering_filter2D.cl new file mode 100644 index 000000000..f96676689 --- /dev/null +++ b/modules/ocl/src/opencl/filtering_filter2D.cl @@ -0,0 +1,370 @@ +/*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) 2010-2013, Advanced Micro Devices, Inc., all rights reserved. +// Third party copyrights are property of their respective owners. +// +// 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 BORDER_REPLICATE +//BORDER_REPLICATE: aaaaaa|abcdefgh|hhhhhhh +#define ADDR_L(i, l_edge, r_edge) ((i) < (l_edge) ? (l_edge) : (i)) +#define ADDR_R(i, r_edge, addr) ((i) >= (r_edge) ? (r_edge)-1 : (addr)) +#define ADDR_H(i, t_edge, b_edge) ((i) < (t_edge) ? (t_edge) :(i)) +#define ADDR_B(i, b_edge, addr) ((i) >= (b_edge) ? (b_edge)-1 :(addr)) +#endif + +#ifdef BORDER_REFLECT +//BORDER_REFLECT: fedcba|abcdefgh|hgfedcb +#define ADDR_L(i, l_edge, r_edge) ((i) < (l_edge) ? -(i)-1 : (i)) +#define ADDR_R(i, r_edge, addr) ((i) >= (r_edge) ? -(i)-1+((r_edge)<<1) : (addr)) +#define ADDR_H(i, t_edge, b_edge) ((i) < (t_edge) ? -(i)-1 : (i)) +#define ADDR_B(i, b_edge, addr) ((i) >= (b_edge) ? -(i)-1+((b_edge)<<1) : (addr)) +#endif + +#ifdef BORDER_REFLECT_101 +//BORDER_REFLECT_101: gfedcb|abcdefgh|gfedcba +#define ADDR_L(i, l_edge, r_edge) ((i) < (l_edge) ? -(i) : (i)) +#define ADDR_R(i, r_edge, addr) ((i) >= (r_edge) ? -(i)-2+((r_edge)<<1) : (addr)) +#define ADDR_H(i, t_edge, b_edge) ((i) < (t_edge) ? -(i) : (i)) +#define ADDR_B(i, b_edge, addr) ((i) >= (b_edge) ? -(i)-2+((b_edge)<<1) : (addr)) +#endif + +//blur function does not support BORDER_WRAP +#ifdef BORDER_WRAP +//BORDER_WRAP: cdefgh|abcdefgh|abcdefg +#define ADDR_L(i, l_edge, r_edge) ((i) < (l_edge) ? (i)+(r_edge) : (i)) +#define ADDR_R(i, r_edge, addr) ((i) >= (r_edge) ? (i)-(r_edge) : (addr)) +#define ADDR_H(i, t_edge, b_edge) ((i) < (t_edge) ? (i)+(b_edge) : (i)) +#define ADDR_B(i, b_edge, addr) ((i) >= (b_edge) ? (i)-(b_edge) : (addr)) +#endif + +#ifdef EXTRA_EXTRAPOLATION // border > src image size +#ifdef BORDER_CONSTANT +// None +#elif defined BORDER_REPLICATE +#define EXTRAPOLATE(x, y, minX, minY, maxX, maxY) \ + { \ + x = max(min(x, maxX - 1), minX); \ + y = max(min(y, maxY - 1), minY); \ + } +#elif defined BORDER_WRAP +#define EXTRAPOLATE(x, y, minX, minY, maxX, maxY) \ + { \ + if (x < minX) \ + x -= ((x - maxX + 1) / maxX) * maxX; \ + if (x >= maxX) \ + x %= maxX; \ + if (y < minY) \ + y -= ((y - maxY + 1) / maxY) * maxY; \ + if (y >= maxY) \ + y %= maxY; \ + } +#elif defined(BORDER_REFLECT) || defined(BORDER_REFLECT_101) +#define EXTRAPOLATE_(x, y, minX, minY, maxX, maxY, delta) \ + { \ + if (maxX - minX == 1) \ + x = minX; \ + else \ + do \ + { \ + if (x < minX) \ + x = -(x - minX) - 1 + delta; \ + else \ + x = maxX - 1 - (x - maxX) - delta; \ + } \ + while (x >= maxX || x < minX); \ + \ + if (maxY - minY == 1) \ + y = minY; \ + else \ + do \ + { \ + if (y < minY) \ + y = -(y - minY) - 1 + delta; \ + else \ + y = maxY - 1 - (y - maxY) - delta; \ + } \ + while (y >= maxY || y < minY); \ + } +#ifdef BORDER_REFLECT +#define EXTRAPOLATE(x, y, minX, minY, maxX, maxY) EXTRAPOLATE_(x, y, minX, minY, maxX, maxY, 0) +#elif defined(BORDER_REFLECT_101) +#define EXTRAPOLATE(x, y, minX, minY, maxX, maxY) EXTRAPOLATE_(x, y, minX, minY, maxX, maxY, 1) +#endif +#else +#error No extrapolation method +#endif +#else +#define EXTRAPOLATE(x, y, minX, minY, maxX, maxY) \ + { \ + int _row = y - minY, _col = x - minX; \ + _row = ADDR_H(_row, 0, maxY - minY); \ + _row = ADDR_B(_row, maxY - minY, _row); \ + y = _row + minY; \ + \ + _col = ADDR_L(_col, 0, maxX - minX); \ + _col = ADDR_R(_col, maxX - minX, _col); \ + x = _col + minX; \ + } +#endif + +#if USE_DOUBLE +#pragma OPENCL EXTENSION cl_khr_fp64:enable +#define FPTYPE double +#define CONVERT_TO_FPTYPE CAT(convert_double, VEC_SIZE) +#else +#define FPTYPE float +#define CONVERT_TO_FPTYPE CAT(convert_float, VEC_SIZE) +#endif + +#if DATA_DEPTH == 0 +#define BASE_TYPE uchar +#elif DATA_DEPTH == 1 +#define BASE_TYPE char +#elif DATA_DEPTH == 2 +#define BASE_TYPE ushort +#elif DATA_DEPTH == 3 +#define BASE_TYPE short +#elif DATA_DEPTH == 4 +#define BASE_TYPE int +#elif DATA_DEPTH == 5 +#define BASE_TYPE float +#elif DATA_DEPTH == 6 +#define BASE_TYPE double +#else +#error data_depth +#endif + +#define __CAT(x, y) x##y +#define CAT(x, y) __CAT(x, y) + +#define uchar1 uchar +#define char1 char +#define ushort1 ushort +#define short1 short +#define int1 int +#define float1 float +#define double1 double + +#define convert_uchar1_sat_rte convert_uchar_sat_rte +#define convert_char1_sat_rte convert_char_sat_rte +#define convert_ushort1_sat_rte convert_ushort_sat_rte +#define convert_short1_sat_rte convert_short_sat_rte +#define convert_int1_sat_rte convert_int_sat_rte +#define convert_float1 +#define convert_double1 + +#if DATA_DEPTH == 5 || DATA_DEPTH == 6 +#define CONVERT_TO_TYPE CAT(CAT(convert_, BASE_TYPE), VEC_SIZE) +#else +#define CONVERT_TO_TYPE CAT(CAT(CAT(convert_, BASE_TYPE), VEC_SIZE), _sat_rte) +#endif + +#define VEC_SIZE DATA_CHAN + +#define VEC_TYPE CAT(BASE_TYPE, VEC_SIZE) +#define TYPE VEC_TYPE + +#define SCALAR_TYPE CAT(FPTYPE, VEC_SIZE) + +#define INTERMEDIATE_TYPE CAT(FPTYPE, VEC_SIZE) + +struct RectCoords +{ + int x1, y1, x2, y2; +}; + +//#define DEBUG +#ifdef DEBUG +#define DEBUG_ONLY(x) x +#define ASSERT(condition) do { if (!(condition)) { printf("BUG in boxFilter kernel (global=%d,%d): " #condition "\n", get_global_id(0), get_global_id(1)); } } while (0) +#else +#define DEBUG_ONLY(x) (void)0 +#define ASSERT(condition) (void)0 +#endif + + +inline INTERMEDIATE_TYPE readSrcPixel(int2 pos, __global TYPE *src, const unsigned int srcStepBytes, const struct RectCoords srcCoords +#ifdef BORDER_CONSTANT + , SCALAR_TYPE borderValue +#endif + ) +{ +#ifdef BORDER_ISOLATED + if(pos.x >= srcCoords.x1 && pos.y >= srcCoords.y1 && pos.x < srcCoords.x2 && pos.y < srcCoords.y2) +#else + if(pos.x >= 0 && pos.y >= 0 && pos.x < srcCoords.x2 && pos.y < srcCoords.y2) +#endif + { + __global TYPE* ptr = (__global TYPE*)((__global char*)src + pos.x * sizeof(TYPE) + pos.y * srcStepBytes); + return CONVERT_TO_FPTYPE(*ptr); + } + else + { +#ifdef BORDER_CONSTANT + return borderValue; +#else + int selected_col = pos.x; + int selected_row = pos.y; + + EXTRAPOLATE(selected_col, selected_row, +#ifdef BORDER_ISOLATED + srcCoords.x1, srcCoords.y1, +#else + 0, 0, +#endif + srcCoords.x2, srcCoords.y2 + ); + + // debug border mapping + //printf("pos=%d,%d --> %d, %d\n", pos.x, pos.y, selected_col, selected_row); + + pos = (int2)(selected_col, selected_row); + if(pos.x >= 0 && pos.y >= 0 && pos.x < srcCoords.x2 && pos.y < srcCoords.y2) + { + __global TYPE* ptr = (__global TYPE*)((__global char*)src + pos.x * sizeof(TYPE) + pos.y * srcStepBytes); + return CONVERT_TO_FPTYPE(*ptr); + } + else + { + // for debug only + DEBUG_ONLY(printf("BUG in boxFilter kernel\n")); + return (FPTYPE)(0.0f); + } +#endif + } +} + +// INPUT PARAMETER: BLOCK_SIZE_Y (via defines) + +__kernel +__attribute__((reqd_work_group_size(LOCAL_SIZE, 1, 1))) +void filter2D(__global TYPE *src, const unsigned int srcStepBytes, const int4 srcRC, + __global TYPE *dst, const unsigned int dstStepBytes, const int4 dstRC, +#ifdef BORDER_CONSTANT + SCALAR_TYPE borderValue, +#endif + __constant FPTYPE* kernelData // transposed: [KERNEL_SIZE_X][KERNEL_SIZE_Y2_ALIGNED] + ) +{ + const struct RectCoords srcCoords = {srcRC.s0, srcRC.s1, srcRC.s2, srcRC.s3}; // for non-isolated border: offsetX, offsetY, wholeX, wholeY + struct RectCoords dstCoords = {dstRC.s0, dstRC.s1, dstRC.s2, dstRC.s3}; + + const int local_id = get_local_id(0); + const int x = local_id + (LOCAL_SIZE - (KERNEL_SIZE_X - 1)) * get_group_id(0) - ANCHOR_X; + const int y = get_global_id(1) * BLOCK_SIZE_Y; + + INTERMEDIATE_TYPE data[KERNEL_SIZE_Y]; + __local INTERMEDIATE_TYPE sumOfCols[LOCAL_SIZE]; + + int2 srcPos = (int2)(srcCoords.x1 + x, srcCoords.y1 + y - ANCHOR_Y); + + int2 pos = (int2)(dstCoords.x1 + x, dstCoords.y1 + y); + __global TYPE* dstPtr = (__global TYPE*)((__global char*)dst + pos.x * sizeof(TYPE) + pos.y * dstStepBytes); // Pointer can be out of bounds! + bool writeResult = (local_id >= ANCHOR_X && local_id < LOCAL_SIZE - (KERNEL_SIZE_X - 1 - ANCHOR_X) && + pos.x >= dstCoords.x1 && pos.x < dstCoords.x2); + +#if BLOCK_SIZE_Y > 1 + bool readAllpixels = true; + int sy_index = 0; // current index in data[] array + + dstCoords.y2 = min(dstCoords.y2, pos.y + BLOCK_SIZE_Y); + for (; + pos.y < dstCoords.y2; + pos.y++, + dstPtr = (__global TYPE*)((__global char*)dstPtr + dstStepBytes)) +#endif + { + ASSERT(pos.y < dstCoords.y2); + + for ( +#if BLOCK_SIZE_Y > 1 + int sy = readAllpixels ? 0 : -1; sy < (readAllpixels ? KERNEL_SIZE_Y : 0); +#else + int sy = 0, sy_index = 0; sy < KERNEL_SIZE_Y; +#endif + sy++, srcPos.y++) + { + data[sy + sy_index] = readSrcPixel(srcPos, src, srcStepBytes, srcCoords +#ifdef BORDER_CONSTANT + , borderValue +#endif + ); + } + + INTERMEDIATE_TYPE total_sum = 0; + for (int sx = 0; sx < KERNEL_SIZE_X; sx++) + { + { + __constant FPTYPE* k = &kernelData[KERNEL_SIZE_Y2_ALIGNED * sx +#if BLOCK_SIZE_Y > 1 + + KERNEL_SIZE_Y - sy_index +#endif + ]; + INTERMEDIATE_TYPE tmp_sum = 0; + for (int sy = 0; sy < KERNEL_SIZE_Y; sy++) + { + tmp_sum += data[sy] * k[sy]; + } + + sumOfCols[local_id] = tmp_sum; + barrier(CLK_LOCAL_MEM_FENCE); + } + + int id = local_id + sx - ANCHOR_X; + if (id >= 0 && id < LOCAL_SIZE) + total_sum += sumOfCols[id]; + + barrier(CLK_LOCAL_MEM_FENCE); + } + + if (writeResult) + { + ASSERT(pos.y >= dstCoords.y1 && pos.y < dstCoords.y2); + *dstPtr = CONVERT_TO_TYPE(total_sum); + } + +#if BLOCK_SIZE_Y > 1 + readAllpixels = false; +#if BLOCK_SIZE_Y > KERNEL_SIZE_Y + sy_index = (sy_index + 1 <= KERNEL_SIZE_Y) ? sy_index + 1 : 1; +#else + sy_index++; +#endif +#endif // BLOCK_SIZE_Y == 1 + } +} diff --git a/modules/ocl/src/opencl/filtering_laplacian.cl b/modules/ocl/src/opencl/filtering_laplacian.cl deleted file mode 100644 index ea22967df..000000000 --- a/modules/ocl/src/opencl/filtering_laplacian.cl +++ /dev/null @@ -1,381 +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 -// Pang Erping, erping@multicorewareinc.com -// Jia Haipeng, jiahaipeng95@gmail.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*/ - -/////////////////////////////////////////////////////////////////////////////////////////////////// -/////////////////////////////////Macro for border type//////////////////////////////////////////// -///////////////////////////////////////////////////////////////////////////////////////////////// -#ifdef BORDER_REPLICATE - -//BORDER_REPLICATE: aaaaaa|abcdefgh|hhhhhhh -#define ADDR_L(i, l_edge, r_edge) ((i) < (l_edge) ? (l_edge) : (i)) -#define ADDR_R(i, r_edge, addr) ((i) >= (r_edge) ? (r_edge)-1 : (addr)) -#define ADDR_H(i, t_edge, b_edge) ((i) < (t_edge) ? (t_edge) : (i)) -#define ADDR_B(i, b_edge, addr) ((i) >= (b_edge) ? (b_edge)-1 :(addr)) -#endif - -#ifdef BORDER_REFLECT -#define ADDR_L(i, l_edge, r_edge) ((i) < (l_edge) ? ((l_edge)<<1)-(i)-1 : (i)) -#define ADDR_R(i, r_edge, addr) ((i) >= (r_edge) ? -(i)-1+((r_edge)<<1) : (addr)) -#define ADDR_H(i, t_edge, b_edge) ((i) < (t_edge) ? ((t_edge)<<1)-(i)-1 : (i)) -#define ADDR_B(i, b_edge, addr) ((i) >= (b_edge) ? -(i)-1+((b_edge)<<1) : (addr)) -#endif - -#ifdef BORDER_REFLECT_101 -//BORDER_REFLECT_101: gfedcb|abcdefgh|gfedcba -#define ADDR_L(i, l_edge, r_edge) ((i) < (l_edge) ? ((l_edge)<<1)-(i) : (i)) -#define ADDR_R(i, r_edge, addr) ((i) >= (r_edge) ? -(i)-2+((r_edge)<<1) : (addr)) -#define ADDR_H(i, t_edge, b_edge) ((i) < (t_edge) ? ((t_edge)<<1)-(i) : (i)) -#define ADDR_B(i, b_edge, addr) ((i) >= (b_edge) ? -(i)-2+((b_edge)<<1) : (addr)) -#endif - -#ifdef IMG_C_1_0 -#define T_IMG uchar -#define T_IMGx4 uchar4 -#define T_IMG_C1 uchar -#define CONVERT_TYPE convert_uchar_sat -#define CONVERT_TYPEx4 convert_uchar4_sat -#endif -#ifdef IMG_C_4_0 -#define T_IMG uchar4 -#define T_IMGx4 uchar16 -#define T_IMG_C1 uchar -#define CONVERT_TYPE convert_uchar4_sat -#define CONVERT_TYPEx4 convert_uchar16_sat -#endif -#ifdef IMG_C_1_5 -#define T_IMG float -#define T_IMGx4 float4 -#define T_IMG_C1 float -#define CONVERT_TYPE convert_float -#define CONVERT_TYPEx4 convert_float4 -#endif -#ifdef IMG_C_4_5 -#define T_IMG float4 -#define T_IMGx4 float16 -#define T_IMG_C1 float -#define CONVERT_TYPE convert_float4 -#define CONVERT_TYPEx4 convert_float16 -#endif - -#ifndef CN -#define CN 1 -#endif - -#if CN == 1 -#define T_SUM float -#define T_SUMx4 float4 -#define CONVERT_TYPE_SUM convert_float -#define CONVERT_TYPE_SUMx4 convert_float4 -#define SUM_ZERO (0.0f) -#define SUM_ZEROx4 (0.0f, 0.0f, 0.0f, 0.0f) -#define VLOAD4 vload4 -#define SX x -#define SY y -#define SZ z -#define SW w -#elif CN == 4 -#define T_SUM float4 -#define T_SUMx4 float16 -#define CONVERT_TYPE_SUM convert_float4 -#define CONVERT_TYPE_SUMx4 convert_float16 -#define SUM_ZERO (0.0f, 0.0f, 0.0f, 0.0f) -#define SUM_ZEROx4 (0.0f, 0.0f, 0.0f, 0.0f, 0.0f, 0.0f, 0.0f, 0.0f, 0.0f, 0.0f, 0.0f, 0.0f, 0.0f, 0.0f, 0.0f, 0.0f) -#define VLOAD4 vload16 -#define SX s0123 -#define SY s4567 -#define SZ s89ab -#define SW scdef -#endif - -#ifndef FILTER_SIZE -#define FILTER_SIZE 3 -#endif - -#define LOCAL_GROUP_SIZE 16 - -#define LOCAL_WIDTH ((FILTER_SIZE/2)*2 + LOCAL_GROUP_SIZE) -#define LOCAL_HEIGHT ((FILTER_SIZE/2)*2 + LOCAL_GROUP_SIZE) - -#define FILTER_RADIUS (FILTER_SIZE >> 1) - -__kernel void filter2D( - __global T_IMG *src, - __global T_IMG *dst, - int src_step, - int dst_step, - __constant float *mat_kernel, - __local T_IMG *local_data, - int wholerows, - int wholecols, - int src_offset_x, - int src_offset_y, - int dst_offset_x, - int dst_offset_y, - int cols, - int rows, - int operate_cols -) -{ - int groupStartCol = get_group_id(0) * get_local_size(0); - int groupStartRow = get_group_id(1) * get_local_size(1); - - int localCol = get_local_id(0); - int localRow = get_local_id(1); - int globalCol = groupStartCol + localCol; - int globalRow = groupStartRow + localRow; - const int src_offset = mad24(src_offset_y, src_step, src_offset_x); - const int dst_offset = mad24(dst_offset_y, dst_step, dst_offset_x); - -#ifdef BORDER_CONSTANT - for(int i = localRow; i < LOCAL_HEIGHT; i += get_local_size(1)) - { - int curRow = groupStartRow + i; - for(int j = localCol; j < LOCAL_WIDTH; j += get_local_size(0)) - { - int curCol = groupStartCol + j; - if(curRow < FILTER_RADIUS - src_offset_y || (curRow - FILTER_RADIUS) >= wholerows - src_offset_y|| - curCol < FILTER_RADIUS - src_offset_x || (curCol - FILTER_RADIUS) >= wholecols - src_offset_x) - { - local_data[(i) * LOCAL_WIDTH + j] = 0; - } - else - { - local_data[(i) * LOCAL_WIDTH + j] = src[(curRow - FILTER_RADIUS) * src_step + curCol - FILTER_RADIUS + src_offset]; - } - } - } -#else - for(int i = localRow; i < LOCAL_HEIGHT; i += get_local_size(1)) - { - int curRow = groupStartRow + i; - - curRow = ADDR_H(curRow, FILTER_RADIUS - src_offset_y, wholerows - src_offset_y); - - curRow = ADDR_B(curRow - FILTER_RADIUS, wholerows - src_offset_y, curRow - FILTER_RADIUS); - - for(int j = localCol; j < LOCAL_WIDTH; j += get_local_size(0)) - { - int curCol = groupStartCol + j; - curCol = ADDR_L(curCol, FILTER_RADIUS - src_offset_x, wholecols - src_offset_x); - curCol = ADDR_R(curCol - FILTER_RADIUS, wholecols - src_offset_x, curCol - FILTER_RADIUS); - if(curRow < wholerows && curCol < wholecols) - { - local_data[(i) * LOCAL_WIDTH + j] = src[(curRow) * src_step + curCol + src_offset]; - } - } - } -#endif - - barrier(CLK_LOCAL_MEM_FENCE); - if(globalRow < rows && globalCol < cols) - { - T_SUM sum = (T_SUM)(SUM_ZERO); - int filterIdx = 0; - for(int i = 0; i < FILTER_SIZE; i++) - { - int offset = (i + localRow) * LOCAL_WIDTH; - - for(int j = 0; j < FILTER_SIZE; j++) - { - sum += CONVERT_TYPE_SUM(local_data[offset + j + localCol]) * mat_kernel[filterIdx++]; - } - } - dst[(globalRow)*dst_step + (globalCol) + dst_offset] = CONVERT_TYPE(sum); - } -} - -/// following is specific for 3x3 kernels - -////////////////////////////////////////////////////////////////////////////////////////////////////// -/////////////////////////////Macro for define elements number per thread///////////////////////////// -//////////////////////////////////////////////////////////////////////////////////////////////////// - -#define ANX 1 -#define ANY 1 - -#define ROWS_PER_GROUP 4 -#define ROWS_PER_GROUP_BITS 2 -#define ROWS_FETCH (ROWS_PER_GROUP + ANY + ANY) //(ROWS_PER_GROUP + anY * 2) - -#define THREADS_PER_ROW 64 -#define THREADS_PER_ROW_BIT 6 - -#define ELEMENTS_PER_THREAD 4 -#define ELEMENTS_PER_THREAD_BIT 2 - -#define LOCAL_MEM_STEP 260 //divup((get_local_size(0) + anX * 2), 4) * 4 - -/////////////////////////////////////////////////////////////////////////////////////////////////// -/////////////////////////////////////////8uC1//////////////////////////////////////////////////////// -//////////////////////////////////////////////////////////////////////////////////////////////////// - -__kernel void filter2D_3x3( - __global T_IMG *src, - __global T_IMG *dst, - int src_step, - int dst_step, - __constant float *mat_kernel, - __local T_IMG *local_data, - int wholerows, - int wholecols, - int src_offset_x, - int src_offset_y, - int dst_offset_x, - int dst_offset_y, - int cols, - int rows, - int operate_cols -) -{ - int gX = get_global_id(0); - int gY = get_global_id(1); - - int lX = get_local_id(0); - - int groupX_size = get_local_size(0); - int groupX_id = get_group_id(0); - -#define dst_align (dst_offset_x & 3) - int cols_start_index_group = src_offset_x - dst_align + groupX_size * groupX_id - ANX; - int rows_start_index = src_offset_y + (gY << ROWS_PER_GROUP_BITS) - ANY; - - if((gY << 2) < rows) - { - for(int i = 0; i < ROWS_FETCH; ++i) - { - if((rows_start_index - src_offset_y) + i < rows + ANY) - { -#ifdef BORDER_CONSTANT - int selected_row = rows_start_index + i; - int selected_cols = cols_start_index_group + lX; - - T_IMG data = src[mad24(selected_row, src_step, selected_cols)]; - int con = selected_row >= 0 && selected_row < wholerows && selected_cols >= 0 && selected_cols < wholecols; - data = con ? data : (T_IMG)(0); - local_data[mad24(i, LOCAL_MEM_STEP, lX)] = data; - - if(lX < (ANX << 1)) - { - selected_cols = cols_start_index_group + lX + groupX_size; - - data = src[mad24(selected_row, src_step, selected_cols)]; - con = selected_row >= 0 && selected_row < wholerows && selected_cols >= 0 && selected_cols < wholecols; - data = con ? data : (T_IMG)(0); - local_data[mad24(i, LOCAL_MEM_STEP, lX) + groupX_size] = data; - } -#else - int selected_row = ADDR_H(rows_start_index + i, 0, wholerows); - selected_row = ADDR_B(rows_start_index + i, wholerows, selected_row); - - int selected_cols = ADDR_L(cols_start_index_group + lX, 0, wholecols); - selected_cols = ADDR_R(cols_start_index_group + lX, wholecols, selected_cols); - - T_IMG data = src[mad24(selected_row, src_step, selected_cols)]; - - local_data[mad24(i, LOCAL_MEM_STEP, lX)] = data; - - if(lX < (ANX << 1)) - { - selected_cols = cols_start_index_group + lX + groupX_size; - selected_cols = ADDR_R(selected_cols, wholecols, selected_cols); - - data = src[mad24(selected_row, src_step, selected_cols)]; - local_data[mad24(i, LOCAL_MEM_STEP, lX) + groupX_size] = data; - } -#endif - } - } - } - barrier(CLK_LOCAL_MEM_FENCE); - - int process_col = groupX_size * groupX_id + ((lX % THREADS_PER_ROW) << 2); - if(((gY << 2) < rows) && (process_col < operate_cols)) - { - int dst_cols_start = dst_offset_x; - int dst_cols_end = dst_offset_x + cols; - int dst_cols_index = (dst_offset_x + process_col) & 0xfffffffc; - - int dst_rows_end = dst_offset_y + rows; - int dst_rows_index = dst_offset_y + (gY << ROWS_PER_GROUP_BITS) + (lX >> THREADS_PER_ROW_BIT); - dst = dst + mad24(dst_rows_index, dst_step, dst_cols_index); - - T_IMGx4 dst_data = *(__global T_IMGx4 *)dst; - - T_SUMx4 sum = (T_SUMx4)SUM_ZEROx4; - T_IMGx4 data; - - for(int i = 0; i < FILTER_SIZE; i++) - { -#pragma unroll - for(int j = 0; j < FILTER_SIZE; j++) - { - if(dst_rows_index < dst_rows_end) - { - int local_row = (lX >> THREADS_PER_ROW_BIT) + i; - int local_cols = ((lX % THREADS_PER_ROW) << ELEMENTS_PER_THREAD_BIT) + j; - - data = VLOAD4(0, (__local T_IMG_C1 *)(local_data + local_row * LOCAL_MEM_STEP + local_cols)); - sum = sum + (mat_kernel[i * FILTER_SIZE + j] * CONVERT_TYPE_SUMx4(data)); - } - } - } - - if(dst_rows_index < dst_rows_end) - { - T_IMGx4 tmp_dst = CONVERT_TYPEx4(sum); - tmp_dst.SX = ((dst_cols_index + 0 >= dst_cols_start) && (dst_cols_index + 0 < dst_cols_end)) ? - tmp_dst.SX : dst_data.SX; - tmp_dst.SY = ((dst_cols_index + 1 >= dst_cols_start) && (dst_cols_index + 1 < dst_cols_end)) ? - tmp_dst.SY : dst_data.SY; - tmp_dst.SZ = ((dst_cols_index + 2 >= dst_cols_start) && (dst_cols_index + 2 < dst_cols_end)) ? - tmp_dst.SZ : dst_data.SZ; - tmp_dst.SW = ((dst_cols_index + 3 >= dst_cols_start) && (dst_cols_index + 3 < dst_cols_end)) ? - tmp_dst.SW : dst_data.SW; - *(__global T_IMGx4 *)dst = tmp_dst; - } - } -} diff --git a/modules/ocl/test/test_filters.cpp b/modules/ocl/test/test_filters.cpp index 3cf7d37b8..a8583b28a 100644 --- a/modules/ocl/test/test_filters.cpp +++ b/modules/ocl/test/test_filters.cpp @@ -160,8 +160,8 @@ OCL_TEST_P(LaplacianTest, Accuracy) { random_roi(); - Laplacian(src_roi, dst_roi, -1, ksize, scale); // TODO FIXIT , 0, borderType); - ocl::Laplacian(gsrc_roi, gdst_roi, -1, ksize, scale); // TODO FIXIT , 0, borderType); + Laplacian(src_roi, dst_roi, -1, ksize, scale, 0, borderType); + ocl::Laplacian(gsrc_roi, gdst_roi, -1, ksize, scale, 0, borderType); Near(); } @@ -298,7 +298,7 @@ OCL_TEST_P(Filter2D, Mat) kernel *= 1.0 / (double)(ksize * ksize); cv::filter2D(src_roi, dst_roi, -1, kernel, anchor, 0.0, borderType); - ocl::filter2D(gsrc_roi, gdst_roi, -1, kernel, anchor, /* TODO FIXIT 0.0,*/ borderType); + ocl::filter2D(gsrc_roi, gdst_roi, -1, kernel, anchor, 0.0, borderType); Near(); }